CN115510612A - Method for predicting remaining service life of lithium ion battery - Google Patents

Method for predicting remaining service life of lithium ion battery Download PDF

Info

Publication number
CN115510612A
CN115510612A CN202211018092.7A CN202211018092A CN115510612A CN 115510612 A CN115510612 A CN 115510612A CN 202211018092 A CN202211018092 A CN 202211018092A CN 115510612 A CN115510612 A CN 115510612A
Authority
CN
China
Prior art keywords
convolution
data
output
service life
battery
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.)
Pending
Application number
CN202211018092.7A
Other languages
Chinese (zh)
Inventor
杨伟
庞晓贤
王昱杰
潘卉楠
刘芝婷
范浩森
郑文芝
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.)
Guangzhou University
Original Assignee
Guangzhou University
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 Guangzhou University filed Critical Guangzhou University
Priority to CN202211018092.7A priority Critical patent/CN115510612A/en
Publication of CN115510612A publication Critical patent/CN115510612A/en
Pending legal-status Critical Current

Links

Images

Classifications

    • GPHYSICS
    • G06COMPUTING; CALCULATING OR COUNTING
    • G06FELECTRIC DIGITAL DATA PROCESSING
    • G06F30/00Computer-aided design [CAD]
    • G06F30/20Design optimisation, verification or simulation
    • GPHYSICS
    • G06COMPUTING; CALCULATING OR COUNTING
    • G06NCOMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00Computing arrangements based on biological models
    • G06N3/02Neural networks
    • G06N3/08Learning methods
    • GPHYSICS
    • G06COMPUTING; CALCULATING OR COUNTING
    • G06FELECTRIC DIGITAL DATA PROCESSING
    • G06F2111/00Details relating to CAD techniques
    • G06F2111/08Probabilistic or stochastic CAD
    • GPHYSICS
    • G06COMPUTING; CALCULATING OR COUNTING
    • G06FELECTRIC DIGITAL DATA PROCESSING
    • G06F2119/00Details relating to the type or aim of the analysis or the optimisation
    • G06F2119/04Ageing analysis or optimisation against ageing

Landscapes

  • Engineering & Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • Theoretical Computer Science (AREA)
  • General Physics & Mathematics (AREA)
  • Evolutionary Computation (AREA)
  • General Engineering & Computer Science (AREA)
  • Artificial Intelligence (AREA)
  • Computational Linguistics (AREA)
  • Health & Medical Sciences (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Computer Hardware Design (AREA)
  • Biomedical Technology (AREA)
  • Biophysics (AREA)
  • Geometry (AREA)
  • Data Mining & Analysis (AREA)
  • General Health & Medical Sciences (AREA)
  • Molecular Biology (AREA)
  • Computing Systems (AREA)
  • Mathematical Physics (AREA)
  • Software Systems (AREA)
  • Charge And Discharge Circuits For Batteries Or The Like (AREA)

Abstract

The invention relates to the field of lithium ion battery capacity detection, and discloses a method for predicting the remaining service life of a lithium ion battery, which comprises the following steps: respectively extracting V, I and T data in the charging process of each battery, and processing the data into matrix shapes required by two subnets; inputting data into an AFSC subnet, and adaptively giving a weight to each element in each charging characteristic spectrogram; the third step: inputting data into a ConvLSTM subnet to obtain hidden information of which the characteristic spectrogram fuses 20 kinds of cycle states; the contributions of the two subnets are fused through the two multilayer perceptrons, a high-accuracy early life prediction value is provided, and a predictor is guided to carry out RUL prediction; the life predictor remains muted until the percentage remaining useful life reaches a threshold value, and the predictor is activated to predict the remaining useful life. The method and the device improve the capability of the model for identifying the aging characteristics of the battery and provide high-precision prediction of the remaining service life of the battery.

Description

Method for predicting remaining service life of lithium ion battery
Technical Field
The invention relates to the field of lithium ion battery capacity detection, in particular to a method for predicting the remaining service life of a lithium ion battery.
Background
In recent years, lithium ion batteries have achieved significant advantages in terms of high energy density, reduced memory effect, low self-discharge rate, long life cycle, and the like. The method plays an important role in a plurality of fields such as electric automobiles, portable electronic products, aerospace, intelligent power systems and the like. However, over time, lithium ion batteries inevitably suffer from aging phenomena and performance degradation, which manifest as a decrease in capacity and an increase in internal resistance. Battery aging can lead to battery leakage, insulation damage, partial shorting. If the detection is not timely, more serious conditions may result. The method and the technology for predicting the service life and managing the health of the battery are used for evaluating the reliability of a system, detecting early faults and predicting the progress of the faults under the condition of an actual life cycle, so that a user can make a maintenance decision in advance and the loss caused by unexpected faults is prevented. Therefore, accurate prediction of the residual service life has very important practical significance for monitoring the health state of the battery, timely carrying out failure replacement and guaranteeing the use safety.
However, as the capacity fade of the battery is a typical long-time sequence data, it becomes a serious challenge to predict the capacity fade tendency at the later stage by using the limited early cycle data. In the current prediction models, most models extract HI, such as internal resistance, discharge power, alternating current impedance and the like, from the charging and discharging processes, the HI is usually difficult to obtain on line, and the HI, such as the discharge power and the like, is influenced by random load of equipment and changes variably, so that more data noise is brought to the models, and poor robustness is caused. Furthermore, the variations of the various HIs are very small during the early cycling of the battery, which makes the modeling work for early life prediction more challenging. Furthermore, the battery data set size of the training model greatly affects the generalization performance of the model.
In summary, the current lithium ion battery remaining service life prediction model has key challenges that the long-term dependence and gradient explosion problems caused by long-term sequence prediction cannot be solved, the prediction accuracy of the remaining service life is not high, the robustness is weak, the generalization performance of the model is poor and the like due to the key challenges that the model is interfered by random workload, the small difference in early cycle data cannot be recognized and predicted, and the used data set is small, so that a lithium ion battery remaining service life prediction method is provided.
Disclosure of Invention
Technical problem to be solved
Aiming at the defects of the prior art, the invention provides a method for predicting the remaining service life of a lithium ion battery, so as to solve the problems.
(II) technical scheme
In order to achieve the above purpose, the invention provides the following technical scheme:
a method for predicting the residual service life of a lithium ion battery comprises the following steps:
the first step is as follows: respectively extracting V, I and T data in the charging process of each battery as HI for measuring the aging trend of the battery, and processing the data into a matrix shape required by two subnets;
the second step is that: inputting data into an AFSC (active back-end separation) subnet, adaptively adjusting the contribution degrees of V, I and T to the model and endowing corresponding weights to the model respectively through deep convolution, global attention, point convolution and local attention, traversing sampling data of different cycles under the same characteristic by a local attention mechanism, and endowing a weight to each element in each charging characteristic spectrogram in a self-adaptive manner;
the third step: inputting data into a ConvLSTM subnet, performing primary processing on the data through deep convolution, adaptively adjusting the input weight of feature data in 20 circulation states by global attention, determining the retention, forgetting and output of features in an input time sequence through a gate structure by the ConvLSTM before local attention is embedded, and fusing hidden information of the 20 circulation states into an obtained feature spectrogram;
the fourth step: the contributions of the two subnets are fused through the two multilayer perceptrons, and a high-accuracy early life predicted value N is provided EOL And guiding the predictor to carry out RUL prediction;
the fifth step: the life predictor remains silent until the percentage remaining service life N RUL,% To a threshold of 10%, 7.5%, 5% or 2.5%, the predictor is activated for prediction of remaining useful life.
Preferably, the first step is specifically performed according to the following steps:
s1: setting an experimental data set consisting of 124 commercial lithium iron phosphate/graphite batteries, wherein the rated capacity is 1.1Ah, and the rated voltage is 3.3V;
s2: cycling to failure under 72 charging strategies and a fixed discharge rate in an Abin charging and discharging cabinet with 48 channels and a thermostat with the temperature of 30 ℃;
s3: according to the following steps of 8:2 is divided randomly into a training set and a test set;
s4: the first 5 cycles of each battery are taken as the most healthy measuring standard of the battery and are recorded as the initial state, and the data of 15 cycles before the prediction starting point are recorded as the real-time state.
Preferably, the second step has the steps of:
s1: deep convolution is carried out on an original 3-dimensional matrix, one convolution kernel is responsible for one channel, one channel is taken as an explanation and is marked as X (i,V) To X (i,V) Zero padding is performed. From X (i,V) Taking out the convolved sub-matrix, denoted X (i,V) (n), the convolution process can be expressed as:
Figure BDA0003812798350000031
wherein the content of the first and second substances,
Figure BDA0003812798350000032
and
Figure BDA0003812798350000033
are each from the k-th 1 The weight and offset of each 2-D convolution kernel, an | _ Hadamard product,
Figure BDA0003812798350000034
is X (i,V) (n) and (k) 1 The operation result of each convolution kernel, in order to perform individual convolution for each curve, the width and step of the convolution kernel are set to 1, the number of convolution kernels per convolution is 1, and X is (i,V) (n) the kth of the output having the same shape as the convolution kernel 1 The 2-D feature map may be expressed as:
Figure BDA0003812798350000035
the same operation is performed for the I and T channels, the output result will be batch normalized, max pooling and leak Rectified Linear Unit activation, the four layers are usually treated as one convolution Unit operation, and the leak ReLU activation function (α = 0.05) is as follows:
Figure BDA0003812798350000036
s2: defining the output volume as V, a global average pooling layer is used to obtain the attention scores s of the three channels n And adaptively adjusting the contributions of the N channel and the N channel to the model, and taking the characteristic mapping of the n channel in V as V n ,s n The calculation is as follows:
Figure BDA0003812798350000037
normalizing the obtained attention scores to obtain a final global weight factor
Figure BDA0003812798350000038
The global weight factor is used for self-adaptively giving the weight of the model to different feature mappings, so that the model can focus on feature extraction of important variables, and the calculation is as follows:
Figure BDA0003812798350000039
by connecting g n (n =1,2,. Cndot., K) will result in a weighted output volume
Figure BDA0003812798350000041
S3: to pair
Figure BDA0003812798350000042
Taking the sub-matrix of the mth convolution kernel size as the mark in zero filling
Figure BDA0003812798350000043
The convolution result of the point convolution can be expressed as:
Figure BDA0003812798350000044
Figure BDA0003812798350000045
is that
Figure BDA0003812798350000046
And kth 2 The operation results of the convolution kernels have the same shape before and after convolution,
Figure BDA0003812798350000047
and
Figure BDA0003812798350000048
are each from the k-th 2 Weight and offset of each 3-D convolution kernel, kth of output 2 The 2-D feature map may be expressed as:
Figure BDA0003812798350000049
s4: representing the output volume of the previous convolutional layer as F, defining a matrix A with the same shape as F, and defining an element A in A i,j For elements in a corresponding feature map
Figure BDA00038127983500000410
The attention weight value of (1) generates an attention weight matrix A through two full-connection layers, and the corresponding element A i,j Can be expressed as:
Figure BDA00038127983500000411
where δ and ω are weights, b and c are offsets, n Fc For the number of neurons, the lower corner is labeled as the index of the element in the matrix, g (-) and f (-) represent the hyperbolic tangent function and Sigmoid function, respectively, as follows:
Figure BDA00038127983500000412
Figure BDA00038127983500000413
the Sigmoid function controls the elements of the weight matrix within 0-1 and controls the elements by the Hadamard product of the two matrixes
Size of information flow entering next layer network, output volume L of output l As follows:
L l =A⊙F l
preferably, the local attention in the third step is to further extract useful information from the feature spectrogram after the convolution LSTM processing, so that the model set is focused on common features which all data frames have.
Preferably, the third step comprises the following specific steps:
the input matrix is embedded into a global attention mechanism after two deep convolution unit operations, the dimensionality of the obtained input volume is increased to be a 5-dimensional tensor to meet the input of ConvLSTM, and key equations of gate functions and data stream transmission in ConvLSTM cells are as follows:
f t =σ(W Xf *X t +W hf *h t-1 +b f );
i t =σ(W Xi *X t +W hg *h t-1 +b i );
g t =tanh(W Xg *X t +W hg *h t-1 +b g );
o t =σ(W Xo *X t +W ho *h t-1 +b o );
C t =f t ⊙C t-1 +i t ⊙g t
h t =o t ⊙tanh(C t );
wherein f represents the output of the forgetting gate, i represents the output of the input gate, o represents the output of the output gate, g represents the candidate memory, C represents the cell state of ConvLSTM, h is the hidden layer output, X represents the input, W X~ And W h~ Is a 2-D convolution kernel, the lower subscript t indicates the corresponding time, the symbol "+" indicates convolution operation, "" indicates hadamard product, the output of ConvLSTM adds local attention to become the output of subnet 2.
Preferably, the fourth step specifically includes:
connecting the outputs of the two subnets and defining as
Figure BDA0003812798350000051
The residual service life of the battery is predicted and calculated by fusing the contribution values of the two subnets through the two multilayer perceptronsThe following:
Figure BDA0003812798350000052
(III) advantageous effects
Compared with the prior art, the method for predicting the remaining service life of the lithium ion battery has the following beneficial effects:
1. the method for predicting the remaining service life of the lithium ion battery is based on the battery charging process data modeling, eliminates the interference of random workload on the model, and enables the model to have stronger robustness, practicability and generalization performance.
2. According to the method for predicting the remaining service life of the lithium ion battery, the V data, the I data and the T data which can be directly measured are used as HI, the data are simple and easy to obtain, and are strongly related to the service life of the battery, so that the real-time monitoring and the service life prediction of the health condition of the battery are facilitated.
3. The method for predicting the remaining service life of the lithium ion battery firstly provides a predictor which can be used for early prediction of the service life of the battery and RUL.
4. According to the method for predicting the remaining service life of the lithium ion battery, the AFSC subnet is good at adaptively adjusting the input weight of the early cycle data from the early cycle data and fusing different characteristics; the ConvLSTM subnet is good at capturing space-time characteristics from later-period cycle data, reduces redundancy of memory cells, effectively captures long-term dependence of long-sequence data, and reduces gradient explosion. The fusion of the two contributes to accurate prediction of early and late RUL simultaneously.
5. According to the method for predicting the remaining service life of the lithium ion battery, the addition of the attention mechanism improves the extraction of important information by the model, the capability of discarding useless information and the prediction precision of the model is further improved.
Drawings
FIG. 1 is a schematic flow chart of a method for predicting remaining service life according to an embodiment of the present invention;
FIG. 2 is a graph of capacity fade for 124 battery data sets used in an embodiment of the present invention;
FIG. 3 shows the difference in regularity of examples V, I and T of the present invention at different cycles;
FIG. 4 is a diagram of the data input structure of two sub-networks according to an embodiment of the present invention;
FIG. 5 is a schematic diagram of a model for predicting remaining service life of a battery according to an embodiment of the present invention;
FIG. 6 is a schematic diagram of a ConvLSTM structure according to an embodiment of the present invention;
FIG. 7 shows the early prediction performance of a model for battery life at different prediction start points according to an embodiment of the present invention;
FIG. 8 shows the RUL prediction performance of the model at different early warning points according to an embodiment of the present invention.
Detailed Description
The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the drawings in the embodiments of the present invention, and it is obvious that the described embodiments are only a part of the embodiments of the present invention, and not all of the embodiments. All other embodiments, which can be obtained by a person skilled in the art without making any creative effort based on the embodiments in the present invention, belong to the protection scope of the present invention.
Examples
Referring to fig. 1 to 8, a method for predicting remaining service life of a lithium ion battery according to an embodiment of the present invention includes the following steps:
step 1, respectively extracting V, I and T data in the charging process of each battery as HI for measuring the aging trend of the battery, and processing the data into a matrix shape required by two subnets.
The step 1 is implemented according to the following steps:
step 1.1, an experimental data set consists of 124 commercial lithium iron phosphate/graphite batteries, the rated capacity is 1.1Ah, and the rated voltage is 3.3V. These cells were cycled to failure in a 48-channel Abin charge-discharge cabinet and 30 ℃ incubator under 72 charge strategies and a fixed discharge rate. Data sets were as per 8: a ratio of 2 is randomly divided into a training set and a test set, which means that there will be 99 accumulated battery data with 52 charging strategies and 25 with 20 charging strategies for estimation of RUL and verification of model generalization performance, respectively.
Fig. 2 shows a graph of the capacity fade for 124 battery data sets.
Step 1.2, the V, I, T data of the battery show different degrees of deviation along with the circulation. Moreover, batteries with different service lives already show obvious differences in the circulation of the initial state of the battery, and the regular differences are strongly correlated with the service life of the battery. Thus, the input raw data should exhibit differences in different cycle states. The first 5 cycles of each cell were taken as the most healthy cell metric and recorded as the initial state. The data 15 cycles before the start of the prediction is recorded as the real-time status and used to compare with the initial status of the battery. The data input structure can be used for simultaneously carrying out early prediction on the service life of the battery and real-time prediction on the residual service life in the later period, the selection of a prediction starting point is depended, and the flexibility of a model is very high. For example, when the early life prediction is performed by using the V, I and T data of the first 20 cycles, the prediction start =20, the data structure is [1,5] + [6,20] (initial state + real-time state), and the number in the interval represents the cycle number used. For RUL prediction, for a 1000 cycle life battery, if the prediction start =800, the data structure is [1,5] + [786,800].
FIG. 3 shows the regularity differences between V, I and T at different cycles.
Fig. 4 shows the data input structure for two subnets.
Step 2, inputting data into an AFSC subnet, and adaptively adjusting the contribution degrees of V, I and T to the model and endowing corresponding weights to the model through the steps of deep convolution, global attention, point convolution, local attention and the like; the local attention mechanism traverses sampling data of different cycles under the same characteristic, and each element in each charging characteristic spectrogram is self-adaptively endowed with a weight, so that the AFSC subnet has high sensitivity to the change rule under each time sampling point.
The step 2 is implemented according to the following steps:
step 2.1, deepening the original 3-dimensional matrixDegree convolution, one convolution kernel is responsible for one channel. One of the channels (voltage) is taken as an explanation and is marked as X (i,V) . To avoid loss of edge data, for X (i,V) And (6) zero padding is carried out. From X (i,V) Taking out the convolved sub-matrices, denoted X (i,V) (n), the convolution process can be expressed as:
Figure BDA0003812798350000081
wherein, the first and the second end of the pipe are connected with each other,
Figure BDA0003812798350000082
and
Figure BDA0003812798350000083
are each from the kth 1 The weight and offset of each 2-D convolution kernel, an, is a Hadamard product.
Figure BDA0003812798350000084
Is X (i,V) (n) and (k) 1 The result of the convolution kernel operation. To convolve each curve individually, the width and step of the convolution kernel is set to 1, and the number of convolution kernels per convolution is 1. And X (i,V) (n) is the same shape as the convolution kernel. Kth of output 1 The 2-D feature map may be expressed as:
Figure BDA0003812798350000085
the same operation is carried out on the I channel and the T channel, and the output result is subjected to batch standardization, maximum pooling and Leaky Rectified Linear Unit (Leaky ReLU) activation so as to relieve the gradient explosion phenomenon in the training process. The above four layers are generally treated as one convolution unit operation. The leakage ReLU activation function (α = 0.05) is as follows:
Figure BDA0003812798350000086
step 2.2, define the output volume of the above step as V, and the global average pooling layer is used to obtain the attention scores s of the three channels n To adaptively adjust their contribution to the model. Taking the feature mapping of the nth channel in V as V n ,s n The calculation is as follows:
Figure BDA0003812798350000087
normalizing the obtained attention scores to obtain a final global weight factor
Figure BDA0003812798350000088
The global weight factor is used for self-adaptively giving the weight of the model to different feature mappings, so that the model can focus on feature extraction of important variables, and the calculation is as follows:
Figure BDA0003812798350000089
by connecting g n (n =1,2,. Cndot., K) will result in a weighted output volume
Figure BDA00038127983500000810
Used as input for the next layer.
Step 2.3, for
Figure BDA00038127983500000811
Taking the sub-matrix of the mth convolution kernel size as the mark in zero filling
Figure BDA00038127983500000812
The convolution result of the point convolution can be expressed as:
Figure BDA00038127983500000813
in a similar manner to that described above,
Figure BDA0003812798350000091
is that
Figure BDA0003812798350000092
And kth 2 The operation results of the convolution kernels have the same shape before and after convolution.
Figure BDA0003812798350000093
And
Figure BDA0003812798350000094
are each from the k-th 2 Weights and offsets of the individual 3-D convolution kernels. Kth of output 2 The 2-D feature map may be expressed as:
Figure BDA0003812798350000095
step 2.4, represent the output volume of the previous convolutional layer as F. In order to explore the contribution of data of each time sampling point in a direct measurement variable curve to a model, a matrix A with the same shape as that of F is defined, and an element A in the matrix A i,j For elements in a corresponding feature map
Figure BDA0003812798350000096
Attention weighting value of (1). The attention weight matrix A is generated by two fully-connected layers in this text, the corresponding element A i,j Can be expressed as:
Figure BDA0003812798350000097
where δ and ω are weights, b and c are offsets, n Fc The number of neurons. The lower corner is labeled as the index of the element in the matrix. g (-) and f (-) represent hyperbolic tangent function and Sigmoid function, respectively, as follows:
Figure BDA0003812798350000098
Figure BDA0003812798350000099
the Sigmoid function controls the elements of the weight matrix within 0-1, controls the size of information flow entering the next layer of network through the Hadamard product of the two matrixes, and outputs an output volume L l As follows:
L l =A⊙F l (11)
FIG. 5 shows a schematic structure of a novel lithium ion battery residual service life prediction model based on AFSC-ConvLSTM.
Step 3, inputting data into a ConvLSTM subnet, firstly carrying out primary processing on the data through deep convolution, and globally paying attention to adaptively adjust the input weight of the characteristic data in 20 circulation states; before local attention embedding, convLSTM determines the retention, forgetting and output of features in an input time sequence through a gate structure, and the obtained feature spectrogram fuses hidden information of 20 cycle states. Local attention is useful information for further extracting the feature spectrogram after convolution LSTM processing, so that the model is focused on common features of all data frames.
Step 3 is specifically implemented according to the following steps:
step 3.1, similar to the subnet 1, the input matrix is embedded into the global attention mechanism after two deep convolution unit operations, which is not described herein again. The obtained input volume is increased in dimension to become a 5-dimensional tensor to satisfy the input of ConvLSTM. Key equations for gating and data flow transmission within ConvLSTM cells are as follows:
f t =σ(W Xf *X t +W hf *h t-1 +b f ) (12)
i t =σ(W Xi* X t +W hg *h t-1 +b i ) (13)
g t =tanh(W Xg *X t +W hg *h t-1 +b g ) (14)
o t =σ(W Xo *X t +W ho *h t-1 +b o ) (15)
C t =f t ⊙C t-1 +i t ⊙g t (16)
h t =o t ⊙tanh(C t ) (17)
wherein f represents the output of the forgetting gate; i represents the input gate output; o represents the output of the output gate; g represents a candidate memory; c represents the cellular state of ConvLSTM; h is hidden layer output; x represents an input; w is a group of X~ And W h~ Is a 2-D convolution kernel; the lower subscript t indicates the corresponding time instant. The symbol "+" represents a convolution operation and "" represents a hadamard product. The output of ConvLSTM becomes the output of subnet 2 after adding local attention, which is not described herein.
FIG. 6 shows a schematic of the structure of ConvLSTM.
And 4, fusing the contributions of the two subnets through two multilayer perceptrons to provide a high-accuracy early life prediction value N EOL And direct the predictor to perform RUL prediction.
The step 4 specifically comprises the following steps:
connecting the outputs of the two subnets and defining as
Figure BDA0003812798350000101
And (3) predicting the residual service life of the battery by fusing the contribution values of the two subnets through the two multilayer perceptrons, wherein the calculation is as follows:
Figure BDA0003812798350000102
where the notation is similar to equation (8), except that f (-) in equation (18) are each expressed as an activation function as shown in equation (3).
FIG. 7 shows the early prediction performance of the model for battery life at different prediction starting points according to the disclosed method.
Step 5, the life predictor keeps silencing until the percentage of the remaining service lifeMin N RUL,% (relative to N) EOL In other words) to a threshold of 10%, 7.5%, 5% or 2.5%, the predictor is activated to make a prediction of the remaining useful life.
FIG. 8 shows the RUL prediction performance of the model at different early warning points according to the disclosed method.
The key points of the above embodiments of the present invention are: extracting V, I and T data in the charging process of each battery as HI; the HI is parallelly input into AFSC and ConvLSTM sub-networks, the AFSC adaptively adjusts the input weight of early cycle data and performs fusion among different characteristics, and the ConvLSTM is good at capturing space-time characteristics from later cycle data and capturing long-term dependence in an HI sequence; a global attention mechanism and a local attention mechanism are respectively embedded into the two subnetworks, the global attention mechanism is used for identifying key data frames in the V, I and T matrixes, the local attention mechanism is used for accurately selecting important characteristics of each data sampling point under the key data frames, and the capability of a model for identifying the aging characteristics of the battery is improved; and the output of the last two subnets is output through the model constructed by the two multilayer perceptrons, so that high-precision prediction of the remaining service life of the battery is provided.
Although embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions and alterations can be made in these embodiments without departing from the principles and spirit of the invention, the scope of which is defined in the appended claims and their equivalents.

Claims (6)

1. A method for predicting the remaining service life of a lithium ion battery is characterized by comprising the following steps:
the first step is as follows: respectively extracting V, I and T data in the charging process of each battery as HI for measuring the aging trend of the battery, and processing the data into a matrix shape required by two sub-networks;
the second step is that: inputting data into an AFSC (active back-end separation) subnet, adaptively adjusting the contribution degrees of V, I and T to the model and endowing corresponding weights to the model respectively through deep convolution, global attention, point convolution and local attention, traversing sampling data of different cycles under the same characteristic by a local attention mechanism, and endowing a weight to each element in each charging characteristic spectrogram in a self-adaptive manner;
the third step: inputting data into a ConvLSTM subnet, performing primary processing on the data through deep convolution, adaptively adjusting the input weight of feature data in 20 circulation states by global attention, determining the retention, forgetting and output of features in an input time sequence through a gate structure by the ConvLSTM before local attention is embedded, and fusing hidden information of the 20 circulation states into an obtained feature spectrogram;
the fourth step: the contributions of the two subnets are fused through the two multilayer perceptrons, and a high-accuracy early life predicted value N is provided EOL And guiding the predictor to carry out RUL prediction;
the fifth step: the life predictor remains muted until the percentage remaining useful life X RUL,% To a threshold of 10%, 7.5%, 5% or 2.5%, the predictor is activated to make a prediction of the remaining useful life.
2. The method for predicting the remaining service life of a lithium ion battery according to claim 1, wherein: the first step is specifically implemented according to the following steps:
s1: setting an experimental data set consisting of 124 commercial lithium iron phosphate/graphite batteries, wherein the rated capacity is 1.1Ah, and the rated voltage is 3.3V;
s2: cycling to failure under 72 charging strategies and a fixed discharge rate in an Abin charging and discharging cabinet with 48 channels and a thermostat with the temperature of 30 ℃;
s3: according to the following steps of 8:2 is divided randomly into a training set and a test set;
s4: the first 5 cycles of each battery are taken as the most healthy measuring standard of the battery and are recorded as the initial state, and the data of 15 cycles before the prediction starting point are recorded as the real-time state.
3. The method for predicting the remaining service life of a lithium ion battery according to claim 1, wherein: the second step has the steps of:
s1: carrying out deep convolution on an original 3-dimensional matrix, wherein one convolution kernel is responsible for one channel, and one channel is taken as an explanation and is marked as X (i,V) To X (i,V) And (6) zero padding is carried out. From X (i,V) Taking out the convolved sub-matrices, denoted X (i,V) (n), the convolution process can be expressed as:
Figure FDA0003812798340000021
wherein, the first and the second end of the pipe are connected with each other,
Figure FDA00038127983400000211
and
Figure FDA00038127983400000212
are each from the k-th 1 The weight and offset of each 2-D convolution kernel, a hadamard product,
Figure FDA0003812798340000022
is X (i,V) (n) and (k) 1 The operation result of each convolution kernel, in order to perform individual convolution for each curve, the width and step of the convolution kernel are set to 1, the number of convolution kernels per convolution is 1, and X is (i,V) (n) the kth of the output having the same shape as the convolution kernel 1 The 2-D feature map may be expressed as:
Figure FDA0003812798340000023
the same operation is performed for the I and T channels, the output result will be batch normalized, max pooling and leak Rectified Linear Unit activation, the four layers are usually treated as one convolution Unit operation, and the leak ReLU activation function (α = 0.05) is as follows:
Figure FDA0003812798340000024
s2: defining the output volume as V, a global average pooling layer is used to obtain the attention scores s of the three channels n Taking the characteristic mapping of the nth channel in V as V by adaptively adjusting the contribution of the channel to the model n ,s n The calculation is as follows:
Figure FDA0003812798340000025
normalizing the obtained attention scores to obtain a final global weight factor
Figure FDA0003812798340000026
The global weight factor is used for self-adaptively giving the weight of the model to different feature mappings, so that the model can focus on feature extraction of important variables, and the calculation is as follows:
Figure FDA0003812798340000027
by connecting g n (n =1,2, \ 8230;, K) will result in the weighted output volume
Figure FDA0003812798340000028
S3: to pair
Figure FDA0003812798340000029
Taking the sub-matrix of the mth convolution kernel size as the mark in zero filling
Figure FDA00038127983400000210
The convolution result of the point convolution can be expressed as:
Figure FDA0003812798340000031
Figure FDA0003812798340000032
is that
Figure FDA0003812798340000033
And k is 2 The operation results of the convolution kernels have the same shape before and after convolution,
Figure FDA0003812798340000034
and
Figure FDA0003812798340000035
are each from the k-th 2 Weight and offset of each 3-D convolution kernel, kth of output 2 The 2-D feature map may be expressed as:
Figure FDA0003812798340000036
s4: representing the output volume of the previous convolutional layer as F, defining a matrix A with the same shape as F, and defining an element A in A i,j For elements in a corresponding feature map
Figure FDA0003812798340000037
The attention weight value of (1) generates an attention weight matrix A through two full-connection layers, and the corresponding element A i,j Can be expressed as:
Figure FDA0003812798340000038
where δ and ω are weights, b and c are offsets, n Fc For the number of neurons, the lower corner is labeled as the index of the element in the matrix, g (-) and f (-) represent the hyperbolic tangent function and Sigmoid function, respectively, as follows:
Figure FDA0003812798340000039
Figure FDA00038127983400000310
the Sigmoid function controls the elements of the weight matrix within 0-1, controls the size of information flow entering the next layer of network through the Hadamard product of the two matrixes, and outputs an output volume L l As follows:
L l =A⊙F l
4. the method for predicting the remaining service life of the lithium ion battery according to claim 1, wherein: the local attention in the third step is used for further extracting useful information from the feature spectrogram after convolution LSTM processing, so that the model set is focused on common features of all data frames.
5. The method for predicting the remaining service life of the lithium ion battery according to claim 1, wherein: the third step comprises the following specific steps:
the input matrix is embedded into a global attention mechanism after two deep convolution unit operations, the dimension of the obtained input volume is increased to be 5-dimensional tensor to meet the input of ConvLSTM, and the key equations of gate functions and data stream transmission in ConvLSTM cells are as follows:
f t =σ(W Xf *X t +W hf *h t-1 +b f );
i t =σ(W Xi* X t +W hg *h t-1 +b i );
g t =tanh(W Xg *X t +W hg *h t-1 +b g );
o t =σ(W Xo *X t +W ho *h t-1 +b o );
C t =f t ⊙C t-1 +i t ⊙g t
h t =o t ⊙tanh(C t );
wherein f represents the output of the forgetting gate, i represents the output of the input gate, o represents the output of the output gate, g represents the candidate memory, C represents the cell state of ConvLSTM, h is the hidden layer output, X represents the input, W X~ And W h~ Is a 2-D convolution kernel, the subscript t indicates the corresponding time, the symbol "+" indicates a convolution operation, "" indicates a hadamard product, and the output of ConvLSTM becomes the output of subnet 2 after being added to local attention.
6. The method for predicting the remaining service life of the lithium ion battery according to claim 1, wherein: the fourth step specifically includes:
connecting the outputs of the two subnets and defining as
Figure FDA0003812798340000041
And (3) predicting the residual service life of the battery by fusing the contribution values of the two subnets through the two multilayer perceptrons, wherein the calculation is as follows:
Figure FDA0003812798340000042
CN202211018092.7A 2022-08-24 2022-08-24 Method for predicting remaining service life of lithium ion battery Pending CN115510612A (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN202211018092.7A CN115510612A (en) 2022-08-24 2022-08-24 Method for predicting remaining service life of lithium ion battery

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN202211018092.7A CN115510612A (en) 2022-08-24 2022-08-24 Method for predicting remaining service life of lithium ion battery

Publications (1)

Publication Number Publication Date
CN115510612A true CN115510612A (en) 2022-12-23

Family

ID=84501507

Family Applications (1)

Application Number Title Priority Date Filing Date
CN202211018092.7A Pending CN115510612A (en) 2022-08-24 2022-08-24 Method for predicting remaining service life of lithium ion battery

Country Status (1)

Country Link
CN (1) CN115510612A (en)

Cited By (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN116027204A (en) * 2023-02-20 2023-04-28 山东大学 Lithium battery residual service life prediction method and device based on data fusion
CN117556261A (en) * 2024-01-08 2024-02-13 浙江大学 MCNN-based diaphragm pump check valve service life prediction method and system

Cited By (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN116027204A (en) * 2023-02-20 2023-04-28 山东大学 Lithium battery residual service life prediction method and device based on data fusion
CN116027204B (en) * 2023-02-20 2023-06-20 山东大学 Lithium battery residual service life prediction method and device based on data fusion
CN117556261A (en) * 2024-01-08 2024-02-13 浙江大学 MCNN-based diaphragm pump check valve service life prediction method and system
CN117556261B (en) * 2024-01-08 2024-05-14 浙江大学 MCNN-based diaphragm pump check valve service life prediction method and system

Similar Documents

Publication Publication Date Title
CN115510612A (en) Method for predicting remaining service life of lithium ion battery
CN107797067B (en) Lithium ion battery life migration prediction method based on deep learning
Lee et al. Soft computing for battery state-of-charge (BSOC) estimation in battery string systems
CN111929602B (en) Single battery leakage or micro-short circuit quantitative diagnosis method based on capacity estimation
CN111652348A (en) Power battery pack fault fusion diagnosis method and system based on improved CNN
CN110703101B (en) Lithium ion battery inter-partition cycle capacity decline prediction method
CN110941929A (en) Battery health state assessment method based on ARMA and Elman neural network combined modeling
CN113406521B (en) Lithium battery health state online estimation method based on feature analysis
CN112051511A (en) Power battery state of health estimation method and system based on multichannel technology
CN112946522A (en) On-line monitoring method for short-circuit fault in battery energy storage system caused by low-temperature working condition
CN115656857A (en) Multi-scale fusion prediction method for remaining service life of lithium ion battery
CN106372272B (en) Lithium battery capacity and service life prediction method based on generalized degradation model and multi-scale analysis
CN113109715A (en) Battery health condition prediction method based on feature selection and support vector regression
CN114636932A (en) Method and system for predicting remaining service life of battery
CN113777515A (en) Electric vehicle charging safety early warning method
CN113820615A (en) Battery health degree detection method and device
CN114966436A (en) Lithium battery state of charge prediction method, device, equipment and readable storage medium
CN112731185A (en) Group matching method, group matching device and group matching system of lithium ion batteries
CN113791351B (en) Lithium battery life prediction method based on transfer learning and difference probability distribution
CN117169743A (en) Battery health state estimation method and device based on partial data and model fusion
CN116754960A (en) Cloud SOH estimation and residual life prediction method, device and program for sodium ion battery
CN116449244A (en) Lithium battery aging track early prediction method and system based on single discharge curve
CN115575840A (en) Early prediction method for remaining service life of lithium ion battery
Liu et al. A comparative research on power transformer fault diagnosis based on several artificial neural networks
CN114839539A (en) Lithium battery SOH estimation method based on multilevel sequence information self-adaptive fusion

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