Lithium precipitation state recognition method and system based on stress fluctuation feature extraction

Abstract

The application relates to the technical field of artificial intelligence, and discloses a lithium precipitation state recognition method and system of a battery based on stress fluctuation feature extraction. The method comprises the following steps: obtaining original strain time series data of battery circulation, and performing standardization processing to obtain standardized strain time series data; calculating a strain change rate according to the standardized strain time series data, and adaptively determining a relaxation baseline time set and a relaxation baseline mean value; calculating a local mechanical bending degree under multiple scales according to the standardized strain time series data and the relaxation baseline mean value, and combining baseline deviation normalization to obtain a stress relaxation deformation rate sequence; constructing a deformation rate time weight according to the stress relaxation deformation rate sequence, and calculating a dynamic bimodality index; and performing feature fusion and classification processing according to the stress relaxation deformation rate sequence and the dynamic bimodality index to obtain a battery lithium precipitation state recognition result. The application can realize high-precision and high-robustness online recognition of the lithium precipitation state of a lithium battery.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence technology, specifically to a method and system for identifying lithium plating state in batteries based on stress fluctuation feature extraction. Background Technology

[0002] During the charging process of lithium-ion batteries, especially under abusive conditions such as low temperature, high rate, or overcharging, metallic lithium is easily deposited on the surface of the negative electrode, a phenomenon known as lithium plating. Lithium plating not only causes irreversible loss of active lithium and accelerates capacity decay, but more seriously, the deposited lithium dendrites may pierce the separator, causing internal short circuits and ultimately leading to catastrophic safety accidents such as thermal runaway. Therefore, developing an accurate, sensitive, and online monitoring method for early identification of lithium plating is a key technological requirement for ensuring the safe operation of batteries throughout their entire lifecycle.

[0003] Existing methods for identifying lithium plating status in batteries mainly suffer from the following problems:

[0004] 1. Existing methods mostly rely on manually setting relaxation time windows or fixed operating condition templates, which are difficult to adapt to the reality that the relaxation stage will drift on the time axis under the entire battery life cycle and changing operating conditions, resulting in unstable reference benchmarks.

[0005] 2. Commonly used strain characteristic analysis methods either only use first-order difference, which are easily affected by the overall charging and discharging expansion trend, or only extract abnormal features at a single spatiotemporal scale, and cannot take into account the multi-morphological mechanical responses such as short-term rebound spikes and continuous plateau distortions caused by lithium plating.

[0006] 3. Traditional strain waveform statistics usually apply equal weight to all moments in the full cycle. Local key abnormal features caused by lithium plating are easily averaged or masked by a large amount of stable data in the normal stage, resulting in insufficient sensitivity to identify early or trace amounts of lithium plating.

[0007] 4. Existing detection models often separate manual feature extraction from the classifier. Effective prior information, such as the overall waveform anomaly level, is only used in the final classification stage. This fails to deeply involve the model in the feature selection and enhancement process, limiting the model's ability to focus on learning key lithium plating patterns and its ability to transfer knowledge across different scenarios. Summary of the Invention

[0008] To address the aforementioned shortcomings in existing technologies, this invention provides a battery lithium plating state identification method and system based on stress fluctuation feature extraction. By constructing a complete technical chain from physical signal adaptive preprocessing, multi-scale mechanical feature extraction, physical prior quantization to prior-driven deep learning identification, it achieves high-precision and robust online identification of lithium battery lithium plating state.

[0009] To achieve the above-mentioned objectives, the technical solution adopted by this invention is as follows:

[0010] In a first aspect, the present invention proposes a method for identifying the lithium plating state of a battery based on stress fluctuation feature extraction, comprising the following steps:

[0011] The raw strain time series data of battery cycling is obtained, and the raw strain time series data is standardized to obtain standardized strain time series data.

[0012] The strain change rate is calculated based on the standardized strain time series data, and the relaxation baseline time set and relaxation baseline mean are adaptively determined.

[0013] Based on the standardized strain time series data and the mean of the relaxation baseline, the degree of local mechanical bending is calculated at multiple scales and normalized by the baseline deviation to obtain the stress relaxation deformation rate sequence.

[0014] Based on the stress relaxation strain rate sequence, a strain rate time weight is constructed, and a dynamic bimodality index characterizing the heavy tail and multi-peak distortion of the strain waveform is calculated.

[0015] Based on the stress relaxation deformation rate sequence and the dynamic bimodality index, feature fusion and classification are performed to obtain the battery lithium plating state identification result.

[0016] Secondly, this invention proposes a battery lithium plating state identification system based on stress fluctuation feature extraction, which applies the battery lithium plating state identification method based on stress fluctuation feature extraction as described above, including:

[0017] The data acquisition module is used to acquire the raw strain time series data of the battery cycle and to standardize the raw strain time series data to obtain standardized strain time series data.

[0018] The baseline determination module is used to calculate the strain change rate based on the standardized strain time series data and adaptively determine the relaxation baseline time set and the relaxation baseline mean.

[0019] The deformation rate calculation module is used to calculate the local mechanical bending degree at multiple scales based on the standardized strain time series data and the mean value of the relaxation baseline, and combine the baseline deviation normalization to obtain the stress relaxation deformation rate sequence.

[0020] The bimodality calculation module is used to construct the deformation rate time weight based on the stress relaxation deformation rate sequence, and calculate the dynamic bimodality index characterizing the heavy tail and multi-peak distortion of the strain waveform.

[0021] The state recognition module is used to perform feature fusion and classification processing based on the stress relaxation deformation rate sequence and the dynamic bimodality index to obtain the battery lithium plating state recognition result.

[0022] The present invention has the following beneficial effects:

[0023] This invention adaptively determines the relaxation baseline, automatically tracking relaxation time drift under different operating conditions and aging stages, eliminating baseline instability caused by manually fixed time windows. Through multi-scale stress relaxation deformation rate extraction, it can simultaneously capture two types of mechanical anomalies: short-term rebound spikes and long-term plateau distortions. By introducing deformation rate time weights through a dynamic bimodality index, it directly couples global waveform anomaly judgment with lithium plating sensitive moments, significantly improving the detection sensitivity of trace lithium plating. Through a dynamic bimodality prior gating mechanism and a deformation rate risk-weighted pooling mechanism, it injects physical prior depth into the feature selection and attention allocation process of the neural network, greatly enhancing the model's generalization ability under varying operating conditions and across scenarios. Attached Figure Description

[0024] Figure 1 This is a schematic diagram of a battery lithium plating state identification method based on stress fluctuation feature extraction according to the present invention.

[0025] Figure 2 This is a schematic diagram of real-time battery strain timing data in this invention;

[0026] Figure 3 This is a schematic diagram of the standardized strain time series data and adaptive relaxation baseline in this invention;

[0027] Figure 4 This is a schematic diagram of the stress relaxation deformation rate sequence in this invention;

[0028] Figure 5 This is a schematic diagram illustrating the relationship between the dynamic bimodality index and the predicted probability of lithium plating state in this invention. Detailed Implementation

[0029] The specific embodiments of the present invention are described below to enable those skilled in the art to understand the present invention. However, it should be understood that the present invention is not limited to the scope of the specific embodiments. For those skilled in the art, various changes are obvious as long as they are within the spirit and scope of the present invention as defined and determined by the appended claims. All inventions utilizing the concept of the present invention are protected.

[0030] like Figure 1 As shown, an embodiment of the present invention provides a battery lithium plating state identification method based on stress fluctuation feature extraction, comprising the following steps S1 to S5:

[0031] S1. Obtain the raw strain time series data of battery cycling, and perform standardization processing on the raw strain time series data to obtain standardized strain time series data;

[0032] In an optional embodiment of the present invention, during charging, resting, and discharging, lithium insertion and delithiation of the electrode material can cause minor deformations in the battery casing, electrode assembly, or structural constraints. When lithium plating occurs, metallic lithium deposits on the negative electrode surface, inducing local volume expansion, interfacial stress concentration, and relaxation response distortion. This causes the original battery strain time series data to exhibit local rebound, slope reversal, multi-peak fluctuations, or plateau distortion at the end of charging, the resting stage, or the beginning of discharging. The present invention extracts real-time battery strain time series data corresponding to the current charge-discharge cycle from the online strain monitoring data stream of the battery under test and obtains the strain sequence of the test cycle using a time-axis normalized resampling method. , The sampling time is specified. In practical implementation, if the online system cannot wait for the complete cycle to end, a test cycle segment can be formed at the end of charging and during the resting phase, and this test cycle segment can be mapped to a local time axis of fixed length.

[0033] like Figure 2 As shown, this illustrates the real-time battery strain time-series data collected by the strain sensor during a complete charge-discharge cycle. Figure 2 The curve shows a distinct local rebound peak at the end of the charging process, indicating that the strain signal deviates from the normal expansion trend. The horizontal axis represents time in hours; the vertical axis represents the original strain value in microstrain.

[0034] S2. Calculate the strain change rate based on the standardized strain time series data, and adaptively determine the relaxation baseline time set and the relaxation baseline mean.

[0035] In an optional embodiment of the present invention, step S2 calculates the strain change rate based on the standardized strain time series data and adaptively determines the relaxation baseline time set and the relaxation baseline mean, including:

[0036] Calculate the first-order rate of change of adjacent sampling points based on the standardized strain time series data, and obtain the median of the first-order rate of change;

[0037] The longest continuous segment with an absolute value of the first-order rate of change less than a preset threshold and a continuous length that meets the minimum continuous length requirement is selected as the relaxation baseline time set.

[0038] The mean of the standardized strain values ​​is calculated within the relaxation baseline time set to obtain the relaxation baseline mean.

[0039] In this embodiment, the cyclic strain sequence can reflect the mechanical response of the battery during lithium-ion insertion and extraction. However, the cyclic strain sequence also includes normal charge-discharge expansion trends, temperature drift, sensor zero-point offset, and random noise. This invention first performs zero-point correction, outlier replacement, and smoothing normalization on the cyclic strain sequence to obtain normalized strain time-series data. Then, it adaptively determines the relaxation baseline time set based on the first-order rate of change of the normalized strain time-series data and calculates the mean of the relaxation baseline. Next, it extracts the local curvature response related to the relaxation baseline at multiple half-window scales to obtain the stress relaxation deformation rate sequence. Finally, it constructs a deformation rate time weight based on the stress relaxation deformation rate sequence and calculates a dynamic bimodal index to generate a joint representation of lithium plating fluctuation characteristics.

[0040] Since manually preset relaxation time windows are easily affected by differences in cycle length, sampling frequency, charge / discharge rate, and resting time, the relaxation references for different samples may be inconsistent. Therefore, this invention filters out segments with relatively slow strain changes and long continuous times by using the rate of change of adjacent sampling points, and uses these segments as the relaxation baseline time set, so that subsequent local rebound strength calculations have a stable reference. The specific steps are as follows:

[0041] S201, regarding the first Cyclic strain sequence of a battery cycle sample Perform zero-point calibration. Specifically, take the value before the loop begins. The mean of each sampling point is used as the zero-point offset. ,Right now Then, from the cyclic strain sequence Subtract zero offset The zero-point correction strain sequence was obtained. ,Right now .in, This indicates the number of initial sampling points used for zero-point estimation, which can be 10, 20, or 30. It refers to the first The zero-point offset of each battery cycle sample is used to characterize the overall offset caused by the initial bias of the sensor and the mounting preload. It refers to the first The battery cycle sample at the ... The zero-point correction strain value at each sampling time refers to the strain value after deducting the zero-point offset.

[0042] Furthermore, this embodiment also includes a zero-point correction strain sequence. Perform outlier replacement. Specifically, based on the current sampling time... Set the length to the center. Within the local neighborhood of the sampled point, the local median and local standard deviation are calculated. If the deviation of the current sampled point from the local median is greater than three times the local standard deviation, the current sampled point is identified as an isolated outlier, and the isolated outlier is replaced by linear interpolation of adjacent normal sampled points. This represents the neighborhood length for anomaly detection, and can be 7, 9, or 11. This processing can reduce isolated spikes caused by sensor transients, electromagnetic interference, or packet loss, preventing these isolated spikes from being amplified in subsequent second-order difference calculus.

[0043] S202. Calculate the in-sample mean and in-sample standard deviation of the zero-point corrected strain sequence after outlier replacement, and perform standard deviation normalization (also known as Z-score normalization or zero-mean normalization) to obtain standardized strain time series data. , It refers to the first The battery cycle sample at the ... The standardized strain values ​​at each sampling time point, i.e. strain time series data after zero-point correction, outlier replacement and standard deviation normalization, have the characteristics of consistent sampling length, uniform numerical scale and comparability between samples.

[0044] In an alternative implementation, if the original sensor noise is high, this embodiment can add low-pass smoothing before standard deviation normalization, for example, by using a moving average filter with a length of 5 or a second-order Butterworth low-pass filter. Low-pass smoothing is only used to suppress high-frequency acquisition noise and should not over-smooth local bounce details at the end of charging and during the resting phase.

[0045] S203, Based on standardized strain time series data Calculate the first-order rate of change between adjacent sampling points to obtain the strain rate of change sequence. ,Right now .in, Indicates the first The battery cycle sample at the ... The strain rate of change value at each sampling time is used to characterize the variation range of the standardized strain value between adjacent sampling points; Indicates the first The battery cycle sample at the ... Standardized strain values ​​at each sampling time point. Strain rate sequence. The smaller the value, the more gradual the strain change is near the sampling time, and the more likely it corresponds to a static relaxation or slow recovery section.

[0046] Furthermore, this embodiment also includes the strain rate of change sequence. Calculate the median ( Within the range), the median global strain rate of change is obtained. , It refers to the first The median global strain rate of change of each battery cycle sample is used to provide a rate of change reference that does not rely on manual time windows, reducing the impact of isolated spikes and short-term abnormal fluctuations on the judgment of relaxation sections.

[0047] S204, will satisfy The sampling time is used as a candidate relaxation time, and the longest continuous time interval is searched among all candidate relaxation times to obtain the relaxation baseline time set. , It refers to the first The set of relaxation baseline times for each battery cycle sample, i.e., the set of sampling times where strain changes are relatively gradual and the continuous time is relatively long. This represents the screening ratio coefficient for the relaxation zone, used to control the tolerance of the strain change rate at the relaxation candidate time. It can be 0.5, 0.8, or 1.0.

[0048] In practical implementation, from Start sequential scan to ;when At that time, the first Each sampling time is added to the current continuous candidate segment; when When the scan is complete, the current consecutive candidate segment ends and the statistics restart; after the scan is complete, the consecutive candidate segment containing the most sampling points is selected as the relaxation baseline time set. When multiple consecutive candidate segments of the same length exist, the consecutive candidate segment with the smallest variance of the standardized strain values ​​within the segment is selected as the relaxation baseline time set. To improve the stability of the relaxation baseline.

[0049] For example, suppose a standardized strain time series data Number of sampling points The calculated median global strain rate of change Screening ratio coefficient That is, the rate of change threshold The scanning process is as follows:

[0050] exist to Between, 11 consecutive points This forms candidate segment A;

[0051] exist to Between, 3 consecutive points This forms candidate segment B;

[0052] exist to Between, 26 consecutive points This forms candidate segment C;

[0053] Candidate segment C contains the most sampling points (26), and is therefore selected as the relaxation baseline time set. If segment D also contains 26 sampling points, then the original normalized strain values ​​for segments C and D need to be calculated separately. Given the variance, select the segment with the smaller variance as... This is to ensure that the strain value in this section is more stable.

[0054] Furthermore, set a minimum continuous length. , This refers to the minimum number of sampling points that the relaxation baseline time set must contain, which can be 30 or 60. If no continuous candidate segment satisfies the minimum continuity length... Then in the standardized strain time series data The length used in the middle is A sliding window is used to calculate the mean strain rate of change within each sliding window, and the sliding window with the smallest mean strain rate of change is selected as the relaxation baseline time set. This fallback approach ensures that each battery cycle sample obtains a set of relaxation baseline times that can be used for subsequent normalization reference.

[0055] S205, in the relaxation baseline time set The mean of the normalized strain value is calculated internally to obtain the mean of the relaxation baseline. , It refers to the first The mean of the relaxation baseline of the nth battery cycle sample, representing the... The reference strain level of a single battery cycle sample during the strain relaxation and stabilization phase, i.e. .in, Represents the set of relaxation baseline times The number of sampling points included is used to normalize the mean of the standardized strain values ​​within the relaxation baseline time set.

[0056] For example, a single cycle is resampled to 1200 sampling points, and the relaxation zone screening ratio coefficient is... minimum continuous length If the strain rate of change of adjacent sampling points in a continuous segment is consistently less than 0.5 times the median global strain rate of change, and the segment contains 180 sampling points, then this continuous segment is taken as the relaxation baseline time set. The average of the standardized strain values ​​corresponding to the 180 sampling points is used to obtain the mean of the relaxation baseline. This is used to provide a stable reference for subsequent identification of local rebound mutations, reducing the impact of different cycle durations and different sensor installation deviations.

[0057] It should be noted that the relaxation baseline time set of the present invention is not a fixed time period specified by humans, but is adaptively determined by the first-order rate of change of the standardized strain time series data itself. It can adapt to the problem of changes in the position of the relaxation stage under different charging rates, different resting times and different aging stages, so that the subsequent stress relaxation strain rate calculation does not depend on a fixed working condition template.

[0058] like Figure 3 As shown, the standardized strain time series data and adaptive relaxation baseline are displayed. The mean of the relaxation baseline is marked with a horizontal dashed line, and the relaxation baseline time set is indicated by a light-colored area. The horizontal axis represents the resampling time, which is unitless (representing the sampling point number after resampling); the vertical axis represents the standardized strain value, which is dimensionless.

[0059] S3. Based on the standardized strain time series data and the mean value of the relaxation baseline, calculate the local mechanical bending degree at multiple scales and combine it with the baseline deviation normalization to obtain the stress relaxation deformation rate sequence.

[0060] In an optional embodiment of the present invention, step S3 calculates the local mechanical bending degree at multiple scales based on the standardized strain time series data and the mean of the relaxation baseline, and combines this with baseline deviation normalization to obtain a stress relaxation deformation rate sequence, including:

[0061] The second-order difference response is calculated on the standardized strain time series data at multiple half-window scales;

[0062] Scale decay processing is performed on the second-order difference responses at different half-window scales to obtain the scale-decayed second-order difference responses;

[0063] Calculate the deviation between the current standardized strain value and the mean of the relaxation baseline, and construct a relaxation baseline deviation normalization term;

[0064] The stress relaxation deformation rate sequence is obtained by summing the second-order difference response after scale attenuation with the normalized term of the relaxation baseline deviation point by point.

[0065] Since the mechanical anomalies induced by lithium plating may manifest as short-term local rebound and local slope reversal, or as strain plateau distortion lasting for a long period, using only the first-order difference is easily affected by the overall charge-discharge expansion trend; using only the second-order difference at a single scale is prone to missing anomalous responses of different durations. Therefore, this invention uses the mean of the relaxation baseline as a reference to calculate the local curvature response at multiple half-window scales, and obtains the stress relaxation strain rate sequence through scale attenuation and relaxation baseline deviation normalization. The specific steps are as follows:

[0066] S301, using standardized strain time series data and relaxation baseline mean As input, at each sampling time Multiple half-window sizes were set up nearby. .in, This represents the half-window scale index, used to characterize the time offset scale relative to the center sampling time during second-order difference calculation. ; This represents the maximum half-window scale, used to control the time range covered by the multi-scale relaxation analysis; it can be 6, 8, or 12. When the sampling frequency is 1Hz and... At that time, the maximum half-window scale corresponds to the local response range 8 seconds before and after the current sampling time.

[0067] S302, regarding the first The battery cycle sample at the ... The sampling time, the first The local curvature response at a half-window scale is calculated to obtain the second-order difference response. , It refers to the first The battery cycle sample at the ... The sampling time, the first The second-order difference response at a half-window scale is used to characterize the curvature of the standardized strain time series data within this local time range, i.e. .in, Indicates the first The battery cycle sample at the ... The standardized strain values ​​at each sampling time are used to characterize the local strain changes to the right of the central sampling time; Indicates the first The battery cycle sample at the ... The standardized strain values ​​at each sampling time are used to characterize the local strain change to the left of the central sampling time. Second-order difference response. The larger the value, the more likely there is local bounce, slope reversal, or plateau distortion near the sampling time.

[0068] It should be noted that when or This caused the sampling location to exceed the standardized strain time series data. When the effective range is reached, the corresponding second-order difference response will be... Setting it to 0 ensures that the stress relaxation strain rate sequence has the same length as the standardized strain time series data and avoids the introduction of false anomalies by boundary extrapolation.

[0069] S303, Second-order difference response at different half-window scales Scale attenuation processing is performed to ensure that short-term sharp rebounds and long-term plateau distortions maintain a reasonable contribution ratio in the same cumulative amount. Specifically, for the first... Second-order difference response at a half-window scale Divided by the scale attenuation term To complete the scale attenuation process; among which, The scale decay index is used to reduce the masking effect of large-scale second-order difference responses on short-term local anomalies. It can be taken as 1.0 to 2.0, with 1.5 being preferred.

[0070] In specific implementation, for the first Second-order difference response at a half-window scale Divided by the scale attenuation term ,Right now This yields the scale-attenuated response, and this operation directly affects the response value in subsequent operations on all... When summing, divide the large-scale response factor by a larger factor. This is attenuated, thus balancing the contributions of responses at different scales.

[0071] For example, let the maximum half-window size be set. Scale decay index At a certain moment The second-order difference responses at the three scales were calculated as follows: , , Without scale attenuation, the total response contribution is: Large-scale contributions account for the largest proportion. After performing scale attenuation processing, the attenuation terms are as follows: , , The contribution of the decayed response is It is evident that, after scale attenuation, the contribution of small-scale (sharp rebound) response is highlighted, while the contribution of large-scale response is effectively suppressed.

[0072] It should be noted that the smaller half-window scale More sensitive to sharp local rebound, larger half-window scale More sensitive to distortion on slow platforms. Through Scale decay can prevent large-scale responses from having too high a weight in the accumulation due to their wide span, while still retaining the ability to perceive persistent relaxation distortions.

[0073] S304, Based on the standardized strain value at the current sampling time With relaxation baseline mean The deviation between them is used to construct a normalized term for the relaxed baseline deviation. , It is the first The battery cycle sample at the ... The normalized term for the relaxation baseline deviation at each sampling time is used to adjust the sensitivity of the local curvature response to the relaxation phase, i.e. .in, Indicates the first The degree of deviation of the standardized strain value at each sampling time from the mean of the relaxation baseline.

[0074] It should be noted that the closer the current sampling time is to the relaxation baseline, The smaller the value, the easier it is to retain the contribution of local springback mutations in subsequent stress relaxation deformation rate calculations; when the current sampling time is far from the relaxation baseline, The increase in macroscopic charging expansion or discharging contraction trends suppresses the interference of local anomaly calculations.

[0075] S305. Combine the second-order difference responses, scale decay terms, and relaxation baseline deviation normalization terms corresponding to multiple half-window scales to calculate the stress relaxation deformation rate value point by point. , It is the first The battery cycle sample at the ... The stress relaxation strain rate at each sampling time is used to characterize the intensity of the local rebound abrupt change in the strain curve relative to the relaxation baseline near that sampling time. The calculation method is expressed as follows: . The larger the value, the more drastic the local curvature change at that location, and the higher the probability of abnormal fluctuations related to lithium plating.

[0076] Furthermore, this embodiment provides for By performing the above calculations point by point, the stress relaxation strain rate sequence is obtained. , It is the first The stress relaxation deformation rate sequence of each battery cycle sample is used to preserve the location, intensity, and duration of lithium plating-related mechanical anomalies on the time axis.

[0077] For example, if standardized strain time series data If a localized rebound phenomenon occurs at the end of the charging process, characterized by "first decreasing, then briefly rising, and then continuing to decrease," then the rebound will occur near the center of the rebound. , and Second-order difference response Significantly increased; if this position is also close to the mean of the relaxation baseline. Relaxation baseline bias normalization term If the value is small, the stress relaxation deformation rate value This process creates local peak values. Through this treatment, the subtle local mechanical anomalies induced by lithium plating can be highlighted from the overall strain variation trend.

[0078] It should be noted that the stress relaxation deformation rate sequence of the present invention is not a normal second-order difference sequence, but is obtained by introducing scale decay and relaxation baseline deviation normalization on the basis of multi-scale curvature response. It can simultaneously focus on short-term rebound and long-term plateau distortion, and reduce the interference of normal macroscopic expansion trend far from the relaxation stage on lithium plating identification.

[0079] like Figure 4 As shown, the stress relaxation deformation rate sequence is obtained after scale decay and relaxation baseline deviation normalization. The curves exhibit significant peaks at the end of charging, the resting stage, and the beginning of discharging, with the peak regions highlighted. The stress relaxation deformation rate can accurately locate local rebound abrupt changes and plateau distortions related to lithium plating, demonstrating its inhibitory effect on normal expansion trends and its high sensitivity to lithium plating mechanical anomalies. The horizontal axis represents the resampling time (unitless); the vertical axis represents the stress relaxation deformation rate value (dimensionless).

[0080] S4. Construct a time weight for the strain rate based on the stress relaxation strain rate sequence, and calculate a dynamic bimodal index that characterizes the heavy tail and multi-peak distortion of the strain waveform.

[0081] In an optional embodiment of the present invention, step S4 constructs a deformation rate time weight based on the stress relaxation deformation rate sequence and calculates a dynamic bimodal index characterizing the heavy tail and multi-peak distortion of the strain waveform, including:

[0082] The average stress relaxation rate value is calculated based on the stress relaxation rate sequence.

[0083] Deformation rate time weights are constructed based on the stress relaxation deformation rate value and the average stress relaxation deformation rate value at each sampling time.

[0084] A fourth-order central deviation term is constructed based on the central deviation between the standardized strain time series data and the mean of the full-cycle standardized strain;

[0085] The deformation rate time weight is introduced into the fourth-order central deviation term, and the variance squared term is used for scaling normalization to obtain the dynamic bimodal index.

[0086] Stress relaxation strain rate sequences can locate local rebound abrupt changes, but individual local peaks may originate from normal electrochemical relaxation fluctuations or acquisition noise. To determine whether standardized strain time-series data exhibits heavy-tailed patterns, multi-peak splitting, or anomalous distortion at the overall waveform level, this invention constructs a strain rate time weight based on the stress relaxation strain rate sequence and incorporates this time weight into the calculation of the fourth-order central moment to obtain a dynamic bimodality index. The specific steps are as follows:

[0087] S401, Based on the stress relaxation deformation rate sequence Calculate the mean stress relaxation rate. , It refers to the first The average stress relaxation rate of each battery cycle sample is used to perform relative scale normalization on the stress relaxation rate values ​​at each sampling time. .in, This represents the time summation index, used to traverse all sampling times in the stress relaxation rate sequence; Indicates the first The battery cycle sample at the ... The stress relaxation deformation rate value at each sampling time.

[0088] S402. Based on the comparison between the stress relaxation deformation rate value and the average stress relaxation deformation rate value at each sampling time, the deformation rate time weight is obtained. , It refers to the first The battery cycle sample at the ... The deformation rate time weight at each sampling time is used to increase the contribution of moments with significant rebound abrupt changes in the dynamic bimodality index calculation. . The base value is 1, when the first Stress relaxation deformation rate at each sampling time Higher than the mean stress relaxation strain rate At that time, deformation rate time weight Increasing this makes the subsequent fourth-order central moments pay more attention to abnormal waveform deviations at this sampling time.

[0089] In one implementation, to avoid excessive weighting of extreme noise points, the deformation rate time weight can be adjusted. Perform upper limit pruning, with the pruning upper limit denoted as . , This represents the maximum allowable value for the time weight of the deformation rate, which can be 5 or 10. If the calculated... Then let This treatment can reduce the excessive influence of single-point noise on the overall dynamic bimodality index.

[0090] S403, Calculate standardized strain time series data Full-cycle standardized strain mean , It refers to the first The normalized strain mean of each battery cycle sample is used to calculate the deviation of each sampling time relative to the overall strain level. .

[0091] Further, calculate the first The center deviation at each sampling time And by raising the central deviation to the fourth power, we obtain the fourth-order central deviation term. Fourth-order central deviation term It can amplify abnormal strain fluctuations that are far from the overall mean, making local rebound, multi-peak distortion, and heavy-tailed waveforms more prominent in the overall statistics.

[0092] S404, Weight the deformation rate by time By introducing a fourth-order central bias term and using a variance squared term for scaling, a dynamic bikurtosis index is obtained. , represented as:

[0093]

[0094] in, Indicates the first The dynamic bimodality index of each battery cycle sample is used to characterize the heavy-tailedness and multi-peak distortion of the standardized strain time-series data; the larger the value, the more the strain relaxation waveform deviates from the normal unimodal stationary distribution, and the more obvious the lithium plating-related anomalies. The item is the first The strain variance of each battery cycle sample is used to measure the overall fluctuation of the standardized strain time series data relative to the full cycle standardized strain mean; the squared variance term is used to scale the weighted fourth-order central bias, making the dynamic bimodality index comparable among different battery cycle samples.

[0095] It should be noted that conventional kurtosis does not distinguish between normal fluctuation periods and abnormal rebound periods, and is easily diluted by the normal expansion trend of the entire cycle. The dynamic bimodality index of this invention... Time weighting by deformation rate Increasing the contribution of local rebound mutation moments to the overall statistics and coupling the overall waveform anomaly statistics with lithium-sensitive local moments can reduce noise misjudgment caused by simply relying on the maximum peak value and improve the overall anomaly separability of early trace lithium-plating samples.

[0096] This embodiment can determine the dynamic bimodal reference threshold based on the dynamic bimodal index distribution of undeposited lithium samples in the training set. , This is used to assist in determining whether there are obvious overall waveform anomalies in the test samples. Specifically, battery cycle samples with a lithium plating state label of 0 are selected from the training set, and the dynamic bimodality index of these non-lithium-plated samples is calculated (calculated in steps 1 to 4 of S203), and the 95th percentile is taken as the dynamic bimodality reference threshold. .

[0097] For example, if the dynamic bimodality index of the unlithiated samples in the training set is mainly distributed between 2.5 and 4.0, and the 95th percentile is 4.2, then the following settings can be configured: When the dynamic bimodality index of the sample to be tested... consistently higher And the stress relaxation strain rate sequence When a concentrated peak occurs at the end of charging or during the resting stage, it indicates that the sample under test has a higher risk of lithium plating.

[0098] S5. Based on the stress relaxation deformation rate sequence and the dynamic bimodality index, feature fusion and classification processing are performed to obtain the battery lithium plating state identification result.

[0099] In an optional embodiment of the present invention, step S5 performs feature fusion and classification processing based on the stress relaxation deformation rate sequence and the dynamic bimodality index to obtain the battery lithium plating state identification result, including:

[0100] The stress relaxation deformation rate sequence and the dynamic bimodal index are input into a trained time-domain perceptual neural network, and the lithium plating state identification result is output through the dynamic bimodal prior gating mechanism and the deformation rate risk weighted pooling mechanism.

[0101] While using the stress relaxation rate sequence alone can preserve the time location of anomalies, it lacks information on the overall waveform anomaly degree; and using the dynamic bimodal index alone can reflect the overall waveform distortion, but it cannot pinpoint the time of anomaly occurrence. Therefore, this invention jointly expresses the stress relaxation rate sequence and the dynamic bimodal index to obtain a joint representation of lithium plating fluctuation characteristics. The specific steps are as follows:

[0102] S501, the first The stress relaxation rate values ​​of each battery cycle sample at all sampling times are arranged in chronological order to obtain the stress relaxation rate sequence. , It refers to the first The stress relaxation deformation rate sequence of a battery cycle sample, with dimension [ ]. ,Right now .in; Indicates the transpose sign. Stress relaxation rate sequence. Each location in the data is compared with the standardized strain time series data. The sampling times correspond one-to-one, thus preserving the time location, intensity, and duration of local rebound mutations related to lithium plating.

[0103] S502, Stress relaxation rate sequence Convert to row vector and with dynamic bimodality index By concatenating along the feature dimension, a joint representation of lithium plating fluctuation characteristics is obtained. , It refers to the first A joint representation of lithium plating fluctuation characteristics from individual battery cycle samples is used to simultaneously characterize local rebound abrupt changes and global waveform anomalies, with a dimension of [missing information]. ,Right now .in, The feature concatenation operator is used to combine sequence features and scalar features into a unified feature representation.

[0104] For example, when a single battery cycling sample contains 1200 sampling points, the stress relaxation deformation rate sequence The dimension is Dynamic bimodality index As scalars, the joint representation of lithium plating fluctuation characteristics The dimension is 1201. The first 1200 positions are used to describe temporal location features, and the last position is used to describe overall waveform anomaly statistical features. This joint representation provides a physically clear prior, temporally clear, and sample-comparable input for the subsequent time-domain perceptual neural network.

[0105] It should be noted that the joint representation of lithium plating fluctuation characteristics Instead of directly inputting the original strain values ​​into the model, the multi-scale local curvature anomaly and overall waveform distortion information are fused together. This reduces the model's dependence on the original amplitude scale, sensor installation bias, and overall expansion differences in different cycles, thereby improving the physical interpretability and cross-condition stability of lithium plating identification.

[0106] S503, stress relaxation deformation rate sequences can characterize the temporal location and intensity of mechanical anomalies related to lithium plating, but they may still contain normal electrochemical relaxation fluctuations. This invention constructs a time-domain-aware neural network, which includes a multi-resolution time-domain feature encoding module, a dynamic bimodal prior gating module, a deformation rate risk pooling module, and a lithium plating state classification module. Specifically, the multi-resolution time-domain feature encoding module is used to simultaneously extract short-term rebound spikes and long-term plateau distortions; the dynamic bimodal prior gating module is used to convert the overall waveform anomaly degree into channel-level modulation information; the deformation rate risk pooling module is used to increase the contribution of high deformation rate moments to sample-level features; and the lithium plating state classification module is used to output the predicted probability of the lithium plating state.

[0107] Lithium plating-induced strain anomalies may manifest as short-duration, narrow-pulse rebound or long-duration strain plateau distortion. This invention constructs a multi-resolution temporal feature encoding module, inputting the stress relaxation strain rate sequence into local-scale convolutional branches and context-scale convolutional branches respectively to obtain a multi-resolution temporal feature map. The specific steps are as follows:

[0108] (1) The first Stress relaxation deformation rate sequence of a battery cycle sample Converted to a single-channel time series tensor that can be received by a one-dimensional convolutional network , It refers to the first The single-channel deformation rate input tensor for each battery cycle sample, with size [value missing]. The first dimension is the number of input channels (the input is only the stress relaxation deformation rate sequence). This is a physical quantity, hence the number of input channels is 1), and the second dimension is the time length; among which, , indicating that the single-channel deformation rate input tensor in the 1st channel and the 2nd channel. The input values ​​at each sampling time point. Each time position in this input tensor corresponds to a stress relaxation rate value, which is used to provide the network with temporal location information of local springback abrupt changes related to lithium plating.

[0109] (2) Input the single-channel deformation rate into the tensor The input local-scale convolution branch performs one-dimensional convolution, zero-padding, batch normalization, and linear rectified activation along the time dimension to obtain the local-scale temporal feature map. Local-scale temporal feature map It is used to capture narrow-pulse rebound, short-duration slope inversion, and sharp deformation rate peaks. The kernel length of the local-scale convolution branch is [not specified]. You can choose 3, which represents the number of output channels. You can choose 32, step size 1, and zero padding length. Thus, the output time length is maintained at .

[0110] In practice, the local scale convolution branch uses the convolution weight tensor and bias vector The deformation rate changes within a small number of adjacent sampling points are weighted and summed, and then batch normalized and a linear rectified activation function is applied to output a local-scale time-domain feature map. ,Right now .in, The convolution weight tensor, representing the local scale convolution branch, is a trainable parameter used to extract short-time abrupt responses in the stress relaxation deformation rate sequence. The bias vector representing the local-scale convolution branch is a trainable parameter used to adjust the channel offset of the local-scale convolution output. This represents the linear rectified activation function, used to introduce nonlinear expressive power and suppress negative response noise; This represents a one-dimensional convolution operation. This indicates batch normalization operation.

[0111] (3) Input the same single-channel deformation rate into the tensor The context-scale convolution branch is input, and one-dimensional convolution, zero-padding, batch normalization, and linear rectified activation are performed along the time dimension to obtain the context-scale temporal feature map. Context-scale temporal feature map This technique is used to capture persistent plateau distortions, wide-range multi-peak splitting, and slow deformation rate bulges. The kernel length of the context-scale convolution branch is also considered. You can choose 15, which represents the number of output channels. You can choose 32, step size 1, and zero padding length. Thus, the output time length is maintained at .

[0112] In its implementation, the context-scale convolution branch uses the convolution weight tensor and bias vector The deformation rate variation over a wide time range is weighted and extracted, and then processed by batch normalization and a linear rectified activation function to output a context-scale time-domain feature map. ,Right now .in, The convolution weight tensor representing the context-scale convolution branch is a trainable parameter used to extract strain plateau distortion and slow waveform shift over a wide time range. The bias vector representing the context-scale convolution branch is a trainable parameter used to adjust the channel offset of the context-scale convolution output.

[0113] (4) Local scale time domain feature map and context-scale temporal feature maps By stitching along the channel dimension, a multi-resolution temporal feature map is obtained. , It refers to the first Multi-resolution temporal feature maps of individual battery cycle samples are used to simultaneously characterize short-time rebound spikes and long-time relaxation distortions, with a size of [size missing]. ,Right now .in, This represents the total number of channels in the multi-resolution temporal feature map. 64 is a good choice; This represents the channel splicing operator, used to merge the local scale branch output and the context scale branch output into a unified feature map.

[0114] For example, the local scale convolution kernel length At this time, the network mainly focuses on sharp bounces within a small number of adjacent sampling points; context-scale convolution kernel length In this way, the network can perceive the slow rise or multi-peak splitting of strain plateaus over a wider time range. The two types of convolutional branches together improve the network's ability to distinguish early lithium plating mechanical anomalies, avoiding the problem of insufficient adaptation of a single convolutional scale to anomaly responses of different durations.

[0115] The dynamic bimodality index can reflect whether the overall strain waveform has heavy tails, multi-peak splitting, and anomalous distortion. However, if the dynamic bimodality index is only used in the final classification stage, the network front-end feature extraction process may still focus on all channels equally. This invention constructs a dynamic bimodality prior gating module, converts the dynamic bimodality index into a standardized dynamic bimodality prior, generates channel-level gating vectors through a trainable nonlinear mapping, and then modulates the multi-resolution time-domain feature map channel by channel. The specific steps are as follows:

[0116] (1) Based on the dynamic bimodality statistic in the training set, the first... Dynamic bimodality index of individual battery cycle samples Standardization is performed to obtain the standardized dynamic bimodal prior. , It refers to the first A standardized dynamic bimodal prior for each battery cycle sample is used to stabilize the prior scale between different samples. In one implementation, the standardization method uses standard deviation normalization (also known as Z-score normalization or zero-mean normalization).

[0117] It should be noted that during standardization, the mean and standard deviation of the dynamic bimodal index are calculated only from the training set. The validation and test sets directly use the mean and standard deviation of the dynamic bimodal index fixed during the training phase during inference to avoid leakage of validation or test set information.

[0118] (2) Standardize the dynamic bimodality prior The input dynamic bimodal prior embedding module is passed sequentially through a first fully connected layer, a linear rectified activation function, and a second fully connected layer to obtain the channel-level gating vector. , It refers to the first The channel-level gated vector of each battery cycle sample, with dimension [missing information]. This is used to control the enhancement or suppression intensity of each channel in the multi-resolution temporal feature map, i.e. The first fully connected layer incorporates a 1-dimensional normalized dynamic bimodal prior. Mapping to the latent prior space, latent dimension A value of 16 can be used; the second fully connected layer maps the hidden prior space to a length of 16. The channel-level gating vector ensures that each channel of the multi-resolution temporal feature map receives a gating coefficient adjusted by a dynamic bimodality index. Wherein; The weight matrix of the first fully connected layer of the dynamic bimodal prior embedding module is a trainable parameter with dimension O(n). ; The bias vector of the first fully connected layer of the dynamic bimodal prior embedding module is denoted as , which are trainable parameters with dimension . ; The weight matrix of the second fully connected layer of the dynamic bimodal prior embedding module is a trainable parameter with dimension O(n). ; The bias vector of the second fully connected layer of the dynamic bimodal prior embedding module is represented by a trainable parameter of dimension . ; This represents a logical function (i.e., the Sigmoid function) used to compress gated mapping values ​​to the range of 0 to 1; Used to adjust the gating range to 0 to 2. When the gating coefficient is less than 1, the corresponding channel response is suppressed, and when the gating coefficient is greater than 1, the corresponding channel response is enhanced.

[0119] In one alternative implementation, to improve stability during the initial training phase, it can be... Initializing to 0 enables the dynamic bimodal prior embedding module to output channel-level gating coefficients close to 1 in the early stages of training. This ensures that the multi-resolution temporal feature map is close to an unmodulated state in the early stages of training, avoiding premature suppression of the effective temporal response due to random gating.

[0120] (3) Gating the channel-level control vector Replication along the time dimension Next, the size is obtained as The channel-level gated matrix is ​​then used in conjunction with the multi-resolution temporal feature map. Element-wise multiplication yields the bimodal prior modulation feature map. , It refers to the first Bimodal prior modulation feature maps of individual battery cycle samples were used to highlight channel responses related to lithium plating waveform distortion and reduce normal relaxation fluctuation interference. The size is [size missing]. .

[0121] In specific implementation, for the first The bimodal prior modulation feature map of the battery cycle sample in the th The first channel, the first The feature values ​​at each sampling time are then: .in, Indicates the channel index. ; Indicates the first The multi-resolution temporal feature map of the battery cycle samples in the first... The first channel, the first Feature values ​​at each sampling time; Indicates the first The channel-level gating vector of the battery cycle sample in the th... Gating coefficients on each channel; Indicates the first The bimodal prior modulation feature map of the battery cycle sample in the th The first channel, the first The feature values ​​at each sampling time.

[0122] It should be noted that the dynamic bimodality index reflects whether there are heavy-tailed and multi-peak anomalies in the overall strain waveform. The channel-level gating mechanism transforms the overall anomaly information into channel selection ability in the feature extraction stage, so that the network does not need to rely entirely on sample labels to learn attention allocation from scratch. It can guide the network to prioritize retaining feature channels related to rebound peaks, multi-peak splitting and plateau distortion in the early stage of training, thereby improving the critical lithium plating sample identification ability.

[0123] For example, when the dynamic bimodality index of a certain battery cycle sample is significantly higher than the distribution range of the unlithiated sample, the standardized dynamic bimodality prior is used. By taking a larger positive value, the dynamic bimodal prior embedding module can improve the gating coefficient of channels related to rebound peaks, multi-peak splitting, and plateau distortion, enabling these channels to contribute more in subsequent classification. When the dynamic bimodal index of a battery cycle sample is within the normal range, the dynamic bimodal prior embedding module can keep the channels related to normal relaxation background close to 1 or moderately reduce them, thereby suppressing the interference of background fluctuations on the classification results.

[0124] The bimodal prior modulation feature map contains lithium plating-related responses of different channels at different time positions. This invention constructs a deformation rate risk pooling module, constructs a normalized time risk weight based on the stress relaxation deformation rate sequence, and uses this normalized time risk weight to perform weighted pooling on the bimodal prior modulation feature map to obtain sample-level time-domain feature vectors. The specific steps are as follows:

[0125] (1) Based on the stress relaxation deformation rate sequence Generate normalized time risk weights , It refers to the first The battery cycle sample at the ... The normalized temporal risk weight for each sampling moment is used to determine the contribution ratio of that sampling moment in the sample-level feature aggregation. Specifically, the deformation rate time weight is first... As the unnormalized time weights, the unnormalized time weights of all sampling times are then normalized, i.e. .in, The item is the first The sum of the time weights of the deformation rate at all sampling times for each battery cycle sample. After normalization, the sum of the normalized time risk weights at each sampling time is close to 1, and higher deformation rate times receive higher pooling contributions.

[0126] (2) Using normalized time risk weights Bimodal prior modulation feature map Perform weighted pooling along the time dimension to obtain the deformation rate risk pooling feature vector. , It refers to the first The deformation rate risk pooling feature vector of each battery cycle sample, with dimension [missing information]. It is used to summarize the lithium plating-related channel responses at high-risk time locations.

[0127] In practical implementation, the deformation rate risk pooling feature vector is targeted at the first... The feature values ​​on each channel are then: .in, The deformation rate risk pooling feature vector represents the first... Feature values ​​on each channel.

[0128] It should be noted that, compared with ordinary average pooling, this weighted pooling can increase the contribution of local rebound moments at the end of charging, the resting stage, or the beginning of discharging, and reduce the dilution of abnormal response during long normal stages.

[0129] (3) Bimodal prior modulation feature map Perform max pooling along the time dimension to obtain the maximum response feature vector. , It refers to the first The maximum response feature vector of each battery cycle sample, with dimension [ ]. This is used to retain the strongest abnormal response that has occurred in the entire loop for each channel.

[0130] In specific implementation, for the first The maximum response feature vector of a battery cycle sample is then: .in, This represents the maximum value operator, used to extract the maximum eigenvalue over time. The maximum response eigenvector can supplement the ability of weighted pooling to capture strong single-point anomalies, and is suitable for identifying lithium plating rebound spikes with extremely short durations.

[0131] (4) Pool the deformation rate risk feature vector Maximum response eigenvector and standardized dynamic bimodal prior By concatenating along the feature dimensions, we obtain the sample-level lithium plating discrimination feature vector. , It refers to the first The sample-level lithium plating discrimination feature vector of each battery cycle sample is used to simultaneously characterize the weighted abnormal response at high-risk moments, the maximum abnormal response throughout the cycle, and the overall waveform anomaly prior. The dimension is [missing information]. ,Right now .in, This represents the feature concatenation operator, used to merge multiple feature vectors or scalars into a unified feature vector.

[0132] This invention constructs a lithium plating state classification module, which classifies lithium plating at the sample level into lithium plating discrimination feature vectors. Inputting the data into a single-layer fully connected classifier yields the log-probability output value of the lithium plating state. , It refers to the first The log-probability output value of the lithium plating state of each battery cycle sample is used to characterize the original discriminative strength of the classifier before the logistic function mapping. Then, the predicted probability of lithium plating state is obtained through a logic function. , It refers to the first The predicted probability that a battery cycle sample belongs to the lithium plating state, with a value ranging from 0 to 1. The higher the value, the higher the probability that the model judges the battery cycle sample to be in the lithium plating state. .in, The classification layer weight vector represents trainable parameters used to map the sample-level lithium plating discrimination feature vector to the log-probability output value of the lithium plating state. Its dimension is... ; The classification layer bias is a trainable parameter used to adjust the overall offset of the log-probability output value of the lithium plating state.

[0133] like Figure 5 As shown, the relationship between the dynamic bimodality index and the predicted probability of lithium plating state is analyzed. The dynamic bimodality index is plotted on the horizontal axis, and the predicted probability of lithium plating state is plotted on the vertical axis. The color of the points changes with the bimodality value. The scatter plots show a clear positive correlation trend; the higher the bimodality, the closer the predicted probability is to 1. There is a strong correlation between the overall waveform anomaly and the lithium plating confidence level output by the neural network. The horizontal axis of the graph represents the dynamic bimodality index, which is dimensionless; the vertical axis represents the predicted probability of lithium plating state, which is dimensionless (range 0 to 1).

[0134] This embodiment uses a set of battery cycle samples labeled with lithium plating status to supervise the training of a time-domain perceptual neural network. The invention first collects raw battery strain time-series data under different charge-discharge cycle conditions. Specifically, it collects raw battery strain time-series data corresponding to a single charge-discharge cycle using a strain sensor, and labels each cycle sample with a lithium plating status, forming a set of labeled battery cycle samples for feature construction and model training. The specific steps are as follows:

[0135] (1) Strain sensors are placed on the surface of the battery under test, the module constraint plate, the outer support structure of the electrode group, or near the cell expansion sensitive location to collect raw battery strain time-series data during charging, resting, and discharging processes. The strain sensors can be resistance strain gauges, fiber optic strain sensors, or thin-film pressure-strain composite sensors; the sampling frequency is denoted as... , This refers to the number of strain sampling points collected per second, which can be 1Hz, 5Hz, or 10Hz.

[0136] The collected first The original battery strain time series data of each battery cycle sample are denoted as: ,in, This represents the battery cycle sample index, used to distinguish samples from different batteries, different cycle counts, or different operating conditions. ; This indicates the total number of battery cycle samples; Indicates the index of the original sampling time; Indicates the first The battery cycle sample at the ... The strain sampling values ​​at each original sampling time are used to characterize the mechanical deformation response of the battery at that time.

[0137] (2) Based on the battery cycle conditions, record the start time of charging, end time of charging, start time of rest, end time of rest, and end time of discharging. Divide the original battery strain time series data corresponding to each battery cycle sample into single cycles to obtain cyclic strain segments. If the duration of different cycles is inconsistent, perform time axis normalization resampling on the cyclic strain segments to obtain cyclic strain sequences of consistent length. .in, This represents the time index after resampling, used to indicate the discrete sampling time within a single loop. ; This represents the number of sampling points after resampling a single battery cycle sample. It is used to characterize the sequence length of a single charge-discharge cycle on a unified time axis and can be 600, 1200, or 2400. Indicates the first The battery cycle sample at the ... The cyclic strain values ​​at each resampling time point serve as the foundational input for subsequent standardization processing, relaxation baseline determination, and lithium plating fluctuation characteristic construction.

[0138] In practical implementation, linear interpolation, cubic spline interpolation, or piecewise proportional resampling can be used to unify cyclic strain segments of different durations to a single value. Each sampling point. If you want to retain the details of the end of charging and the resting stage, you can resample according to the stage ratio. For example, map the charging stage to the first 60% of the sampling points, the resting stage to the middle 20% of the sampling points, and the initial stage of discharging to the last 20% of the sampling points, so that the lithium plating sensitive stage has a relatively consistent time position in different samples.

[0139] It should be noted that the core input used in this invention is a cyclic strain sequence. It is not required that voltage, current, and temperature data be input simultaneously. Voltage, current, and temperature data can be used as operating condition records for sample segmentation, label confirmation, or subsequent feature expansion, but do not change the main process of lithium plating identification based on stress relaxation deformation rate and dynamic bimodality prior in this invention.

[0140] In this embodiment, the lithium plating state is labeled for each battery cycle sample to obtain a lithium plating state label. Lithium plating status label It refers to the first The actual lithium plating category label for each battery cycle sample is labeled with a value of 0 or 1, where 0 indicates that lithium plating was not detected and 1 indicates that lithium plating was detected.

[0141] In one implementation, the lithium plating status label can be confirmed through disassembly and observation, detection of lithium metal deposition on the negative electrode surface, abnormal electrochemical incremental capacity, abnormal voltage recovery at the end of charging, low-temperature high-rate induction experiments, or a combination of multiple detection results. For example, if a battery is disassembled after being charged at a low temperature and a high rate, and a visible silvery-white metallic lithium deposit is found on the surface of the negative electrode, the corresponding cycle or adjacent cycle is marked as a lithium-plated state; if no lithium deposit is found during disassembly and the electrochemical curve does not show any abnormalities related to lithium plating, it is marked as a non-lithium-plated state.

[0142] It should be noted that the lithium plating status label This data is used to supervise subsequent training of the time-domain perceptual neural network. To improve the model's generalization ability, the labeled battery cycle sample set should cover different ambient temperatures, different charging rates, different states of charge ranges, different aging stages, and different individual batteries. For example, it can cover temperature conditions such as -10℃, 0℃, and 25℃, charging rate conditions such as 0.5C, 1C, and 2C, and include samples of new batteries, moderately aged batteries, and batteries with high cycle counts. By acquiring data under multiple operating conditions, the risk of the model only remembering the differences in strain amplitude at a single temperature or single charging rate can be reduced.

[0143] Cyclic strain sequence Lithium plating status label Combine to form a tagged battery cycling sample set ,Right now .in, Indicates the first The cyclic strain sequence of a battery cycle sample is composed of... A sequence arranged in chronological order; Indicates the first Lithium plating status labels for individual battery cycle samples.

[0144] Each battery cycle sample undergoes the above steps to generate a joint representation of stress relaxation deformation rate sequence, dynamic bimodality index, and lithium plating fluctuation characteristics. Finally, the predicted probability of lithium plating state is obtained, and the network parameters are updated using class-weighted binary cross-entropy loss. The specific steps are as follows:

[0145] (1) For the first training batch A battery cycle sample was used to read its lithium plating state tag. Lithium plating state prediction probability , It refers to the first batch in the training batch The predicted probability of lithium plating state for each battery cycle sample. This represents the sample index within the training batch. ; This indicates the number of samples in the training batch, which can be 16, 32, or 64. Indicates the first in the training batch The true lithium plating status label of a battery cycle sample.

[0146] (2) Set class weights based on the number of lithium-plated samples and the number of non-lithium-plated samples in the training set. Specifically, the number of lithium-plated samples is denoted as... The number of samples without lithium deposition is denoted as The weight of lithium plating category is denoted as The weight of the undeposited lithium category is denoted as ,but , .in, This indicates the number of samples in the training set with a lithium plating state label of 1. This indicates the number of samples in the training set with a lithium plating state label of 0. If the number of lithium plating samples is less than the number of non-lithium plating samples, then... Greater than During training, the model will pay more attention to misclassification of lithium-plated samples.

[0147] (3) Predicting probability based on lithium plating state and the actual lithium plating status label Calculate the category-weighted binary classification cross-entropy loss function , represented as:

[0148]

[0149] in, The category-weighted binary cross-entropy loss value of the training batch is used to measure the difference between the predicted lithium plating state and the actual lithium plating state label, and serves as the target for network parameter optimization. This represents the natural logarithm function, and in machine learning tasks, the default base is the natural constant. Used to prevent This leads to numerical errors. Category-weighted binary cross-entropy loss can alleviate the class imbalance problem caused by a large number of unlithiated samples and a small number of lithium-plated samples in actual data collection.

[0150] (4) Convert the convolution weight tensor of the local scale convolution branch and bias vector Convolution weight tensors of context-scale convolution branches and bias vector Weight matrix of dynamic bimodal prior embedding module and Bias vector of dynamic bimodal prior embedding module and Classification layer weight vector and classification layer bias As trainable parameters, they are iteratively updated through gradient backpropagation.

[0151] In one implementation, the optimizer employs an adaptive moment estimation optimization algorithm, with an initial learning rate of 0.001, a maximum number of training epochs ranging from 100 to 200, and a training batch size of... Set the value to 32. If the validation set loss does not decrease for 10 consecutive rounds, training can be stopped early; alternatively, if the validation set loss does not decrease for 5 consecutive rounds, the learning rate can be multiplied by 0.5 to improve the convergence stability in the later stages of training.

[0152] (5) After each training round, the time-domain perceptual neural network is evaluated using the validation set, and the decision threshold for predicting the lithium plating state probability is determined based on the receiver operating characteristic curve of the validation set. , This refers to the threshold for determining the predicted probability of lithium plating state, used to convert the predicted probability of lithium plating state into a judgment result of lithium plating state. It can be set to 0.5, or 0.4 to 0.7 depending on the needs of early warning. If more emphasis is placed on the early lithium plating recall rate, a lower threshold that achieves the preset requirement can be selected; if more emphasis is placed on reducing false alarms, a higher threshold that achieves the preset requirement can be selected.

[0153] It should be noted that this invention extracts short-term strain rebound abrupt changes and long-term relaxation plateau distortions through multi-resolution temporal coding, incorporates the overall waveform anomaly degree into the feature extraction process through a dynamic bimodal prior gating mechanism, enhances the sample-level contribution of high-risk time locations through a deformation rate risk pooling mechanism, and completes the training of the lithium plating state recognition model through category-weighted binary cross-entropy loss. This combined approach enables the model to simultaneously possess the ability to perceive local temporal anomalies, utilize global waveform anomalies, and provide physical interpretability.

[0154] After completing the training of the time-domain perceptual neural network, the trained lithium plating identification model is deployed to the battery testing system, battery management system or battery safety monitoring platform. The system continuously collects raw battery strain time series data during battery cycling and generates a joint representation of lithium plating fluctuation features according to the same processing flow as the training stage. Finally, it outputs the lithium plating state prediction probability and lithium plating risk warning results.

[0155] This invention also provides a battery lithium plating state identification system based on stress fluctuation feature extraction, which applies the battery lithium plating state identification method based on stress fluctuation feature extraction as described above, including:

[0156] The data acquisition module is used to acquire the raw strain time series data of the battery cycle and to standardize the raw strain time series data to obtain standardized strain time series data.

[0157] The baseline determination module is used to calculate the strain change rate based on the standardized strain time series data and adaptively determine the relaxation baseline time set and the relaxation baseline mean.

[0158] The deformation rate calculation module is used to calculate the local mechanical bending degree at multiple scales based on the standardized strain time series data and the mean value of the relaxation baseline, and combine the baseline deviation normalization to obtain the stress relaxation deformation rate sequence.

[0159] The bimodality calculation module is used to construct the deformation rate time weight based on the stress relaxation deformation rate sequence, and calculate the dynamic bimodality index characterizing the heavy tail and multi-peak distortion of the strain waveform.

[0160] The state recognition module is used to perform feature fusion and classification processing based on the stress relaxation deformation rate sequence and the dynamic bimodality index to obtain the battery lithium plating state recognition result.

[0161] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0162] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0163] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0164] Specific embodiments have been used to illustrate the principles and implementation methods of this invention. The descriptions of the embodiments above are only for the purpose of helping to understand the method and core ideas of this invention. At the same time, for those skilled in the art, there will be changes in the specific implementation methods and application scope based on the ideas of this invention. Therefore, the content of this specification should not be construed as a limitation of this invention.

[0165] Those skilled in the art will recognize that the embodiments described herein are intended to help the reader understand the principles of the invention, and should be understood that the scope of protection of the invention is not limited to such specific statements and embodiments. Those skilled in the art can make various other specific modifications and combinations based on the technical teachings disclosed in this invention without departing from the spirit of the invention, and these modifications and combinations are still within the scope of protection of this invention.

Claims

1. A method for identifying the lithium plating state of a battery based on stress fluctuation feature extraction, characterized in that, Includes the following steps: The raw strain time series data of battery cycling is obtained, and the raw strain time series data is standardized to obtain standardized strain time series data. The strain change rate is calculated based on the standardized strain time series data, and the relaxation baseline time set and relaxation baseline mean are adaptively determined. Based on the standardized strain time series data and the mean of the relaxation baseline, the degree of local mechanical bending is calculated at multiple scales and normalized by the baseline deviation to obtain the stress relaxation deformation rate sequence. Based on the stress relaxation strain rate sequence, a strain rate time weight is constructed, and a dynamic bimodality index characterizing the heavy tail and multi-peak distortion of the strain waveform is calculated. Based on the stress relaxation deformation rate sequence and the dynamic bimodality index, feature fusion and classification are performed to obtain the battery lithium plating state identification result. The step involves calculating the local mechanical bending degree at multiple scales based on the standardized strain time series data and the mean of the relaxation baseline, and then combining this with baseline deviation normalization to obtain the stress relaxation deformation rate sequence, specifically including: The second-order difference response is calculated on the standardized strain time series data at multiple half-window scales; Scale decay processing is performed on the second-order difference responses at different half-window scales to obtain the scale-decayed second-order difference responses; Calculate the deviation between the current standardized strain value and the mean of the relaxation baseline, and construct a relaxation baseline deviation normalization term; The stress relaxation deformation rate sequence is obtained by summing the second-order difference response after scale attenuation with the normalized term of the relaxation baseline deviation point by point. The step of constructing a deformation rate time weight based on the stress relaxation deformation rate sequence and calculating a dynamic bimodal index characterizing the heavy tail and multi-peak distortion of the strain waveform specifically includes: The average stress relaxation rate value is calculated based on the stress relaxation rate sequence. Deformation rate time weights are constructed based on the stress relaxation deformation rate value and the average stress relaxation deformation rate value at each sampling time. A fourth-order central deviation term is constructed based on the central deviation between the standardized strain time series data and the mean of the full-cycle standardized strain; The deformation rate time weight is introduced into the fourth-order central deviation term, and the variance squared term is used for scaling normalization to obtain the dynamic bimodal index.

2. The battery lithium plating state identification method based on stress fluctuation feature extraction according to claim 1, characterized in that, Based on the standardized strain time series data, the strain change rate is calculated, and the relaxation baseline time set and relaxation baseline mean are adaptively determined, including: Calculate the first-order rate of change of adjacent sampling points based on the standardized strain time series data, and obtain the median of the first-order rate of change; The longest continuous segment with an absolute value of the first-order rate of change less than a preset threshold and a continuous length that meets the minimum continuous length requirement is selected as the relaxation baseline time set. The mean of the standardized strain values ​​is calculated within the relaxation baseline time set to obtain the relaxation baseline mean.

3. The battery lithium plating state identification method based on stress fluctuation feature extraction according to claim 1, characterized in that, The step of performing feature fusion and classification based on the stress relaxation deformation rate sequence and the dynamic bimodality index to obtain the battery lithium plating state identification result includes: The stress relaxation deformation rate sequence and the dynamic bimodal index are input into a trained time-domain perceptual neural network, and the lithium plating state identification result is output through the dynamic bimodal prior gating mechanism and the deformation rate risk weighted pooling mechanism.

4. The battery lithium plating state identification method based on stress fluctuation feature extraction according to claim 3, characterized in that, The time-domain sensing neural network includes: The multi-resolution temporal feature encoding module is used to extract multi-resolution temporal feature maps that simultaneously characterize short-time rebound spikes and long-time relaxation distortions based on the stress relaxation deformation rate sequence and the dynamic bimodality index. A dynamic bimodality prior gating module is used to modulate the dynamic bimodality index into a bimodality prior modulation feature map; The deformation rate risk pooling module is used to perform weighted pooling on the bimodal prior modulation feature map to obtain a sample-level lithium plating discrimination feature vector. The lithium plating state classification module is used to output the battery lithium plating state identification result based on the sample-level lithium plating discrimination feature vector.

5. The battery lithium plating state identification method based on stress fluctuation feature extraction according to claim 4, characterized in that, The multi-resolution temporal feature encoding module extracts a multi-resolution temporal feature map that simultaneously characterizes short-time rebound spikes and long-time relaxation distortion based on the stress relaxation deformation rate sequence and the dynamic bimodality index, including: The stress relaxation deformation rate sequence is converted into a single-channel time series tensor. Local-scale convolution is performed on the single-channel deformation rate input tensor to obtain a local-scale temporal feature map; Perform context-scale convolution on the single-channel deformation rate input tensor to obtain a context-scale temporal feature map; The local-scale temporal feature map and the context-scale temporal feature map are concatenated along the channel dimension to obtain the multi-resolution temporal feature map.

6. The battery lithium plating state identification method based on stress fluctuation feature extraction according to claim 5, characterized in that, The dynamic bimodality prior gating module modulates the dynamic bimodality index into a bimodality prior modulation feature map, including: The dynamic bimodality index is standardized to obtain the standardized dynamic bimodality prior. The standardized dynamic bimodal prior is input into the dynamic bimodal prior embedding module, and then passed through the first fully connected layer, the linear rectified activation function, and the second fully connected layer to obtain the channel-level gating vector. The channel-level gated vector is copied along the time dimension to obtain the channel-level gated matrix, and the channel-level gated matrix is ​​multiplied element-wise with the multi-resolution time-domain feature map to obtain the bimodal prior modulation feature map.

7. The battery lithium plating state identification method based on stress fluctuation feature extraction according to claim 6, characterized in that, The deformation rate risk pooling module performs weighted pooling on the bimodal prior modulation feature map to obtain a sample-level lithium plating discrimination feature vector, including: A normalized time risk weight is generated based on the stress relaxation deformation rate sequence; The normalized time risk weights are used to perform weighted pooling on the bimodal prior modulation feature map along the time dimension to obtain the deformation rate risk pooling feature vector. Max pooling is performed along the time dimension on the bimodal prior modulation feature map to obtain the maximum response feature vector; The deformation rate risk pooling feature vector, the maximum response feature vector, and the standardized dynamic bimodal prior are concatenated along the feature dimension to obtain the sample-level lithium plating discrimination feature vector.

8. A battery lithium plating state identification system based on stress fluctuation feature extraction, employing the battery lithium plating state identification method based on stress fluctuation feature extraction as described in any one of claims 1 to 7, characterized in that, include: The data acquisition module is used to acquire the raw strain time series data of the battery cycle and to standardize the raw strain time series data to obtain standardized strain time series data. The baseline determination module is used to calculate the strain change rate based on the standardized strain time series data and adaptively determine the relaxation baseline time set and the relaxation baseline mean. The deformation rate calculation module is used to calculate the local mechanical bending degree at multiple scales based on the standardized strain time series data and the mean value of the relaxation baseline, and combine the baseline deviation normalization to obtain the stress relaxation deformation rate sequence. The bimodality calculation module is used to construct the deformation rate time weight based on the stress relaxation deformation rate sequence, and calculate the dynamic bimodality index characterizing the heavy tail and multi-peak distortion of the strain waveform. The state recognition module is used to perform feature fusion and classification processing based on the stress relaxation deformation rate sequence and the dynamic bimodality index to obtain the battery lithium plating state recognition result.

Images