Lithium ion battery health state prediction method and equipment based on fusion degradation mechanism and deep learning, and medium

Abstract

The invention discloses a lithium ion battery health state prediction method and device based on fusion degradation mechanism and deep learning, and a medium, and relates to the field of artificial intelligence and battery health state prediction. The invention aims to solve the problem that an existing battery SOH prediction method is limited. The method comprises the following steps: taking the capacity and characteristics of a tested lithium ion battery as the input of a prediction model, and obtaining a health state prediction result of the tested lithium ion battery according to the prediction model; the characteristics of the tested lithium ion battery are related to the degradation mechanism of the tested lithium ion battery; the prediction model comprises a channel attention mechanism, a time step attention mechanism based on a sliding window and a long short-term memory network which are cascaded in sequence. According to the invention, the dual attention mechanism is cascaded before the LSTM network, so that the feature data can fully emphasize important features and time steps before being input into the LSTM. The influence of outliers and noise on key features is reduced by means of a self-adaptive robustness loss function, and the prediction precision is improved.

Description

Technical Field

[0001] This application belongs to the field of artificial intelligence and battery state of health (SOH) prediction. Background Art

[0002] As a key enabler of new energy technology development, lithium-ion battery technology plays an irreplaceable role in promoting energy transition, achieving carbon neutrality, and improving energy efficiency. However, existing technologies struggle to fully explain lithium-ion battery performance degradation. Therefore, accurately measuring battery health is crucial for maintaining the stability and safety of lithium-ion battery-related equipment.

[0003] Battery SOH is an important indicator reflecting the degree of degradation and health status of lithium-ion batteries. Existing battery SOH prediction technologies are roughly divided into data-driven, model-based, and model-data-driven methods. Model-based methods are based on physical degradation processes of known physical and chemical models, can reveal the inherent mechanisms of battery performance degradation, and are highly explanatory. However, the desire to accurately model electrochemically through model methods is limited by the complex electrochemical reactions within the battery. Data-driven methods are not affected by the complex electrochemical reactions within the battery. By fully mining the historical degradation data of the battery and fully fitting the nonlinear relationship between input and output, they can predict the battery SOH. Model-data-driven fusion methods are still limited by the shortcomings of a single model, and data quality and model performance directly affect the final results. Summary of the Invention

[0004] This application aims to solve the problem of limitations of existing battery SOH prediction methods. It now provides a lithium-ion battery health status prediction method, equipment and medium based on the fusion of degradation mechanism and deep learning. By combining physical information related to battery degradation with feature engineering technology, these physical information are converted into fusion features that can express the battery working status, thereby achieving accurate prediction of battery health status.

[0005] The first aspect of the present application provides a lithium-ion battery health status prediction method based on the integration of degradation mechanism and deep learning, including:

[0006] Using the capacity and characteristics of the lithium-ion battery under test as inputs to a prediction model, and obtaining a health status prediction result of the lithium-ion battery under test according to the prediction model;

[0007] The characteristic of the lithium-ion battery under test is related to the degradation mechanism of the lithium-ion battery under test;

[0008] The prediction model includes a channel attention mechanism, a sliding window-based time step attention mechanism and a long short-term memory network that are cascaded in sequence.

[0009] In one possible design, the characteristics of the lithium-ion battery under test include: constant current charging time, the voltage corresponding to the peak value of the IC curve during the battery charging phase, and the battery internal resistance.

[0010] In one possible design, before the capacity and characteristics of the lithium-ion battery under test are input into the prediction model, the characteristics of the lithium-ion battery under test are standardized.

[0011] In one possible design, the standardizing of the characteristics of the lithium-ion battery under test includes mapping different characteristic data to an interval of [0, 1].

[0012] In one possible design, the loss function of the prediction model is expressed as follows:

[0013]

[0014] Among them, α and c are coordination parameters, is the i-th feature under the k-th charge-discharge cycle, for The estimated value of n f is the number of features.

[0015] In one possible design, the channel attention mechanism includes:

[0016] A global average pooling operation is performed on each feature channel of the input to extract the global information of each channel; a one-dimensional convolution layer is used to perform one-dimensional convolution processing on the global information of each channel; the weight of each channel output by the one-dimensional convolution layer is scaled to the interval [0, 1] through an activation function to generate a scaled weight matrix; the input data of the channel attention mechanism and the scaled weight matrix are channel-weighted to obtain the output of the channel attention mechanism.

[0017] In one possible design, the sliding window-based time-step attention mechanism includes:

[0018] Performing a one-dimensional convolution operation on the input of the sliding window-based time-step attention mechanism;

[0019] Performing a nonlinear transformation on the result of the one-dimensional convolution operation through a ReLU activation function to obtain a transformation result;

[0020] Performing a one-dimensional convolution operation on the transformation result so that the weight matrix outputted therefrom has the same number of input feature channels as the time-step attention mechanism based on the sliding window;

[0021] The input of the sliding window-based time-step attention mechanism is weighted by time steps using the weight matrix to obtain the output of the sliding window-based time-step attention mechanism.

[0022] In a possible design, the long short-term memory network includes: an input gate, a forget gate, and an output gate;

[0023] The calculation process of the long short-term memory network is as follows:

[0024]

[0025] Among them, the intermediate variable i k 、f k and o k are the outputs of the input gate, forget gate, and output gate respectively, c k and h k Respectively represent the cell state and hidden state of the long short-term memory network input, σ(·) represents the Sigmoid activation function, W f 、W i 、W C and W o are all weight matrices, b f 、b i 、b C 、b o are all deviations, and k represents the serial number of the charge and discharge cycle.

[0026] A second aspect of the present application provides a lithium-ion battery health status prediction device based on the fusion of degradation mechanism and deep learning. The lithium-ion battery health status prediction device based on the fusion of degradation mechanism and deep learning includes a processor and a memory. The memory stores at least one instruction, and the at least one instruction is loaded and executed by the processor to implement the above-mentioned lithium-ion battery health status prediction method based on the fusion of degradation mechanism and deep learning.

[0027] A third aspect of the present application provides a computer storage medium, wherein the computer storage medium stores at least one instruction, and the at least one instruction is loaded and executed by a processor to implement the above-mentioned lithium-ion battery health status prediction method based on the fusion of degradation mechanism and deep learning.

[0028] Beneficial effects of this application:

[0029] When predicting the SOH of a battery, this application starts from the internal degradation mechanism of the battery and selects physical features that are highly correlated with the degradation phenomenon that occurs during the cycle of battery use. In terms of neural network model construction, a dual attention mechanism that integrates channel attention ECA and time step attention TSA based on a sliding time window is integrated into the long short-term memory network LSTM, a neural network commonly used in time series prediction. This dual attention mechanism is cascaded before the LSTM network, so that the feature data can fully emphasize important features and time steps before entering the LSTM. In addition, the adaptive robustness loss function is used to reduce the impact of outliers and noise on key features, thereby improving prediction accuracy. BRIEF DESCRIPTION OF THE DRAWINGS

[0030] Figure 1 This is a flowchart of a lithium-ion battery health status prediction method based on the fusion of degradation mechanism and deep learning;

[0031] Figure 2 This is the charging current-time curve of the CS2_35 battery;

[0032] Figure 3 This is the charge and discharge IC curve of the CS2_35 battery;

[0033] Figure 4 This is the internal resistance aging curve of the CS2_35 battery;

[0034] Figure 5 The framework diagram of the constructed model;

[0035] Figure 6 The following are the SOH prediction results of four lithium batteries, where (a) represents CS2_35, (b) represents CS2_36, (c) represents CS2_37, and (d) represents CS2_38. DETAILED DESCRIPTION

[0036] The following will be combined with the drawings in the embodiments of the present application to clearly and completely describe the technical solutions in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, rather than all of the embodiments. Based on the embodiments in the present application, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of this application. It should be noted that, in the absence of conflict, the embodiments in the present application and the features in the embodiments can be combined with each other.

[0037] Deep learning neural networks play an important role in battery SOH prediction, especially in processing large-scale, high-dimensional data and complex nonlinear relationships. Thanks to the advancement of sensor technology and battery management systems, more and more high-frequency data (such as voltage, current, temperature, capacity, etc.) can be recorded. Deep learning models can effectively mine the complex patterns and potential information in these data to achieve high-precision prediction of battery SOH. The degradation process of batteries is usually a complex nonlinear system caused by the interaction of multiple physical and chemical reactions. Traditional linear or simple nonlinear models may find it difficult to capture these complex degradation characteristics, while deep learning models (such as convolutional neural networks, recurrent neural networks, etc.) have powerful nonlinear fitting capabilities and can capture the potential complex relationships in battery performance changes.

[0038] In view of this, the embodiment of the present application provides a lithium-ion battery health status prediction method based on the fusion of degradation mechanism and deep learning, in order to solve the above problems. Figure 1 , the scheme of the implementation method of this application is described in detail.

[0039] Specific embodiment 1: The lithium-ion battery health status prediction method based on the integration of degradation mechanism and deep learning described in this embodiment includes:

[0040] Using the capacity and characteristics of the lithium-ion battery under test as inputs to a prediction model, and obtaining a health status prediction result of the lithium-ion battery under test according to the prediction model;

[0041] The characteristic of the lithium-ion battery under test is related to the degradation mechanism of the lithium-ion battery under test;

[0042] The prediction model includes a channel attention mechanism, a sliding window-based time step attention mechanism and a long short-term memory network that are cascaded in sequence.

[0043] In one embodiment, the characteristics of the lithium-ion battery under test include: constant current charging time, the voltage corresponding to the peak value of the IC curve during the battery charging phase, and the internal resistance of the battery.

[0044] In one embodiment, before the capacity and characteristics of the lithium-ion battery under test are input into the prediction model, the characteristics of the lithium-ion battery under test are standardized.

[0045] In one embodiment, the standardizing the characteristics of the lithium-ion battery under test includes:

[0046] Map different feature data to the interval [0,1].

[0047] In one embodiment, the loss function of the prediction model is expressed as follows:

[0048]

[0049] Among them, α and c are coordination parameters, is the i-th feature under the k-th charge-discharge cycle, for The estimated value of n f is the number of features.

[0050] In one embodiment, the channel attention mechanism includes:

[0051] Perform global average pooling on each feature channel of the input to extract the global information of each channel;

[0052] Using a one-dimensional convolution layer to perform one-dimensional convolution processing on the global information of each channel;

[0053] Scale the weight of each channel output by the one-dimensional convolutional layer to the interval [0, 1] through an activation function to generate a scaled weight matrix;

[0054] The input data of the channel attention mechanism is channel-weighted with the scaled weight matrix to obtain the output of the channel attention mechanism.

[0055] In one embodiment, the sliding window-based time-step attention mechanism includes:

[0056] Performing a one-dimensional convolution operation on the input of the sliding window-based time-step attention mechanism;

[0057] Performing a nonlinear transformation on the result of the one-dimensional convolution operation through a ReLU activation function to obtain a transformation result;

[0058] Performing a one-dimensional convolution operation on the transformation result so that the weight matrix outputted therefrom has the same number of input feature channels as the time-step attention mechanism based on the sliding window;

[0059] The input of the sliding window-based time-step attention mechanism is weighted by time steps using the weight matrix to obtain the output of the sliding window-based time-step attention mechanism.

[0060] In one embodiment, the long short-term memory network includes: an input gate, a forget gate, and an output gate;

[0061] The calculation process of the long short-term memory network is as follows:

[0062]

[0063] Among them, the intermediate variable i k 、fk and o k are the outputs of the input gate, forget gate, and output gate respectively, c k and h k Respectively represent the cell state and hidden state of the long short-term memory network input, σ(·) represents the Sigmoid activation function, W f 、W i 、W C and W o are all weight matrices, b f 、b i 、b C 、b o are all deviations, and k represents the serial number of the charge and discharge cycle.

[0064] To further introduce the embodiments of this application, Figure 1 A method for predicting the health status of lithium-ion batteries based on the fusion of degradation mechanism and deep learning is provided, including steps (1) to (4). The numbering of each step does not necessarily limit the order in which they are executed. Each step / unit is described in detail below:

[0065] The dataset used in this implementation study was provided by the Center for Advanced Lifecycle Engineering (CALCE) at the University of Maryland. The experiment used constant current charging (CC) mode until the battery voltage reached 4.2V, followed by CV mode charging until the battery current dropped to 20mA. The battery was then discharged in CC mode until the battery voltage dropped to 2.7V. Similarly, end of life (EOL) occurs when the capacity drops to 70% of the rated capacity. The above method was implemented using four batteries provided by CALCE (CS2_35, CS2_36, CS2_37, and CS2_38).

[0066] (1) Feature selection

[0067] In the actual operation of lithium-ion batteries, the discharge conditions of the batteries vary and are not easily controlled by various external factors, while the charging conditions of the batteries are relatively stable. The discharge method mostly depends on the load conditions, but the charging method is more regular, generally a multi-stage constant current (CC) charging or constant current-constant voltage (CV) charging method. Therefore, some physical characteristics of the battery during charging are selected to estimate the battery SOH. Generally speaking, the rated capacity C0 of the battery is a fixed value, so the capacity C0 of the kth charge and discharge cycle is k It is sufficient to represent the health status of the current cycle. The SOH of the battery is defined as:

[0068]

[0069] The capacity and the extracted features are subjected to outlier removal operations to obtain the original time series data X. The time series data X consists of capacity and features, and the expression is as follows:

[0070] X=(C,F1,F2,F3,...,Fn) T (2),

[0071] Among them, C is the capacity of the battery and Fn is the nth feature.

[0072] This embodiment takes the CS2_35 battery provided by CALCE as an example, and plots its voltage-time curve, IC curve, and internal resistance curve during the charging stage. Figure 2 、 3 and 4. From the voltage-time curve, it can be seen that with the increase of the number of charge and discharge times, the constant current charging time for the battery voltage to reach 4.2V is gradually decreasing. It can be inferred that the constant current charging time is closely related to the aging of the battery, and the constant current charging time is recorded as F1.

[0073] The battery IC curve is the curve of the change of current with voltage during the battery charging and discharging process. Its calculation formula is as follows:

[0074]

[0075] Where Q is the capacity, V is the voltage, and are the battery capacities corresponding to voltages V1 and V2 respectively.

[0076] The dV / dt in the CACLE data set is given. Considering the characteristics of the data set, the formula (3) for calculating the IC curve is transformed into:

[0077]

[0078] Where i is the current.

[0079] As can be seen from formula (4), the IC curve reflects the relationship between capacity and voltage. Furthermore, the voltage corresponding to the peak point of the IC curve during the charging phase changes regularly with the number of battery cycles. Therefore, the voltage corresponding to the peak point of the charging phase during each charge-discharge cycle is selected as F2; the internal resistance of the battery is recorded as F3.

[0080] In order to verify the effectiveness of the selected features, the Spearman's Rank Correlation Coefficient (SSC) is introduced to calculate the correlation between features and capacity.

[0081] SSC is defined as the ranked version of the Pearson correlation coefficient:

[0082]

[0083] Among them, ρ s is the value of SSC, d i Represents the rank difference of the corresponding values of two variables, n is the number of samples. ρ s The closer the absolute value of is to 1, the stronger the correlation between the feature and capacity is. The calculation results are shown in Table 1.

[0084] Table 1 Correlation calculation of F1, F2 and F3 of four batteries with Cap

[0085] CS2_35 CS2_36 CS2_37 CS2_38 F1 0.982981 0.979168 0.979498 0.969837 F2 -0.938852 -0.966789 -0.949564 -0.909406 F3 -0.956208 -0.945015 -0.940269 -0.35804

[0086] The calculation of SSC verifies that F1, F2 and F3 are highly correlated with the capacity attenuation. F1, F2 and F3 and the capacity form the original characteristic sequence, where the characteristic vector X of the kth charge and discharge cycle is k Defined as:

[0087] X k =(C k ,F1 k ,F2 k ,F3 k ) T (6).

[0088] Due to the differences in the numerical ranges and units between different features, it is necessary to normalize the input features and map the different feature data to the interval [0,1] to resolve the impact of the numerical range and unit on the training and convergence of the model and improve the prediction accuracy. This implementation adopts the minimum-maximum ratio method for normalization, and the normalization formula is:

[0089]

[0090] in, is the normalized i-th feature data, is the original i-th feature data, and They are The minimum and maximum values of the corresponding features.

[0091] The time series data is divided into multiple samples according to the specified time window size. A sliding window method is used. Local time segments are extracted as input features, and the target values related to these segments are predicted. Assuming the total length of the dataset sequence X is l, and the input time step of the sliding window is w, the entire dataset sequence can generate lw sequences of length w. The input of the mth time window is [Xm ,X m+1 ,...,X m+w ], the target value is X m+w+1 .

[0092] (2) Loss function of fusion mechanism characteristics

[0093] In order to reduce the impact of outliers or noise on prediction accuracy and improve the generalization ability and robustness of the prediction model, robustness is used as the loss function. The preliminary robustness loss function is defined as follows:

[0094]

[0095] Here, c can be regarded as a coordination parameter, which is used to adjust the scale of the curve of the loss function f(x, α, c) at x = 0. This adaptive robust loss function mainly controls the robustness of the loss function through the hyperparameter α, that is, α can not only be manually set to a constant value to represent other loss functions, but can also be expressed as a suitable loss function based on the data characteristics. From the perspective of the formula, α is not defined at α = 0 (Cauchy Loss), α = 2 (L2 Loss), and α = -∞ (Welsch Loss), but can be approximated by taking the limit of the definition formula. Therefore, the final adaptive robust loss function can be expressed as the following piecewise function set form:

[0096] Taking the limit of f(x,α,c),

[0097] The expression of the first-order partial derivative of the adaptive robustness loss function with respect to x is as follows:

[0098]

[0099] Traditional loss functions, such as mean squared error (MSE), are very sensitive to outliers because they calculate the square of the error. If there are noise or outliers in the data, these outliers will have a huge impact on model training. From the definition of SOH, we can see that the health state SOH of the kth cycle is k The capacity C of the kth cycle k In order to focus on the capacity feature, the robustness loss function is used, and the MSE loss function is used for the other features.

[0100] Based on the above analysis, a highly robust loss function for the fusion mechanism is proposed:

[0101]

[0102] is the i-th feature under the k-th charge-discharge cycle, for Both α and c are coordination parameters, which are dynamically adjusted according to the distribution characteristics of the input data. f is the number of features, and in this implementation, it is set to 3. Robust loss functions typically weight the loss to reduce the contribution of outliers to the overall loss. During training, α and c are adjusted based on the output of the current loss function or the distribution of the data (e.g., the size of the residual, the variance of the error, etc.). If the model encounters a large number of outliers or noise, these hyperparameters may be increased to make the loss function more robust and reduce the impact of outliers.

[0103] (3) Neural network model based on dual attention mechanism

[0104] like Figure 5 As shown in the figure, the neural network model based on the dual attention mechanism combines efficient channel attention (ECA) and sliding window-based time step attention (TSA), respectively adjusting the importance of input data at the feature channel and time step level. These attention mechanisms enable the model to focus more on important features and time steps when processing time series data, thereby improving prediction accuracy. Finally, the long short-term memory (LSTM) network is responsible for capturing dynamic patterns in the time series and outputting the final prediction results.

[0105] (1) ECA channel attention mechanism

[0106] The ECA channel attention mechanism effectively models the dependencies between feature channels. For different features corresponding to different channels, ECA dynamically assigns channel weights, allowing the network to focus more on key features while suppressing less important ones. Compared to the SE (Squeeze-and-Excitation) channel attention mechanism, ECA implements channel-level attention allocation by replacing fully connected layers with a simple one-dimensional convolution operation, significantly reducing the number of parameters and achieving a good balance between computational efficiency and performance.

[0107] The calculation process of ECA is as follows:

[0108]

[0109] The operation process of ECA is as follows Figure 5 As shown, it includes the following steps:

[0110] 1) For each feature channel X of the input i Perform a global average pooling operation GAP(·) to extract the global information of each channel.

[0111] 2) Use a one-dimensional convolution layer Conv1D k To learn the interdependence between channels.

[0112] The kernel size k of the one-dimensional convolutional layer is adaptively determined based on the number of channels to ensure computational efficiency. Kernel size k = |log2(C) + b| × γ, where C is the number of channels of the input feature; b is the offset used to adjust the translation of the calculation results, usually set to b = 1; and γ is the scaling factor used to control the range of the convolution kernel size, here set to γ = 2.

[0113] 3) After the one-dimensional convolution, the output of the one-dimensional convolution is processed by the activation function σ(·) (sigmoid or tanh), and the weight of each channel is scaled to the interval [0,1] to reflect the importance of each channel.

[0114] 4) Finally, the original input data X can be channel-weighted with the generated weight matrix x′ to obtain the ECA output X C , in this way, we can focus on important features.

[0115] (2) TSA time step attention mechanism

[0116] The TSA time-step attention mechanism is introduced after the ECA channel attention mechanism. When predicting the target value within a sliding window, it dynamically assigns weights to different time steps, allowing the model to focus more on the time steps that are most critical to the current prediction result. This improves prediction accuracy while allowing users to intuitively understand which time steps the model focuses on during prediction, enhancing model interpretability. This time-step attention mechanism mainly consists of two one-dimensional convolution operations.

[0117] The calculation process of TSA is as follows:

[0118]

[0119] 1) Conv1D (1) Represents the first layer of one-dimensional convolution, which maps the data of different feature channels corresponding to each time step in the window through one-dimensional convolution, reducing the number of channels while also reducing the number of model parameters. The features extracted by convolution are more compact and some redundant information is removed.

[0120] 2) After the first layer of one-dimensional convolution, a nonlinear transformation is performed through the ReLU (Rectified Linear Unit) activation function to obtain the transformation result Z, which enables the model to fit more complex functions and enhance the model's expressive power.

[0121] 3) Conv1D(2) Represents the second layer of one-dimensional convolution, which is responsible for generating outputs with the same number of input feature channels, so that each channel at each time step is assigned an attention weight. This weight controls the degree of influence of the information of the node at that time step on the final output.

[0122] 4) Finally, the output X of ECA is adjusted by the generated weight matrix β C Perform weighting of the time steps to obtain the output X of TSA s , which enables the model to dynamically adjust the attention of each time step according to the information of each time step in the input data, and enhance the sensitivity to key time steps.

[0123] (3) LSTM network

[0124] The Long Short-Term Memory (LSTM) network inherits the inherent advantages of recurrent neural networks in time series prediction tasks. At the same time, its clever gating mechanism solves the problem of gradient vanishing or gradient exploding that recurrent neural networks face during back propagation. Figure 5 As shown in the figure, the LSTM gating mechanism mainly includes three gates (input gate i k 、Forget Gate k and output gate o k ) to manage the flow of information and state updates (cell state c k and hidden state h k ). These gates can selectively "pass" or "block" information to control which information is retained or discarded.

[0125] The main calculation process of LSTM is as follows:

[0126]

[0127] Where σ(·) represents the Sigmoid activation function, W f 、W i 、W C 、W o Represent the corresponding weight matrix, b f 、b i 、b C 、b o They represent the corresponding deviations respectively.

[0128] The extracted feature data is dynamically weighted through a dual attention module and then fed into an LSTM to learn the dynamic dependencies of the time series. The LSTM processes the input data step by step, ultimately extracting a hidden state that represents the temporal information. The output of the last time step in the last LSTM layer is taken and mapped to the final output space through a fully connected layer.

[0129] (IV) Leave-one-out cross-validation

[0130] Because leave-one-out cross-validation was used, CS2_35 served as the validation set, and CS2_36, CS2_37, and CS2_38 served as the training set. This process was repeated until every battery sample had been used as the validation set. Finally, the average of all validation results was taken to measure model performance. This method fully utilizes the data and avoids bias caused by improper data partitioning.

[0131] The parameters of the ECA-TSA-LSTM deep neural network optimization algorithm, which incorporates a dual attention mechanism, are set, including the convolution kernel size in TSA, the number of LSTM layers, the hidden layer dimension of each LSTM layer, and the optimizer's learning rate. The training set is then fed into the network for training. The parameters of the training model are adjusted using the validation set data, and a satisfactory parameter combination is selected to form the final training model. In this embodiment, the ECA convolution kernel size is adaptive based on the number of feature channels and does not require setting. The TSA convolution kernel size is set to 3; the number of LSTM layers is set to 2, the hidden layer dimension is set to 64, and the number of input channels is set to 4; the sliding window size is set to 64; the optimizer learning rate is set to 0.001; and the number of training runs is initially set to 500. The training set is then fed into the network for training.

[0132] The three evaluation metrics used to evaluate the SOH prediction performance are: root mean square error (RMSE), relative error (RE), and mean absolute error (MAE). The three evaluation metrics are defined as follows:

[0133]

[0134] The SOH prediction results of four lithium batteries are as follows Figure 6 The prediction errors (%) calculated for the prediction results are shown in Table 2.

[0135] Table 2 Comparison of SOH prediction errors (%) of different methods on the CACLE dataset

[0136]

[0137] It can be seen that the method proposed in this embodiment has a high accuracy in predicting battery SOH. Compared with the conventional LSTM model trained with MSE for SOH prediction using a single feature, RMSE, RE, and MAE are all improved. The main reasons for the performance improvement are as follows:

[0138] 1) The selected physical features can fully characterize the battery degradation process.

[0139] 2) The channel attention module is introduced to dynamically adjust the importance of each feature dimension by learning the relationship between different feature dimensions. This not only focuses on the relationship between the remaining features and capacity, but also focuses on the relationship between different features. This allows the network to pay more attention to key features while suppressing unimportant features.

[0140] 3) A sliding window-based time step attention module is introduced. When fitting the target value within a window, weights are dynamically assigned to different time steps, so that the model pays more attention to the time steps that are critical to the current prediction results. This improves the prediction accuracy while allowing users to intuitively understand which time steps the model focuses on during prediction, thereby enhancing the model's interpretability.

[0141] 4) A more efficient adaptive robustness loss function is used during training. By reducing the impact of outliers on the gradient update of network parameters, the system's adaptability to outliers such as noise data when performing prediction tasks is significantly improved.

[0142] Specific embodiment 2: The lithium-ion battery health status prediction device based on the fusion of degradation mechanism and deep learning described in this embodiment includes a processor and a memory, and the memory stores at least one instruction. The at least one instruction is loaded and executed by the processor to implement the lithium-ion battery health status prediction method based on the fusion of degradation mechanism and deep learning as described in specific embodiment 1.

[0143] Specific embodiment three: A computer storage medium described in this embodiment stores at least one instruction, and the at least one instruction is loaded and executed by a processor to implement the lithium-ion battery health status prediction method based on the fusion of degradation mechanism and deep learning as described in specific embodiment one.

[0144] Although the present application is described herein with reference to specific embodiments, it should be understood that these embodiments are merely illustrative of the principles and applications of the present application. It should therefore be understood that many modifications may be made to the illustrative embodiments, and that other arrangements may be devised, without departing from the spirit and scope of the present application as defined by the appended claims. It should be understood that the various dependent claims and features described herein may be combined in ways other than those described in the original claims. It should also be understood that features described in conjunction with individual embodiments may be used in other described embodiments.

Claims

1. A lithium-ion battery health status prediction method based on the fusion of degradation mechanism and deep learning, characterized by: include: Using the capacity and characteristics of the lithium-ion battery under test as inputs to a prediction model, and obtaining a health status prediction result of the lithium-ion battery under test according to the prediction model; The characteristic of the lithium-ion battery under test is related to the degradation mechanism of the lithium-ion battery under test; The prediction model includes a channel attention mechanism, a sliding window-based time step attention mechanism and a long short-term memory network that are cascaded in sequence.

2. The lithium-ion battery health status prediction method based on fusion degradation mechanism and deep learning according to claim 1 is characterized in that: The characteristics of the lithium-ion battery under test include: constant current charging time, the voltage corresponding to the peak value of the IC curve during the battery charging phase, and the internal resistance of the battery.

3. The lithium-ion battery health status prediction method based on fusion degradation mechanism and deep learning according to claim 2 is characterized in that: Before the capacity and characteristics of the lithium-ion battery under test are input into the prediction model, the characteristics of the lithium-ion battery under test are standardized.

4. The lithium-ion battery health status prediction method based on fusion degradation mechanism and deep learning according to claim 3 is characterized in that: The standardization process of the characteristics of the lithium-ion battery under test includes: Map different feature data to the interval [0,1].

5. The lithium-ion battery health status prediction method based on fusion degradation mechanism and deep learning according to claim 3 is characterized in that: The loss function expression of the prediction model is as follows: Among them, α and c are coordination parameters, is the i-th feature under the k-th charge-discharge cycle, for The estimated value of n f is the number of features.

6. The lithium-ion battery health status prediction method based on fusion degradation mechanism and deep learning according to claim 3 is characterized in that: The channel attention mechanism includes: Perform global average pooling on each feature channel of the input to extract the global information of each channel; Using a one-dimensional convolution layer to perform one-dimensional convolution processing on the global information of each channel; Scale the weight of each channel output by the one-dimensional convolutional layer to the interval [0, 1] through an activation function to generate a scaled weight matrix; The input data of the channel attention mechanism is channel-weighted with the scaled weight matrix to obtain the output of the channel attention mechanism.

7. The lithium-ion battery health status prediction method based on fusion degradation mechanism and deep learning according to claim 3 is characterized in that: The sliding window-based time-step attention mechanism includes: Performing a one-dimensional convolution operation on the input of the sliding window-based time-step attention mechanism; Performing a nonlinear transformation on the result of the one-dimensional convolution operation through a ReLU activation function to obtain a transformation result; Performing a one-dimensional convolution operation on the transformation result so that the weight matrix outputted therefrom has the same number of input feature channels as the time-step attention mechanism based on the sliding window; The input of the sliding window-based time-step attention mechanism is weighted by time steps using the weight matrix to obtain the output of the sliding window-based time-step attention mechanism.

8. The lithium-ion battery health status prediction method based on fusion degradation mechanism and deep learning according to claim 3 is characterized in that: The long short-term memory network includes: an input gate, a forget gate and an output gate; The calculation process of the long short-term memory network is as follows: Among them, the intermediate variable i k 、f k and o k are the outputs of the input gate, forget gate, and output gate respectively, c k and h k Respectively represent the cell state and hidden state of the long short-term memory network input, σ(·) represents the Sigmoid activation function, W f 、W i 、W C and W o are all weight matrices, b f 、b i 、b C 、b o are all deviations, and k represents the serial number of the charge and discharge cycle.

9. A lithium-ion battery health status prediction device based on the fusion of degradation mechanism and deep learning, characterized by: The lithium-ion battery health status prediction device based on the fusion of degradation mechanism and deep learning includes a processor and a memory, wherein the memory stores at least one instruction, and the at least one instruction is loaded and executed by the processor to implement the lithium-ion battery health status prediction method based on the fusion of degradation mechanism and deep learning as described in any one of claims 1 to 8.

10. A computer storage medium, characterized in that The computer storage medium stores at least one instruction, which is loaded and executed by the processor to implement the lithium-ion battery health status prediction method based on the fusion of degradation mechanism and deep learning as described in any one of claims 1 to 8.

Images