A battery state of health identification method, device, apparatus and medium

Abstract

The application provides a battery health state recognition method, device and equipment and a medium, and relates to the technical field of artificial intelligence, and comprises the following steps: acquiring multi-channel time sequence data of a battery to be recognized; inputting the multi-channel time sequence data into a physical feature enhancement model, adaptively extracting time sequence data corresponding to each channel, and fusing a plurality of features extracted to obtain a fused feature sequence; inputting the fused feature sequence into an adaptive gated recurrent model based on cumulative charge guidance, recursively deriving a hidden state of the fused feature sequence to obtain a terminal hidden state; and inputting the terminal hidden state into a classification model to obtain a health state recognition result of the battery to be recognized. The technical scheme of the embodiment of the application can improve the accuracy of feature recognition, avoid the problem that a pure data-driven model learns false time-dependent relationships in a long-term aging scenario, and improve the accuracy of the battery health state recognition result.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence technology, and in particular to a method, apparatus, device, and medium for identifying battery health status. Background Technology

[0002] With the rapid development of the new energy vehicle industry, the health status of power batteries, as a core component, directly affects the vehicle's range, safety performance, and residual value. Accurately identifying whether a battery is in a healthy, sub-healthy, or degraded state is crucial for battery management systems, used car evaluation, secondary utilization, and safety early warning systems. However, battery health degradation is a complex electrochemical process, externally characterized by the coordinated changes in multiple physical quantities such as voltage, current, temperature, and accumulated charge. These physical quantities exhibit highly coupled and nonlinear degradation patterns over time.

[0003] Existing methods typically input multi-channel time-series data directly into convolutional or recurrent networks for health status classification or regression prediction. This relies on manual feature engineering, and the quality of feature design directly impacts model performance. Furthermore, it struggles to fully utilize the dynamic changes within the time-series data. Treating multi-channel time-series data as a two-dimensional matrix and automatically extracting local patterns through convolutional neural networks can automatically learn features. However, the receptive field of the convolutional kernel is limited, making it difficult to capture long-term dependencies, and it doesn't consider the differentiated characteristics of each physical channel. Using Long Short-Term Memory (LSTM) networks or gated recurrent units to model time-series data can capture long-term dependencies over time and performs well in time-series prediction tasks. However, their gating mechanisms are entirely data-driven, and in long-term battery degradation scenarios, they are prone to learning time-series dependencies that do not align with the actual charging and discharging depth and rate due to changes in data distribution. Summary of the Invention

[0004] In view of this, the purpose of the present invention is to provide a battery health status identification method, apparatus, device and medium. By designing dedicated adaptive extraction methods for the degradation characteristics corresponding to different physical features, the accuracy of feature identification is improved. By constructing an adaptive gated loop model guided by accumulated charge, the problem of pure data-driven models learning false time dependencies in long-term aging scenarios is avoided, thereby improving the accuracy of battery health status identification results.

[0005] In a first aspect, embodiments of the present invention provide a method for identifying battery health status, including: Acquire multi-channel time-series data of the battery to be identified; The multi-channel time-series data is input into the physical feature enhancement model, and adaptive feature extraction is performed on the time-series data corresponding to each channel. The extracted features are then fused to obtain a fused feature sequence. The fused feature sequence is input into an adaptive gated loop model guided by cumulative power consumption, and the hidden state of the fused feature sequence is recursively calculated to obtain the final hidden state. The endpoint hidden state is input into the classification model to obtain the health status identification result of the battery to be identified.

[0006] In a preferred embodiment of the present invention, the above-mentioned input of the multi-channel time-series data into the physical feature enhancement model, adaptive feature extraction of the time-series data corresponding to each channel, and fusion of the extracted features to obtain a fused feature sequence include: Extract the timing data corresponding to the voltage channel, current channel, temperature channel and cumulative charge channel from the multi-channel timing data respectively; Plateau period identification is performed on the time series data corresponding to the voltage channel, and the identified plateau period information is used to perform convolutional encoding on the time series data corresponding to the voltage channel to obtain the plateau period sensing feature sequence. Multi-scale difference operations are performed on the time-series data corresponding to the current channel to obtain the current transient response feature sequence; For the time-series data corresponding to the temperature channel and the time-series data corresponding to the cumulative power channel, the coupling correlation and incremental response ratio are extracted according to the preset local window to obtain the coupling feature sequence; The plateau period sensing feature sequence, the current transient response feature sequence, and the coupling feature sequence are weighted and fused to obtain a fused feature sequence.

[0007] In a preferred embodiment of the present invention, the above-described plateau period identification of the time-series data corresponding to the voltage channel, and the convolutional encoding of the identified plateau period information on the time-series data corresponding to the voltage channel to obtain a plateau period sensing feature sequence, includes: The timing data corresponding to the voltage channel is subjected to first-order difference point by point to obtain the voltage change sequence; Perform a local moving average on the voltage change sequence to obtain a local voltage change intensity sequence; The plateau period threshold is determined based on the voltage change sequence; Each change intensity value in the local voltage change intensity sequence is compared with the plateau period threshold to determine the plateau period sensing modulation coefficient corresponding to each change intensity value; One-dimensional convolutional encoding is performed on the time-series data corresponding to the voltage channel, and the convolutional output is enhanced point by point using the plateau period sensing modulation coefficient to obtain the plateau period sensing feature sequence.

[0008] In a preferred embodiment of the present invention, the above-mentioned multi-scale difference operation on the time-series data corresponding to the current channel to obtain the current transient response feature sequence includes: According to the preset step size, the timing data corresponding to the current channel is forward differentiald to obtain the first differential result; According to the preset step size, the timing data corresponding to the current channel is subjected to backward differential to obtain the second differential result; The first difference result and the second difference result are concatenated to obtain the current transient response feature sequence.

[0009] In a preferred embodiment of the present invention, for the time-series data corresponding to the temperature channel and the time-series data corresponding to the cumulative charge channel, coupling correlation and incremental response ratio are extracted according to a preset local window to obtain a coupling feature sequence, including: For the time-series data corresponding to the temperature channel and the time-series data corresponding to the cumulative power channel, a local window is established with the acquisition time corresponding to each time-series data as the center. Within the local window, the coupling correlation is determined based on the standard deviation and covariance of the time-series data corresponding to the temperature channel and the time-series data corresponding to the cumulative power channel, respectively. The incremental response ratio is determined based on the local increments corresponding to the time-series data of the temperature channel and the time-series data of the cumulative power channel, respectively. The coupling correlation and incremental response ratio determined within multiple local windows are concatenated to obtain the coupling feature sequence.

[0010] In a preferred embodiment of the present invention, the above-mentioned inputting the fused feature sequence into an adaptive gated loop model guided by accumulated power, and performing hidden state recursion on the fused feature sequence to obtain the final hidden state, includes: Extract the time series data corresponding to the cumulative power consumption channel from the multi-channel time series data; The time-series data corresponding to the cumulative power channel is compressed to obtain the cumulative power guiding coefficient sequence; The fused feature sequence is input into the adaptive gated recurrent model guided by cumulative power consumption, and converted into an input vector through an input mapping layer; Based on a preset time window, determine the cumulative power increment rate sequence corresponding to the time-series data of the cumulative power channel; The cumulative power guiding coefficient sequence and the cumulative power increment rate sequence are respectively centered and linearly combined to obtain the physical modulation term; Based on the physical modulation term, the update gate pre-activation value and reset gate pre-activation value determined according to the standard gated loop unit are updated to obtain the update gate and reset gate after physical guidance; Based on the update gate and reset gate after the physical guidance, the hidden state of the fused feature sequence is recursively calculated to obtain the final hidden state.

[0011] In a preferred embodiment of the present invention, the above-mentioned inputting the endpoint hidden state into the classification model to obtain the health status identification result of the battery to be identified includes: The endpoint hidden state is input into the classification model to obtain classification scores corresponding to multiple state categories; The classification scores corresponding to each state category are normalized to obtain the probability value corresponding to each state category. The health status identification result of the battery to be identified is determined based on the probability value corresponding to each of the aforementioned state categories.

[0012] Secondly, embodiments of the present invention also provide a battery health status identification device, comprising: The data acquisition module is used to acquire multi-channel time-series data of the battery to be identified; The feature extraction module is used to input the multi-channel time series data into the physical feature enhancement model, perform adaptive feature extraction on the time series data corresponding to each channel, and fuse the extracted features to obtain a fused feature sequence. The gated loop module is used to input the fused feature sequence into the adaptive gated loop model guided by the cumulative power, and to perform hidden state recursion on the fused feature sequence to obtain the final hidden state; The identification module is used to input the endpoint hidden state into the classification model to obtain the health status identification result of the battery to be identified.

[0013] Thirdly, embodiments of the present invention also provide an electronic device, including a processor and a memory, wherein the memory stores computer-executable instructions that can be executed by the processor, and the processor executes the computer-executable instructions to implement the battery health status identification method of the first aspect described above.

[0014] Fourthly, embodiments of the present invention also provide a computer-readable storage medium storing computer-executable instructions, which, when invoked and executed by a processor, cause the processor to implement the battery health status identification method of the first aspect described above.

[0015] The embodiments of the present invention bring the following beneficial effects: This invention provides a battery health status identification method. By using a physical feature enhancement module, a dedicated adaptive extraction method is designed for the degradation characteristics corresponding to different physical features. This enables the model to specifically capture the degradation patterns of different features, improving the interpretability and robustness of the features. By constructing an adaptive gated loop model guided by accumulated charge, the charge / discharge depth and rate are used as physical priors to directly intervene in the update logic of the gated unit. This ensures that the time-series memory behavior conforms to the actual degradation patterns of the battery, avoiding the problem of pure data-driven models learning false time dependencies in long-term aging scenarios.

[0016] Other features and advantages of the invention will be set forth in the following description, or some features and advantages may be inferred from the description or determined without doubt, or may be learned by practicing the techniques described above.

[0017] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, preferred embodiments are described below in detail with reference to the accompanying drawings. Attached Figure Description

[0018] To more clearly illustrate the specific embodiments of the present invention or the technical solutions in the prior art, the drawings used in the description of the specific embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0019] Figure 1 A flowchart of a battery health status identification method provided in an embodiment of the present invention; Figure 2 A flowchart of another battery health status identification method provided in an embodiment of the present invention; Figure 3 A flowchart illustrating another battery health status identification method provided in an embodiment of the present invention; Figure 4 This is a schematic diagram of the feature space distribution obtained using the technical solution of the present invention; Figure 5 This is a schematic diagram of the feature space distribution obtained using the baseline method; Figure 6 The curves showing the changes in the guiding coefficient and the increment rate over normalization time; Figure 7 The curve showing the variation of the physical modulation term with normalized time; Figure 8 The curve showing the change of the update threshold with or without physical guidance as a function of normalized time; Figure 9This is a schematic diagram of the structure of a battery health status identification device provided in an embodiment of the present invention; Figure 10 This is a schematic diagram of the structure of an electronic device provided in an embodiment of the present invention. Detailed Implementation

[0020] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0021] With the rapid development of the new energy vehicle industry, the health status of power batteries, as a core component, directly affects the vehicle's range, safety performance, and residual value. Accurately identifying whether a battery is in a healthy, sub-healthy, or degraded state is crucial for battery management systems, used car evaluation, secondary utilization, and safety early warning systems. However, battery health degradation is a complex electrochemical process, externally characterized by the coordinated changes in multiple physical quantities such as voltage, current, temperature, and accumulated charge. These physical quantities exhibit highly coupled and nonlinear degradation patterns over time.

[0022] Based on this, the battery health status identification method provided in this embodiment of the invention improves the accuracy of feature identification and avoids the problem of pure data-driven models learning false time dependencies in long-term aging scenarios, thereby improving the accuracy of battery health status identification results.

[0023] To facilitate understanding of this embodiment, a battery health status identification method disclosed in this embodiment of the invention will first be described in detail.

[0024] Example 1 This invention provides a method for identifying battery health status. Figure 1 This is a flowchart illustrating a battery health status identification method provided in an embodiment of the present invention. Figure 1 As shown, the battery health status identification method may include the following steps: Step S101: Obtain multi-channel timing data of the battery to be identified.

[0025] The battery to be identified refers to a battery whose health status (healthy, sub-healthy, or deteriorating) needs to be determined. This is typically found in practical applications such as new energy vehicles and energy storage systems. Multi-channel time-series data refers to a data sequence of multiple physical quantities changing over time, recorded synchronously at a fixed sampling frequency during the battery's charging and discharging process. Each physical quantity constitutes a "channel," and all channels are strictly aligned in time. In this embodiment of the invention, four channels are typically included: voltage, current, temperature, and accumulated charge.

[0026] Specifically, standard charge-discharge tests can be performed on the battery to be identified, or multi-channel time-series data can be collected under actual operating conditions. The data acquisition process must ensure a fixed sampling frequency, synchronously recording voltage (terminal voltage changes), current (magnitude and direction of charge / discharge current), temperature (battery surface or internal temperature), and cumulative charge (total charge or discharge obtained through current integration). The acquisition duration should cover a complete charge-discharge cycle or a typical operating condition segment to fully reflect the battery's dynamic behavior. For example, the number of sampling points can be fixed at 5600. The acquired multi-channel time-series data may contain noise or outliers. In practical applications, data cleaning (e.g., removing sensor fault points) and normalization (e.g., minimum-maximum value normalization) are typically performed first to eliminate the influence of differences in dimensions and amplitudes between different physical quantities, and to remove obvious erroneous data points caused by sensor malfunctions or communication interruptions.

[0027] Step S102: Input the multi-channel time series data into the physical feature enhancement model, perform adaptive feature extraction on the time series data corresponding to each channel, and fuse the extracted features to obtain a fused feature sequence.

[0028] Multi-channel time-series data of batteries for new energy vehicles includes four physical channels: voltage, current, temperature, and cumulative charge. Each channel exhibits different but interconnected change patterns during the health degradation process. If conventional convolutional networks are used to process the four channels uniformly, key information with clear physical significance, such as voltage plateau period, current transient response, and the coupling relationship between temperature and cumulative charge, is easily overlooked, resulting in unstable extracted time-series features and unclear classification boundaries.

[0029] The physical feature enhancement model is a neural network structure with specialized processing modules designed for the inherent characteristics of different channels. Instead of employing a uniform feature extraction method, it extracts features with clear correlations to battery degradation based on the physical meaning of voltage, current, temperature, and accumulated charge. For example, it can extract the voltage plateau length for voltage, transient response anomalies for current, and coupling anomalies between temperature rise and accumulated charge for temperature and accumulated charge. The extracted feature sequences are concatenated according to feature dimensions (e.g., 1-dimensional for voltage, 4-dimensional for current, and 2-dimensional for temperature-electric coupling, for a total of 7 dimensions). Then, for each sampling point, a small attention network adaptively learns the importance weights of each feature dimension at each sampling point. Finally, a weighted sum is performed to obtain a single-channel fused feature sequence. This fused feature sequence changes over time, and the value at each sampling point integrates the most useful physical information for that sampling point. For the voltage channel, plateau period detection, slope analysis, wavelet transform, or convolutional coding can be used to highlight the steady segment shape; for the current channel, multi-scale difference, statistical features, or transient mode learning can be used to capture response characteristics; for the temperature and cumulative charge channels, the coupling relationship can be modeled by local correlation, incremental response ratio, or regression slope.

[0030] Step S103: Input the fused feature sequence into the adaptive gated loop model guided by cumulative power, and perform hidden state recursion on the fused feature sequence to obtain the endpoint hidden state.

[0031] The adaptive gated loop model guided by cumulative charge level utilizes the battery's cumulative charge level channel (representing the total charge or discharge) as a physical prior to guide the update behavior of the gated loop unit. The cumulative charge level reflects the charge / discharge depth and process stage, and is a key factor in battery aging. The adaptive gated loop model directly adds the cumulative charge level and its rate of change to the gate pre-activation value, making the gating opening and closing more consistent with the actual working patterns of the battery (e.g., when the cumulative charge level is low, it tends to retain historical states; when the cumulative charge level is high and progresses rapidly, it tends to update the state). After the fused feature sequence is input into the adaptive gated loop model guided by cumulative charge level, it is progressively pushed forward according to the update mechanism of the gated loop unit, from the first time position to the last time position, to obtain the endpoint hidden state.

[0032] Step S104: Input the endpoint hidden state into the classification model to obtain the health status identification result of the battery to be identified.

[0033] The classification model maps the endpoint hidden state to the battery health state (at least two of the following: healthy, sub-healthy, and declining). This can be implemented using various methods such as fully connected layers with Softmax, support vector machines, gradient boosting trees, or ensemble classifiers. Specifically, the endpoint hidden state is input into a pre-trained classification layer. The classification layer obtains classification scores in three dimensions through a fully connected layer, corresponding to the healthy, sub-healthy, and declining categories, respectively. The category with the highest classification score is taken as the health state identification result for the battery to be identified.

[0034] Furthermore, inputting the endpoint hidden state into the classification model to obtain the health status identification result of the battery to be identified includes: inputting the endpoint hidden state into the classification model to obtain classification scores corresponding to multiple status categories; normalizing the classification scores corresponding to each status category to obtain the probability value corresponding to each status category; and determining the health status identification result of the battery to be identified based on the probability value corresponding to each status category.

[0035] Specifically, the endpoint hidden state is input into the classification model, which employs a fully connected structure. The model maps the endpoint hidden state to a three-dimensional output, corresponding to classification scores for three categories: healthy, sub-healthy, and declining. The weight matrix of the classification layer is denoted as... The size is 3×64; the bias vector is denoted as The size is 3.

[0036] Softmax normalization is performed on the 3D classification scores to obtain the class prediction probability vector. , represents the probability distribution of the battery to be identified belonging to three categories: healthy, sub-healthy, and declining. The health status identification result of the battery to be identified is taken from the category with the highest probability value.

[0037] This invention provides a battery health status identification method. By using a physical feature enhancement module, a dedicated adaptive extraction method is designed for the degradation characteristics corresponding to different physical features. This enables the model to specifically capture the degradation patterns of different features, improving the interpretability and robustness of the features. By constructing an adaptive gated loop model guided by accumulated charge, the charge / discharge depth and rate are used as physical priors to directly intervene in the update logic of the gated unit. This ensures that the time-series memory behavior conforms to the actual degradation patterns of the battery, avoiding the problem of pure data-driven models learning false time dependencies in long-term aging scenarios.

[0038] Example 2 This invention also provides another method for identifying battery health status; this method is implemented based on the method in the above embodiments; the method focuses on describing the specific implementation of inputting the multi-channel time-series data into the physical feature enhancement model, performing adaptive feature extraction on the time-series data corresponding to each channel, and fusing the extracted multiple features to obtain a fused feature sequence.

[0039] Figure 2 A flowchart of another battery health status identification method provided in an embodiment of the present invention is shown below. Figure 2 As shown, the battery health status identification method may include the following steps: Step S201: Obtain multi-channel timing data of the battery to be identified.

[0040] Define the multi-channel timing data as X, with a size of N×C, where N=5600 represents the number of time sampling points, C=4 represents the number of physical channels, and the 4 physical channels are denoted as voltage channels. Current path Temperature channel and cumulative power channel Each channel is a one-dimensional time series with a length of 5600; To avoid the impact of different units on subsequent data processing, the four physical channels are first normalized. The normalization method can be either minimum-maximum value normalization or mean-standard deviation normalization.

[0041] Step S202: Extract the timing data corresponding to the voltage channel, current channel, temperature channel and cumulative charge channel from the multi-channel timing data.

[0042] Extracting voltage channels from multi-channel timing data X X is a one-dimensional voltage sequence of length 5600, representing the change in the battery's terminal voltage throughout the sampling process; the current channel is extracted from the multi-channel time-series data X. X is a one-dimensional current sequence of length 5600, representing the changes in charge and discharge current of the battery at each sampling time; the temperature channel is extracted from the multi-channel time-series data X. and cumulative power channel Among them, temperature channel It is a one-dimensional temperature sequence of length 5600, and a cumulative charge channel. It is a one-dimensional cumulative electricity sequence with a length of 5600.

[0043] Step S203: Plateau period identification is performed on the time series data corresponding to the voltage channel, and the identified plateau period information is used to perform convolutional encoding on the time series data corresponding to the voltage channel to obtain the plateau period sensing feature sequence.

[0044] Voltage channel is one of the most sensitive signals in identifying the health status of batteries in new energy vehicles. Healthy batteries typically have a long and stable voltage plateau period during stable charging or discharging, while the plateau period of degraded batteries is shorter and the fluctuations within the plateau are greater. Conventional convolution only focuses on local numerical patterns and is difficult to highlight the plateau period. The plateau period refers to the time interval in voltage time-series data where voltage changes very slowly (close to constant). During battery charging and discharging, healthy batteries typically have a long and stable voltage plateau period, while the plateau period of degraded batteries is shorter and the fluctuations are greater. Plateau period identification is usually achieved by calculating the absolute value of the first-order difference, local smoothing, and comparing it with an adaptive threshold. This embodiment first identifies the plateau period and then uses the plateau period information to guide convolutional encoding, enabling the model to focus on voltage morphology changes directly related to the health status.

[0045] Furthermore, plateau period identification is performed on the time-series data corresponding to the voltage channel. The identified plateau period information is then used to perform convolutional encoding on the time-series data corresponding to the voltage channel to obtain a plateau period-aware feature sequence. This includes: performing first-order difference on the time-series data corresponding to the voltage channel point by point to obtain a voltage change sequence; performing a local moving average on the voltage change sequence to obtain a local voltage change intensity sequence; determining a plateau period threshold based on the voltage change sequence; comparing each change intensity value in the local voltage change intensity sequence with the plateau period threshold to determine the plateau period-aware modulation coefficient corresponding to each change intensity value; performing one-dimensional convolutional encoding on the time-series data corresponding to the voltage channel, and using the plateau period-aware modulation coefficient to enhance the convolution output point by point to obtain the plateau period-aware feature sequence.

[0046] Specifically, for the voltage channel In other words, the first-order difference is calculated point-by-point from the time-series data corresponding to the voltage channel to obtain the voltage change at adjacent time points, which is used to determine whether the current time point is in a plateau period. In specific implementation, the first-order difference at position t (i.e., the t-th sampling point) can be defined as the difference between the current sampling point and the next sampling point; when there is no next sampling point at the end of the time-series data, the end value can be directly copied as the first-order difference, or the nearest neighbor padding method can be used to handle the boundary. The sequence formed by the first-order differences at all positions is taken as the voltage change sequence.

[0047] Perform a local moving average on the absolute value of the first-order difference, that is, perform a local moving average on the absolute value of each value in the sequence of changes in electrical quantity to obtain a sequence of local voltage change intensity. The local moving window length can be 9, 11 or 15. The smaller the local voltage change intensity, the more gradual the voltage change near the current position is, and the more likely it is in a plateau period.

[0048] The plateau threshold is adaptively determined based on the statistical results of the absolute values ​​of the first-order differences in the voltage change sequence for the first 10% of the time period. This allows the threshold to automatically adjust with the rate of change of the reference voltage for different battery samples, preventing the fixed threshold from becoming ineffective across different samples. In practice, the absolute values ​​of the first-order differences corresponding to the first 560 sampling points can be taken first, and their mean and standard deviation can be calculated. Then, the "mean plus 0.5 times the standard deviation" can be used as the plateau threshold. Alternatively, a certain quantile value of this data portion can be directly used as the plateau threshold.

[0049] Each value in the local voltage change intensity sequence (called the local voltage change intensity value) is compared with the plateau threshold to generate the plateau sensing modulation coefficient. This represents the degree of enhancement of the voltage plateau period at position t, and its value is preferably set to the range of 0.5 to 1.0. When the intensity of the local voltage change at position t is lower than the plateau period threshold, it indicates that position t is closer to the plateau period, and the voltage can be increased. Set to close to 1.0; when the intensity of the local voltage change at position t is higher than the plateau threshold, it indicates that position t is more likely to be in a rapid change phase, and can be... Set it to be close to 0.5.

[0050] In practical implementation, to avoid abrupt changes at platform boundaries, Preferably, a linear mapping or sigmoid mapping is used for continuous generation, rather than using only hard determination of 0 and 1. In this embodiment of the invention, a modulation coefficient between 0.5 and 1.0 can be generated through a continuous mapping function. Specifically, the local voltage change intensity value at each position t is first calculated. And an adaptive plateau threshold θ, then using a linear mapping or a Sigmoid mapping to... Smooth mapping to the [0.5, 1.0] interval: when When the value is far below the threshold, the mapping value is close to 1.0, indicating that the position is in a stable plateau period, and feature representation should be enhanced; when When the value is much higher than the threshold, the mapping value is close to 0.5, indicating that the location is in a rapidly changing segment, and the enhancement intensity should be reduced; when Near the threshold, the mapping value transitions continuously between 0.5 and 1.0, avoiding abrupt changes at the plateau boundary; thus, the obtained modulation coefficients... As point-wise weights, the voltage features after convolutional encoding are enhanced element-wise, enabling the model to achieve a higher response during the plateau period and maintain the basic response during the non-plateau period.

[0051] For voltage channel One-dimensional convolutional coding is performed, and the convolutional output is enhanced point-by-point using plateau-sensing modulation coefficients to obtain the voltage plateau-sensing feature sequence. , representing the single-channel timing feature after plateau-stage bootstrapping, with a length of 5600.

[0052] In practical implementation, in order to simultaneously cover the stable shape inside the platform and the transition shape at the platform boundary, the convolution kernel length is preferably 17, corresponding to a relative position range of 8 sampling points before and after.

[0053] It should be noted that the embodiments of the present invention do not simply learn the voltage pattern automatically through convolution, but first locate the plateau period based on the voltage change rate, and then increase the feature extraction intensity of the plateau period position, so that the voltage plateau length, plateau stability and the upward trend of the plateau end can be more clearly entered into the model.

[0054] Step S204: Perform multi-scale difference operation on the time series data corresponding to the current channel to obtain the current transient response feature sequence.

[0055] The current channel reflects the battery's response speed, stability, and fluctuation characteristics to charge and discharge commands. Healthy batteries typically exhibit smoother current changes, while degraded batteries are more prone to overshoot, oscillations, and hysteresis during rate changes or operating condition switching. To fully characterize and utilize transient response features, this invention does not directly perform a single convolution on the current channel. Instead, it first constructs multi-scale differential features and then forms a current transient response feature sequence. For example, forward and backward differentials of different step sizes are calculated for the current time-series data. Examples include forward differentials with step sizes of 1 and 3 (reflecting instantaneous changes and short-term trends) and backward differentials (reflecting the magnitude of changes relative to historical data). By concatenating the differentials across multiple scales, a set of multi-dimensional features describing the current transient response (rise rate, fall rate, oscillation, hysteresis) is formed, i.e., the current transient response feature sequence. Each position in the current transient response feature sequence corresponds to a multi-dimensional vector (e.g., 4-dimensional), characterizing the transient change characteristics of the current at that moment. Degraded batteries are more prone to overshoot, oscillations, and hysteresis during rate switching or load changes, and these anomalies are reflected in this feature sequence.

[0056] Furthermore, multi-scale difference operations are performed on the time-series data corresponding to the current channel to obtain a current transient response feature sequence, including: performing forward difference on the time-series data corresponding to the current channel according to a preset step size to obtain a first difference result; performing backward difference on the time-series data corresponding to the current channel according to a preset step size to obtain a second difference result; and concatenating the first difference result and the second difference result to obtain the current transient response feature sequence.

[0057] Specifically, with each position t as the center, forward differences with step size 1 and step size 3 are calculated respectively to obtain the first difference result. The forward difference is used to characterize the change trend of the current position to the short future time. The step size of 1 is used to capture instantaneous changes, and the step size of 3 is used to capture short-term trend changes. When t+1 or t+3 exceeds the end of the sequence, the boundary can be filled by copying the end value.

[0058] Centered on each position t, backward differences with step sizes of 1 and 3 are calculated respectively to obtain the second difference result. The backward difference is used to characterize the change magnitude of the current position relative to the historical time. The step size of 1 is used to describe the change magnitude of the most recent step, and the step size of 3 is used to describe a smoother historical change trend. When t-1 or t-3 is less than the starting point of the sequence, the boundary can be padded by copying the starting value.

[0059] The "forward difference with step size 1, backward difference with step size 1, forward difference with step size 3, and backward difference with step size 3" are concatenated according to the feature dimension to obtain the current transient response feature sequence. The size is 5600×4, where each position t corresponds to a 4-dimensional feature, which describes the upward change, backward change, short-term forward trend and short-term backward trend of the current position.

[0060] Current transient response characteristic sequence A further linear scaling or one-dimensional 1×1 convolution is performed to adjust the numerical range of the four difference features, making them more suitable for fusion with other physical features.

[0061] It should be noted that separating the current rise rate, fall rate, local oscillations, and short-time hysteresis from the original sequence makes it easier to identify abnormal responses of a degraded battery under rate switching or load changes.

[0062] Step S205: Extract coupling correlation and incremental response ratio from the time-series data corresponding to the temperature channel and the time-series data corresponding to the cumulative power channel according to a preset local window to obtain a coupling feature sequence.

[0063] There is a clear physical coupling between the temperature channel and the accumulated charge channel. As the accumulated charge increases, the temperature usually rises synchronously. When the battery degrades, the internal resistance increases, the temperature rise rate accelerates, and the same change in accumulated charge corresponds to more pronounced temperature fluctuations. To reflect this coupling relationship, this invention does not process the two channels separately, but instead calculates their correlation and response intensity within a local window to obtain the dual-channel coupling characteristics. The preset local window refers to a fixed length (e.g., 20 sampling points) taken forward and backward from each position on both the temperature channel and the accumulated charge channel. Within this window, the relationship between temperature and accumulated charge is statistically analyzed to capture their coupling dynamics over a local time period.

[0064] Furthermore, for the time-series data corresponding to the temperature channel and the cumulative power channel, coupling correlation and incremental response ratio are extracted according to a preset local window to obtain a coupling feature sequence, including: establishing a local window centered on the acquisition time of each time-series data for the time-series data corresponding to the temperature channel and the cumulative power channel; determining the coupling correlation based on the standard deviation and covariance of the time-series data corresponding to the temperature channel and the cumulative power channel, respectively, within the local window; determining the incremental response ratio based on the local increments of the time-series data corresponding to the temperature channel and the cumulative power channel, respectively; and concatenating the coupling correlation and incremental response ratio determined in multiple local windows to obtain the coupling feature sequence.

[0065] A local window is established with each time position t as the center. The radius of the local window is denoted as L, which represents how many sampling points are taken before and after the current position to participate in the local statistics. In a convenient implementation method, L=20, that is, the window length is 41. If the window exceeds the sequence boundary, the nearest neighbor padding method can be used.

[0066] Within a local window, the mean and standard deviation of the temperature channel and the cumulative energy channel are calculated separately. Then, their local covariance is calculated, and the covariance is normalized by the product of their standard deviations to obtain the local coupling correlation value. The local coupling correlation value is used to describe whether the changes in cumulative energy and temperature are synchronized near that time position. The larger the value, the stronger the synchronization between the two. In addition, to avoid the instability of the value due to the standard deviation being too small, a very small positive number can be added to the denominator. , Can be taken .

[0067] Calculate the local increments of the temperature channel and the cumulative charge channel at the current position, and further calculate the "ratio of local temperature increment to local charge increment" to obtain the incremental response coefficient, which is used to characterize the average degree of temperature change when the cumulative charge increases by one unit.

[0068] In practical implementation, to reduce the impact of noise, the local increment can be the difference between the data collected at the current sampling point and the data collected at the previous sampling point. Alternatively, the mean of the data collected at the first 3 or 5 sampling points can be determined, and the local increment can be the difference between the current sampling point and the mean. In addition, a very small positive number is also added to the denominator. To prevent division by zero.

[0069] By concatenating the local coupling correlation values ​​and incremental response coefficients along the feature dimensions, the coupling feature sequence of the temperature channel and the accumulated charge channel is obtained. The size is 5600×2, and each time position corresponds to a 2D feature, which represents synchronicity and response intensity respectively.

[0070] It should be noted that, in this embodiment of the invention, temperature change is no longer treated as a separate thermal information, but rather temperature change and the cumulative charge propagation process are analyzed together in the same local window, which makes it easier to identify the decay feature of "abnormally rapid temperature rise under the same cumulative charge propagation".

[0071] Step S206: The plateau period sensing feature sequence, the current transient response feature sequence, and the coupling feature sequence are weighted and fused to obtain a fused feature sequence.

[0072] In this embodiment of the invention, a channel attention fusion mechanism can be used to automatically assign each feature (voltage plateau period sensing feature sequence) to the following features: Current transient response characteristic sequence Coupled characteristic sequence of temperature channel and cumulative charge channel Time-related weights are assigned to obtain the final fused feature sequence.

[0073] Specifically, the voltage plateau period sensing feature sequence 1D features, current transient response feature sequence The 4D features, temperature channel and cumulative charge channel coupling feature sequence The 2D features are concatenated according to their feature dimensions to obtain a multi-physical concatenated feature sequence. The size is 5600×7, with each position corresponding to a 7-dimensional physical enhancement feature.

[0074] For each position t, take the multi-physical splicing feature sequence. The 7-dimensional vector at this location is input into a small attention network, which outputs 7 unnormalized weight scores.

[0075] In one implementation, a small attention network can adopt a structure of "fully connected layer + ReLU activation + fully connected layer", with an input dimension of 7, an intermediate hidden dimension of 14 or 16, and an output dimension of 7.

[0076] Furthermore, Softmax normalization is performed on the 7 unnormalized weight scores to obtain 7 attention weights. The c-th attention weight is defined as _c_. , representing the importance of the c-th feature channel at position t, satisfying that the sum of the 7 attention weights is 1.

[0077] Using 7 attention weights to concatenate multi-physical feature sequences Perform weighted summation to obtain the fused feature sequence. , is a single-channel timing sequence of length 5600, represented as .

[0078] in, Represents the fusion feature sequence The eigenvalue at position t; (t,c) represents the multi-physics splicing feature sequence. The value at position t, in the c-th channel; This represents the attention weight of the c-th feature channel at position t.

[0079] It should be noted that, through the channel attention fusion mechanism, the model can automatically adjust the intensity of attention to features such as voltage plateau, current transients and temperature rise coupling according to the initial, middle and final stages of charging and different health states, thereby improving the ability to distinguish the battery health state under complex operating conditions.

[0080] Step S207: Input the fused feature sequence into an adaptive gated loop model guided by accumulated power, and perform hidden state recursion on the fused feature sequence to obtain the endpoint hidden state.

[0081] Step S208: Input the endpoint hidden state into the classification model to obtain the health status identification result of the battery to be identified.

[0082] The battery health status identification method provided in this invention enhances the perception of voltage channel plateau periods, enabling the model to accurately capture the voltage plateau length, stability, and end-rise trend closely related to health status, avoiding the shortcomings of conventional convolution which only focuses on local values ​​and ignores the overall shape. The multi-scale differential features of the current channel separate the rise rate, fall rate, local oscillations, and short-term lags from the original sequence, making it easier to identify abnormal responses of degraded batteries under rate switching or load changes. The coupling features of temperature and accumulated charge (local correlation and incremental response ratio) push temperature rise and charge... By analyzing the data within the same local window, the degradation characteristic of "abnormal temperature rise under the same charge" can be keenly detected, rather than simply treating temperature as isolated thermal information. The weighted fusion based on the attention mechanism can automatically adjust the intensity of attention to features such as voltage plateau, current transients, and temperature rise coupling according to the initial, middle, and final stages of charging and different health states. This enables the model to maintain stable discrimination ability under complex operating conditions, significantly improving the interpretability of features, the fit to battery degradation patterns, and the robustness under multiple operating conditions, providing high-quality input for subsequent cumulative charge-guided time series modeling.

[0083] Example 3 This invention also provides another battery health status identification method; this method is implemented based on the method in the above embodiments; the method focuses on describing the specific implementation of inputting the fused feature sequence into an adaptive gated loop model guided by cumulative power, and performing hidden state recursion on the fused feature sequence to obtain the final hidden state.

[0084] Figure 3 A flowchart of another battery health status identification method provided in an embodiment of the present invention is shown below. Figure 3 As shown, the battery health status identification method may include the following steps: Step S301: Obtain multi-channel timing data of the battery to be identified.

[0085] Step S302: Input the multi-channel time series data into the physical feature enhancement model, perform adaptive feature extraction on the time series data corresponding to each channel, and fuse the extracted features to obtain a fused feature sequence.

[0086] Step S303: Extract the timing data corresponding to the cumulative power channel from the multi-channel timing data.

[0087] It is understandable that battery health status is essentially accumulated and evolved over time. Relying solely on local features is insufficient to characterize the long-term differences between the three states of health, sub-health, and degradation. Conventional gated cyclic cells can process long-term data, but their gating updates are entirely data-driven and do not utilize the physical prior of accumulated charge, which directly corresponds to the depth of charge and discharge. Therefore, they are still insufficient in modeling the gradual degradation of the health status.

[0088] In this embodiment of the invention, the accumulated power channel is used as a physical guidance signal to construct an adaptive gated loop unit based on the accumulated power guidance. This makes the model more inclined to retain historical states when the accumulated power is low, and more inclined to absorb current abnormal features when the accumulated power is high and the progress is fast, so that the time series modeling process is more in line with the actual working law of the battery.

[0089] Specifically, the cumulative power channel is extracted from the multi-channel time-series data X. This is a one-dimensional sequence of length 5600, which increases during the charging phase, representing the total amount of electricity accumulated or discharged by the battery within the current sample. It should be noted that this is the accumulated electricity channel. It is a sequence that has undergone normalization, eliminating the differences in absolute cumulative energy values ​​among different samples.

[0090] Step S304: Compress the time-series data corresponding to the cumulative power channel to obtain the cumulative power guiding coefficient sequence.

[0091] Specifically, to make the changes in the guiding signal smoother, the normalized cumulative charge sequence is further compressed using Sigmoid to obtain the cumulative charge guiding coefficient sequence. , is a one-dimensional guide sequence of length 5600, used to represent the depth of charge / discharge corresponding to each time position, and, as time progresses, The overall trend is monotonous and increasing.

[0092] Step S305: Input the fused feature sequence into the adaptive gated loop model guided by cumulative power consumption, and convert it into an input vector through the input mapping layer.

[0093] Fuse feature sequences The input is sequentially given to an adaptive gated recurrent model guided by cumulative battery power, due to the fusion of feature sequences. Since there is only one scalar feature at each location, it is first mapped to a 64-dimensional input vector through an input mapping layer. The mapped input vector is denoted as . , represents the 64-dimensional input representation of the gated loop unit at position t, with the hidden state dimension fixed at 64, and the initial hidden state denoted as . It can be initialized as a vector of all zeros.

[0094] Step S306: Determine the cumulative power increment rate sequence corresponding to the time-series data of the cumulative power channel according to the preset time window.

[0095] In the cumulative power channel The local increment rate is calculated using a fixed-length window to obtain the cumulative charge increment rate sequence. This represents the average propulsion speed of the accumulated charge within a local time range, reflecting the charging and discharging intensity of the current stage. The window length is denoted as... This represents the number of sampling points looked back at. In a convenient implementation method, =100, and when the position t is less than 100, the starting point can be fixed as the first sampling point, or the average increment rate can be calculated using the "actual available length".

[0096] In the specific implementation, in the cumulative power channel The local incremental rate is calculated using a fixed-length window. Specifically, firstly, the normalized cumulative charge channel... Extract a one-dimensional sequence of length 5600, and then set a fixed-length lookback window. (For example, 100 sampling points) are used to calculate the local increment rate at each position t. For each position t in the sequence, the previous value is taken. sampling points (if t is less than 10 ... Then, all available points from the 1st sampling point to the tth sampling point are taken. The cumulative charge difference between the starting point and the endpoint within this window is calculated, and then divided by the actual number of time steps within the window (i.e., the number of sampling intervals from the starting point to the endpoint). This yields the average advancing speed of this position within a local time range. This speed reflects the rate of change of the cumulative charge within this local time period, i.e., the charging and discharging intensity. Furthermore, this speed sequence is linearly normalized to obtain the cumulative charge increment speed sequence. The numerical range was compressed to between 0 and 1 to eliminate the dimensional effects caused by differences in charge / discharge rates between different samples. Furthermore, the cumulative charge increment rate sequence was... Perform another linear normalization to stabilize its value between 0 and 1, so as to avoid the value being too large during subsequent gating modulation, which would cause training instability.

[0097] It should be noted that, Used to express "which stage of the charging and discharging process has been reached". Both are used to express "how fast the current stage is progressing" and together determine whether the current moment should favor preserving historical information or updating current information.

[0098] Step S307: The cumulative power guiding coefficient sequence and the cumulative power increment rate sequence are respectively centered and linearly combined to obtain the physical modulation term.

[0099] Based on the cumulative electricity guiding coefficient and cumulative electricity increment rate Constructing physical modulation terms, specifically, first, the cumulative charge guiding coefficient... and cumulative electricity increment rate The physical modulation term is obtained by performing a centralization process (e.g., subtracting 0.5 from each term) and then linearly combining the terms according to a preset ratio. This physical modulation term is used to indicate whether it is more suitable to enhance the influence of the current input at this moment.

[0100] In an easy-to-implement approach, when linearly combining components according to a preset ratio, the weight of the cumulative power guiding coefficient is 1.0, and the weight of the cumulative power increment rate is 0.5.

[0101] It should be noted that, This is the result of applying Sigmoid compression to the normalized cumulative charge sequence. The output range of the Sigmoid function is (0,1), therefore... The numerical range is between 0 and 1; This is the cumulative charge increment rate after linear normalization, and its value range is also stabilized between 0 and 1. If not centered, both values ​​would be in the range [0,1], and their linear combination would necessarily be a non-negative number (minimum 0). This means that the physical modulation term can only provide "zero or positive" adjustment, and cannot provide "negative" adjustment. However, the design goal of this invention is: when the cumulative charge is large and the progress is fast, the physical modulation term should be positive to enhance the activity of the update gate and reset gate (absorbing new information); when the cumulative charge is small and the progress is slow, the physical modulation term should be negative to suppress the update gate and reset gate (preserving historical information). Therefore, by subtracting 0.5, and If the range of values ​​changes from [0,1] to [-0.5,0.5], then their linear combination result can be a negative number, zero, or a positive number, thereby realizing bidirectional adjustment of the gating pre-activation value (it can be increased or decreased), making the guidance mechanism more flexible and in line with physical intuition.

[0102] Step S308: Based on the physical modulation term, update the update gate pre-activation value and reset gate pre-activation value determined according to the standard gating loop unit to obtain the update gate and reset gate after physical guidance.

[0103] Based on the current input vector Hidden state from the previous moment The update gate pre-activation value and the reset gate pre-activation value are calculated separately according to the standard gating loop unit method. Both the update gate pre-activation value and the reset gate pre-activation value are 64-dimensional vectors, which are used to determine how much historical information to retain and how much historical information to reset at the current moment.

[0104] The physical modulation term is added to the pre-activation values ​​of the update gate and the reset gate, respectively, and then constrained to between 0 and 1 by the Sigmoid function to obtain the update gate and the reset gate after physical boot.

[0105] It should be noted that if a sigmoid function is applied first and then the gating value is directly amplified, the gating value is prone to exceeding 1, which is detrimental to training stability. Adding the physical modulation term to the pre-activated gating value before uniformly passing it through the sigmoid function makes it easier to maintain numerical stability and also achieves the design goal of "enhancing updates when the accumulated energy is large and the progress is fast." When the accumulated energy guiding coefficient... Smaller and cumulative power increment rate When the value is small, the physical modulation term is small or even negative. In this case, the update gate after physical guidance tends to decrease, and the model is more inclined to retain the existing hidden state; when the cumulative energy guidance coefficient is small... Larger and cumulative power increment rate When the value is large, the physical modulation term increases, and the update gate and reset gate after physical guidance become more active. The model is more inclined to absorb the current input and refresh historical information. By directly embedding the monotonic physical quantity of accumulated power into the gating mechanism, the gating recurrent unit no longer relies solely on data distribution to automatically learn time dependencies, but is guided by a clear charge and discharge depth, thereby improving the model's ability to characterize long-term degradation trajectories.

[0106] Step S309: Based on the update gate and reset gate after physical guidance, perform hidden state recursion on the fused feature sequence to obtain the endpoint hidden state.

[0107] The reset door, guided by physical means, restores the previous hidden state. Element-by-element filtering is performed to obtain historical information that participated in the generation of the current candidate state. The larger the reset gate, the more historical details are retained; the smaller the reset gate, the more inclined to ignore historical influences.

[0108] In practice, element-by-element filtering refers to using a reset gate guided by physical principles. (64-dimensional vector) for the hidden state at the previous time step (64-dimensional vector) multiply corresponding positions, that is Here, ⊙ represents element-wise multiplication; specifically, it resets the value of the k-th element in the gate. With the k-th element in the hidden state Multiplying the values ​​yields the amount of historical information retained in that dimension. When the value of a certain dimension of the reset gate is close to 1, the historical information in the corresponding dimension is almost completely retained. When it is close to 0, the historical information in the corresponding dimension is suppressed. This allows the model to control the inflow of historical information at a fine granular level of the feature dimension, rather than uniformly filtering the entire state vector.

[0109] Use the current input vector The candidate hidden state is input into the candidate state generation layer along with the historical information filtered by the reset gate. After linear transformation and hyperbolic tangent activation function, the candidate hidden state is obtained. The candidate hidden state represents the new state proposal value formed by "current input + filtered historical information" at the current time.

[0110] The update gate, guided by physical means, hides the state from the previous moment. The hidden state at the current time is obtained by performing element-wise weighted fusion with the candidate hidden states. A larger update gate indicates that the current position is more inclined to adopt new information; a smaller update gate indicates that the current position is more inclined to continue the historical state.

[0111] Following the time sequence from t=1 to t=5600, the final hidden state is obtained. , is a 64-dimensional vector that represents the cumulative result of all timing information during the entire charging and discharging process.

[0112] Based on this, feature sequences are fused. It is further converted into a deep temporal representation with long-term memory, and the update process of this representation is always constrained by the physical prior of the accumulated charge. Therefore, it is more suitable for describing the entire process of the evolution of batteries for new energy vehicles from healthy to sub-healthy and then to deterioration.

[0113] Furthermore, the adaptive gated loop unit based on accumulated charge guidance constructed in this embodiment of the invention directly embeds the physical process of accumulated charge into the update logic of the gated loop unit, making the temporal memory behavior of the model conform to the physical laws of battery charge and discharge depth and rate changes. The specific structure is as follows: Input layer: fuses feature sequences (Single-channel timing) is converted into a 64-dimensional input vector through an input mapping layer. The hidden state dimension is fixed at 64.

[0114] Physical guidance coefficient calculation: Extracting the normalized cumulative charge channel from the original sample The guiding coefficient is obtained by Sigmoid compression. (Characterizing the charging and discharging phase), and simultaneously calculating the local incremental rate through a fixed-length window. (Characterizing propulsion speed), the two are centered (reduced by 0.5) and linearly combined to form a physical modulation term.

[0115] Gating modulation: The pre-activation values ​​of the update gate and the reset gate are calculated according to the standard gating loop unit. Then, the physical modulation term is directly added to these two pre-activation values. After Sigmoid activation, the update gate and the reset gate are obtained after physical guidance, so as to realize the intervention of physical prior on the gating mechanism. State Update: Utilize the reset gate after guidance to restore the hidden state from the previous moment. Perform element-by-element filtering (multiplying corresponding positions), combined with the current input. Generate candidate hidden states; then utilize the updated gate after bootstrapping. The hidden state at the current time is obtained by weighted fusion with the candidate states. ; Output layer: Recursively calculate to the final hidden state. (64 dimensions), serving as deep temporal features for subsequent classification.

[0116] Step S310: Input the endpoint hidden state into the classification model to obtain the health status identification result of the battery to be identified.

[0117] End point hidden state Although multiple physical features and long-term degradation information have been integrated, the three states of health, sub-health, and decline may still overlap in the feature space. In particular, the sub-health state is located in the transition range and is prone to intersecting with the health or decline state. When only a normal Softmax classification layer is used, the model focuses more on the probability fitting of the output layer, making it difficult to compress intra-class differences and widen the inter-class distance.

[0118] In this embodiment of the invention, learnable class centers and class center constraint loss are introduced during the classification training phase to hide the endpoint state of samples of the same class. By clustering towards the corresponding class center in the 64-dimensional feature space and imposing higher constraints on boundary samples farther from the class center, the clarity and robustness of the classification boundary are improved.

[0119] Specifically, before the classification training phase, the first step is to collect raw samples of new energy vehicle batteries and construct a dataset. The data collected includes new energy vehicle batteries in different health states, covering the entire lifecycle from new batteries to severely degraded batteries. During the data collection process, the batteries are placed under standard charge-discharge test conditions to simulate typical charge-discharge behaviors in actual vehicle operation, including constant current charging, constant voltage charging, and dynamic discharging. Within each sampling period, data from four physical channels—voltage, current, temperature, and cumulative charge—are collected simultaneously. The data from these four channels are strictly aligned in time to form a complete raw sample.

[0120] To ensure the samples fully reflect the dynamic changes of the battery during charging and discharging, each sample contains a fixed number of 5600 sampling points, meaning each sample includes multi-channel data from 5600 consecutive sampling moments. During data acquisition, data is recorded at a fixed sampling frequency to ensure the continuity and synchronization of the four physical quantities—voltage, current, temperature, and accumulated charge—over time. The raw data is first cleaned to remove obviously erroneous data points caused by sensor malfunctions or communication interruptions. Then, the cleaned data is divided into samples, with each of the 5600 consecutively collected sampling points grouped into a single sample.

[0121] During dataset construction, each sample was labeled according to the actual health status of the battery. The labeling categories were divided into three types: healthy, sub-healthy, and degraded. The healthy category corresponds to battery samples with a capacity retention rate above 80%, no significant increase in internal resistance, and good charge / discharge performance. The sub-healthy category corresponds to battery samples with a capacity retention rate between 60% and 80%, a certain degree of increase in internal resistance, and slight degradation in charge / discharge performance, but not yet reaching the end-of-life standard. The degraded category corresponds to battery samples with a capacity retention rate below 60%, significantly increased internal resistance, severely degraded charge / discharge performance, and having reached or approached the end-of-life standard. The labeling process was determined comprehensively based on the battery's capacity test results, internal resistance test results, and cycle life test data to ensure the accuracy and consistency of the labeling.

[0122] After the annotation is completed, all samples are divided into training set and test set according to the preset ratio. The training set is used for subsequent model training, and the test set is used to evaluate the recognition performance of the model.

[0123] During the classification training phase, a learnable class center vector is defined for each of the three health status categories, with each health category corresponding to a class center vector. Sub-health category corresponding to class center vector The class center vector corresponding to the decline category The dimensions of the three class center vectors are all 64, and they correspond to the endpoint hidden state. The dimensions are consistent.

[0124] To avoid the feature distribution being too random at the beginning of model training, which would cause distortion in class center initialization, we first pre-trained steps S102, S103 and the classification layer using only cross-entropy loss. The number of pre-training rounds can be 1 to 3, or several training batches; the purpose is to first hide the endpoint state. It has preliminary separability.

[0125] After pre-training, the endpoint hidden states h_5600 of the first 10% of the training set are extracted. Then, the samples are grouped according to the true labels, and the mean of the endpoint hidden states for each class is calculated. The mean is used as the initial value of the class center for the corresponding class. If the number of samples of a certain class in the first 10% of the samples is insufficient, samples of that class can be added from subsequent training samples before calculating the mean.

[0126] class center vector , , It is added to the model as a trainable parameter and updated synchronously with the network parameters during subsequent joint training.

[0127] For each sample in each training batch, read its endpoint hidden state. and real labels ,in, This represents the 64-dimensional endpoint hidden state of the i-th sample. This represents the true class label of the i-th sample, with a value of 0, 1, or 2.

[0128] Based on real labels Select the corresponding class center vector It calculates the Euclidean distance from the current sample to the corresponding class center. The larger the Euclidean distance, the more the sample deviates from the main feature distribution of the class.

[0129] Furthermore, within the current training batch, the average distance from samples of the same category to the class center of that category is calculated and used as the reference distance for that category in the current batch. If there is only one sample of a certain category in the current batch, the moving average distance of the previous few batches can be used as the reference distance to ensure the stability of the weight calculation.

[0130] The adaptive weight of the sample is calculated based on the ratio of the current sample distance to the distance of similar reference samples. , where represents the penalty strength of the i-th sample in the class center constraint loss; the distance from the current sample to the corresponding class center is denoted as . The average distance between samples of the same category in the current batch is denoted as . ;but .

[0131] In practical implementation, to avoid training fluctuations caused by excessively large weights for individual outlier samples, the following can be used: Limit it to a preset range, for example, between 0.5 and 2.0.

[0132] Use sample adaptive weights The class center constraint loss is obtained by weighting the squared class center distances of each sample and averaging them over the entire training batch. This method is used to compress intra-class differences and focus on optimizing boundary samples that are far from the class center, thereby improving the transition boundary between healthy, sub-healthy, and declining states. It should be noted that this embodiment of the invention does not perform intra-class shrinkage uniformly on all samples, but rather imposes stronger constraints on samples that are "more likely to be misclassified and have a higher degree of outlier status." Therefore, it is more suitable for solving the problem of blurred boundaries between the sub-healthy state and the other two states.

[0133] Hide the finish line The input is a classification model, which uses a fully connected structure to map the 64-dimensional input to a 3-dimensional output, corresponding to classification scores for three categories: healthy, sub-healthy, and declining.

[0134] Softmax normalization is performed on the 3D classification scores to obtain the class prediction probability vector. Let represent the probability distribution of the i-th sample belonging to one of the three categories: healthy, sub-healthy, or declining. The probability vector is predicted based on the category. and real labels Calculate the cross-entropy loss to obtain the classification loss. Classification loss With class center constraint loss We perform a weighted summation to obtain the total loss. , represented as Where λ represents the weighting coefficient of the class center constraint loss, used to balance classification accuracy and feature space compactness. In a convenient implementation approach, λ = 0.1.

[0135] Gradient descent is used to jointly train the physical feature enhancement model, the adaptive gated recurrent model guided by cumulative charge, the classification model, and the class center vectors end-to-end. The optimizer can be the Adam optimizer, and the learning rate can be set to 0.001.

[0136] In one embodiment, a comparative experiment on feature space distribution is conducted to reveal how this method improves the separability of the three states of health, sub-health, and decline from the perspective of feature representation. Figure 4 This is a schematic diagram of the feature space distribution obtained using the technical solution of the present invention. Figure 5 This is a schematic diagram of the feature space distribution obtained using the baseline method. The experiment used principal component analysis to reduce the 64-dimensional endpoint hidden state output by the model to a two-dimensional space, facilitating visualization of intra-class compactness and inter-class separation. Figure 5 The diagram shows the feature space distribution obtained by the baseline method without introducing class center constraint loss. Its model structure is the same as the technical solution in this embodiment, but it only uses cross-entropy loss during training, does not include a class center constraint loss term, and does not explicitly optimize the feature space. Figure 4-5 As shown, the three colors represent three categories of samples: healthy, sub-healthy, and declining. Figure 4 As shown, the three categories of samples—healthy, sub-healthy, and declining—each form three separate clusters. Each cluster has a high density of samples, with clear intervals between clusters. Furthermore, the confidence ellipse around each cluster has a small area, indicating that intra-class differences are effectively compressed and inter-class boundaries are clear. Figure 5 As shown, there is significant overlap among the three types of samples, particularly with blurred boundaries between the sub-healthy, healthy, and declining categories. The large areas of the confidence ellipses indicate substantial intra-class differences and difficulty in distinguishing between classes. Experimental results demonstrate that the learnable class center and class center constraint loss employed in this invention, by applying stronger constraints to boundary samples through adaptive weights, effectively enhance the discriminative power of the feature space, making samples of different health states more compact and separated within the feature space.

[0137] In one embodiment, an analysis experiment of the gating update mechanism guided by accumulated charge is conducted to verify how the physical guidance gating loop unit proposed in this embodiment of the invention uses the prior accumulated charge to adjust the gating behavior. The experiment plots the accumulated charge guidance coefficient, local increment rate, physical modulation term, and update threshold with and without physical guidance on the same coordinate system to demonstrate the impact of the guidance mechanism on gating updates. Figure 6 The curves show the changes in the guiding coefficient and increment rate over normalization time. For example... Figure 6 As shown, the guiding coefficient increases monotonically from zero to one with normalization time, characterizing the gradual deepening of charge and discharge depth; the incremental rate curve fluctuates, reflecting the change in instantaneous speed during charging. Figure 7 This is the curve showing the variation of the physical modulation term with normalized time. For example... Figure 7 The physical modulation term is determined by both the guiding coefficient and the incremental rate. It is negative in the early stage of charging and discharging and turns positive in the middle and late stages, indicating that the modulation term can adjust the gating pre-activation value bidirectionally according to the charging and discharging stage and rate. Figure 8 This is a curve showing the change of the update threshold with and without physical guidance over normalized time. For example... Figure 8 The results show that the update gate without physical guidance is driven solely by data, exhibiting a random fluctuation pattern that is unrelated to the physical laws governing the charging and discharging process. The update gate with physical guidance, however, maintains a low value during the negative phase of the physical modulation term and increases significantly after the physical modulation term turns positive. This indicates that the model tends to retain historical states when the accumulated charge is low and the progress is slow, while more actively absorbing current input features when the accumulated charge is high and the progress is fast. In the coordinate axis, the normalized time is a dimensionless relative time, and the coefficient values, modulation terms, and gate values ​​are all dimensionless numerical values, ranging from zero to one or symmetrically distributed on both sides of zero. Experimental results demonstrate that this invention embeds the physical prior of accumulated charge into the update logic of the gated recurrent unit, making the model's temporal memory behavior conform to the physical laws governing changes in battery charging and discharging depth and rate, thereby more accurately depicting the long-term evolution of the health state.

[0138] The battery health status identification method provided in this invention uses a physically guided hidden state recursion. The endpoint hidden state not only integrates multi-channel physical enhancement features, but also carries long-term dependency information consistent with the actual charge and discharge depth and rate process of the battery. Compared with the purely data-driven GRU / LSTM, this solution can effectively avoid learning false time dependencies in long-term degradation scenarios, and significantly improve the identification accuracy and generalization robustness of the three states of health, sub-health, and decay (especially the transitional sub-health).

[0139] Example 4 Corresponding to the above method embodiments, this invention provides a battery health status identification device. Figure 9 This is a schematic diagram of a battery health status identification device provided in an embodiment of the present invention, as shown below. Figure 9 As shown, the battery health status identification device may include: Data acquisition module 401 is used to acquire multi-channel time-series data of the battery to be identified; The feature extraction module 402 is used to input the multi-channel time series data into the physical feature enhancement model, perform adaptive feature extraction on the time series data corresponding to each channel, and fuse the extracted features to obtain a fused feature sequence. The gated loop module 403 is used to input the fused feature sequence into the adaptive gated loop model guided by the cumulative power, and to perform hidden state recursion on the fused feature sequence to obtain the final hidden state; The identification module 404 is used to input the endpoint hidden state into the classification model to obtain the health status identification result of the battery to be identified.

[0140] The battery health status identification device provided in this invention uses a physical feature enhancement module to design dedicated adaptive extraction methods for the degradation characteristics corresponding to different physical features. This enables the model to specifically capture the degradation patterns of different features, improving the interpretability and robustness of the features. By constructing an adaptive gated loop model guided by accumulated charge, the charging and discharging depth and rate are used as physical priors to directly intervene in the update logic of the gated unit. This makes the time-series memory behavior conform to the actual degradation patterns of the battery, avoiding the problem of pure data-driven models learning false time dependencies in long-term aging scenarios.

[0141] In some embodiments, the feature extraction module 402 is further configured to: Extract the timing data corresponding to the voltage channel, current channel, temperature channel and cumulative charge channel from the multi-channel timing data respectively; Plateau period identification is performed on the time series data corresponding to the voltage channel, and the identified plateau period information is used to perform convolutional encoding on the time series data corresponding to the voltage channel to obtain the plateau period sensing feature sequence. Multi-scale difference operations are performed on the time-series data corresponding to the current channel to obtain the current transient response feature sequence; For the time-series data corresponding to the temperature channel and the time-series data corresponding to the cumulative power channel, the coupling correlation and incremental response ratio are extracted according to the preset local window to obtain the coupling feature sequence; The plateau period sensing feature sequence, the current transient response feature sequence, and the coupling feature sequence are weighted and fused to obtain a fused feature sequence.

[0142] In some embodiments, the feature extraction module 402 is further configured to: The timing data corresponding to the voltage channel is subjected to first-order difference point by point to obtain the voltage change sequence; Perform a local moving average on the voltage change sequence to obtain a local voltage change intensity sequence; The plateau period threshold is determined based on the voltage change sequence; Each change intensity value in the local voltage change intensity sequence is compared with the plateau period threshold to determine the plateau period sensing modulation coefficient corresponding to each change intensity value; One-dimensional convolutional encoding is performed on the time-series data corresponding to the voltage channel, and the convolutional output is enhanced point by point using the plateau period sensing modulation coefficient to obtain the plateau period sensing feature sequence.

[0143] In some embodiments, the feature extraction module 402 is further configured to: According to the preset step size, the timing data corresponding to the current channel is forward differentiald to obtain the first differential result; According to the preset step size, the timing data corresponding to the current channel is subjected to backward differential to obtain the second differential result; The first difference result and the second difference result are concatenated to obtain the current transient response feature sequence.

[0144] In some embodiments, the feature extraction module 402 is further configured to: For the time-series data corresponding to the temperature channel and the time-series data corresponding to the cumulative power channel, a local window is established with the acquisition time corresponding to each time-series data as the center. Within the local window, the coupling correlation is determined based on the standard deviation and covariance of the time-series data corresponding to the temperature channel and the time-series data corresponding to the cumulative power channel, respectively. The incremental response ratio is determined based on the local increments corresponding to the time-series data of the temperature channel and the time-series data of the cumulative power channel, respectively. The coupling correlation and incremental response ratio determined within multiple local windows are concatenated to obtain the coupling feature sequence.

[0145] In some embodiments, the gated loop module 403 is further configured to: Extract the time series data corresponding to the cumulative power consumption channel from the multi-channel time series data; The time-series data corresponding to the cumulative power channel is compressed to obtain the cumulative power guiding coefficient sequence; The fused feature sequence is input into the adaptive gated recurrent model guided by cumulative power consumption, and converted into an input vector through an input mapping layer; Based on a preset time window, determine the cumulative power increment rate sequence corresponding to the time-series data of the cumulative power channel; The cumulative power guiding coefficient sequence and the cumulative power increment rate sequence are respectively centered and linearly combined to obtain the physical modulation term; Based on the physical modulation term, the update gate pre-activation value and reset gate pre-activation value determined according to the standard gated loop unit are updated to obtain the update gate and reset gate after physical guidance; Based on the update gate and reset gate after the physical guidance, the hidden state of the fused feature sequence is recursively calculated to obtain the final hidden state.

[0146] In some embodiments, the identification module 404 is further configured to: The endpoint hidden state is input into the classification model to obtain classification scores corresponding to multiple state categories; The classification scores corresponding to each state category are normalized to obtain the probability value corresponding to each state category. The health status identification result of the battery to be identified is determined based on the probability value corresponding to each of the aforementioned state categories.

[0147] The device provided in this embodiment of the invention has the same implementation principle and technical effect as the aforementioned method embodiment. For the sake of brevity, any parts not mentioned in the device embodiment can be referred to the corresponding content in the aforementioned method embodiment.

[0148] Example 5 This invention also provides an electronic device for running the above-described battery health status identification method; see [link to previous document]. Figure 10The diagram shows the structure of an electronic device, which includes a memory 500 and a processor 501. The memory 500 is used to store one or more computer instructions, which are executed by the processor 501 to implement the battery health status identification method described above.

[0149] Furthermore, Figure 10 The electronic device shown also includes a bus 502 and a communication interface 503. The processor 501, the communication interface 503 and the memory 500 are connected via the bus 502.

[0150] The memory 500 may include high-speed random access memory (RAM) and may also include non-volatile memory, such as at least one disk storage device. Communication between this system network element and at least one other network element is achieved through at least one communication interface 503 (which can be wired or wireless), such as the Internet, wide area network, local area network, metropolitan area network, etc. The bus 502 can be an ISA bus, PCI bus, or EISA bus, etc. The bus can be divided into address bus, data bus, control bus, etc. For ease of representation, Figure 10 The symbol is represented by a single double-headed arrow, but this does not mean that there is only one bus or one type of bus.

[0151] Processor 501 may be an integrated circuit chip with signal processing capabilities. In implementation, each step of the above method can be completed by the integrated logic circuitry in the hardware of processor 501 or by instructions in software form. Processor 501 can be a general-purpose processor, including a Central Processing Unit (CPU), a Network Processor (NP), etc.; it can also be a Digital Signal Processor (DSP), an Application Specific Integrated Circuit (ASIC), a Field-Programmable Gate Array (FPGA), or other programmable logic devices, discrete gate or transistor logic devices, or discrete hardware components. It can implement or execute the methods, steps, and logic block diagrams disclosed in the embodiments of this invention. The general-purpose processor can be a microprocessor or any conventional processor. The steps of the methods disclosed in the embodiments of this invention can be directly manifested as execution by a hardware decoding processor, or execution by a combination of hardware and software modules in the decoding processor. The software module can reside in a readily available storage medium in the art, such as random access memory, flash memory, read-only memory, programmable read-only memory, electrically erasable programmable memory, or registers. This storage medium is located in memory 500, and processor 501 reads information from memory 500 and, in conjunction with its hardware, completes the steps of the method described in the foregoing embodiments.

[0152] This invention also provides a computer-readable storage medium storing computer-executable instructions. When these computer-executable instructions are called and executed by a processor, they cause the processor to implement the aforementioned battery health status identification method. For specific implementation details, please refer to the method embodiments, which will not be repeated here.

[0153] The computer program product for battery health status identification provided in this embodiment of the invention includes a computer-readable storage medium storing non-volatile program code executable by a processor. The instructions included in the program code can be used to execute the methods described in the preceding method embodiments. For specific implementation details, please refer to the method embodiments, which will not be repeated here.

[0154] Those skilled in the art will understand that, for the sake of convenience and brevity, the specific working processes of the systems, devices, and units described above can be referred to the corresponding processes in the foregoing method embodiments, and will not be repeated here.

[0155] In the several embodiments provided by this invention, it should be understood that the disclosed systems, apparatuses, and methods can be implemented in other ways. The apparatus embodiments described above are merely illustrative. For example, the division of units is only a logical functional division, and in actual implementation, there may be other division methods. Furthermore, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Additionally, the coupling or direct coupling or communication connection shown or discussed may be through some communication interfaces; the indirect coupling or communication connection between devices or units may be electrical, mechanical, or other forms.

[0156] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0157] In addition, the functional units in the various embodiments of the present invention can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit.

[0158] If the aforementioned functions are implemented as software functional units and sold or used as independent products, they can be stored in a processor-executable, non-volatile, computer-readable storage medium. Based on this understanding, the technical solution of this invention, essentially, or the part that contributes to the prior art, or a portion of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0159] Finally, it should be noted that the above-described embodiments are merely specific implementations of the present invention, used to illustrate the technical solutions of the present invention, and not to limit it. The scope of protection of the present invention is not limited thereto. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that any person skilled in the art can still modify or easily conceive of changes to the technical solutions described in the foregoing embodiments within the technical scope disclosed in the present invention, or make equivalent substitutions for some of the technical features; and these modifications, changes, or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention, and should all be covered within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A method for identifying battery health status, characterized in that, include: Acquire multi-channel timing data of the battery to be identified; The multi-channel time-series data is input into the physical feature enhancement model, and adaptive feature extraction is performed on the time-series data corresponding to each channel. The extracted features are then fused to obtain a fused feature sequence. The fused feature sequence is input into an adaptive gated loop model guided by cumulative power consumption, and the hidden state of the fused feature sequence is recursively calculated to obtain the final hidden state. The endpoint hidden state is input into the classification model to obtain the health status identification result of the battery to be identified; Specifically, the fused feature sequence is input into an adaptive gated loop model guided by accumulated battery power, and the hidden state of the fused feature sequence is recursively calculated to obtain the final hidden state, including: Extract the time series data corresponding to the cumulative power consumption channel from the multi-channel time series data; The time-series data corresponding to the cumulative power channel is compressed to obtain the cumulative power guiding coefficient sequence; The fused feature sequence is input into the adaptive gated recurrent model guided by cumulative power consumption, and converted into an input vector through an input mapping layer; Based on a preset time window, determine the cumulative power increment rate sequence corresponding to the time-series data of the cumulative power channel; The cumulative power guiding coefficient sequence and the cumulative power increment rate sequence are respectively centered and linearly combined to obtain the physical modulation term; Based on the physical modulation term, the update gate pre-activation value and reset gate pre-activation value determined according to the standard gated loop unit are updated to obtain the update gate and reset gate after physical guidance; Based on the update gate and reset gate after the physical guidance, the hidden state of the fused feature sequence is recursively calculated to obtain the final hidden state.

2. The method according to claim 1, characterized in that, The multi-channel time-series data is input into the physical feature enhancement model. Adaptive feature extraction is performed on the time-series data corresponding to each channel, and the extracted features are fused to obtain a fused feature sequence, including: Extract the timing data corresponding to the voltage channel, current channel, temperature channel and cumulative charge channel from the multi-channel timing data respectively; Plateau period identification is performed on the time series data corresponding to the voltage channel, and the identified plateau period information is used to perform convolutional encoding on the time series data corresponding to the voltage channel to obtain the plateau period sensing feature sequence. Multi-scale difference operations are performed on the time-series data corresponding to the current channel to obtain the current transient response feature sequence; For the time-series data corresponding to the temperature channel and the time-series data corresponding to the cumulative power channel, the coupling correlation and incremental response ratio are extracted according to the preset local window to obtain the coupling feature sequence; The plateau period sensing feature sequence, the current transient response feature sequence, and the coupling feature sequence are weighted and fused to obtain a fused feature sequence.

3. The method according to claim 2, characterized in that, Plateau period identification is performed on the time series data corresponding to the voltage channel. The identified plateau period information is then used to perform convolutional encoding on the time series data corresponding to the voltage channel to obtain a plateau period sensing feature sequence, including: The timing data corresponding to the voltage channel is subjected to first-order difference point by point to obtain the voltage change sequence; Perform a local moving average on the voltage change sequence to obtain a local voltage change intensity sequence; The plateau period threshold is determined based on the voltage change sequence; Each change intensity value in the local voltage change intensity sequence is compared with the plateau period threshold to determine the plateau period sensing modulation coefficient corresponding to each change intensity value; One-dimensional convolutional encoding is performed on the time-series data corresponding to the voltage channel, and the convolutional output is enhanced point by point using the plateau period sensing modulation coefficient to obtain the plateau period sensing feature sequence.

4. The method according to claim 2, characterized in that, Multi-scale difference operations are performed on the time-series data corresponding to the current channel to obtain the current transient response feature sequence, including: According to the preset step size, the timing data corresponding to the current channel is forward differentiald to obtain the first differential result; According to the preset step size, the timing data corresponding to the current channel is subjected to backward differential to obtain the second differential result; The first difference result and the second difference result are concatenated to obtain the current transient response feature sequence.

5. The method according to claim 2, characterized in that, For the time-series data corresponding to the temperature channel and the time-series data corresponding to the cumulative charge channel, coupling correlation and incremental response ratio are extracted according to a preset local window to obtain a coupling feature sequence, including: For the time-series data corresponding to the temperature channel and the time-series data corresponding to the cumulative power channel, a local window is established with the acquisition time corresponding to each time-series data as the center. Within the local window, the coupling correlation is determined based on the standard deviation and covariance of the time-series data corresponding to the temperature channel and the time-series data corresponding to the cumulative power channel, respectively. The incremental response ratio is determined based on the local increments corresponding to the time-series data of the temperature channel and the time-series data of the cumulative power channel, respectively. The coupling correlation and incremental response ratio determined within multiple local windows are concatenated to obtain the coupling feature sequence.

6. The method according to claim 1, characterized in that, The endpoint hidden state is input into the classification model to obtain the health status identification result of the battery to be identified, including: The endpoint hidden state is input into the classification model to obtain classification scores corresponding to multiple state categories; The classification scores corresponding to each state category are normalized to obtain the probability value corresponding to each state category. The health status identification result of the battery to be identified is determined based on the probability value corresponding to each of the aforementioned state categories.

7. A battery health status identification device, characterized in that, include: The data acquisition module is used to acquire multi-channel time-series data of the battery to be identified; The feature extraction module is used to input the multi-channel time series data into the physical feature enhancement model, perform adaptive feature extraction on the time series data corresponding to each channel, and fuse the extracted features to obtain a fused feature sequence. The gated loop module is used to input the fused feature sequence into the adaptive gated loop model guided by the cumulative power, and to perform hidden state recursion on the fused feature sequence to obtain the final hidden state; The identification module is used to input the endpoint hidden state into the classification model to obtain the health status identification result of the battery to be identified; The gated loop module is further configured to: extract time-series data corresponding to the cumulative power channel from the multi-channel time-series data; compress the time-series data corresponding to the cumulative power channel to obtain a cumulative power guidance coefficient sequence; input the fused feature sequence into the adaptive gated loop model based on cumulative power guidance, and convert it into an input vector through an input mapping layer; determine the cumulative power increment rate sequence corresponding to the time-series data corresponding to the cumulative power channel according to a preset time window; perform centering processing and linear combination on the cumulative power guidance coefficient sequence and the cumulative power increment rate sequence respectively to obtain a physical modulation term; update the update gate pre-activation value and reset gate pre-activation value determined according to the standard gated loop unit according to the physical modulation term to obtain the update gate and reset gate after physical guidance; and perform hidden state recursion on the fused feature sequence according to the update gate and reset gate after physical guidance to obtain the endpoint hidden state.

8. An electronic device, characterized in that, The method includes a processor and a memory, the memory storing computer-executable instructions that can be executed by the processor, the processor executing the computer-executable instructions to implement the battery health status identification method according to any one of claims 1 to 6.

9. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores computer-executable instructions, which, when invoked and executed by a processor, cause the processor to implement the battery health status identification method according to any one of claims 1 to 6.

Images