Lithium ion battery health life evaluation method and system based on feature image

Abstract

The invention discloses a lithium ion battery health life assessment method and system based on a feature image, and the method comprises the steps: pre-training a battery health life assessment model through the sample data of a source domain, and building a mapping relation between the feature image and the health life of a lithium ion battery, and the feature image is generated based on an incremental capacity feature; introducing transfer learning and utilizing sample data of a target domain to optimize a battery health life evaluation model; and extracting incremental capacity features for the to-be-evaluated lithium ion battery in the target domain, generating a feature image of the lithium ion battery, and inputting the feature image into the optimized battery health life evaluation model to obtain the corresponding health life. According to the method, the lithium ion battery health life evaluation model obtained based on source domain data training can better adapt to the battery characteristics and working conditions of the target domain so as to solve the problems of battery aging and differences influenced by various external environment factors, and therefore the accuracy of lithium ion battery health life evaluation on the target domain is improved.

Description

Technical Field

[0001] The present invention belongs to the technical field of batteries, and particularly relates to a method and system for evaluating the health life of lithium-ion batteries based on feature images. Background Art

[0002] Due to advantages such as high energy density, light weight, and long life, lithium-ion batteries are widely used in fields such as portable electronic devices, electric vehicles, energy storage systems, and military equipment. However, with the aging of the battery and the influence of various external environmental factors, its available capacity and stability gradually decrease, which may lead to safety risks such as leakage or short circuit. Accurately evaluating the state of health (SOH) of lithium-ion batteries is crucial for ensuring the stability and reliability of related devices. However, in practical applications, current feature extraction technologies face challenges such as complex processes and difficulties in capturing the dynamic evolution of time series data by models. Existing SOH estimation methods are mainly divided into two categories: model-based and data-driven. The former, such as electrochemical models and equivalent circuit models, have problems such as complexity, poor adaptability, or the need for a large amount of computing resources and prior knowledge; although the latter is highly flexible, traditional one-dimensional feature analysis methods also have limitations when dealing with battery data. Therefore, a new and more effective method for evaluating the SOH of lithium-ion batteries is needed. Summary of the Invention

[0003] The technical problem to be solved by the present invention: In view of the above problems of the prior art, a method and system for evaluating the health life of lithium-ion batteries based on feature images are provided. The present invention aims to enable the lithium-ion battery health life evaluation model trained based on source domain data to better adapt to the battery characteristics and working conditions of the target domain to solve the differences in battery aging and the influence of various external environmental factors, thereby improving the accuracy of the lithium-ion battery health life evaluation in the target domain.

[0004] To solve the above technical problems, the technical solution adopted by the present invention is as follows: A method for evaluating the health life of lithium-ion batteries based on feature images, comprising the following steps: using the sample data of lithium-ion batteries in the source domain with similar electrochemical characteristics but different working conditions to pre-train a battery health life evaluation model, so that the battery health life evaluation model establishes a mapping relationship between the feature image of the lithium-ion battery and its corresponding health life, and the feature image of the lithium-ion battery is generated based on the incremental capacity feature of the lithium-ion battery; introducing transfer learning to optimize the battery health life evaluation model using the sample data of the lithium-ion battery in the target domain; extracting the incremental capacity feature for the lithium-ion battery to be evaluated in the target domain and generating a feature image of the lithium-ion battery, and inputting the feature image of the lithium-ion battery into the optimized battery health life evaluation model to obtain the corresponding health life.

[0005] Optionally, the generation of the feature image of the lithium-ion battery includes: S101, collect the capacity and voltage of the lithium-ion battery at different cycle times through multiple cycle periods respectively; S102, determine the corresponding capacity change Δ V corresponding to different preset fixed voltage intervals Δ Q , and calculate the incremental capacity of each sample point corresponding to each cycle time according to the following formula:

[0006] In the above formula, IC is the incremental capacity, dQ / dV is the derivative of the capacity with respect to the voltage, Δ Q is the capacity change of the lithium-ion battery; Δ V is the preset fixed voltage interval; S103, calculate the average value of the incremental capacity of all cycle periods at a certain cycle time according to the following formula:

[0007] where, is the average value of the incremental capacity of all cycle periods at a certain cycle time, is the number of cycle periods, is the incremental capacity of the i-th cycle period of the lithium-ion battery, and finally obtain the incremental capacity curve represented by the time series composed of the average values of the incremental capacities at each cycle time of the lithium-ion battery; S104, perform smoothing processing on the incremental capacity curve represented by the time series composed of the average values of the incremental capacities at each cycle time of the lithium-ion battery by using the Kalman filtering method; S105, convert the incremental capacity curve represented by the time series composed of the average values of the incremental capacities at each cycle time of the smoothed lithium-ion battery into a two-dimensional image by using the relative position matrix.

[0008] Optionally, converting the smoothed battery incremental capacity curve into a two-dimensional image by using the relative position matrix in step S105 includes: S201, perform zscore normalization on the average value of the incremental capacity in the time series composed of the average values of the incremental capacities at each cycle time of the smoothed lithium-ion battery; S202, reduce the dimension of the average value of the incremental capacity of the standard normal distribution obtained after zscore normalization according to the following formula: , , where, is the i-th data value obtained after dimension reduction, is the dimension reduction factor, is the average value of the j-th incremental capacity after zscore normalization, is the number of cycle times within a cycle period, is the dimension of the data value after dimensionality reduction; S203. Construct a relative position matrix of size as shown in the following formula based on the data values obtained after dimensionality reduction : , wherein, ~ are respectively the 1st to th data values obtained after dimensionality reduction; S204. Apply min-max normalization to convert the relative position matrix into a grayscale value matrix as the converted two-dimensional image: , wherein, is the grayscale value matrix, min represents taking the minimum value, and max represents taking the maximum value.

[0009] Optionally, the battery health life evaluation model is a ResNet neural network model. The processing of the input characteristic image of the lithium-ion battery by the ResNet neural network model includes: extracting features from the characteristic image of the lithium-ion battery through a residual unit, and the residual unit includes a plurality of cascaded residual blocks; and the output features of the residual unit are globally average pooled through a global average pooling layer, and then classified or regressed through a classification regression module composed of multiple fully connected layers and activation function layers to obtain the health life of the lithium-ion battery.

[0010] Optionally, the functional expression of the residual block is: Or , wherein, is the output feature of the residual block, is the input feature of the residual block, represents the residual mapping of the input feature of the residual block based on the weight parameter , is a linear mapping, and is adopted when the residual mapping and the input feature dimension of the residual block are the same; is adopted when the residual mapping and the input feature dimension of the residual block are different; the residual mapping is composed of a plurality of cascaded convolutional modules, and each level of convolutional module is composed of a series-connected convolutional layer, a batch normalization layer and an activation function.

[0011] Optionally, when pre-training the battery health life evaluation model, it further includes determining the optimal model parameters of the battery health life evaluation model by using Bayesian optimization technology and fine-tuning the model.

[0012] Optionally, when introducing transfer learning to optimize the battery health life evaluation model by using the sample data of lithium-ion batteries in the target domain, it includes, for the pre-trained battery health life evaluation model, dividing the multiple cascaded residual blocks in the residual unit into the bottom-layer residual blocks and the high-layer residual blocks, freezing the parameters of the bottom-layer residual blocks so that they do not participate in backpropagation, and adding a random dropout layer after the fully connected layer to randomly discard some neurons to enhance the generalization ability of the model, and using the sample data of lithium-ion batteries in the target domain, through the given loss function and a learning rate smaller than that when pre-training the battery health life evaluation model to backpropagate and optimize the parameters of the global average pooling layer of the high-layer residual blocks and the fully connected layer in the classification and regression module, the loss function is the loss function used when pre-training the battery health life evaluation model with an additional L2 regularization term added, and the functional expression of the loss function is: , where is the weight parameter, is the L2 regularization term, is the number of layers of the battery health life evaluation model, is the layer parameter of the battery health life evaluation model of the L2 norm; when backpropagating and optimizing the parameters of the global average pooling layer of the high-layer residual blocks and the fully connected layer in the classification and regression module through the given loss function and a learning rate smaller than that when pre-training the battery health life evaluation model, it includes using two iteration-round thresholds epoch1 and epoch2 to control the end of training, where the iteration-round threshold epoch1 is smaller than the iteration-round threshold epoch2. If the loss value of the loss function has not decreased or the decrease amount is less than the preset threshold for more than the threshold epoch1 consecutive iteration rounds, the learning rate is further decreased. The further decrease of the learning rate means subtracting a preset adjustment amount from the currently used learning rate or multiplying by a parameter less than 1 to make the value of the learning rate smaller; if the loss value of the loss function has not decreased or the decrease amount is less than the preset threshold for more than the threshold epoch2 consecutive iteration rounds, it is determined that the training is over, and the battery health life evaluation model obtained at this time is the optimized battery health life evaluation model.

[0013] In addition, the present invention also provides a lithium-ion battery health life evaluation system based on feature images, including a microprocessor and a memory connected to each other, and the microprocessor is programmed or configured to execute the lithium-ion battery health life evaluation method based on feature images.

[0014] In addition, the present invention also provides a computer-readable storage medium, in which a computer program or instruction is stored, and the computer program or instruction is programmed or configured to execute the lithium-ion battery health life evaluation method based on feature images through a processor.

[0015] In addition, the present invention also provides a computer program product, including a computer program or instruction, and the computer program or instruction is programmed or configured to execute the lithium-ion battery health life evaluation method based on feature images through a processor.

[0016] Compared with the prior art, the present invention can mainly achieve the following beneficial effects: The present invention includes using the sample data in the source domain to pre-train a battery health life evaluation model to establish a mapping relationship between the feature images and the health life of lithium-ion batteries, and the feature images are generated based on incremental capacity features; introducing transfer learning to use the sample data in the target domain to optimize the battery health life evaluation model; extracting incremental capacity features for the lithium-ion battery to be evaluated in the target domain and generating the feature images of the lithium-ion battery, and inputting them into the optimized battery health life evaluation model to obtain the corresponding health life, so that the lithium-ion battery health life evaluation model trained based on the source domain data can better adapt to the battery characteristics and working conditions in the target domain to solve the differences in battery aging and the influence of various external environmental factors, thereby improving the accuracy of lithium-ion battery health life evaluation in the target domain. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] Figure 1 It is a schematic diagram of the basic process of the method in the embodiment of the present invention.

[0018] Figure 2 It is the IC curve of Cell-1 in the present invention embodiment within the voltage range of 3.7V to 4V.

[0019] Figure 3 It is the two-dimensional image of the IC curve at the 2000th cycle in the present invention embodiment.

[0020] Figure 4 It is the two-dimensional image of the IC curve at the 4000th cycle in the present invention embodiment.

[0021] Figure 5 It is the two-dimensional image of the IC curve at the 7000th cycle in the present invention embodiment.

[0022] Figure 6The flowchart of Bayesian optimization in the embodiments of the present invention. Detailed implementation manners

[0023] In order to enable those skilled in the art of the present technology to better understand the technical solutions of the present invention, the technical solutions of the present invention will be further described in detail below with reference to the accompanying drawings in the embodiments of the present invention.

[0024] As Figure 1 shown, the method for evaluating the health life of a lithium-ion battery based on a feature image in this embodiment includes the following steps: S1. Use the sample data of lithium-ion batteries with similar electrochemical characteristics but different operating conditions in the source domain to pre-train the battery health life evaluation model, so that the battery health life evaluation model establishes a mapping relationship between the feature image of the lithium-ion battery and its corresponding health life (SOH). The feature image of the lithium-ion battery is generated based on the incremental capacity feature of the lithium-ion battery; S2. Introduce transfer learning to optimize the battery health life evaluation model using the sample data of the lithium-ion battery in the target domain; S3. Extract the incremental capacity feature for the lithium-ion battery to be evaluated in the target domain and generate the feature image of the lithium-ion battery, and input the feature image of the lithium-ion battery into the optimized battery health life evaluation model to obtain the corresponding health life.

[0025] As Figure 1 shown, the generation of the feature image of the lithium-ion battery mainly includes three stages of data processing: extracting battery data, feature extraction, and two-dimensional conversion. Specifically, the generation of the feature image of the lithium-ion battery in this embodiment includes: S101. Collect the capacity and voltage of the lithium-ion battery at different cycle times through multiple cycle periods respectively; S102. Determine the capacity change amount Δ V corresponding to different preset fixed voltage intervals Δ Q , calculate the derivative of the charge and discharge capacity with respect to the battery voltage dQ / dV , and convert the change of the capacity into obvious and easy-to-identify peak features. These peaks indicate the phase balance between the anode and cathode of the battery and indirectly reflect its internal chemical reaction mechanism. Therefore, the incremental capacity of each sample point corresponding to each cycle time can be calculated according to the following formula: . . In the above formula, IC is the incremental capacity, dQ / dV is the derivative of the capacity with respect to the voltage, Δ Q is the capacity change amount of the lithium-ion battery; Δ V is the preset fixed voltage interval; S103. Calculate the average incremental capacity of all cycle periods at a certain cycle time according to the following formula:

[0026] where, is the average incremental capacity of all cycle periods at a certain cycle time, is the number of cycle periods, is the incremental capacity of the i-th cycle period of the lithium-ion battery. Finally, an incremental capacity curve (i.e., IC curve) represented by a time series of the average incremental capacity of each cycle time of the lithium-ion battery is obtained. Taking the aging data of the Oxford University Cell-1 battery as an example, the IC curve of Cell-1 extracted in this embodiment in the voltage range of 3.7V to 4V is as shown in Figure 2 . The Pearson correlation coefficient (PCC) can be used to objectively analyze the degree of linear relationship between the average incremental capacity and the battery health life. After determining each average incremental capacity, calculate the Pearson correlation coefficient : : , where, n is the number of samples, x i and y i are the i-th observed values of the average incremental capacity X and the battery SOH value Y respectively, i is the i-th observation value, and are the sample means of the variables and respectively; S104. Use the Kalman filtering method to smooth the incremental capacity curve represented by the time series of the average incremental capacity of each cycle time of the lithium-ion battery; S105. Use the relative position matrix (RPM) to convert the incremental capacity curve represented by the time series of the average incremental capacity of each cycle time of the smoothed lithium-ion battery into a two-dimensional image. Taking the data of the 2000th cycle, the 4000th cycle, and the 7000th cycle of the battery as an example, convert the one-dimensional incremental capacity data into a two-dimensional image through the relative position matrix, as shown in Figure 3 , Figure 4 and Figure 5 respectively.

[0027] In step S104 of this embodiment, the Kalman filtering method is used to smooth the incremental capacity curve represented by the time series of the average incremental capacity at each cycle moment of the lithium-ion battery. When smoothing the original IC curve, the data points of the IC curve are regarded as measurement values. By continuously iterating the above process and using the prediction and update mechanisms of the Kalman filter, the state estimation value is gradually adjusted according to the dynamic change law of the data and the characteristics of the measurement noise, so as to smooth the IC curve, reduce the influence of noise and abnormal fluctuations on the basis of retaining the original information, and make the processed curve more conducive to subsequent feature extraction and analysis work. The prediction mechanism of the Kalman filter is as follows: , , wherein, is the estimated incremental capacity at the previous moment k moment, is the incremental capacity at the previous moment k-1 ; is the state transition matrix, which can be expressed as the change rate of the incremental capacity; is the control input matrix, is k the voltage at the moment, is the covariance matrix estimated at the previous moment k moment, is the covariance matrix at the previous moment k-1 ; is the process noise covariance matrix; The update mechanism of the Kalman filter is as follows: , , , wherein, K k is the Kalman gain, which is used to balance the weights of the predicted value and the measured value. P k is the updated covariance matrix, which reflects the new uncertainty of the state estimation; is the observation matrix, is the observation noise covariance matrix, is k the actual incremental capacity at the moment; is the identity matrix.

[0028] In this embodiment, converting the smoothed battery incremental capacity curve into a two-dimensional image by using the relative position matrix in step S105 includes: S201. Normalize the average incremental capacity at each cycle time of the lithium-ion battery after smoothing processing by zscore. It can be expressed as: , , where, is the time series, ~ are the average incremental capacities at the 1st to nth cycle times, is the average incremental capacity at the i th after zscore normalization, is the average incremental capacity at the i th cycle time, and are the mean and standard deviation of the average incremental capacities in the time series respectively; S202. Dimensionally reduce the average incremental capacity of the standard normal distribution obtained after zscore normalization according to the following formula: , , where, is the i th data value obtained after dimensional reduction, is the dimensional reduction factor, is the average incremental capacity at the j th after zscore normalization, is the number of cycle times in a cycle period, is the dimension of the data values after dimensional reduction; S203. Construct a relative position matrix of size as shown in the following formula based on the data values obtained after dimensional reduction: , where, ~ are the 1st to th data values obtained after dimensional reduction respectively; Each pair of timestamps in the time series is connected by the relative position matrix to obtain their relative positions. Each row and each column of the relative position matrix is referenced by a certain timestamp and contains the information of the entire time series; S204. Apply min-max normalization to convert the relative position matrix into a grayscale value matrix as the converted two-dimensional image: , wherein, is the grayscale value matrix, min represents taking the minimum value, and max represents taking the maximum value.

[0029] As Figure 1 shown, in this embodiment, the battery health life evaluation model is a ResNet neural network model (ResNet network). For the dataset obtained by processing the source domain data, it can be classified into a training set and a test set through classification. The ResNet neural network model is pre-trained using the training set and the test set, and then through transfer learning, the final ResNet neural network model applicable to the target domain can be obtained. The health life (SOH) of a lithium-ion battery is the result of comparing the current available capacity of the battery with its initial factory rated capacity, and can be expressed as:

[0030] wherein, is the health life (SOH) of the lithium-ion battery, is the current available capacity of the battery, is the initial factory rated capacity. Through the above, the health life can be calculated as the label used during the training of the ResNet neural network model, enabling the battery health life evaluation model to establish a mapping relationship between the characteristic image of the lithium-ion battery and its corresponding health life (SOH), thereby using the incremental capacity curve to determine the health life (SOH) of the battery.

[0031] In this embodiment, the processing of the characteristic image of the lithium-ion battery by the ResNet neural network model includes: extracting features from the characteristic image of the lithium-ion battery through a residual unit, and the residual unit includes a plurality of cascaded residual blocks (Residual Block); and after the output features of the residual unit are globally average pooled through a global average pooling layer, they are then classified or regressed through a classification and regression module composed of multiple fully connected layers and activation function layers (sigmoid function) to obtain the health life of the lithium-ion battery. The residual unit of the ResNet neural network introduces a shortcut connection structure (ShortcutConnection), which can effectively promote the information propagation in the network, thereby suppressing the network degradation phenomenon. The functional expression of the residual block in this embodiment is: or , wherein, is the output feature of the residual block, is the input feature of the residual block, represents the residual mapping of the input feature of the residual block based on the weight parameter , is a linear mapping, which is adopted when the residual mapping of the residual block is the same as the input feature dimension ; when the residual mapping of the residual block is different from the input feature dimension, a linear mapping can be added through the shortcut connection w s to match the dimensions; the constant mapping path from the input to the output of the residual block is called the short connection. The feature information contained in the input is retained through the short connection, thereby reducing the attenuation of effective information during the forward propagation process. In this embodiment, the residual mapping is composed of a cascade of multiple convolutional modules, and each convolutional module is composed of a series-connected convolutional layer, a batch normalization layer, and an activation function. The network also introduces a global pooling layer to compress the spatial dimension of the feature map into a fixed length, and a fully connected layer is used to reduce and increase the dimension of the feature vector. The convolutional layer is used to extract local features of the image, including multiple convolutional kernels, and each convolutional kernel can capture different feature patterns. The pooling layer is used to reduce the spatial dimension of the feature map, reducing the amount of computation and the risk of overfitting. The fully connected layer is responsible for comprehensively processing and classifying the extracted features. In ResNet, the initial convolutional layer and pooling layer are used to perform preliminary feature extraction and dimensionality reduction on the input image, and then multiple residual blocks are used to gradually extract high-level features. Finally, the global average pooling layer and the fully connected layer output the classification or regression results. In this embodiment, according to the objective of battery SOH evaluation, the number of neurons in the output layer is set to 1 to output the SOH estimation value of the battery. When inputting training data for training, the two-dimensional image features obtained from step S2 and their corresponding true battery SOH values are used as training samples. The training data is randomly shuffled and grouped, and divided into a training set, a validation set, and a test set according to a certain ratio. For example, 70% is used for training, 15% is used for validation, and 15% is used for testing.

[0032] In this embodiment, when pre-training the battery health life evaluation model, it also includes using Bayesian optimization technology to determine the optimal model parameters of the battery health life evaluation model and fine-tuning the model. Bayesian optimization technology is used to determine the optimal hyperparameters when adjusting the model. After placing the optimal parameters, the model is fine-tuned to ensure that the performance in the estimation results reaches the highest level. Finally, the trained neural network model is used to determine the battery SOH, and a comprehensive quantitative test is performed on the estimation value obtained by the model. Bayesian optimization technology is used to determine the optimal hyperparameters when adjusting the model. After placing the optimal parameters, the model is fine-tuned to ensure that the performance in the estimation results reaches the highest level. The trained neural network model is used to determine the battery SOH, and a comprehensive quantitative test is performed on the estimation value obtained by the model. As Figure 6 shown, the optimization process of Bayesian optimization includes: (1) Initialization n hyperparameter points X n, evaluate its corresponding objective function value y n ; It should be noted that the objective function of Bayesian optimization is a well-known function in the art, so its implementation will not be elaborated here; (2) Construct a Gaussian process model: , where is the posterior distribution of the Gaussian process; f is the objective function, assuming it follows a Gaussian process distribution. is the posterior mean function, representing the expected value of the objective function at point X n and y n under the condition of given data x . is the posterior mean function, is the posterior variance function, representing the uncertainty of the objective function at point X n and y n under the condition of given data x . And there is: , , where , and are kernel functions, used to measure the similarity between two points. is the kernel matrix, representing the similarity between hyperparameter points . (3) Select a new hyperparameter point , including: calculating the acquisition function and finding the hyperparameter point x n+1 that maximizes the acquisition function: Expected Improvement is one of the most commonly used acquisition functions in Bayesian optimization. The functional expression of Expected Improvement is as follows: , where is the prediction performance (posterior mean) of the model under hyperparameter x , is the uncertainty (posterior standard deviation) of the model under hyperparameter x . is the currently known optimal model performance. is an adjustable parameter for controlling exploration and exploitation. and are the cumulative distribution function CDF and probability density function PDF of the standard normal distribution, used to calculate the expected improvement.x is an arbitrary point in the hyperparameter space. x n+1 is the next most promising hyperparameter point selected by the acquisition function. x n+1 is the point selected from the hyperparameter space by maximizing the acquisition function and added to X n to update to the new hyperparameter point and the update function expression is: X n+1 = X n { x n+1}, y n+1 = y n { y n+1}, , where x n+1 is the next hyperparameter point selected by the acquisition function, f ( x n+1 ) is the value of the objective function at the point x n+1 ;

[0033] The updated Gaussian process posterior distribution is: , The updated posterior mean and variance can be calculated by the following formulas: , ; (5) Repeat steps (2)-(4) until the stopping condition is met.

[0034] It should be noted that the Bayesian optimization technique is an existing technology. In this embodiment, only the application of the Bayesian optimization technique is involved, and no improvement to the Bayesian optimization technique itself is made. Therefore, its detailed details will not be elaborated here. Through the above steps, Bayesian optimization can efficiently search for the optimal solution in the hyperparameter space. By using the existing evaluation results, it can more specifically explore the potential hyperparameter regions, so as to find the optimal hyperparameter combination faster and improve the model performance. The model performance can select the mean absolute error (MAE), root mean square error (RMSE), mean absolute percentage error (MAPE), mean square error (MSE), and absolute error (AE) as evaluation indicators according to needs to correctly quantify the effectiveness of the network model in estimating the SOH. The calculation function expressions of the above evaluation indicators are: , , , , , wherein, N is the number of samples; y i and respectively represent the actual value of the i th SOH and the estimated value of the i th SOH.

[0035] In this embodiment, when introducing transfer learning to optimize the battery health life evaluation model by using the sample data of lithium-ion batteries in the target domain, it includes, for the pre-trained battery health life evaluation model, dividing the multiple cascaded residual blocks in the residual unit into the bottom-layer residual blocks and the high-layer residual blocks, freezing the parameters of the bottom-layer residual blocks so that they do not participate in backpropagation (because the features learned in these stages are still relatively general), while the features learned in the high-layer residual blocks in these stages are more specific and need to adapt to the target domain data, and a random dropout layer is added after the fully connected layer to randomly discard some neurons to enhance the generalization ability of the model, and using the sample data of lithium-ion batteries in the target domain, through the given loss function And use a learning rate smaller than that during the pre-training of the battery health life evaluation model to backpropagate and optimize the parameters of the fully connected layers in the high-level residual block, global average pooling layer, and classification and regression module, so that the model can better adapt to the battery characteristics and operating conditions of the target domain and improve the SOH evaluation accuracy on the target domain. During the fine-tuning process, closely monitor the performance metrics of the model on the target domain validation set, and adjust the fine-tuning strategy and parameters according to the changes in the metrics to ensure that the model achieves the best performance on the target domain. The monitored performance metrics include MAE, RMSE, MAPE, and training / validation loss, which are used to evaluate the prediction accuracy and stability of the model.

[0036] In this embodiment, the loss function is the loss function used during the pre-training of the battery health life evaluation model (which can be selected as needed, such as using cross-entropy loss, etc.) with an additional L2 regularization term added. The function expression of the loss function is: , where are weight parameters, is the L2 regularization term, is the number of layers of the battery health life evaluation model, is the layer parameter of the battery health life evaluation model L2 norm; when backpropagating and optimizing the parameters of the fully connected layers in the high-level residual block, global average pooling layer, and classification and regression module through the given loss function and a learning rate smaller than that during the pre-training of the battery health life evaluation model, it includes using two iteration round thresholds epoch1 and epoch2 to control the end of training. Among them, the round threshold epoch1 is smaller than the round threshold epoch2. If the loss value of the loss function does not decrease or the decrease amount is less than the preset threshold for more than the threshold epoch1 consecutive iteration rounds, then further reduce the learning rate. The further reduction of the learning rate means subtracting a preset adjustment amount from the currently used learning rate or multiplying by a parameter less than 1 to make the value of the learning rate smaller; if the loss function ​​​​​​​If the loss value does not decrease or the decrease amount is less than the preset threshold, it is determined that the training is over. At this time, the obtained battery health life evaluation model is the optimized battery health life evaluation model. By setting the learning rate to one-tenth of that during pre-training (for example, if the pre-training learning rate is 0.01, the fine-tuning learning rate is 0.001). Within the range of the epoch threshold epoch1, if the validation loss does not decrease in multiple epochs, the learning rate can be gradually reduced, each time reduced to half of the current learning rate, and the learning rate scheduler is used to dynamically adjust the learning rate; then within the range of the epoch threshold epoch2, if the validation loss does not decrease in multiple epochs, the training is stopped in advance to prevent overfitting. In this embodiment, transfer learning is introduced to improve the performance and reliability of the model under different batteries and working conditions, and provide more effective technical support for the health management of lithium-ion batteries: Select a battery data set with similar electrochemical characteristics to the target domain battery but under different working conditions as the source domain data. Perform the same data collection, preprocessing, and feature transformation steps on the source domain data as above to obtain the two-dimensional image features of the source domain and the corresponding SOH labels. Use the source domain data to pre-train the constructed ResNet network model. During the pre-training process, use the same network structure and hyperparameter settings as in the model training stage, and let the model learn the general features and patterns of battery health state evaluation through a large amount of source domain data, so as to initialize the parameters of the model. Apply the pre-trained model to the target domain data. According to the characteristics of the target domain data, fine-tune the model. Specifically, use a smaller learning rate, adjust the learning rate to one-tenth of that during pre-training, and at the same time freeze the parameters of some underlying convolutional layers of the model, and only update the parameters of the high-level fully connected layer and ResNet layer of the model. Through backpropagation and parameter update on the target domain data, the model can better adapt to the battery characteristics and working conditions of the target domain, and improve the SOH evaluation accuracy on the target domain.

[0037] In summary, the lithium-ion battery health life assessment method based on feature images in this embodiment first obtains the data information of the lithium-ion battery pack, extracts the battery incremental capacity (IC) features, and uses the relative position matrix (RPM) to convert the one-dimensional IC data into a two-dimensional image. A ResNet neural network model is constructed based on the two-dimensional image features, and the training data is input for training. After the training is completed, the test data is input into the model to carry out SOH assessment. During this process, the Bayesian optimization technique is used to determine the optimal hyperparameters and fine-tune the model. Finally, transfer learning is introduced. A battery data set with similar electrochemical characteristics but different working conditions is selected as the source domain data for pre-training the model, and then applied to the target domain data and fine-tuned according to its characteristics, so as to improve the performance and reliability of the model under different batteries and working conditions, and provide effective technical support for the health management of lithium-ion batteries. The lithium-ion battery health life assessment method based on feature images in this embodiment effectively solves the problems existing in the existing lithium-ion battery SOH estimation methods, such as complex processes, difficulty for the model to capture dynamic evolution, and limitations of traditional one-dimensional feature analysis.

[0038] In addition, this embodiment also provides a lithium-ion battery health life assessment system based on feature images, including a microprocessor and a memory connected to each other, and the microprocessor is programmed or configured to execute the lithium-ion battery health life assessment method based on feature images.

[0039] In addition, this embodiment also provides a computer-readable storage medium, in which a computer program or instruction is stored, and the computer program or instruction is programmed or configured to execute the lithium-ion battery health life assessment method based on feature images through a processor.

[0040] In addition, this embodiment also provides a computer program product, including a computer program or instruction, and the computer program or instruction is programmed or configured to execute the lithium-ion battery health life assessment method based on feature images through a processor.

[0041] Those skilled in the art should understand that the technical solutions provided by the embodiments of the present invention can be in the form of a method, a system, or a computer program product. Therefore, the present invention can be implemented in the form of a complete hardware embodiment, a complete software embodiment, or an embodiment combining software and hardware aspects. Moreover, the present invention can be implemented in the form of a computer program product implemented on one or more computer-readable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) that contain computer-usable program code. The present invention is described with reference to the flowcharts and / or block diagrams of methods, devices (systems), and computer program products according to the embodiments of the present invention. It should be understood that each process and / or block in the flowchart and / or block diagram, as well as the combination of processes and / or blocks in the flowchart and / or block diagram, can be realized by computer program instructions. These computer program instructions can be provided to the processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing devices to generate a machine, so that the instructions executed by the processor of the computer or other programmable data processing devices generate means for realizing the functions specified in the process Figure 1 one process or multiple processes and / or blocks Figure 1 one block or multiple blocks. These computer program instructions can also be stored in a computer-readable memory that can direct a computer or other programmable data processing device to work in a specific manner, so that the instructions stored in the computer-readable memory generate a manufactured article including instruction means, and the instruction means realizes the functions specified in the process Figure 1 one process or multiple processes and / or blocks Figure 1 one block or multiple blocks. These computer program instructions can also be loaded onto a computer or other programmable data processing device, so that a series of operation steps are executed on the computer or other programmable device to generate a computer-implemented process. Thus, the instructions executed on the computer or other programmable device provide steps for realizing the functions specified in the process Figure 1 one process or multiple processes and / or blocks Figure 1 one block or multiple blocks.

[0042] The above is only the preferred embodiment of the present invention, and the protection scope of the present invention is not limited to the above embodiments. All technical solutions falling within the idea of the present invention belong to the protection scope of the present invention. It should be noted that for those of ordinary skill in the art, without departing from the principle of the present invention, several improvements and refinements should also be regarded as the protection scope of the present invention.

Claims

1. A lithium-ion battery health life assessment method based on feature images, characterized in that: The steps include: Pre-training a battery health life assessment model using sample data of lithium-ion batteries with similar electrochemical characteristics but different working conditions in the source domain, so that the battery health life assessment model establishes a mapping relationship between a characteristic image of the lithium-ion battery and its corresponding health life, wherein the characteristic image of the lithium-ion battery is generated based on the incremental capacity feature of the lithium-ion battery; Introducing transfer learning to optimize the battery health life assessment model using lithium-ion battery sample data in the target domain; For the lithium-ion battery to be evaluated in the target domain, the incremental capacity features are extracted and the feature image of the lithium-ion battery is generated. The feature image of the lithium-ion battery is input into the optimized battery health life assessment model to obtain the corresponding health life.

2. The method for evaluating the healthy life of a lithium-ion battery based on a characteristic image according to claim 1, characterized in that: The generation of the characteristic image of the lithium-ion battery includes: S101, collecting the capacity and voltage of the lithium-ion battery at different cycle times through multiple cycle periods; S102, determining different preset fixed voltage intervals Δ according to the capacity and voltage at different cycle times V The corresponding capacity change Δ Q , and calculate the incremental capacity of the sample point corresponding to each cycle time according to the following formula: In the above formula, IC is the incremental capacity, dQ / dV is the derivative of capacity with respect to voltage, Δ Q is the capacity change of lithium-ion battery; Δ V A preset fixed voltage range; S103, calculating the average value of the incremental capacity of all cycle periods at a certain cycle time according to the following formula: Among them, is the average value of the incremental capacity of all cycle periods at a certain cycle time, is the number of cycles, is the incremental capacity of the lithium-ion battery in the ith cycle, and finally the incremental capacity curve represented by the time series composed of the average value of the incremental capacity at each cycle moment of the lithium-ion battery is obtained; S104, using a Kalman filter method to smooth an incremental capacity curve represented by a time series consisting of an average value of incremental capacity at each cycle moment of the lithium-ion battery; S105, using a relative position matrix, converting an incremental capacity curve represented by a time series consisting of average incremental capacities of the lithium-ion battery at each cycle time after smoothing into a two-dimensional image.

3. The method for evaluating the healthy life of a lithium-ion battery based on a characteristic image according to claim 2, characterized in that: The step S105 of converting the smoothed battery incremental capacity curve into a two-dimensional image using the relative position matrix includes: S201, performing zscore normalization on the average value of the incremental capacity in the time series formed by the average value of the incremental capacity at each cycle moment of the lithium-ion battery after smoothing; S202, the average value of the incremental capacity of the standard normal distribution obtained after normalizing zscore is reduced according to the following formula: ， ， in, is the i-th data value obtained after dimensionality reduction, is the dimensionality reduction factor, is the average value of the j-th incremental capacity after normalization of zscore, is the number of cycle moments in a cycle period,. . is the dimension of the data value after dimensionality reduction; S203, construct the following formula based on the data value obtained after dimensionality reduction: The relative position matrix : ， in, ~ They are the first to data values; S204, applying min-max normalization to the relative position matrix Convert to a grayscale value matrix as the converted 2D image: ， in, It is a gray value matrix, min means taking the minimum value, and max means taking the maximum value.

4. The method for evaluating the healthy life of a lithium-ion battery based on a characteristic image according to claim 1, characterized in that: The battery health life assessment model is a ResNet neural network model. The ResNet neural network model processes the input feature image of the lithium-ion battery by: extracting features from the feature image of the lithium-ion battery through a residual unit, wherein the residual unit includes a plurality of cascade-connected residual blocks; and the output features of the residual unit are globally averaged pooled through a global average pooling layer, and then classified or regressed through a multi-level classification and regression module consisting of a fully connected layer and an activation function layer to obtain the health life of the lithium-ion battery.

5. The method for evaluating the healthy life of a lithium-ion battery based on a characteristic image according to claim 4, characterized in that: The functional expression of the residual block is: or , in, is the output feature of the residual block, is the input feature of the residual block, Represents weight-based parameters Input features to the residual block The residual mapping of For linear mapping, when the residual mapping of the residual block is the same as the input feature dimension, ; Used when the residual map of the residual block is different from the input feature dimension The residual map is composed of a plurality of convolution modules cascaded together, and each level of convolution modules is composed of a series of convolution layers, batch normalization layers and activation functions.

6. The method for evaluating the healthy life of a lithium-ion battery based on a characteristic image according to claim 1, characterized in that: The pre-training of the battery health life assessment model also includes using Bayesian optimization technology to determine the optimal model parameters of the battery health life assessment model and fine-tune the model.

7. The method for evaluating the healthy life of a lithium-ion battery based on a characteristic image according to claim 4, characterized in that: The introduction of transfer learning to optimize the battery health life assessment model using lithium-ion battery sample data in the target domain includes dividing the multiple cascade-connected residual blocks in the residual unit into bottom-level residual blocks and high-level residual blocks for the pre-trained battery health life assessment model, freezing the parameters of the bottom-level residual blocks so that they do not participate in back propagation, and adding a random dropout layer after the fully connected layer to randomly drop some neurons to enhance the generalization ability of the model, and using the lithium-ion battery sample data in the target domain, based on a given loss function and a smaller learning rate than that used in the pre-trained battery health life assessment model to back-propagate and optimize the parameters of the global average pooling layer of the high-level residual block and the fully connected layer in the classification regression module. The loss function The loss function used in the pre-trained battery health life evaluation model Based on the L2 regularization term, the loss function The function expression is: ， in, is the weight parameter, is the L2 regularization term, The number of layers for the battery health life assessment model, The battery health life assessment model Layer Parameters The L2 norm of When backpropagating and optimizing the parameters of the global average pooling layer of the high-level residual block and the fully connected layer in the classification regression module with a smaller learning rate than the pre-trained battery health life assessment model, the end of training is controlled by using two iteration rounds of epoch thresholds epoch1 and epoch2, where the epoch threshold epoch1 is smaller than the epoch threshold epoch2. If the loss function exceeds the threshold epoch1 for consecutive iteration rounds, If the loss value does not decrease or the decrease is less than the preset threshold, the learning rate is further reduced. Further reducing the learning rate means subtracting a preset adjustment amount from the current learning rate or multiplying it by a parameter less than 1 to make the learning rate smaller. If the loss function exceeds the threshold epoch2 for two consecutive iterations, If the loss value does not decrease or the decrease is less than the preset threshold, the training is determined to be completed. At this time, the battery health life assessment model obtained is the optimized battery health life assessment model.

8. A lithium-ion battery health life assessment system based on feature images, comprising a microprocessor and a memory connected to each other, characterized in that: The microprocessor is programmed or configured to execute the lithium-ion battery health life assessment method based on feature images as described in any one of claims 1 to 7.

9. A computer-readable storage medium having a computer program or instruction stored therein, characterized in that: The computer program or instruction is programmed or configured to execute the lithium-ion battery health life assessment method based on feature images as described in any one of claims 1 to 7 through a processor.

10. A computer program product comprising a computer program or instructions, characterized in that The computer program or instruction is programmed or configured to execute the lithium-ion battery health life assessment method based on feature images as described in any one of claims 1 to 7 through a processor.

Images