Battery state of health estimation method and device based on multi-dimensional operating condition feature fusion

Abstract

The application provides a battery health state evaluation method and device based on multi-dimensional working condition feature fusion, and relates to the technical field of battery detection, and comprises the following steps: constructing a single cycle time sequence data matrix of a current battery cycle stage; performing adaptive time sequence decomposition and reconstruction guided by a degradation trend on the single cycle time sequence data matrix according to a current cycle order indicating the current battery cycle stage and a historical battery health state of a previous battery cycle stage in the single cycle time sequence data matrix, to obtain a reconstructed feature matrix; determining a degradation synergy vector according to a channel statistic quantity in the reconstructed feature matrix; inputting the reconstructed feature matrix and the degradation synergy vector into a degradation gate time sequence prediction model to obtain an initial battery health state prediction value of the current battery cycle stage; and performing degradation constraint smoothing on the initial battery health state prediction value to obtain a battery health state prediction value. The technical scheme of the embodiment of the application can improve the accuracy of battery health state prediction.

Description

Technical Field

[0001] This invention relates to the field of battery testing technology, and in particular to a method and apparatus for assessing battery health status based on the fusion of multi-dimensional operating condition features. Background Technology

[0002] With the large-scale application of electrochemical energy storage and power battery technologies, accurate assessment of battery health status has become a core issue for ensuring system safety, extending service life, and optimizing operation and maintenance strategies. During long-term cycling, battery capacity gradually declines. This decline is influenced by a combination of factors, including charge / discharge rate, temperature conditions, and cycle depth, and the aging mechanisms and degradation sensitivity characteristics differ significantly at different cycle stages. In the early stages of cycling, capacity decline is relatively gradual, with degradation signals often hidden in localized voltage and current fluctuations. However, in the later stages of cycling, macroscopic trends such as increased internal resistance and prolonged constant-voltage charging time gradually become the dominant characteristics of capacity decline.

[0003] Existing methods typically employ uniform standardization or normalization for all monitoring channels, weakening the responses of strong degradation indicator channels such as internal resistance and constant voltage charging time. Furthermore, they struggle to express the coupling relationships between channels related to the degradation process. Traditional time-series decomposition or filtering methods often use preset smoothing parameters or convolution scales, leading to the loss of early cycle detail information or insufficient modeling of late cycle trends, making it difficult to effectively capture the changing patterns of degradation features throughout the entire lifecycle. Regular time-series convolutional or recurrent neural networks use fixed network weights and channel response patterns for all cycle samples, failing to dynamically adjust the importance and temporal dependence of feature channels based on the degree of capacity degradation, thus limiting their adaptability to different degradation rates. Summary of the Invention

[0004] In view of this, the purpose of this invention is to provide a battery health status assessment method and device based on multi-dimensional operating condition feature fusion. By designing a channel gating mechanism driven by degradation cooperative vector, the output channels of each layer of separable convolution with dilatation depth are dynamically modulated in terms of importance. When faced with batteries with different aging levels, the internal parameters are adaptively changed, which breaks through the limitation that conventional fixed convolution response cannot cover the differences in degradation rate throughout the entire life cycle and improves the accuracy of battery health status prediction.

[0005] In a first aspect, embodiments of the present invention provide a battery health status assessment method based on multi-dimensional operating condition feature fusion, comprising: Based on the collected multi-channel sampling data of the battery during the current battery cycle stage, a single-cycle time-series data matrix for the current battery cycle stage is constructed. Based on the current cycle order of the current battery cycle stage and the historical battery health status of the previous battery cycle stage indicated in the single-cycle time-series data matrix, the single-cycle time-series data matrix is ​​subjected to adaptive time-series decomposition and reconstruction guided by degradation trend to obtain the reconstructed feature matrix. The degradation collaboration vector is determined based on the channel statistics in the reconstructed feature matrix; The reconstructed feature matrix and the degradation co-vector are input into the degradation-gated temporal prediction model to obtain the initial battery health state prediction value of the current battery cycle stage; wherein, the convolution channels and temporal attention weights in the degradation-gated temporal prediction model are dynamically modulated by the degradation co-vector; The initial battery health state prediction value is smoothed by degradation constraints to obtain the battery health state prediction value.

[0006] In a preferred embodiment of the present invention, the above-mentioned adaptive temporal decomposition and reconstruction of the single-cycle time-series data matrix guided by degradation trends, based on the current cycle order indicating the current battery cycle stage and the historical battery health status of the previous battery cycle stage in the single-cycle time-series data matrix, yields a reconstructed feature matrix, including: Obtain the channel mutual information coupling weights obtained during the training phase of the degradation-gated time series prediction model; Based on the single-cycle time-series data matrix, which indicates the current cycle order of the current battery cycle stage, the historical battery health status of the previous battery cycle stage, and the channel mutual information coupling weight, the single-cycle time-series data matrix is ​​standardized to obtain a fused standardized feature matrix. Based on the current cycle order and the historical battery health status, the fused standardized feature matrix is ​​subjected to adaptive temporal decomposition and reconstruction guided by degradation trends to obtain the reconstructed feature matrix.

[0007] In a preferred embodiment of the present invention, the single-cycle time-series data matrix is ​​standardized based on the single-cycle time-series data matrix, which indicates the current cycle order of the current battery cycle stage, the historical battery health status of the previous battery cycle stage, and the channel mutual information coupling weight, to obtain a fused standardized feature matrix, including: Standardization is performed on each channel in the single-cycle time-series data matrix to obtain the basic standardized feature matrix; Based on the current cycle order of the current battery cycle stage and the historical battery health status of the previous battery cycle stage indicated in the single-cycle timing data matrix, the channel degradation modulation coefficient corresponding to each channel in the single-cycle timing data matrix is ​​determined. Based on the channel degradation modulation coefficients, the basic normalized feature matrix is ​​differentially enhanced to obtain the degradation-enhanced normalized feature matrix; Based on the channel mutual information coupling weights, the basic normalized feature matrix and the degradation-enhanced normalized feature matrix are fused to obtain the fused normalized feature matrix.

[0008] In a preferred embodiment of the present invention, the above-described adaptive temporal decomposition and reconstruction of the fused standardized feature matrix based on the current cycle order and the historical battery health status, guided by degradation trends, yields a reconstructed feature matrix, including: Determine the adaptive smoothing factor based on the current loop order; Based on the adaptive smoothing factor, the fused standardized feature matrix is ​​recursively smoothed to obtain the trend feature matrix; The fluctuation residual matrix is ​​determined based on the fused standardized feature matrix and the trend feature matrix; The reconstructed feature matrix is ​​determined based on the fluctuation residual matrix and the trend feature matrix.

[0009] In a preferred embodiment of the present invention, determining the degenerate collaborative vector based on the channel statistics in the reconstructed feature matrix includes: Obtain the channel statistics corresponding to the reconstructed feature matrix and the cyclic scalar corresponding to the single-cycle time series data matrix; The cyclic scalar and the channel statistics are concatenated to obtain the cyclic statistical feature vector; Obtain the mean vector and standard deviation vector of the cyclic statistical features obtained during the training phase of the degenerate gated time series prediction model; The degradation collaboration vector is determined based on the mean vector of the cyclic statistical features, the standard deviation vector of the cyclic statistical features, and the cyclic statistical feature vector.

[0010] In a preferred embodiment of the present invention, the above-mentioned inputting the reconstructed feature matrix and the degradation cooperative vector into the degradation-gated time-series prediction model to obtain the initial battery health state prediction value of the current battery cycle stage includes: The reconstructed feature matrix and the degenerate co-vector are input into a depthwise separable temporal convolutional encoder. Temporal features are extracted using multi-layer dilated depthwise separable convolutions, and channel-level enhancement is performed on the output of each convolution layer to obtain a degenerate gated temporal feature matrix. The degradation collaborative vector is input into the attention query projection layer to obtain the degradation condition query vector; The temporal attention weights are determined based on the degradation condition query vector and the degradation gating temporal feature matrix. The degenerate gated temporal features are weighted and summed according to the temporal attention weights to obtain the temporal context vector; The temporal context vector and the degenerate co-vector are concatenated along the feature dimension to obtain the regression input vector; The regression input vector is input into a fully connected regression network to obtain the initial battery health status prediction value for the current battery cycle stage.

[0011] In a preferred embodiment of the present invention, the above-mentioned degradation constraint smoothing of the initial battery health state prediction value to obtain the battery health state prediction value includes: Obtain historical battery health status prediction values ​​from the previous battery cycle stage; The initial battery health state prediction value is constrained by the variation range of the historical battery health state prediction value to obtain the constrained prediction value. The battery health state prediction value is obtained by exponentially smoothing the constraint prediction value and the historical battery health state prediction value.

[0012] In a preferred embodiment of the present invention, the above-mentioned construction of a single-cycle time-series data matrix for the current battery cycle stage based on the collected multi-channel sampling data of the battery during the current battery cycle stage includes: According to the preset sampling frequency, multiple physical quantity data of the battery during the current battery cycle stage are collected to obtain multi-channel sampling data; According to the preset sampling length, the multi-channel sampled data is resampled at equal intervals to obtain a single-cycle time-series data matrix.

[0013] In a preferred embodiment of the present invention, determining the adaptive smoothing factor based on the current loop order includes: Obtain the maximum loop order obtained by the degradation-gated time series prediction model during the training phase; The normalized loop process is determined based on the current loop order and the maximum loop order; Based on the normalization cycle process and the preset minimum smoothing factor, maximum smoothing factor, and smoothing factor growth rate coefficient, the adaptive smoothing factor is determined.

[0014] Secondly, embodiments of the present invention also provide a battery health status assessment device based on multi-dimensional operating condition feature fusion, comprising: The data acquisition module is used to construct a single-cycle time-series data matrix for the current battery cycle stage based on the multi-channel sampling data of the battery collected during the current battery cycle stage. The feature extraction module is used to perform adaptive time series decomposition and reconstruction guided by degradation trend on the single-cycle time series data matrix based on the current cycle order indicating the current battery cycle stage and the historical battery health status of the previous battery cycle stage in the single-cycle time series data matrix, so as to obtain the reconstructed feature matrix. The degradation co-vector determination module is used to determine the degradation co-vector based on the channel statistics in the reconstructed feature matrix; The prediction module is used to input the reconstructed feature matrix and the degradation co-vector into the degradation-gated time-series prediction model to obtain the initial battery health state prediction value of the current battery cycle stage; wherein, the convolution channels and temporal attention weights in the degradation-gated time-series prediction model are dynamically modulated by the degradation co-vector; The constraint smoothing module is used to perform degradation constraint smoothing on the initial battery health state prediction value to obtain the battery health state prediction value.

[0015] 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 assessment method based on multi-dimensional operating condition feature fusion described in the first aspect above.

[0016] Fourthly, embodiments of the present invention also provide a computer-readable storage medium storing computer-executable instructions. When the computer-executable instructions are invoked and executed by a processor, the computer-executable instructions cause the processor to implement the battery health status assessment method based on multi-dimensional operating condition feature fusion described in the first aspect.

[0017] The embodiments of the present invention bring the following beneficial effects: This invention provides a battery health status assessment method based on multi-dimensional operating condition feature fusion. By constructing a standardized time-series data matrix, introducing time-series decomposition and reconstruction based on cyclic process adaptive adjustment, and extracting a collaborative vector representing the global degradation state, this method dynamically modulates the channel gating factor and temporal attention distribution of a deep separable convolutional network using this collaborative vector. This achieves end-to-end adaptive evolution of feature extraction and model inference as the battery ages. It effectively overcomes the fundamental defect of traditional methods where fixed preprocessing parameters and static network weights cannot simultaneously consider the sensitivity of early local details and the accuracy of late-stage macroscopic trend modeling. By processing the initial battery health status prediction value through amplitude pruning and exponential smoothing based on physical degradation laws, it not only significantly reduces the prediction error during the rapid capacity decay stage throughout the entire life cycle but also fundamentally suppresses abnormal jumps that violate the irreversible laws of electrochemistry. This results in an output sequence with high accuracy, strong smoothness, and physical interpretability, providing a highly reliable, plug-and-play online health status assessment and early warning capability for battery management systems.

[0018] 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.

[0019] 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

[0020] 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.

[0021] Figure 1 A flowchart illustrating a battery health status assessment method based on multi-dimensional operating condition feature fusion, provided for an embodiment of the present invention; Figure 2 A flowchart of another battery health status assessment method based on multi-dimensional operating condition feature fusion provided in an embodiment of the present invention; Figure 3 This is a schematic diagram illustrating the fusion characteristics and trends of internal resistance channels; Figure 4 This is a schematic diagram of the residual fluctuation of the internal resistance channel; Figure 5 A flowchart illustrating another battery health status assessment method based on multi-dimensional operating condition feature fusion provided in this embodiment of the invention; Figure 6 A comparison chart showing the original predicted battery health status and the final output value after degradation constraint smoothing; Figure 7 A flowchart illustrating another battery health status assessment method based on multi-dimensional operating condition feature fusion provided in an embodiment of the present invention; Figure 8 This is a schematic diagram of the attention weight distribution in the early recurrent temporal domain. Figure 9 This is a schematic diagram of the attention weight distribution in the late-cycle temporal domain. Figure 10 This is a schematic diagram of the distribution of the first-layer channel gating factors; Figure 11 This is a schematic diagram of the distribution of the gate factors for the second-layer channel; Figure 12 A schematic diagram of the structure of a battery health status assessment device based on multi-dimensional operating condition feature fusion provided in an embodiment of the present invention; Figure 13 This is a schematic diagram of the structure of an electronic device provided in an embodiment of the present invention. Detailed Implementation

[0022] 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.

[0023] During long-term cycling, battery capacity gradually declines. This decline is influenced by a combination of factors, including charge / discharge rate, temperature conditions, and cycle depth. Furthermore, the aging mechanisms and degradation susceptibility characteristics differ significantly across different battery cycle stages. In the early stages of cycling, capacity decline is relatively gradual, with degradation signals often hidden in localized voltage and current fluctuations. However, in the later stages, macroscopic trends such as increased internal resistance and prolonged constant-voltage charging time gradually become the dominant indicators of capacity decline. Accurately assessing the current health status of batteries from charge / discharge cycle monitoring data and adapting to the characteristic evolution patterns of different degradation stages is a pressing issue for battery management systems and predictive maintenance.

[0024] Traditional temporal decomposition or filtering methods often employ preset smoothing parameters or convolution scales, failing to adaptively adjust them according to the battery's cycling stage. This results in the loss of detailed information in early cycles or insufficient modeling of trends in later cycles, making it difficult to effectively capture the changing patterns of degradation features throughout the entire lifespan. Conventional temporal convolutional or recurrent neural networks use fixed network weights and channel response patterns for all cycle samples, failing to incorporate the current cycle's macroscopic degradation state as conditional information. This prevents dynamic adjustment of feature channel importance and temporal dependence based on the degree of capacity degradation, limiting their adaptability to different degradation rates.

[0025] Based on this, the present invention provides a battery health status assessment method based on multi-dimensional operating condition feature fusion, which can ensure that the internal parameters change adaptively when facing batteries with different aging levels. This overcomes the limitation that conventional fixed convolution response cannot cover the differences in degradation rate throughout the entire life cycle, and improves the accuracy of battery health status prediction.

[0026] To facilitate understanding of this embodiment, a battery health status assessment method based on multi-dimensional operating condition feature fusion, as disclosed in this embodiment of the invention, will first be described in detail.

[0027] Example 1 This invention provides a battery health status assessment method based on multi-dimensional operating condition feature fusion. Figure 1 This is a flowchart illustrating a battery health status assessment method based on multi-dimensional operating condition feature fusion, provided as an embodiment of the present invention. Figure 1As shown, the battery health status assessment method based on multi-dimensional operating condition feature fusion may include the following steps: Step S101: Based on the collected multi-channel sampling data of the battery during the current battery cycle stage, construct a single-cycle time-series data matrix for the current battery cycle stage.

[0028] A battery cycle stage refers to the completion of a full charge-discharge cycle (e.g., from fully charged to discharged, or from constant current to constant voltage until fully charged). During battery cycle testing, the battery will undergo multiple cycle tests, with each cycle corresponding to a battery cycle stage. As the number of cycle tests increases, the battery exhibits different degradation sensitivities in voltage, current, temperature, charging capacity, internal resistance, constant current charging time, and constant voltage charging time at different battery cycle stages. This invention organizes the multi-channel monitoring data collected during the current battery cycle stage into a fixed-length single-cycle time-series data matrix. The current battery cycle stage refers to the cycle test the battery is currently performing, simply referred to as the current cycle. Multi-channel acquired data refers to multiple physical quantities synchronously acquired on a unified time axis. In this invention embodiment, the multi-channel acquired data includes multiple physical quantities such as voltage, current, temperature, charging capacity, internal resistance, constant current charging time, constant voltage charging time, cycle number, and the battery health status of the previous cycle. It can be understood that one channel corresponds to one physical quantity, and these acquired physical quantities can describe the battery state from different dimensions such as electrochemical, thermal, and impedance.

[0029] Specifically, based on the collected multi-channel sampling data of the battery during the current battery cycle stage, a single-cycle time-series data matrix for the current battery cycle stage is constructed, including: collecting multiple physical quantity data of the battery during the current battery cycle stage according to a preset sampling frequency to obtain multi-channel sampling data; and resampling the multi-channel sampling data at equal intervals according to a preset sampling length to obtain a single-cycle time-series data matrix.

[0030] During battery cycle testing, voltage, current, temperature, charging capacity, internal resistance, constant current charging time, constant voltage charging time, cycle number, and battery health status from the previous cycle are collected at a preset sampling frequency. The sampling frequency can be determined based on the data output frequency of the battery testing equipment, for example, 1Hz, 2Hz, or 10Hz. Specifically, voltage characterizes the change in battery terminal voltage during the current cycle; current characterizes the change in charge / discharge rate during the current cycle; temperature characterizes the change in thermal state during the current cycle; charging capacity characterizes the cumulative charging capacity during the current cycle; internal resistance characterizes the battery polarization and impedance state during the current cycle; constant current charging time characterizes the duration of the constant current phase; constant voltage charging time characterizes the duration of the constant voltage phase; cycle number characterizes the position of the current battery cycle within the battery's lifespan, also known as the current cycle sequence; and battery health status from the previous cycle characterizes the battery's capacity retention level before the current cycle, also known as the historical battery health status from the previous cycle.

[0031] During battery cycle testing, if the sampling lengths are inconsistent across different battery cycle stages, a uniform sampling length can be preset. If the sampling length of the original multi-channel sampling data acquired in the current cycle differs from the preset sampling length, the data for each channel in the original multi-channel sampling data is resampled at equal intervals to obtain multi-channel sampling data that matches the preset sampling length. This ensures that the multi-channel sampling data obtained in each battery cycle stage is uniform. Each time step. Among them, This represents the total number of time steps for multi-channel sampled data, and can be 128, 256, or 512. In a convenient implementation method... Take 128; Equal-interval resampling can be achieved using linear interpolation, that is, first normalize the original time axis of the current loop to the range of 0 to 1, and then uniformly set the time axis within the range of 0 to 1. The system identifies several target sampling points and obtains the target sampling value based on the linear interpolation results of adjacent original sampling points. If the target sampling point is located outside the original sampling range, a boundary value filling method can be used, that is, the original sampling value closest to the boundary can be used for filling.

[0032] The data from the nine channels after resampling are concatenated column-wise to obtain the single-cycle time-series data matrix for the current loop. Single-cycle time series data matrix This refers to a two-dimensional matrix formed by arranging the nine monitoring channels within the current cycle according to a fixed time step, with a size of [missing information]. ,in, In this embodiment of the invention, the total number of channels is indicated. The nine channels are, in order: voltage channel, current channel, temperature channel, charging capacity channel, internal resistance channel, constant current charging time channel, constant voltage charging time channel, cycle number channel, and battery health status channel for the previous cycle. (Definition) Represents a single-cycle time series data matrix The Middle Time step The original sampled value of the channel, This represents the time step index, with a value range of 1 to... , This represents the channel index, with values ​​ranging from 1 to... .

[0033] It should be noted that the cycle number and the battery health status of the previous cycle can be set as constant channels within the current cycle. For example, the current cycle number is... At that time, the cyclic sequence number channel is in all The values ​​of each time step are all The battery health status of the previous cycle in the current cycle was: At that time, the battery health status channel of the previous cycle was in all The values ​​of each time step are all For the first cycle, since there is no previous cycle battery health status, the previous cycle battery health status can be set to the nominal battery health status. , When expressed as a percentage, 100 can be used.

[0034] Step S102: Based on the current cycle order indicating the current battery cycle stage and the historical battery health status of the previous battery cycle stage in the single-cycle time-series data matrix, the single-cycle time-series data matrix is ​​subjected to adaptive time-series decomposition and reconstruction guided by degradation trend to obtain the reconstructed feature matrix.

[0035] The current cycle sequence describes the number of battery cycle stages that the battery has completed, also known as the current cycle number. The historical battery health state of the previous battery cycle stage refers to the predicted battery health state obtained from the previous battery cycle stage. Different channels in the single-cycle time-series data matrix X have different physical dimensions and different degradation sensitivities. Voltage, current, and temperature reflect the current operating conditions more, while internal resistance, constant voltage charging time, and charging capacity are more sensitive to capacity degradation. The cycle number and the battery health state of the previous cycle can provide the degradation stage of the current cycle. If a uniform standardization method is directly applied to all channels, it is easy to weaken the response of degradation-sensitive channels and make it difficult to express the operating condition coupling relationship between channels.

[0036] First, standardization is performed on each channel to eliminate the influence of dimensions. Then, based on the current cycle order and the historical battery health status of the previous battery cycle stage, a "degradation sensitivity coefficient" is set for each channel as a channel degradation modulation coefficient. This coefficient is used to differentially scale the standardized single-cycle time-series data matrix, so that channels that are more sensitive to aging (such as internal resistance and constant voltage charging time) are numerically amplified, resulting in a fused standardized feature matrix.

[0037] The trend smoothing strength is determined using the cycle number. This strength monotonically increases with the number of cycles, meaning the older the battery, the stronger the smoothing. This trend smoothing strength is then used to recursively smooth the fused normalized feature matrix, extracting a trend feature matrix reflecting the long-term degradation trend. Simultaneously, the difference between the fused normalized feature matrix and the trend feature matrix is ​​calculated to obtain the fluctuation residual matrix. Finally, the trend feature matrix is ​​added to the fluctuation residual matrix to obtain the reconstructed feature matrix.

[0038] Step S103: Determine the degenerate collaborative vector based on the channel statistics in the reconstructed feature matrix.

[0039] Channel statistics refer to global statistical indicators extracted from each channel in the reconstructed feature matrix along the time dimension. These mainly include the channel mean, standard deviation, and time regression slope, used to characterize the overall characteristic level and trend of the battery within the current cycle. For example, the mean, standard deviation, and time regression slope can be extracted for each channel in the reconstructed feature matrix. Simultaneously, cycle-level scalars with clear physical meaning can be extracted from the reconstructed feature matrix, including the total constant-current / constant-voltage charging time, mean internal resistance, historical battery health status from the previous battery cycle stage, and normalized cycle progress. The channel statistics are concatenated with the cycle-level scalars to form a high-dimensional cycle-level statistical feature vector. To eliminate the influence of dimensions, the cycle-level statistical feature vector is typically standardized and subjected to principal component analysis (PCA) for dimensionality reduction, ultimately compressed to a stable numerical range, resulting in a low-dimensional, dense degradation co-vector.

[0040] Step S104: Input the reconstructed feature matrix and the degradation cooperative vector into the degradation gating time-series prediction model to obtain the initial battery health state prediction value of the current battery cycle stage.

[0041] In the degradation-gated temporal prediction model, the convolutional channels and temporal attention weights are dynamically modulated by the degradation co-vector. The front end of the degradation-gated temporal prediction model employs degradation-gated depthwise separable temporal convolutional encoding. The reconstructed feature matrix, as input, is first processed through multiple layers of dilated depthwise separable convolutions. These dilated convolutions exponentially expand the receptive field without increasing the number of parameters by increasing the dilation rate, capturing long-range temporal dependencies. Pointwise convolutions (1×1 convolutions) are responsible for fusing information from different channels. Degradation co-vectors are used to generate channel gating factors through fully connected layers. These factors are multiplied element-wise with the output of each convolutional layer, achieving dynamic filtering and enhancement of local features based on the global degradation state. After residual alignment mapping, degradation-gated temporal features are obtained.

[0042] The backend of the degradation-gated time-series prediction model employs temporal attention pooling and regression. The degradation co-vector is input into the attention query projection layer to generate a degradation conditional query vector. The similarity between this query and the features (Keys) at each time step is calculated, and then normalized using Softmax to obtain temporal attention weights. These weights are then used to perform a weighted summation of the features at each time step, aggregating a temporal context vector. Finally, the temporal context vector and the degradation co-vector are concatenated and input into a fully connected regression network to output the initial battery health state prediction.

[0043] Step S105: Perform degradation constraint smoothing on the initial battery health state prediction value to obtain the battery health state prediction value.

[0044] Degradation constraint smoothing is used to reduce prediction jumps caused by single-cycle measurement noise and ensure that the online output conforms to the slow degradation pattern of battery capacity. According to electrochemical principles, battery capacity decays irreversibly; even with a slight capacity rebound, it should not increase significantly. Therefore, degradation constraint smoothing sets the maximum increase (e.g., 0.3%) and maximum decrease (e.g., 2.0%) of battery capacity in each battery cycle. If the initial battery health state prediction suddenly increases by 5% compared to the historical battery health state of the previous battery cycle, degradation constraint smoothing will force it to "pull back" to the allowable increase range. The battery health state prediction after this round of constraints is weighted and fused with the historical battery health state of the previous battery cycle, filtering out high-frequency random noise to obtain the final battery health state prediction. The weight values ​​used in the weighted summation can be set according to actual conditions.

[0045] This invention provides a battery health status assessment method based on multi-dimensional operating condition feature fusion. By constructing a standardized time-series data matrix, introducing time-series decomposition and reconstruction based on cyclic process adaptive adjustment, and extracting a collaborative vector representing the global degradation state, this method dynamically modulates the channel gating factor and temporal attention distribution of a deep separable convolutional network using this collaborative vector. This achieves end-to-end adaptive evolution of feature extraction and model inference as the battery ages. It effectively overcomes the fundamental defect of traditional methods where fixed preprocessing parameters and static network weights cannot simultaneously consider the sensitivity of early local details and the accuracy of late-stage macroscopic trend modeling. By processing the initial battery health status prediction value through amplitude pruning and exponential smoothing based on physical degradation laws, it not only significantly reduces the prediction error during the rapid capacity decay stage throughout the entire life cycle but also fundamentally suppresses abnormal jumps that violate the irreversible laws of electrochemistry. This results in an output sequence with high accuracy, strong smoothness, and physical interpretability, providing a highly reliable, plug-and-play online health status assessment and early warning capability for battery management systems.

[0046] Example 2 This invention also provides another battery health status assessment method based on multi-dimensional operating condition feature fusion. This method is implemented on the basis of the method in the above embodiments. The method focuses on describing the specific implementation of adaptive temporal decomposition and reconstruction of the single-cycle time-series data matrix guided by degradation trend based on the current cycle order indicating the current battery cycle stage and the historical battery health status of the previous battery cycle stage in the single-cycle time-series data matrix, to obtain the reconstructed feature matrix.

[0047] Figure 2 A flowchart of another battery health status assessment method based on multi-dimensional operating condition feature fusion provided in this embodiment of the invention is shown below. Figure 2 As shown, the battery health status assessment method based on multi-dimensional operating condition feature fusion may include the following steps: Step S201: Based on the collected multi-channel sampling data of the battery during the current battery cycle stage, construct a single-cycle time-series data matrix for the current battery cycle stage.

[0048] Step S202: Obtain the channel mutual information coupling weights obtained during the training phase of the degradation gating time series prediction model.

[0049] Before training the degradation-gated time series prediction model, based on the training sample set The channel mutual information coupling weights are calculated offline. These weights characterize the dependence strength between different physical channels related to the degradation process and are not used as neural network parameter updates during model training.

[0050] In this embodiment of the invention, a training sample set is composed of multiple single-cycle samples and their actual battery health status labels. Training sample set This refers to a dataset consisting of multiple labeled, cyclically generated samples, represented as... .in, Indicates the training sample index. This represents the total number of training samples. Indicates the first The single-cycle time-series data matrix corresponding to each training sample Indicates the first Each training sample corresponds to a real battery health state label. A single-cycle sample refers to multi-channel sampling data collected during historical cycle stages between the current battery cycle stages. Real battery health state labels are constructed based on the battery health state test values ​​from historical cycle stages. Real battery health status label This refers to the capacity maintenance level during this historical cycling phase, used to supervise the training of the battery health state prediction model. In practical implementation, it can be based on the measured discharge capacity during the historical cycling phase. and rated capacity Calculate the true battery health status label, i.e. .in, This represents the measured discharge capacity during the historical cycle phase. Indicates the battery's rated capacity. This label represents the actual battery health status during historical cycle stages, expressed as a percentage. If the testing system uses the initial capacity as the health status benchmark, it can also be... Replace with the initial stable cycle capacity , This represents the average discharge capacity of the battery during the initial stable cycles.

[0051] It should be noted that the training sample set The training should cover different battery cycle stages, temperature conditions, charge / discharge rates, and degradation levels. For example, training samples could include early, mid, and late cycles, as well as normal temperature (25°C), high temperature, and low temperature conditions. Multi-condition coverage reduces the risk of the model only remembering amplitude differences under a specific condition, allowing subsequent models to learn stable characteristics related to battery degradation mechanisms.

[0052] In one implementation, the training sample set is first... Cyclic-level channel statistics are extracted for each channel of each single-cycle sample, forming a cyclic-level channel statistics sequence. These statistics include the channel mean, channel standard deviation, and channel time regression slope. The channel time regression slope is obtained by performing a least-squares linear fit between the time step index and the channel's sampled values, characterizing the overall trend of the channel's change over time within a single cycle (one battery cycle). Then, for the target channel... Non-self channel , training sample set Middle target channel Cyclic channel statistics sequence and non-self channels Mutual information estimation is performed on the cyclic channel statistics sequence to obtain the mutual information intensity. , Characterizing target channel and non-self channels The strength of the degradation-related dependency between them The larger the value, the more likely it is to be a non-self channel. For the target channel It provides more comprehensive supplementary information.

[0053] In this implementation, mutual information estimation can be achieved using binning statistics. Specifically, the target channel... The cyclic channel statistics sequence is divided into several intervals, and non-self channels are excluded. The cyclic channel statistics sequence is divided into several intervals. The frequency of occurrence of joint intervals and univariate intervals is statistically analyzed, and the mutual information intensity is calculated from the joint probability and marginal probability. The number of bins can be between 10 and 30. A larger number of bins can be used when the sample size is large, and a smaller number of bins can be used when the sample size is small, in order to avoid the probability estimation being too sparse.

[0054] Furthermore, for each target channel The mutual information strengths of all non-self channels are normalized to obtain the channel mutual information coupling weights. , is represented as: ; in, Indicates the first Channel to the first The normalized mutual information coupling weights of the channels are used to control the first... The passage leads to the first The strength of the channel injected information; Indicates the exception to the target channel Other channel indexes; This represents a very small constant to prevent the denominator from being zero; for example, it can be 0.0001. Indicates the target channel The sum of mutual information strengths with all non-self channels. If the target channel All corresponding mutual information strengths are close to 0, which allows the target channel to be... The corresponding channel mutual information coupling weights are set to a uniform distribution to avoid numerical instability.

[0055] Step S203: Based on the single-cycle time-series data matrix indicating the current cycle order of the current battery cycle stage, the historical battery health status of the previous battery cycle stage, and the channel mutual information coupling weight, the single-cycle time-series data matrix is ​​standardized to obtain a fused standardized feature matrix.

[0056] For each channel in the single-cycle time-series data matrix, Z-score normalization is performed individually to obtain a basic normalized value for each channel. Based on the current cycle order (e.g., cycle 800) and historical battery health status (e.g., 80%), a degradation modulation coefficient is determined. The influence weight of the degradation modulation coefficient varies for different channels. For each channel, the basic normalized value of non-channels is multiplied by the mutual information coupling weight, and then added back to the basic normalized value of that channel to obtain the fused normalized feature matrix.

[0057] Specifically, based on the indications in the single-cycle time-series data matrix indicating the current cycle order of the current battery cycle stage, the historical battery health status of the previous battery cycle stage, and the channel mutual information coupling weights, the single-cycle time-series data matrix is ​​standardized to obtain a fused standardized feature matrix. This includes: performing standardization on each channel in the single-cycle time-series data matrix to obtain a basic standardized feature matrix; determining the channel degradation modulation coefficient corresponding to each channel in the single-cycle time-series data matrix based on the indications in the current cycle order of the current battery cycle stage and the historical battery health status of the previous battery cycle stage; performing differential enhancement on the basic standardized feature matrix based on the channel degradation modulation coefficients to obtain a degradation-enhanced standardized feature matrix; and fusing the basic standardized feature matrix and the degradation-enhanced standardized feature matrix based on the channel mutual information coupling weights to obtain a fused standardized feature matrix.

[0058] From a single-cycle time series data matrix The sampled values ​​of each of the nine channels are read, and the within-sample mean and within-sample standard deviation are calculated for each channel. For the first... Channel, in-sample mean By the The passage is in all The average of the sampled values ​​at each time step is used to obtain the within-sample standard deviation. By the The passage is in all The values ​​are calculated from the sampled values ​​at each time step. Indicates the current loop iteration number. The overall level of the channel is used to perform centralized processing; Indicates the current loop iteration number. The fluctuation intensity of the channel is used to perform scale normalization.

[0059] In specific implementation, for the first Time step raw sampled value of the channel By performing centralization and scale normalization, the basic standardized eigenvalues ​​are obtained. Specifically, this is achieved through Z-score standardization, which ensures that the data for all features are distributed around the value of 0, and the standard deviation becomes 1, thereby eliminating the influence of different physical units and making them numerically comparable.

[0060] Through this processing, the sampled values ​​of different dimensions such as voltage channel, current channel, temperature channel, and internal resistance channel are converted to a comparable numerical range, providing a unified input basis for subsequent degradation modulation and channel coupling.

[0061] Read the current cycle number from the cycle number channel. Read the battery health status from the previous cycle's battery health status channel. Wherein, the current loop number Used to characterize the cycling progress of the current single-cycle sample in the battery life cycle, and the battery health status of the previous cycle. This characterizes the capacity preservation level prior to the current loop. Let the maximum loop number in the training set be... , From the training sample set The cycle number is obtained by statistically analyzing all single-cycle samples and used to normalize the current cycle number.

[0062] In the specific implementation, first through Calculate the normalized loop process , This refers to the normalized loop progress of the current loop. The larger the value, the closer the current loop is to a later loop in the training set. Then, based on the nominal battery health state... Battery health status compared to the previous cycle Calculate the difference in normalized degradation stages , is represented as: ; in, This indicates the degree of degradation of the battery relative to its nominal state of health in the current cycle. The larger the value, the more significant the battery capacity degradation.

[0063] Furthermore, a degradation sensitivity coefficient is set for each channel. , Characterizing the first The strength of the channel's response to changes in the degradation phase is used to control the first... The magnitude of enhancement as the channel deepens during the degradation phase. For internal resistance channels and constant voltage charging time channels, A larger value can be taken; for temperature channels and current channels, A smaller value can be taken. Then, based on the difference in the normalized degradation stage... and normalization cycle process Generation channel degradation modulation coefficient , is represented as: ; in, Indicates the first The channel degradation modulation coefficients of the channel are used to modify the first channel. The basic standardized eigenvalues ​​of the channel are enhanced during the degradation phase. This represents the degradation modulation steepness coefficient, used to control the growth rate of the channel degradation modulation coefficient when the battery health status declines in the previous cycle; it can be set to 3. This represents the hyperbolic tangent function, used to limit the degradation modulation amplitude within a stable range and avoid excessive amplification of eigenvalues.

[0064] In one implementation, the degradation sensitivity coefficient of the internal resistance channel A value of 0.80 can be used for the degradation sensitivity coefficient of the constant voltage charging time channel. A value of 0.70 can be used for the degradation sensitivity coefficient of the charging capacity channel. A value of 0.50 can be used for the degradation sensitivity coefficient of the temperature channel. A value of 0.05 is acceptable. This setting indicates that the internal resistance, constant voltage charging time, and charging capacity have a stronger response to battery degradation, while temperature is more susceptible to environmental conditions and should not be excessively amplified.

[0065] Based on channel degradation modulation coefficient For basic standardized eigenvalues Differential enhancement is performed to obtain the degradation-enhancement standardized eigenvalues. In practical implementation, it can be made ,in, Indicates the first Time step The degradation enhancement normalization eigenvalues ​​of the channel are used to highlight the channel response that is more sensitive to the current degradation stage.

[0066] It should be noted that this operation normalizes the loop process in the early stages of the loop. Smaller, channel degradation modulation coefficient Approaching 1, each channel mainly maintains the basic standardized results; the normalization cycle process is performed in the late cycle. The normalized characteristics of the internal resistance channel, constant voltage charging time channel, and charging capacity channel are moderately enhanced, which improves the model's ability to sense capacity degradation signals. The battery health status of the previous cycle is more significantly reduced.

[0067] The basic normalized eigenvalues ​​of non-self channels are injected into the target channel according to the channel mutual information coupling weights to obtain the fused normalized feature matrix. Fusion of standardized feature matrix This refers to the single-cycle temporal feature matrix after differential degradation enhancement and mutual information injection, with a size of [missing information]. Furthermore, the definition Represents the fusion normalized feature matrix The Middle Time step The fusion-normalized eigenvalues ​​of the channels are calculated as follows: ; in, Indicates the first The mutual information injection strength of the channels is used to control the injection of information from other channels into the third channel. The amplitude of the channel can be between 0.05 and 0.20; Indicates a non-self channel index; Indicates the first Time step The basic standardized characteristic values ​​of the channel.

[0068] It should be noted that when the target channel is the internal resistance channel, if the mutual information strength between the constant voltage charging time channel and the charging capacity channel and the internal resistance channel is large as calculated during the training phase, then the constant voltage charging time channel and the charging capacity channel will be injected into the internal resistance channel with a larger channel mutual information coupling weight. This ensures that the internal resistance channel not only retains its own degradation enhancement information but also includes charging process information related to capacity degradation. This processing reduces the fragmentation of degradation information caused by isolated processing of single channels and improves the ability of subsequent time-series models to express the battery capacity degradation mechanism.

[0069] It should also be noted that the channel mutual information coupling weights are determined during the training phase based on the training sample set. The channel mutual information coupling weights, which are saved during the training phase, are used in the offline calculation, verification phase, and online prediction phase. They are not recalculated based on the samples to be predicted, so as to ensure that the feature processing methods are consistent across different phases.

[0070] Step S204: Based on the current cycle order and the historical battery health status, perform adaptive temporal decomposition and reconstruction of the fused standardized feature matrix guided by degradation trend to obtain the reconstructed feature matrix.

[0071] Based on the current loop order, an adaptive smoothing factor is calculated using a preset exponential function. Recursive smoothing is then performed on the fused standardized feature matrix to obtain the trend feature matrix. The difference between the fused standardized feature matrix and the trend feature matrix is ​​used as the fluctuation residual matrix. The fluctuation residual matrix is ​​selectively enhanced using preset residual enhancement coefficients, and the enhancement result is added to the trend feature matrix to obtain the reconstructed feature matrix.

[0072] Specifically, the fused normalized feature matrix is ​​subjected to adaptive temporal decomposition and reconstruction guided by degradation trends based on the current cycle order and the historical battery health status to obtain a reconstructed feature matrix. This includes: determining an adaptive smoothing factor based on the current cycle order; recursively smoothing the fused normalized feature matrix based on the adaptive smoothing factor to obtain a trend feature matrix; determining a fluctuation residual matrix based on the fused normalized feature matrix and the trend feature matrix; and determining the reconstructed feature matrix based on the fluctuation residual matrix and the trend feature matrix.

[0073] Furthermore, determining the adaptive smoothing factor based on the current loop order includes: obtaining the maximum loop order obtained by the degenerate gated time series prediction model during the training phase; determining the normalized loop process based on the current loop order and the maximum loop order; and determining the adaptive smoothing factor based on the normalized loop process, as well as preset minimum smoothing factor, maximum smoothing factor, and smoothing factor growth rate coefficient.

[0074] Based on the current loop number And the maximum cycle order of the training set (also known as the maximum cycle number of the training set). Calculate the normalization loop process And according to the normalization loop process Generate adaptive smoothing factor Adaptive smoothing factor Used to control the degree to which trend components retain information from historical time steps; The larger the value, the smoother the trend extraction, and the more obviously short-term fluctuations are suppressed; The smaller the value, the closer the trend extraction is to the current time step input, and the more local details are preserved.

[0075] In practical implementation, it is based on a normalized loop process. Preset minimum smoothing factor Maximum smoothing factor and the growth rate coefficient of the smoothing factor Calculate the adaptive smoothing factor , is represented as: ; in, This represents the minimum value of the adaptive smoothing factor, used to constrain the minimum smoothing intensity of the early cycles, and can be set to 0.60. This represents the maximum value of the adaptive smoothing factor, used to constrain the highest smoothing intensity of late-cycle cycles; it can be set to 0.95. This represents the growth rate coefficient of the smoothing factor, which is used to control the rate at which the adaptive smoothing factor increases with the cycle process, and can be taken as 2. This represents the natural exponential function.

[0076] It should be noted that the battery degradation is lower in the early stages of cycling, and the adaptive smoothing factor... near The reconstructed features can retain more local details; the battery degradation is higher in late cycles, and the adaptive smoothing factor... Approaching Reconstructing features can yield a more stable expression of trends.

[0077] With fusion of standardized feature matrices As input, recursive smoothing is performed on each channel to obtain the trend feature matrix. Trend feature matrix This refers to the fusion of standardized feature matrices. The single-cycle trend matrix obtained after performing adaptive recursive smoothing has a size of . Trend feature matrix The Middle Time step The trend characteristic value of the channel is used to characterize the first... The channel exhibits a relatively stable temporal variation trend within the current single cycle.

[0078] In the specific implementation, first the first The recursive initial value of the channel is set to Then, starting from the second time step, the calculation is performed recursively to achieve the fusion and standardization of the feature matrix. The channel-wise, time-step-wise adaptive recursive smoothing is represented as: ; in, Trend feature matrix The Middle Time step The trend characteristic value of the channel; Trend feature matrix The Middle Time step The channel's trend characteristic value is used to convey historical smoothing information to the current time step.

[0079] It should be noted that, through recursive smoothing, the standardized feature matrix is ​​fused. High-frequency disturbances are moderately suppressed, while stable trends with degenerative significance, such as slow voltage changes, capacity changes, internal resistance changes, and constant voltage charging time changes, are preserved, making subsequent temporal convolutional coding less susceptible to interference from accidental sampling fluctuations.

[0080] Calculate the fusion normalized feature matrix With trend feature matrix The difference between them yields the fluctuation residual matrix. Fluctuation residual matrix This refers to the fusion of standardized feature matrices. Relative to trend feature matrix The short-term deviation matrix, with size . .definition Indicates the first Time step The channel's fluctuation residual value is used to characterize the short-term dynamic changes of the current time step relative to the smooth trend, including local voltage changes, local current changes, temperature disturbances, and aging-sensitive abnormal fluctuations.

[0081] In practical implementation, the fluctuation residual value It is obtained by calculating the difference between the fused standardized eigenvalues ​​and the corresponding trend eigenvalues, i.e. .

[0082] Furthermore, based on the difference in normalized degradation stages Generate residual enhancement coefficient Residual enhancement coefficient Used to control the fluctuation residual matrix The contribution strength in the reconstructed feature matrix is ​​enhanced by boosting the base amplitude through fluctuation residuals. and degradation bias modulation slope Perform calculations, that is .in, This represents the basic magnitude of the fluctuation residual enhancement, used to control the maximum contribution intensity of short-term operating condition fluctuations in the reconstructed characteristics, and can be taken as 0.50; This represents the degradation bias modulation slope, used to control the rate of increase in the amplitude of the fluctuation residual when the battery health status declines in the previous cycle. It can be set to 4.

[0083] Trend feature matrix With residual enhancement coefficient Adjusted fluctuation residual matrix Add them together to obtain the reconstructed feature matrix. Reconstruct the feature matrix This refers to the single-cycle time-series feature matrix after degradation trend smoothing and degradation-sensitive fluctuation enhancement, with a size of [missing information]. It is used as input to the degradation-gated time series prediction model.

[0084] in, Represents the reconstructed feature matrix The Middle Time step The reconstructed eigenvalues ​​of the channel are obtained by using the trend feature matrix. With residual enhancement coefficient Modulated wave residual matrix To perform the calculation by addition, that is... .

[0085] In one embodiment, if the current cycle number of an earlier loop is small, then the adaptive smoothing factor... The residual enhancement factor is close to 0.60. Relatively small, reconstructing the feature matrix It retains more local details without excessively amplifying fluctuations; if the current cycle number of a late cycle is large and the battery health status of the previous cycle has significantly declined, then the adaptive smoothing factor... Approaching 0.95, trend characteristic matrix Smoother, while the residual enhancement factor Abnormal fluctuations during the growth and rapid capacity decay phases can be used to reconstruct the feature matrix. The middle part is highlighted.

[0086] It should be noted that in recursive smoothing or subsequent convolution calculations, when the time index exceeds 1... When the range is large, boundary value padding can be used, that is, indices less than 1 use the value of the first time step, and indices greater than 1 use the value of the first time step. The index uses the first The value of the time step is used to keep the output timing length constant.

[0087] In an implementable example, we analyze an example of internal resistance channel timing decomposition. Figure 3 This diagram illustrates the fusion and trend characteristics of the internal resistance channel. It plots the fusion-standardized characteristic values ​​of the internal resistance channel and the trend characteristic values ​​obtained through recursive smoothing. The horizontal axis represents the time step (dimensionless), and the vertical axis represents the characteristic value (dimensionless). Figure 4This diagram illustrates the residual fluctuations within the internal resistance channel, showing the corresponding residual values. The horizontal axis represents the time step (dimensionless), and the vertical axis represents the residual fluctuation value (dimensionless). The adaptive time series decomposition demonstrates the ability of a single-cycle internal dynamics to be decomposed into stable trends and short-term fluctuations. It proves that trend characteristics can reflect slow changes with degenerative significance, while residual characteristics retain local operating condition disturbances and aging-sensitive fluctuations.

[0088] Step S205: Determine the degradation collaboration vector based on the channel statistics in the reconstructed feature matrix.

[0089] Step S206: Input the reconstructed feature matrix and the degradation cooperative vector into the degradation gating time-series prediction model to obtain the initial battery health state prediction value of the current battery cycle stage.

[0090] Step S207: Perform degradation constraint smoothing on the initial battery health state prediction value to obtain the battery health state prediction value.

[0091] The battery health status assessment method based on multi-dimensional operating condition feature fusion provided in this invention proposes a differentiated channel standardization and mutual information enhancement method. It generates channel degradation modulation coefficients based on the cycle number and the health status of the previous cycle, selectively enhancing degradation-sensitive channels. Simultaneously, by offline calculation of the mutual information coupling weights between channels, operating condition-related channel information is injected into the target channel, eliminating differences in physical dimensions while preserving degradation-sensitive responses, overcoming the shortcomings of traditional unified standardization in weakening key degradation signals. A smoothing factor that adaptively adjusts with the cycle progress is introduced to perform recursive smoothing on single-cycle time-series data to separate stable trends from short-term fluctuations. Furthermore, the fluctuation residuals are selectively enhanced based on the difference in degradation stages, allowing reconstructed features to retain local details in the early stages and highlight abnormal fluctuations in rapid capacity decay in the later stages, solving the problem that fixed smoothing strategies cannot simultaneously meet the information needs of different degradation stages.

[0092] Example 3 This invention also provides another battery health status assessment method based on multi-dimensional operating condition feature fusion; this method is implemented on the basis of the method in the above embodiments; the method focuses on describing the specific implementation of determining the degradation cooperative vector according to the channel statistics in the reconstructed feature matrix, and performing degradation constraint smoothing on the initial battery health status prediction value to obtain the battery health status prediction value.

[0093] Figure 5 A flowchart of another battery health status assessment method based on multi-dimensional operating condition feature fusion provided by an embodiment of the present invention is shown below. Figure 5 As shown, the battery health status assessment method based on multi-dimensional operating condition feature fusion may include the following steps: Step S301: Based on the collected multi-channel sampling data of the battery during the current battery cycle stage, construct a single-cycle time-series data matrix for the current battery cycle stage.

[0094] Step S302: Based on the current cycle order indicating the current battery cycle stage and the historical battery health status of the previous battery cycle stage in the single-cycle time-series data matrix, the single-cycle time-series data matrix is ​​subjected to adaptive time-series decomposition and reconstruction guided by degradation trend to obtain the reconstructed feature matrix.

[0095] Step S303: Obtain the channel statistics corresponding to the reconstructed feature matrix and the cyclic scalar corresponding to the single-cycle time series data matrix.

[0096] For reconstructing the feature matrix For each channel, extract the channel mean, channel standard deviation, and channel time regression slope. For the ... Channel, channel mean from to The average is obtained and used to characterize the first... The overall response level of the channel in the current loop; the channel standard deviation is determined by... to Calculated, used to characterize the The channel's volatility in the current cycle; the channel's time regression slope from time step index 1 to... With the The reconstructed eigenvalues ​​of the channels are obtained by least-squares linear fitting, and are used to characterize the ... The overall upward or downward trend of the channel in the current cycle.

[0097] In practical implementation, least squares linear fitting can use the time step index as the independent variable, and the 1st... The reconstructed eigenvalues ​​of the channel are used as the dependent variable to fit a straight line, and the slope of the line is the channel time regression slope. A positive channel time regression slope indicates that the channel generally increases within a single cycle, while a negative channel time regression slope indicates that the channel generally decreases within a single cycle.

[0098] From a single-cycle time series data matrix Extract the cyclic scalars with clear physical meaning, and then combine the cyclic scalars with the reconstructed feature matrix. The channel statistics are concatenated to obtain the cyclic statistical feature vector. Cyclic statistical eigenvectors It refers to a one-dimensional feature vector that centrally expresses the current macroscopic operating condition and degradation state of the cycle.

[0099] In one implementation, the cyclic statistical feature vector This includes the mean and standard deviation of each channel, the time regression slope of each channel, the total constant current charging time, the total constant voltage charging time, the mean internal resistance, the battery health status of the previous cycle, and the normalized cycle progress. The total constant current charging time is derived from the single-cycle time-series data matrix. The constant current charging time channel is in the first stage. The sampling values ​​at each time step are obtained; the total constant voltage charging time is determined by the single-cycle time-series data matrix. The constant voltage charging time channel is in the first stage. The sampling values ​​at each time step are obtained; the mean internal resistance is obtained from the single-cycle time-series data matrix. The average of the sampled values ​​of the medium-impedance channel over all time steps is obtained; the normalized loop process is determined by the current loop number. Divide by the maximum cycle number of the training set get.

[0100] In one embodiment, as an example, suppose the current sample being processed is a battery sample in the late cycle, whose reconstructed feature matrix... Dimensions ,from The extracted channel statistics, taking the internal resistance channel (assumed to be the 5th channel) as an example: Channel mean: For internal resistance channels across all Reconstructed feature values ​​at each time step to Calculate the average; if the result is 0.85, then this value represents the overall response level of the internal resistance of the cycle. Channel standard deviation: Calculate the standard deviation of all reconstructed eigenvalues ​​of the internal resistance channel. If the result is 0.12, this value represents the degree of fluctuation of the internal resistance of the cycle. Channel time regression slope: indexed by time step to Using the internal resistance channel reconstruction feature value as the independent variable and the corresponding internal resistance channel reconstruction feature value as the dependent variable, a least squares linear fit is performed. If the slope is 0.01, it indicates that the internal resistance generally shows an upward trend within the cycle. From a single-cycle time series data matrix The extracted cycle-level scalars, taking the same late-cycle sample as an example: Total constant current charging time: from matrix In the constant current charging time channel, take the first... The sampled value at each time step; assuming that the channel records the real-time cumulative time, the value in the last column is the total time. If it is 1600 seconds, then this value is 1600. Total constant voltage charging time: Similar to the method for obtaining the total constant current charging time, the time is taken from the constant voltage charging time channel. The sampled value at each time step; if it is 800 seconds, then this value is 800. Mean internal resistance: from the matrix In the internal resistance channel, the sampled values ​​of all time steps are directly averaged. If the value is 0.045 ohms, then this value is 0.045. Previous cycle battery health status: Read from any position in the previous cycle battery health status channel (constant channel), assuming it is... ; Normalized loop process: Assuming the current loop number Maximum cycle number of the training set Then the normalized loop process .

[0101] Step S304: Concatenate the cyclic scalar with the channel statistics to obtain the cyclic statistical feature vector.

[0102] Concatenating all cyclic scalars and channel statistics forms the cyclic statistical feature vector of the sample. It should be noted that the optimal values ​​for the total constant current charging time, the total constant voltage charging time, and the average internal resistance are derived from the single-cycle timing data matrix. Extracted from the matrix to preserve the direct meaning of the corresponding physical quantities; the mean, standard deviation, and time regression slope of each channel are preferably derived from the reconstructed feature matrix. Extracted from [the data] to reflect the temporal expression after the reconstruction of the degradation trend.

[0103] Step S305: Obtain the mean vector and standard deviation vector of the cyclic statistical features obtained during the training phase of the degradation-gated time series prediction model.

[0104] Based on training sample set Calculate the mean vector of the cyclic statistical features of all single-cycle samples in the training set. and the standard deviation vector of cyclic statistical features in the training set Training set cyclic statistical feature mean vector Used for cyclic statistical feature vectors Centering is performed on the standard deviation vector of the cyclic statistical features of the training set. Used for cyclic statistical feature vectors Scale normalization is performed to avoid the dimensional differences of statistical features at different cyclic levels affecting subsequent projections.

[0105] Step S306: Determine the degradation collaboration vector based on the mean vector of the cyclic statistical features, the standard deviation vector of the cyclic statistical features, and the cyclic statistical feature vector.

[0106] Furthermore, for the training sample set Principal component analysis is performed on the cyclic statistical eigenvectors after centering and scaling to obtain the principal component projection matrix. Principal component projection matrix The number of rows is determined and fixed offline during the training data preprocessing stage. The number of columns equals the number of cyclic statistical eigenvectors. Dimensions. The dimension of the degradation co-vector is used to control the representation capacity of global degradation information, and can be 16 or 32.

[0107] Specifically, the loop-level statistical feature vector of the current loop. According to the mean vector of cyclic statistical features of the training set and the standard deviation vector of cyclic statistical features in the training set Standardize, then project using principal component matrix Projecting and compressing the numerical range using the hyperbolic tangent function yields the degenerate co-vector. ,Right now .in, Represents the degenerate collaborative vector, with dimension . , is used to characterize the global degradation stage of the current loop and serves as the conditional input for degradation-gated temporal convolutional coding and degradation-conditional temporal attention pooling; This indicates that element-wise standardization is performed on the cyclic statistical feature vector; This represents the hyperbolic tangent function, used to compress the projection result to a stable numerical range.

[0108] It should be noted that the principal component projection matrix Training set cyclic statistical feature mean vector and the standard deviation vector of cyclic statistical features in the training set All of these are determined during the training data preprocessing stage and remain unchanged during neural network training, thus preventing the degradation of the co-vectors. It can provide stable cyclic prior information, avoiding excessive perturbation of global degradation information during the end-to-end training process.

[0109] Step S307: Input the reconstructed feature matrix and the degradation cooperative vector into the degradation gating time-series prediction model to obtain the initial battery health state prediction value of the current battery cycle stage.

[0110] Step S308: Obtain the historical battery health status prediction value of the previous battery cycle stage.

[0111] Specifically, the historical battery health state prediction values ​​from the previous battery cycle stage can be retrieved from the database. These historical battery health state prediction values ​​refer to the battery health state prediction values ​​obtained during battery cycles prior to the current battery cycle stage. These historical battery health state prediction values ​​are obtained by smoothing the output of the degradation-gated time-series prediction model after applying degradation constraints.

[0112] Step S309: Constrain the change range of the initial battery health state prediction value based on the historical battery health state prediction value to obtain the constrained prediction value.

[0113] Specifically, assuming the historical battery health state prediction value read from the previous battery cycle stage, i.e., the online battery health state prediction value of the previous battery cycle is... The maximum allowable increase in a single cycle is The maximum allowable decrease in a single cycle is .in, To allow for small capacity recovery or measurement errors, a range of 0.2 to 0.5 percentage points can be used; To limit abnormal decline in single-cycle predictions, a range of 0.5 to 2.0 percentage points can be selected, depending on the battery type and cycling conditions. Therefore, based on the upper and lower limits of the single-cycle prediction variation range constraints, the constrained prediction value is obtained. This operation first converts the original predicted value Limit to the predicted value of the previous battery cycle Subtract the maximum allowable decrease Within the range, then take the result and... Plus the maximum allowable increase The smaller of the two is represented as: ; in, This represents the predicted value after a single-cycle variation range constraint, i.e., the constrained predicted value. Describes the minimum value function; This represents the maximum value function.

[0114] Step S310: Perform exponential smoothing on the constraint prediction value and the historical battery health state prediction value to obtain the battery health state prediction value.

[0115] The predicted battery health status, i.e., the online battery health status prediction, is obtained by using exponential smoothing. , is represented as: ; in, This represents the final online battery health status prediction value output by the current loop. This represents the online smoothing coefficient, used to control the fusion ratio between the current prediction result and the output result of the previous battery cycle, and can be taken as 0.3 to 0.7.

[0116] In one embodiment, if the online battery health status prediction value of the previous battery cycle Original battery health status prediction value Maximum allowable increase in a single cycle If the initial predicted value is 92.3, then the constrained predicted value will not exceed 92.3; if the original predicted battery health value is... Maximum allowable descent in a single cycle If this condition is met, the constrained predicted value will not be lower than 91.0. This processing can reduce online prediction jumps caused by abnormal sampling or short-term operating condition disturbances.

[0117] Output online battery health status prediction value And online battery health status prediction values Save the battery health status as the previous cycle's value for the next cycle. (If the online battery health status prediction value...) Below the preset health status alarm threshold If this occurs, a battery health status alarm will be output. This indicates the battery health status alarm threshold, which can be set according to the application scenario, for example, 80.

[0118] In one embodiment, the original battery health state prediction values ​​for 50 consecutive online cycles are compared with the final output values ​​after degradation constraint smoothing in the form of a line graph. This smoothing strategy can effectively suppress prediction jumps caused by measurement noise, making the output sequence more consistent with the physical law of slow battery capacity degradation. Figure 6 A comparison chart showing the original predicted battery health status and the final output value after degradation constraint smoothing, as shown below. Figure 6 As shown, the horizontal axis represents the "cycle number" and is dimensionless; the vertical axis represents the "battery health status (%)".

[0119] The battery health status assessment method based on multi-dimensional operating condition feature fusion provided in this invention extracts channel statistics reflecting the internal evolution law of a single cycle from the reconstructed feature matrix and cycle-level scalars with clear physical meaning from the single-cycle time-series data matrix. These two types of features are concatenated to form a cycle-level statistical feature vector. This feature vector is then standardized using the mean vector and standard deviation vector of the cycle-level statistical features pre-saved during the training phase. Furthermore, a degradation co-vector is obtained through principal component analysis dimensionality reduction and hyperbolic tangent function compression. This vector represents the global degradation state of the current cycle in a compact form. Subsequently, in the prediction output stage, the previous... Using historical battery health state predictions from a single battery cycle as constraint anchors, the initial predictions directly output by the model are first subject to hard pruning based on preset maximum increase and decrease amplitudes to ensure that the prediction results do not violate the physical law of irreversible battery capacity decay. Then, the pruned constraint predictions are subjected to an exponentially weighted moving average with the historical predictions to suppress single-sample noise and random fluctuations. Finally, a smooth, stable, and physically reasonable battery health state prediction is output. This achieves end-to-end collaboration from local temporal feature extraction and global state compression to physical law constraint output, effectively ensuring the accuracy, robustness, and engineering applicability of the prediction results.

[0120] Example 4 This invention also provides another battery health status assessment method based on multi-dimensional operating condition feature fusion; this method is implemented on the basis of the method in the above embodiments; the method focuses on describing the specific implementation of determining the degradation cooperative vector according to the channel statistics in the reconstructed feature matrix, and performing degradation constraint smoothing on the initial battery health status prediction value to obtain the battery health status prediction value.

[0121] Figure 7 A flowchart of another battery health status assessment method based on multi-dimensional operating condition feature fusion provided in this embodiment of the invention is shown below. Figure 7 As shown, the battery health status assessment method based on multi-dimensional operating condition feature fusion may include the following steps: Step S401: Based on the collected multi-channel sampling data of the battery during the current battery cycle stage, construct a single-cycle time-series data matrix for the current battery cycle stage.

[0122] Step S402: Based on the current cycle order of the current battery cycle stage and the historical battery health status of the previous battery cycle stage indicated in the single-cycle time-series data matrix, adaptive time-series decomposition and reconstruction guided by degradation trend are performed on the single-cycle time-series data matrix to obtain the reconstructed feature matrix.

[0123] Step S403: Determine the degradation collaboration vector based on the channel statistics in the reconstructed feature matrix.

[0124] Step S404: Input the reconstructed feature matrix and the degenerate co-vector into a depthwise separable temporal convolutional encoder, extract temporal features using multi-layer dilated depthwise separable convolution, and perform channel-level enhancement on the output of each convolution layer to obtain a degenerate gated temporal feature matrix.

[0125] The effective dependency range of battery single-cycle time-series features changes with the degradation stage. In early cycles, degradation information is more reflected in local short-term fluctuations; in late cycles, degradation information is more reflected in trend changes over a longer time range. Conventional temporal convolutional models use fixed convolution scales and fixed channel responses, which are difficult to adapt to the differences in degradation rates at different battery cycling stages, and are prone to insufficient representation of early details or inadequate modeling of late trends.

[0126] By inputting the reconstructed feature matrix Z and the degradation co-vector v into a degradation-gated temporal prediction model, local and extended temporal dependencies are extracted through depthwise separable temporal convolution. Furthermore, by dynamically modulating the convolution channels and temporal attention weights using the degradation co-vector v, adaptive prediction for different degradation stages can be achieved. Specifically, the reconstructed feature matrix... As a temporal input, multi-layer dilated depthwise separable convolutions are used to extract temporal features, which are then processed by degenerate collaborative vectors. Generate channel gating factors to enhance the channel level of each convolutional output, enabling the model to dynamically adjust the importance of different convolutional channels based on the degradation state of the current loop.

[0127] Furthermore, the feature matrix will be reconstructed. As the timing input of layer 0 (denoted as...) ), This refers to the initial time series feature matrix before entering the degenerate-gated time series prediction model, with a size of It is used to provide degenerate gated time series prediction models with single-cycle time series information that has been standardized, trend reconstructed and volatility enhanced.

[0128] Let the total number of convolutional coding layers be... , This indicates the number of depth-separable temporal convolutional layers with degeneracy gating, and can be 2 or 3. The input to the convolutional coding layer is The output is ;in, This represents the convolutional layer index, with values ​​ranging from 1 to... Number of output channels in the first layer 32 is a good number; this refers to the number of output channels in the second layer. 64 can be chosen. The timing characteristics of the last layer of degenerate gating are: The size is ,in, The number of channels represents the timing characteristics of the last layer of degenerate gating, which can be 64.

[0129] In the In the layer, dilated depthwise convolution is first performed on each input channel individually. Dilated depthwise convolution is used to expand the temporal receptive field without significantly increasing the number of parameters, thereby capturing both local short-term changes and long-term trend changes at the same time.

[0130] In practical implementation, let the first... The kernel radius of the layer is , This controls the time range of the convolution kernel coverage; a value of 2 can be used to determine the convolution kernel coverage. The offset position; let the first... The expansion rate of the layer is , This controls the time interval between adjacent sampling points of the convolution kernel. For example, the dilation rate of the first layer can be 1, and the dilation rate of the second layer can be 2. For the... Input channel, at the At the time step, according to the offset index to Read The Middle Time step The input features of each channel are multiplied and summed with the corresponding depthwise convolution kernel parameters to obtain the depthwise convolution result. , is represented as: ; in, Indicates the first Layer Time step The depthwise convolution result of the input channel; This represents the input channel index, with a value ranging from 1 to... ; Indicates the first The number of channels in the layer, when hour ; Indicates the kernel offset index; Indicates the first Layer Input channel at offset position The depthwise convolution kernel parameters at a given point are trainable parameters of the neural network. This represents a boundary clipping function used to clip values ​​exceeding 1 to... The time index of the range is pruned to the nearest boundary.

[0131] Perform pointwise convolution on the depthwise convolution result to obtain the first... The convolutional response features of the layer. Among them, pointwise convolution is used to fuse degradation-related information between different input channels, which is equivalent to performing a trainable linear combination of multiple input channels at each time step.

[0132] In specific implementation, for the first For the output channel, the depthwise convolution results of all input channels are summed using pointwise convolution parameters and weighted, then the convolution bias is added to obtain the convolution response value. , is represented as: ; in, Indicates the first Layer Time step Output channel convolution response value before degradation gating; This represents the output channel index, with a value ranging from 1 to... ; Indicates the first Number of output channels of the layer; Indicates the first From the first layer Input channel to the The pointwise convolution parameters of the output channels are trainable parameters of the neural network; Indicates the first Layer The convolution bias of the output channel is a trainable parameter of the neural network.

[0133] It should be noted that by combining depthwise convolution and pointwise convolution, the model can first extract local temporal patterns within a single physical channel, and then fuse multiple operating condition information in the channel dimension. Compared with ordinary temporal convolution, this can reduce the number of parameters and improve the expression efficiency of multi-channel degradation features.

[0134] Utilizing degenerate collaborative vectors For the first Each output channel of the layer generates a channel gating factor, which is then used to enhance the convolutional response features. The channel gating factor adjusts the importance of different output channels based on the current recurrence's global degradation stage, enabling the model to employ different feature response patterns in early and late recurrences.

[0135] In specific implementation, for the first Layer Output channel setting gated projection vector and gate bias Among them, the gated projection vector The length is Used to degenerate collaborative vectors Mapped to the gating strength of this output channel; gating bias This is used to adjust the gating reference value of the output channel. Therefore, the channel gating factor... Based on degenerate collaborative vector With gated projection vector inner product and gate bias The calculated value, and its enhancement range limited by the Sigmoid function, is expressed as: ; in, Indicates the first Layer The channel gating factor of the output channel; This represents the Sigmoid function, with an output range of 0 to 1. Represents the gated projection vector With degenerate co-vector The inner product between; and All of these are trainable parameters of the neural network. Due to the channel gating factor... The value ranges from 1 to 2, so channel gating can enhance the response of degradation-related channels while avoiding completely suppressing a channel to 0.

[0136] The first Convolutional response value of the layer With the corresponding channel gating factor Multiply and combine with the residual alignment mapping to obtain the first... Layer degradation gating timing characteristics Among them, residual alignment mapping is used to maintain the transitivity of input information when the number of channels changes, thereby improving the stability of multi-layer convolution training.

[0137] In practical implementation, if the first Number of input channels per layer and the first layer If the number of output channels of each layer is the same, then residual alignment mapping is performed. This is an identity mapping; if the number of channels is inconsistent, then a residual alignment mapping is used. Dimension alignment is achieved using 1×1 pointwise convolution or linear projection. Then, the... Layer Time step Degeneracy gate timing characteristics of the output channel The calculation method is expressed as follows: ; in, Indicates the first Layer Time step Degeneracy gate timing characteristics of the output channel; Represents the th after residual alignment mapping Time step Channel residual eigenvalues; This represents a linear rectified activation function, used to preserve the positive response and introduce nonlinear expressive power.

[0138] In one implementation, the expansion ratio of the first layer is 1 to extract local fluctuation features between adjacent time steps; the expansion ratio of the second layer is 2 to extract trend change features over a wider time range. Degenerate Co-vector Channel gating factors are generated for the output channels of both layers, allowing the early loops to emphasize channels related to local fluctuations and the late loops to emphasize channels related to trend changes. Based on this, the temporal coding process can adaptively adjust with the degradation stage, avoiding the problem that a fixed convolutional response cannot accommodate different degradation stages.

[0139] Complete in sequence Degenerate-gated depthwise separable temporal convolutions are used to compute the temporal features of the last degenerate-gated layer. The last layer of degenerate gating timing features This refers to the time-series coded feature matrix after multi-layer degradation conditional channel modulation, with a size of This is used for subsequent degradation condition-driven temporal attention pooling, and can also be called the degradation-gated temporal feature matrix.

[0140] It should be noted that the degenerate-gated depthwise separable temporal convolutional coding does not use fixed channel weights for all cycles, but rather uses degenerate co-vectors. Dynamically generated channel gating factors enable the model to adjust the importance of different convolutional channels based on the current macroscopic degradation state of the loop. Based on this, the model can maintain sensitivity to local details when capacity decay is mild, and enhance its response to trend changes and degradation-sensitive channels when capacity decay is severe.

[0141] Step S405: Input the degradation collaborative vector into the attention query projection layer to obtain the degradation condition query vector.

[0142] Within a single cycle, not all time steps contribute equally to battery health prediction. For example, early cycles may focus more on local voltage changes during the constant current charging phase, while later cycles may focus more on duration changes and internal resistance response during the constant voltage charging phase. The degradation co-vector... Input attention to the projection layer to obtain the degradation condition query vector. The attention query projection layer is used to map the global degradation information of the current loop to the same feature dimension as the degradation gating temporal features of the last layer.

[0143] In the specific implementation, the degenerate collaborative vector is transformed through a linear projection layer. Mapped to degenerate conditional query vector This query vector is used for subsequent attention matching with temporal features, i.e. .in, This represents the degenerate condition query vector, with a length of [length missing]. This is used to select a more important time step based on the current cycle degradation stage; This represents the attention query projection matrix, with a size of [value missing]. , which are trainable parameters of a neural network; This represents the attention query bias vector, with a length of [value missing]. These are trainable parameters of a neural network.

[0144] Through this operation, temporal attention pooling no longer determines weights solely based on the features of the time step itself, but introduces the global degradation state of the current loop, enabling loops at different times to focus on different temporal segments.

[0145] Step S406: Determine the temporal attention weights based on the degradation condition query vector and the degradation gating temporal feature matrix.

[0146] Calculate the degradation condition query vector With the timing characteristics of the last layer of degenerate gating The similarity between feature vectors at each time step is calculated, and the similarity across all time steps is normalized to obtain the temporal attention weights. , Characterizing the first The temporal attention weights of the time steps are used to measure the attention of the first time step. The contribution of time step features to the prediction of the current cycle battery health status is calculated, with the sum of the temporal attention weights for all time steps being 1.

[0147] In the specific implementation, first calculate the first... Attention similarity score at time step ,Right now .in, Indicates the first Attention similarity score at each time step; This indicates that the timing characteristics of the last layer of degenerate gating are in the first... The feature vector at time step has a length of ; Degradation condition query vector With the The inner product similarity between feature vectors at time steps.

[0148] Furthermore, Softmax normalization is performed on the attention similarity scores across all time steps to obtain the temporal attention weights. , is represented as: ; in, This represents the attention normalization summation index, with values ​​ranging from 1 to... ; Indicates the first Attention similarity score at each time step; This represents the maximum attention similarity score across all time steps, used to improve the numerical stability of exponential operations. This represents the natural exponential function.

[0149] Step S407: The degenerate gating temporal features are weighted and summed according to the temporal attention weights to obtain the temporal context vector.

[0150] Based on temporal attention weights Timing characteristics of the last layer of degenerate gating We obtain the temporal context vector by weighted summation of all time steps. ,Right now Temporal context vector This refers to the aggregated time-domain information most relevant to battery health state prediction in the current cycle, with a length of [length value missing]. .because Degenerate Cooperative Vector Driven generation, therefore time-domain context vector It can adaptively emphasize information at different time steps based on the current cycle degradation stage.

[0151] It should be noted that in the early stages of the cycle, the time-domain attention weights can focus on the detailed changes in voltage and current during the initial stage of constant current charging; in the later stages, the time-domain attention weights can focus on the changes in internal resistance and constant voltage charging time during the final stage of constant voltage charging. Based on this, the time information aggregation method can adaptively change with the degradation stage.

[0152] Step S408: Concatenate the temporal context vector and the degenerate co-vector along the feature dimension to obtain the regression input vector.

[0153] Temporal context vector With degenerate co-vector Concatenate along the feature dimensions to obtain the regression input vector. Regression input vector It refers to a prediction input vector that simultaneously contains both the internal timing information of a single loop and the information of the loop-level degradation stage.

[0154] Step S409: Input the regression input vector into a fully connected regression network to obtain the initial battery health state prediction value for the current battery cycle stage.

[0155] Regression input vector Input the fully connected regression network to obtain the battery health state prediction for the current cycle. Battery health status prediction value This refers to the estimation result of the degradation-gated time-series prediction model on the current cycle capacity maintenance level, expressed as a percentage.

[0156] In one implementation, the fully connected regression network consists of two fully connected layers. The first fully connected layer has an output dimension of 32 and is followed by a linear rectification activation function; the second fully connected layer has an output dimension of 1, does not use an activation function, and is used to output the predicted battery health status. To avoid prediction results exceeding a reasonable range, range pruning can be performed after output. Prediction results less than 0 are pruned to 0, and prediction results greater than 120 are pruned to 120. This pruning range retains a small amount of capacity recovery or measurement bias space, while preventing abnormal output from affecting online applications.

[0157] In one embodiment, via Figure 8 and Figure 9 This demonstrates an example of temporal attention weights driven by degradation conditions. Figure 8 This is a schematic diagram of the attention weight distribution in the early recurrent temporal domain. Figure 9 This is a schematic diagram of the attention weight distribution in the late-stage recurrent time domain. (Example:) Figure 8 and Figure 9 As shown, the horizontal axis represents the time step (dimensionless), and the vertical axis represents the attention weights (dimensionless, summing to 1). The attention weights in early loops are mainly concentrated in the earlier time steps, while the attention weights in later loops shift significantly to later time steps. After the degenerate collaborative vector drives temporal attention pooling, the time segments of interest in different degenerate stages exhibit adaptability.

[0158] Step S410: Perform degradation constraint smoothing on the initial battery health state prediction value to obtain the battery health state prediction value.

[0159] It is understandable that during the training phase, the degradation-gated time-series prediction model can obtain a large number of early healthy state samples from actual battery cycling data, while the number of low healthy state samples and samples from the rapid capacity decay stage is relatively small. If ordinary mean square error loss is used, the trained degradation-gated time-series prediction model is prone to bias towards early stable cycles, resulting in larger prediction errors for late healthy states.

[0160] In this embodiment of the invention, a degradation penalty weight determined by the battery health status of the previous cycle is introduced into the loss function, and a monotonic constraint penalty is added so that samples with low health status receive a higher error penalty.

[0161] Specifically, during model training, from the training sample set Selected from A training batch consists of 10 single-cycle samples. This indicates the number of samples in the training batch, which can be 32 or 64. For the nth sample in the training batch... The degradation-gated time-series prediction model outputs a battery health state prediction value from a training sample. The training data provides labels for the actual battery health status. .in, This represents the sample index within the training batch, with a value ranging from 1 to... ; Indicates the first Predicted battery health status values ​​for each training sample; Indicates the first The true battery health status labels of the training samples.

[0162] Furthermore, from the first The battery health status of the previous cycle is read from the previous cycle battery health status channel of each training sample. and according to Calculate the degradation penalty weight Degradation penalty weight Used to control the The influence of the prediction error of each training sample on the loss function depends on the battery health status in the previous cycle, which increases the degradation penalty weight. The larger the value, the more attention the model pays to samples with low health status during training.

[0163] Based on the battery health status of the previous cycle and nominal battery health status Calculate the degradation penalty weight , is represented as: ; in, Indicates the first Degradation penalty weights for each training sample; This represents the degradation weight growth slope hyperparameter, used to control the magnitude of the enhancement of error penalty for low-healthy samples; This represents the maximum value function, used to ensure that the degradation penalty weight is not less than 1. When battery health is expressed as a percentage... A value of 0.02 can be used, meaning that for every 1 percentage point decrease in battery health in the previous cycle, the error penalty weight increases by approximately 0.02.

[0164] In one embodiment, for example, if The battery health status of a training sample in the previous cycle ,and Then the degradation penalty weight The prediction error of this training sample will have a greater impact on parameter updates than the training sample from the previous cycle where the battery health state was close to 100, thereby improving the model's learning strength for low health state stages and rapid capacity decay stages.

[0165] Based on degradation penalty weights Battery health status prediction value and the actual battery health status label Calculate the degenerate weighted mean square error loss , is represented as: ; in, This represents the degradation-weighted mean squared error loss, used to measure the weighted deviation between the model's predicted values ​​and the actual battery health status labels.

[0166] Furthermore, to reduce the abnormal increase in the predicted battery health status compared to the previous cycle, a monotonic constraint penalty loss is introduced. In practical implementation, a small increase due to measurement error or short-term capacity recovery is allowed, and an increase tolerance threshold is set. , This represents the maximum allowable increase in the current cycle's predicted value relative to the previous cycle's battery health state, and can be taken as 0.2 to 0.5 percentage points. The monotonic constraint penalty loss is expressed as: ; in, This represents the monotonic constraint penalty loss, used to suppress predictions that significantly violate the battery degradation trend.

[0167] The total training loss is obtained by weighting the degradation-weighted mean square error loss and the monotonic constraint penalty loss. ,Right now .in, Indicates the total training loss; This represents the monotonic constraint loss weight, used to control the proportion of the monotonic constraint penalty loss in the total training loss, and can be taken as 0.05 to 0.20.

[0168] It should be noted that the monotonic constraint penalty loss does not force all predicted values ​​to decrease strictly, but only for those exceeding the rise tolerance threshold. The abnormal increase is penalized, which can both preserve the small capacity rebound that may occur in battery testing and limit the model output to results that are obviously inconsistent with the degradation law.

[0169] The Adam optimizer is used to iteratively update the trainable parameters of the neural network in the degenerate gated time series prediction model. In one implementation, the initial learning rate of the Adam optimizer can be 0.001, and the weight decay can be... The number of training cycles can be between 100 and 200.

[0170] At the end of each training cycle, the degradation-gated time-series prediction model is evaluated using a validation set. Validation metrics may include mean absolute error, root mean square error, and maximum absolute error. The model parameters corresponding to the lowest mean absolute error on the validation set are saved as the parameters for the final battery health state prediction model.

[0171] It should be noted that the battery health state prediction model consists of a degradation-gated deep separable temporal convolutional encoder, a degradation-conditionally driven temporal attention pooling layer, and a fully connected regression network; firstly, the feature matrix is ​​reconstructed. Input encoder, passing through sequentially The layer consists of degenerate-gated depthwise separable temporal convolutional layers. Within each layer, dilated depthwise convolutions are first performed on the input to extract temporal patterns at different scales. Then, channel information is fused through pointwise convolutions, and finally, degenerate co-vectors are utilized. The generated channel gating factor dynamically enhances the output of this layer's convolution, and the output of this layer is obtained through residual connections. The encoder's final output Entering the temporal attention pooling layer, the degenerate collaborative vector... The mapped query vector is used to calculate the attention weights at each time step, and then... Weighted summation yields the time-domain context vector. Finally, With degenerate co-vector The input is a fully connected regression network consisting of two fully connected layers, which directly outputs the predicted battery health status. In the model, the degenerate collaborative vector It serves as a conditional input throughout the encoder and pooling layers to drive adaptive processing for different degradation stages.

[0172] It should also be noted that the channel mutual information coupling weights and the principal component projection matrix... Training set cyclic statistical feature mean vector Training set cyclic statistical feature standard deviation vector and the maximum cycle number of the training set These parameters are all determined during the training data preprocessing stage and are not considered trainable parameters of the neural network. During model training, they are only determined through the total training loss. Backpropagation updates the trainable parameters of the neural network in the degenerate gated time series prediction model.

[0173] In one embodiment, the layer-by-layer dynamic modulation effect of the degenerate co-vector on the depthwise separable temporal convolutional coding is demonstrated through channel gating factor distribution experiments. Figure 10This is a schematic diagram of the distribution of the first-layer channel gating factors. Figure 11 This is a schematic diagram of the gating factor distribution for the second-layer channel. (Example:) Figure 10 and Figure 11 As shown, the horizontal axis represents the "channel index," and the vertical axis represents the "channel gating factor (unitless)." The first layer has 32 output channels, and the second layer has 64 output channels. Compared to the traditional temporal convolution that uses fixed channel response weights for all cycles, the channel gating factor, dynamically generated by the degradation co-vector, allows the same convolutional network to automatically switch the importance ranking of its feature channels at different degradation stages.

[0174] The battery health status assessment method based on multi-dimensional operating condition feature fusion provided in this invention employs a channel gating mechanism driven by degradation cooperative vectors. This mechanism dynamically modulates the importance of the output channels of each layer of separable convolutions with varying expansion depths. This allows the model to emphasize local fluctuation channels in early cycles and enhance trend change channels in later cycles, achieving adaptive switching of the temporal encoding process according to the degradation stage. This overcomes the limitation of conventional fixed convolutional responses failing to cover the differences in degradation rates throughout the entire lifecycle. The degradation cooperative vectors are used to generate temporal attention query vectors, which are then adaptively weighted and aggregated for single-cycle time steps at different degradation stages. This ensures that early cycles focus on constant current charging details, while later cycles focus on internal resistance changes at the end of constant voltage charging. Simultaneously, degradation penalty weights determined by the health status of the previous cycle and monotonic constraint penalties are introduced into the loss function. This allows the model to impose higher penalties on low-health-state samples and abnormal rise predictions during training, improving the accuracy and physical plausibility of late-stage capacity decay predictions.

[0175] Example 5 Corresponding to the above method embodiments, this invention provides a battery health status assessment device based on multi-dimensional operating condition feature fusion. Figure 12 This is a schematic diagram of a battery health status assessment device based on multi-dimensional operating condition feature fusion provided in an embodiment of the present invention, as shown below. Figure 12 As shown, the battery health status assessment device based on multi-dimensional operating condition feature fusion may include: The data acquisition module 501 is used to construct a single-cycle time-series data matrix for the current battery cycle stage based on the multi-channel sampling data of the battery collected during the current battery cycle stage. The feature extraction module 502 is used to perform adaptive temporal decomposition and reconstruction guided by degradation trend on the single-cycle time-series data matrix based on the current cycle order indicating the current battery cycle stage and the historical battery health status of the previous battery cycle stage in the single-cycle time-series data matrix, so as to obtain the reconstructed feature matrix. The degradation co-vector determination module 503 is used to determine the degradation co-vector based on the channel statistics in the reconstructed feature matrix; The prediction module 504 is used to input the reconstructed feature matrix and the degradation co-vector into the degradation-gated temporal prediction model to obtain the initial battery health state prediction value of the current battery cycle stage; wherein, the convolution channels and temporal attention weights in the degradation-gated temporal prediction model are dynamically modulated by the degradation co-vector. The constraint smoothing module 505 is used to perform degradation constraint smoothing on the initial battery health state prediction value to obtain the battery health state prediction value.

[0176] This invention provides a battery health status assessment device based on multi-dimensional operating condition feature fusion. By constructing a standardized time-series data matrix, introducing time-series decomposition and reconstruction based on cyclic process adaptive adjustment, and extracting a collaborative vector representing the global degradation state, this device dynamically modulates the channel gating factor and temporal attention distribution of a deep separable convolutional network using this collaborative vector. This achieves end-to-end adaptive evolution of feature extraction and model inference as the battery ages. It effectively overcomes the fundamental defect of traditional methods where fixed preprocessing parameters and static network weights cannot simultaneously consider the sensitivity of early local details and the accuracy of late-stage macroscopic trend modeling. By processing the initial battery health status prediction value through amplitude pruning and exponential smoothing based on physical degradation laws, it not only significantly reduces the prediction error during the rapid capacity decay stage throughout the entire life cycle but also fundamentally suppresses abnormal jumps that violate the irreversible laws of electrochemistry. This results in an output sequence with high accuracy, strong smoothness, and physical interpretability, providing a highly reliable, plug-and-play online health status assessment and early warning capability for battery management systems.

[0177] In some embodiments, the feature extraction module 502 is further configured to: Obtain the channel mutual information coupling weights obtained during the training phase of the degradation-gated time series prediction model; Based on the single-cycle time-series data matrix, which indicates the current cycle order of the current battery cycle stage, the historical battery health status of the previous battery cycle stage, and the channel mutual information coupling weight, the single-cycle time-series data matrix is ​​standardized to obtain a fused standardized feature matrix. Based on the current cycle order and the historical battery health status, the fused standardized feature matrix is ​​subjected to adaptive temporal decomposition and reconstruction guided by degradation trends to obtain the reconstructed feature matrix.

[0178] In some embodiments, the feature extraction module 502 is further configured to: Standardization is performed on each channel in the single-cycle time-series data matrix to obtain the basic standardized feature matrix; Based on the current cycle order of the current battery cycle stage and the historical battery health status of the previous battery cycle stage indicated in the single-cycle timing data matrix, the channel degradation modulation coefficient corresponding to each channel in the single-cycle timing data matrix is ​​determined. Based on the channel degradation modulation coefficients, the basic normalized feature matrix is ​​differentially enhanced to obtain the degradation-enhanced normalized feature matrix; Based on the channel mutual information coupling weights, the basic normalized feature matrix and the degradation-enhanced normalized feature matrix are fused to obtain the fused normalized feature matrix.

[0179] In some embodiments, the feature extraction module 502 is further configured to: Determine the adaptive smoothing factor based on the current loop order; Based on the adaptive smoothing factor, the fused standardized feature matrix is ​​recursively smoothed to obtain the trend feature matrix; The fluctuation residual matrix is ​​determined based on the fused standardized feature matrix and the trend feature matrix; The reconstructed feature matrix is ​​determined based on the fluctuation residual matrix and the trend feature matrix.

[0180] In some embodiments, the degradation collaboration vector determination module 503 is further configured to: Obtain the channel statistics corresponding to the reconstructed feature matrix and the cyclic scalar corresponding to the single-cycle time series data matrix; The cyclic scalar and the channel statistics are concatenated to obtain the cyclic statistical feature vector; Obtain the mean vector and standard deviation vector of the cyclic statistical features obtained during the training phase of the degenerate gated time series prediction model; The degradation collaboration vector is determined based on the mean vector of the cyclic statistical features, the standard deviation vector of the cyclic statistical features, and the cyclic statistical feature vector.

[0181] In some embodiments, the prediction module 504 is further configured to: The reconstructed feature matrix and the degenerate co-vector are input into a depthwise separable temporal convolutional encoder. Temporal features are extracted using multi-layer dilated depthwise separable convolutions, and channel-level enhancement is performed on the output of each convolution layer to obtain a degenerate gated temporal feature matrix. The degradation collaborative vector is input into the attention query projection layer to obtain the degradation condition query vector; The temporal attention weights are determined based on the degradation condition query vector and the degradation gating temporal feature matrix. The degenerate gated temporal features are weighted and summed according to the temporal attention weights to obtain the temporal context vector; The temporal context vector and the degenerate co-vector are concatenated along the feature dimension to obtain the regression input vector; The regression input vector is input into a fully connected regression network to obtain the initial battery health status prediction value for the current battery cycle stage.

[0182] In some embodiments, the constraint smoothing module 505 is further configured to: Obtain historical battery health status prediction values ​​from the previous battery cycle stage; The initial battery health state prediction value is constrained by the variation range of the historical battery health state prediction value to obtain the constrained prediction value. The battery health state prediction value is obtained by exponentially smoothing the constraint prediction value and the historical battery health state prediction value.

[0183] In some embodiments, the data acquisition module 501 is further configured to: According to the preset sampling frequency, multiple physical quantity data of the battery during the current battery cycle stage are collected to obtain multi-channel sampling data; According to the preset sampling length, the multi-channel sampled data is resampled at equal intervals to obtain a single-cycle time-series data matrix.

[0184] In some embodiments, the feature extraction module 502 is further configured to: Obtain the maximum loop order obtained by the degradation-gated time series prediction model during the training phase; The normalized loop process is determined based on the current loop order and the maximum loop order; Based on the normalization cycle process and the preset minimum smoothing factor, maximum smoothing factor, and smoothing factor growth rate coefficient, the adaptive smoothing factor is determined.

[0185] 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.

[0186] Example 6 This invention also provides an electronic device for running the above-described battery health status assessment method based on multi-dimensional operating condition feature fusion; see also Figure 13The diagram shows the structure of an electronic device, which includes a memory 600 and a processor 601. The memory 600 stores one or more computer instructions, which are executed by the processor 601 to implement the battery health status assessment method based on the fusion of multi-dimensional operating condition features.

[0187] Furthermore, Figure 13 The electronic device shown also includes a bus 602 and a communication interface 603, with the processor 601, communication interface 603 and memory 600 connected via the bus 602.

[0188] The memory 600 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 603 (which can be wired or wireless), such as the Internet, wide area network, local area network, metropolitan area network, etc. The bus 602 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 13 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.

[0189] Processor 601 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 601 or by instructions in software form. Processor 601 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 600. Processor 601 reads information from memory 600 and, in conjunction with its hardware, completes the steps of the method described in the foregoing embodiments.

[0190] 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 assessment method based on multi-dimensional operating condition feature fusion. For specific implementation details, please refer to the method embodiments, which will not be repeated here.

[0191] The computer program product for a battery health status assessment method based on multi-dimensional operating condition feature fusion provided in this 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.

[0192] 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.

[0193] 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.

[0194] 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.

[0195] 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.

[0196] 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.

[0197] 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 battery health status assessment method based on multi-dimensional operating condition feature fusion, characterized in that, include: Based on the collected multi-channel sampling data of the battery during the current battery cycle stage, a single-cycle time-series data matrix for the current battery cycle stage is constructed. Based on the current cycle order of the current battery cycle stage and the historical battery health status of the previous battery cycle stage indicated in the single-cycle time-series data matrix, the single-cycle time-series data matrix is ​​subjected to adaptive time-series decomposition and reconstruction guided by degradation trend to obtain the reconstructed feature matrix. The degradation collaboration vector is determined based on the channel statistics in the reconstructed feature matrix; The reconstructed feature matrix and the degradation co-vector are input into the degradation-gated temporal prediction model to obtain the initial battery health state prediction value of the current battery cycle stage; wherein, the convolution channels and temporal attention weights in the degradation-gated temporal prediction model are dynamically modulated by the degradation co-vector; The initial battery health state prediction value is smoothed by degradation constraints to obtain the battery health state prediction value.

2. The method according to claim 1, characterized in that, Based on the current cycle order indicating the current battery cycle stage and the historical battery health status of the previous battery cycle stage in the single-cycle time-series data matrix, an adaptive time-series decomposition and reconstruction guided by degradation trends is performed on the single-cycle time-series data matrix to obtain a reconstructed feature matrix, including: Obtain the channel mutual information coupling weights obtained during the training phase of the degradation-gated time series prediction model; Based on the single-cycle time-series data matrix, which indicates the current cycle order of the current battery cycle stage, the historical battery health status of the previous battery cycle stage, and the channel mutual information coupling weight, the single-cycle time-series data matrix is ​​standardized to obtain a fused standardized feature matrix. Based on the current cycle order and the historical battery health status, the fused standardized feature matrix is ​​subjected to adaptive temporal decomposition and reconstruction guided by degradation trends to obtain the reconstructed feature matrix.

3. The method according to claim 2, characterized in that, Based on the single-cycle time-series data matrix indicating the current cycle order of the current battery cycle stage, the historical battery health status of the previous battery cycle stage, and the channel mutual information coupling weights, the single-cycle time-series data matrix is ​​standardized to obtain a fused standardized feature matrix, including: Standardization is performed on each channel in the single-cycle time-series data matrix to obtain the basic standardized feature matrix; Based on the current cycle order of the current battery cycle stage and the historical battery health status of the previous battery cycle stage indicated in the single-cycle timing data matrix, the channel degradation modulation coefficient corresponding to each channel in the single-cycle timing data matrix is ​​determined. Based on the channel degradation modulation coefficients, the basic normalized feature matrix is ​​differentially enhanced to obtain the degradation-enhanced normalized feature matrix; Based on the channel mutual information coupling weights, the basic normalized feature matrix and the degradation-enhanced normalized feature matrix are fused to obtain the fused normalized feature matrix.

4. The method according to claim 2, characterized in that, Based on the current cycle order and the historical battery health status, the fused normalized feature matrix is ​​subjected to adaptive temporal decomposition and reconstruction guided by degradation trends to obtain the reconstructed feature matrix, including: Determine the adaptive smoothing factor based on the current loop order; Based on the adaptive smoothing factor, the fused standardized feature matrix is ​​recursively smoothed to obtain the trend feature matrix; The fluctuation residual matrix is ​​determined based on the fused standardized feature matrix and the trend feature matrix; The reconstructed feature matrix is ​​determined based on the fluctuation residual matrix and the trend feature matrix.

5. The method according to claim 1, characterized in that, Based on the channel statistics in the reconstructed feature matrix, the degenerate collaborative vector is determined, including: Obtain the channel statistics corresponding to the reconstructed feature matrix and the cyclic scalar corresponding to the single-cycle time series data matrix; The cyclic scalar and the channel statistics are concatenated to obtain the cyclic statistical feature vector; Obtain the mean vector and standard deviation vector of the cyclic statistical features obtained during the training phase of the degenerate gated time series prediction model; The degradation collaboration vector is determined based on the mean vector of the cyclic statistical features, the standard deviation vector of the cyclic statistical features, and the cyclic statistical feature vector.

6. The method according to claim 1, characterized in that, The reconstructed feature matrix and the degradation co-vector are input into the degradation-gated time-series prediction model to obtain the initial battery health state prediction value for the current battery cycle stage, including: The reconstructed feature matrix and the degenerate co-vector are input into a depthwise separable temporal convolutional encoder. Multi-layer dilated depthwise separable convolutions are used to extract temporal features, and channel-level enhancement is performed on the output of each convolution layer to obtain a degenerate gated temporal feature matrix. The degradation collaborative vector is input into the attention query projection layer to obtain the degradation condition query vector; The temporal attention weights are determined based on the degradation condition query vector and the degradation gating temporal feature matrix. The degenerate gated temporal features are weighted and summed according to the temporal attention weights to obtain the temporal context vector; The temporal context vector and the degenerate co-vector are concatenated along the feature dimension to obtain the regression input vector; The regression input vector is input into a fully connected regression network to obtain the initial battery health status prediction value for the current battery cycle stage.

7. The method according to claim 1, characterized in that, The initial battery health state prediction is subjected to degradation constraint smoothing to obtain the battery health state prediction value, including: Obtain historical battery health status prediction values ​​from the previous battery cycle stage; The initial battery health state prediction value is constrained by the variation range of the historical battery health state prediction value to obtain the constrained prediction value. The battery health state prediction value is obtained by exponentially smoothing the constraint prediction value and the historical battery health state prediction value.

8. The method according to claim 1, characterized in that, Based on the collected multi-channel sampling data of the battery during the current battery cycle phase, a single-cycle time-series data matrix for the current battery cycle phase is constructed, including: According to the preset sampling frequency, multiple physical quantity data of the battery during the current battery cycle stage are collected to obtain multi-channel sampling data; According to the preset sampling length, the multi-channel sampled data is resampled at equal intervals to obtain a single-cycle time-series data matrix.

9. The method according to claim 4, characterized in that, Determining the adaptive smoothing factor based on the current loop order includes: Obtain the maximum loop order obtained by the degradation-gated time series prediction model during the training phase; The normalized loop process is determined based on the current loop order and the maximum loop order; Based on the normalization cycle process and the preset minimum smoothing factor, maximum smoothing factor, and smoothing factor growth rate coefficient, the adaptive smoothing factor is determined.

10. A battery health status assessment device based on multi-dimensional operating condition feature fusion, characterized in that, include: The data acquisition module is used to construct a single-cycle time-series data matrix for the current battery cycle stage based on the multi-channel sampling data of the battery collected during the current battery cycle stage. The feature extraction module is used to perform adaptive time series decomposition and reconstruction guided by degradation trend on the single-cycle time series data matrix based on the current cycle order indicating the current battery cycle stage and the historical battery health status of the previous battery cycle stage in the single-cycle time series data matrix, so as to obtain the reconstructed feature matrix. The degradation co-vector determination module is used to determine the degradation co-vector based on the channel statistics in the reconstructed feature matrix; The prediction module is used to input the reconstructed feature matrix and the degradation co-vector into the degradation-gated time-series prediction model to obtain the initial battery health state prediction value of the current battery cycle stage; wherein, the convolution channels and temporal attention weights in the degradation-gated time-series prediction model are dynamically modulated by the degradation co-vector; The constraint smoothing module is used to perform degradation constraint smoothing on the initial battery health state prediction value to obtain the battery health state prediction value.

Images