Lithium battery SOH estimation method for intelligent Yellow River underwater vehicle

Abstract

The invention discloses a lithium battery SOH estimation method for an intelligent Yellow River underwater vehicle. The lithium battery SOH estimation method comprises the following steps: acquiring lithium battery data sets under four different working conditions; analyzing the lithium battery data set from the perspective of battery attributes, and constructing attribute features; analyzing attribute features from the perspective of data range, and constructing data range features; analyzing data range features from the perspective of mathematical statistics, and constructing mathematical features; forming a feature library by the attribute features, the data range features and the mathematical features; selecting a highest correlation feature from the feature library by using a highest correlation search method to obtain a model input feature; the model input features are sent to a two-channel interpretable network, and an SOH estimation result is obtained; according to the method, the problem of effective feature input of the lithium battery can be solved, a rich feature library is obtained through multi-angle analysis of a battery data set, feature selection is automatically and transparently carried out through a highest correlation search method, and the prediction precision and the calculation efficiency of an SOH estimation model are improved in cooperation with a dual-channel interpretable network.

Description

Technical Field

[0001] The present invention relates to a lithium-ion battery health status estimation technology in the field of new energy, and in particular to a lithium battery SOH estimation method for a smart Yellow River underwater vehicle. Background Art

[0002] The ecological environment of the Yellow River Basin is complex and changeable. Realizing full-time and full-area dynamic ecological monitoring and governance of the Yellow River is an important task in building the "Smart Yellow River" digital platform. For real-time monitoring of the complex underwater environment of the Yellow River, autonomous underwater vehicles (AUVs) have gradually become the core equipment of the "Smart Yellow River" perception platform due to their high flexibility and wide coverage. However, long-term autonomous underwater operations of AUVs place higher demands on battery performance and safety, especially in turbulent and changeable water environments. As the core power source, the health status (SOH) of lithium batteries directly determines the monitoring efficiency and stability of the autonomous vehicle network system.

[0003] At present, the health status estimation method of lithium batteries usually relies on manual experience for feature selection, with a low degree of automation and a lack of transparency in the estimation process, making it difficult to reveal the mechanism of battery status changes, which affects the reliability of underwater monitoring tasks. At the same time, the complex flow patterns and environmental conditions in the Yellow River waters change frequently, which has a significant impact on battery performance. An automatic and transparent feature selection method is urgently needed to accurately extract and track the key features of the battery health status and improve the accuracy, generalization ability and processing efficiency of battery status estimation.

[0004] Therefore, in order to meet the construction needs of the "Smart Yellow River" project, it is urgent to develop a lithium battery health status estimation method that is suitable for the SD-AUV network and can realize automatic transparent feature selection. The present invention effectively captures the historical data dependencies and recent trends of the battery through an automatic transparent feature selection mechanism, and realizes accurate estimation of the lithium battery SOH, thereby improving the safety, reliability and long-term operating performance of the underwater autonomous vehicle network system, and providing important technical guarantees and support for the underwater stereoscopic perception and ecological monitoring of the "Smart Yellow River". Summary of the invention

[0005] In response to the needs in the prior art, the present invention provides a lithium battery SOH estimation method for smart Yellow River underwater vehicles, aiming to solve the problem of automatic transparent feature selection in the lithium battery SOH estimation task, process the recent trends and historical dependencies of battery data, select favorable features, and speed up network processing efficiency, providing new ideas for the field of lithium battery SOH estimation.

[0006] The lithium battery SOH estimation method for the smart Yellow River underwater vehicle includes the following steps:

[0007] Step 1: Establish a lithium-ion battery cycle life test platform and obtain lithium battery data sets under four different working conditions;

[0008] Step 2: Analyze and process the lithium battery data set to build a feature library, specifically:

[0009] Step 2.1: Analyze the lithium battery dataset from the perspective of battery attributes and construct attribute features;

[0010] Step 2.2: Analyze attribute characteristics from the perspective of data range and construct data range characteristics;

[0011] Step 2.3: Analyze the data range characteristics from the perspective of mathematical statistics and construct mathematical features;

[0012] Step 2.4: compose a feature library with attribute features, data range features, and mathematical features;

[0013] Step 3: Use the highest correlation search method to select the highest correlation feature from the feature library to obtain the model input feature;

[0014] Step 4: Feed the model input features into the dual-channel interpretability network to obtain the SOH estimation results.

[0015] Further: The lithium battery data set is collected from four aging tests, all of which adopt constant current constant voltage (CC-CV) and constant current (CC) discharge mode; in the constant current constant voltage mode, the battery is charged to the upper cut-off voltage using a specified constant current, and then charged using the upper cut-off voltage until the current drops below 0.05C; in the constant current discharge mode, the battery is discharged to a lower cut-off voltage at a specified constant current, and the cycle life test will stop until the capacity drops below 80% of the initial capacity; the operating condition settings of the lithium-ion battery in the aging test conditions are shown in Table 1:

[0016] Among them, L1~L16 are the names of different lithium battery packs, and the lithium battery is a commercial pouch lithium-ion battery with a capacity of 12Ah.

[0017] Further: In step 2.1, the attribute characteristics include voltage, current, capacity, energy, IC curve, capacity change, energy change, average voltage, ohmic internal resistance and polarization resistance, among which ohmic resistance , polarization resistance ; Indicates the current change, Indicates the voltage difference within 1s, Indicates voltage changes within 1s-20s;

[0018] In step 2.2, the attribute features are divided into three parts for feature extraction according to the charging mode: CC stage, CV stage and CC-CV stage, and each charging stage is taken as a battery charging data curve;

[0019] In step 2.3, each battery charging data curve is converted into 8 statistical values, namely: quantile, maximum value, minimum value, mean value, variance, kurtosis, skewness and slope, to capture the shape and position evolution of each cycle curve; quantile reflects the evolution by sampling the value of each quantile position; mean value measures the average level of each curve within a fixed range; variance is used to evaluate the unevenness of data stream distribution, and the larger the variance, the more uneven the distribution; skewness and kurtosis characterize the shape of each curve; slope represents the transverse polarization of the battery;

[0020] In step 2.4, each characteristic attribute sequence of the 10 attribute features is divided into 3 charging segments according to the charging mode, and each charging segment is converted into 8 statistical values. Through the combined statistical method, a comprehensive feature library with 240 mechanical statistical fusion features is generated for the battery under study.

[0021] Further: Step 3 is specifically the following steps:

[0022] Step 3.1: Evaluate the nonlinear correlation between the features in the feature library and the SOH, and select the features whose correlation is greater than the first set threshold;

[0023] Step 3.2: Decentralize the selected features and make the mean of the features 0;

[0024] Step 3.3: Calculate the covariance matrix of the decentralized results;

[0025] Step 3.4: Perform eigenvalue decomposition and singular value decomposition of the covariance matrix to obtain eigenvalues ​​and eigenvectors to determine the variance contribution rate and cumulative variance contribution rate;

[0026] Step 3.5: Select features whose cumulative variance contribution rate is greater than the second set threshold, and calculate the split gains of these features by the maximum gradient boosting method;

[0027] Step 3.6: Average the split gains of the features to get the importance score of each feature;

[0028] Step 3.7: Select several features with the largest importance scores as model input features.

[0029] Further: the dual-channel interpretability network includes a global channel, a local channel, an interactive attention mechanism and a fully connected layer, and step 4 includes the following steps:

[0030] Step 4.1: Perform maximum and minimum normalization on the model input features and then send them to the dual-channel interpretability network;

[0031] Step 4.2: The outputs of the maximum and minimum normalization are processed by the global channel and the local channel respectively and then sent to the interactive attention mechanism;

[0032] Step 4.3: The interactive attention mechanism uses the output of the global channel as the input of its own query operation to locate the key time sequence segments related to the overall health status in the local channel guided by the global degradation trend; and uses the output of the local channel as the input of its own key and value to map the local dynamic response to the global degradation framework; generates the attention weight matrix through the Softmax function to quantify the importance of global and local features;

[0033] Step 4.4: The output of the interactive attention mechanism passes through the fully connected layer to obtain the SOH estimation result.

[0034] Further: The formula of the interactive attention mechanism is as follows: (15)

[0035] in, , and represent linear transformations of queries, keys, and values, respectively, represents the transposition of the key, Represents the dimension of the key. Represents the Softmax activation function.

[0036] Further: the global channel includes a hidden layer, a maximum and minimum normalization layer, a B-spline activation function and a regularization layer, wherein the input feature is 4-dimensional data and the hidden layer is 9 nodes; the input feature of the global channel is subjected to maximum and minimum normalization processing, and the input feature is scaled to a fixed range of [-1,1]; the B-spline function is used as the activation function, and the model structure is optimized through sparse regularization and pruning to obtain the output of the global channel.

[0037] Further: the local channel includes a sliding window, a multi-scale convolution layer, a ReLU activation function and a global pooling layer. The sliding window divides the input features into sub-sequences of fixed length to capture local timing patterns; the multi-scale convolution layer is used to extract local patterns of different time scales in parallel, and the weighted sum of the local area is calculated; the ReLU activation function is used to set the negative values ​​of the input features to zero and retain the positive values ​​to focus on the positive area, and screen out local features that are strongly correlated with battery aging; the global pooling operation is used to compress the content extracted by the multi-scale convolution into a mean to retain the global timing pattern.

[0038] The beneficial effects of the present invention are as follows: by analyzing and processing lithium battery data sets under four different working conditions, the problem of effective feature input of lithium batteries under different types, working conditions and materials can be solved, and a rich feature library is obtained by analyzing the battery data set from multiple angles, and then automatic and transparent feature selection is achieved through the highest correlation search method, which effectively reduces redundant data input, extracts highly correlated features, solves the problems of recent trends and historical dependencies in lithium battery data, and provides new insights into the field of lithium-ion battery SOH estimation tasks. BRIEF DESCRIPTION OF THE DRAWINGS

[0039] Figure 1 is a flow chart of the present invention;

[0040] Figure 2 It is a data processing process diagram of the present invention;

[0041] Figure 3 It is the dual-channel interpretability network in the present invention. DETAILED DESCRIPTION

[0042] The present invention is described in detail below in conjunction with the accompanying drawings. The embodiments of the present invention are described in detail below, and examples of the embodiments are shown in the accompanying drawings, wherein the same or similar reference numerals throughout represent the same or similar elements or elements with the same or similar functions. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present invention, and cannot be interpreted as limiting the present invention. The directional terms such as left, middle, right, top, and bottom in the embodiments of the present invention are only relative concepts or are based on the normal use state of the product, and should not be considered as restrictive.

[0043] Lithium battery SOH estimation method for smart Yellow River underwater vehicles, combined with Figure 1 and Figure 2 As shown, the following steps are included:

[0044] Step 1: Establish a lithium-ion battery cycle life test platform and obtain lithium battery data sets under four different working conditions;

[0045] The lithium battery data set is collected from four aging tests, all of which adopt constant current constant voltage (CC-CV) and constant current (CC) discharge mode; in the constant current constant voltage mode, the battery is charged to the upper cut-off voltage using a specified constant current, and then charged using the upper cut-off voltage until the current drops below 0.05C; in the constant current discharge mode, the battery is discharged to a lower cut-off voltage at a specified constant current, and the cycle life test will stop until the capacity drops below 80% of the initial capacity; the operating condition settings of the lithium-ion battery in the aging test conditions are shown in Table 1:

[0046] Among them, L1~L16 are the names of different lithium battery packs, and the lithium battery is a commercial pouch lithium-ion battery with a capacity of 12Ah. From the perspective of degradation paths under different working conditions, the attenuation of the battery under the same working condition generally remains unchanged. The SOH of the battery under working condition 3 decreases linearly. The other working conditions are characterized by two-stage degradation, with slow linear attenuation in the first stage and rapid attenuation in the second stage, indicating an overall nonlinear behavior. There is a difference between the usage rate and the aging rate, especially the four batteries in working condition 1 decay the fastest. As the aging rate and the length of the cycle life vary greatly between working states, accurate SOH estimates under different working conditions are obtained; the SOH of the battery in this work is defined by the capacity, that is, the ratio of the aged current capacity to the initial capacity, as shown in formula (1): (1)

[0047] in Indicates the current capacity. Indicates the initial capacity.

[0048] The following table shows the basic parameters of 12Ah commercial pouch lithium-ion battery;

[0049] Step 2: Analyze and process the lithium battery data set to construct a feature library;

[0050] Feature construction is the most critical part of feature engineering. Traditional feature extraction methods are often designed and optimized for specific applications, so their generalization ability is relatively weak. When the application scenario or data set changes, the traditional method needs to continue to be optimized. Good feature construction can enhance the generalization ability of the model for unknown data, so that the model can maintain good estimation performance under different working conditions and conditions. The present invention creates a method that can be constructed once and in detail from the lithium battery data set without any changes. The feature library constructed by the feature construction method of the present invention is generated by a three-step construction strategy: attribute features, data range features, and mathematical features. Attribute features are the first step in the construction of the feature library, which determines what type of data to extract. The data range feature details the data range of the battery attribute feature. As the last step of the construction, the mathematical feature is to convert the lithium battery data set slices into features from a mathematical perspective.

[0051] Step 2.1: Analyze the lithium battery dataset from the perspective of battery attributes and construct attribute features;

[0052] Based on the battery attribute perspective, the voltage, current and time information in the battery aging experiment are monitored, and attribute characteristics are introduced from these three methods: direct measurement information, implicit electrochemical information, and physical characteristics information of the battery;

[0053] Direct measurement information is four directly monitored data: voltage (V), current (I), capacity (Q) and energy (E), all of which capture the trend of battery aging at different charging stages. As the polarization effect in thermodynamics decreases, the time in the constant voltage stage increases, causing the slope of the current curve to gradually decrease and the curve to expand to the right during aging. Similarly, the time to reach the upper cut-off voltage is also reduced, and the voltage curve in the constant current stage shows an increasing slope and shrinks to the left. The above changes are subsequently converted into characteristic values ​​combined with mathematical perspectives to reveal the thermodynamic and kinetic changes in battery life. The temperature rises and then falls in a cycle. Therefore, if the battery temperature is detected during the aging process, the battery temperature can be used as a directly measured parameter.

[0054] The implicit electrochemical information contains three parameters, namely IC, ΔQ and ΔE, which are introduced into the property characteristics as sufficient data for aging diagnosis. The voltage and capacity curve of 1C current is analyzed, where the voltage rises monotonically over a 15% voltage range (3.5v–3.75V) and can be charged up to 50%. By processing the voltage and capacity curve into a capacity increment (IC) curve, a deeper understanding of battery aging can be achieved. IC converts the voltage plateau on the voltage curve into identifiable peaks, which represent the key factors affecting battery aging. The calculation formula of the IC curve is as follows: (2)

[0055] in Indicates the voltage from Change to The capacity change when Indicates the number of sampling times, represents the sampling time, is the current, Represents differential.

[0056] The curve is the difference between the capacity of the old battery and the capacity of the new battery. By analyzing the curve, we can intuitively get the capacity dissipation near 3.5V, 3.65V and 3.78V. The voltage position of the capacity dissipation platform coincides with the voltage position of the reaction peak.

[0057] The curve reflects the capacity loss in the full voltage range between cycles, while the IC curve finds that the capacity loss mainly occurs during the electrochemical phase change process. The above phenomenon is shown in formula (3): (3)

[0058] The curve is also the total change of the full-cycle IC curve, where and New battery and old battery The electrochemical significance of the ΔE(V) curve and resemblance.

[0059] The physical performance parameters use three variables, average voltage, ohmic resistance and polarization resistance, as the final part of the aging parameters. The average voltage is determined by the Gibbs free energy exchange reaction of lithium ions and can characterize the thermodynamic decay of the battery. The intensity of the exchange reaction changes during the aging process and is also reflected in the average voltage. The average voltage formula is shown in (4): (4)

[0060] in, and Represents the end capacity and initial capacity respectively.

[0061] Ohmic resistance and polarization resistance can describe the dynamic performance of the battery. Ohmic resistance is usually considered to be the ratio of the instantaneous voltage change to the corresponding current change. This paper uses the voltage difference within 1s after applying the current ΔV1s for calculation, as shown in formula (5): (5)

[0062] in, Indicates the change in current.

[0063] Since lithium ion transfer is smaller than electron transfer, there is a polarization effect, as shown in formula (6): (6)

[0064] The ratio of the voltage change to the current within 1s-20s is taken as the polarization resistance.

[0065] It can be seen that the property characteristics include voltage (V), current (I), capacity (Q), energy (E), IC curve, capacity change (∆Q), energy change (∆E), average voltage, ohmic internal resistance and polarization resistance;

[0066] Step 2.2: Analyze attribute characteristics from the perspective of data range and construct data range characteristics;

[0067] The discharge mode depends on the load conditions, but the charging mode is relatively regular, generally a multi-level constant current or constant current constant voltage mode. Therefore, the data range of this article is divided according to charging. The evolution of measurement parameters such as current, voltage, capacity, and energy at different stages reflects the increase in polarization and thermodynamic losses during battery aging. Therefore, the battery reaches the cut-off voltage earlier, while the battery reaches the cut-off charging current later. Since battery aging occurs at any charging moment, the entire operating data can be divided into multiple segments according to the state to capture changes in different operating stages. For example, suppose that the battery is charged in multiple stages at a constant current. In this case, the battery degradation can be tracked by the data evolution of each stage, and each charging segment can be used as a data range. According to the charging mode, the attribute features are divided into three parts for feature extraction: CC stage, CV stage, and CC-CV stage, and each charging segment is used as a battery charging data curve;

[0068] Step 2.3: Analyze the data range characteristics from the perspective of mathematical statistics and construct mathematical features; convert each battery charging data curve into 8 statistical values, namely: quantile, maximum value, minimum value, mean value, variance, kurtosis, skewness and slope to capture the shape and position evolution of each cycle curve; quantile reflects the evolution by sampling the value of each quantile position; the mean value measures the average level of each curve within a fixed range; variance is used to evaluate the unevenness of the data stream distribution, and the larger the variance, the more uneven the distribution; skewness and kurtosis characterize the shape of each curve; for the battery voltage curve, due to the increase in thermodynamic loss and polarization, the larger the skewness indicates that the slope of the high end of the curve is larger and the percentage of the plateau area is smaller; the slope indicates the lateral polarization of the battery;

[0069] Step 2.4: Attribute features, data range features, and mathematical features are combined into a feature library; each of the 10 attribute features is divided into three charging segments according to the charging mode, and each charging segment is converted into eight statistical values. Through the combined statistical method, a comprehensive feature library with 240 mechanical statistical fusion features is generated for the studied battery;

[0070] Step 3: Use the highest correlation search method to select the highest correlation feature from the feature library to obtain the model input feature;

[0071] The purpose of feature selection is to fully automate the selection of features applicable to any battery. Traditional feature selection methods cannot extract the most representative features. Moreover, too many features may lead to network overfitting and increase the network computational burden. Therefore, the present invention designs a new feature selection engineering based on the highest correlation search method, selects the most representative four feature data from the lithium-ion battery charging and discharging data, and realizes the feature engineering innovation of automatic and transparent feature selection.

[0072] The specific steps include:

[0073] Step 3.1: Evaluate the nonlinear correlation between the features in the feature library and SOH (battery health status), and select features whose correlation is greater than the first set threshold; Specifically: The present invention first evaluates the nonlinear correlation between the extracted features and SOH, and the nonlinear correlation screening results are normalized so that the range is between [0,1], where 1 represents a complete linear correlation, and 0 represents no correlation between the feature and SOH. The calculation of the nonlinear correlation is as shown in formula (7): (7)

[0074] Where X represents the feature variable, Y represents the target variable (i.e., SOH), p(x, y) is the joint probability distribution, and p(x) and p(y) are the marginal probability distributions. In order to facilitate comparison and analysis, the present invention standardizes the features in each feature library so that the results are on the same scale. The present invention selects 24 features with a correlation greater than 0.9, which means that these features have a high correlation. In this way, features with low correlation will be discarded, reducing data input and retaining high-correlation features that affect lithium battery aging factors.

[0075] Step 3.2: Decentralize the selected features and make the mean of the features 0;

[0076] The high-correlation features after nonlinear correlation screening are still high-dimensional data. High-dimensional data has dimensionality disasters and redundancy. Redundant data will increase the computing burden of the network and slow down the computing speed. In order to reduce the information coupling between these features, the dimensionality reduction method is then used to reduce the dimensionality. The present invention designs a novel contribution dimensionality reduction method to identify important features in battery data. This method converts high-dimensional data into low-dimensional data, solves the problem of multicollinearity, and retains important information to better capture the essential characteristics of battery aging data. A new set of variables is generated by performing an orthogonal transformation on the high-correlation features after correlation analysis of lithium-ion batteries. Subsequently, features with high contribution rates are retained to encapsulate key information, and features with low contribution rates are discarded to achieve data dimensionality reduction and reduce redundancy. The operation process is as follows:

[0077] First, perform a decentralization operation to make the data mean 0, as shown in the formula: (8)

[0078] in The number of samples with correlation greater than 0.85 is 14. , each of which are features with correlation greater than 0.9.

[0079] Step 3.3: Calculate the covariance matrix of the decentralized result; the covariance matrix is: (9)

[0080] in, yes The transpose of the covariance matrix results in .

[0081] Step 3.4: Covariance matrix Eigenvalue decomposition and singular value decomposition to obtain eigenvalues and the eigenvector To determine the variance contribution rate and cumulative variance contribution rate; (10) (11)

[0082] in, Represents the feature vector The variance contribution of Indicates the number of health characteristics, represents the number of eigenvectors, Represents the feature vector The characteristic value of Represents the cumulative variance contribution. In this work, the contribution threshold is selected as 1 / 24, which is the average contribution of each feature under ideal conditions. Features below the threshold are defined as invalid features. The core of this process is to ensure that the amount of information contained in the principal component is 90% of the original time-effect space by setting the cumulative variance rate of steps 10% to 90%. At this time, the features can effectively capture the main features of the data, and 12 features remain after selection. This step eliminates feature redundant information and reduces the computational burden of the subsequent SOH estimation model.

[0083] Step 3.5: Select features whose cumulative variance contribution rate is greater than the second set threshold, and calculate the split gains of these features by the maximum gradient boosting method;

[0084] Although the features with high correlation can be screened out after correlation analysis and contribution dimensionality reduction, it is not possible to determine which features are the most influential for battery SOH estimation. The present invention designs a maximum gradient boosting method to select the most relevant features. The maximum gradient boosting method adds a regularization term to the loss function in the traditional gradient boosting decision tree, and because some loss functions are difficult to calculate derivatives, the maximum gradient boosting method of the present invention uses the second-order Taylor expansion of the loss function as the fitting of the loss function. The feature set can be further optimized to remove redundant information, so that the network training speed is more efficient, and the SOH estimation model performs well in preventing overfitting and improving generalization ability. The maximum gradient boosting method is an integrated learning algorithm based on gradient boosting, which measures the importance of features by the split gain of the decision tree. The importance score of a feature can represent the frequency of the number of times a feature is split in the tree structure or the gain brought by the split. The more times a feature is split in the tree structure or the greater the gain brought by the split, the higher its importance score in SOC estimation.

[0085] During the training process of the maximum gradient boosting method, the algorithm automatically calculates the importance score of each feature. Specifically, the maximum gradient boosting feature importance is calculated based on the information gain in each tree: (12)

[0086] in, is feature k, Show features Split gain in decision tree j;

[0087] Step 3.6: Get the importance score of each feature by averaging the split gains of all features;

[0088] Step 3.7: Features with higher scores are more representative in SOH estimation. The present invention selects the four features with the largest scores as model input features;

[0089] Step 4: Feed the model input features into the dual-channel interpretability network, where Figure 3 As shown in the figure, the dual-channel interpretability network includes a global channel (KAN), a local channel (1D-CNN), an interactive attention mechanism, and a fully connected layer, specifically including the following steps:

[0090] Step 4.1: Normalize the model input features to the maximum and minimum values, eliminate the dimension difference, and then send them to the dual-channel interpretability network;

[0091] Step 4.2: The output of maximum and minimum normalization is processed by the global channel and the local channel respectively and then sent to the interactive attention mechanism; the global channel includes a hidden layer, a maximum and minimum normalization layer, a B-spline activation function and a regularization layer, where the model input feature is 4-dimensional data and the hidden layer is 9 nodes, which effectively reduces the dimension of the model input feature while retaining important information structure; the input feature of the global channel is processed by maximum and minimum normalization, and the input feature is scaled to a fixed range of [-1,1] to improve the calculation stability of the model; the B-spline function is used as the activation function, and the model structure is optimized through sparse regularization and pruning to reduce input redundancy and obtain the output of the global channel. The local channel includes a sliding window, a multi-scale convolution layer, a ReLU activation function, and a global pooling layer. The sliding window divides the input features into subsequences of fixed length. The window length is set to 32 and the step size is 10 to capture local temporal patterns. The multi-scale convolution layer is used to extract local patterns of different time scales in parallel and calculate the weighted sum of the local area. The ReLU activation function is used to set the negative values ​​of the input features to zero and retain the positive values ​​to focus on the positive area, screen out local features that are strongly related to battery aging, and introduce nonlinear expression capabilities to the model. The global pooling operation is used to compress the content extracted by the multi-scale convolution to the mean to retain the global temporal pattern.

[0092] Step 4.3: The interactive attention mechanism uses the output of the global channel as the input of its own query operation to guide the global degradation trend and locate the key time sequence segments related to the overall health status in the local channel; and uses the output of the local channel as the input of its own key and value to map the local dynamic response to the global degradation framework, enhancing the model's sensitivity to abnormal events; generates the attention weight matrix through the Softmax function to quantify the importance of global and local features; the formula of the interactive attention mechanism is as follows: (15)

[0093] in, , and represent linear transformations of queries, keys, and values, respectively, represents the transposition of the key, Represents the dimension of the key. Represents the Softmax activation function;

[0094] Step 4.4: The output of the interactive attention mechanism passes through the fully connected layer to obtain the SOH estimation result.

[0095] In view of the fact that the traditional weight parameters will be replaced by univariate function parameters at the edge of the network. This paper proposes a new type of SOH estimation network: dual-channel interpretable network (KAN-1DCNN). The design idea of ​​the dual-channel structure is to divide the network into local channels and global channels. The model input features are automatically learned by the fully connected layer after the dual-channel structure and the features from the two channels are fused to obtain the estimated SOH estimation result. The local channel uses one-dimensional convolution to extract local features and focuses on capturing local details. The global channel uses the interpretable layer to process global context information. The dual-channel interpretable network can help the model maintain sensitivity to local features and grasp global information at the same time. The dual-channel interpretable network does not rely on fixed activation functions. Instead, it uses learnable univariate functions, thereby enhancing its ability to adapt to complex data patterns. In the dual-channel interpretable network, each node does not perform any nonlinear transformation when summarizing the outputs of these functions. This feature enables the SOH estimation model to improve prediction accuracy and computational efficiency while maintaining a high level of model interpretability.

[0096] This paper uses four evaluation indicators to comprehensively evaluate the performance of the proposed KAN-1DCNN. These indicators help to objectively measure the prediction ability of the network and reveal its accuracy and robustness from different perspectives. The definitions of the four evaluation indicators are as follows:

[0097] Root Mean Square Error (RMSE): Root mean square error is the square root of the mean square error, which is used to measure the average difference between the network's predicted value and the true value, that is, the square root of the prediction error. RMSE is more sensitive to outliers and can be used to measure the accuracy of the network. Mean absolute error (MAE): Mean absolute error is the average of the absolute errors between the predicted value and the true value, which measures the average error of the prediction. Unlike MSE, MAE does not amplify the impact of large errors, so it can better reflect the overall accuracy of the prediction. Mean percentage absolute error (MAPE): Mean percentage absolute error is the average of the relative differences between the predicted value and the true value, expressed as a percentage. It measures the relative error of the network in different data ranges and can reflect the relative accuracy of the prediction. Maximum absolute error (MAXE): Maximum absolute error is the maximum absolute difference between the predicted value and the true value, which identifies the prediction error of the network in the worst case. MAXE is particularly sensitive to outliers and helps to understand the biggest risk of the network in prediction.

[0098] The present invention uses four different working condition data sets obtained from two test platforms, using L1, L2, L3, L5, L6, L7, L9, L10, L11, L13, L14 and L15 as training sets, and using L4, L8, L12 and L16 as test sets. In order to train the network, the present invention selects the mean absolute error function as the loss function of the network. The hyperparameters in this experiment are determined as follows: batchsize=32, epoch=160, learning rate=0.00225. The present invention uses the gradient-based AdamW optimization algorithm, which is a variant of the Adam optimization algorithm. The original Adam algorithm is improved by adding a weight decay mechanism, making the optimization algorithm more flexible and helping to better select hyperparameters.

[0099] The accuracy of the proposed KAN-1DCNN was verified by comparative experiments, and it was compared with multiple existing networks (TCN, LSTM and BiGRU). In this series of comparative experiments, the present invention maintained a unified experimental parameter setting and the same environmental conditions to ensure the comparability of the experimental results. The experimental results are shown in Table 3.

[0100] Table 3 Comparative experimental results

[0101] Observing the experimental results, it can be seen that KAN-1DCNN shows significant advantages in all evaluation indicators. Taking the L4 dataset as an example, compared with TCN and BiGRU, KAN-1DCNN improves RMSE, MAE, MAPE and MAXE by 82%, 77%, 81% and 63% respectively. For LSTM, KAN-1DCNN improves RMSE, MAE, MAPE and MAXE by 72%, 81%, 78% and 62% respectively. For BiGRU, KAN-1DCNN improves RMSE, MAE, MAPE and MAXE by 79%, 80%, 82% and 52% respectively.

[0102] The KAN-1DCNN network performs well in prediction accuracy and provides a reliable basis for battery health status assessment and management. The comprehensive comparative test results show that the KAN-1DCNN network is accurate and stable on the test set. Compared with existing networks such as TCN, LSTM and BiGRU, KAN-1DCNN shows significant advantages in multiple evaluation indicators. This strengthens the excellent performance of KAN-1DCNN in lithium-ion battery health status estimation and provides strong support and guidance for the optimization of battery management and maintenance strategies.

[0103] In the robustness experiment, the present invention evaluates the impact of noise on the performance of the KAN-1DCNN network, aiming to explore its performance under different noise levels. In the experiment, the present invention introduced noise of different amplitudes (50mV, 100mV and 150mV) to simulate the uncertainty of real battery data. By analyzing the experimental results, we can gain an in-depth understanding of the stability and robustness of the KAN-1DCNN network in the face of noise. To ensure the reliability of the experiment, this content adopts the above parameter settings and conducts experiments under the same environmental conditions. The results are shown in Table 4.

[0104] Table 4 Robustness experimental results

[0105] It is obvious from the result table that the performance of each network decreases with the increase of noise level. However, KAN-1DCNN still maintains a high prediction accuracy at each noise level. Taking the test set L8 as an example, as the noise level increases from 50mV to 150mV, the MAE of KAN-1DCNN only increases from 0.55% to 1.64%. This shows that KAN-1DCNN is highly robust to noise and can resist the influence of data uncertainty to a certain extent. In addition, the team noticed that KAN-1DCNN performed well on the L4 dataset. Although the performance of all networks decreased at high noise levels, KAN-1DCNN still maintained a relatively low evaluation index value, confirming its stability under different data qualities. The robustness experimental results further verified the excellence of KAN-1DCNN in dealing with noise. Compared with other networks, KAN-1DCNN showed stronger stability and robustness, which provided strong support for its reliability in actual battery data scenarios. This is of great significance for the practical application of battery data management, especially when external factors are involved.

[0106] The above shows and describes the basic principles, main features and advantages of the present invention. It should be understood by those skilled in the art that the present invention is not limited to the above embodiments. The above embodiments and descriptions are only for explaining the principles of the present invention. Without departing from the spirit and scope of the present invention, the present invention may have various changes and improvements, which fall within the scope of the present invention. The scope of protection of the present invention is defined by the attached claims and their equivalents.

Claims

1. A lithium battery SOH estimation method for a smart Yellow River underwater vehicle, characterized by: The following steps are involved: Step 1: Establish a lithium-ion battery cycle life test platform and obtain lithium battery data sets under four different working conditions; Step 2: Analyze and process the lithium battery data set to build a feature library, specifically: Step 2.1: Analyze the lithium battery dataset from the perspective of battery attributes and construct attribute features; Step 2.2: Analyze attribute characteristics from the perspective of data range and construct data range characteristics; Step 2.3: Analyze the data range characteristics from the perspective of mathematical statistics and construct mathematical features; Step 2.4: compose a feature library with attribute features, data range features, and mathematical features; Step 3: Use the highest correlation search method to select the highest correlation feature from the feature library to obtain the model input feature; Step 4: Feed the model input features into the dual-channel interpretability network to obtain the SOH estimation results.

2. The lithium battery SOH estimation method for the smart Yellow River underwater vehicle according to claim 1 is characterized by: The lithium battery data set is collected from four aging tests, all of which adopt constant current constant voltage (CC-CV) and constant current (CC) discharge methods; In constant current and constant voltage mode, the battery is charged to the upper cut-off voltage using a specified constant current, and then charged using the upper cut-off voltage until the current drops below 0.05C; in constant current discharge mode, the battery is discharged to a lower cut-off voltage at a specified constant current, and the cycle life test will stop until the capacity drops below 80% of the initial capacity; the operating condition settings of lithium-ion batteries in the aging test conditions are shown in Table 1: ； Among them, L1~L16 are the names of different lithium battery packs, and the lithium battery is a commercial pouch lithium-ion battery with a capacity of 12Ah.

3. The lithium battery SOH estimation method for the smart Yellow River underwater vehicle according to claim 1 is characterized by: In step 2.1, the attribute characteristics include voltage, current, capacity, energy, IC curve, capacity change, energy change, average voltage, ohmic internal resistance and polarization resistance, among which ohmic resistance , polarization resistance ; Indicates the current change, Indicates the voltage difference within 1s, Indicates voltage changes within 1s-20s; In step 2.2, the attribute features are divided into three parts for feature extraction according to the charging mode: CC stage, CV stage and CC-CV stage, and each charging stage is taken as a battery charging data curve; In step 2.3, each battery charging data curve is converted into 8 statistical values, namely: quantile, maximum value, minimum value, mean value, variance, kurtosis, skewness and slope, to capture the shape and position evolution of each cycle curve; quantile reflects the evolution by sampling the value of each quantile position; mean value measures the average level of each curve within a fixed range; variance is used to evaluate the unevenness of data stream distribution, and the larger the variance, the more uneven the distribution; skewness and kurtosis characterize the shape of each curve; slope represents the transverse polarization of the battery; In step 2.4, each characteristic attribute sequence of the 10 attribute features is divided into 3 charging segments according to the charging mode, and each charging segment is converted into 8 statistical values. Through the combined statistical method, a comprehensive feature library with 240 mechanical statistical fusion features is generated for the battery under study.

4. The lithium battery SOH estimation method for the smart Yellow River underwater vehicle according to claim 1 is characterized by: Step 3 is specifically as follows: Step 3.1: Evaluate the nonlinear correlation between the features in the feature library and the SOH, and select the features whose correlation is greater than the first set threshold; Step 3.2: Decentralize the selected features and make the mean of the features 0; Step 3.3: Calculate the covariance matrix of the decentralized results; Step 3.4: Perform eigenvalue decomposition and singular value decomposition of the covariance matrix to obtain eigenvalues ​​and eigenvectors to determine the variance contribution rate and cumulative variance contribution rate; Step 3.5: Select features whose cumulative variance contribution rate is greater than the second set threshold, and calculate the split gains of these features by the maximum gradient boosting method; Step 3.6: Average the split gains of the features to get the importance score of each feature; Step 3.7: Select several features with the largest importance scores as model input features.

5. The lithium battery SOH estimation method for the smart Yellow River underwater vehicle according to claim 1 is characterized by: The two-channel interpretability network includes a global channel, a local channel, an interactive attention mechanism, and a fully connected layer. Step 4 includes the following steps: Step 4.1: Perform maximum and minimum normalization on the model input features and then send them to the dual-channel interpretability network; Step 4.2: The outputs of the maximum and minimum normalization are processed by the global channel and the local channel respectively and then sent to the interactive attention mechanism; Step 4.3: The interactive attention mechanism uses the output of the global channel as the input of its own query operation to locate the key time sequence segments related to the overall health status in the local channel guided by the global degradation trend; and uses the output of the local channel as the input of its own key and value to map the local dynamic response to the global degradation framework; generates the attention weight matrix through the Softmax function to quantify the importance of global and local features; Step 4.4: The output of the interactive attention mechanism passes through the fully connected layer to obtain the SOH estimation result.

6. The lithium battery SOH estimation method for the smart Yellow River underwater vehicle according to claim 5 is characterized by: The formula of the interactive attention mechanism is as follows: （15）； in, , and represent linear transformations of queries, keys, and values, respectively, represents the transposition of the key, represents the dimension of the key, Represents the Softmax activation function.

7. The lithium battery SOH estimation method for the smart Yellow River underwater vehicle according to claim 5 is characterized by: The global channel includes a hidden layer, a maximum and minimum normalization layer, a B-spline activation function and a regularization layer, where the input feature is 4-dimensional data and the hidden layer is 9 nodes; the input feature of the global channel is subjected to maximum and minimum normalization processing and is scaled to a fixed range of [-1, 1]; the B-spline function is used as the activation function, and the model structure is optimized through sparse regularization and pruning to obtain the output of the global channel.

8. The lithium battery SOH estimation method for the smart Yellow River underwater vehicle according to claim 5 is characterized by: The local channel includes a sliding window, a multi-scale convolution layer, a ReLU activation function, and a global pooling layer. The sliding window divides the input features into subsequences of fixed length to capture local timing patterns. The multi-scale convolution layer is used to extract local patterns of different time scales in parallel and calculate the weighted sum of the local area. The ReLU activation function is used to set the negative values ​​of the input features to zero and retain the positive values ​​to focus on the positive area and screen out local features that are strongly correlated with battery aging. The global pooling operation is used to compress the content extracted by the multi-scale convolution into a mean to retain the global timing pattern.

Images