Evaluation method of lithium ion battery health state model combining DAE and self-attention mechanism

Abstract

The invention relates to a model and method for evaluating the health state of a lithium ion battery, and the method comprises the steps: 1, extracting parameters, which are easy to monitor and obtain, of the lithium ion battery in practical application, carrying out the normalization processing of the data, and inputting the data into a subsequent network model; and step 2, adding noise in the input data, setting the number of hidden layers to determine an output dimension, coding and reconstructing the noise-containing data, and reducing the noise influence in the measurement process of the lithium ion battery. And step 3, constructing a training sample of the LSTM network by using a sliding window algorithm, learning dependency among data through a hidden layer of the LSTM network, and reserving important information. And step 4, inputting the feature quantity after activation processing of the hidden layer of the LSTM network into the self-attention layer, calculating the attention weight of each feature, and amplifying the feature value which has great influence on the target value. And 5, fusing the feature information of the previous layer, and mapping the fused feature information to an output space to obtain a model output result.

Description

Technical Field

[0001] The present invention provides a method for evaluating a lithium-ion battery health status model combining DAE and self-attention mechanism, which relates to a model and method for evaluating the health status of a lithium-ion battery. Background Art

[0002] Lithium-ion batteries are widely used in technical fields such as electric vehicles and portable electronic devices due to their high power density, long cycle life, and small size. However, as they age and lose capacity during use, they limit the effective range of electric vehicles. When the battery overheats or shorts, it can even cause catastrophic consequences such as explosions. Accurately estimating its health status, performing maintenance and replacement are of great significance to ensuring the safe and stable operation of the system and avoiding accidents.

[0003] At present, there are two methods for estimating the health status of batteries, one is model-based and the other is data-driven. The model-based method can reflect the changes in the physical and chemical aspects of the battery, but the structure is complex and depends on the accuracy and robustness of the built model. The data-driven method does not require the construction of a physical model, but only requires the extraction of features through a large amount of historical data to learn the trend of battery capacity attenuation. Both evaluation methods have corresponding shortcomings, so it is necessary to design a new method to better meet actual evaluation needs. Summary of the invention

[0004] The state of health (SOH) of a lithium-ion battery can characterize the usage of the battery. Accurate estimation can reflect the health of the battery and provide a basis for battery replacement. As the battery capacity is difficult to measure directly during application, this application proposes a lithium-ion battery SOH estimation method that combines a denoising autoencoder (DAE, hereinafter referred to as DAE) with a self-attention mechanism.

[0005] The steps of the evaluation method of the lithium-ion battery health status model combining DAE and self-attention mechanism in this application are as follows:

[0006] Establish the input layer, DAE layer, LSTM layer, self-attention layer, fully connected layer and output layer in sequence. The specific steps are as follows;

[0007] Step (1) Setting the input layer: extracting parameters that are easy to monitor in practical applications of lithium-ion batteries, normalizing the data, reducing the value range to [0, 1], and inputting them into the subsequent network model.

[0008] Step (2) setting the DAE layer: adding noise to the input data, setting the number of hidden layers to determine the output dimension, encoding and reconstructing the noisy data, and reducing the impact of noise during the lithium-ion battery measurement process.

[0009] Step (3) Setting up the LSTM layer: Using the sliding window algorithm to construct the training samples of the LSTM network, the data format is changed to the number of samples, step size and number of features. The dependencies between data are learned through the hidden layer of the LSTM network to retain important information.

[0010] Step (4) Setting the self-attention layer: Input the feature quantity after activation processing of the hidden layer of the LSTM network into the self-attention layer, calculate the query vector, key vector and value vector, and then calculate the attention weight of each feature, and amplify the feature value that has a greater impact on the target value.

[0011] Step (5) sets the fully connected layer: the feature information of the previous layer is fused and mapped to the output space to obtain the model output result, that is, the output layer.

[0012] Specifically, the parameters described in step (1) are the voltage and current of the lithium-ion battery in the charge and discharge curve, specifically the following eight health factors: the relationship between discharge voltage and discharge time, average discharge voltage, discharge voltage drop at equal time intervals, equal voltage rise charging time, equal time interval charging voltage rise, total discharge time, total discharge voltage drop and terminal voltage value.

[0013] Specifically, in step (2), the DAE layer is a denoising autoencoder, which includes an encoder, a hidden layer and a decoder; during the working process, the input data xi is first compressed and encoded into a latent space representation, that is, the data of the hidden layer, by the encoder, and then the data of the hidden layer is decoded and reconstructed into an output by the decoder.

[0014] The calculation formula for the encoding and decoding process is:

[0015] h=f(wx+b) (1)

[0016]

[0017] The denoising autoencoder adds q to the input D Binary random noise enables DAE to learn more robust features from noisy data x, which is calculated by the above formulas (1) and (2), reducing its sensitivity to small disturbances.

[0018] The denoising autoencoder is a multi-layer neural network for unsupervised learning. No additional target value information is required during training, and a back-propagation algorithm is used to make the output value approximately equal to the input value.

[0019] Specifically, after completing step (2), after adding noise to the original data through the denoising autoencoder to improve the robustness of the system, the number of hidden layers in the DAE layer is set to determine the subsequent dimensions, and the original input is described with less data.

[0020] Specifically, in step (3), the LSTM network is input, and the information that has a greater impact on the battery health is retained through the gated unit. Then, the weights of the eight battery health factors after dimensionality reduction are calculated through the self-attention layer, the internal relationship of the feature quantity is captured, and the key information characterizing the battery degradation is highlighted.

[0021] LSTM is an improved recurrent neural network that is used to solve the problems of gradient vanishing and gradient exploding that occur during the training process of RNN. Compared with the RNN neural network that only contains one tanh layer, LSTM consists of three parts: input gate, forget gate, and output gate. Among them, the input gate determines what needs to be saved in the current input and adds the state of the previous moment to the memory unit to form new memory information. The forget gate determines what information needs to be forgotten, and the output gate determines what information needs to be output based on the updated state.

[0022] x t represents the input at time t, h t represents the output at time t, f t represents the forget gate, i t represents the input gate, o t represents the output gate, C t Represents a memory unit, passing control information to the next moment. LSTM adds or deletes information to the memory unit through the gating unit.

[0023] Specifically, in step (4), a self-attention layer is set by letting x1, x2, ..., x n is the input data; h1, h2, …, h n The self-attention mechanism calculates the query vector q by mapping the input to different spaces. i , key vector k i With the value vector v i .

[0024] In the present invention, firstly, health factors are constructed from the voltage, current and time curves during the battery charging and discharging process, namely, eight categories: average discharge voltage, discharge voltage drop at equal time intervals, charging time with equal voltage rise, charging voltage rise at equal time intervals, total discharge time, total discharge voltage drop and terminal voltage value.

[0025] The denoising autoencoder is used for dimensionality reduction to improve the robustness of the model and reduce the impact of multicollinearity on the results. The data is then used as the input of the long short-term memory (LSTM) network, which is used to solve the problem of long-term dependency that the traditional recurrent neural network (RNN) cannot solve. Finally, the self-attention mechanism is introduced to extract the internal correlation of the data and establish an evaluation model for the health status.

[0026] By extracting the voltage and current parameters from the charge and discharge curves of lithium-ion batteries as input, DAE is used to add noise to the input to improve the robustness of the model and reduce the dimension. Based on the LSTM model, the self-attention mechanism is introduced to capture the internal connections between data and improve the overall accuracy of the model. Finally, the experiment proves that this method can predict the nonlinear attenuation process of capacity degradation of different types of batteries with high accuracy, and the root mean square error is maintained at about 1%.

[0027] A computer device includes a memory and a processor, wherein the memory stores a computer program, and is characterized in that the memory stores the computer program, and when the processor executes the computer program, the evaluation method of the lithium-ion battery health status model combining DAE and self-attention mechanism is implemented.

[0028] A computer-readable storage medium stores a computer program thereon, wherein the storage medium includes a stored program, wherein when the program is run, the device where the storage medium is located is controlled to execute the evaluation method of the lithium-ion battery health status model combining DAE and self-attention mechanism.

[0029] The key to achieving health status assessment is to extract key parameters that characterize capacity degradation from lithium-ion battery charge and discharge monitoring data. In order to address the problems of multicollinearity and low estimation accuracy among the extracted variables, this application proposes a health status assessment method that combines denois-ing autoencoder (DAE) with self-attention (SA). BRIEF DESCRIPTION OF THE DRAWINGS

[0030] Figure 1 The flowchart of the lithium-ion battery health status model combining DAE and self-attention mechanism of the present invention;

[0031] Figure 2 Schematic diagram of the structure of the autoencoder in the lithium-ion battery health status model combining DAE and self-attention mechanism in the present invention.

[0032] Figure 3This is a schematic diagram of the structure of the denoising autoencoder in the lithium-ion battery health status model that combines DAE and self-attention mechanism in the present invention.

[0033] Figure 4 This is a schematic diagram of the LSTM structure in the lithium-ion battery health status model that combines DAE and self-attention mechanism in the present invention.

[0034] Figure 5 Schematic diagram of the self-attention mechanism in the lithium-ion battery health status model combining DAE and self-attention mechanism in the present invention. DETAILED DESCRIPTION

[0035] The following is attached to the instruction manual Figure 1-5 The present invention is further described in detail:

[0036] The steps of the evaluation method of the lithium-ion battery health status model combining DAE and self-attention mechanism in this application are as follows:

[0037] See attached Figure 1 , establish the input layer, DAE layer, LSTM layer, self-attention layer, fully connected layer and output layer in sequence. The specific steps are as follows;

[0038] Step (1) Setting the input layer: extracting parameters that are easy to monitor in practical applications of lithium-ion batteries, normalizing the data, reducing the value range to [0, 1], and inputting them into the subsequent network model.

[0039] Step (2) setting the DAE layer: adding noise to the input data, setting the number of hidden layers to determine the output dimension, encoding and reconstructing the noisy data, and reducing the impact of noise during the lithium-ion battery measurement process.

[0040] Step (3) Setting up the LSTM layer: Using the sliding window algorithm to construct the training samples of the LSTM network, the data format is changed to the number of samples, step size and number of features. The dependencies between data are learned through the hidden layer of the LSTM network to retain important information.

[0041] Step (4) Setting the self-attention layer: Input the feature quantity after activation processing of the hidden layer of the LSTM network into the self-attention layer, calculate the query vector, key vector and value vector, and then calculate the attention weight of each feature, and amplify the feature value that has a greater impact on the target value.

[0042] Step (5) sets the fully connected layer: the feature information of the previous layer is fused and mapped to the output space to obtain the model output result, that is, the output layer.

[0043] Specifically, the parameters described in step (1) are the voltage and current of the lithium-ion battery in the charge and discharge curve, specifically the following eight health factors: the relationship between discharge voltage and discharge time, average discharge voltage, discharge voltage drop at equal time intervals, equal voltage rise charging time, equal time interval charging voltage rise, total discharge time, total discharge voltage drop and terminal voltage value.

[0044] The capacity decay of lithium-ion batteries is usually determined by its complex internal reactions and external conditions. The direct factors that can characterize its degradation trend are usually capacity and internal resistance, but such characteristic quantities are difficult to measure directly in actual processes, so the capacity decay is usually characterized by the selection of indirect health factors. The selection of indirect health factors is generally the easily measured charge and discharge voltage, current and time. In the present invention, a total of eight groups of health factors are extracted, namely the relationship between discharge voltage and discharge time, average discharge voltage, discharge voltage drop at equal time intervals, charging time with equal voltage rise, charging voltage rise at equal time intervals, total discharge time, total discharge voltage drop and terminal voltage value.

[0045] See attached Figure 2 and attached Figure 3 Specifically, in step (2), the DAE layer is a denoising autoencoder, which includes an encoder, a hidden layer and a decoder; during the working process, the input data xi is first compressed and encoded into a latent space representation, that is, the data of the hidden layer, by the encoder, and then the data of the hidden layer is decoded and reconstructed into an output by the decoder.

[0046] The calculation formula for the encoding and decoding process is:

[0047] h=f(wx+b) (1)

[0048]

[0049] The denoising autoencoder adds q to the input D Binary random noise enables DAE to learn more robust features from noisy data x, which is calculated by the above formulas (1) and (2), reducing its sensitivity to small disturbances.

[0050] The denoising autoencoder is a multi-layer neural network for unsupervised learning. No additional target value information is required during training, and a back-propagation algorithm is used to make the output value approximately equal to the input value.

[0051] Specifically, after completing step (2), after adding noise to the original data through the denoising autoencoder to improve the robustness of the system, the number of hidden layers in the DAE layer is set to determine the subsequent dimensions, and the original input is described with less data.

[0052] Specifically, in step (3), the LSTM network is input, and the information that has a greater impact on the battery health is retained through the gated unit. Then, the weights of the eight battery health factors after dimensionality reduction are calculated through the self-attention layer, the internal relationship of the feature quantity is captured, and the key information characterizing the battery degradation is highlighted.

[0053] LSTM is an improved recurrent neural network that is used to solve the problems of gradient vanishing and gradient exploding that occur during the training process of RNN. Compared with the RNN neural network that only contains one tanh layer, LSTM consists of three parts: input gate, forget gate, and output gate. Among them, the input gate determines what needs to be saved in the current input and adds the state of the previous moment to the memory unit to form new memory information. The forget gate determines what information needs to be forgotten, and the output gate determines what information needs to be output based on the updated state.

[0054] Figure 5 In, x t represents the input at time t, h t represents the output at time t, f t represents the forget gate, i t represents the input gate, o t represents the output gate, C t Represents a memory unit, passing control information to the next moment. LSTM adds or deletes information to the memory unit through the gating unit.

[0055] Specifically, in step (4), a self-attention layer is set by letting x1, x2, ..., x n is the input data; h1, h2, …, h n The self-attention mechanism calculates the query vector q by mapping the input to different spaces. i , key vector k i With the value vector v i .

[0056] The human brain cannot process all information at the same time with limited resources, and can only focus on specific places. If the neural network wants to process more information, it must increase the complexity of the model. In order to solve this problem, the attention mechanism comes on the stage. It does not need to pay attention to the entire amount of information, but only focuses on the features of key areas, which can avoid a lot of useless calculations. The self-attention mechanism is a variant of the attention mechanism. Its core idea is to capture the connection between long-distance information within the time series and improve the accuracy of the model output.

[0057] Firstly, health factors are constructed from the voltage, current and time curves during the battery charging and discharging process, namely, eight categories: average discharge voltage, discharge voltage drop at equal time intervals, charging time with equal voltage rise, charging voltage rise at equal time intervals, total discharge time, total discharge voltage drop and terminal voltage value.

[0058] The denoising autoencoder is used for dimensionality reduction to improve the robustness of the model and reduce the impact of multicollinearity on the results. The data is then used as the input of the long short-term memory (LSTM) network, which is used to solve the problem of long-term dependency that the traditional recurrent neural network (RNN) cannot solve. Finally, the self-attention mechanism is introduced to extract the internal correlation of the data and establish an evaluation model for the health status.

[0059] By extracting the voltage and current parameters from the charge and discharge curves of lithium-ion batteries as input, DAE is used to add noise to the input to improve the robustness of the model and reduce the dimension. Based on the LSTM model, the self-attention mechanism is introduced to capture the internal connections between data and improve the overall accuracy of the model. Finally, the experiment proves that this method can predict the nonlinear attenuation process of capacity degradation of different types of batteries with high accuracy, and the root mean square error is maintained at about 1%.

Claims

1. A method for evaluating a lithium-ion battery health status model combining DAE and self-attention mechanism, characterized in that: The steps are: Establish the input layer, DAE layer, LSTM layer, self-attention layer, fully connected layer and output layer in sequence. The specific steps are as follows; Step (1) setting the input layer: extracting parameters that are easy to monitor in practical applications of lithium-ion batteries, normalizing the data, reducing the value range to [0, 1], and inputting them into the subsequent network model; Step (2) setting the DAE layer: adding noise to the input data, setting the number of hidden layers to determine the output dimension, encoding and reconstructing the noisy data, and reducing the noise impact during the lithium-ion battery measurement process; Step (3) Setting up the LSTM layer: Using the sliding window algorithm to construct the training samples of the LSTM network, the data format is changed to the number of samples, step size and number of features, and the dependencies between data are learned through the hidden layer of the LSTM network to retain important information; Step (4) Setting the self-attention layer: Input the feature quantity after activation processing of the hidden layer of the LSTM network into the self-attention layer, calculate the query vector, key vector and value vector, and then calculate the attention weight of each feature, and amplify the feature value that has a greater impact on the target value; Step (5) sets the fully connected layer: the feature information of the previous layer is fused and mapped to the output space to obtain the model output result, that is, the output layer.

2. The evaluation method of the lithium-ion battery health status model combining DAE and self-attention mechanism according to claim 1 is characterized in that: Specifically, the parameters described in step (1) are the voltage and current of the lithium-ion battery in the charge and discharge curve, specifically the following eight health factors: the relationship between discharge voltage and discharge time, average discharge voltage, discharge voltage drop at equal time intervals, equal voltage rise charging time, equal time interval charging voltage rise, total discharge time, total discharge voltage drop and terminal voltage value.

3. The evaluation method of the lithium-ion battery health status model combining DAE and self-attention mechanism according to claim 1 is characterized in that: In step (2), the DAE layer is a denoising autoencoder, which includes an encoder, a hidden layer and a decoder. During the working process, the input data xi is first compressed and encoded into a latent space representation, i.e., the data of the hidden layer, by the encoder, and then the data of the hidden layer is decoded and reconstructed into an output by the decoder. The calculation formula for the encoding and decoding process is: h=f(wx+b) (1) The denoising autoencoder adds q to the input D Binomial random noise makes DAE from noisy data More robust features are learned from the above formula (1) and formula (2), which reduces its sensitivity to small disturbances.

4. The method for evaluating the health status model of a lithium-ion battery combining DAE and self-attention mechanism according to claim 3 is characterized in that: The denoising autoencoder is a multi-layer neural network for unsupervised learning. No additional target value information is required during training, and a back-propagation algorithm is used to make the output value approximately equal to the input value.

5. The method for evaluating the health status model of lithium-ion batteries by combining DAE and self-attention mechanism according to claim 3, characterized in that: In step (2), after adding noise to the original data through the denoising autoencoder to improve the robustness of the system, the number of hidden layers in the DAE layer is set to determine the subsequent dimensions and use less data to describe the original input.

6. The method for evaluating the health status model of a lithium-ion battery combining DAE and self-attention mechanism according to claim 2, characterized in that: In step (3), the LSTM network is input and the information that has a greater impact on battery health is retained through the gated unit. Then, the weights of the eight battery health factors after dimensionality reduction are calculated through the self-attention layer to capture the internal relationship of the feature quantity and highlight the key information that characterizes battery degradation.

7. The method for evaluating the health status model of a lithium-ion battery combining DAE and self-attention mechanism according to claim 6, characterized in that: The LSTM consists of three parts: input gate, forget gate and output gate. The input gate determines the content of the current input that needs to be saved and adds the state of the previous moment to the memory unit to form new memory information. The forget gate determines which information needs to be forgotten. The output gate determines the information to be output according to the updated state. x t represents the input at time t, h t represents the output at time t, f t represents the forget gate, i t represents the input gate, o t represents the output gate, C t Represents a memory unit, passing control information to the next moment. LSTM adds or deletes information to the memory unit through the gating unit.

8. The method for evaluating the health status model of lithium-ion batteries by combining DAE and self-attention mechanism according to claim 1, characterized in that: In step (4), a self-attention layer is set by letting x1, x2, ..., x n is the input data; h1, h2, …, h n is the corresponding output; the self-attention mechanism calculates the query vector q by mapping the input to different spaces i , key vector k i With the value vector v i .

9. A computer device comprising a memory and a processor, wherein the memory stores a computer program, wherein: The memory stores a computer program, and when the processor executes the computer program, it implements the evaluation method of the lithium-ion battery health status model combining DAE and self-attention mechanism.

10. A computer-readable storage medium having a computer program stored thereon, the storage medium comprising a stored program, wherein: When the program is running, the device where the storage medium is located is controlled to execute the evaluation method of the lithium-ion battery health status model combining DAE and self-attention mechanism.

Images