Lithium ion battery time series data cleaning method, system and storage medium

Abstract

The present application relates to a kind of lithium ion battery timing data cleaning method, system and storage medium, the method is aimed at the physical characteristic distortion problem caused by sensor noise and communication packet loss of battery timing data, first by embedding Butler-Volmer kinetics equation and diffusion model of mass conservation law, noisy or missing data is reconstructed reversely, strictly follow electrochemical reaction mechanism to generate data, then by constructing cross-scale spatiotemporal feature pyramid, utilize inflation causal convolution and time-varying graph network synchronous capture millisecond-level transient fluctuation and cycle-level degradation trend, and solve multiscale feature fragmentation problem by the adaptive attention mechanism of capacity jump period.The present application ensures that the data after cleaning has high fidelity and physical interpretability by the complete scheme from data denoising to mechanism reconstruction.

Description

Technical Field

[0001] This invention relates to the field of data processing technology, and in particular to a method, system and storage medium for cleaning time-series data of lithium-ion batteries. Background Technology

[0002] Accurate assessment of the state of health (SOH) of lithium-ion batteries is crucial for ensuring safe operation and extending battery life, relying on in-depth analysis of multi-dimensional sensor data such as voltage, current, and temperature. However, in practical applications, data quality defects and the challenge of cross-scale feature extraction severely restrict assessment accuracy, and existing technologies still have significant limitations.

[0003] Regarding data quality, sensor noise and long-term data gaps are the main challenges. Sensor noise often masks subtle changes within the battery; for example, current noise can accumulate and amplify errors in state-of-charge (SOC) estimation, and temperature noise may misjudge precursors to thermal runaway. Traditional filtering methods tend to simultaneously filter out valuable dynamic features. For long-term data gaps, traditional restoration methods such as linear interpolation and polynomial fitting have significant drawbacks: linear or polynomial interpolation fills gaps based solely on mathematical smoothness, easily creating false capacity rebounds that violate thermodynamic laws or non-physical voltage abrupt changes in the battery capacity decay trajectory. While standard Generative Adversarial Networks (GANs) or Denoising Diffusion Probabilistic Models (DDPMs) can generate realistic data distributions, the voltage-current curves they generate often violate Ohm's law or the conservation of lithium-ion mass (e.g., voltage fluctuations when there is no current input), rendering the data unusable for subsequent electrochemical simulations.

[0004] In terms of cross-scale feature extraction, existing technologies struggle to effectively bridge the gap between millisecond-level dynamic responses and cycle-level degradation trends. Battery aging is a process where the accumulation of millisecond-level electrochemical reactions (such as lithium plating) leads to macroscopic capacity decay (months, years). Existing methods typically treat these two types of features separately, failing to establish a causal mapping between microscopic transient anomalies (such as voltage spikes) and macroscopic lifespan drops, resulting in delayed early warnings of early failures. In summary, existing technologies have significant shortcomings in ensuring physical consistency of data repair and mining the correlation of cross-scale features, and a new technical solution is urgently needed to address these issues. Summary of the Invention

[0005] This invention addresses the problem of data distortion caused by existing methods of processing lithium-ion battery data.

[0006] To address the aforementioned technical problems, in a first aspect, the present invention provides a method for cleaning lithium-ion battery timing data, comprising the following steps: Acquire the time-series data generated during the operation of the lithium-ion battery as the raw data; Gaussian noise is added to the original data by forward diffusion to obtain noisy sequence data. Then, the noisy sequence data is reverse-denoised by a preset denoising network, and physical constraints are introduced to obtain electrochemical time series data. The local transient features and global degradation features of the electrochemical time series data are extracted using a spatiotemporal feature pyramid model. Then, the local transient features and global degradation features are weighted by a cross-scale attention mechanism under the guidance of the number of cycles and the capacity decay rate to perform cross-scale feature fusion and obtain fused features. The fusion features are linearly mapped to the SEI film thickness corresponding to the lithium-ion battery to output interpretable higher-order features. The interpretable high-order features are input into a preset decoding network for inverse reconstruction, and lithium-ion battery time-series cleaning data with the same dimension as the original data is output.

[0007] Furthermore, the process of adding Gaussian noise to the original data via forward diffusion to obtain noisy sequence data, followed by inverse denoising of the noisy sequence data using a preset denoising network and introducing physical constraints to obtain electrochemical time-series data, includes the following sub-steps: Gaussian noise is progressively added to the original data via forward diffusion to generate the noisy sequence data. A mass conservation equation is introduced as a physical constraint for the forward diffusion process, constraining the lithium-ion concentration in the lithium-ion battery. This sub-step satisfies the following relationship: ; ; in, The original data, For the noisy sequence data, This is the noise dispatch coefficient, used to control the noise addition rate; Standard Gaussian noise; The noisy sequence data is reverse-denoised using the preset denoising network to remove noise and obtain the electrochemical time-series data. The Butler-Volmer equation is introduced as the physical constraint for the reverse denoising process. The Butler-Volmer equation constrains the relationship between the generation current and overpotential of the lithium-ion battery during the reverse denoising process. This sub-step satisfies the following relationship: ; in, For noise reduction loss, For the preset denoising network pair, This is the diffusion time step.

[0008] Furthermore, the step of adding Gaussian noise to the original data through forward diffusion to obtain noisy sequence data, and then performing inverse denoising on the noisy sequence data through a preset denoising network and introducing physical constraints to obtain electrochemical time-series data, also includes the step of: The denoising loss generated during the inverse denoising process and the physical constraints are weighted and allocated using a preset total loss function. The preset total loss function is defined as follows: It satisfies the following relationship: ; in, For the physical constraints, These are the weighting coefficients for physical constraints, used to balance the denoising loss and the physical constraint loss.

[0009] Furthermore, the process of extracting local transient features and global degradation features from the electrochemical time-series data using a spatiotemporal feature pyramid model, and then performing cross-scale feature fusion by assigning weights to the local transient features and global degradation features based on the number of cycles and capacity decay rate using a cross-scale attention mechanism to obtain the fused features, includes the following sub-steps: At the bottom layer of the spatiotemporal feature pyramid model, the electrochemical time series data is used to extract features by dilated causal convolution according to the dilation factor covering transient fluctuations over different time spans, thereby obtaining the local transient features. In the middle layer of the spatiotemporal feature pyramid model, a correlation graph of voltage, temperature and internal resistance in the electrochemical time series data is constructed through a time-varying graph attention network. The edge weights of the correlation graph are calculated using sliding window mutual information, and the node information of the correlation graph is aggregated through multi-head attention. The output of the correlation graph, combined with the edge weights and the node information, is used as the global degradation feature. The current cycle number of the lithium-ion battery is injected into the spatiotemporal feature pyramid model. With capacity decay rate The number of loops With the capacity decay rate The feature is converted into a conditional vector with the same dimension as the feature through the embedding layer, and the conditional vector is used as an external guiding condition, which is then input into the cross-scale attention mechanism along with the local transient feature and the global degradation feature. At the higher levels of the spatiotemporal feature pyramid model, the weights corresponding to the local transient features and the global degenerate features are dynamically allocated through the cross-scale attention mechanism. The weight allocation is adaptively adjusted by the external guiding conditions, and the local transient features and the global degenerate features are fused across scales according to the weights to obtain the fused features.

[0010] Furthermore, based on the aforementioned external guiding conditions, the fusion feature is defined as follows: The sub-step of dynamically assigning weights corresponding to the local transient features and the global degenerate features through the cross-scale attention mechanism, and fusing the local transient features and the global degenerate features according to the weights, satisfies the following relationship: ; Where softmax is the activation function. T This is the matrix transpose operator. The query vector for the local transient features. , These are the key vector and value vector of the global degradation feature, respectively. For feature dimensions; The weights of the local transient features are adjusted by a preset adaptive strategy, which is to increase the weight ratio of the local transient features during the capacity drop phase of the lithium-ion battery.

[0011] Furthermore, the step of linearly mapping the fusion features to the SEI film thickness corresponding to the lithium-ion battery to output interpretable higher-order features includes the following sub-steps: The fusion feature is linearly mapped to the SEI film thickness corresponding to the lithium-ion battery, and this process satisfies the following relationship: ; in, For SEI film thickness, This is the weight matrix. This is a bias term.

[0012] Furthermore, in the sub-step of linearly mapping the fused features to the SEI film thickness corresponding to the lithium-ion battery, the weight matrix is ​​constrained through supervised learning. and the bias term This makes the fusion characteristics strongly correlated with the thickening process of the SEI film.

[0013] Secondly, the present invention also provides a lithium-ion battery timing data cleaning system, comprising: The data acquisition module is used to acquire time-series data generated during the operation of lithium-ion batteries as raw data; The data processing module is used to add Gaussian noise to the original data through forward diffusion to obtain noisy sequence data. Then, the noisy sequence data is reversed through a preset denoising network, and physical constraints are introduced to obtain electrochemical time series data. The feature extraction module is used to extract local transient features and global degradation features from the electrochemical time series data through a spatiotemporal feature pyramid model, and to perform cross-scale feature fusion by assigning weights to the local transient features and the global degradation features under the guidance of the number of cycles and the capacity decay rate through a cross-scale attention mechanism to obtain fused features; An enhanced interpretation module is used to perform linear mapping processing between the fused features and the SEI film thickness corresponding to the lithium-ion battery, and output interpretable high-order features. The data output module is used to input the interpretable high-order features into a preset decoding network for inverse reconstruction and output lithium-ion battery time-series cleaning data with the same dimension as the original data.

[0014] Thirdly, the present invention also provides a computer device, including: a memory, a processor, and a lithium-ion battery timing data cleaning program stored in the memory and executable on the processor, wherein when the processor executes the lithium-ion battery timing data cleaning program, it implements the steps of the lithium-ion battery timing data cleaning method as described in any of the above embodiments.

[0015] Fourthly, the present invention also provides a storage medium storing a lithium-ion battery timing data cleaning program, wherein when the lithium-ion battery timing data cleaning program is executed by a processor, it implements the steps of the lithium-ion battery timing data cleaning method as described in any of the above embodiments.

[0016] Compared with existing technologies, this invention incorporates electrode dynamics (Butler-Volmer) and mass conservation constraints into the reverse generation process of the diffusion model, fundamentally eliminating the generation of noisy data that violates physical laws. This allows the repaired data to directly support high-precision electrochemical mechanism analysis. Simultaneously, by constructing a spatiotemporal feature pyramid that integrates millisecond-level transients and cyclic-level trends, and designing an adaptive attention enhancement strategy for capacity-draining phases, the capture rate of microscopic electrochemical anomalies is improved. Furthermore, by forcibly mapping high-dimensional features from deep learning to physical state quantities such as SEI film thickness, this invention achieves a closed loop of data cleaning, feature extraction, and mechanism inversion, endowing the pure data-driven model with clear physical semantics and greatly enhancing the credibility of engineering applications. Attached Figure Description

[0017] Figure 1 This is a flowchart illustrating the steps of the lithium-ion battery timing data cleaning method provided in this embodiment of the invention. Figure 2 This is a schematic diagram comparing the effect of the lithium-ion battery timing data cleaning method provided in this embodiment of the invention on voltage curve repair with existing methods; Figure 3 This is a schematic diagram comparing the effect of the lithium-ion battery time-series data cleaning method provided in this embodiment of the invention on capacity decay trajectory repair with existing methods; Figure 4 This is a graph showing the interaction between the cross-scale feature attention weight distribution and the capacity decay acceleration during the execution of the lithium-ion battery time-series data cleaning method provided in this embodiment of the invention. Figure 5 This is a schematic diagram of the structure of the lithium-ion battery timing data cleaning system provided in an embodiment of the present invention; Figure 6 This is a schematic diagram of the structure of a computer device provided in an embodiment of the present invention. Detailed Implementation

[0018] 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.

[0019] Example 1 Please refer to Figure 1 , Figure 1 This is a schematic flowchart of the lithium-ion battery time-series data cleaning method provided in an embodiment of the present invention. The lithium-ion battery time-series data cleaning method includes the following steps: S101. Obtain the timing data generated during the operation of the lithium-ion battery as the raw data.

[0020] In this embodiment of the invention, the raw data should include dynamic monitoring parameters related to the battery's operating state, such as voltage, current, temperature, and internal resistance. This data can be acquired by sensors connected to the lithium-ion battery.

[0021] S102. Gaussian noise is added to the original data by forward diffusion to obtain noisy sequence data. Then, the noisy sequence data is reverse-denoised by a preset denoising network, and physical constraints are introduced to obtain electrochemical time series data.

[0022] Step S102 includes the following sub-steps: Gaussian noise is progressively added to the original data via forward diffusion to generate the noisy sequence data. A mass conservation equation is introduced as a physical constraint for the forward diffusion process, constraining the lithium-ion concentration in the lithium-ion battery. This sub-step satisfies the following relationship: ; ; in, The original data, For the noisy sequence data, This is the noise dispatch coefficient, used to control the noise addition rate; Standard Gaussian noise; The noisy sequence data is inversely denoised using the preset denoising network to remove noise from the noisy sequence data, thereby obtaining the electrochemical time-series data. In this embodiment of the invention, the preset denoising network adopts a U-Net structure based on time-embedding, which includes 4 downsampling convolutional layers and 4 upsampling convolutional layers, with each layer incorporating diffusion time steps. The network uses sinusoidal positional encoding; the convolutional kernel size is 3×1, the stride is 2, and the activation function is Swish; the network input is noisy sequence data. and time step The output is the predicted noise. The Butler-Volmer equation is introduced as a physical constraint for the reverse denoising process. This equation constrains the relationship between the generated current and overpotential of the lithium-ion battery during the reverse denoising process, and this sub-step satisfies the following relationship: ; in, For noise reduction loss, For the preset denoising network pair, This is the diffusion time step.

[0023] The Butler-Volmer equation is a method used to constrain generated currents. With overpotential The electrochemical kinetic equations that satisfy the electrode reaction kinetics relationship serve to ensure that the current generated during data restoration conforms to the basic laws of electrode reactions. They are expressed as follows: ; in The exchange current density is determined by the material properties. , The anode and cathode transfer coefficients are, , , These are Faraday constant, gas constant, and temperature, respectively.

[0024] The mass conservation equation is used to constrain lithium-ion concentration. The diffusion process can be expressed as follows: ; in is the lithium-ion diffusion coefficient.

[0025] Specifically, the process of adding Gaussian noise to the original data through forward diffusion to obtain noisy sequence data, then performing inverse denoising on the noisy sequence data using a preset denoising network and introducing physical constraints to obtain electrochemical time-series data, further includes the following steps: The denoising loss generated during the inverse denoising process and the physical constraints are weighted and allocated using a preset total loss function. The preset total loss function is defined as follows: It satisfies the following relationship: ; in, For the physical constraints, These are the weighting coefficients for physical constraints, used to balance the denoising loss and the physical constraint loss.

[0026] Specifically, the physical constraints This is a regularization term used in the preset denoising network loss function to penalize generated data that violates physical laws. Its complete calculation formula is as follows:

[0027] in, The current prediction value generated by the preset denoising network at the current diffusion time step is... This is the corresponding predicted overpotential value; The complete expression for the Butler-Volmer equation is... , used to calculate dynamic residuals; For the predicted lithium-ion concentration distribution, The lithium-ion diffusion coefficient is... For the Laplace operator, Used to calculate concentration diffusion residuals; and These are the sub-weighting coefficients for the kinetic residual term and the concentration-diffusion residual term, respectively, satisfying... ; Preferably, , The value of is determined by optimizing the battery chemistry system on the training set using a grid search method, and the typical value range is . ∈ [0.3, 0.7], ∈ [0.3, 0.7], in this embodiment both are taken as 0.5. Physical constraints It directly acts on the electrochemical state prediction value output by the preset denoising network at each diffusion time step, forcing its gradient update direction to conform to the electrochemical reaction mechanism of the real battery.

[0028] In step S102, this embodiment of the invention employs a progressive inverse denoising approach. First, Gaussian noise is added to the original data to generate a noisy sequence. Then, a denoising network is trained to predict the noise and optimize the error. Simultaneously, physical constraints such as the Butler-Volmer equation and the mass conservation equation are embedded. Through joint optimization using the total loss function, cleaned data that combines data fidelity and physical consistency is finally obtained. This step repairs noise and missing data in the original data and ensures that the generated electrochemical time-series data strictly conforms to electrochemical laws.

[0029] S103. The local transient features and global degradation features of the electrochemical law time series data are extracted by using a spatiotemporal feature pyramid model. The local transient features and global degradation features are then weighted by a cross-scale attention mechanism under the guidance of the number of cycles and the capacity decay rate to perform cross-scale feature fusion and obtain fused features.

[0030] In this embodiment of the invention, the spatiotemporal feature pyramid is a feature extraction unit used to fuse millisecond-level transient fluctuations and cyclic-level degradation trends of the battery. Specifically, step S103 includes the following sub-steps: At the bottom layer of the spatiotemporal feature pyramid model, the electrochemical time series data is used to extract features by dilated causal convolution according to the dilation factor covering transient fluctuations over different time spans, thereby obtaining the local transient features. In the middle layer of the spatiotemporal feature pyramid model, a correlation graph of voltage, temperature and internal resistance in the electrochemical time series data is constructed through a time-varying graph attention network. The edge weights of the correlation graph are calculated using sliding window mutual information, and the node information of the correlation graph is aggregated through multi-head attention. The output of the correlation graph, combined with the edge weights and the node information, is used as the global degradation feature. The current cycle number of the lithium-ion battery is injected into the spatiotemporal feature pyramid model. With capacity decay rate The number of loops With the capacity decay rate The feature is converted into a conditional vector with the same dimension as the feature through the embedding layer. This conditional vector is then used as an external guiding condition and input into the cross-scale attention mechanism along with the local transient feature and the global degradation feature.

[0031] At the higher levels of the spatiotemporal feature pyramid model, the weights corresponding to the local transient features and the global degenerate features are dynamically allocated through the cross-scale attention mechanism. The weight allocation is adaptively adjusted by the external guiding conditions, and the local transient features and the global degenerate features are fused across scales according to the weights to obtain the fused features.

[0032] In this embodiment of the invention, the dilation factor of the dilated causal convolution used in the bottom layer of the spatiotemporal feature pyramid model is... They cover transient fluctuations with time spans of 1ms, 2ms, 4ms, and 8ms, respectively.

[0033] Time-varying graphical attention network (TV-GAT) is used to dynamically construct voltage. ,temperature Internal resistance The association graph is used to calculate edge weights through sliding window mutual information. Its logic can be expressed as: ; in For mutual information, For information entropy, and These are two node variables in the correlation graph, representing time-series parameters of the lithium-ion battery (such as voltage). ,temperature Internal resistance ); The node information obtained by multi-head attention aggregation It can be represented as: ; in For single-head attention output, This is the output weight matrix.

[0034] Furthermore, based on the aforementioned external guiding conditions, the fusion feature is defined as follows: The sub-step of dynamically assigning weights corresponding to the local transient features and the global degenerate features through the cross-scale attention mechanism, and fusing the local transient features and the global degenerate features according to the weights, satisfies the following relationship: ; Where softmax is the activation function. T This is the matrix transpose operator. The query vector for the local transient features. , These are the key vector and value vector of the global degradation feature, respectively. For feature dimensions; The weights of the local transient features are adjusted by a preset adaptive strategy, which is to increase the weight ratio of the local transient features during the capacity drop phase of the lithium-ion battery.

[0035] Specifically, during implementation, the preset adaptive strategy is as follows: real-time calculation of the lithium-ion battery's performance in the near term... The first derivative of the capacity decay rate (i.e., decay acceleration) within each cycle is taken in this embodiment of the invention. N =10; when this acceleration is continuous The number of cycles exceeds a preset safety threshold (this threshold is taken as 3 times the standard deviation of the mean capacity decay acceleration, or the absolute value is greater than 0.02 mAh / cycle²). In this embodiment of the invention, the threshold is... M =3, or when the current state of health (SOH) of the battery drops to 80% of the rated capacity and the relative growth rate of internal resistance exceeds 5% within 10 adjacent cycles (i.e., it is determined to be a sudden increase in internal resistance), the battery is determined to enter the capacity drop stage, thereby adaptively triggering the increase of the local transient characteristic weight ratio.

[0036] Through step S103, this embodiment of the invention accurately focuses on abnormal information such as precursors to lithium plating using a spatiotemporal feature pyramid model, solving the problem of multi-scale feature fragmentation and providing comprehensive and crucial feature support for subsequent health assessment.

[0037] S104. Perform a linear mapping process between the fusion feature and the SEI film thickness corresponding to the lithium-ion battery to output an interpretable high-order feature.

[0038] The SEI film (Solid Electrolyte Interface) is a passivation film that naturally forms on the surface of the negative electrode material during the first charge and discharge of a battery. SEI film thickness refers to the physical thickness of this passivation film. During battery cycling, the SEI film gradually thickens due to volume changes during charge and discharge, continuous electrolyte decomposition, and other factors. An excessively thick SEI film increases lithium-ion migration resistance, leading to battery capacity decay and increased internal resistance, which is a significant indicator of battery performance degradation. Since the SEI film thickness cannot be directly measured using destructive physical methods under actual operating conditions, in this embodiment of the invention, the SEI film thickness label data for the lithium-ion battery can be obtained through any of the following non-destructive methods: Dynamic estimates are obtained by joint simulation calculation using a pseudo-two-dimensional (P2D) electrochemical mechanism model (specifically the Doyle-Fuller-Newman model) and the charge-discharge curves of batteries from the same batch. Based on electrochemical impedance spectroscopy (EIS) test data, the semi-circular impedance characteristics in the mid-to-high frequency region were extracted, and the equivalent SEI film thickness was calculated according to the positive correlation between the impedance value and the SEI film thickness. The calibration empirical values ​​from publicly available battery material degradation experimental datasets were used.

[0039] This invention establishes a linear correlation between fusion features and SEI film thickness, aiming to enable model features to directly reflect this key degradation mechanism and enhance the physical interpretability of battery health status assessment.

[0040] Step S104 includes the following sub-steps: The fusion feature is linearly mapped to the SEI film thickness corresponding to the lithium-ion battery, and this process satisfies the following relationship: ; in, For SEI film thickness, This is the weight matrix. For bias terms; Specifically, in the step of linearly mapping the fusion features to the SEI film thickness corresponding to the lithium-ion battery, the weight matrix is ​​constrained through supervised learning. and the bias term This makes the fusion characteristics strongly correlated with the thickening process of the SEI film.

[0041] The main purpose of step S104 is to establish an explicit correlation between high-order features and battery degradation mechanisms. Through feature mapping to physical quantities, a linear correlation between fused features and SEI film thickness is forced. Supervised learning is used to constrain parameters, ensuring a strong correlation between features and the SEI film thickening process. The number of iterations... With capacity decay rate In step S103, a spatiotemporal feature pyramid model has been injected to guide the cross-scale attention mechanism to adaptively adjust the weight allocation; in step S104, the fused features after weight adjustment are received, the SEI film thickness linear mapping is performed, and the interpretable high-order features are output.

[0042] S105. Input the interpretable high-order features into a preset decoding network for inverse reconstruction, and output lithium-ion battery time-series cleaning data with the same dimension as the original data.

[0043] In this embodiment of the invention, the preset decoding network adopts a U-Net structure that is symmetrical to the preset denoising network. It gradually restores the temporal resolution through upsampling convolution and finally outputs cleaned temporal data with the same sampling frequency and dimension as the original data.

[0044] To demonstrate the data repair and feature extraction effects of the method proposed in each stage of this invention, please refer to... Figure 2 , Figure 2 This illustration shows a comparison between the effect of the lithium-ion battery time-series data cleaning method proposed in this invention on voltage curve repair and existing methods. In this embodiment, the terminal voltage monitoring data of a single cell in the NASA battery dataset during the 80th to 120th cycle interval is selected. Gaussian noise is artificially injected into the original real data, and about 15% of the data points are randomly deleted to simulate communication packet loss. The data is repaired using linear interpolation, standard DDPM, and the method proposed in this embodiment, and then compared with the real voltage curve.

[0045] Figure 2 In the graph, the horizontal axis represents time (s), and the vertical axis represents the terminal voltage (V). From Figure 2 It can be clearly seen that linear interpolation repair produces a straight transition in the data missing section, and the noise area fluctuates significantly; although standard DDPM repair has a good overall fit, non-physical voltage rebound that violates electrochemical laws appears in the local magnified area; while the method proposed in the embodiments of this invention is highly consistent with the marked real curve in both global trend and local details, with a smooth transition and no abnormal abrupt changes.

[0046] Please refer to Figure 3 , Figure 3 This illustration shows a comparison between the lithium-ion battery time-series data cleaning method proposed in this invention and existing methods for capacity decay trajectory repair. In this invention, discharge capacity data from 200 charge-discharge cycles of the same battery are selected, and data loss and noise disturbances are randomly introduced into the original trajectory. The repair is performed using polynomial fitting, standard GAN, and the method proposed in this invention.

[0047] Figure 3 In the graph, the horizontal axis represents the number of cycles (Cycle), and the vertical axis represents the discharge capacity (Ah). From... Figure 3 It can be seen that the polynomial fitting method exhibits a significant upward convexity in the region with a large missing value (around the 120th cycle), which is a false recovery of capacity and violates the laws of thermodynamics; the standard GAN method fluctuates greatly in the drop zone (around the 150th-180th cycle) and fails to characterize the inflection point; while the method proposed in this embodiment of the invention has a thick solid line that decreases monotonically throughout, accurately capturing the drop inflection point around the 142nd cycle.

[0048] Please refer to Figure 4 , Figure 4 The diagram illustrates the interaction between the cross-scale feature attention weight distribution and the capacity decay acceleration during the execution of the lithium-ion battery time-series data cleaning method proposed in this embodiment of the invention. Figure 4In the graph, the horizontal axis represents the number of cycles, the left vertical axis represents the attention weight (0-1), and the right vertical axis represents the capacity decay acceleration (mAh / cycle²). The weight distribution is presented as a stacked area map. To distinguish between the two regions in the black and white image, diagonal lines filled with textures in different directions are used: the diagonal line filled area from the lower left to the upper right represents the weight proportion of local transient features, and the diagonal line filled area from the upper left to the lower right represents the weight proportion of global degradation features. The sum of the two is always 1. The capacity decay acceleration is represented by a broken line with square markers; the plunge acceleration threshold is marked with a horizontal dashed line, and the capacity plunge stage interval is highlighted with an unfilled dashed box.

[0049] from Figure 4 It can be observed that in the early stage of the method proposed in this embodiment, the proportion of the lower left-upper right diagonal region (local transient weight) is relatively low; as the decay acceleration increases and exceeds the threshold, this region expands significantly, indicating that the preset denoising model adaptively increases the attention to micro-transient fluctuations.

[0050] To further verify the effectiveness of the lithium-ion battery time-series data cleaning method provided in this embodiment of the invention, this embodiment of the invention also compares and verifies the proposed method with existing technologies using the following experimental data. The comparative experiment uses a publicly available lithium battery aging dataset (NASA battery dataset).

[0051] First, a comparative experiment on data cleaning performance was conducted. To verify the effectiveness of forward diffusion and the introduction of physical constraints in inverse denoising, the method provided in this embodiment of the invention was compared with existing linear interpolation methods, standard generative adversarial networks (GANs), and the standard denoising diffusion probabilistic model (DDPM). Verification metrics included the physical consistency of the repaired data (i.e., the proportion of data violating the Butler-Volmer equation) and the root mean square error (RMSE) compared to the true value. The experimental results are shown in Table 1 below.

[0052] Table 1 Comparison of data cleaning effects experimental data

[0053] As can be seen from the data in Table 1, the method proposed in this embodiment of the invention introduces physical constraints such as the Butler-Volmer equation and mass conservation into the preset denoising network, which significantly reduces the RMSE to 0.032 and reduces the proportion of data that violates physical laws to 0.8%, which is significantly better than other existing technologies in comparison, and ensures the high fidelity and physical significance of data repair.

[0054] Secondly, a comparative experiment was conducted on feature extraction performance and early warning capabilities. Addressing the problem of multi-scale feature fragmentation, this invention validated the feature extraction and early warning capabilities of the spatiotemporal feature pyramid model. The proposed method was compared with single-scale convolutional neural networks (CNN), long short-term memory networks (LSTM), and the Transformer model, with the comparison metrics being the lead time for capacity plunge warnings (number of cycles) and anomaly detection accuracy. The experimental results are shown in Table 2 below.

[0055] Table 2. Experimental data comparing feature extraction performance and early warning.

[0056] As can be seen from the data in Table 2, the embodiments of the present invention effectively integrate local transient features and global degradation features through the spatiotemporal feature pyramid model, so that the early warning of capacity drop reaches 23 cycles, the anomaly detection accuracy is improved to 94%, and a causal mapping between microscopic transient anomalies and macroscopic lifetime drop is successfully established.

[0057] Finally, ablation experiments were conducted for verification. To verify the technical contribution of each core design element (physical constraints, spatiotemporal feature pyramid, and SEI film thickness mapping) in the pre-defined denoising network proposed in this embodiment, an ablation experiment was performed for evaluation. The evaluation metric was the RMSE of the cleaned data. The experimental configuration and results are shown in Table 3 below. The baseline model is the basic denoising diffusion network after removing the physical constraints, spatiotemporal feature pyramid, and SEI film thickness mapping in this embodiment. Its structure is consistent with the standard DDPM, but the training dataset is the same. As a reference, the RMSE of the standard DDPM under the same experimental conditions is 0.041.

[0058] Table 3 Ablation Experiment Data

[0059] It can be seen that the RMSE is high when only the baseline model is used, but the RMSE steadily decreases after gradually introducing physical constraints, a spatiotemporal feature pyramid model, and linear mapping processing with SEI film thickness. This fully demonstrates the technical synergy advantages brought about by the combination of the core designs in the preset denoising network proposed in this embodiment of the invention, and ultimately achieves data cleaning with both high fidelity and physical interpretability.

[0060] Compared with existing technologies, this invention incorporates electrode dynamics (Butler-Volmer) and mass conservation constraints into the reverse generation process of the diffusion model, fundamentally eliminating the generation of noisy data that violates physical laws. This allows the repaired data to directly support high-precision electrochemical mechanism analysis. Simultaneously, by constructing a spatiotemporal feature pyramid that integrates millisecond-level transients and cyclic-level trends, and designing an adaptive attention enhancement strategy for capacity-draining phases, the capture rate of microscopic electrochemical anomalies is improved. Furthermore, by forcibly mapping high-dimensional features from deep learning to physical state quantities such as SEI film thickness, this invention achieves a closed loop of data cleaning, feature extraction, and mechanism inversion, endowing the pure data-driven model with clear physical semantics and greatly enhancing the credibility of engineering applications.

[0061] Example 2 This invention also provides a lithium-ion battery timing data cleaning system 200, please refer to... Figure 5 , Figure 5 This is a schematic diagram of the structure of a lithium-ion battery time-series data cleaning system provided in an embodiment of the present invention, which includes: Data acquisition module 201 is used to acquire time-series data generated during the operation of lithium-ion batteries as raw data; The data processing module 202 is used to add Gaussian noise to the original data by forward diffusion to obtain noisy sequence data. Then, the noisy sequence data is reversed by a preset denoising network and physical constraints are introduced to obtain electrochemical time series data. Feature extraction module 203 is used to extract local transient features and global degradation features of the electrochemical law time series data through a spatiotemporal feature pyramid model, and to perform cross-scale feature fusion by assigning weights to the local transient features and the global degradation features under the guidance of the number of cycles and capacity decay rate through a cross-scale attention mechanism to obtain fused features; The enhanced interpretation module 204 is used to perform linear mapping processing between the fused features and the SEI film thickness corresponding to the lithium-ion battery, and output interpretable high-order features. The data output module 205 is used to input the interpretable high-order features into a preset decoding network for inverse reconstruction and output lithium-ion battery time-series cleaning data with the same dimension as the original data.

[0062] The lithium-ion battery time-series data cleaning system 200 can implement the steps in the lithium-ion battery time-series data cleaning method in the above embodiments and achieve the same technical effect. Referring to the description in the above embodiments, it will not be repeated here.

[0063] Example 3 This invention also provides a computer device, please refer to... Figure 6 , Figure 6 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 lithium-ion battery timing data cleaning program stored in the memory 302 and capable of running on the processor 301.

[0064] The processor 301 calls the lithium-ion battery timing data cleaning program stored in the memory 302 and executes the steps in the lithium-ion battery timing data cleaning method provided in this embodiment of the invention. Please refer to... Figure 1 Specifically, it includes the following steps: S101. Obtain the timing data generated during the operation of the lithium-ion battery as the raw data.

[0065] S102. Gaussian noise is added to the original data by forward diffusion to obtain noisy sequence data. Then, the noisy sequence data is reverse-denoised by a preset denoising network, and physical constraints are introduced to obtain electrochemical time series data.

[0066] Step S102 includes the following sub-steps: Gaussian noise is progressively added to the original data via forward diffusion to generate the noisy sequence data. A mass conservation equation is introduced as a physical constraint for the forward diffusion process, constraining the lithium-ion concentration in the lithium-ion battery. This sub-step satisfies the following relationship: ; ; in, The original data, For the noisy sequence data, This is the noise dispatch coefficient, used to control the noise addition rate; Standard Gaussian noise; The noisy sequence data is reverse-denoised using the preset denoising network to remove noise and obtain the electrochemical time-series data. The Butler-Volmer equation is introduced as the physical constraint for the reverse denoising process. The Butler-Volmer equation constrains the relationship between the generation current and overpotential of the lithium-ion battery during the reverse denoising process. This sub-step satisfies the following relationship: ; in, For noise reduction loss, For the preset denoising network pair, This is the diffusion time step.

[0067] Specifically, the process of adding Gaussian noise to the original data through forward diffusion to obtain noisy sequence data, then performing inverse denoising on the noisy sequence data using a preset denoising network and introducing physical constraints to obtain electrochemical time-series data, further includes the following steps: The denoising loss generated during the inverse denoising process and the physical constraints are weighted and allocated using a preset total loss function. The preset total loss function is defined as follows: It satisfies the following relationship: ; in, For the physical constraints, These are the weighting coefficients for physical constraints, used to balance the denoising loss and the physical constraint loss.

[0068] S103. The local transient features and global degradation features of the electrochemical law time series data are extracted by using a spatiotemporal feature pyramid model. The local transient features and global degradation features are then weighted by a cross-scale attention mechanism under the guidance of the number of cycles and the capacity decay rate to perform cross-scale feature fusion and obtain fused features.

[0069] Step S103 includes the following sub-steps: At the bottom layer of the spatiotemporal feature pyramid model, the electrochemical time series data is used to extract features by dilated causal convolution according to the dilation factor covering transient fluctuations over different time spans, thereby obtaining the local transient features. In the middle layer of the spatiotemporal feature pyramid model, a correlation graph of voltage, temperature and internal resistance in the electrochemical time series data is constructed through a time-varying graph attention network. The edge weights of the correlation graph are calculated using sliding window mutual information, and the node information of the correlation graph is aggregated through multi-head attention. The output of the correlation graph, combined with the edge weights and the node information, is used as the global degradation feature. The current cycle number of the lithium-ion battery is injected into the spatiotemporal feature pyramid model. With capacity decay rate The number of loops With the capacity decay rate The feature is converted into a conditional vector with the same dimension as the feature through the embedding layer, and the conditional vector is used as an external guiding condition, which is then input into the cross-scale attention mechanism along with the local transient feature and the global degradation feature. At the higher levels of the spatiotemporal feature pyramid model, the weights corresponding to the local transient features and the global degenerate features are dynamically allocated through the cross-scale attention mechanism. The weight allocation is adaptively adjusted by the external guiding conditions, and the local transient features and the global degenerate features are fused across scales according to the weights to obtain the fused features.

[0070] Furthermore, based on the aforementioned external guiding conditions, the fusion feature is defined as follows: The sub-step of dynamically assigning weights corresponding to the local transient features and the global degenerate features through the cross-scale attention mechanism, and fusing the local transient features and the global degenerate features according to the weights, satisfies the following relationship: ; Where softmax is the activation function. T This is the matrix transpose operator. The query vector for the local transient features. , These are the key vector and value vector of the global degradation feature, respectively. For feature dimensions; The weights of the local transient features are adjusted by a preset adaptive strategy, which is to increase the weight ratio of the local transient features during the capacity drop phase of the lithium-ion battery.

[0071] S104. Perform a linear mapping process between the fusion feature and the SEI film thickness corresponding to the lithium-ion battery to output an interpretable high-order feature.

[0072] Step S104 includes the following sub-steps: The fusion feature is linearly mapped to the SEI film thickness corresponding to the lithium-ion battery, and this process satisfies the following relationship: ; in, For SEI film thickness, This is the weight matrix. This is a bias term.

[0073] Specifically, in the sub-step of linearly mapping the fused features to the SEI film thickness corresponding to the lithium-ion battery, the weight matrix is ​​constrained through supervised learning. and the bias term This makes the fusion characteristics strongly correlated with the thickening process of the SEI film.

[0074] S105. Input the interpretable high-order features into a preset decoding network for inverse reconstruction, and output lithium-ion battery time-series cleaning data with the same dimension as the original data.

[0075] The computer device 300 provided in this embodiment of the invention can implement the steps in the lithium-ion battery timing data cleaning method as described in the above embodiments, and can achieve the same technical effect. Referring to the description in the above embodiments, it will not be repeated here.

[0076] Example 4 This invention also provides a storage medium storing a lithium-ion battery timing data cleaning program. When the lithium-ion battery timing data cleaning program is executed by a processor, it implements the various processes and steps in the lithium-ion battery timing data cleaning method provided in this invention and can achieve the same technical effect. To avoid repetition, it will not be described again here.

[0077] Those skilled in the art will understand that all or part of the processes in the above embodiments can be implemented using hardware (such as a mobile phone, computer, server, air conditioner, or network device) related to a lithium-ion battery timing data cleaning program or instructions. 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.

[0078] 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.

[0079] 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 under the guidance of the present invention without departing from the spirit and scope of the claims. All such changes are within the protection scope of the present invention.

Claims

1. A method for cleaning time-series data of lithium-ion batteries, characterized in that, Includes the following steps: Acquire the time-series data generated during the operation of the lithium-ion battery as the raw data; Gaussian noise is added to the original data by forward diffusion to obtain noisy sequence data. Then, the noisy sequence data is reverse-denoised by a preset denoising network, and physical constraints are introduced to obtain electrochemical time series data. The local transient features and global degradation features of the electrochemical time series data are extracted using a spatiotemporal feature pyramid model. Then, the local transient features and global degradation features are weighted by a cross-scale attention mechanism under the guidance of the number of cycles and the capacity decay rate to perform cross-scale feature fusion and obtain fused features. The fusion features are linearly mapped to the SEI film thickness corresponding to the lithium-ion battery to output interpretable higher-order features. The interpretable high-order features are input into a preset decoding network for inverse reconstruction, and lithium-ion battery time-series cleaning data with the same dimension as the original data is output.

2. The lithium-ion battery time-series data cleaning method according to claim 1, characterized in that, The steps of adding Gaussian noise to the original data through forward diffusion to obtain noisy sequence data, then performing inverse denoising on the noisy sequence data using a preset denoising network and introducing physical constraints to obtain electrochemical time-series data include the following sub-steps: Gaussian noise is progressively added to the original data via forward diffusion to generate the noisy sequence data. A mass conservation equation is introduced as a physical constraint for the forward diffusion process, constraining the lithium-ion concentration in the lithium-ion battery. This sub-step satisfies the following relationship: ； ； in, The original data, For the noisy sequence data, This is the noise dispatch coefficient, used to control the noise addition rate; Standard Gaussian noise; The noisy sequence data is reverse-denoised using the preset denoising network to remove noise and obtain the electrochemical time-series data. The Butler-Volmer equation is introduced as the physical constraint for the reverse denoising process. The Butler-Volmer equation constrains the relationship between the generation current and overpotential of the lithium-ion battery during the reverse denoising process. This sub-step satisfies the following relationship: ； in, For noise reduction loss, For the preset denoising network pair, This is the diffusion time step.

3. The lithium-ion battery time-series data cleaning method according to claim 2, characterized in that, The process of adding Gaussian noise to the original data via forward diffusion to obtain noisy sequence data, and then performing inverse denoising on the noisy sequence data using a preset denoising network and introducing physical constraints to obtain electrochemical time-series data, further includes the following steps: The denoising loss generated during the inverse denoising process and the physical constraints are weighted and allocated using a preset total loss function. The preset total loss function is defined as follows: It satisfies the following relationship: ； in, For the physical constraints, These are the weighting coefficients for physical constraints, used to balance the denoising loss and the physical constraint loss.

4. The lithium-ion battery time-series data cleaning method according to claim 1, characterized in that, The steps of extracting local transient features and global degradation features from the electrochemical time-series data using a spatiotemporal feature pyramid model, and then fusing these features across scales by assigning weights to them based on the number of cycles and capacity decay rate using a cross-scale attention mechanism, to obtain the fused features, include the following sub-steps: At the bottom layer of the spatiotemporal feature pyramid model, the electrochemical time series data is used to extract features by dilated causal convolution according to the dilation factor covering transient fluctuations over different time spans, thereby obtaining the local transient features. In the middle layer of the spatiotemporal feature pyramid model, a correlation graph of voltage, temperature and internal resistance in the electrochemical time series data is constructed through a time-varying graph attention network. The edge weights of the correlation graph are calculated using sliding window mutual information, and the node information of the correlation graph is aggregated through multi-head attention. The output of the correlation graph, combined with the edge weights and the node information, is used as the global degradation feature. The current cycle number of the lithium-ion battery is injected into the spatiotemporal feature pyramid model. With capacity decay rate The number of loops With the capacity decay rate The feature is converted into a conditional vector with the same dimension as the feature through the embedding layer, and the conditional vector is used as an external guiding condition, which is then input into the cross-scale attention mechanism along with the local transient feature and the global degradation feature. At the higher levels of the spatiotemporal feature pyramid model, the weights corresponding to the local transient features and the global degenerate features are dynamically allocated through the cross-scale attention mechanism. The weight allocation is adaptively adjusted by the external guiding conditions, and the local transient features and the global degenerate features are fused across scales according to the weights to obtain the fused features.

5. The lithium-ion battery time-series data cleaning method according to claim 4, characterized in that, Based on the aforementioned external guiding conditions, the fusion feature is defined as follows: The sub-step of dynamically assigning weights corresponding to the local transient features and the global degenerate features through the cross-scale attention mechanism, and fusing the local transient features and the global degenerate features according to the weights, satisfies the following relationship: ； Where softmax is the activation function. T This is the matrix transpose operator. The query vector for the local transient features. , These are the key vector and value vector of the global degradation feature, respectively. For feature dimensions; The weights of the local transient features are adjusted by a preset adaptive strategy, which is to increase the weight ratio of the local transient features during the capacity drop phase of the lithium-ion battery.

6. The lithium-ion battery time-series data cleaning method according to claim 1, characterized in that, The step of performing a linear mapping between the fused features and the SEI film thickness corresponding to the lithium-ion battery to output interpretable higher-order features includes the following sub-steps: The fusion feature is linearly mapped to the SEI film thickness corresponding to the lithium-ion battery, and this process satisfies the following relationship: ； in, For SEI film thickness, This is the weight matrix. This is a bias term.

7. The lithium-ion battery time-series data cleaning method according to claim 6, characterized in that, In the sub-step of linearly mapping the fused features to the SEI film thickness corresponding to the lithium-ion battery, the weight matrix is ​​constrained through supervised learning. and the bias term This makes the fusion characteristics strongly correlated with the thickening process of the SEI film.

8. A lithium-ion battery time-series data cleaning system, characterized in that, include: The data acquisition module is used to acquire time-series data generated during the operation of lithium-ion batteries as raw data; The data processing module is used to add Gaussian noise to the original data through forward diffusion to obtain noisy sequence data. Then, the noisy sequence data is reversed through a preset denoising network, and physical constraints are introduced to obtain electrochemical time series data. The feature extraction module is used to extract local transient features and global degradation features from the electrochemical time series data through a spatiotemporal feature pyramid model, and to perform cross-scale feature fusion by assigning weights to the local transient features and the global degradation features under the guidance of the number of cycles and the capacity decay rate through a cross-scale attention mechanism to obtain fused features; An enhanced interpretation module is used to perform linear mapping processing between the fused features and the SEI film thickness corresponding to the lithium-ion battery, and output interpretable high-order features. The data output module is used to input the interpretable high-order features into a preset decoding network for inverse reconstruction and output lithium-ion battery time-series cleaning data with the same dimension as the original data.

9. A computer device, characterized in that, include: The system includes a memory, a processor, and a lithium-ion battery timing data cleaning program stored in the memory and executable on the processor. When the processor executes the lithium-ion battery timing data cleaning program, it implements the steps of the lithium-ion battery timing data cleaning method as described in any one of claims 1-7.

10. A storage medium, characterized in that, The storage medium stores a lithium-ion battery timing data cleaning program, which, when executed by a processor, implements the steps of the lithium-ion battery timing data cleaning method as described in any one of claims 1-7.

Images