Battery SOH prediction method based on MH-MOE and EIS

Abstract

The present invention discloses a battery SOH prediction method based on MH-MOE and EIS, tests and collects battery EIS and SOH data, constructs an MH-MOE model based on a multi-head attention mechanism, and the MH-MOE model based on the multi-head attention mechanism includes 3 expert sub-models, 3 gated network models and 3 attention mechanism calculation modules. The MH-MOE model based on the multi-head attention mechanism is trained and verified to predict the battery SOH. After the MH-MOE model based on the multi-head attention mechanism is trained, the trained model is deployed under the premise that the model effect is good, and the trained model is used to predict the battery SOH. The present invention can predict the health status of the battery more efficiently and accurately, and adapt to complex and changeable battery working conditions.

Description

Technical Field

[0001] The invention relates to a battery SOH prediction method, in particular to a battery SOH prediction method based on MH-MOE and EIS, belonging to the technical field of battery detection. Background Art

[0002] The health of the battery, namely the "State of Health" (SOH), is a key indicator to measure the performance degradation and remaining service life of the battery. The importance of SOH is mainly reflected in the following aspects: it reveals the degree of battery aging, provides a basis for judging when to replace the battery, and helps to evaluate the remaining capacity of the battery. In addition, SOH plays a vital role in optimizing battery management strategies, analyzing potential battery failures, evaluating battery safety, and predicting the remaining service life of the battery. Predicting the state of health (SOH) of the battery is of great importance in practical applications.

[0003] Electrochemical impedance spectroscopy (EIS) is an effective method for analyzing the electrochemical reactions inside batteries. When the active substances in the battery decrease and the performance declines, the EIS curve will also change. Therefore, using EIS data to predict SOH is a feasible method.

[0004] Typically, we use EIS data to train machine learning algorithms, such as neural networks or support vector machines, to capture the changing patterns of battery SOH and make predictions.

[0005] Traditional machine learning models often use a single model to fit the entire data set, which is inadequate when faced with complex and highly heterogeneous data. Such models have limited generalization capabilities, are prone to overfitting problems, and are inefficient when processing large-scale data. Summary of the invention

[0006] The technical problem to be solved by the present invention is to provide a battery SOH prediction method based on MH-MOE and EIS to improve the battery SOH prediction accuracy.

[0007] In order to solve the above technical problems, the technical solution adopted by the present invention is:

[0008] A battery SOH prediction method based on MH-MOE and EIS comprises the following steps:

[0009] S1. Test and collect battery EIS and SOH data;

[0010] S2. Construct a MH-MOE model based on a multi-head attention mechanism. The MH-MOE model based on a multi-head attention mechanism includes three expert sub-models, three gated network models, and three attention mechanism calculation modules.

[0011] S3, training and verifying the MH-MOE model based on the multi-head attention mechanism;

[0012] S4. Predict battery SOH. After the MH-MOE model based on the multi-head attention mechanism is trained, the trained model is deployed on the premise that the model effect is good, and the trained model is used to predict the battery SOH.

[0013] Furthermore, the step S1 is specifically as follows: testing and collecting battery EIS data by an impedance spectrum tester, and the collected frequency range is 10 -2 Hz-10 5 Hz, the collected battery EIS data includes real battery EIS data real_1-real_m and imaginary battery EIS data imaginary_1-imaginary_m, m is the frequency number of battery EIS data, and SOH data is collected through the battery management system.

[0014] Furthermore, the expert sub-model adopts a CNN-LSTM model, which is composed of a CNN model and a LSTM model. The CNN model extracts spatial features of input data, and the LSTM model extracts temporal features of battery EIS data.

[0015] Furthermore, the CNN model includes a convolution layer and a maximum pooling layer, the activation functions of the convolution layer and the maximum pooling layer are both relu functions, the number of filters is set to 32, and the kernel size is 1; there is a 1*6*6 rectangle in the convolution layer as input data and becomes the input feature map, the input feature map is convolved with a 3*3 convolution kernel to obtain a 1*4*4 output feature map result; the pooling layer performs nonlinear downsampling on the input, and takes the maximum value or average value in the local neighborhood as the output, the pooling layer input feature map is the output feature map result of the convolution layer, the pooling layer input feature map size is 4*4, the maximum pooling and average pooling convolution kernel sizes are both 2*2, and the pooling layer output feature map size is 2*2.

[0016] Furthermore, the number of neurons in the LSTM model is 32, the activation functions are all relu functions, and the LSTM model includes an input gate, a forget gate, a state unit, an output gate, an activation function and an LSTM output layer.

[0017] Furthermore, the gated network models all adopt a multi-layer perceptron model, and the gated network models include a gated network model input layer, a gated network model middle layer and a gated network model output layer. The gated network model input layer receives the input battery EIS data, the dimension of the gated network model middle layer is 32, and the gated network model output layer outputs a 3-dimensional vector, and the 3-dimensional vectors correspond to the weights of the three expert sub-models respectively, and the sum of the weights of the three expert sub-models is 1.

[0018] Furthermore, the calculation steps of the attention mechanism calculation module are:

[0019] Step 1: Calculate the attention score;

[0020] First, calculate the similarity between each query vector and all key vectors:

[0021] attention_score i =query·key i

[0022] Among them, attention_score i is the i-th attention score, query is the query vector, key i is the i-th key vector;

[0023] Step 2: Calculate attention weight;

[0024] The obtained attention score is converted into a probability form through the softmax function, which is called the attention weight:

[0025] attention_weights=softmax(attention_scores)=

[0026] Among them, attention_weights is the attention weight, attention_scores is the attention score, softmax() is the softmax function, the softmax function ensures that all weights add up to 1, and attention_score i is the i-th attention score, attention_score j is the jth attention score;

[0027] Step 3: Calculate the weighted value

[0028] Multiply each value by its corresponding attention weight and add the results to get the final attention output:

[0029]

[0030] Among them, attention_output is the attention output, attention_weight i is the i-th attention weight, value i is the i-th value vector;

[0031] The comprehensive features are used as the input data of the attention mechanism calculation layer to obtain the attention output.

[0032] Furthermore, the feature representations output by the three expert sub-models are weighted and summed with the expert sub-model weights output by the three gating networks to obtain three weighted feature representations.

[0033] The three weighted feature representations are averaged to obtain the mean vector of the feature representation;

[0034] Connect the mean vector of feature representation to three attention mechanism calculation modules respectively;

[0035] Concatenate the output vectors of the three attention mechanism calculation modules;

[0036] Connect the concatenated output vector to a fully connected layer;

[0037] Connect the fully connected layer to a linear regression layer to predict SOH.

[0038] Furthermore, the step S3 is specifically as follows: implement the MH-MOE model code based on the multi-head attention mechanism on the Keras platform and train it, with the number of training epochs=100 and the number of batch samples batch_size=32; the optimizer for model training is the adam algorithm, and the loss function is the mean square error; the training data is divided into a training set and a test set in a ratio of 8:2, the model is trained on the training set, and the model is verified on the test set.

[0039] Compared with the prior art, the present invention has the following advantages and effects:

[0040] 1. The present invention uses the battery impedance EIS to predict the battery SOH based on the multi-head attention mechanism MH-MOE model, which has the advantages of diversified feature extraction brought by the multi-head gating mechanism, multi-angle feature learning and parallel computing of the multi-head attention mechanism, and the ability of the expert network to focus on different modes. Through these highlights, the MH-MOE model based on the multi-head attention mechanism can predict the health status of the battery more efficiently and accurately, and adapt to complex and changeable battery working conditions;

[0041] 2. The present invention improves the feature extraction capability through a multi-head gating mechanism. Multiple gating networks (heads) in the MH-MOE model can process input data in parallel. Each head can focus on extracting different features of the input data and capture complex information in the battery impedance spectrum. The multi-head mechanism can combine multiple feature representations to improve the model's sensitivity to changes in the battery health state, thereby improving the accuracy of SOH prediction.

[0042] 3. The present invention is based on a multi-head attention mechanism. Each head can learn different features of the input sequence, which helps the model capture more complex and richer information. At the same time, since each head is calculated independently, multi-head attention can take advantage of parallel computing to improve computing efficiency.

[0043] 4. Each expert network of the present invention can focus on different battery impedance patterns or characteristics, thereby improving the modeling capabilities of different battery states; through the collaboration of different expert networks, the MH-MOE model can handle complex patterns and nonlinear relationships in battery impedance data, thereby improving the robustness of SOH prediction. BRIEF DESCRIPTION OF THE DRAWINGS

[0044] Figure 1 It is a flow chart of the battery SOH prediction method based on MH-MOE and EIS of the present invention.

[0045] Figure 2 It is a structural diagram of the battery SOH prediction model based on MH-MOE and EIS of the present invention.

[0046] Figure 3 Schematic diagram of the convolutional layer of the present invention.

[0047] Figure 4 Schematic diagram of the pooling layer of the present invention.

[0048] Figure 5 It is a structural diagram of the LSTM model of the present invention.

[0049] Figure 6 It is a structural diagram of the gated network model of the present invention.

[0050] Figure 7 It is a structural diagram of the attention mechanism computing module of the present invention.

[0051] Figure 8 It is a schematic diagram of EIS and SOH data of a battery according to an embodiment of the present invention.

[0052] Fig. 9 Schematic diagram of the battery SOH prediction result according to an embodiment of the present invention.

[0053] Fig.10 It is a bar chart of the battery SOH prediction results of the embodiment of the present invention. DETAILED DESCRIPTION

[0054] In order to elaborate on the technical scheme adopted by the present invention to achieve the predetermined technical purpose, the technical scheme in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only partial embodiments of the present invention, rather than all embodiments, and the technical means or technical features in the embodiments of the present invention can be replaced without paying creative work. The present invention will be described in detail below with reference to the drawings and in conjunction with the embodiments.

[0055] like Figure 1 As shown, a battery SOH prediction method based on MH-MOE and EIS of the present invention comprises the following steps:

[0056] S1. Test and collect battery EIS and SOH data.

[0057] The battery EIS data was tested and collected by impedance spectroscopy test instrument, and the frequency range of the data was 10 -2 Hz-10 5 Hz, the collected battery EIS data includes real battery EIS data real_1-real_m and imaginary battery EIS data imaginary_1-imaginary_m, m is the number of frequencies of battery EIS data, and the number of frequencies is generally 5-60. SOH data is collected through the battery management system.

[0058] S2. Construct a MH-MOE model based on the multi-head attention mechanism, such as Figure 2 As shown in the figure, the MH-MOE model based on the multi-head attention mechanism contains 3 expert sub-models, 3 gating network models and 3 attention mechanism calculation modules.

[0059] Each expert sub-model adopts the CNN-LSTM model, which is composed of a CNN model and a LSTM model. The CNN model extracts the spatial features of the input data, and the LSTM model extracts the temporal features of the battery EIS data.

[0060] CNN can automatically learn the spatial correlation of input data and extract high-level abstract features.

[0061] The CNN model contains a convolution layer and a maximum pooling layer. The activation functions of the convolution layer and the maximum pooling layer are both relu functions. The number of filters is set to 32 and the kernel size is 1.

[0062] like Figure 3As shown in the figure, there is a 1*6*6 rectangle in the convolution layer as input data and becomes the input feature map. The input feature map is convolved through a 3*3 convolution kernel, that is, the tensor product is calculated starting from the 3*3 receiving field area in the upper left corner of the input, and moves to the right after each calculation with a stride of 1. After completing this calculation, a 1*4*4 output feature map result is obtained.

[0063] The pooling layer performs nonlinear downsampling on the input, and takes the maximum value (MaxPooling) or average value (Avg Pooling) in the local neighborhood as the output. The input feature map of the pooling layer is the output feature map result of the convolutional layer.

[0064] like Figure 4 As shown in the figure, the input feature map size of the pooling layer is 4*4, the maximum pooling and average pooling convolution kernel sizes are both 2*2, and the output feature map size of the pooling layer is 2*2.

[0065] Each convolution layer can extract high-level features of the input data, and the pooling layer implements downsampling to reduce the data dimension. The present invention preferably uses the maximum pooling method.

[0066] The LSTM model is a special RNN that can learn long-term dependencies and capture the dynamic changes of time series data.

[0067] The number of neurons in the LSTM model is 32, and the activation function is the relu function.

[0068] like Figure 5 As shown in the figure, the structure of the LSTM model contains the following components:

[0069] (1) Input gate: determines which parts of the current input information are allowed to enter the state unit.

[0070] (2) Forget gate: determines which information in the state unit needs to be forgotten or removed.

[0071] (3) Status unit: The core component used to store historical status information.

[0072] (4) Output gate: determines the information output to the next hidden layer.

[0073] (5) Activation function (Sigmoid, Tanh): used for nonlinear transformation.

[0074] (6) LSTM output layer: LSTM output data corresponding to the input data.

[0075] The LSTM model controls the flow of information entry, retention, and output by opening and closing the input gate, forget gate, and output gate, thereby realizing the memory and processing of time series information. Compared with simple RNN, LSTM can better capture the long-term dependencies in time series data, process longer time series, and has more powerful modeling capabilities.

[0076] like Figure 6 As shown in the figure, the gated network models all adopt a multi-layer perceptron model. The gated network models include a gated network model input layer, a gated network model middle layer and a gated network model output layer. The gated network model input layer receives the input battery EIS data, the dimension of the gated network model middle layer is 32, and the gated network model output layer outputs a 3-dimensional vector. The 3-dimensional vectors correspond to the weights of the three expert sub-models respectively, and the sum of the weights of the three expert sub-models is 1.

[0077] like Figure 7 As shown, the calculation steps of the attention mechanism calculation module are:

[0078] Step 1: Calculate the attention score;

[0079] First, calculate the similarity between each query vector and all key vectors:

[0080] attention_score i =query·key i

[0081] Among them, attention_score i is the i-th attention score, query is the query vector, key i is the i-th key vector;

[0082] Step 2: Calculate attention weight;

[0083] The obtained attention score is converted into a probability form through the softmax function, which is called the attention weight:

[0084] attention_weights=softmax(attention_scores)=

[0085] Among them, attention_weights is the attention weight, attention_scores is the attention score, softmax() is the softmax function, the softmax function ensures that all weights add up to 1, and attention_score i is the i-th attention score, attention_score j is the jth attention score;

[0086] Step 3: Calculate the weighted value

[0087] Multiply each value by its corresponding attention weight and add the results to get the final attention output:

[0088]

[0089] Among them, attention_output is the attention output, attention_weight i is the i-th attention weight, value i is the i-th value vector;

[0090] The comprehensive features are used as the input data of the attention mechanism calculation layer to obtain the attention output.

[0091] The input data are input into three expert sub-models and three gating network models respectively.

[0092] The three expert sub-models extract corresponding features from different angles, and the three gating network models independently calculate the weight of each expert sub-model.

[0093] The feature representations output by the three expert sub-models are weighted and summed with the expert sub-model weights output by the three gating networks to obtain three weighted feature representations.

[0094] The three weighted feature representations are averaged to obtain the mean vector of the feature representation.

[0095] The mean vector of the feature representation is connected to the three attention mechanism calculation modules respectively.

[0096] Concatenate the output vectors of the three attention mechanism calculation modules.

[0097] Connect the concatenated output vector to a fully connected layer.

[0098] Finally, the fully connected layer is connected to a linear regression layer to predict the battery SOH.

[0099] S3. Train and verify the MH-MOE model based on the multi-head attention mechanism.

[0100] The MH-MOE model code based on the multi-head attention mechanism is implemented and trained on the Keras platform. The number of training epochs=100 and the number of batch samples batch_size=32 are used during training. The optimizer for model training is the adam algorithm, and the loss function is the mean square error. The training data is divided into a training set and a test set in a ratio of 8:2. The model is trained on the training set and verified on the test set.

[0101] S4. Predict battery SOH. After the MH-MOE model based on the multi-head attention mechanism is trained, the trained model is deployed on the premise that the model effect is good, and the trained model is used to predict the battery SOH.

[0102] The present invention tests the effect on a public data set and uses the impedance data therein to predict the battery SOH. Figure 8 As shown in the figure, there are 1358 battery EIS and SOH data. EIS has 120 features, and the feature numbers are f1 to f120. The first 60 features are the real part data of the impedance data corresponding to the frequency from high to low, and the last 60 features are the imaginary part data of the impedance data corresponding to the frequency from high to low.

[0103] like Figure 8 As shown, the mean of the test set of battery capacity is 30.6249, the standard deviation is 4.1309, and the mean of the prediction set is 30.5339, the standard deviation is 4.1291, which shows that the data distribution of the prediction results is basically consistent with the data distribution of the real data.

[0104] The root mean square error between the prediction set and the test set of battery SOH is only 0.176, indicating that the prediction effect of the model is good.

[0105] like Fig. 9 As shown in the figure, the solid line represents the true value, the dotted line represents the predicted value, and the solid square represents the error value. The numbers represent the true values. It can be seen from the figure that the prediction effect is good.

[0106] The present invention uses the battery impedance EIS to predict the battery SOH based on the MH-MOE model of the multi-head attention mechanism. It has the advantages of diversified feature extraction brought by the multi-head gating mechanism, multi-angle feature learning of the multi-head attention mechanism, parallel computing, and the ability of the expert network to focus on different modes. Through these highlights, the MH-MOE model based on the multi-head attention mechanism can more efficiently and accurately predict the health status of the battery and adapt to complex and changeable battery working conditions. The present invention improves the feature extraction capability through the multi-head gating mechanism. The multiple gated networks (heads) in the MH-MOE model can process the input data in parallel, and each head can focus on extracting different features of the input data to capture the complex miscellaneous information; the multi-head mechanism can combine multiple feature representations to improve the model's sensitivity to changes in the battery's health state, thereby improving the accuracy of SOH prediction; the present invention is based on a multi-head attention mechanism, and each head can learn different features of the input sequence, which helps the model capture more complex and richer information. At the same time, since each head is calculated independently, multi-head attention can take advantage of parallel computing to improve computing efficiency; each expert network of the present invention can focus on different battery impedance modes or features, thereby improving the modeling capabilities of different battery states; through the collaboration of different expert networks, the MH-MOE model can handle complex patterns and nonlinear relationships in battery impedance data, thereby improving the robustness of SOH prediction.

[0107] The above is only a preferred embodiment of the present invention and does not limit the present invention in any form. Although the present invention has been disclosed as a preferred embodiment as above, it is not used to limit the present invention. Any technician familiar with this profession can make some changes or modify the technical contents disclosed above into equivalent embodiments without departing from the scope of the technical solution of the present invention. However, any simple modification, equivalent replacement and improvement made to the above embodiments without departing from the content of the technical solution of the present invention, based on the technical essence of the present invention, within the spirit and principles of the present invention, still fall within the protection scope of the technical solution of the present invention.

Claims

1. A battery SOH prediction method based on MH-MOE and EIS, characterized in that The following steps are involved: S1. Test and collect battery EIS and SOH data; S2. Construct an MH-MOE model based on a multi-head attention mechanism. The MH-MOE model based on a multi-head attention mechanism includes three expert sub-models, three gated network models, and three attention mechanism calculation modules. The gated network models all adopt a multi-layer perceptron model. The gated network models include a gated network model input layer, a gated network model middle layer and a gated network model output layer. The gated network model input layer receives the input battery EIS data. The dimension of the gated network model middle layer is 32. The gated network model output layer outputs a 3-dimensional vector. The 3-dimensional vectors correspond to the weights of the three expert sub-models respectively, and the sum of the weights of the three expert sub-models is 1. The feature representations output by the three expert sub-models are weighted and summed with the weights of the expert sub-models output by the three gating networks to obtain three weighted feature representations; The three weighted feature representations are averaged to obtain the mean vector of the feature representation; Connect the mean vector of feature representation to three attention mechanism calculation modules respectively; Concatenate the output vectors of the three attention mechanism calculation modules; Connect the concatenated output vector to a fully connected layer; Connect the fully connected layer to a linear regression layer to predict SOH; S3, training and verifying the MH-MOE model based on the multi-head attention mechanism; S4. Predict battery SOH. After the MH-MOE model based on the multi-head attention mechanism is trained, the trained model is deployed on the premise that the model effect is good, and the trained model is used to predict the battery SOH.

2. The battery SOH prediction method based on MH-MOE and EIS according to claim 1, characterized in that: The step S1 specifically includes: testing and collecting battery EIS data through an impedance spectrum tester, and the collected frequency range is 10 -2 Hz-10 5 Hz, the collected battery EIS data includes real battery EIS data real_1-real_m and imaginary battery EIS data imaginary_1-imaginary_m, m is the frequency number of battery EIS data, and SOH data is collected through the battery management system.

3. The battery SOH prediction method based on MH-MOE and EIS according to claim 1, characterized in that: The expert sub-model adopts a CNN-LSTM model, which is composed of a CNN model and an LSTM model. The CNN model extracts spatial features of input data, and the LSTM model extracts temporal features of battery EIS data.

4. The battery SOH prediction method based on MH-MOE and EIS according to claim 3 is characterized in that: The CNN model includes a convolution layer and a maximum pooling layer. The activation functions of the convolution layer and the maximum pooling layer are both relu functions, the number of filters is set to 32, and the kernel size is 1; there is a 1*6*6 rectangle in the convolution layer as input data and becomes the input feature map. The input feature map is convolved with a 3*3 convolution kernel to obtain a 1*4*4 output feature map result; the pooling layer performs nonlinear downsampling on the input, and takes the maximum value or average value in the local neighborhood as the output. The input feature map of the pooling layer is the output feature map result of the convolution layer. The size of the input feature map of the pooling layer is 4*4, the size of the maximum pooling and average pooling convolution kernels are both 2*2, and the size of the output feature map of the pooling layer is 2*2.

5. The battery SOH prediction method based on MH-MOE and EIS according to claim 3, characterized in that: The number of neurons in the LSTM model is 32, and the activation functions are all relu functions. The LSTM model includes an input gate, a forget gate, a state unit, an output gate, an activation function and an LSTM output layer.

6. The battery SOH prediction method based on MH-MOE and EIS according to claim 1, characterized in that: The calculation steps of the attention mechanism calculation module are: Step 1: Calculate the attention score; First, calculate the similarity between each query vector and all key vectors: attention_score i =query·key i Among them, attention_score i is the i-th attention score, query is the query vector, key i is the i-th key vector; Step 2: Calculate attention weight; The obtained attention score is converted into a probability form through the softmax function, which is called the attention weight: attention_weights=softmax(attention_scores)= ; Among them, attention_weights is the attention weight, attention_scores is the attention score, softmax() is the softmax function, the softmax function ensures that all weights add up to 1, and attention_score i is the i-th attention score, attention_score j is the jth attention score; Step 3: Calculate the weighted value Multiply each value by its corresponding attention weight and add the results to get the final attention output: ； Among them, attention_output is the attention output, attention_weight i is the i-th attention weight, value i is the i-th value vector; The comprehensive features are used as the input data of the attention mechanism calculation layer to obtain the attention output.

7. The battery SOH prediction method based on MH-MOE and EIS according to claim 1, characterized in that: The step S3 is specifically as follows: implementing the MH-MOE model code based on the multi-head attention mechanism on the Keras platform and training it, with the number of training epochs=100 and the number of batch samples batch_size=32; the optimizer for model training is the adam algorithm, and the loss function is the mean square error; the training data is divided into a training set and a test set in a ratio of 8:2, the model is trained on the training set, and the model is verified on the test set.

Images