CN109033450B - Elevator equipment fault prediction method based on deep learning - Google Patents

Elevator equipment fault prediction method based on deep learning Download PDF

Info

Publication number
CN109033450B
CN109033450B CN201810962887.0A CN201810962887A CN109033450B CN 109033450 B CN109033450 B CN 109033450B CN 201810962887 A CN201810962887 A CN 201810962887A CN 109033450 B CN109033450 B CN 109033450B
Authority
CN
China
Prior art keywords
elevator
fault
time
sequence
model
Prior art date
Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
Active
Application number
CN201810962887.0A
Other languages
Chinese (zh)
Other versions
CN109033450A (en
Inventor
王莉
江海洋
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
Taiyuan University of Technology
Original Assignee
Taiyuan University of Technology
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by Taiyuan University of Technology filed Critical Taiyuan University of Technology
Priority to CN201810962887.0A priority Critical patent/CN109033450B/en
Publication of CN109033450A publication Critical patent/CN109033450A/en
Application granted granted Critical
Publication of CN109033450B publication Critical patent/CN109033450B/en
Active legal-status Critical Current
Anticipated expiration legal-status Critical

Links

Images

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/04—Architecture, e.g. interconnection topology
    • G06N3/049—Temporal neural networks, e.g. delay elements, oscillating neurons or pulsed inputs
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/04—Architecture, e.g. interconnection topology
    • G06N3/044—Recurrent networks, e.g. Hopfield networks
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F2216/00—Indexing scheme relating to additional aspects of information retrieval not explicitly covered by G06F16/00 and subgroups
    • G06F2216/03—Data mining

Landscapes

  • Engineering & Computer Science (AREA)
  • Theoretical Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • General Health & Medical Sciences (AREA)
  • Computing Systems (AREA)
  • Biomedical Technology (AREA)
  • Biophysics (AREA)
  • Computational Linguistics (AREA)
  • Data Mining & Analysis (AREA)
  • Evolutionary Computation (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Molecular Biology (AREA)
  • Artificial Intelligence (AREA)
  • General Engineering & Computer Science (AREA)
  • General Physics & Mathematics (AREA)
  • Mathematical Physics (AREA)
  • Software Systems (AREA)
  • Health & Medical Sciences (AREA)
  • Elevator Control (AREA)
  • Management, Administration, Business Operations System, And Electronic Commerce (AREA)

Abstract

The invention relates to an elevator equipment fault prediction method based on deep learning, belonging to the technical field of elevator fault prediction; the technical problem to be solved is to provide a method for timely and accurately predicting the elevator fault type and the elevator fault time; in order to solve the technical problems, the method comprises the following specific steps: firstly, collecting elevator fault record information and establishing a real-time elevator fault information base; then processing the elevator fault information into an event sequence and a time sequence; respectively taking the event sequence and the time sequence as input data of the double LSTM, and obtaining output embedding of the two sequences through iterative training of a recurrent neural network; training to obtain the background knowledge of the intensity function and the nonlinear representation of the historical influence by combining the joint layer with the two output embeddings; finally, predicting the type and time of the elevator fault according to the representation result of the intensity function; the elevator maintenance system can assist elevator maintenance personnel to take related preventive measures as soon as possible, and fault events are avoided.

Description

Elevator equipment fault prediction method based on deep learning
Technical Field
The invention discloses an elevator equipment fault prediction method based on deep learning, and belongs to the technical field of elevator fault prediction.
Background
With the increasing number of high-rise buildings, the quality of elevators is receiving much attention. The daily life of people is affected by elevators which are stopped in failure, unsmooth in operation and even have accidents. The method for reducing the failure rate of the elevator and timely and accurately detecting and removing the failure needs further research. Most of the existing elevator fault diagnosis and detection methods are detection methods for a mechanical system, an electrical control system and a safety protection system of an elevator by combining the structure and the principle of the elevator, but a large amount of analysis time is consumed.
Disclosure of Invention
The invention overcomes the defects of the prior art, and the technical problem to be solved is to provide a timely and accurate elevator equipment fault prediction method based on deep learning, so that the purpose of predicting the type and time of the elevator fault is achieved, a maintenance master can detect regularly according to the prediction result, and the occurrence rate of the elevator fault is reduced.
In order to solve the technical problems, the invention adopts the technical scheme that: the elevator equipment fault prediction method based on deep learning comprises the following steps:
step 1, establishing a real-time elevator fault information base, and screening effective elevator fault information as a preprocessing information sequence of a network according to four principles of fault information correlation, complete information, non-repeated information and non-human error operation of the fault information, wherein the elevator fault information comprises fault recording information and elevator basic information, and the fault recording information comprises: elevator fault type, fault reason and fault time, elevator basic information includes: the production date of the elevator, the position of the elevator, the type of the elevator and the service life of the elevator;
step 2, constructing a time sequence, which comprises two characteristics: 1) counting the failure times of various types of elevators, 2) basic information of the elevators, and combining the two characteristics to form a time sequence;
and 3, constructing an event sequence, wherein the event sequence comprises two characteristics: 1) the data recording sequence of the elevator fault types is arranged in sequence according to the fault occurrence time, 2) the time interval between two adjacent fault events, and the two characteristics are combined to form an event sequence;
step 4, constructing an LSTM neural network;
step 5, training a time sequence and an event sequence by using a double LSTM neural network to obtain a background knowledge representation and a history influence representation of the intensity function;
step 6, integrating background knowledge representation and historical influence representation through the joint layer;
step 7, predicting elevator fault types through a fault type prediction layer by using the strength function learned through double LSTM, and quantizing the loss values predicted by categories by using a classification loss layer;
step 8, predicting elevator fault time through a fault time prediction layer by using the strength function learned through double LSTM, and quantizing the loss value predicted by time by using a regression loss layer;
step 9, continuously and iteratively training a neural network model based on the step 7 and the step 8 to obtain an optimal network model, and then respectively predicting the type of the elevator fault and the elevator fault time through a fault type prediction layer and a fault time prediction layer by using the trained optimal model;
and step 10, perfecting the optimization model, updating the elevator fault information base in real time, inputting the real-time updated data into the model to test the accuracy of the model, and correcting the perfection model according to the actual feedback condition.
Specifically, the "constructing a time series" in step 2 includes: assuming that the tested elevators have M types, the effective fault types have N types, and the number of the elevator fault time windows is N, the number of the k-th fault types of the jth time window of the ith elevator can be recorded as the number of the k-th fault types of the jth time window of the ith elevator
Figure BDA0001774231350000021
1) Counting the times of various fault types by taking one week as a time window
Figure BDA0001774231350000022
2) The elevator basic information, the two characteristics are combined to form time series data, and the time series data are specifically expressed as follows:
Figure BDA0001774231350000023
wherein M represents the elevator model, d represents the elevator production date, l represents the elevator position, and N represents the number of elevator failure time windows, and double transverse lines are taken as examples in the representation to represent the statistical times of N failure types and the representation unit of the elevator basic information in the first time window of the Mth elevator.
Specifically, the "constructing an event sequence" in step 3 includes: 1) storing all elevator information separately according to different elevator ids, converting elevator fault time into timestamp data, arranging elevator fault types according to fault occurrence time for each elevator, 2) calculating interval time of adjacent fault events of each elevator, and 3) forming a minimum event sequence unit by the fault type T and the timestamp interval I, wherein the minimum event sequence unit is specifically represented as follows:
Figure BDA0001774231350000024
the elevator id represents the unique identifier of the elevator-the elevator number.
Specifically, the method of "constructing the LSTM neural network" in step 4 is as follows:
the specific formula of the recurrent neural network variant LSTM used in the invention is defined as follows:
it＝g(Wi xt+Uiyt-1+Vict-1+bi)，
ft＝g(Wfxt+Ufyt-1+Vfct-1+bf),
ct＝ftct-1+it⊙tanh(Wcxt+Ucyt-1+bc),
ot＝g(Woxt+Uoyt-1+Voct+bo),
yt＝ot⊙tanh(ct)
the above is the formula for the calculation of input gate, forget gate, cell state, and output gate in LSTM, where itFormula for calculating the input gate, ftFormula for calculating a forgetting gate, ctRepresenting the cell state calculation formula, otAnd ytFormula of calculation, o, which jointly represents the output gatetIndicating which part of the cell state is to be output using the g function, and finally ytIndicates that the cell state is treated with tanh and otMultiplying is the part which is determined to be output; wherein U and V are parameter matrixes to be learned in neural network training respectively, which represent the sequential multiplication of array elements, the function g adopts a sigmoid activation function,
Figure BDA0001774231350000031
a sequence of inputs is represented that is,
Figure BDA0001774231350000032
representing the generation of the hidden layer states, the above formula can be simplified as: (y)t,ct)＝LSTM(xtyt-1+ct-1)。
The LSTM construction process comprises the following specific steps:
(1) network initialization: determining the number i of input layer nodes, the number j of hidden layer nodes, the number k of output layer nodes and the state dimension of a unit of the network, initializing the connection weight among neurons of the input layer, the hidden layer and the output layer, and initializing the connection weight W of input gates, forgetting gates, output gates and cell unitsi,Wf,WO,WcInitializing the threshold bf,bc,bo,biGiving a learning rate and a neuron excitation function;
(2) calculating the output value of each neuron: f for LSTMt,it,ct,ot,ht；
(3) And (3) error calculation: inversely calculating an error term of each neuron according to the prediction output and the expected output matrix;
(4) updating the weight value: updating the network connection weight W according to the network prediction errori,Wf,WO,Wc；
(5) Updating a threshold value: updating network node threshold b according to network prediction errorf,bc,bo,bi；
(6) Judging whether the operation is finished or not, and if not, returning to the step (2);
(7) after finishing, using the trained double LSTM neural network to enter the step 5;
specifically, the method of "training the time sequence and the event sequence using the dual LSTM neural network" in step 5 is as follows:
the specific formula is as follows:
Figure BDA0001774231350000033
Figure BDA0001774231350000041
in the above formula
Figure BDA0001774231350000042
Which represents a time series of the images of the object,
Figure BDA0001774231350000043
represents a sequence of events, wherein ziRepresenting the type of event in the sequence of events, tiTime stamps indicating the occurrence of the events, the two sequences being used to learn background knowledge and historical impact, respectively.
Specifically, the method of "fusing the background knowledge representation and the historical influence representation through the joint layer" in the step 6 is as follows: adopting the tanh function as the joint function of the join layer to construct the nonlinear mapping of the strength function of the point process,
the formula of the nonlinear mapping of the point process intensity function constructed by the joint layer join layer is as follows:
Figure BDA0001774231350000044
specifically, in step 7, "predicting elevator fault types by fault type prediction layers using the strength function learned by the dual LSTM, and quantizing the loss values predicted by categories using the classification loss layers" includes:
the calculation formulas of the failure prediction main class and the failure prediction subclass are as follows:
Ut＝softMax(Wuet+bU)
ut＝softMax(Wu[et,Ut]+bu)
in the above formula, U and U represent the main and sub-classes, respectively, and WuAnd buRepresenting a parameter matrix to be learned by the model in the training process; first, e is divided by utilizing softMax functiontPredicting failure dominant categories as inputUt(ii) a Then e is processed again by utilizing the softMax functiontAnd the main class prediction value UtThe predicted value of the subclass is calculated as input.
Specifically, in step 8, "the elevator fault time is predicted by the fault time prediction layer by using the strength function learned by the double LSTM. The time-predicted loss values are quantified using a regression loss layer. The method specifically comprises the following steps:
st＝Wset+bs
wherein s istIs a time stamp, W, for each eventsAnd bsRepresenting the parameter matrix to be learned by the model during the training process.
Specifically, in the step 9, "predicting the type of the elevator fault and the elevator fault time through the fault type prediction layer and the fault time prediction layer by using a trained optimal model" specifically includes:
where the loss of the entire model is the sum of the time prediction loss and the event type prediction loss. And predicting the event type by using a cross entropy loss function, and predicting the time stamp by using a square loss function. The specific loss function formula is as follows:
Figure BDA0001774231350000045
in the above formula, N represents the total number of nodes, and the event points are indexed by l, i.e. l represents the second event, WuA parameter matrix representing the class of events learned by the neural network model,
Figure BDA0001774231350000046
is the time stamp of the next event point,
Figure BDA0001774231350000047
representing historical information of time points, adopting a Gaussian penalty function for a time loss function part:
Figure BDA0001774231350000051
wherein σ represents the uniform variance, and σ is taken2＝10；
And training the whole model in a circulating way by minimizing loss to finally obtain the optimal model.
Specifically, the method of "perfecting the optimization model" in the step 10 is as follows: and obtaining an elevator fault prediction model through a training set, then obtaining a test set by utilizing a real-time updated fault information base, testing the accuracy of the model based on the test set, and perfecting the optimization model in real time.
Through the steps, the construction of an elevator system fault prediction model based on a point process strength function of double LSTM can be completed, time sequence and event sequence data are formed by preprocessing elevator fault data, namely, fault representation can be obtained through a double LSTM neural network model, a nonlinear representation function of the strength function is further obtained by combining two embedding methods, the whole model is trained circularly, and finally the fault time and the fault type of the elevator can be predicted through the strength function of the point process. The method is suitable for solving the problem of elevator fault type and time prediction in practical problems, can help an elevator maintainer to predict the type and time of the elevator to be in fault by using the known elevator fault information, can make preventive measures in advance, avoids the occurrence of elevator faults, reduces the loss and danger caused by the elevator faults, and has high practical application value.
In summary, the invention discloses a method for predicting elevator equipment faults based on deep learning, which provides a data mining algorithm for respectively training an event sequence and a time sequence by using double LSTM and then combining a point process intensity function aiming at online operation measurement data of high-density sampling in an elevator system. The method and the system start from the data perspective, remotely monitor the fault elevator in real time based on the elevator historical fault information sequence, adopt the neural network model to dig out the internal rules of the elevator fault from the massive elevator alarm information, and accurately predict the type and time of the elevator fault in time, thereby assisting the elevator maintenance personnel to take related preventive measures as soon as possible and avoiding the occurrence of fault events.
LSTM (Long Short-term memory) is a Long Short-term memory network, a time recurrent neural network, suitable for processing and predicting important events with relatively Long intervals and delays in time series. LSTM differs from RNN mainly in that it incorporates a "processor" in the algorithm that determines the usefulness of the information, and this processor-oriented architecture is called a cell. Three doors, namely an input door, a forgetting door and an output door, are placed in one cell. A message enters the LSTM network and may be determined to be useful based on rules. Only the information which is in accordance with the algorithm authentication is left, and the information which is not in accordance with the algorithm authentication is forgotten through a forgetting door. It has been proven that LSTM is an effective technique to solve the problem of long order dependence, and the universality of this technique is very high. Various researchers put forward their own variable versions according to the LSTM, so that the LSTM can deal with the vertical problem of the great diversity.
Application of RNN to sequence data: from an application scenario, RNN, as a building block of the present invention, can be applied to two kinds of sequence data: the time sequence and the event sequence can be cooperatively modeled, and are specifically introduced as follows:
(1) time series: refers to a kind of synchronization sequence for recording information statistics of related data in the same time interval. Time series can capture time-varying features in a recent time window in time, so RNNs often use time series as input for relevant sequence prediction problems. For example, a video frame can be considered as a time series data, and we can analyze its intrinsic relation based on the historical video frame to predict the content of the next frame. RNN has been widely used in the fields of video analysis and speech recognition.
(2) The sequence of events: refers to an asynchronous sequence of event random features recorded by time stamps. The event sequence takes the randomly generated time stamp as the input of the RNN, so that the event sequence can capture the long-term dependence relationship among the events more efficiently. Therefore, the invention uses the double LSTM to carry out cooperative modeling on the two sequence data, can capture the updating rule of synchronous information and the asynchronous randomness of burst information, and avoids the limitation brought by parameter assumption by utilizing the nonlinear mapping of the condition intensity function generated by the recurrent neural network.
The point process refers to a time series or a stochastic model of a time series. The point process is a mathematical framework of a modeling sequence recommendation algorithm, and measures the dynamic variability of the point process by using an intensity function. The point process is divided into an event point process and an object point process. Event point process: an earthquake or other catastrophic event; an access event of the server; a factory manufactured off-grade product event; traffic accident incidents that occur at the three-way road, etc. The marked point process is used for predicting the occurrence rule problem of earthquake and aftershock at first; and (3) object point process: the location of the car on the highway; DNA gene, etc.
Introduction of the development history of the intensity function: the point process is used as a mathematical framework of a modeling sequence model, and the dynamic variability of the point process is measured by adopting a conditional strength function. Definition of the intensity function: in the time window [ t + dt), λ (t) represents the time at which the event H is historicalt＝{zi,ti|tiThe occurrence probability of a new event under the premise that < t } occurs is as follows:
Figure BDA0001774231350000061
wherein E (dN (t) Ht) Is shown in history HtOn the basis of the expected value of the number of events occurring within the time interval t + dt). The conditional intensity function plays a key role in the point process sequence prediction algorithm, and different point processes in the parameterized form of the intensity function have different effects. The variation of the intensity function of the point process is as follows: a poisson process; reinforcing the Poisson process; a hokes process; an active site process; a self-correction point process; the RMTPP model; a TRPP model.
The parameterized form of the intensity function consists of two parts: background knowledge and historical characteristics. The above methods are summarized in table 1. From table 1, it can be seen that the former five intensity functions are parameter models artificially assumed according to prior knowledge, and have certain limitations, and the models cannot completely conform to complex dynamic changes of the real sequence problem. While the RMTPP model learns the parameters from the historical features of the event sequence using LSTM. Although the background characteristics of the time series are ignored, the RMTPP model also achieves good effect as a semi-parameterized intensity function; the TRPP model uses dual LSTM to model time and event sequences, respectively, as background knowledge and historical information of intensity functions. Therefore, the model is constructed by mainly using the ideas of the last two models.
TABLE 1 Point Process Strength function
Figure BDA0001774231350000071
Compared with the prior art, the invention has the following beneficial effects.
1. The invention combines a neural network LSTM and a point process for elevator fault prediction, provides a method for respectively training an event sequence and a time sequence by using double LSTMs, and then combines a data mining algorithm of a point process intensity function, and from the perspective of data, based on an elevator historical fault information sequence, learns the change rule of the elevator fault type by nonlinear representation of the neural network by adopting the internal rule contained in the neural network model training data, thereby achieving the purpose of predicting the elevator fault type and the elevator fault time, enabling a maintenance master to regularly detect according to the prediction result, and reducing the occurrence rate of the elevator fault.
2. The method utilizes mass data of elevator faults and utilizes the neural network to learn the occurrence internal rule of the elevator faults, thereby avoiding misjudgment of artificial hypothesis and ensuring that the model is more accurate and efficient.
3. According to the invention, the model utilizes the real-time updated fault information base, so that the accuracy of the model can be verified in time, and the model can be optimized in time, so that the model has timeliness.
Drawings
FIG. 1 is a flow chart of the general steps of the present invention.
FIG. 2 is a time series and a time series co-modeling diagram of the present invention.
Fig. 3 is a frame diagram of the type of failure prediction and time of failure prediction in the present invention.
Detailed Description
The present invention will be described in further detail with reference to the accompanying drawings.
As shown in fig. 1, the elevator equipment fault prediction method based on deep learning of the present invention includes the following steps:
step 1, establishing a real-time elevator fault information base, and screening effective elevator fault information as a preprocessing information sequence of a network according to four principles of fault information correlation, complete information, non-repeated information and non-human error operation of the fault information, wherein the elevator fault information comprises fault recording information and elevator basic information, and the fault recording information comprises: elevator fault type, fault reason and fault time, elevator basic information includes: the production date of the elevator, the position of the elevator, the type of the elevator and the service life of the elevator.
Step 2, constructing a time sequence, which comprises two characteristics: 1) counting the failure times of various types of elevators, 2) basic information of the elevators, and combining the two characteristics to form a time sequence;
specifically, the "constructing a time series" includes: assuming that the tested elevators have M types, the effective fault types have N types, and the number of the elevator fault time windows is N, the number of the k-th fault types of the jth time window of the ith elevator can be recorded as the number of the k-th fault types of the jth time window of the ith elevator
Figure BDA0001774231350000081
1) Counting the times of various fault types by taking one week as a time window
Figure BDA0001774231350000082
2) The elevator basic information, the two characteristics are combined to form time series data, and the time series data are specifically expressed as follows:
Figure BDA0001774231350000083
wherein M represents the elevator model, d represents the elevator production date, l represents the elevator position, and N represents the number of elevator failure time windows, and double transverse lines are taken as examples in the representation to represent the statistical times of N failure types and the representation unit of the elevator basic information in the first time window of the Mth elevator.
And 3, constructing an event sequence, wherein the event sequence comprises two characteristics: 1) the data recording sequence of the elevator fault types is arranged in sequence according to the fault occurrence time, 2) the time interval between two adjacent fault events, and the two characteristics are combined to form an event sequence;
specifically, the "construction of the event sequence" is performed by: 1) storing all elevator information separately according to different elevator ids, converting elevator fault time into timestamp data, arranging elevator fault types according to fault occurrence time for each elevator, 2) calculating interval time of adjacent fault events of each elevator, and 3) forming a minimum event sequence unit by the fault type T and the timestamp interval I, wherein the minimum event sequence unit is specifically represented as follows:
Figure BDA0001774231350000084
the elevator id represents the unique identifier of the elevator-the elevator number.
Step 4, constructing an LSTM neural network;
the specific method comprises the following steps:
the specific formula of the recurrent neural network variant LSTM used in the invention is defined as follows:
it＝g(Wi xt+Uiyt-1+Vict-1+bi)，
ft＝g(Wfxt+Ufyt-1+Vfct-1+bf),
ct＝ftct-1+it⊙tanh(Wcxt+Ucyt-1+bc),
ot＝g(Woxt+Uoyt-1+Voct+bo),
yt＝ot⊙tanh(ct)
the above is the formula for the calculation of input gate, forget gate, cell state, and output gate in LSTM, where itFormula for calculating the input gate, ftFormula for calculating a forgetting gate, ctRepresenting the cell state calculation formula, otAnd ytFormula of calculation, o, which jointly represents the output gatetIndicating which part of the cell state is to be output using the g function, and finally ytIndicates that the cell state is treated with tanh and otMultiplying is the part which is determined to be output; wherein U and V are parameter matrixes to be learned in neural network training respectively, which represent the sequential multiplication of array elements, the function g adopts a sigmoid activation function,
Figure BDA0001774231350000091
a sequence of inputs is represented that is,
Figure BDA0001774231350000092
representing the generation of the hidden layer states, the above formula can be simplified as: (y)t,ct)＝LSTM(xtyt-1+ct-1)。
The LSTM construction process comprises the following specific steps:
(1) network initialization: determining the number i of input layer nodes, the number j of hidden layer nodes, the number k of output layer nodes and the state dimension of a unit of the network, initializing the connection weight among neurons of the input layer, the hidden layer and the output layer, and initializing the connection weight W of input gates, forgetting gates, output gates and cell unitsi,Wf,WO,WcInitializing the threshold bf,bc,bo,biGiving a learning rate and a neuron excitation function;
(2) calculating the output value of each neuron: f for LSTMt,it,ct,ot,ht；
(3) And (3) error calculation: inversely calculating an error term of each neuron according to the prediction output and the expected output matrix;
(4) updating the weight value: updating the network connection weight W according to the network prediction errori,Wf,WO,Wc；
(5) Updating a threshold value: updating network node threshold b according to network prediction errorf,bc,bo,bi；
(6) Judging whether the operation is finished or not, and if not, returning to the step (2);
(7) after finishing, using the trained double LSTM neural network to enter the step 5;
step 5, training a time sequence and an event sequence by using a double LSTM neural network to obtain a background knowledge representation and a history influence representation of the intensity function;
specifically, the method for training the time sequence and the event sequence by using the double LSTM neural network comprises the following steps:
the specific formula is as follows:
Figure BDA0001774231350000101
Figure BDA0001774231350000102
in the above formula
Figure BDA0001774231350000103
Which represents a time series of the images of the object,
Figure BDA0001774231350000104
represents a sequence of events, wherein ziRepresenting the type of event in the sequence of events, tiTime stamps indicating the occurrence of the events, the two sequences being used to learn background knowledge and historical impact, respectively.
Step 6, integrating background knowledge representation and historical influence representation through the joint layer;
the specific method comprises the following steps: adopting the tanh function as the joint function of the join layer to construct the nonlinear mapping of the strength function of the point process,
the formula of the nonlinear mapping of the point process intensity function constructed by the joint layer join layer is as follows:
Figure BDA0001774231350000105
step 7, predicting elevator fault types through a fault type prediction layer by using the strength function learned through double LSTM, and quantizing the loss values predicted by categories by using a classification loss layer;
the specific method comprises the following steps:
the calculation formulas of the main class and the subclass of the fault prediction are as follows:
Ut＝softMax(Wuet+bU)
ut＝softMax(Wu[et,Ut]+bu)
in the above formula, U and U represent the main and sub-classes, respectively, and WuAnd buRepresenting a parameter matrix to be learned by the model in the training process; first, e is divided by utilizing softMax functiontPredicting a failure primary category U as inputt(ii) a Then e is processed again by utilizing the softMax functiontAnd the main class prediction value UtThe predicted value of the subclass is calculated as input.
Step 8, predicting elevator fault time through a fault time prediction layer by using the strength function learned through double LSTM, and quantizing the loss value predicted by time by using a regression loss layer;
the specific method comprises the following steps:
st＝Wset+bs
wherein s istIs a time stamp, W, for each eventsAnd bsRepresenting the parameter matrix to be learned by the model during the training process.
Step 9, continuously and iteratively training a neural network model based on the step 7 and the step 8 to obtain an optimal network model, and then respectively predicting the type of the elevator fault and the elevator fault time through a fault type prediction layer and a fault time prediction layer by using the trained optimal model;
specifically, the method for respectively predicting the type of the elevator fault and the elevator fault time through the fault type prediction layer and the fault time prediction layer by using the trained optimal model comprises the following steps:
where the loss of the entire model is the sum of the time prediction loss and the event type prediction loss. And predicting the event type by using a cross entropy loss function, and predicting the time stamp by using a square loss function. The specific loss function formula is as follows:
Figure BDA0001774231350000111
in the above formula, N represents the total number of nodes, and the event points are indexed by l, i.e. l represents the second event, WuA parameter matrix representing the class of events learned by the neural network model,
Figure BDA0001774231350000112
is the time stamp of the next event point,
Figure BDA0001774231350000113
representing historical information of time points, adopting a Gaussian penalty function for a time loss function part:
Figure BDA0001774231350000114
wherein σ represents the uniform variance, and σ is taken2＝10；
And training the whole model in a circulating way by minimizing loss to finally obtain the optimal model.
And step 10, perfecting the optimization model, updating the elevator fault information base in real time, inputting the real-time updated data into the model to test the accuracy of the model, and correcting the perfection model according to the actual feedback condition.
Specifically, the method for "perfecting the optimization model" comprises the following steps: and obtaining an elevator fault prediction model through a training set, then obtaining a test set by utilizing a real-time updated fault information base, testing the accuracy of the model based on the test set, and perfecting the optimization model in real time.
The above steps 2 and 3 are shown in fig. 2, which shows the generation of the time series and the event series, and the two series are modeled cooperatively. Because the time sequence consists of basic information and statistical information, the time sequence carries background knowledge information of data, the asynchronously and randomly generated time stamp sequence of the event sequence carries bursty information, and more complete information connotation can be captured only by cooperatively modeling the two sequences.
Fig. 3 is a main block diagram of elevator fault type prediction and time of failure prediction, as shown in steps 5-9.
The embodiments of the present invention have been described in detail with reference to the accompanying drawings, but the present invention is not limited to the above embodiments, and various changes can be made within the knowledge of those skilled in the art without departing from the gist of the present invention.

Claims (10)

1. The elevator equipment fault prediction method based on deep learning is characterized by comprising the following steps:
step 1, establishing a real-time elevator fault information base, and screening effective elevator fault information as a preprocessing information sequence of a network according to four principles of fault information correlation, complete information, non-repeated information and non-human error operation of the fault information, wherein the elevator fault information comprises fault recording information and elevator basic information, and the fault recording information comprises: elevator fault type, fault reason and fault time, elevator basic information includes: the production date of the elevator, the position of the elevator, the type of the elevator and the service life of the elevator;
step 2, constructing a time sequence, which comprises two characteristics: 1) counting the failure times of various types of elevators, 2) basic information of the elevators, and combining the two characteristics to form a time sequence;
and 3, constructing an event sequence, wherein the event sequence comprises two characteristics: 1) the data recording sequence of the elevator fault types is arranged in sequence according to the fault occurrence time, 2) the time interval between two adjacent fault events, and the two characteristics are combined to form an event sequence;
step 4, constructing an LSTM neural network;
step 5, training a time sequence and an event sequence by using a double LSTM neural network to obtain a background knowledge representation and a history influence representation of the intensity function;
step 6, integrating background knowledge representation and historical influence representation through the joint layer;
step 7, predicting elevator fault types through a fault type prediction layer by using the strength function learned through double LSTM, and quantizing the loss values predicted by categories by using a classification loss layer;
step 8, predicting elevator fault time through a fault time prediction layer by using the strength function learned through double LSTM, and quantizing the loss value predicted by time by using a regression loss layer;
step 9, continuously and iteratively training a neural network model based on the step 7 and the step 8 to obtain an optimal network model, and then respectively predicting the type of the elevator fault and the elevator fault time through a fault type prediction layer and a fault time prediction layer by using the trained optimal model;
and step 10, perfecting the optimization model, updating the elevator fault information base in real time, inputting the real-time updated data into the model to test the accuracy of the model, and correcting the perfection model according to the actual feedback condition.
2. The deep learning based elevator equipment failure prediction method of claim 1, wherein:
the method of "constructing a time series" in step 2 is as follows: assuming that the tested elevators have M types, the effective fault types have N types, and the number of the elevator fault time windows is N, the number of the k-th fault types of the jth time window of the ith elevator can be recorded as the number of the k-th fault types of the jth time window of the ith elevator
Figure FDA0003234677060000011
1) Counting the times of various fault types by taking one week as a time window
Figure FDA0003234677060000012
2) The elevator basic information, the two characteristics are combined to form time series data, and the time series data are specifically expressed as follows:
Figure FDA0003234677060000021
wherein M represents the elevator model, d represents the elevator production date, l represents the elevator position, and N represents the number of elevator failure time windows, and double transverse lines are taken as examples in the representation to represent the statistical times of N failure types and the representation unit of the elevator basic information in the first time window of the Mth elevator.
3. The deep learning based elevator equipment failure prediction method of claim 1, wherein:
the method of "constructing an event sequence" in step 3 is as follows: 1) storing all elevator information separately according to different elevator ids, converting elevator fault time into timestamp data, arranging elevator fault types according to fault occurrence time for each elevator, 2) calculating interval time of adjacent fault events of each elevator, and 3) forming a minimum event sequence unit by the fault type T and the timestamp interval I, wherein the minimum event sequence unit is specifically represented as follows:
Figure FDA0003234677060000022
the elevator id represents the unique identifier of the elevator-the elevator number.
4. The deep learning based elevator equipment failure prediction method of claim 1, wherein:
the method for constructing the LSTM neural network in the step 4 comprises the following steps:
the specific formula of the recurrent neural network variant LSTM used is defined as follows:
it＝g(Wixt+Uiyt-1+Vict-1+bi)，
ft＝g(Wfxt+Ufyt-1+Vfct-1+bf),
ct＝ftct-1+it⊙tanh(Wcxt+Ucyt-1+bc),
ot＝g(Woxt+Uoyt-1+Voct+bo),
yt＝ot⊙tanh(ct)
the above is the formula for the calculation of input gate, forget gate, cell state, and output gate in LSTM, where itFormula for calculating the input gate, ftFormula for calculating a forgetting gate, ctRepresenting the cell state calculation formula, otAnd ytFormula of calculation, o, which jointly represents the output gatetIndicating which part of the cell state is to be output using the g function, and finally ytIndicates that the cell state is treated with tanh and otMultiplying is the part which is determined to be output; wherein U and V are parameter matrixes to be learned in neural network training respectively, which represent the sequential multiplication of array elements, the function g adopts a sigmoid activation function,
Figure FDA0003234677060000031
a sequence of inputs is represented that is,
Figure FDA0003234677060000032
representing the generation of the hidden layer states, the above formula can be simplified as: (y)t,ct)＝LSTM(xtyt-1+ct-1)。
5. The deep learning based elevator equipment failure prediction method of claim 1, wherein:
the method for training the time sequence and the event sequence by using the double LSTM neural network in the step 5 comprises the following steps:
the specific formula is as follows:
Figure FDA0003234677060000033
Figure FDA0003234677060000034
wherein z isiRepresenting the type of event in the sequence of events, tiTime stamps indicating the occurrence of the events, the two sequences being used to learn background knowledge and historical impact, respectively.
6. The deep learning based elevator equipment failure prediction method of claim 1, wherein:
the method for fusing the background knowledge representation and the historical influence representation through the joint layer in the step 6 comprises the following steps: adopting the tanh function as the joint function of the join layer to construct the nonlinear mapping of the strength function of the point process,
the formula of the nonlinear mapping of the point process intensity function constructed by the joint layer join layer is as follows:
Figure FDA0003234677060000035
7. the deep learning based elevator equipment failure prediction method of claim 1, wherein:
in the step 7, "predicting the elevator fault type through a fault type prediction layer by using the strength function learned through the double LSTM, and quantizing the loss value predicted by the category by using the classification loss layer" includes:
the calculation formulas of the failure prediction main class and the failure prediction subclass are as follows:
Ut＝softMax(Wuet+bU)
ut＝softMax(Wu[et,Ut]+bu)
in the above formula, U and U represent the main and sub-classes, respectively, and WuAnd buRepresenting a parameter matrix to be learned by the model in the training process; first, e is divided by utilizing softMax functiontPredicting a failure primary category U as inputt(ii) a Then e is processed again by utilizing the softMax functiontAnd the main class prediction value UtThe predicted value of the subclass is calculated as input.
8. The deep learning based elevator equipment failure prediction method of claim 6, wherein:
in the step 8, the elevator fault time is predicted through a fault time prediction layer by using the strength function learned through double LSTMs; the method for quantizing the loss value predicted by time by using the regression loss layer specifically comprises the following steps:
st＝Wset+bs
wherein s istIs a time stamp, W, for each eventsAnd bsRepresenting the parameter matrix to be learned by the model during the training process.
9. The deep learning based elevator equipment failure prediction method of claim 1, wherein:
in the step 9, "the type of the elevator fault and the elevator fault time are predicted respectively through the fault type prediction layer and the fault time prediction layer by using the trained optimal model", the method specifically comprises the following steps:
wherein the loss of the entire model is the sum of the time prediction loss and the event type prediction loss; predicting the event type by using a cross entropy loss function, and predicting the timestamp by using a square loss function; the specific loss function formula is as follows:
Figure FDA0003234677060000041
in the above formula, N represents the total number of nodes, and the event points are indexed by l, i.e. l represents the second event, WuA parameter matrix representing the class of events learned by the neural network model,
Figure FDA0003234677060000042
is the time stamp of the next event point,
Figure FDA0003234677060000043
history information representing the point in time of the time,
Figure FDA0003234677060000044
and (3) representing a fault prediction subclass of the event point, adopting a Gaussian penalty function for the time loss function part:
Figure FDA0003234677060000045
wherein σ represents the uniform variance, and σ is taken2＝10；
And training the whole model in a circulating way by minimizing loss to finally obtain the optimal model.
10. The deep learning based elevator equipment failure prediction method of claim 1, wherein:
the method for "perfecting the optimization model" in the step 10 comprises the following steps: and obtaining an elevator fault prediction model through a training set, then obtaining a test set by utilizing a real-time updated fault information base, testing the accuracy of the model based on the test set, and perfecting the optimization model in real time.
CN201810962887.0A 2018-08-22 2018-08-22 Elevator equipment fault prediction method based on deep learning Active CN109033450B (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN201810962887.0A CN109033450B (en) 2018-08-22 2018-08-22 Elevator equipment fault prediction method based on deep learning

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN201810962887.0A CN109033450B (en) 2018-08-22 2018-08-22 Elevator equipment fault prediction method based on deep learning

Publications (2)

Publication Number Publication Date
CN109033450A CN109033450A (en) 2018-12-18
CN109033450B true CN109033450B (en) 2021-11-05

Family

ID=64627980

Family Applications (1)

Application Number Title Priority Date Filing Date
CN201810962887.0A Active CN109033450B (en) 2018-08-22 2018-08-22 Elevator equipment fault prediction method based on deep learning

Country Status (1)

Country Link
CN (1) CN109033450B (en)

Families Citing this family (28)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN110009384A (en) * 2019-01-07 2019-07-12 阿里巴巴集团控股有限公司 Predict the method and device of operational indicator
CN109782603A (en) * 2019-02-03 2019-05-21 中国石油大学(华东) The detection method and monitoring system of rotating machinery coupling fault
CN109978026B (en) * 2019-03-11 2021-03-09 浙江新再灵科技股份有限公司 Elevator position detection method and system based on LSTM network
CN110135241B (en) * 2019-03-27 2021-08-31 浙江新再灵科技股份有限公司 Statistical analysis system of elevator stroke
CN110018425B (en) * 2019-04-10 2021-01-29 北京理工大学 Power battery fault diagnosis method and system
CN110008928A (en) * 2019-04-16 2019-07-12 杭州再灵云梯信息科技有限公司 The analysis method to elevator accident based on deep learning method
CN110264270B (en) * 2019-06-20 2020-09-01 京东数字科技控股有限公司 Behavior prediction method, behavior prediction device, behavior prediction equipment and storage medium
CN110348513A (en) * 2019-07-10 2019-10-18 北京华电天仁电力控制技术有限公司 A kind of Wind turbines failure prediction method based on deep learning
CN110580213A (en) * 2019-09-16 2019-12-17 浪潮软件股份有限公司 Database anomaly detection method based on cyclic marking time point process
CN111027591B (en) * 2019-11-13 2022-07-12 西安交通大学 Node fault prediction method for large-scale cluster system
CN110996066B (en) * 2019-12-19 2021-12-24 浙江中控技术股份有限公司 Accident backtracking method and device
CN111931819A (en) * 2020-07-13 2020-11-13 江苏大学 Machine fault prediction and classification method based on deep learning
CN112101431A (en) * 2020-08-30 2020-12-18 西南电子技术研究所（中国电子科技集团公司第十研究所） Electronic equipment fault diagnosis system
CN112083244B (en) * 2020-08-30 2022-10-28 西南电子技术研究所（中国电子科技集团公司第十研究所） Integrated intelligent diagnosis system for faults of avionic equipment
CN112232370A (en) * 2020-09-16 2021-01-15 中国电子产品可靠性与环境试验研究所（（工业和信息化部电子第五研究所）（中国赛宝实验室）） Fault analysis and prediction method for engine
CN112560338B (en) * 2020-12-10 2022-03-25 东北大学 Complex industrial system intelligent forecasting method, device, equipment and storage medium based on adaptive deep learning
CN113111585A (en) * 2021-04-15 2021-07-13 德州欧瑞电子通信设备制造有限公司 Intelligent cabinet fault prediction method and system and intelligent cabinet
CN113112004B (en) * 2021-04-21 2022-12-06 上海电力大学 Time-space prediction method based on Hox process and graph neural network
CN115385209B (en) * 2021-05-25 2023-08-29 太原理工大学 Alarm and password release method for abnormal opening of elevator door
CN113489744B (en) * 2021-07-27 2022-05-10 哈尔滨工业大学 Internet of things attack pattern recognition method based on hoxophone multivariate process modeling
CN113834184B (en) * 2021-08-18 2023-03-21 青岛海尔空调器有限总公司 Control method and device for air conditioner and server
CN113793227B (en) * 2021-09-16 2023-10-31 中国电子科技集团公司第二十八研究所 Intelligent human-like perception and prediction method for social network event
CN113884809A (en) * 2021-09-28 2022-01-04 国网黑龙江省电力有限公司 Secondary equipment fault positioning method and system based on neural network
CN114266201B (en) * 2022-03-01 2022-07-22 杭州市特种设备检测研究院(杭州市特种设备应急处置中心) Self-attention elevator trapping prediction method based on deep learning
CN114626467A (en) * 2022-03-17 2022-06-14 湖南优湖科技有限公司 Feature cross elevator trapping time series prediction model construction method based on deep learning, obtained model and prediction method
CN117034157B (en) * 2023-10-08 2024-01-12 广州健新科技有限责任公司 Hydropower equipment fault identification method and system combining multimodal operation data
CN117131457B (en) * 2023-10-26 2024-01-26 杭州海兴泽科信息技术有限公司 AI model-based electric power big data acquisition and processing method and system
CN118153919A (en) * 2024-05-10 2024-06-07 江苏中天互联科技有限公司 Equipment maintenance arrangement method and related equipment

Citations (7)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
WO2003097506A1 (en) * 2002-05-14 2003-11-27 Otis Elevator Company Neural network detection of obstructions within and motion toward elevator doors
CN101833596A (en) * 2010-03-31 2010-09-15 桂林电子科技大学 Two-damage quantitative identification method of rectangular and round section beam structure
CN102491140A (en) * 2011-12-27 2012-06-13 阳西县电梯配件有限公司 Characteristic-signal-based elevator safety checking device and elevator safety checking method
CN105844050A (en) * 2016-04-12 2016-08-10 吉林大学 Time correlation based replacement time calculation method for numerical control machine tool components
CN106125713A (en) * 2016-06-14 2016-11-16 西安交通大学 The assessment of reliability growth and Forecasting Methodology in the case of a kind of Interval censoring
CN106516923A (en) * 2016-08-31 2017-03-22 江苏鸿信***集成有限公司 Elevator running failure prediction method based on technology of Internet of Things
CN107992693A (en) * 2017-12-08 2018-05-04 中国北方发动机研究所(天津) A kind of cylinder head reliable life evaluation method

Patent Citations (7)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
WO2003097506A1 (en) * 2002-05-14 2003-11-27 Otis Elevator Company Neural network detection of obstructions within and motion toward elevator doors
CN101833596A (en) * 2010-03-31 2010-09-15 桂林电子科技大学 Two-damage quantitative identification method of rectangular and round section beam structure
CN102491140A (en) * 2011-12-27 2012-06-13 阳西县电梯配件有限公司 Characteristic-signal-based elevator safety checking device and elevator safety checking method
CN105844050A (en) * 2016-04-12 2016-08-10 吉林大学 Time correlation based replacement time calculation method for numerical control machine tool components
CN106125713A (en) * 2016-06-14 2016-11-16 西安交通大学 The assessment of reliability growth and Forecasting Methodology in the case of a kind of Interval censoring
CN106516923A (en) * 2016-08-31 2017-03-22 江苏鸿信***集成有限公司 Elevator running failure prediction method based on technology of Internet of Things
CN107992693A (en) * 2017-12-08 2018-05-04 中国北方发动机研究所(天津) A kind of cylinder head reliable life evaluation method

Non-Patent Citations (2)

* Cited by examiner, † Cited by third party
Title
Sensor System for Fault Detection Identification and Accommodation of Elevator of UAV;P Panitsrisit;《SICE Annual Conference 2011》;20110918;第1035-1040页 *
基于LSTM循环神经网络的故障时间序列预测;王鑫 等;《北京航空航天大学学报》;20180430;第44卷(第4期);第772-783页 *

Also Published As

Publication number Publication date
CN109033450A (en) 2018-12-18

Similar Documents

Publication Publication Date Title
CN109033450B (en) Elevator equipment fault prediction method based on deep learning
CN108197845B (en) Transaction index abnormity monitoring method based on deep learning model LSTM
CN109934337B (en) Spacecraft telemetry data anomaly detection method based on integrated LSTM
CN116757534B (en) Intelligent refrigerator reliability analysis method based on neural training network
CN111899510A (en) Intelligent traffic system flow short-term prediction method and system based on divergent convolution and GAT
CN110555230B (en) Rotary machine residual life prediction method based on integrated GMDH framework
CN111461551B (en) Deep learning and SPC criterion-based electric submersible pump fault early warning method
CN112838946B (en) Method for constructing intelligent sensing and early warning model based on communication network faults
CN109471698B (en) System and method for detecting abnormal behavior of virtual machine in cloud environment
CN112101431A (en) Electronic equipment fault diagnosis system
CN111445010B (en) Distribution network voltage trend early warning method based on evidence theory fusion quantum network
CN111126658A (en) Coal mine gas prediction method based on deep learning
CN110210169A (en) A kind of shield machine failure prediction method based on LSTM
CN110991776A (en) Method and system for realizing water level prediction based on GRU network
CN114266278A (en) Dual-attention-network-based method for predicting residual service life of equipment
CN117193222A (en) Intelligent quality control system based on industrial Internet of things and big data and control method thereof
CN113688770B (en) Method and device for supplementing long-term wind pressure missing data of high-rise building
CN116894364A (en) Post-earthquake damage rapid prediction method for high-speed railway track bridge system
CN115470850A (en) Water quality abnormal event recognition early warning method based on pipe network water quality time-space data
CN113536671B (en) Lithium battery life prediction method based on LSTM
CN113988210A (en) Method and device for restoring distorted data of structure monitoring sensor network and storage medium
CN117010683A (en) Operation safety risk prediction method based on hybrid neural network and multiple agents
CN116432839A (en) Backfill cost prediction method, device and equipment based on BP neural network
CN114401135B (en) Internal threat detection method based on LSTM-Attention user and entity behavior analysis technology
CN115392618A (en) AHP-EW and AE-RNN fusion-based food safety risk early warning model and establishment method thereof

Legal Events

Date Code Title Description
PB01 Publication
PB01 Publication
SE01 Entry into force of request for substantive examination
SE01 Entry into force of request for substantive examination
GR01 Patent grant
GR01 Patent grant