Battery remaining life prediction method, system, and related devices

Abstract

The application is suitable for the technical field of battery health management, and particularly relates to a battery residual life prediction method, a system and related equipment. The battery residual life prediction method comprises the following steps: obtaining battery monitoring data based on a battery multi-parameter monitoring network, and preprocessing the battery monitoring data; performing spatial parameter correlation modeling and time key event detection on a multi-dimensional input tensor; performing frequency domain decoupling multi-scale analysis on a space-time joint feature vector; dynamically fusing fine-grained fluctuation features and coarse-grained trend features based on a physical constraint gating mechanism; establishing an adversarial training framework, taking the fused features as the input of a predictor in the adversarial training framework to predict the battery residual life; and optimizing the initial prediction result through the cooperation of a discriminator in the adversarial training framework and the physical constraint. Compared with the prior art, the application improves the prediction accuracy of the battery residual life.

Description

Technical Field

[0001] This invention relates to the field of battery health management technology, and in particular to a method, system and related equipment for predicting the remaining life of a battery. Background Technology

[0002] With the rapid development of the new energy industry, lithium-ion batteries, with their advantages of high energy density, long cycle life, and low self-discharge rate, have become core energy storage components in new energy vehicles, grid-scale energy storage systems, and portable electronic devices. During cyclic charging and discharging, lithium-ion batteries undergo irreversible electrochemical aging, manifested as capacity decay and increased internal resistance, ultimately leading to performance failure and even safety accidents such as thermal runaway and internal short circuits. Remaining Useful Life (RUL) refers to the number of cycles or operating time required for a battery to degrade from its current state to its end-of-life threshold. Accurate prediction of RUL is a core function of Battery Health Management (BHM) systems, playing a crucial role in ensuring the safe operation of battery systems, optimizing maintenance strategies, and improving the economic benefits throughout the battery's lifecycle.

[0003] Current methods for predicting the relative aging resistance (RUL) of lithium-ion batteries are mainly divided into two categories: physical model methods and data-driven methods. Physical model methods construct prediction models based on the battery's electrochemical mechanisms and aging kinetic equations, possessing clear physical interpretability. However, these models rely on the identification of a large number of precise electrochemical parameters and are difficult to adapt to the nonlinear aging process under complex operating conditions with multiple stress couplings, resulting in poor generalization ability in practical engineering scenarios. Data-driven methods, on the other hand, do not require in-depth analysis of the complex electrochemical reactions inside the battery and can directly learn degradation patterns from monitoring data, making them the mainstream research direction in the industry. Deep learning models based on Convolutional Neural Networks (CNNs) and Long Short-Term Memory (LSTMs) networks have been widely applied.

[0004] The existing mainstream data-driven RUL prediction technologies mainly include the following categories:

[0005] Firstly, the CNN-LSTM hybrid model extracts local degradation features from the battery charge-discharge curve using CNN, and then captures the long-term temporal dependence of capacity decay using LSTM. However, this type of model does not explicitly model the dynamic interaction between spatial dimensions (coupling relationships of multiple parameters such as voltage, temperature, and internal resistance) and temporal dimensions during battery degradation, resulting in the loss of cross-dimensional correlation features and the inability to capture the global performance degradation phenomenon caused by local parameter anomalies.

[0006] Secondly, time-series prediction models that introduce attention mechanisms optimize the weight allocation of LSTM through time attention mechanisms, focusing on the time-series information of key capacity decay stages to improve prediction accuracy. However, such models only focus on key nodes in the time-series dimension, ignoring the influence of the spatial distribution of sensor parameters on the degradation mode, and do not combine the battery electrochemical mechanism to quantify the influence weight of different physical parameters on the degradation process, resulting in insufficient ability to capture core degradation-related features.

[0007] Third, multi-scale CNN prediction models extract degradation features at different time granularities through convolutional kernels of different scales to adapt to the degradation patterns of the entire battery life cycle. However, such models lack dedicated processing branches for different frequency features and cross-scale feature adaptive fusion mechanisms, making it difficult to effectively integrate the differentiated features of early slow degradation and later accelerated aging of the battery. This results in insensitivity to early weak decay signs and susceptibility to local noise interference in long-term predictions.

[0008] In summary, existing data-driven lithium-ion battery RUL prediction technologies generally suffer from the following core shortcomings: First, there is a severe lack of physical interpretability, with the connection between mechanism and phenomenon completely severed. Existing purely data-driven methods rely solely on statistical feature modeling of sensor data, without incorporating prior knowledge of battery electrochemical mechanisms. They cannot establish causal relationships between physical states such as internal resistance growth and polarization changes and capacity decay phenomena, resulting in predictions that become "black box outputs" that are difficult to accept in engineering and maintenance scenarios. Furthermore, they cannot identify abnormal degradation patterns that violate physical common sense.

[0009] Second, effective feature extraction is difficult under mixed noise interference, and frequency domain and time domain features are not decoupled. Monitoring data during actual battery operation includes various interference components such as load fluctuations, measurement noise, and capacity regeneration. Existing single-scale models do not perform frequency domain decoupling processing on the signals, resulting in the confusion between long-term degradation trends and short-term random fluctuations. This makes it impossible to accurately extract the weak attenuation signs in the early stages of battery aging, and long-term predictions are easily affected by local noise interference, leading to serious deviations.

[0010] Third, spatiotemporal feature modeling lacks guidance from physical mechanisms, leading to inaccurate localization of key degradation nodes. Existing spatiotemporal modeling methods mostly focus only on the physical location of sensors or key temporal nodes, without explicitly quantifying the correlation weight between physical parameters and spatiotemporal anomalies by incorporating electrochemical mechanisms. This makes it impossible to distinguish the importance differences between core parameter nodes strongly correlated with degradation mechanisms and ordinary monitoring nodes, resulting in insufficient efficiency and accuracy in extracting core degradation features.

[0011] Fourth, multi-scale feature fusion lacks physical constraints and has poor adaptability across degradation stages. Existing simple multi-scale models lack dedicated processing branches and physical gating mechanisms for decoupled features, making it impossible to achieve accurate matching of "long-term trends - low-frequency features" and "short-term fluctuations - high-frequency features." This results in insensitivity to weak signals in the early, gradual degradation stage of the battery and poor adaptability to the drastic fluctuations in the later, accelerated degradation stage, failing to cover the complex degradation patterns throughout the battery's entire life cycle.

[0012] The aforementioned deficiencies are causally related: the lack of physical interpretability prevents the model from distinguishing between trends and noise; the fragmentation of spatiotemporal features leads to inaccurate localization of key degradation nodes; and the lack of physical constraints in multi-scale fusion results in poor cross-stage adaptability. No single improvement can simultaneously solve these interconnected technical problems.

[0013] Currently, the aforementioned individual technical shortcomings (lack of physical interpretability, frequency-domain temporal coupling, spatiotemporal feature fragmentation, and unconstrained multi-scale fusion) are generally considered independent problems in the industry, and are alleviated separately using local solutions such as data augmentation, attention mechanisms, and multi-scale CNNs. However, no technical solution has yet systematically and deeply coupled physical model-driven state parameter extraction, frequency-domain decoupled multi-scale analysis, physical parameter-based gated adaptive fusion, and adversarial training embedded with physical laws. This is because the aforementioned technical branches belong to different disciplines such as electrochemical mechanism modeling, signal processing, and deep learning, resulting in technical barriers. Furthermore, those skilled in the art generally believe that data-driven methods are incompatible with physical models, and there is a technical bias that "mechanism-data fusion leads to model complexity and decreased generalization." Therefore, there is a lack of technical inspiration to organically integrate the above methods.

[0014] Therefore, there is an urgent need for a new method, system, and related equipment for predicting the remaining battery life to solve the above-mentioned technical problems. Summary of the Invention

[0015] This invention provides a method, system, and related equipment for predicting the remaining life of a battery, aiming to provide a method for predicting the remaining life of a lithium-ion battery that combines high prediction accuracy, strong physical interpretability, excellent anti-interference ability, and full life cycle adaptability.

[0016] In a first aspect, the present invention provides a method for predicting the remaining life of a battery, the method comprising the following steps: S1. Obtain battery monitoring data based on the battery multi-parameter monitoring network, and preprocess the battery monitoring data to obtain a multi-dimensional input tensor; S2. Perform spatial parameter correlation modeling and time key event detection on the multidimensional input tensor to obtain a spatiotemporal joint feature vector; S3. Perform frequency domain decoupling multi-scale analysis on the spatiotemporal joint feature vector to extract fine-grained fluctuation features and coarse-grained trend features respectively; S4. Based on a physical constraint gating mechanism, the fine-grained fluctuation features and the coarse-grained trend features are dynamically fused to obtain fused features; S5. Establish an adversarial training framework, use the fused features as input to the predictor in the adversarial training framework to predict the remaining battery life, and obtain an initial prediction result; optimize the initial prediction result by cooperating with the discriminator and physical constraints in the adversarial training framework to obtain the remaining battery life prediction result.

[0017] Preferably, step S1 includes the following sub-steps: S11. Perform sliding window segmentation on the battery monitoring data, extract time-series data and capacity labels, and obtain basic feature data; S12. Standardize and correct outlier data for the basic feature data, and decouple frequency and time domains by synchronous discrete wavelet transform to obtain low-frequency approximate components and high-frequency detail components. S13. Based on a simplified equivalent circuit electrochemical model, the current and voltage data of the battery are acquired in real time to obtain physical state parameters; wherein, the physical state parameters include ohmic internal resistance, polarization voltage, rate of change of internal resistance, and theoretical state of charge. S14. Stack the basic feature data, the low-frequency approximation component, the high-frequency detail component, and the physical state parameters in a dimensional manner to obtain the multidimensional input tensor.

[0018] Preferably, step S2 includes the following sub-steps: S21. Based on the physical topology and mechanism correlation matrix of the battery multi-parameter monitoring network, establish a parameter spatial location encoding matrix, extract the spatial coupling relationship between parameters in the multidimensional input tensor through a graph convolutional neural network, and fuse it with the parameter spatial location encoding matrix to obtain a spatial weight matrix. S22. Obtain the battery capacity decay sequence, and calculate the time attention weights for the low-frequency approximation component and the high-frequency detail component based on the multi-head attention mechanism; wherein, the low-frequency approximation component captures the global degradation trend and temporal dependency of the battery capacity decay sequence through long-range attention focusing; the high-frequency detail component identifies the time node corresponding to the short-circuit precursor in the battery through local attention focusing; and the battery charge-discharge cycle characteristics are incorporated into the battery capacity decay sequence through timestamp encoding to obtain the time attention score matrix; S23. Perform tensor multiplication on the spatial weight matrix and the temporal attention score matrix to obtain the spatiotemporal joint feature vector.

[0019] Preferably, step S3 includes the following sub-steps: S31. A trend feature extraction model is established based on dilated convolution and causal long short-term memory network. The low-frequency approximation component and physical state parameter in the spatiotemporal joint feature vector are extracted by the trend feature extraction model to obtain the coarse-grained trend feature. S32. A fluctuation feature extraction model is established based on a one-dimensional convolutional and bidirectional long short-term memory network. The high-frequency detail components in the spatiotemporal joint feature vector are extracted using the fluctuation feature extraction model to obtain the fine-grained fluctuation features.

[0020] Preferably, step S4 includes the following sub-steps: S41. Using the rate of change of internal resistance in the physical state parameters as a gating signal, the contribution weights of the coarse-grained trend feature and the fine-grained fluctuation feature are dynamically adjusted through the physical constraint gating mechanism, and feature fusion is performed to obtain the fused feature.

[0021] Preferably, step S5 includes the following sub-steps: S51. Construct the adversarial training framework; wherein, the predictor in the adversarial training framework adopts a temporal convolutional layer stacking structure; the discriminator in the adversarial training framework adopts a one-dimensional spectral normalized convolutional network. S52. Use the fused features as input to the predictor to predict the remaining battery life and obtain an initial prediction result; S53. Establish a joint loss function, and optimize the initial prediction result using the joint loss function and the discriminator to obtain the battery remaining life prediction result.

[0022] Preferably, the joint loss function satisfies the following condition: ; in, Denotes the joint loss function, This represents the mean squared error loss. Indicating resistance to loss, Indicates the loss of physical consistency. and All represent weighting coefficients. The discriminator is only involved in calculating the adversarial loss. This is used to constrain the predicted lifetime distribution to be consistent with the actual sample distribution; physical consistency loss. The system is independently constructed based on the electrochemical degradation law of the battery and does not depend on the output of the discriminator; the two losses are weighted by coefficients. After weighting, compared with the basic regression loss The total loss is obtained by summing the results, thereby achieving dual collaborative optimization of data distribution constraints and physical mechanism constraints.

[0023] Secondly, the present invention also provides a battery remaining life prediction system, the battery remaining life prediction system comprising: The physical enhancement preprocessing module is used to acquire battery monitoring data based on the battery multi-parameter monitoring network, preprocess the battery monitoring data, and obtain a multi-dimensional input tensor. The physical perception spatiotemporal module is used to perform spatial parameter correlation modeling and key temporal event detection on the multidimensional input tensor to obtain a spatiotemporal joint feature vector. The frequency domain decoupled multi-scale module is used to perform frequency domain decoupled multi-scale analysis on the spatiotemporal joint feature vector, and extract fine-grained fluctuation features and coarse-grained trend features respectively. The physical constraint fusion module is used to dynamically fuse the fine-grained fluctuation features and the coarse-grained trend features based on a physical constraint gating mechanism to obtain fused features; The physical constraint adversarial prediction module is used to establish an adversarial training framework. The fused features are used as inputs to the predictor in the adversarial training framework to predict the remaining battery life and obtain an initial prediction result. The initial prediction result is optimized by the discriminator in the adversarial training framework in collaboration with physical constraints to obtain the remaining battery life prediction result.

[0024] Thirdly, the present invention also provides a computer device, comprising: a memory, a processor, and a battery remaining life prediction program stored in the memory and executable on the processor, wherein the processor, when executing the battery remaining life prediction program, implements the steps of the battery remaining life prediction method as described in any one of the preceding claims.

[0025] Fourthly, the present invention also provides a computer-readable storage medium storing a battery remaining life prediction program, which, when executed by a processor, implements the steps of the battery remaining life prediction method as described in any of the above embodiments.

[0026] Compared with existing technologies, this invention integrates physical mechanism features into preprocessed battery monitoring data, constructing dynamic correlations based on spatial topological relationships and temporal evolution patterns. This overcomes the limitations of the fragmented processing of existing technologies ("spatial separation-temporal isolation") and compensates for the lack of physical interpretability in purely data-driven models. Based on a physical constraint gating mechanism, it integrates fine-grained fluctuation features and coarse-grained trend features on demand, adapting to the complex degradation patterns of batteries, from linear decay to nonlinear abrupt changes. Based on predictor-discriminator adversarial training with physical consistency loss, it significantly reduces the cumulative error effect by constraining the initial prediction results from both data distribution and mechanistic laws. This invention achieves accurate characterization of the battery's entire lifecycle degradation trajectory, especially under actual operating conditions such as high noise, multi-parameter coupling, and variable temperature and load, exhibiting higher early warning sensitivity, long-term prediction robustness, and result reliability. Attached Figure Description

[0027] The present invention will now be described in detail with reference to the accompanying drawings. The above and other aspects of the present invention will become clearer and more readily understood through the detailed description following the accompanying drawings. In the drawings: Figure 1 This is a flowchart of the battery remaining life prediction method provided in the embodiments of the present invention; Figure 2 This is a comparison curve of the remaining life prediction results of lithium-ion batteries provided in the embodiments of the present invention; Figure 3 This is a bar chart of ablation experiments provided in an embodiment of the present invention; Figure 4 This is a schematic diagram of the battery remaining life prediction system provided in an embodiment of the present invention; Figure 5 This is a schematic diagram of the structure of a computer device provided in an embodiment of the present invention. Detailed Implementation

[0028] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0029] Example 1 Please refer to Figure 1 This invention provides a method for predicting the remaining life of a battery, the method comprising the following steps: S1. Obtain battery monitoring data based on the battery multi-parameter monitoring network, and preprocess the battery monitoring data to obtain a multi-dimensional input tensor.

[0030] In this embodiment, the battery monitoring network includes a voltage sensor, a temperature sensor, and a current sensor. The monitoring data covers the charge and discharge voltage curves of ternary lithium batteries (NMC), the temperature at each monitoring point, the charge and discharge current, the cycle capacity, and the number of cycles. The data is collected under variable temperature and variable load conditions.

[0031] In this embodiment of the invention, step S1 includes the following sub-steps: S11. Perform sliding window segmentation on the battery monitoring data, extract time series data and capacity labels, and obtain basic feature data.

[0032] Specifically, a sliding window of length 20 is used to segment the data, and the voltage time series, temperature mean series, current integral value and time series capacity label of each cycle are extracted to ensure that the window covers the complete features of a single cycle, providing comprehensive basic data support for subsequent feature extraction.

[0033] S12. The basic feature data is standardized and outlier correction is performed, and the frequency domain and time domain are decoupled by synchronous discrete wavelet transform to obtain low-frequency approximate components and high-frequency detail components.

[0034] Specifically, the standardization process employs Z-score standardization (standard deviation standardization) to eliminate dimensional and scale differences between different feature data, ensuring all features are on the same order of magnitude and preventing some feature values ​​from dominating model learning due to excessively large (or small) values. Outlier correction uses the 3σ criterion to eliminate measurement errors and extreme data, ensuring the authenticity and consistency of the data. Discrete wavelet transform is performed based on the db4 wavelet basis to decompose voltage and capacity data into low-frequency approximate components (including the long-term irreversible capacity decay trend) and high-frequency detail components (including transient changes such as load fluctuations and capacity regeneration), achieving the separation of useful signals from interference signals.

[0035] S13. Based on the simplified equivalent circuit electrochemical model, the current and voltage data of the battery are acquired in real time to obtain the physical state parameters; wherein, the physical state parameters include ohmic internal resistance, polarization voltage, rate of change of internal resistance and theoretical state of charge.

[0036] Specifically, a second-order RC equivalent circuit model (simplified electrochemical model) is adopted. Real-time current and voltage data are used to identify ohmic internal resistance, polarization resistance, rate of change of internal resistance, polarization capacitance and theoretical state of charge online, and output physical state parameter sequence to inject prior knowledge of physical mechanism into the model and make up for the "black box" defect of pure data-driven model.

[0037] S14. Stack the basic feature data, the low-frequency approximation component, the high-frequency detail component, and the physical state parameters in a dimensional manner to obtain the multidimensional input tensor.

[0038] Specifically, the standardized data, discrete wavelet transform decoupled components, and physical state parameters are stacked according to the channel dimension to form a multidimensional input tensor with dimensions of [number of samples, window length, number of feature channels]. This enables the initial fusion of data-driven features and physical mechanism features, providing rich multi-source information for subsequent spatiotemporal feature modeling.

[0039] S2. Perform spatial parameter correlation modeling and time key event detection on the multidimensional input tensor to obtain a spatiotemporal joint feature vector.

[0040] In this embodiment of the invention, the spatiotemporal joint feature vector simultaneously embeds the parameter space topological relationship and the temporal evolution law. The spatiotemporal joint feature vector integrates the mechanism prior. Step S2 includes the following sub-steps: S21. Based on the physical topology and mechanism correlation matrix of the battery multi-parameter monitoring network, establish a parameter spatial location encoding matrix, extract the spatial coupling relationship between parameters in the multidimensional input tensor through a graph convolutional neural network, and fuse it with the parameter spatial location encoding matrix to obtain a spatial weight matrix.

[0041] Specifically, the physical topology of the battery multi-parameter detection network contains important information. The physical locations of the six temperature sensors have a thermal conduction effect, and their topological relationships directly affect the judgment of the battery degradation state. Based on this physical topology, a topological adjacency matrix is ​​constructed, and combined with a learnable mechanism correlation matrix, a parameter spatial location encoding matrix is ​​generated, allowing the model to understand the relative positional relationships and physical associations between the monitoring points.

[0042] A two-layer graph convolutional neural network is used to capture the spatial coupling relationship between monitoring parameters using learnable spatial convolutional kernels, which are then fused with the position encoding matrix to obtain a 64-dimensional spatial weight matrix. The mechanism correlation matrix can quantify the causal relationship between physical parameters such as ohmic internal resistance growth and polarization voltage change and spatial anomalies such as local temperature rise and voltage imbalance, so that the model prioritizes spatial features that are strongly correlated with degradation mechanisms, thereby improving the pertinence of spatial modeling.

[0043] S22. Obtain the battery capacity decay sequence, and calculate the time attention weights for the low-frequency approximation component and the high-frequency detail component based on the multi-head attention mechanism; wherein, the low-frequency approximation component captures the global degradation trend and temporal dependency of the battery capacity decay sequence through long-range attention focusing; the high-frequency detail component identifies the time node corresponding to the short-circuit precursor in the battery through local attention focusing; and the battery charge-discharge cycle characteristics are incorporated into the battery capacity decay sequence through timestamp encoding to obtain the time attention score matrix.

[0044] Specifically, long-range dependency refers to the correlation between charge-discharge cycles that are far apart in the battery capacity decay sequence; the time nodes corresponding to the short circuit precursor in the battery include the capacity inflection point, the node where the fault precursor appears, etc., and their occurrence is the result of the cumulative effect of multiple early cycles.

[0045] A four-head attention mechanism is adopted to calculate attention weights for low-frequency approximation components (focusing on long-term trends) and high-frequency detail components (focusing on recent fluctuations) to achieve targeted attention to different frequency features. Timestamp encoding transforms the number of cycles into sinusoidal position features and incorporates them into the capacity decay sequence to more clearly understand the physical meaning of time distance, avoid misjudging short-term fluctuations as long-term trends, and obtain a 64-dimensional frequency-domain sensitive time attention score matrix to improve the accuracy of long-range dependency analysis.

[0046] S23. Perform tensor multiplication on the spatial weight matrix and the temporal attention score matrix to obtain the spatiotemporal joint feature vector.

[0047] Specifically, the interaction between the spatial weight matrix and the temporal attention score matrix is ​​realized through tensor dot product, and the output is a spatiotemporal joint feature vector with dimensions of [number of samples, window length, 64]. This feature contains not only spatial parameter correlation information, but also incorporates temporal key event features and physical mechanism priors, thus possessing stronger discriminative power and reliability.

[0048] S3. Perform frequency domain decoupling multi-scale analysis on the spatiotemporal joint feature vector to extract fine-grained fluctuation features and coarse-grained trend features respectively.

[0049] In this embodiment of the invention, the coarse-grained trend feature is used to characterize the long-term degradation baseline of the battery, and the fine-grained fluctuation feature is used to capture local transient anomalies. Step S3 includes the following sub-steps: S31. Establish a trend feature extraction model based on dilated convolution and causal long short-term memory network. Extract features from the low-frequency approximation components and physical state parameters in the spatiotemporal joint feature vector through the trend feature extraction model to obtain the coarse-grained trend features.

[0050] Specifically, this branch is a trend-following branch, consisting of two layers of dilated convolutions and one layer of causal long short-term memory network. The dilated convolutions form a hierarchical receptive field by stacking multiple convolutional layers with different dilation factors, covering a wider time range. The causal convolutions serve as a preprocessing layer, providing the long short-term memory network with feature sequences that conform to temporal causality.

[0051] This branch takes low-frequency components and physical parameters from the spatiotemporal joint features as input, removes high-frequency noise interference, smoothly fits the physical baseline of capacity decay, and outputs 64-dimensional coarse-grained trend features to capture global degradation trends across cycles (such as the capacity decay slope every 100 cycles).

[0052] S32. A fluctuation feature extraction model is established based on a one-dimensional convolutional and bidirectional long short-term memory network. The high-frequency detail components in the spatiotemporal joint feature vector are extracted using the fluctuation feature extraction model to obtain the fine-grained fluctuation features.

[0053] Specifically, this branch is a fluctuation capture branch, consisting of two layers of high-resolution one-dimensional convolution and one layer of bidirectional long short-term memory network. It uses high temporal resolution one-dimensional dilated convolution to expand the receptive field and capture the weak fluctuations implied by early capacity decay in the spatiotemporal joint feature vector (such as subtle voltage perturbations caused by lithium deposition). At the same time, it uses the bidirectional long short-term memory network to analyze the positive and negative dependencies of local sequences, specifically extracting transient signals such as capacity regeneration and subtle voltage perturbations, and outputting 64-dimensional fine-grained fluctuation features.

[0054] S4. Based on the physical constraint gating mechanism, the fine-grained fluctuation features and the coarse-grained trend features are dynamically fused to obtain fused features.

[0055] In this embodiment of the invention, step S4 includes the following sub-steps: S41. Using the internal resistance change rate in the physical state parameters as a gating signal, the contribution weights of the coarse-grained trend feature and the fine-grained fluctuation feature are dynamically adjusted through the physical constraint gating mechanism, and feature fusion is performed to obtain a fused feature. The magnitude of the internal resistance change rate directly maps to the degradation stage of the battery: when the internal resistance change rate is below a preset threshold, it is determined to be a smooth degradation period, and the weight of the coarse-grained trend feature is increased; when the internal resistance change rate is higher than or equal to the preset threshold, it is determined to be an accelerated degradation period, and the weight of the fine-grained fluctuation feature is increased.

[0056] Specifically, temporal interpolation and alignment processing is used to solve the problems of temporal misalignment and missing features in the two branches, ensuring that the sequences are comparable on the time axis; spatial projection transformation processing is used to solve the problems of spatial correlation ambiguity and dimensional redundancy, and to enhance the response to key spatial patterns. Through these two processes, data format differences are eliminated and key feature correlations are strengthened, providing more standardized and easier-to-learn inputs for subsequent fusion.

[0057] Specifically, the gating signal uses the rate of change of ohmic internal resistance (Δohmic internal resistance / Δt), a physical parameter that can accurately reflect the battery degradation stage: when the rate of change is less than 0.01Ω / cycle (the period of gradual capacity decay and stable physical parameters), the coarse-grained feature weight is set to 0.7 and the fine-grained feature weight is set to 0.3, focusing on guiding long-term trends; when the rate of change is greater than or equal to 0.01Ω / cycle (the period of abnormal degradation and abrupt changes in physical parameters), it is automatically adjusted to a coarse-grained weight of 0.3 and a fine-grained weight of 0.7, focusing on capturing transient anomalies. A 64-dimensional fusion feature is obtained through weighted summation, achieving synergistic optimization of "trend guidance + anomaly focusing".

[0058] S5. Establish an adversarial training framework, use the fused features as input to the predictor in the adversarial training framework to predict the remaining battery life, and obtain an initial prediction result; optimize the initial prediction result by cooperating with the discriminator and physical constraints in the adversarial training framework to obtain the remaining battery life prediction result.

[0059] In this embodiment of the invention, adversarial loss is calculated by the discriminator in the adversarial training framework, and physical consistency loss is constructed by combining electrochemical mechanisms. These two losses, along with the mean squared error loss, constitute a joint loss function. The initial prediction result is iteratively optimized using backpropagation of the joint loss to obtain the battery remaining life prediction result. Step S5 includes the following sub-steps: S51. Construct the adversarial training framework; wherein, the predictor in the adversarial training framework adopts a temporal convolutional layer stacking structure; the discriminator in the adversarial training framework adopts a one-dimensional spectral normalized convolutional network. S52. Use the fused features as input to the predictor to predict the remaining battery life and obtain an initial prediction result; S53. Establish a joint loss function, and optimize the initial prediction result using the joint loss function and the discriminator to obtain the battery remaining life prediction result. Specifically, calculate the adversarial loss based on the discriminator output. Combined with physical consistency loss Construct a joint loss function, use backpropagation based on the joint loss function to update the model parameters, correct the initial prediction results, and output the final battery remaining life prediction results.

[0060] The joint loss function satisfies the following condition: ; in, Denotes the joint loss function, Mean squared error loss is used to measure the deviation between the predicted value and the actual remaining lifetime. To combat the loss, cross-entropy loss is used to constrain the statistical distribution of the prediction results to be consistent with the actual data; To mitigate the loss of physical consistency, we must ensure, from a mechanistic perspective, that the prediction results conform to electrochemical laws. , The weighting coefficient is adjusted according to the operating conditions in this embodiment. , Physical consistency loss In this model, monotonicity constraints penalize positive capacity growth except during capacity regeneration, ensuring the predicted trend conforms to the physical laws of battery degradation. Boundary constraints, based on the Arrhenius equation, set an upper limit of the decay rate at 0.02 Ah / cycle, penalizing predicted values ​​exceeding the physical limit. The discriminator only participates in calculating adversarial loss. This is used to constrain the predicted lifetime distribution to be consistent with the actual sample distribution; physical consistency loss. The algorithm is independently constructed based on the electrochemical degradation law of batteries and does not rely on the discriminator output; the two losses are weighted by coefficients. After weighting, compared with the basic regression loss The total loss is obtained by summing the results, achieving dual collaborative optimization of data distribution constraints and physical mechanism constraints.

[0061] Regarding the joint loss function mentioned above, the adversarial loss... loss of physical consistency The expressions are as follows:

[0062] in, Total sample size; For the first Battery pack's true remaining life label; The initial predicted value output by the predictor; The true / false probability of the discriminator network output is represented by a binary classification cross-entropy construct to construct an adversarial loss, thus bringing the predicted results closer to the data distribution of the actual lifetime.

[0063] The physical consistency loss is a weighted composite of the monotonicity loss and the decay rate boundary loss.

[0064] in, For sub-loss weighting coefficients; The capacity monotonicity loss is expressed as: (in, The time series length for a single sample; For the first Circular prediction capacity; The indicator function is used only if the current loop does not belong to the capacity regeneration set. (This takes effect immediately, penalizing capacity increases during non-regeneration phases to ensure the overall capacity decay trend conforms to electrochemical laws.) This is the decay rate boundary constraint loss. Wherein, The interval between adjacent cycles; To mitigate predictions that violate physical limits, a penalty is applied when the predicted capacity decay rate exceeds the maximum decay rate defined by the Arrhenius formula. The model is trained iteratively for 200 epochs using the Adam optimizer, and the model parameters are optimized through backpropagation using a joint loss function to continuously correct prediction biases, ultimately outputting stable remaining lifetime predictions.

[0065] This invention deeply couples physical model-driven state parameter extraction, frequency domain decoupled multi-scale feature extraction, physical parameter-based gated adaptive fusion, and adversarial training embedded with physical laws, forming a complete technical path of "signal decoupling, feature modeling, weight allocation, and result constraint." This architecture is specifically designed to solve the technical challenge of existing technologies in simultaneously tracking long-term trends and capturing early, subtle signs under complex operating conditions.

[0066] This approach achieves a synergistic effect among multiple technical features: frequency domain decoupling provides a prerequisite for differentiated processing in multi-head attention mechanisms; specialized multi-scale extraction branches provide targeted inputs for physical constraint gating fusion; and physical consistency loss imposes mechanistic constraints on the output of the adversarial training framework. These features work together to enable the model to dynamically adjust its focus based on battery degradation stages, achieving adaptive prediction of the entire battery lifecycle and producing unexpected technical effects that cannot be achieved by a single technical feature.

[0067] This invention successfully extracts early, subtle signs of battery degradation from strong noise through frequency domain decoupling and a dedicated fluctuation branch. Simultaneously, by using physical state parameters as guidance and physical consistency loss constraints, the prediction results have a clear electrochemical mechanism basis, effectively solving the problems of "black box" operation and unreliable results in purely data-driven models in existing technologies. In this embodiment, to better understand and verify the effectiveness of the proposed battery remaining life prediction method, this embodiment uses the NASA battery dataset and the CALCE battery dataset for multi-dimensional testing, as detailed below: Please refer to Figure 2 , Figure 2 This is a comparison curve of the remaining life prediction results of lithium-ion batteries provided in the embodiments of the present invention; the method of the present invention is compared with the traditional CNN-LSTM model, the simple spatiotemporal attention model, and the data-driven model without physical constraints. The root mean square error (RMSE) and mean absolute error (MAE) are used as evaluation metrics in the experiment. The performance comparison is shown in Table 1 below: Table 1: Performance comparison of different prediction methods on the NASA dataset

[0068] Experimental data show that the method of this invention significantly outperforms the comparative scheme in prediction accuracy, with substantial reductions in both RMSE and MAE. Particularly noteworthy is the improvement in the physical consistency violation rate, thanks to the superior physical consistency loss function. Due to the constraints, the prediction curve generated by this invention strictly follows the monotonicity of capacity decay, almost eliminating the phenomenon of sudden capacity increases that violate physical laws. Specifically, the physical consistency violation rate of this invention is only 0.2%, far superior to the 18.4% and 12.1% of the comparative schemes, achieving a balance between high accuracy and high physical reliability.

[0069] Please also refer to Figure 3 , Figure 3 This is a bar chart of ablation experiments provided in this embodiment of the invention; to verify the contribution of each innovative module to the system performance, ablation experiments were conducted, and the experimental results are shown in Table 2: Table 2: Ablation Experiment Results of the Core Module of the Invention

[0070] Note: S3 (frequency domain multi-scale analysis) and S4 (physical gating fusion) have been integrated into the overall verification of the "Experiment B → Invention" stage and cannot be separated separately. As shown in the table above, this invention has successfully achieved a paradigm shift from multi-scale data fitting to accurate prediction driven by both physical and data aspects by deeply empowering the physical mechanism throughout the entire process of input, feature extraction, and optimization. This significantly improves the reliability and interpretability of the prediction results under complex working conditions.

[0071] Compared with existing technologies, this invention integrates physical mechanism features into preprocessed battery monitoring data, constructing dynamic correlations based on spatial topological relationships and temporal evolution patterns. This overcomes the limitations of the fragmented processing of existing technologies ("spatial separation-temporal isolation") and compensates for the lack of physical interpretability in purely data-driven models. Based on a physical constraint gating mechanism, it integrates fine-grained fluctuation features and coarse-grained trend features on demand, adapting to the complex degradation patterns of batteries, from linear decay to nonlinear abrupt changes. Based on predictor-discriminator adversarial training with physical consistency loss, it significantly reduces the cumulative error effect by constraining the initial prediction results from both data distribution and mechanistic laws. This invention achieves accurate characterization of the battery's entire lifecycle degradation trajectory, especially under actual operating conditions such as high noise, multi-parameter coupling, and variable temperature and load, exhibiting higher early warning sensitivity, long-term prediction robustness, and result reliability.

[0072] Example 2 This invention also provides a battery remaining life prediction system, please refer to... Figure 4 , Figure 4 This is a schematic diagram of the battery remaining life prediction system 200 provided in an embodiment of the present invention, which includes: 201. Physical enhancement preprocessing module, used to acquire battery monitoring data based on battery multi-parameter monitoring network, preprocess the battery monitoring data to obtain multidimensional input tensors; 202. A physical perception spatiotemporal module is used to perform spatial parameter correlation modeling and key temporal event detection on the multidimensional input tensor to obtain a spatiotemporal joint feature vector; 203. Frequency domain decoupling multi-scale module, used to perform frequency domain decoupling multi-scale analysis on the spatiotemporal joint feature vector, and extract fine-grained fluctuation features and coarse-grained trend features respectively; 204. Physical constraint fusion module, used to dynamically fuse the fine-grained fluctuation features and the coarse-grained trend features based on a physical constraint gating mechanism to obtain fused features; 205. A physical constraint adversarial prediction module is used to establish an adversarial training framework, use the fused features as input to the predictor in the adversarial training framework to predict the remaining battery life, and obtain an initial prediction result; through the discriminator in the adversarial training framework and the physical constraints, the initial prediction result is optimized to obtain the remaining battery life prediction result.

[0073] The battery remaining life prediction system 200 can implement the steps in the battery remaining life prediction method in the above embodiments and achieve the same technical effect, as described in the above embodiments, and will not be repeated here.

[0074] Example 3 This invention also provides a computer device, please refer to... Figure 5 , Figure 5 This is a schematic diagram of the structure of a computer device provided in an embodiment of the present invention. The computer device 300 includes: a memory 302, a processor 301, and a battery remaining life prediction program stored in the memory 302 and capable of running on the processor 301.

[0075] The processor 301 calls the battery remaining life prediction program stored in the memory 302 and executes the steps in the battery remaining life prediction method provided in this embodiment of the invention. Please refer to... Figure 1 Specifically, it includes the following steps: S1. Obtain battery monitoring data based on the battery multi-parameter monitoring network, and preprocess the battery monitoring data to obtain a multi-dimensional input tensor.

[0076] In this embodiment, the battery monitoring network includes a voltage sensor, a temperature sensor, and a current sensor. The monitoring data covers the charge and discharge voltage curves of ternary lithium batteries (NMC), the temperature at each monitoring point, the charge and discharge current, the cycle capacity, and the number of cycles. The data is collected under variable temperature and variable load conditions.

[0077] In this embodiment of the invention, step S1 includes the following sub-steps: S11. Perform sliding window segmentation on the battery monitoring data, extract time series data and capacity labels, and obtain basic feature data.

[0078] Specifically, a sliding window of length 20 is used to segment the data, and the voltage time series, temperature mean series, current integral value and time series capacity label of each cycle are extracted to ensure that the window covers the complete features of a single cycle, providing comprehensive basic data support for subsequent feature extraction.

[0079] S12. The basic feature data is standardized and outlier correction is performed, and the frequency domain and time domain are decoupled by synchronous discrete wavelet transform to obtain low-frequency approximate components and high-frequency detail components.

[0080] Specifically, the standardization process employs Z-score standardization (standard deviation standardization) to eliminate dimensional and scale differences between different feature data, ensuring all features are on the same order of magnitude and preventing some feature values ​​from dominating model learning due to excessively large (or small) values. Outlier correction uses the 3σ criterion to eliminate measurement errors and extreme data, ensuring the authenticity and consistency of the data. Discrete wavelet transform is performed based on the db4 wavelet basis to decompose voltage and capacity data into low-frequency approximate components (including the long-term irreversible capacity decay trend) and high-frequency detail components (including transient changes such as load fluctuations and capacity regeneration), achieving the separation of useful signals from interference signals.

[0081] S13. Based on the simplified equivalent circuit electrochemical model, the current and voltage data of the battery are acquired in real time to obtain the physical state parameters; wherein, the physical state parameters include ohmic internal resistance, polarization voltage, rate of change of internal resistance and theoretical state of charge.

[0082] Specifically, a second-order RC equivalent circuit model (simplified electrochemical model) is adopted. Real-time current and voltage data are used to identify ohmic internal resistance, polarization resistance, rate of change of internal resistance, polarization capacitance and theoretical state of charge online, and output physical state parameter sequence to inject prior knowledge of physical mechanism into the model and make up for the "black box" defect of pure data-driven model.

[0083] S14. Stack the basic feature data, the low-frequency approximation component, the high-frequency detail component, and the physical state parameters in a dimensional manner to obtain the multidimensional input tensor.

[0084] Specifically, the standardized data, discrete wavelet transform decoupled components, and physical state parameters are stacked according to the channel dimension to form a multidimensional input tensor with dimensions of [number of samples, window length, number of feature channels]. This enables the initial fusion of data-driven features and physical mechanism features, providing rich multi-source information for subsequent spatiotemporal feature modeling.

[0085] S2. Perform spatial parameter correlation modeling and time key event detection on the multidimensional input tensor to obtain a spatiotemporal joint feature vector.

[0086] In this embodiment of the invention, the spatiotemporal joint feature vector incorporates prior mechanistic knowledge, and step S2 includes the following sub-steps: S21. Based on the physical topology and mechanism correlation matrix of the battery multi-parameter monitoring network, establish a parameter spatial location encoding matrix, extract the spatial coupling relationship between parameters in the multidimensional input tensor through a graph convolutional neural network, and fuse it with the parameter spatial location encoding matrix to obtain a spatial weight matrix.

[0087] Specifically, the physical topology of the battery multi-parameter detection network contains important information. The physical locations of the six temperature sensors have a thermal conduction effect, and their topological relationships directly affect the judgment of the battery degradation state. Based on this physical topology, a topological adjacency matrix is ​​constructed, and combined with a learnable mechanism correlation matrix, a parameter spatial location encoding matrix is ​​generated, allowing the model to understand the relative positional relationships and physical associations between the monitoring points.

[0088] A two-layer graph convolutional neural network is used to capture the spatial coupling relationship between monitoring parameters using learnable spatial convolutional kernels, which are then fused with the position encoding matrix to obtain a 64-dimensional spatial weight matrix. The mechanism correlation matrix can quantify the causal relationship between physical parameters such as ohmic internal resistance growth and polarization voltage change and spatial anomalies such as local temperature rise and voltage imbalance, so that the model prioritizes spatial features that are strongly correlated with degradation mechanisms, thereby improving the pertinence of spatial modeling.

[0089] S22. Obtain the battery capacity decay sequence, and calculate the time attention weights for the low-frequency approximation component and the high-frequency detail component based on the multi-head attention mechanism; wherein, the low-frequency approximation component captures the global degradation trend and temporal dependency of the battery capacity decay sequence through long-range attention focusing; the high-frequency detail component identifies the time node corresponding to the short-circuit precursor in the battery through local attention focusing; and the battery charge-discharge cycle characteristics are incorporated into the battery capacity decay sequence through timestamp encoding to obtain the time attention score matrix.

[0090] Specifically, long-range dependency refers to the correlation between charge-discharge cycles that are far apart in the battery capacity decay sequence; the time nodes corresponding to the short circuit precursor in the battery include the capacity inflection point, the node where the fault precursor appears, etc., and their occurrence is the result of the cumulative effect of multiple early cycles.

[0091] A four-head attention mechanism is adopted to calculate attention weights for low-frequency approximation components (focusing on long-term trends) and high-frequency detail components (focusing on recent fluctuations) to achieve targeted attention to different frequency features. Timestamp encoding transforms the number of cycles into sinusoidal position features and incorporates them into the capacity decay sequence to more clearly understand the physical meaning of time distance, avoid misjudging short-term fluctuations as long-term trends, and obtain a 64-dimensional frequency-domain sensitive time attention score matrix to improve the accuracy of long-range dependency analysis.

[0092] S23. Perform tensor multiplication on the spatial weight matrix and the temporal attention score matrix to obtain the spatiotemporal joint feature vector.

[0093] Specifically, the interaction between the spatial weight matrix and the temporal attention score matrix is ​​realized through tensor dot product, and the output is a spatiotemporal joint feature vector with dimensions of [number of samples, window length, 64]. This feature contains not only spatial parameter correlation information, but also incorporates temporal key event features and physical mechanism priors, thus possessing stronger discriminative power and reliability.

[0094] S3. Perform frequency domain decoupling multi-scale analysis on the spatiotemporal joint feature vector to extract fine-grained fluctuation features and coarse-grained trend features respectively.

[0095] In this embodiment of the invention, step S3 includes the following sub-steps: S31. Establish a trend feature extraction model based on dilated convolution and causal long short-term memory network. Extract features from the low-frequency approximation components and physical state parameters in the spatiotemporal joint feature vector through the trend feature extraction model to obtain the coarse-grained trend features.

[0096] Specifically, this branch is a trend-following branch, consisting of two layers of dilated convolutions and one layer of causal long short-term memory network. The dilated convolutions form a hierarchical receptive field by stacking multiple convolutional layers with different dilation factors, covering a wider time range. The causal convolutions serve as a preprocessing layer, providing the long short-term memory network with feature sequences that conform to temporal causality.

[0097] This branch takes low-frequency components and physical parameters from the spatiotemporal joint features as input, removes high-frequency noise interference, smoothly fits the physical baseline of capacity decay, and outputs 64-dimensional coarse-grained trend features to capture global degradation trends across cycles (such as the capacity decay slope every 100 cycles).

[0098] S32. A fluctuation feature extraction model is established based on a one-dimensional convolutional and bidirectional long short-term memory network. The high-frequency detail components in the spatiotemporal joint feature vector are extracted using the fluctuation feature extraction model to obtain the fine-grained fluctuation features.

[0099] Specifically, this branch is a fluctuation capture branch, consisting of two layers of high-resolution one-dimensional convolution and one layer of bidirectional long short-term memory network. It uses high temporal resolution one-dimensional dilated convolution to expand the receptive field and capture the weak fluctuations implied by early capacity decay in the spatiotemporal joint feature vector (such as subtle voltage perturbations caused by lithium deposition). At the same time, it uses the bidirectional long short-term memory network to analyze the positive and negative dependencies of local sequences, specifically extracting transient signals such as capacity regeneration and subtle voltage perturbations, and outputting 64-dimensional fine-grained fluctuation features.

[0100] S4. Based on the physical constraint gating mechanism, the fine-grained fluctuation features and the coarse-grained trend features are dynamically fused to obtain fused features.

[0101] In this embodiment of the invention, step S4 includes the following sub-steps: S41. Using the rate of change of internal resistance in the physical state parameters as a gating signal, the contribution weights of the coarse-grained trend feature and the fine-grained fluctuation feature are dynamically adjusted through the physical constraint gating mechanism, and feature fusion is performed to obtain the fused feature.

[0102] Specifically, temporal interpolation and alignment processing is used to solve the problems of temporal misalignment and missing features in the two branches, ensuring that the sequences are comparable on the time axis; spatial projection transformation processing is used to solve the problems of spatial correlation ambiguity and dimensional redundancy, and to enhance the response to key spatial patterns. Through these two processes, data format differences are eliminated and key feature correlations are strengthened, providing more standardized and easier-to-learn inputs for subsequent fusion.

[0103] Specifically, the gating signal uses the rate of change of ohmic internal resistance (Δohmic internal resistance / Δt), a physical parameter that can accurately reflect the battery degradation stage: when the rate of change is less than 0.01Ω / cycle (the period of gradual capacity decay and stable physical parameters), the coarse-grained feature weight is set to 0.7 and the fine-grained feature weight is set to 0.3, focusing on guiding long-term trends; when the rate of change is greater than or equal to 0.01Ω / cycle (the period of abnormal degradation and abrupt changes in physical parameters), it is automatically adjusted to a coarse-grained weight of 0.3 and a fine-grained weight of 0.7, focusing on capturing transient anomalies. A 64-dimensional fusion feature is obtained through weighted summation, achieving synergistic optimization of "trend guidance + anomaly focusing".

[0104] S5. Establish an adversarial training framework, use the fused features as input to the predictor in the adversarial training framework to predict the remaining battery life, and obtain an initial prediction result; optimize the initial prediction result by cooperating with the discriminator and physical constraints in the adversarial training framework to obtain the remaining battery life prediction result.

[0105] In this embodiment of the invention, adversarial loss is calculated by the discriminator in the adversarial training framework, and physical consistency loss is constructed by combining electrochemical mechanisms. These two losses, along with the mean squared error loss, constitute a joint loss function. The initial prediction result is iteratively optimized using backpropagation of the joint loss to obtain the battery remaining life prediction result. Step S5 includes the following sub-steps: S51. Construct the adversarial training framework; wherein, the predictor in the adversarial training framework adopts a temporal convolutional layer stacking structure; the discriminator in the adversarial training framework adopts a one-dimensional spectral normalized convolutional network. S52. Use the fused features as input to the predictor to predict the remaining battery life and obtain an initial prediction result; S53. Establish a joint loss function, and optimize the initial prediction result using the joint loss function and the discriminator to obtain the battery remaining life prediction result.

[0106] The joint loss function satisfies the following condition: ; in, Denotes the joint loss function, Mean squared error loss is used to measure the deviation between the predicted value and the actual remaining lifetime. To combat the loss, cross-entropy loss is used to constrain the statistical distribution of the prediction results to be consistent with the actual data; To mitigate the loss of physical consistency, we must ensure, from a mechanistic perspective, that the prediction results conform to electrochemical laws. , The weighting coefficient is adjusted according to the operating conditions in this embodiment. , Physical consistency loss In this model, monotonicity constraints penalize positive capacity growth except during capacity regeneration, ensuring the predicted trend conforms to the physical laws of battery degradation. Boundary constraints, based on the Arrhenius equation, set an upper limit of the decay rate at 0.02 Ah / cycle, penalizing predicted values ​​exceeding the physical limit. The discriminator only participates in calculating adversarial loss. This is used to constrain the predicted lifetime distribution to be consistent with the actual sample distribution; physical consistency loss. The algorithm is independently constructed based on the electrochemical degradation law of batteries and does not rely on the discriminator output; the two losses are weighted by coefficients. After weighting, compared with the basic regression loss The total loss is obtained by summing the results, achieving dual collaborative optimization of data distribution constraints and physical mechanism constraints.

[0107] Regarding the joint loss function mentioned above, the adversarial loss... loss of physical consistency The expressions are as follows: ; in, Total sample size; For the first Battery pack's true remaining life label; The initial predicted value output by the predictor; The true / false probability of the discriminator network output is represented by a binary classification cross-entropy construct to construct an adversarial loss, thus bringing the predicted results closer to the data distribution of the actual lifetime.

[0108] The physical consistency loss is a weighted composite of the monotonicity loss and the decay rate boundary loss. ; in, For sub-loss weighting coefficients; The capacity monotonicity loss is expressed as: (in, The time series length for a single sample; For the first Circular prediction capacity; The indicator function is used only if the current loop does not belong to the capacity regeneration set. (This takes effect immediately, penalizing capacity increases during non-regeneration phases to ensure the overall capacity decay trend conforms to electrochemical laws.) This is the decay rate boundary constraint loss. Wherein, The interval between adjacent cycles; The maximum decay rate is defined by the Arrhenius formula. When the predicted capacity decay rate exceeds this threshold, a penalty is imposed to avoid predictions that violate physical limits.

[0109] By using the Adam optimizer for 200 iterations of training and optimizing the model parameters through backpropagation of the joint loss function, prediction bias is continuously corrected, and finally a stable remaining lifetime prediction result is output.

[0110] The computer device 300 provided in this embodiment of the invention can implement the steps in the battery remaining life prediction method as described in the above embodiments, and can achieve the same technical effect. Refer to the description in the above embodiments, which will not be repeated here.

[0111] Example 4 This invention also provides a computer-readable storage medium storing a battery remaining life prediction program. When executed by a processor, the battery remaining life prediction program implements the various processes and steps in the battery remaining life prediction method provided in this invention and achieves the same technical effect. To avoid repetition, it will not be described again here.

[0112] Those skilled in the art will understand that all or part of the processes in the above embodiments can be implemented by a computer program instructing related hardware. The program can be stored in a computer-readable storage medium, and when executed, it can include the processes of the embodiments of the above methods. The storage medium can be a magnetic disk, optical disk, read-only memory (ROM), or random access memory (RAM), etc.

[0113] It should be noted that, in this document, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Unless otherwise specified, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes that element.

[0114] Through the above description of the embodiments, those skilled in the art can clearly understand that the methods of the above embodiments can be implemented by means of software plus necessary general-purpose hardware platforms. Of course, they can also be implemented by hardware, but in many cases the former is a better implementation method. Based on this understanding, the technical solution of the present invention, or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product is stored in a storage medium (such as ROM / RAM, magnetic disk, optical disk) and includes several instructions to cause a terminal (which may be a mobile phone, computer, server, air conditioner, or network device, etc.) to execute the methods described in the various embodiments of the present invention.

[0115] The embodiments of the present invention have been described above with reference to the accompanying drawings. The disclosed embodiments are merely preferred embodiments of the present invention. However, the present invention is not limited to the specific embodiments described above. The specific embodiments described above are merely illustrative and not restrictive. Those skilled in the art can make many equivalent changes in form without departing from the spirit and scope of the claims of the present invention, and all such changes are within the protection scope of the present invention.

Claims

1. A method for predicting the remaining life of a battery, characterized in that, The battery remaining life prediction method includes the following steps: S1. Obtain battery monitoring data based on the battery multi-parameter monitoring network, and preprocess the battery monitoring data to obtain a multi-dimensional input tensor; S2. Perform spatial parameter correlation modeling and time key event detection on the multidimensional input tensor to obtain a spatiotemporal joint feature vector; S3. Perform frequency domain decoupling multi-scale analysis on the spatiotemporal joint feature vector to extract fine-grained fluctuation features and coarse-grained trend features respectively; S4. Based on a physical constraint gating mechanism, the fine-grained fluctuation features and the coarse-grained trend features are dynamically fused to obtain fused features; S5. Establish an adversarial training framework, use the fused features as input to the predictor in the adversarial training framework to predict the remaining battery life, and obtain an initial prediction result; optimize the initial prediction result by cooperating with the discriminator and physical constraints in the adversarial training framework to obtain the remaining battery life prediction result.

2. The battery remaining life prediction method as described in claim 1, characterized in that, Step S1 includes the following sub-steps: S11. Perform sliding window segmentation on the battery monitoring data, extract time-series data and capacity labels, and obtain basic feature data; S12. Standardize and correct outlier data for the basic feature data, and decouple frequency and time domains by synchronous discrete wavelet transform to obtain low-frequency approximate components and high-frequency detail components. S13. Based on a simplified equivalent circuit electrochemical model, the current and voltage data of the battery are acquired in real time to obtain physical state parameters; wherein, the physical state parameters include ohmic internal resistance, polarization voltage, rate of change of internal resistance, and theoretical state of charge. S14. Stack the basic feature data, the low-frequency approximation component, the high-frequency detail component, and the physical state parameters in a dimensional manner to obtain the multidimensional input tensor.

3. The battery remaining life prediction method as described in claim 2, characterized in that, Step S2 includes the following sub-steps: S21. Based on the physical topology and mechanism correlation matrix of the battery multi-parameter monitoring network, establish a parameter spatial location encoding matrix, extract the spatial coupling relationship between parameters in the multidimensional input tensor through a graph convolutional neural network, and fuse it with the parameter spatial location encoding matrix to obtain a spatial weight matrix. S22. Obtain the battery capacity decay sequence, and calculate the time attention weights for the low-frequency approximation component and the high-frequency detail component based on the multi-head attention mechanism; wherein, the low-frequency approximation component captures the global degradation trend and temporal dependency of the battery capacity decay sequence through long-range attention focusing; the high-frequency detail component identifies the time node corresponding to the short-circuit precursor in the battery through local attention focusing; and the battery charge-discharge cycle characteristics are incorporated into the battery capacity decay sequence through timestamp encoding to obtain the time attention score matrix; S23. Perform tensor multiplication on the spatial weight matrix and the temporal attention score matrix to obtain the spatiotemporal joint feature vector.

4. The battery remaining life prediction method as described in claim 3, characterized in that, Step S3 includes the following sub-steps: S31. A trend feature extraction model is established based on dilated convolution and causal long short-term memory network. The low-frequency approximation component and physical state parameter in the spatiotemporal joint feature vector are extracted by the trend feature extraction model to obtain the coarse-grained trend feature. S32. A fluctuation feature extraction model is established based on a one-dimensional convolutional and bidirectional long short-term memory network. The high-frequency detail components in the spatiotemporal joint feature vector are extracted using the fluctuation feature extraction model to obtain the fine-grained fluctuation features.

5. The battery remaining life prediction method as described in claim 4, characterized in that, Step S4 includes the following sub-steps: S41. Using the rate of change of internal resistance in the physical state parameters as a gating signal, the contribution weights of the coarse-grained trend feature and the fine-grained fluctuation feature are dynamically adjusted through the physical constraint gating mechanism, and feature fusion is performed to obtain the fused feature.

6. The battery remaining life prediction method as described in claim 1, characterized in that, Step S5 includes the following sub-steps: S51. Construct the adversarial training framework; wherein, the predictor in the adversarial training framework adopts a temporal convolutional layer stacking structure; the discriminator in the adversarial training framework adopts a one-dimensional spectral normalized convolutional network. S52. Use the fused features as input to the predictor to predict the remaining battery life and obtain an initial prediction result; S53. Establish a joint loss function, and optimize the initial prediction result using the joint loss function and the discriminator to obtain the battery remaining life prediction result.

7. The battery remaining life prediction method as described in claim 6, characterized in that, The joint loss function satisfies the following condition: ； in, Denotes the joint loss function, This represents the mean squared error loss. Indicating resistance to loss, Indicates the loss of physical consistency. and All of these represent weighting coefficients.

8. A battery remaining life prediction system, characterized in that, The battery remaining life prediction system includes: The physical enhancement preprocessing module is used to acquire battery monitoring data based on the battery multi-parameter monitoring network, preprocess the battery monitoring data, and obtain a multi-dimensional input tensor. The physical perception spatiotemporal module is used to perform spatial parameter correlation modeling and key temporal event detection on the multidimensional input tensor to obtain a spatiotemporal joint feature vector. The frequency domain decoupled multi-scale module is used to perform frequency domain decoupled multi-scale analysis on the spatiotemporal joint feature vector, and extract fine-grained fluctuation features and coarse-grained trend features respectively. The physical constraint fusion module is used to dynamically fuse the fine-grained fluctuation features and the coarse-grained trend features based on a physical constraint gating mechanism to obtain fused features; The physical constraint adversarial prediction module is used to establish an adversarial training framework. The fused features are used as inputs to the predictor in the adversarial training framework to predict the remaining battery life and obtain an initial prediction result. The initial prediction result is optimized by the discriminator in the adversarial training framework in collaboration with physical constraints to obtain the remaining battery life prediction result.

9. A computer device, characterized in that, include: The system includes a memory, a processor, and a battery remaining life prediction program stored in the memory and executable on the processor, wherein the processor, when executing the battery remaining life prediction program, implements the steps of the battery remaining life prediction method as described in any one of claims 1-7.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a battery remaining life prediction program, which, when executed by a processor, implements the steps of the battery remaining life prediction method as described in any one of claims 1-7.

Images