Disclosure of Invention
The purpose is as follows: in order to overcome the defects in the prior art, the invention provides a remote fault early warning method and a remote fault early warning device which are suitable for an electric automobile charging process, so that the problem of fault early warning in the electric automobile charging process is solved.
The technical scheme is as follows: in order to solve the technical problems, the technical scheme adopted by the invention is as follows:
in a first aspect, a remote fault early warning method suitable for an electric vehicle charging process includes the following steps:
and acquiring related data of the charging process.
And preprocessing the obtained relevant data of the charging process to obtain the preprocessed relevant data of the charging process.
And judging the relevant data of the preprocessed charging process, and judging whether the charging process is abnormal.
When the charging process of the electric automobile is abnormal, fault diagnosis is carried out on the charging process of the electric automobile, and the fault type of the charging process is diagnosed.
And matching the specific fault of the current charging according to the fault type through a fault data knowledge base.
As a preferred scheme, the determining the relevant data of the pre-processed charging process to determine whether the charging process is abnormal includes the following steps:
and training the modal learning network, acquiring parameters of the modal learning network, substituting the parameters of the modal learning network into the modal learning network, and acquiring the trained modal learning network.
And training the modal discrimination network, acquiring parameters of the modal discrimination network, substituting the parameters of the modal discrimination network into the modal discrimination network, and acquiring the trained modal discrimination network.
And constructing a modal early warning network according to the trained modal learning network and the trained modal discrimination network, training the modal early warning network, acquiring parameters of the modal early warning network, substituting the parameters of the modal early warning network into the modal early warning network to obtain the trained modal early warning network, inputting relevant data of the preprocessed charging process into the trained modal early warning network, and judging whether the charging process is abnormal.
As a preferred scheme, the method for training the modal learning network to obtain the parameters of the modal learning network and substituting the parameters of the modal learning network into the modal learning network to obtain the trained modal learning network comprises the following steps:
step 1: initializing a modal learning network unit, including the number M of neurons contained in the input layer
in The number of neurons in the a-th hidden layer
Number M of neurons included in output layer
O A =1,2, 3., a denotes the total number of hidden layers of the modal learning network.
And 2, step: initializing input weight matrix of a-th hidden layer
When a =1, is selected>
The size of the matrix is
An input weight matrix representing a first hidden layer; when a > 1, is selected>
The size of the matrix is
An input weight matrix representing the a-th hidden layer.
And step 3: computing the output matrix P of the a-th hidden layer
a (ii) a When a =1, the number of the bits is set to a =1,
when a is greater than 1, the ratio of a,
wherein f is an activation function>
Normalizing the data input by the modal learning network in the training process and corresponding to the charging process corresponding to the abnormal charging process, wherein the normalized data comprise the charging voltage, the charging current, the charging power, the charging module temperature, the charging gun temperature, the power battery monomer temperature, the voltage and the current, T is the number of the electric vehicles, M
in Number of correlated data after normalization for correlated data of charging process, P
a-1 The output matrix representing the a-1 st hidden layer.
And 4, step 4: estimating output weight matrix W of output layer by pseudo-inverse method
out I.e. W
out = Y pinv (P). Wherein, pinv (#) represents a pseudo-inverse calculation,
an output target matrix of the network in the training process is learned for the mode,
the output matrix synthesis set of the input layer and the hidden layer is obtained. Training output of the modal learning network->
Is [1, 1., 1 ]]And representing that in the modal learning network, the input relevant electrical sampling value data is defined as the absence of the early warning information, and 1 represents that the charging state of the electric vehicle is normal.
And 5: and substituting the parameters of the modal learning network into the modal learning network to obtain the trained modal learning network.
As a preferred scheme, the method for training the modal discrimination network to obtain the parameters of the modal discrimination network and substituting the parameters of the modal discrimination network into the modal discrimination network to obtain the trained modal discrimination network comprises the following steps:
step 1: the initialization mode discrimination network unit comprises a neuron number N contained in an input layer
in The number N of neurons included in the hidden layer
H And the number N of neurons included in the output layer
O Input weight matrix of mode discrimination network input layer
And 2, step: computing the output matrix of the hidden layer by
I.e. based on>
Wherein f is an activation function; />
Y is the training output of the modal learning network, X
normal And the related data of the corresponding charging process representing the normal charging process comprise charging voltage, charging current, charging power, charging module temperature, charging gun temperature, and power battery monomer temperature, voltage and current. T is the number of electric vehicles, N
in Number of relevant data for a corresponding charging process which is a normal charging process>
Representing a matrix transpose.
And step 3: estimating an output weight matrix of a modal discrimination network output layer by solving the following optimization problem J
Namely:
wherein ,Y
* Representing an output target matrix of the mode discrimination network in the training process, | | × | zero calculation
2 Is represented by
2 Norm, α
i,j Connecting weight of ith neuron representing hidden layer of modal discrimination network and jth neuron of output layer
The regularization coefficients of (a). The mode discrimination network outputs a target matrix ^ according to the training process>
Is [1, 1., 1, 0., 0 ]]In the mode discrimination network, the input related electrical sampling value is defined as containing early warning information, 1 represents that the charging state of the electric automobile is normal, and 0 represents that the charging state of the electric automobile is abnormal.
And 4, step 4: and substituting the parameters of the modal discrimination network into the modal discrimination network to obtain the trained modal discrimination network.
As a preferred scheme, the method comprises the following steps of constructing a modal early warning network according to a trained modal learning network and a modal discrimination network, training the modal early warning network, acquiring parameters of the modal early warning network, substituting the parameters of the modal early warning network into the modal early warning network to obtain the trained modal early warning network, inputting relevant data of a charging process into the trained modal early warning network, and judging whether the charging process is abnormal, wherein the method comprises the following steps:
step 1: the initialization mode early warning network unit comprises a neuron number M contained in an input layer
in The number of neurons contained in the a-th hidden layer
a =1,2, 3.., a, number of neurons M contained in the a +1 st hidden layer
O And the number N of neurons contained in the A +2 th hidden layer
H And outputNumber of neurons N contained in layer
O 。
And 2, step: initializing a connection weight matrix W of an input layer and a 1 st hidden layer of a modal early warning network unit
in 1,2 A-1 th hidden layer and a connection weight matrix W of the a-th hidden layer
in a,a+1 A is more than or equal to 2 and less than A-1, and the connection weight matrix W of the A-th hidden layer and the A + 1-th hidden layer
in A,A+1 The connection weight matrix W of the A +1 th hidden layer and the A +2 th hidden layer
in A+1,A+2 Output weight matrix of output layer
In the process of constructing the modal early warning network, the connection weight matrix W of the input layer and the 1 st hidden layer
in 1,2 Input weight matrix equal to modal learning network input layer and 1 st hidden layer
Connection weight matrix W of a-1 th hidden layer and a-th hidden layer
in a,a+1 Connection weight matrix ≧ which equals the a-1 th hidden layer of the modal learning network>
A is more than or equal to 2 and less than A-1, and the connection weight matrix W of the A-th hidden layer and the A + 1-th hidden layer
in A,A+1 Output weight matrix W equal to output layer of modal learning network
out The connection weight matrix W of the A +1 th hidden layer and the A +2 th hidden layer
in A+1,A+2 Input weight matrix equal to mode decision network input layer->
And step 3: and substituting the parameters of the modal early warning network into the modal early warning network to obtain the trained modal early warning network.
And 4, step 4: and acquiring related data of the current electric vehicle charging process, taking the related data of the current electric vehicle charging process as input, and calculating the trained modal early warning network in a forward direction to obtain an early warning result of the charging process.
Preferably, the output weight matrix of the mode discrimination network
Is selected based on the output weight matrix of the jth neuron of (a)>
An estimation process, comprising:
wherein ,Y
j * Expressed as a target matrix of the j-th neuron of the corresponding output layer in the training process of the modal discrimination network,
an output weight matrix representing the jth neuron of the output layer corresponding to the modal discrimination network, and->
An output weight element, alpha, representing the ith neuron of the A +2 th hidden layer and the jth neuron of the output layer in the modal discrimination network
i,j And a regularization coefficient of an output weight element of an ith neuron of an A +2 th hidden layer and a jth neuron of an output layer in the modal discrimination network, wherein A represents the total number of the hidden layers of the modal learning network.
Modal discrimination network output weight matrix
Is selected based on the output weight matrix of the jth neuron of (a)>
The estimation process is further derived as:
wherein ,Y
1 * Expressed as the target matrix of the 1 st neuron of the mode discrimination network corresponding to the output layer in the training process,
an output weight matrix representing the 1 st neuron of the corresponding output layer of the modal discrimination network, and->
Output weight element, alpha, representing the ith neuron of the A +2 th hidden layer and the 1 st neuron of the output layer in the modal discrimination network
i,1 The regularization coefficients represent output weight elements of the ith neuron of an A +2 th hidden layer and the 1 st neuron of an output layer in the modal discrimination network; y is
2 * A target matrix, expressed as the 2 nd neuron of the modal discrimination network corresponding to the output layer during the training process, is asserted>
An output weight matrix representing the 2 nd neuron of the corresponding output layer of the modal discrimination network, and->
Output weight element, α, representing the ith neuron of the A +2 th hidden layer and the 2 nd neuron of the output layer in the modal discrimination network
i,2 The regularization coefficients represent output weight elements of the ith neuron of an A +2 th hidden layer and the 2 nd neuron of an output layer in the modal discrimination network; />
Expressed as N corresponding to output layer of modal discrimination network in training process
H A target matrix of individual neurons, based on the evaluation of the neuron>
Representing modalitiesJudging the Nth output layer corresponding to the network
H An output weight matrix for each neuron, < > >>
Representing the ith neuron of the A +2 hidden layer and the Nth neuron of the output layer in the modal discrimination network
H An output weight element for each neuron, < > or >>
Representing the ith neuron of the A +2 hidden layer and the Nth neuron of the output layer in the modal discrimination network
H Regularization coefficients of output weight elements of the individual neurons.
Modal discrimination network estimation output weight matrix
Is selected based on the output weight matrix of the jth neuron of (a)>
Can be described as solving for N for J
H An optimization problem, namely:
wherein ,Jj Denotes the jth optimization problem, j =1,2,3 H (ii) a The jth optimization problem can be further derived as:
wherein ,
the revised output weight matrix of the jth neuron of the corresponding output layer of the representation mode discrimination network is used for being greater than or equal to>
Indicating mouldThe state judges the revised output matrix of the network hidden layer, | | × | count
1/2 Represents l
1/2 And (4) norm. Y is
j * And expressing the target matrix of the j-th neuron corresponding to the output layer in the training process of the modal discrimination network.
The solving process of the jth optimization problem is as follows:
step 1: initializing a weight matrix
And zero matrix->
and step 3: label i =1 is defined.
And 4, step 4: judging whether i is equal to N O (ii) a If not, i = i +1, and step 5 is entered; if so, go to step 7.
step 6: calculated by the following formula
Namely that
wherein ,Δi 、Α i and Βi The expression is:
wherein ,Y
j * (t) revised data of a target output matrix of a jth neuron of an output layer expressing the tth input data,
the revised output data of the qth neuron of the t input data in the hidden layer of the modal discrimination network is represented, and arccos (—) represents an arccosine function; t is the number of electric automobiles. />
And the revised output weight matrix represents the j-th neuron of the corresponding output layer of the q-th neuron of the hidden layer of the modal discrimination network.
And 7: judgment of
If yes, executing step 8, otherwise, executing step 4.
And 8: output weight matrix
As a preferred scheme, the charging data of the electric vehicle and the charging pile 1-3 times before and after the current charging is used as the data monitoring range in the fault diagnosis process as the relevant data of the current electric vehicle charging process.
As a preferred scheme, when an abnormal charging process occurs in the charging process of the electric vehicle, the fault diagnosis is performed on the charging process of the electric vehicle, and the fault type of the charging process is diagnosed, including the following steps:
step 1: initializing network parameters of the fault diagnosis network, including the number S of neurons in the input layer in And number of hidden layer neurons S H The number of neurons in the output layer is1。
Step 2: initializing fault diagnosis network input weight matrix
And step 3: obtaining a trained fault input sample x and a fault target label
And 4, step 4: estimating fault diagnosis network output weight matrix by correlation variable method
And 5: inputting the weight matrix
Output weight matrix->
Substituting into the fault diagnosis network, and inputting a sample x and a fault target label->
And training the fault diagnosis network to obtain the trained fault diagnosis network.
And 6: and taking relevant data of the charging process corresponding to the abnormal charging process as input, and calculating the trained fault diagnosis network in a forward direction to obtain the fault type of the charging process.
Preferably, the calculation process of the correlation variable method is as follows:
step 1: defining activation tag sets
Inactive set of tags O = {1,2,3
H Are multiplied by
Step 2: the correlation of the variables for the first iteration is calculated by
Namely:
wherein ,N
s For a preset number of variables to be screened,
represents the output weight matrix evaluated in l iterations, and>
a transpose operation representing a matrix; />
and />
Representing a data matrix with an active data set and an inactive data set, respectively.
And 3, step 3: estimating an output weight matrix estimated in the l-th iteration
Namely:
wherein, λ is a regularization coefficient, and I is an identity matrix;
and 4, step 4: after the variable correlation of the first iteration is calculated, the correlation value of the variable is calculated
Removed from the data set O and added to the data set Θ;
and 5: obtaining an optimal output weight matrix from the following formula according to Akaike information criteria, namely
Wherein l is 1,2, 823060, 8230S H /N s 。
And 6: weight matrix of output estimation
Preferably, the fault types include a fault of a charging pile and a battery fault of an electric vehicle.
As a preferred scheme, the matching of the specific fault of the current charging according to the fault type through the fault data knowledge base comprises the following steps:
and establishing an electric vehicle charging fault data knowledge base.
And matching the related data and the fault type of the charging process with the faults in the charging fault data knowledge base of the electric automobile to obtain the specific faults of the charging.
The electric vehicle charging fault data knowledge base comprises: fill electric pile trouble knowledge base, power battery trouble knowledge base.
Fill electric pile trouble knowledge base, including equipment storehouse information and trouble storehouse information, wherein equipment storehouse information: charging pile equipment name, equipment model, version and rated parameters; for charging pile faults, experts already summarize common faults of equipment through charging pile fault detection and analysis and historical fault data to form fault library information; the fault bank information includes: fault codes, fault types, fault phenomena and fault record lists;
the power battery fault knowledge base comprises equipment base information and fault base information, wherein the equipment base information is as follows: the name, the model, the version and the rated parameters of the battery equipment; for the power battery fault, the expert summarizes common faults of the equipment through power battery fault detection analysis and historical fault data to form fault library information; the fault bank information includes: fault code, fault type, fault phenomenon, fault record table.
In a second aspect, a remote fault early warning device suitable for an electric vehicle charging process includes the following modules:
and the data acquisition controller is used for acquiring the relevant data of the charging process through the data acquisition unit.
And the digital signal processor is used for carrying out noise reduction, abnormal value detection and normalization processing on the obtained related data of the charging process to obtain the preprocessed related data of the charging process.
And the fault early warning controller is used for judging the relevant data of the preprocessed charging process and judging whether the charging process is abnormal.
And the fault diagnosis controller is used for diagnosing faults in the electric vehicle charging process and diagnosing the fault type of the charging process when the abnormal charging process occurs in the electric vehicle charging process.
And the database controller is used for matching specific faults of the charging according to the fault types through the fault data knowledge base.
As a preferred scheme, the fault early warning controller comprises the following functions:
and training the modal learning network, acquiring parameters of the modal learning network, substituting the parameters of the modal learning network into the modal learning network, and acquiring the trained modal learning network.
And training the modal discrimination network, acquiring parameters of the modal discrimination network, substituting the parameters of the modal discrimination network into the modal discrimination network, and acquiring the trained modal discrimination network.
And constructing a modal early warning network according to the trained modal learning network and the trained modal discrimination network, training the modal early warning network, acquiring parameters of the modal early warning network, substituting the parameters of the modal early warning network into the modal early warning network to obtain the trained modal early warning network, inputting the preprocessed charging process related data into the trained modal early warning network, and judging whether the charging process is abnormal.
As a preferred scheme, training a modal learning network, obtaining parameters of the modal learning network, substituting the parameters of the modal learning network into the modal learning network, and obtaining the trained modal learning network, including the following steps:
step 1: initializing a modal learning network unit, including the number M of neurons contained in the input layer
in The number of neurons in the a-th hidden layer
The number M of neurons included in the output layer
O A =1,2, 3., a denotes the total number of hidden layers of the modal learning network.
And 2, step: initializing input weight matrix of a-th hidden layer
When a =1, is selected>
The size of the matrix is
An input weight matrix representing a first hidden layer; when a > 1, is selected>
The size of the matrix is
The input weight matrix representing the a-th hidden layer.
And 3, step 3: computing the output matrix P of the a-th hidden layer
a (ii) a When a =1, the number of the bits is set to a =1,
when a is greater than 1, the ratio of a,
wherein f is an activation function->
Normalization data of relevant data of the charging process corresponding to the abnormal charging process input by the modal learning network in the training process, wherein the data comprises charging voltage, charging current, charging power, charging module temperature, charging gun temperature and power battery monomer temperature, voltage and current, T is the number of the electric vehicles, M is the number of the electric vehicles
in Number of correlated data after normalization for correlated data of charging process, P
a-1 The output matrix representing the a-1 st hidden layer.
And 4, step 4: estimating output weight matrix W of output layer by pseudo-inverse method
out I.e. W
out = Y pinv (P). Wherein, pinv (#) represents a pseudo-inverse calculation,
an output target matrix of the network in the training process is learned for the mode,
the output matrix synthesis set of the input layer and the hidden layer is obtained. Training output of the modal learning network->
Is [1, 1., 1 ]]And representing that in the modal learning network, the input relevant electrical sampling value data is defined as the absence of the early warning information, and 1 represents that the charging state of the electric vehicle is normal.
And 5: and substituting the parameters of the modal learning network into the modal learning network to obtain the trained modal learning network.
As a preferred scheme, training a mode discrimination network, acquiring parameters of the mode discrimination network, substituting the parameters of the mode discrimination network into the mode discrimination network, and acquiring the trained mode discrimination network, including the following steps:
step 1: the initialization mode discrimination network unit comprises a neuron number N contained in an input layer
in The number N of neurons included in the hidden layer
H And the number N of neurons included in the output layer
O Input weight matrix of mode discrimination network input layer
Step 2: computing the output matrix of the hidden layer by
I.e. is>
Wherein f is an activation function; />
Y is the training output of the modal learning network, X
normal The relevant data of the corresponding charging process of the normal charging process comprise charging voltage, charging current, charging power, charging module temperature, charging gun temperature, and power battery monomer temperature, voltage and current. T is the number of electric vehicles, N
in Number of relevant data for a corresponding charging process which is a normal charging process>
Representing a matrix transpose.
And step 3: estimating an output weight matrix of a modal discrimination network output layer by solving the following optimization problem J
Namely:
wherein ,Y
* Representing an output target matrix of the mode discrimination network in the training process, | × | | non-woven phosphor
2 Represents l
2 Norm, α
i,j Connecting weight of ith neuron representing hidden layer of modal discrimination network and jth neuron of output layer
The regularization coefficients of (a). The mode discrimination network outputs a target matrix ^ according to the training process>
Is [1, 1., 1, 0., 0 ]]And representing that in the mode judging network, the input related electrical sampling value is defined to contain early warning information, 1 represents that the charging state of the electric automobile is normal, and 0 represents that the charging state of the electric automobile is abnormal.
And 4, step 4: and substituting the parameters of the modal discrimination network into the modal discrimination network to obtain the trained modal discrimination network.
As a preferred scheme, the method comprises the following steps of constructing a modal early warning network according to a trained modal learning network and a modal discrimination network, training the modal early warning network, acquiring parameters of the modal early warning network, substituting the parameters of the modal early warning network into the modal early warning network to obtain the trained modal early warning network, inputting relevant data of a preprocessed charging process into the trained modal early warning network, and judging whether the charging process is abnormal, wherein the method comprises the following steps:
step 1: the initialization mode early warning network unit comprises a neuron number M contained in an input layer
in The number of neurons contained in the a-th hidden layer
a =1,2,3.., a +1 th hidden layer contains the number M of neurons
O And the number N of neurons contained in the A +2 th hidden layer
H And the number N of neurons included in the output layer
O 。
Step 2: initializing a connection weight matrix W of an input layer and a 1 st hidden layer of a modal early warning network unit
in 1,2 A-1 th hidden layer and the a th hidden layer connection weight matrix W
in a,a+1 A is more than or equal to 2 and less than A-1, and the connection weight matrix W of the A-th hidden layer and the A + 1-th hidden layer
in A,A+1 The connection weight matrix W of the A +1 th hidden layer and the A +2 nd hidden layer
in A+1,A+2 Output weight matrix of output layer
In the process of constructing the modal early warning network, the connection weight matrix W of the input layer and the 1 st hidden layer
in 1,2 Input weight matrix equal to input layer and 1 st hidden layer of modal learning network
Connection weight matrix W of a-1 th hidden layer and a-th hidden layer
in a,a+1 Connection weight matrix & -1 hidden layer of a modal learning network>
A is more than or equal to 2 and less than A-1, and the connection weight matrix W of the A-th hidden layer and the A + 1-th hidden layer
in A,A+1 Output weight matrix W equal to output layer of modal learning network
out The connection weight matrix W of the A +1 th hidden layer and the A +2 th hidden layer
in A+1,A+2 Input weight matrix equal to mode decision network input layer->
And 3, step 3: and substituting the parameters of the modal early warning network into the modal early warning network to obtain the trained modal early warning network.
And 4, step 4: and acquiring related data of the current electric vehicle charging process, taking the related data of the current electric vehicle charging process as input, and calculating the trained modal early warning network in a forward direction to obtain an early warning result of the charging process.
Preferably, the output weight matrix of the mode discrimination network
In the jth neuron of (a) < x >>
An estimation process, comprising:
wherein ,Y
j * Expressed as the target matrix of the j-th neuron of the output layer corresponding to the modal discrimination network in the training process,
an output weight matrix representing the jth neuron of the output layer corresponding to the modal discrimination network, and->
Output weight element, alpha, representing the ith neuron of the A +2 th hidden layer and the jth neuron of the output layer in the modal discrimination network
i,j And a regularization coefficient of an output weight element of an ith neuron of an A +2 th hidden layer and a jth neuron of an output layer in the modal discrimination network is represented, wherein A represents the total number of the hidden layers of the modal learning network.
Modal discrimination network output weight matrix
Is selected based on the output weight matrix of the jth neuron of (a)>
The estimation process is further derived as:
wherein ,Y
1 * Expressed as a target matrix of the 1 st neuron of the corresponding output layer of the modal discrimination network in the training process,
an output weight matrix representing the 1 st neuron of the corresponding output layer of the modal discrimination network, and->
Output weight element, α, representing the ith neuron of the A +2 th hidden layer and the 1 st neuron of the output layer in the modal discrimination network
i,1 The regularization coefficients represent output weight elements of the ith neuron of an A +2 th hidden layer and the 1 st neuron of an output layer in the modal discrimination network; y is
2 * A target matrix, denoted as the 2 nd neuron of the modal discrimination network corresponding to the output layer during training, in combination with a selection criterion>
An output weight matrix representing the 2 nd neuron of the corresponding output layer of the modal discrimination network, and->
Output weight element, α, representing the ith neuron of the A +2 th hidden layer and the 2 nd neuron of the output layer in the modal discrimination network
i,2 The regularization coefficients represent output weight elements of the ith neuron of the A +2 th hidden layer and the 2 nd neuron of the output layer in the modal discrimination network; />
Expressed as the Nth output layer corresponding to the mode discrimination network in the training process
H A target matrix of individual neurons, based on the evaluation of the neuron>
Nth of output layer corresponding to expression mode discrimination network
H The output weight matrix of the individual neurons,
representing the ith neuron of the A +2 th hidden layer and the Nth neuron of the output layer in the modal discrimination network
H An output weight element for each neuron, based on the number of neurons in the neuron's input>
Representing the ith neuron of the A +2 hidden layer and the Nth neuron of the output layer in the modal discrimination network
H Regularization coefficients of output weight elements of individual neurons.
Modal discrimination network estimation output weight matrix
Is selected based on the output weight matrix of the jth neuron of (a)>
Can be described as J solving for N
H The optimization problems are as follows:
wherein ,Jj Denotes the jth optimization problem, j =1,2,3 H (ii) a The jth optimization problem can be further derived as:
wherein ,
the revised output weight matrix of the jth neuron of the corresponding output layer of the representation mode discrimination network is used for being greater than or equal to>
Revised output matrix representing modal discriminant network hidden layer (| + | non-conducting phosphor)
1/2 Is represented by
1/2 And (4) norm. Y is
j * And expressing the target matrix of the j-th neuron corresponding to the output layer in the training process of the modal discrimination network.
The solving process of the jth optimization problem is as follows:
step 1: initializing a weight matrix
And zero matrix->
and 3, step 3: label i =1 is defined.
And 4, step 4: judging whether i is equal to N O (ii) a If not, i = i +1, entering step 5; if yes, go to step 7.
and 6: calculated by the following formula
Namely that
wherein ,Δi 、Α i and Βi The expression is:
wherein ,Y
j * (t) revised data expressing a target output matrix of the jth input data at the jth neuron in the output layer,
represents the mode discrimination of the t-th input dataRevised output data of the qth neuron of the network hidden layer, arccos (×) representing an arccosine function; t is the number of the electric automobiles. />
And the revised output weight matrix represents the j-th neuron of the corresponding output layer of the q-th neuron of the hidden layer of the modal discrimination network.
And 7: judgment of
If yes, executing step 8, otherwise, executing step 4.
And step 8: output weight matrix
As a preferred scheme, the charging data of the electric vehicle and the charging pile 1-3 times before and after the current charging is used as the data monitoring range in the fault diagnosis process as the relevant data of the current electric vehicle charging process.
Preferably, the fault diagnosis controller includes the following functions:
step 1: initializing network parameters of a fault diagnosis network, including input layer neuron number S in And number of hidden layer neurons S H And the number of neurons in the output layer is 1.
Step 2: initializing fault diagnosis network input weight matrix
And step 3: obtaining a trained fault input sample x and a fault target label
And 4, step 4: estimating fault diagnosis network output weight matrix by correlation variable method
And 5: inputting the weight matrix
Output weight matrix>
Substituting into the fault diagnosis network, and inputting a sample x and a fault target label->
And training the fault diagnosis network to obtain the trained fault diagnosis network.
Step 6: and (4) taking relevant data of the charging process corresponding to the abnormal charging process as input, and calculating the trained fault diagnosis network in a forward direction to obtain the fault type of the charging process.
Preferably, the calculation process of the correlation variable method is as follows:
step 1: defining activation tag sets
Set of inactive labels O = {1,2,3
H Is and->
Step 2: the variable dependence of the l-th iteration is calculated by
Namely:
wherein ,N
s The number of variables to be screened is preset,
represents the output weight matrix evaluated in l iterations, and>
a transpose operation representing a matrix; />
and />
Representing a data matrix with an active data set and an inactive data set, respectively.
And 3, step 3: estimating an output weight matrix estimated in the l-th iteration
Namely:
wherein, lambda is a regularization coefficient, and I is an identity matrix;
and 4, step 4: after calculating the variable correlation of the first iteration, will
Removed from the data set O and added to the data set Θ;
and 5: obtaining an optimal output weight matrix from the following formula according to Akaike information criteria, namely
Wherein l is 1,2, 823060, 8230S H /N s 。
And 6: weight matrix of output estimation
Preferably, the fault types include a fault of a charging pile and a battery fault of an electric vehicle.
Preferably, the database controller includes the following functions:
and establishing an electric vehicle charging fault data knowledge base.
And matching the related data and the fault type of the charging process with the faults in the charging fault data knowledge base of the electric automobile to obtain the specific faults of the charging.
The electric vehicle charging fault data knowledge base comprises: fill electric pile trouble knowledge base, power battery trouble knowledge base.
Fill electric pile trouble knowledge base, including equipment storehouse information and trouble storehouse information, wherein equipment storehouse information: charging pile equipment name, equipment model, version and rated parameters; for charging pile faults, experts already summarize common faults of equipment through charging pile fault detection and analysis and historical fault data to form fault library information; the fault bank information includes: fault codes, fault types, fault phenomena and fault record lists;
the power battery fault knowledge base comprises equipment base information and fault base information, wherein the equipment base information comprises: the name, the model, the version and the rated parameters of the battery equipment; for power battery faults, experts already summarize common faults of equipment through power battery fault detection analysis and historical fault data to form fault library information; the fault bank information includes: fault code, fault type, fault phenomenon, fault record table.
Has the beneficial effects that: according to the remote fault early warning method and device applicable to the electric vehicle charging process, fault diagnosis is carried out by analyzing the charging process data of the electric vehicle and utilizing historical data and logic decision judgment, the fault early warning network of the electric vehicle is constructed through the learning network with the game ability, a judgment threshold value is not required to be set, effective fault information can be provided for operation and maintenance personnel and electric vehicle users, and the safety of the users and equipment is guaranteed.
Detailed Description
The present invention will be further described with reference to the following examples.
As shown in fig. 1, a first embodiment of a remote fault early warning method applied to an electric vehicle charging process includes the following steps:
and acquiring related data of the charging process through a data acquisition unit.
And carrying out noise reduction, abnormal value detection and normalization processing on the obtained relevant data of the charging process to obtain the preprocessed relevant data of the charging process.
And judging the relevant data of the preprocessed charging process, and judging whether the charging process is abnormal.
When the charging process of the electric automobile is abnormal, fault diagnosis is carried out on the charging process of the electric automobile, and the fault type of the charging process is diagnosed.
And matching specific faults of the charging according to the fault types through a fault data knowledge base.
The relevant data of the charging process at least comprises the following data: charging voltage, charging current, charging power, charging module temperature, charging gun temperature, and power battery cell temperature, voltage, current.
As shown in fig. 2, the determining the data related to the charging process to determine whether the charging process is abnormal includes: and the three sub-networks are respectively a mode learning network, a mode judging network and a mode early warning network.
Firstly, training a modal learning network, acquiring parameters of the modal learning network, substituting the parameters of the modal learning network into the modal learning network, and acquiring the trained modal learning network. Namely:
step 1: initializing a modal learning network unit comprising the number M of neurons contained in an input layer
in The number of neurons in the a-th hidden layer
The number M of neurons included in the output layer
O A =1,2, 3., a denotes the total number of hidden layers of the modal learning network.
Step 2: initializing input weight matrix of a-th hidden layer
When a =1, is selected>
The size of the matrix is
An input weight matrix representing a first hidden layer; when a > 1, is selected>
The size of the matrix is
An input weight matrix representing the a-th hidden layer.
And 3, step 3: computing the output matrix P of the a-th hidden layer
a (ii) a When a =1, the control unit is configured to,
when a is greater than 1, the ratio of a,
wherein f is an activation function->
Normalization data of relevant data of the charging process corresponding to the abnormal charging process input by the modal learning network in the training process, wherein the data comprises charging voltage, charging current, charging power, charging module temperature, charging gun temperature and power battery monomer temperature, voltage and current, T is the number of the electric vehicles, M is the number of the electric vehicles
in Normalizing the data relating to the charging processNumber, P
a-1 The output matrix representing the a-1 st hidden layer.
And 4, step 4: estimating output weight matrix W of output layer by pseudo-inverse method
out I.e. W
out = Y pinv (P). Wherein pinv (—) represents a pseudo-inverse calculation,
an output target matrix of the network in the training process is learned for the mode,
the output matrix synthesis set of the input layer and the hidden layer is obtained. Training output of the modal learning network->
Is [1, 1., 1 ]]And representing that in the modal learning network, the input relevant electrical sampling value data is defined as the absence of the early warning information, and 1 represents that the charging state of the electric automobile is normal. />
And 5: and substituting the parameters of the modal learning network into the modal learning network to obtain the trained modal learning network.
Through the steps, a well-trained modal learning network can be obtained, and the forward game of the charging state of the electric automobile is mainly based on the electric charging voltage, the charging current, the charging power, the temperature of a charging module, the temperature of a charging gun, the temperature, the voltage and the current of a power battery monomer and other related data of the electric automobile.
Secondly, training the modal discrimination network, acquiring parameters of the modal discrimination network, substituting the parameters of the modal discrimination network into the modal discrimination network, and acquiring the trained modal discrimination network. Namely:
step 1: the initialization mode discrimination network unit comprises a neuron number N contained in an input layer
in The number N of neurons included in the hidden layer
H And the number N of neurons included in the output layer
O Input weight matrix of mode discrimination network input layer
Step 2: computing the output matrix of the hidden layer by
I.e. is>
Wherein f is an activation function; />
Y is the training output of the modal learning network, X
normal And the related data of the corresponding charging process representing the normal charging process comprise charging voltage, charging current, charging power, charging module temperature, charging gun temperature, and power battery monomer temperature, voltage and current. T is the number of electric vehicles, N
in Number of relevant data for a corresponding charging process which is a normal charging process>
Representing a matrix transpose.
And 3, step 3: estimating an output weight matrix of a modal discrimination network output layer by solving the following optimization problem J
Namely:
wherein ,Y
* Representing an output target matrix of the mode discrimination network in the training process, | × | | non-woven phosphor
2 Is represented by
2 Norm, α
i,j Connecting weight of ith neuron representing hidden layer of modal discrimination network and jth neuron of output layer
The regularization coefficients of (a). The output purpose of the mode discrimination network in the training processMark matrix>
Is [1, 1., 1, 0., 0 ]]And representing that in the mode judging network, the input related electrical sampling value is defined to contain early warning information, 1 represents that the charging state of the electric automobile is normal, and 0 represents that the charging state of the electric automobile is abnormal.
And 4, step 4: and substituting the parameters of the modal discrimination network into the modal discrimination network to obtain the trained modal discrimination network.
After the steps, a well-trained training discrimination network is obtained, and the reverse game of the charging state of the electric automobile is mainly based on the electric charging voltage, the charging current, the charging power, the temperature of a charging module, the temperature of a charging gun, the temperature, the voltage, the current and other related data of the power battery monomer.
Further, the output weight matrix of the mode discrimination network
In the jth neuron of (a) < x >>
An estimation process, comprising:
wherein ,Y
j * Expressed as the target matrix of the j-th neuron of the output layer corresponding to the modal discrimination network in the training process,
an output weight matrix representing the jth neuron of the output layer corresponding to the modal discrimination network, and->
Representing the ith neuron of the A +2 th hidden layer and the jth neuron of the output layer in the modal discrimination networkOutput weight element of neuron, alpha
i,j And a regularization coefficient of an output weight element of an ith neuron of an A +2 th hidden layer and a jth neuron of an output layer in the modal discrimination network is represented, wherein A represents the total number of the hidden layers of the modal learning network.
Further, a mode discrimination network output weight matrix
Output weight matrix of jth neuron
The estimation process is further derived as:
wherein ,Y
1 * Expressed as a target matrix of the 1 st neuron of the corresponding output layer of the modal discrimination network in the training process,
an output weight matrix representing the 1 st neuron of the corresponding output layer of the modal discrimination network, and->
Output weight element, alpha, representing the ith neuron of the A +2 th hidden layer and the 1 st neuron of the output layer in the modal discrimination network
i,1 The regularization coefficients represent output weight elements of the ith neuron of an A +2 th hidden layer and the 1 st neuron of an output layer in the modal discrimination network; y is
2 * A target matrix, expressed as the 2 nd neuron of the modal discrimination network corresponding to the output layer during the training process, is asserted>
An output weight matrix representing the 2 nd neuron of the corresponding output layer of the modal discrimination network, and->
Output weight element, α, representing the ith neuron of the A +2 th hidden layer and the 2 nd neuron of the output layer in the modal discrimination network
i,2 The regularization coefficients represent output weight elements of the ith neuron of an A +2 th hidden layer and the 2 nd neuron of an output layer in the modal discrimination network; />
Expressed as N corresponding to output layer of modal discrimination network in training process
H A target matrix for individual neurons>
Nth of output layer corresponding to expression mode discrimination network
H An output weight matrix for each neuron, < > >>
Representing the ith neuron of the A +2 th hidden layer and the Nth neuron of the output layer in the modal discrimination network
H An output weight element for each neuron, based on the number of neurons in the neuron's input>
Representing the ith neuron of the A +2 hidden layer and the Nth neuron of the output layer in the modal discrimination network
H Regularization coefficients of output weight elements of individual neurons.
Modal discrimination network estimation output weight matrix
Is selected based on the output weight matrix of the jth neuron of (a)>
Can be described as J solving for N
H The optimization problems are as follows:
wherein ,Jj Denotes the jth optimization problem, j =1,2,3 H (ii) a The jth optimization problem can be further derived as:
wherein ,
a revised output weight matrix representing the jth neuron in the output layer corresponding to the modal discrimination network, and->
Revised output matrix representing modal discriminant network hidden layer (| + | non-conducting phosphor)
1/2 Represents l
1/2 And (4) norm. Y is
j * And expressing the target matrix of the j-th neuron corresponding to the output layer in the training process of the modal discrimination network.
The solving process of the jth optimization problem is as follows:
step 1: initializing a weight matrix
And zero matrix>
and step 3: label i =1 is defined.
And 4, step 4: judging whether i is equal to N O (ii) a If not, i = i +1, and step 5 is entered; if so, go to step 7.
and 6: calculated by the following formula
Namely, it is
wherein ,Δi 、Α i and Βi The expression is:
wherein ,Y
j * (t) revised data expressing a target output matrix of the jth input data at the jth neuron in the output layer,
the revised output data of the qth neuron of the t input data in the hidden layer of the modal discrimination network is represented, and arccos (×) represents an inverse cosine function; t is the number of electric automobiles. />
And the revised output weight matrix of the jth neuron of the output layer corresponding to the qth neuron of the hidden layer of the modal discrimination network is represented. />
And 7: judgment of
If yes, executing step 8, otherwise, executing step 4.
And step 8: output weight matrix
The above calculation process solves the problem of j-th output weight matrix estimation of the output weight matrix estimated by the modal discrimination network. By analogy, the estimation problem of the output weight matrix estimated by the modal discrimination network can be solved, and the output weight matrix estimated by the modal discrimination network can be obtained.
And finally, constructing a modal early warning network according to the trained modal learning network and the trained modal discrimination network, training the modal early warning network, acquiring parameters of the modal early warning network, substituting the parameters of the modal early warning network into the modal early warning network to obtain the trained modal early warning network, inputting relevant data of a charging process into the trained modal early warning network, and judging whether the charging process is abnormal, namely:
step 1: the initialization mode early warning network unit comprises a neuron number M contained in an input layer
in The number of neurons included in the a-th hidden layer
a =1,2, 3.., a, number of neurons M contained in the a +1 st hidden layer
O And the number N of the neurons contained in the A +2 th hidden layer
H And the number N of neurons included in the output layer
O 。
Step 2: connection weight matrix W of input layer and 1 st hidden layer of initialization mode early warning network unit
in 1,2 A-1 th hidden layer and a connection weight matrix W of the a-th hidden layer
in a,a+1 A is more than or equal to 2 and less than A-1, and the connection weight matrix W of the A-th hidden layer and the A + 1-th hidden layer
in A,A+1 The connection weight matrix W of the A +1 th hidden layer and the A +2 th hidden layer
in A+1,A+2 Output weight matrix of output layer
In the process of constructing the modal early warning network, the connection weight matrix W of the input layer and the 1 st hidden layer
in 1,2 Equal modal learning netInput weight matrix of the input layer and the 1 st hidden layer
Connection weight matrix W of a-1 th hidden layer and a-th hidden layer
in a,a+1 Connection weight matrix ≧ which equals the a-1 th hidden layer of the modal learning network>
A is more than or equal to 2 and less than A-1, and the connection weight matrix W of the A-th hidden layer and the A + 1-th hidden layer
in A,A+1 Output weight matrix W equal to output layer of modal learning network
out The connection weight matrix W of the A +1 th hidden layer and the A +2 th hidden layer
in A+1,A+2 Input weight matrix equal to mode decision network input layer->
And 3, step 3: and substituting the parameters of the modal early warning network into the modal early warning network to obtain the trained modal early warning network.
And 4, step 4: and acquiring related data of the current electric vehicle charging process, taking the related data of the current electric vehicle charging process as input, and calculating the trained modal early warning network in a forward direction to obtain an early warning result of the charging process.
The process of identifying the vehicle pile fault according to the early warning result of the charging process is as follows: if the voltage, the current, the power, the charging module and the charging gun output by the charging pile have fault early warning, judging that the charging pile has a fault or is about to have a fault; if the temperature, the voltage and the current of the power battery of the electric automobile have fault early warning, the electric automobile is judged to have a fault or to be subjected to the fault.
Furthermore, the charging data of the charging automobile and the charging pile 1-3 times before and after the current charging is used as the data monitoring range of the fault diagnosis process for the related data of the current charging process of the electric automobile, namely, the charging data of the charging automobile and the charging pile 1-3 times before the current charging is subjected to early warning judgment by a fault early warning method, and the charging data of the charging automobile and the charging pile 1-3 times after the current charging is subjected to early warning judgment.
When the charging process of the electric automobile is abnormal, fault diagnosis is carried out on the charging process of the electric automobile, and the fault type of the charging process is diagnosed, and the method comprises the following steps:
step 1: initializing network parameters of a fault diagnosis network, including input layer neuron number S in And number of hidden layer neurons S H And the number of neurons in the output layer is 1.
Step 2: initializing fault diagnosis network input weight matrix
And 3, step 3: obtaining a trained fault input sample x and a fault target label
And 4, step 4: estimating fault diagnosis network output weight matrix by correlation variable method
And 5: inputting the weight matrix
Output weight matrix>
Substituting into the fault diagnosis network, and inputting a sample x and a fault target label into the fault diagnosis network>
And training the fault diagnosis network to obtain the trained fault diagnosis network.
And 6: and (4) taking relevant data of the charging process corresponding to the abnormal charging process as input, and calculating the trained fault diagnosis network in a forward direction to obtain the fault type of the charging process.
Further, the calculation process of the correlation variable method is as follows:
step 1: defining activation tag sets
Set of inactive labels O = {1,2,3
H Are multiplied by
And 2, step: the correlation of the variables for the first iteration is calculated by
Namely:
wherein ,N
s The number of variables to be screened is preset,
represents the output weight matrix evaluated in l iterations, and>
a transpose operation representing a matrix; />
and />
Representing a data matrix with an active data set and an inactive data set, respectively.
And 3, step 3: estimating an output weight matrix estimated in the l-th iteration
Namely:
wherein, lambda is a regularization coefficient, and I is an identity matrix;
and 4, step 4: after calculating the variable correlation of the first iteration, will
Removed from data set O and added to data set Θ;
and 5: obtaining an optimal output weight matrix from the following formula according to Akaike information standard, namely
Wherein l is 1,2, 823060, 8230S H /N s 。
Step 6: weight matrix of output estimation
Further, the fault types comprise a fault of a charging pile and a battery fault of an electric automobile.
Further, matching specific faults of the current charging according to the fault types through a fault data knowledge base includes:
and establishing an electric vehicle charging fault data knowledge base.
And matching the related data and the fault type of the charging process with the faults in the charging fault data knowledge base of the electric automobile to obtain the specific faults of the charging.
The electric vehicle charging fault data knowledge base comprises: fill electric pile trouble knowledge base, power battery trouble knowledge base.
Fill electric pile trouble knowledge base, including equipment storehouse information and trouble storehouse information, wherein equipment storehouse information: charging pile equipment name, equipment model, version and rated parameters; to filling the electric pile trouble, the expert has already summarized the common trouble of equipment through filling electric pile fault detection analysis and with historical fault data, forms trouble storehouse information: fault codes, fault types, fault phenomena, fault record tables (recording module voltage, current, temperature and other information changes corresponding to different fault types of the charging pile);
the power battery fault knowledge base comprises equipment base information and fault base information, wherein the equipment base information is as follows: the name, the model, the version and the rated parameters of the battery equipment; for the power battery faults, experts already summarize common faults of equipment through power battery fault detection analysis and historical fault data to form fault library information: fault codes, fault types, fault phenomena and fault record tables (recording information changes such as voltage, current, temperature and the like corresponding to different fault types of the power battery).
As shown in fig. 3, a second embodiment of the remote fault early warning device suitable for use in an electric vehicle charging process includes the following modules:
and the data acquisition controller is used for acquiring the relevant data of the charging process through the data acquisition unit.
And the digital signal processor is used for carrying out noise reduction, abnormal value detection and normalization processing on the obtained relevant data of the charging process to obtain the preprocessed relevant data of the charging process.
And the fault early warning controller is used for judging the relevant data of the preprocessed charging process and judging whether the charging process is abnormal.
And the fault diagnosis controller is used for diagnosing faults in the electric vehicle charging process and diagnosing the fault type of the charging process when the abnormal charging process occurs in the electric vehicle charging process.
And the database controller is used for matching specific faults of the charging according to the fault types through the fault data knowledge base.
The relevant data of the charging process at least comprises the following data: charging voltage, charging current, charging power, charging module temperature, charging gun temperature, and power battery cell temperature, voltage, current.
The fault early warning controller comprises the following functions:
the method comprises the following steps: and the three sub-networks are respectively a mode learning network, a mode judging network and a mode early warning network.
Firstly, training a modal learning network, acquiring parameters of the modal learning network, substituting the parameters of the modal learning network into the modal learning network, and acquiring the trained modal learning network. Namely:
step 1: initializing a modal learning network unit, including the number M of neurons contained in the input layer
in The number of neurons in the a-th hidden layer
The number M of neurons included in the output layer
O A =1,2, 3., a denotes the total number of hidden layers of the modal learning network.
Step 2: initializing input weight matrix of a-th hidden layer
When a =1, is selected>
The size of the matrix is
An input weight matrix representing a first hidden layer; when a > 1, are present>
The size of the matrix is
An input weight matrix representing the a-th hidden layer.
And step 3: computing the output matrix P of the a-th hidden layer
a (ii) a When a =1, the number of the bits is set to a =1,
when a is greater than 1, the ratio of a,
wherein f is an activation function->
Normalization data of relevant data of the charging process corresponding to the abnormal charging process input by the modal learning network in the training process, wherein the data comprises charging voltage, charging current, charging power, charging module temperature, charging gun temperature and power battery monomer temperature, voltage and current, T is the number of the electric vehicles, M is the number of the electric vehicles
in Number of correlated data after normalization for correlated data of charging process, P
a-1 The output matrix representing the a-1 st hidden layer.
And 4, step 4: estimating output weight matrix W of output layer by pseudo-inverse method
out I.e. W
out = Y pinv (P). Wherein, pinv (#) represents a pseudo-inverse calculation,
an output target matrix of the network in the training process is learned for the mode,
the output matrix composition set for the input layer and the hidden layer. Training output of the modal learning network->
Is [1, 1., 1 ]]And representing that in the modal learning network, the input relevant electrical sampling value data is defined as the absence of the early warning information, and 1 represents that the charging state of the electric vehicle is normal.
And 5: and substituting the parameters of the modal learning network into the modal learning network to obtain the trained modal learning network.
Through the steps, a well-trained modal learning network can be obtained, and the forward game of the charging state of the electric automobile is mainly based on the electric charging voltage, the charging current, the charging power, the temperature of a charging module, the temperature of a charging gun, the temperature, the voltage and the current of a power battery monomer and other related data of the electric automobile.
Secondly, training the modal discrimination network, acquiring parameters of the modal discrimination network, substituting the parameters of the modal discrimination network into the modal discrimination network, and acquiring the trained modal discrimination network. Namely:
step 1: the initialization mode discrimination network unit comprises a neuron number N contained in an input layer
in The number N of neurons in the hidden layer
H And the number N of neurons included in the output layer
O Input weight matrix of mode discrimination network input layer
And 2, step: computing the output matrix of the hidden layer by
I.e. is>
Wherein f is an activation function; />
Y is the training output of the modal learning network, X
normal And the related data of the corresponding charging process representing the normal charging process comprise charging voltage, charging current, charging power, charging module temperature, charging gun temperature, and power battery monomer temperature, voltage and current. T is the number of electric vehicles, N
in The number of relevant data of the corresponding charging process, which is a normal charging process, is->
Representing a matrix transposition.
And 3, step 3: estimating an output weight matrix of a modal discrimination network output layer by solving the following optimization problem J
Namely:
wherein ,Y
* Representing an output target matrix of the mode discrimination network in the training process, | × | | non-woven phosphor
2 Is represented by
2 Norm, α
i,j Connecting weight of ith neuron representing modal discrimination network hidden layer and jth neuron of output layer
The regularization coefficients of (1). The mode discrimination network outputs a target matrix &inthe training process>
Is [1, 1., 1, 0., 0 ]]In the mode discrimination network, the input related electrical sampling value is defined as containing early warning information, 1 represents that the charging state of the electric automobile is normal, and 0 represents that the charging state of the electric automobile is abnormal.
And 4, step 4: and substituting the parameters of the modal discrimination network into the modal discrimination network to obtain the trained modal discrimination network.
After the steps, a training discrimination network with good training is obtained, and the reverse game of the charging state of the electric automobile is mainly based on the electric charging voltage, the charging current, the charging power, the temperature of a charging module, the temperature of a charging gun, the temperature of a power battery monomer, the voltage, the current and other related data of the electric automobile.
Further, the output weight matrix of the mode discrimination network
Is selected based on the output weight matrix of the jth neuron of (a)>
An estimation process, comprising:
wherein ,Y
j * Expressed as the target matrix of the j-th neuron of the output layer corresponding to the modal discrimination network in the training process,
an output weight matrix representing the jth neuron of the output layer corresponding to the modal discrimination network, and->
Output weight element, alpha, representing the ith neuron of the A +2 th hidden layer and the jth neuron of the output layer in the modal discrimination network
i,j And a regularization coefficient of an output weight element of an ith neuron of an A +2 th hidden layer and a jth neuron of an output layer in the modal discrimination network is represented, wherein A represents the total number of the hidden layers of the modal learning network.
Further, the modal discrimination network outputs a weight matrix
Output weight matrix of jth neuron
The estimation process is further derived as:
wherein ,Y
1 * Expressed as the target matrix of the 1 st neuron of the mode discrimination network corresponding to the output layer in the training process,
an output weight matrix representing the 1 st neuron of the corresponding output layer of the modal discrimination network, and->
Output weight element, alpha, representing the ith neuron of the A +2 th hidden layer and the 1 st neuron of the output layer in the modal discrimination network
i,1 Representing modalitiesJudging a regularization coefficient of an output weight element of an ith neuron of an A +2 th hidden layer and a 1 st neuron of an output layer in the network; y is
2 * A target matrix, expressed as the 2 nd neuron of the modal discrimination network corresponding to the output layer during the training process, is asserted>
An output weight matrix representing the 2 nd neuron of the corresponding output layer of the modal discrimination network, and->
Output weight element, α, representing the ith neuron of the A +2 th hidden layer and the 2 nd neuron of the output layer in the modal discrimination network
i,2 The regularization coefficients represent output weight elements of the ith neuron of the A +2 th hidden layer and the 2 nd neuron of the output layer in the modal discrimination network; />
Expressed as the Nth output layer corresponding to the mode discrimination network in the training process
H A target matrix for individual neurons>
Nth of output layer corresponding to expression mode discrimination network
H An output weight matrix for each neuron, < > >>
Representing the ith neuron of the A +2 hidden layer and the Nth neuron of the output layer in the modal discrimination network
H An output weight element for each neuron, < > or >>
Representing the ith neuron of the A +2 hidden layer and the Nth neuron of the output layer in the modal discrimination network
H Regularization coefficients of output weight elements of the individual neurons.
Modal discrimination network estimation output weight matrix
Is selected based on the output weight matrix of the jth neuron of (a)>
Can be described as solving for N for J
H The optimization problems are as follows: />
wherein ,Jj Denotes the jth optimization problem, j =1,2,3 H (ii) a The jth optimization problem can be further derived as:
wherein ,
a revised output weight matrix representing the jth neuron in the output layer corresponding to the modal discrimination network, and->
A revised output matrix representing the modal-discriminating network hidden layer | | | × n calculation
1/2 Represents l
1/2 And (4) norm. Y is
j * And expressing the target matrix of the j-th neuron corresponding to the output layer in the training process of the modal discrimination network.
The solving process of the j-th optimization problem is as follows:
step 1: initializing a weight matrix
And zero matrix->
and 3, step 3: label i =1 is defined.
And 4, step 4: judging whether i is equal to N O (ii) a If not, i = i +1, and step 5 is entered; if yes, go to step 7.
and 6: calculated by the following formula
Namely, it is
wherein ,Δi 、Α i and Βi The expression is:
wherein ,Y
j * (t) revised data expressing a target output matrix of the jth input data at the jth neuron in the output layer,
the revised output data of the qth neuron of the t input data in the hidden layer of the modal discrimination network is represented, and arccos (—) represents an arccosine function; t is the number of electric automobiles. />
And the revised output weight matrix of the jth neuron of the output layer corresponding to the qth neuron of the hidden layer of the modal discrimination network is represented.
And 7: judgment of
If yes, executing step 8, otherwise, executing step 4.
And 8: output weight matrix
The above calculation process solves the jth output weight matrix estimation problem of the output weight matrix estimated by the modal discrimination network. By analogy, the estimation problem of the output weight matrix estimated by the modal discrimination network can be solved, and the output weight matrix estimated by the modal discrimination network can be obtained.
And finally, constructing a modal early warning network according to the trained modal learning network and the trained modal discrimination network, training the modal early warning network, acquiring parameters of the modal early warning network, substituting the parameters of the modal early warning network into the modal early warning network to obtain the trained modal early warning network, inputting relevant data of a charging process into the trained modal early warning network, and judging whether the charging process is abnormal, namely:
step 1: the initialization mode early warning network unit comprises a neuron number M contained in an input layer
in The number of neurons contained in the a-th hidden layer
a =1,2,3.., a +1 th hidden layer contains the number M of neurons
O And the number N of the neurons contained in the A +2 th hidden layer
H And the number N of neurons included in the output layer
O 。
And 2, step: initializing a connection weight matrix W of an input layer and a 1 st hidden layer of a modal early warning network unit
in 1,2 A-1 th hidden layer and the a th hidden layer connection weight matrix W
in a,a+1 ,2≤a<A-1, the connection weight matrix W of the A-th hidden layer and the A + 1-th hidden layer
in A,A+1 The connection weight matrix W of the A +1 th hidden layer and the A +2 th hidden layer
in A+1,A+2 Output weight matrix of output layer
In the process of constructing the modal early warning network, the connection weight matrix W of the input layer and the 1 st hidden layer
in 1,2 Input weight matrix equal to input layer and 1 st hidden layer of modal learning network
Connection weight matrix W of a-1 th hidden layer and a-th hidden layer
in a,a+1 Connection weight matrix ≧ which equals the a-1 th hidden layer of the modal learning network>
A is more than or equal to 2 and less than A-1, and the connection weight matrix W of the A-th hidden layer and the A + 1-th hidden layer
in A,A+1 Output weight matrix W equal to output layer of modal learning network
out The connection weight matrix W of the A +1 th hidden layer and the A +2 th hidden layer
in A+1,A+2 Input weight matrix equal to mode decision network input layer->
And step 3: and substituting the parameters of the modal early warning network into the modal early warning network to obtain the trained modal early warning network.
And 4, step 4: and acquiring relevant data of the current electric vehicle charging process, taking the relevant data of the current electric vehicle charging process as input, and calculating the trained modal early warning network in a forward direction to obtain an early warning result of the charging process.
The process of distinguishing the vehicle pile fault according to the early warning result of the charging process is as follows: if the voltage, the current, the power, the charging module and the charging gun output by the charging pile have fault early warning, judging that the charging pile has a fault or is about to have a fault; if the temperature, the voltage and the current of the power battery of the electric automobile have fault early warning, the electric automobile is judged to have a fault or to be subjected to a fault.
Furthermore, the charging data of the charging automobile and the charging pile 1-3 times before and after the charging of the current electric automobile is used as the data monitoring range of the fault diagnosis process, namely, the charging data of the charging automobile and the charging pile 1-3 times before the charging of the current electric automobile is subjected to early warning judgment through a fault early warning method, and the charging data of the charging automobile and the charging pile 1-3 times after the charging of the current electric automobile is subjected to early warning judgment.
The fault diagnosis controller comprises the following functions:
step 1: initializing network parameters of a fault diagnosis network, including input layer neuron number S in And number of hidden layer neurons S H And the number of neurons in the output layer is 1.
Step 2: initializing fault diagnosis network input weight matrix
And 3, step 3: obtaining a trained fault input sample χ and a fault target label
And 4, step 4: estimating fault diagnosis network output weight matrix by correlation variable method
And 5: inputting the weight matrix
Output weight matrix>
Substituting into fault diagnosis network, inputting X and Y values through faultBarrier target label->
And training the fault diagnosis network to obtain the trained fault diagnosis network.
Step 6: and taking relevant data of the charging process corresponding to the abnormal charging process as input, and calculating the trained fault diagnosis network in a forward direction to obtain the fault type of the charging process.
Further, the calculation process of the correlation variable method is as follows:
step 1: defining activation tag sets
Set of inactive labels O = {1,2,3
H } and
step 2: the variable dependence of the l-th iteration is calculated by
Namely:
wherein ,N
s The number of variables to be screened is preset,
represents the output weight matrix evaluated in l iterations, and>
a transpose operation representing a matrix; />
and />
Representing a data matrix with an active data set and an inactive data set, respectively.
And step 3: estimating an output weight matrix estimated in the l-th iteration
Namely:
wherein, λ is a regularization coefficient, and I is an identity matrix;
and 4, step 4: after the variable correlation of the first iteration is calculated, the correlation value of the variable is calculated
Removed from the data set O and added to the data set Θ;
and 5: obtaining an optimal output weight matrix from the following formula according to Akaike information criteria, namely
Wherein l is 1,2, 823060, 8230S H /N s 。
And 6: weight matrix of output estimation
Further, the fault types comprise a fault of a charging pile and a battery fault of an electric automobile.
The database controller comprises the following functions:
and establishing an electric vehicle charging fault data knowledge base.
And matching the related data and the fault type of the charging process with the faults in the charging fault data knowledge base of the electric automobile to obtain the specific faults of the charging.
The electric vehicle charging fault data knowledge base comprises: fill electric pile trouble knowledge base, power battery trouble knowledge base.
Fill electric pile trouble knowledge base, including equipment storehouse information and trouble storehouse information, wherein equipment storehouse information: charging pile equipment name, equipment model, version and rated parameters; to filling the electric pile trouble, the expert has already summarized the common trouble of equipment through filling electric pile fault detection analysis and with historical fault data, forms trouble storehouse information: fault codes, fault types, fault phenomena and fault recording tables (recording information changes such as module voltage, current, temperature and the like corresponding to different fault types of the charging pile);
the power battery fault knowledge base comprises equipment base information and fault base information, wherein the equipment base information is as follows: the name, the model, the version and the rated parameters of the battery equipment; for the power battery faults, experts already summarize common faults of equipment through power battery fault detection analysis and historical fault data to form fault library information: fault codes, fault types, fault phenomena and fault record tables (recording information changes of voltage, current, temperature and the like corresponding to different fault types of the power battery).
As will be appreciated by one skilled in the art, embodiments of the present invention may be provided as a method, system, or computer program product. Accordingly, the present invention may take the form of an entirely hardware embodiment, an entirely software embodiment or an embodiment combining software and hardware aspects. Furthermore, the present invention may take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, and the like) having computer-usable program code embodied therein.
The present invention is described with reference to flowchart illustrations and/or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each flow and/or block of the flowchart illustrations and/or block diagrams, and combinations of flows and/or blocks in the flowchart illustrations and/or block diagrams, can be implemented by computer program instructions. These computer program instructions may be provided to a processor of a general purpose computer, special purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, create means for implementing the functions specified in the flowchart flow or flows and/or block diagram block or blocks.
These computer program instructions may also be stored in a computer-readable memory that can direct a computer or other programmable data processing apparatus to function in a particular manner, such that the instructions stored in the computer-readable memory produce an article of manufacture including instruction means which implement the function specified in the flowchart flow or flows and/or block diagram block or blocks.
These computer program instructions may also be loaded onto a computer or other programmable data processing apparatus to cause a series of operational steps to be performed on the computer or other programmable apparatus to produce a computer implemented process such that the instructions which execute on the computer or other programmable apparatus provide steps for implementing the functions specified in the flowchart flow or flows and/or block diagram block or blocks.
The above description is only of the preferred embodiments of the present invention, and it should be noted that: it will be apparent to those skilled in the art that various modifications and adaptations can be made without departing from the principles of the invention and these are intended to be within the scope of the invention.