Lithium battery state of health variable working condition migration prediction method based on mask representation

Abstract

The application discloses a lithium battery health state variable working condition migration prediction method based on mask representation, and the method comprises the following steps: inputting current charging and discharging cycle data of a battery to be predicted into a lithium battery health state prediction model to obtain a health state prediction result; the lithium battery health state prediction model is to mask reconstruct the source domain unlabelled charging and discharging data, to approximate the extracted features and the manually extracted features containing rich information after dimension reduction, to guide the network to learn the most representative features from two aspects. Then, the following decoder is discarded, a full connection layer is added to perform the health state prediction task, and a small amount of data of the target domain is used for fine tuning, so as to solve the migration learning problem of the unlabelled source domain data in the lithium battery health state prediction task. The method innovatively introduces mask reconstruction and manual feature approximation to extract features from the unlabelled charging and discharging data, and well solves the self-supervised lithium battery health state prediction problem under variable working conditions.

Description

Technical Field

[0001] This invention belongs to the field of lithium battery health status prediction, and particularly relates to a self-supervised prediction method based on varying operating conditions. Background Technology

[0002] In recent years, lithium-ion batteries have been widely used due to their high energy density, low self-discharge rate, high voltage, long lifespan, and higher reliability. However, the safety issues they bring cannot be underestimated. As one of the main means of ensuring safety, the State of Health (SOH) management of lithium batteries has received increasing attention. SOH, as an indicator of battery aging, reflects the rate of capacity reduction between the actual and nominal capacities. A brand-new battery has an SOH value of 100%. As the battery is used, the SOH gradually decreases. Typically, when the battery capacity drops to 80% of its initial value, i.e., when the SOH is 80%, it is considered to have reached the end of life (EOL) and should be replaced. Generally, accurate SOH can only be measured directly under laboratory conditions; in practical applications, it can only be estimated using other variables such as voltage and current. However, due to the nonlinearity and uncertainty of battery aging, predicting SOH becomes extremely difficult.

[0003] Due to the development of machine learning theory and the improvement of computing power in recent years, data-driven methods have received increasing attention in academia and industry. Data-driven battery SOH prediction methods have been widely applied in the field of lithium battery SOH prediction. By establishing a data-based SOH prediction model without relying on any complex domain knowledge, it has stronger generalization ability.

[0004] Although data-driven methods have received widespread attention in SOH assessment tasks, these methods all assume that training and testing data follow the same distribution. However, in reality, battery operating states vary greatly, which can cause an SOH prediction model trained under one operating condition to fail under another. Therefore, monitoring battery SOH under different operating conditions is a challenging problem.

[0005] In recent years, transfer learning methods have been increasingly adopted to address the SOH prediction problem under the aforementioned variable operating conditions. However, most existing transfer learning methods utilize the idea of ​​domain adaptation, aligning the source and target domain data at the feature level after feature extraction. Since SOH prediction is a regression problem, this alignment method, originally designed for classification problems, may not be effective, and it's difficult to theoretically prove that the aligned features can assist subsequent downstream tasks. Therefore, how to extract features with representational information that are helpful for downstream tasks from a large amount of source domain data, and then fine-tune them using a small amount of target domain data to achieve accurate SOH prediction in the target domain, is a pressing issue with significant theoretical research value and practical implications. Summary of the Invention

[0006] The purpose of this invention is to address the difficulty of State of Health (SOH) prediction under varying operating conditions by providing a lithium battery health state transition prediction method based on mask representation. This method inputs the current charge-discharge cycle data of the battery to be predicted into a lithium battery health state prediction model to obtain the predicted health state. The lithium battery health state prediction model first reconstructs the unlabeled charge-discharge data in the source domain using a mask, and then approximates the extracted features after dimensionality reduction with manually extracted features rich in information, guiding the network to learn the most representative features from both aspects. Then, the subsequent decoder part is discarded, and a fully connected layer is added to perform the health state prediction task. Fine-tuning is performed using a small amount of data from the target domain, thereby solving the problem of unlabeled transfer learning of source domain data in the lithium battery health state prediction task. This method innovatively introduces both mask reconstruction and manual feature approximation to extract features from unlabeled charge-discharge data, effectively solving the problem of self-supervised lithium battery health state prediction under varying operating conditions.

[0007] The objective of this invention is achieved through the following technical solution:

[0008] The method for predicting the health state transition of lithium batteries based on mask representation is as follows:

[0009] The current charge-discharge cycle data of the battery to be predicted is input into the trained lithium battery health state prediction model, and the lithium battery health state prediction model outputs the current lithium battery health state prediction result of the battery to be predicted.

[0010] The lithium battery health status prediction model consists of a feature extractor with self-supervised mask representation and a fully connected layer. It is trained by taking each sample of the collected first training dataset as input and minimizing the error between the predicted lithium battery health status value and the label.

[0011] The self-supervised mask representation feature extractor is constructed by using the feature extractor as the encoder to build a mask feature approximation network with an encoder-decoder structure. Each sample of the collected second training dataset is used as the input of the mask feature approximation network. The network is trained with the goal of minimizing the error between the reconstructed data output by the decoder and the real data, and the error between the features output by the encoder and the corresponding manual features.

[0012] Each sample in the first training dataset contains charge-discharge cycle data and lithium battery health status labels, wherein the operating conditions of the charge-discharge cycle data are the same as those of the charge-discharge cycle data of the battery to be predicted.

[0013] Each sample in the second training dataset contains masked charge-discharge cycle data and corresponding manual features. The operating conditions of the charge-discharge cycle data are different from those of the charge-discharge cycle data of the battery to be predicted. The mask ratio λ∈[0,1], where 0 represents no masking and 1 represents full masking. The manual features are one or more of the following: the duration of constant current charging, the duration of constant voltage charging, the slope and area of ​​the voltage curve before the start of constant voltage charging, and features extracted based on the incremental capacity curve.

[0014] Furthermore, the feature extractor of the self-supervised mask representation consists of a parallel bidirectional gating unit module and a convolutional neural network module.

[0015] Furthermore, the decoder consists of parallel gating units, deconvolutional layers, and fully connected layers.

[0016] Furthermore, the objective of minimizing the error between the reconstructed data output by the decoder and the real data, and the error between the features output by the encoder and the corresponding manual features, is specifically expressed as follows:

[0017]

[0018]

[0019]

[0020] in, The characteristic h represents the encoder output. i With the corresponding manual feature z i The error, Represents the reconstructed data output by the decoder Compared with real data x i The error, α and β are the weights of the two parts of the loss function, W h and b h The values ​​represent the weights and biases of the linear layer, with the subscript i indicating the i-th sample and n representing the total number of samples in the second training dataset.

[0021] Furthermore, the charge / discharge cycle data includes one or more of voltage, current, and temperature data.

[0022] Furthermore, the features extracted based on the incremental capacity curve are the voltage values ​​and dQ / dV values ​​of the two peaks and one trough on the incremental capacity curve.

[0023] Furthermore, masking methods include setting some segments to zero or adding Gaussian white noise.

[0024] Furthermore, during the training process, which uses each sample of the collected first training dataset as input and aims to minimize the error between the predicted lithium battery health status of the model output and the label, the learning rate of the fully connected layer is greater than that of the feature extractor.

[0025] Compared with the prior art, the beneficial effects of the present invention include at least the following:

[0026] A self-supervised mask representation-based method for predicting the health state of lithium-ion batteries under varying operating conditions is proposed. This method involves reconstructing unlabeled charge-discharge data from the source domain using a mask, and approximating the extracted features (after dimensionality reduction) with manually extracted features rich in information. This approach guides the network to learn the most representative features. Then, the subsequent decoder is discarded, and a fully connected layer is added to perform the health state prediction task. Fine-tuning is performed using a small amount of data from the target domain, effectively solving the problem of self-supervised lithium-ion battery health state prediction under varying operating conditions. Attached Figure Description

[0027] Figure 1 This is a flowchart of the method of the present invention.

[0028] Figure 2 This is a structural diagram of a mask feature approximation network.

[0029] Figure 3 This is a diagram showing the effect of predicting the SOH of battery No. 47 after migrating from operating condition 1 to operating condition 3 using the proposed invention. Detailed Implementation

[0030] The present invention will be further described below with reference to the accompanying drawings and specific examples.

[0031] The lithium battery health state transition prediction method based on mask representation of the present invention is as follows:

[0032] The current charge-discharge cycle data of the battery to be predicted is input into the pre-trained lithium battery health state prediction model, and the lithium battery health state prediction model outputs the current lithium battery health state prediction result of the battery to be predicted.

[0033] The lithium battery health status prediction model consists of a self-supervised mask representation feature extractor and a fully connected layer, such as... Figure 1 As shown, it is obtained through training using the following method:

[0034] (1) Collect raw data, including charge-discharge cycle data of the source domain and the target domain. The target domain refers to the data domain with the same operating conditions as the charge-discharge cycle data of the battery to be predicted, and the source domain refers to the data domain with different operating conditions than the charge-discharge cycle data of the battery to be predicted. That is, the operating conditions of the charge-discharge cycle data of the target domain are the same as those of the charge-discharge cycle data of the battery to be predicted, while the operating conditions of the charge-discharge cycle data of the source domain are different from those of the charge-discharge cycle data of the battery to be predicted. Generally, the operating conditions of the charge-discharge cycle data include temperature, charge-discharge conditions (charging current, discharging current, charge-discharge mode), etc. Each charge-discharge cycle sample includes one or more of the data from three sensors: voltage, current, and temperature. In this embodiment, it is preferred to include all three. The data of the source domain does not need a health status label, while the data of the target domain needs to have a corresponding label for the current health status.

[0035] (2) Downsampling the original data, i.e., sampling the original curve at equal intervals, sampling the length of the charge-discharge curve for each cycle to L, and then using the downsampled data sets of the source and target domains respectively. and It means that, among them n and m represent the number of samples in the source and target domains, respectively, and k represents the number of sensors, which is 3 in this paper, namely current, voltage, and temperature sensors. X and Y represent the data space and label space, respectively. This is the first training dataset that was collected.

[0036] (3) Manual feature extraction: Manually extract some representative features from the source domain data. These features contain the core information of the charge-discharge curve and are strongly correlated with the downstream health status prediction task. These features include the duration of constant current charging, the duration of constant voltage charging, the slope and area of ​​the voltage curve before constant voltage charging begins, etc. Taking the extraction of features closely related to battery capacity based on the incremental capacity curve as an example, the details are as follows:

[0037] The original voltage curve for each sample is converted into an incremental capacity curve, and the curve needs to be smoothed. Noise reduction can be achieved using methods such as moving average, SG filtering, or Gaussian filtering.

[0038] Specifically, the voltage, current, and temperature curves generated for each charge-discharge cycle are represented by V, I, and T. The voltage curve is converted into an incremental capacity curve using the following formula:

[0039]

[0040] The vertical axis of the curve is dQ / dV, representing the incremental capacity value, and the horizontal axis is the voltage value. The incremental capacity curve can indicate the degree of battery aging. However, because it involves differential calculations, the curve contains a lot of noise signals, requiring filtering methods to smooth it. Here, Gaussian filtering is used for noise reduction, which can be expressed as:

[0041]

[0042] Where μ is the mean, σ is the standard deviation, and z represents the data point on the incremental capacity curve. After filtering with a Gaussian filter, features closely related to battery capacity are extracted from the smoothed incremental capacity curve. In this embodiment, six feature points are extracted from the curve, namely the voltage values ​​of two peaks and one trough, and the dQ / dV value. The manual feature of the i-th iteration, i.e., the i-th sample, is represented by z. i express.

[0043] (4) Data masking: The source domain data is masked to obscure some information. Data masking methods include setting some segments to zero or adding Gaussian white noise. The masking ratio is adjustable and denoted by λ, where λ∈[0,1], 0 represents no masking, and l represents full masking. Only the source domain data is masked; the target domain data is not masked. The final second training dataset can be represented as... in, It contains partial mask data.

[0044] (5) Establish a mask feature approximation network, such as Figure 2 As shown, the entire system consists of an encoder and a decoder. Both the encoder and decoder can adopt conventional structures. When training the mask feature network, the source domain data, i.e., the second training dataset, is used, and the input is data x. i The training aims to minimize the error between the reconstructed data output by the decoder and the real data, and the error between the features output by the encoder and the corresponding manually generated features. In this embodiment, the encoder consists of a parallel bidirectional gate recurrent unit (BiGRU) module and a convolutional neural network (CNN) module, extracting features from both spatiotemporal perspectives, and then concatenating the features. Using F... B Indicates BiGRU module, F CNN H represents the CNN module. i The characteristics representing splicing are:

[0045] h i =[F B (x i ), F CNN(x i (3)

[0046] Here, [·, ·] represents vector concatenation operation, h i After dimensionality reduction through a linear layer, and combined with the previously extracted manual features z... i To approximate this part, the loss function can be expressed as:

[0047]

[0048] Among them W h and b h Represents the weights and biases of the linear layer.

[0049] Then h i The input consists of a decoder composed of parallel gate recurrent units (GRUs), deconvolutional layers, and fully connected layers (FCs). The addition of fully connected layers is inspired by residual neural networks to prevent the network from becoming too complex and overfitting. The final outputs are summed to obtain clean reconstructed data. Use F G GRU, F D F represents a deconvolutional layer. C If we represent a fully connected layer, then:

[0050]

[0051] Then, by approximating the reconstructed data with the actual clean data, we have:

[0052]

[0053] The overall loss function is:

[0054]

[0055] Here, α and β are the weights of the two parts of the loss function. When α > β, the network focuses more on approximating the extracted features and the manually generated features, while when α < β, the network focuses more on approximating the reconstructed data and the real, clean data.

[0056] Furthermore, for GRU and BiGRU, update gates and reset gates are used to address the gradient vanishing problem during training. The calculation method for each time step t is as follows:

[0057] e t =σ(W e ·[g t-1 x i,t (8)

[0058] rt =σ(W r ·[g t-1 x i,t ]) (9)

[0059]

[0060]

[0061] Where e t and r t These are the update gate and the reset gate, respectively. σ is the sigmoid function, and W... e W r W is the weight matrix, ⊙ represents the Hadamard product (element-wise product), g t It is the hidden state at time step t. These are intermediate parameters; the information at time t-1 is stored... In the diagram, tanh represents the hyperbolic tangent activation function, x i，t Represents the i-th sample data x i Data corresponding to time step t;

[0062] BiGRU considers not only historical information but also future information. It consists of two independent GRUs, one for forward training and one for backward training. The final result is formed by concatenating the two outputs.

[0063]

[0064] in and These represent the features generated by the forward and backward directions, respectively. t For the output of BiGRU, F B (x i ) = o t .

[0065] The CNN module includes convolutional layers and pooling layers; while deconvolution is a special type of forward convolution. It first enlarges the size of the input image by padding with zeros according to a certain ratio, and then rotates the convolution kernel to perform forward convolution. Its calculation method is the same as that of forward convolution.

[0066] (6) Discard the decoder part of the trained mask feature network, retain the encoder as a feature extractor for self-supervised mask representation, and then add a fully connected layer after the feature extractor of self-supervised mask representation to form a lithium battery health state prediction model for predicting lithium battery health state. Then, use the first training dataset, i.e., the data in the target domain (x) j y jThe target domain data, used as input for training and fine-tuning the body model, is not masked. The training objective is to minimize the error between the model's output prediction of lithium battery health status and the label; the loss function is:

[0067]

[0068] Where y j and These represent the true value and the network output value, respectively. Preferably, the fine-tuning operation involves fine-tuning the fully connected layer while simultaneously fine-tuning the previous feature extractor (encoder) with a smaller learning rate than that used for fine-tuning the fully connected layer, ultimately obtaining a trained lithium battery health status prediction model.

[0069] By inputting other data from the target domain into the lithium battery health status prediction model, the predicted value of the battery health status can be obtained.

[0070] In this embodiment, NASA's open-source dataset was selected for experimental verification. The batteries in the dataset were commercially available 18650 lithium batteries. A group of batteries underwent cyclic charging, discharging, and impedance testing at different ambient temperatures. Three different operating conditions were selected to verify the effectiveness of the proposed method. For operating condition 1, all four batteries were cyclically charged, discharged, and subjected to impedance testing at room temperature (24°C), with 168, 168, 168, and 132 charge-discharge cycles respectively. During charging, the batteries were first charged at a constant current (CC) of 1.5A until the voltage reached 4.2V; then, they were charged at a constant voltage (CV) of 2A until the charging current dropped to 20mA. The discharging process included CC mode, discharging at a constant current of 2A until the voltage dropped to 2.7V, 2.5V, 2.2V, and 2.5V respectively. For Condition 2, the experiment was conducted at 43°C, and the discharge process involved a constant current of 4A until the four voltages dropped to 2.0V, 2.2V, 2.5V, and 2.7V, respectively. For Condition 3, the experimental temperature was 4°C, and the discharge current was 1A. Table 1 provides a detailed description of the selected battery data.

[0071] Table 1. Detailed Description of Batteries Under Different Operating Conditions

[0072]

[0073] Six sets of experiments were conducted for each of the three operating conditions. In each set of experiments, one operating condition was selected as the source domain and the other operating condition was selected as the target domain. One battery in the target domain was selected as the fine-tuning data and the remaining batteries were selected as the validation data. The experimental design is shown in Table 2, where C1, C2 and C3 represent operating conditions 1, 2 and 3 respectively, and C1→C2 represents the migration from operating condition 1 to operating condition 2.

[0074] Table 2. Different experimental designs

[0075]

[0076] The state of harmonics (SOH) of a battery can be calculated using the following formula:

[0077]

[0078] Q act and Q nom These represent the battery's actual capacity and nominal capacity, respectively. For a given battery, its Q... nom It is certain, therefore the prediction of SOH can be transformed into the prediction of Q. act The prediction is that a brand new battery has a SOH value of 100%. As the battery is used, the SOH will gradually decrease.

[0079] In the experiment, the evaluation criteria used were root mean square error (RMSE) and mean absolute error (MAE). RMSE measures the deviation between the observed value and the true value, while MAE better reflects the actual situation of the predicted error. Both can be expressed by the following formula:

[0080]

[0081]

[0082] Where y i Represents the true value. This represents the predicted value, and M represents the total number of predicted samples.

[0083] Table 3. Prediction accuracy of SOH using different methods

[0084]

[0085] Table 3 shows the prediction accuracy of SOH under different methods. Three transfer learning methods—Deep Adaptation Network (DAN), Domain Adaptive Neural Network (DaNN), and Deep Domain Confusion (DDC)—were selected for comparison. Additionally, a method that directly models without transfer learning (the model consists of an encoder and fully connected layers, trained using a mixture of target and source domain data) was selected as a control. The feature extractors for all methods were the same. In six experiments, the prediction accuracy of this invention was best at 2.62% RMSE, worst at 4.08%, and average at 3.54%. The best MAE was 2.04%, worst at 3.51%, and average at 2.88%. Its prediction results outperform all other algorithms (best averages of 5.11% and 3.96%), and the accuracy is doubled compared to the direct modeling method, fully demonstrating the effectiveness of this invention. Figure 3 The figure shows the effect of predicting the SOH of battery No. 47 after migrating from operating condition 1 to operating condition 3 using the proposed invention. As can be seen from the figure, the predicted value of the present invention is very close to the actual value, which further proves the effectiveness of the present invention.

[0086] Obviously, the above embodiments are merely illustrative examples for clear explanation and are not intended to limit the implementation. Those skilled in the art will recognize that other variations or modifications can be made based on the above description. It is neither necessary nor possible to exhaustively list all possible implementations here. However, obvious variations or modifications derived therefrom are still within the scope of protection of this invention.

Claims

1. A method for predicting the health state transition of lithium batteries based on mask representation, characterized in that, Specifically: The current charge-discharge cycle data of the battery to be predicted is input into the trained lithium battery health state prediction model, and the lithium battery health state prediction model outputs the current lithium battery health state prediction result of the battery to be predicted. The lithium battery health status prediction model consists of a feature extractor with self-supervised mask representation and a fully connected layer. It is trained by taking each sample of the collected first training dataset as input and minimizing the error between the lithium battery health status prediction value output by the model and the label. The self-supervised mask representation feature extractor is constructed by using the feature extractor as the encoder to build a mask feature approximation network with an encoder-decoder structure. Each sample of the collected second training dataset is used as the input of the mask feature approximation network. The network is trained with the goal of minimizing the error between the reconstructed data output by the decoder and the real data, and the error between the features output by the encoder and the corresponding manual features. Each sample in the first training dataset contains charge-discharge cycle data and lithium battery health status labels, wherein the operating conditions of the charge-discharge cycle data are the same as those of the charge-discharge cycle data of the battery to be predicted. Each sample in the second training dataset contains masked charge-discharge cycle data and corresponding manual features, wherein the operating conditions of the charge-discharge cycle data differ from those of the charge-discharge cycle data of the battery to be predicted, and the masking ratio is [not specified]. 0 represents no masking, 1 represents full masking, and manual features include one or more of the following: constant current charging duration, constant voltage charging duration, slope and area of ​​the voltage curve before constant voltage charging begins, and features extracted based on the incremental capacity curve.

2. The method according to claim 1, characterized in that, The feature extractor of the self-supervised mask representation consists of a parallel bidirectional gating unit module and a convolutional neural network module.

3. The method according to claim 1, characterized in that, The decoder consists of parallel gating units, deconvolutional layers, and fully connected layers.

4. The method according to claim 1, characterized in that, The objective of minimizing the error between the reconstructed data output by the decoder and the real data, and the error between the features output by the encoder and the corresponding manual features, is specifically expressed as follows: in, Characteristics representing the encoder output Corresponding manual features The error, Represents the reconstructed data output by the decoder With real data The error, and The weights of the two parts of the loss function, and Represents the weights and biases of a linear layer, subscript Indicates the first One sample, This represents the total number of samples in the second training dataset.

5. The method according to claim 1, characterized in that, The charge / discharge cycle data includes one or more of the following: voltage, current, and temperature data.

6. The method according to claim 1, characterized in that, The features extracted based on the incremental capacity curve are the voltage values ​​of the two peaks and one trough on the incremental capacity curve. value.

7. The method according to claim 1, characterized in that, Masking methods include setting some segments to zero or adding Gaussian white noise.

8. The method according to claim 1, characterized in that, In the training process, which uses each sample of the first training dataset as input and aims to minimize the error between the predicted lithium battery health status of the model output and the label, the learning rate of the fully connected layer is greater than that of the feature extractor.

Images