Method for detecting health state of battery changing cabinet based on deep learning

Abstract

The invention discloses a deep learning-based battery replacement cabinet health state detection method, and relates to the technical field of battery management, a time sequence feature extraction layer can effectively capture the long-term degradation trend of a battery in a complex environment, and misjudgment caused by short-term fluctuation is avoided; the dynamic weighting layer driven by environmental conditions adaptively focuses on a high-temperature and high-humidity period through an attention mechanism, so that the sensitivity to environmental stress related abnormity is enhanced, and the risk of missing report is reduced; the antagonism training module improves the robustness of the model in an extreme environment by simulating high-temperature and high-humidity disturbance, so that the model can distinguish a real fault from noise interference without depending on manual threshold adjustment; the real-time performance and adaptability of detection are ensured through an online sliding window mechanism and edge deployment, station changes and seasonal fluctuations can be dynamically responded, and the operation and maintenance cost is reduced.

Description

Technical Field

[0001] This invention relates to the field of battery management technology, and in particular to a method for detecting the health status of battery swapping cabinets based on deep learning. Background Technology

[0002] As an energy supply infrastructure for electric vehicles, battery swapping cabinets operate in unstable environments such as high temperature and humidity in summer. Under these conditions, condensation and corrosion are prone to occur inside the cabinet, causing fluctuations in the contact resistance of terminals and busbars, drift of sensor and insulation parameters, and amplification of cell measurement noise through thermo-electric coupling mechanism, thus masking early fault characteristics such as slight increase in internal resistance and localized heating.

[0003] Existing health status detection methods mostly employ data-driven approaches, such as Gaussian processes and statistical thresholds, which use voltage / current / temperature time series to construct probabilistic models to identify anomalies. However, under conditions of high temperature and humidity leading to strong noise, heteroscedasticity, and concept drift, the prior kernels / fixed thresholds relied upon by these methods are insufficiently adaptable to environmental disturbances. On the one hand, it is difficult to distinguish between environmentally induced parameter fluctuations and actual degradation; on the other hand, the stability of the thresholds decreases during sudden environmental changes, leading to missed or false alarms of early faults. Therefore, it is necessary to propose a health status detection method that is adaptive to the environment-degradation coupling in humid and hot and condensation scenarios, in order to improve its robustness and sensitivity in actual operation. Summary of the Invention

[0004] In view of the aforementioned existing problems, the present invention is proposed.

[0005] This invention provides a deep learning-based method for detecting the health status of battery swapping cabinets, which solves the problem of missed early battery faults caused by existing detection methods in high temperature and high humidity environments due to environmental noise interference and insufficient model adaptability.

[0006] To solve the above-mentioned technical problems, the present invention provides the following technical solution:

[0007] This invention provides a deep learning-based method for detecting the health status of battery swapping cabinets, comprising:

[0008] Step S1: Collect multi-source time-series data of batteries in the battery swapping cabinet. The multi-source time-series data includes at least the voltage, current, and temperature of the battery cells or modules, as well as the temperature and relative humidity of the cabinet environment.

[0009] Step S2: Preprocess the multi-source time series data to obtain a standardized time series tensor. The preprocessing includes time alignment, missing value imputation, detrending and deseasonalization, and numerical scaling.

[0010] Step S3: Input the standardized temporal tensor into a deep learning model, which includes a temporal feature extraction layer and an environment-driven dynamic weighting layer, and outputs a health status score.

[0011] Step S4: Based on the health status score and the judgment threshold, determine whether the battery has an early failure.

[0012] Step S5: Deploy the trained deep learning model on the edge computing device of the battery swapping cabinet, and update the health status score online using a sliding window mechanism.

[0013] As a preferred embodiment of the deep learning-based health status detection method for battery swapping cabinets described in this invention, the temporal feature extraction layer adopts a gated recurrent unit network or a one-dimensional temporal convolutional network.

[0014] As a preferred embodiment of the deep learning-based health status detection method for battery swapping cabinets described in this invention, the dynamic weighting layer adopts an attention mechanism and calculates segment-level attention weights for time segments based on an environmental encoding vector composed of ambient temperature and relative humidity.

[0015] The calculation method for the fragment-level attention weights is as follows:

[0016] The normalized temporal tensor is divided into segments using a fixed window. From each segment, we obtain segment-level temporal feature vectors. With the corresponding environment encoding vector This aggregation is accomplished jointly by a temporal feature extraction layer and environmental data statistics.

[0017] Calculate the first in the public attention subspace The matching energy of each segment is used as an attention precursor:

[0018] ,

[0019] in, Indicates the first The matching score of each segment. Indicates the first Temporal feature vectors of each segment, This represents the context encoding vector of the segment. and These represent the mapping matrices to the common subspace, A constant representing the dimension of the common subspace. This represents the score bias scalar. This represents the fragment index, with a value range of 100. ;

[0020] Introduce an amplification factor based on temperature and humidity thresholds to the matching score:

[0021] ,

[0022] in, Indicates the first The threshold modulation coefficients of each segment, and These represent the environmental temperature and relative humidity statistics for this segment, respectively. and These represent the temperature and relative humidity thresholds, respectively. and These represent the amplification coefficients of the two channels, and These represent the slope coefficients of the two channels, respectively. Represents the Sigmoid function;

[0023] By combining invalid fragment masking and normalization, fragment-level attention weights are obtained:

[0024] ,

[0025] in, Indicates the first Attention weights for each segment, Indicates the matching score. Represents the threshold modulation coefficient. Indicates the masking indicator quantity, take This segment is not included in the normalization process. Indicates the total number of segments. This represents a nonnegative constant representing the stable term in the denominator. Indicates the summation index;

[0026] When all segments are masked, a rollback strategy is used; adjustments are made according to the site strategy during seasonal changes. and .

[0027] As a preferred embodiment of the deep learning-based method for detecting the health status of a battery swapping cabinet as described in this invention, the deep learning model adopts adversarial training during the training phase by adding data perturbations simulating high temperature and high humidity conditions to the input time series.

[0028] The data perturbation simulating high temperature and high humidity conditions is generated in the following way:

[0029] Based on historical site data, the empirical distribution and correlation of ambient temperature and relative humidity are estimated, and the amplitude range of the joint disturbance is given:

[0030] Temperature disturbance Unit: °C; humidity disturbance is taken as... The unit is %RH;

[0031] Constructing a two-dimensional AR(1) process:

[0032] ,

[0033] in, Indicates time The joint temperature and humidity perturbation vector, This represents the autocorrelation coefficient, used to control the strength of the time correlation of the disturbance. This represents a zero-mean Gaussian noise vector. This represents a two-dimensional positive definite covariance matrix, the values ​​of which are derived from historical operating condition estimates. Represents the zero vector. Indicates transpose; to satisfy amplitude constraints, after sampling... and Perform interval projection to and ;

[0034] Environmental disturbances are injected into the electrical and thermal channels in a linear coupling manner to form adversarial examples used during training:

[0035] ,

[0036] in, This represents the three-channel sequence of voltage, current, and cell temperature after the disturbance. Represents the original sequence. This represents the coupling matrix, whose elements encode the temperature and humidity sensitivity of each channel. This represents a small-amplitude zero-mean noise vector, used to avoid completely deterministic linear drift;

[0037] A moment-matching-based statistical consistency term is added to the training objective to keep the perturbed environmental trajectory close to the true distribution.

[0038] ,

[0039] in, This represents the statistical consistency loss. These represent the weighting coefficients for the mean and covariance terms, respectively. These represent the perturbation sequences respectively. The estimated empirical mean and covariance, Let these represent the empirical mean and covariance of historical operating conditions, respectively. and Let them represent the 2-norm and the Frobenius norm, respectively;

[0040] During the injection phase, Apply engineering constraints:

[0041] The voltage change rate and cell temperature rise rate are kept within the achievable range of the equipment; any occurrence exceeding this range is unacceptable. or The samples were regressed to boundary values. The symbols and relative magnitudes are derived from historical regression or prior knowledge and are fine-tuned during adversarial training.

[0042] As a preferred embodiment of the deep learning-based method for detecting the health status of battery swapping cabinets described in this invention, the deep learning model is trained using a focus loss function for classification or scoring.

[0043] As a preferred embodiment of the deep learning-based health status detection method for battery swapping cabinets described in this invention, the missing value interpolation adopts temporal interpolation based on time neighborhood trends.

[0044] The de-seasonalization process decomposes and removes environmental fluctuations that include daily or weekly cycles.

[0045] As a preferred embodiment of the deep learning-based health status detection method for battery swapping cabinets described in this invention, the health status score ranges from 0 to 1, and a maintenance alarm is triggered when the score is below a threshold.

[0046] The threshold is determined based on the quantiles of the healthy sample set and is adaptively corrected according to environmental condition coding.

[0047] As a preferred embodiment of the deep learning-based health status detection method for battery swapping cabinets described in this invention, the sliding window mechanism sets a fixed window length and step size, and performs online updates of normalized parameters on the data entering the window on the edge device.

[0048] As a preferred embodiment of the deep learning-based health status detection method for battery swapping cabinets described in this invention, the deep learning model further includes an uncertainty estimation head, which is used to output the uncertainty estimate corresponding to the score, and to suppress alarms or delay judgment when the uncertainty is higher than a set upper limit.

[0049] As a preferred embodiment of the deep learning-based health status detection method for battery swapping cabinets described in this invention, the training process further includes a cross-site or cross-cabinet domain adaptation step, which reduces the impact of data distribution differences between different sites and different cabinets on the health status score through distribution alignment constraints or adversarial domain discriminators.

[0050] The beneficial effects of this invention are as follows: This invention integrates environmental condition-driven mechanisms through a deep learning model, thereby achieving more accurate and reliable battery health status detection.

[0051] The temporal feature extraction layer of this invention can effectively capture the long-term degradation trend of batteries in complex environments, avoiding misjudgments caused by short-term fluctuations. The environmental condition-driven dynamic weighting layer adaptively focuses on high-temperature and high-humidity periods through an attention mechanism, enhancing the sensitivity to environmental stress-related anomalies and thus reducing the risk of missed detections. The adversarial training module improves the robustness of the model in extreme environments by simulating high-temperature and high-humidity disturbances, enabling it to distinguish between real faults and noise interference without relying on manual threshold adjustments. The online sliding window mechanism and edge deployment ensure the real-time and adaptive nature of detection, dynamically responding to site changes and seasonal fluctuations, reducing operation and maintenance costs. In addition, uncertainty estimation and domain adaptation steps further improve the model's generalization ability, ensuring stable performance in different battery swapping cabinets and environments.

[0052] This invention achieves closed-loop management from data acquisition to fault diagnosis through algorithm optimization, which not only improves detection accuracy but also enhances the system's adaptability and reliability, providing technical support for the long-term stable operation of the battery swapping cabinet in challenging scenarios such as high temperature and high humidity. Attached Figure Description

[0053] To more clearly illustrate the technical solutions of the embodiments of this application, the accompanying drawings used in the embodiments will be briefly described below. It should be understood that the following drawings only show some embodiments of this application and should not be regarded as a limitation on the scope of this application.

[0054] Figure 1 This is a flowchart illustrating the deep learning-based health status detection method for battery swapping cabinets in this embodiment. Detailed Implementation

[0055] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0056] All terms used in this application (including technical and scientific terms) have the meanings commonly understood by those skilled in the art, unless otherwise defined. It should be noted that the terms used herein should be interpreted in a manner consistent with the context of this specification, and not in an idealized or overly rigid way.

[0057] For example, the terms “first” and “second” used in this application are only used to distinguish similar objects and differentiate the first object from another object, rather than to describe a specific order or sequence, nor should they be interpreted as indicating or implying relative importance.

[0058] This application proposes a deep learning-based method for detecting the health status of battery swapping cabinets, combining... Figure 1 As shown, the method includes:

[0059] Step S1: Collect multi-source time-series data of the batteries in the battery swapping cabinet. The multi-source time-series data includes at least the voltage, current, and temperature of the battery cells or modules, as well as the temperature and relative humidity of the environment inside the cabinet. In this embodiment, the multi-source time-series data comes from the existing measurement channels of the battery swapping cabinet and is collected with a unified timestamp. The temperature of the battery cells or modules refers to the temperature of the measurement point adjacent to the individual cells or sub-modules. The ambient temperature and relative humidity refer to the measured values ​​of the environmental sensing channels inside the cabinet. The default sampling period is 1 second, which can be adjusted from 0.2 to 5 seconds, depending on the site communication load and historical identification error. When the sampling periods are not synchronized, the timestamps are aligned to the finest granularity and the alignment deviation is limited to no more than 200ms. If the environmental channel is missing for a short period of time, the environmental code for that period is marked as empty and the original electrical channel is retained. If the missing measurement exceeds 300 seconds continuously, that period is entered into the invalid segment set for subsequent masking.

[0060] Step S2: Preprocess the multi-source time series data to obtain a standardized time series tensor. The preprocessing includes time alignment, missing value imputation, detrending and deseasonalization, and numerical scaling.

[0061] Step S3: Input the standardized temporal tensor into the deep learning model. The deep learning model includes a temporal feature extraction layer and an environment-driven dynamic weighting layer, and outputs a health status score. For example, the health status score is a numerical measure of the probability of early failure, ranging from 0 to 1. By default, the score is output once for each sliding step and monotonically calibrated on the validation set to enhance interpretability. When there are outliers in the input sequence that exceed the acceptable range, they are truncated before being fed into the model. The upper and lower limits of truncation are set by the high and low percentiles of the training set distribution.

[0062] Step S4: Based on the health status score and the judgment threshold, determine whether the battery has an early fault. Similarly, the judgment threshold is set to the low percentile of the health sample score distribution by default, usually 5% to 10%, and the minimum duration for the score to continuously meet the threshold condition is not less than 2 sliding steps to avoid jitter. When the uncertainty index is high, the minimum duration is increased to 3 to 4 sliding steps or the observation state is temporarily maintained until the uncertainty drops.

[0063] Step S5: Deploy the trained deep learning model on the edge computing device of the battery swapping cabinet, and update the health status score online using a sliding window mechanism;

[0064] Optionally, the default length of the sliding window is 60~120s and the step size is 10~20s, which is set according to the distribution of site load and event duration. During online updates, the normalized statistics are updated exponentially to reduce the impact of seasonal and site differences, and the default half-life is 30~60min. When insufficient edge resources cause calculation lag, the window length is kept unchanged and the step size is increased to within 30s, while the lag interval is recorded for subsequent supplementary calculations.

[0065] In one embodiment, the temporal feature extraction layer employs a gated recurrent unit network (GRN) or a one-dimensional temporal convolutional network (TCNN) to extract long-term dependency features across windows. Furthermore, the GRN is suitable for scenarios with long correlations, while the TCNN is suitable for scenarios with stable local patterns and requiring parallel inference. The default network depth is 1 to 3 layers, and the channel width is between 32 and 128, determined based on the long-term dependency capture capability and edge latency on the validation set. When the input length is short or the site latency constraint is strict, it can degenerate into a shallow structure to ensure real-time performance.

[0066] In one embodiment, the dynamic weighting layer employs an attention mechanism and calculates segment-level attention weights for time segments based on an environmental encoding vector composed of ambient temperature and relative humidity, in order to highlight the representation of environmental stress-related time periods.

[0067] The calculation method for fragment-level attention weights is as follows:

[0068] The normalized temporal tensor is divided into segments using a fixed window. From each segment, we obtain segment-level temporal feature vectors. With the corresponding environment encoding vector This aggregation is jointly completed by the temporal feature extraction layer and environmental data statistics, providing input for subsequent matching and weighting.

[0069] Calculate the first in the public attention subspace The matching energy of each segment is used as an attention precursor:

[0070] ,

[0071] in, Indicates the first The matching score of each segment. Indicates the first Temporal feature vectors of each segment, This represents the context encoding vector of the segment. and These represent the mapping matrices to the common subspace, A constant representing the dimension of the common subspace. This represents the score bias scalar. This represents the fragment index, with a value range of 100. ;

[0072] An amplification based on temperature and humidity thresholds is introduced into the matching score to highlight environmental stress-related segments:

[0073] ,

[0074] in, Indicates the first The threshold modulation coefficients of each segment, and These represent the environmental temperature and relative humidity statistics (such as the mean within the segment) for that segment. and These represent the temperature and relative humidity thresholds, respectively. and These represent the amplification coefficients of the two channels, and These represent the slope coefficients of the two channels, respectively. Represents the Sigmoid function;

[0075] By combining invalid fragment masking and normalization, fragment-level attention weights are obtained:

[0076] ,

[0077] in, Indicates the first Attention weights for each segment, Indicates the matching score. Represents the threshold modulation coefficient. Indicates the masking indicator quantity, take This segment is not included in the normalization process. Indicates the total number of segments. This represents a nonnegative constant representing the stable term in the denominator. Indicates the summation index;

[0078] When all segments are masked, a rollback strategy is adopted (such as using the most recent valid window or temporary uniform distribution). During seasonal changes, the site strategy can be adjusted accordingly. and The above parameters can be used in conjunction with the online update mechanism;

[0079] In this embodiment, the number of segments is determined by the window length and sampling period, typically ranging from tens to hundreds. The environment encoding vector consists of the environmental temperature and relative humidity statistics within the segment, prioritizing the use of the mean within the segment, and may include the maximum value or rate of change if necessary. The common subspace dimension is set to 16-128 by default, specifically selected based on the matching stability on the validation set. The bias term of the matching score is learned during training. The temperature and humidity thresholds for threshold modulation are recommended to be in the range of 35-45 degrees Celsius and 80%-95% relative humidity, with the amplification and slope coefficients searched within the ranges of 0.5-2.0 and 1.0-10.0, respectively. The masking indicator is determined by the missing rate and anomaly ratio within the segment; it is set to invalid when the missing rate exceeds 30% or the number of anomalies exceeds 20% of the segment length consecutively. The normalization stability term is a positive number not less than one ten-thousandth to avoid an excessively small denominator. When all segments are masked, the weight of the most recent effective window or uniform weight is used for backtracking, and automatic recovery occurs when the next effective window arrives.

[0080] Specifically, this involves a three-stage attention chain consisting of matching, threshold modulation, and masking normalization. The matching stage compares fragment representations with environmental encodings in a unified subspace using bilinear projection followed by inner product, reducing bias caused by differences in the original scale. Introducing dimensional scaling makes the numerical range more stable. The threshold modulation stage uses temperature and humidity as the dominant signals, providing smooth amplification to high-temperature or high-humidity regions through differentiable gate functions, giving higher weights to fragments related to environmental stress, while maintaining an approximately constant response in the normal range, avoiding over-amplification of non-critical fragments. The masking normalization stage handles fragments with missing sensors, poor quality, or deemed invalid, stabilizing the normalization denominator and preventing a single anomalous energy from dominating the overall allocation. This structure is compatible with adversarial training, focus loss, and online sliding window mechanisms, with clear training and deployment paths, facilitating operation on edge devices at low resource costs, and adapting to different sites and seasonal scenarios through a small number of threshold and coefficient adjustments.

[0081] In one embodiment, the deep learning model employs adversarial training during the training phase, which enhances the model's robustness in scenarios with high temperature and humidity risks by adding data perturbations that simulate high temperature and humidity conditions to the input time series.

[0082] The data perturbation for simulating high temperature and high humidity conditions is generated as follows:

[0083] Based on historical site data, the empirical distribution and correlation of ambient temperature and relative humidity are estimated, and the amplitude range of the joint disturbance is given:

[0084] Temperature disturbance Unit: °C; humidity disturbance is taken as... The unit is %RH; it is recommended to select the upper and lower bounds from the high percentile range (such as P90–P99) to cover humid and hot scenarios while preserving physical accessibility;

[0085] To obtain the time-dependent combined high temperature and high humidity perturbation, a two-dimensional AR(1) process is constructed:

[0086] ,

[0087] in, Indicates time The joint temperature and humidity perturbation vector, This represents the autocorrelation coefficient, used to control the strength of the time correlation of the disturbance. This represents a zero-mean Gaussian noise vector. This represents a two-dimensional positive definite covariance matrix, the values ​​of which are derived from historical operating condition estimates. Represents the zero vector. Indicates transpose; to satisfy amplitude constraints, after sampling... and Perform interval projection to and ;

[0088] Environmental disturbances are injected into the electrical and thermal channels in a linear coupling manner to form adversarial examples used during training:

[0089] ,

[0090] in, This represents the three-channel sequence of voltage, current, and cell temperature after the disturbance. Represents the original sequence. This represents the coupling matrix, whose elements encode the temperature and humidity sensitivity of each channel. This represents a small-amplitude zero-mean noise vector, used to avoid completely deterministic linear drift;

[0091] To reflect thermal inertia, it is possible to perform the following during implementation: First, perform short-term memory smoothing (such as an exponential window), then substitute it into the above formula;

[0092] A moment-matching-based statistical consistency term is added to the training objective to keep the perturbed environmental trajectory close to the true distribution.

[0093] ,

[0094] in, This represents the statistical consistency loss. These represent the weighting coefficients for the mean and covariance terms, respectively. These represent the perturbation sequences respectively. The estimated empirical mean and covariance, Let these represent the empirical mean and covariance of historical operating conditions, respectively. and Let them represent the 2-norm and the Frobenius norm, respectively;

[0095] During the injection phase, Apply engineering constraints:

[0096] The voltage change rate and cell temperature rise rate are kept within the achievable range of the equipment; any occurrence exceeding this range is unacceptable. or The samples were regressed to boundary values. The symbols and relative magnitudes are derived from historical regression or prior knowledge and are fine-tuned in adversarial training;

[0097] Specifically, the temporal correlation is controlled by first-order autocorrelation, with a recommended range of 0.6 to 0.95 to align with the gradual changes in the environment; the upper and lower bounds for temperature perturbation are recommended to be ±3 to 8 degrees Celsius, and the upper and lower bounds for humidity perturbation are recommended to be 10% to 20% relative humidity; the elements of the coupling matrix are initialized based on the channel sensitivity obtained from historical regression. The voltage channel typically has a weaker direct response to environmental perturbations, the cell temperature channel has a stronger response, and the current channel is in between; small-amplitude noise is used to avoid completely linear drift, with an amplitude not exceeding one-tenth of the standard deviation of the corresponding channel; the weight of the statistical consistency loss is adjusted on the validation set to match the mean and covariance of the perturbation distribution with historical operating conditions; the proportion of adversarial samples in each training iteration is recommended to be 20% to 50%. When a perturbation causes the engineering constraints to be breached, the corresponding value is projected back to the preset boundary and the sample is retained for training robustness.

[0098] Specifically, the perturbation generation strategy adopts a combination of statistical drive and physical constraints: first, a two-dimensional AR(1) process is used to obtain time-related temperature and humidity joint perturbations, with the amplitude range coming from the high quantile to avoid rare and unreachable extreme values. Then, the environmental perturbations are mapped to voltage, current and cell temperature channels using a coupling matrix, and small-amplitude noise and short-memory smoothing are configured to make the injected sequence have identifiable damp heat characteristics while maintaining continuity with the original trajectory. The statistical consistency term uses moment matching on the mean and covariance of the perturbation trajectory to converge the synthetic samples to the distribution neighborhood of the real working conditions, which is convenient for improving the model's sensitivity and robustness to damp heat and condensation risks in adversarial training. The amplitude and physical consistency rules provide feasible deployment boundaries, which, together with the channel sensitivity obtained from historical regression, limit the injection form, thereby reducing the interference of unreasonable samples on training and improving the reliability of implementation.

[0099] In one embodiment, the classification or scoring training of the deep learning model employs a focus loss function to improve the detection capability in situations where early fault samples are scarce.

[0100] For example, the default modulation factor is around 2 and is adjusted between 1 and 4 to enhance attention to difficult samples; when the classes are imbalanced, the faulty class is given a higher weight, which is usually set in the range of 0.25 to 0.5 to balance recall and false positive rate; when gradient instability occurs in the early stage of training, several rounds are warmed up without modulation before modulation is turned on.

[0101] In one embodiment, missing value imputation uses temporal interpolation based on temporal neighborhood trends; similarly, short-term missing values ​​are mainly imputed using linear trends, and when continuous missing values ​​exceed 60s, fragment extrapolation is used and the fragment is marked as low confidence for subsequent masking; if the imputation result introduces unreasonable fluctuations, it is backed to the endpoint of the adjacent observation interval by amplitude limit.

[0102] Seasonality removal decomposes and eliminates environmental fluctuations that include daily or weekly cycles. Optionally, the mean of the residuals after seasonality removal should be close to zero and the variance should decrease significantly. If the variance decreases by less than 20% after processing, it indicates that the seasonality is not significant, and this step can be skipped automatically to avoid overfitting. When there are large-scale phenomena caused by holidays or extreme weather, a shorter day-ahead window is used to reduce cross-day leakage.

[0103] In one embodiment, the health status score is between 0 and 1, and a maintenance alarm is triggered when the score is below a threshold.

[0104] The threshold is determined based on the quantiles of the healthy sample set and is adaptively corrected according to the environmental conditions. Furthermore, the default threshold corresponds to the 5th percentile of the health score distribution. Under high temperature or high humidity conditions, the threshold is reduced by a fixed amount to improve the detection rate, with a common reduction of 0.02 to 0.05. When the environmental conditions return to the normal range, the threshold is gradually restored to the baseline threshold, with the recovery time usually set to 10 to 30 minutes to avoid frequent switching.

[0105] In one embodiment, the sliding window mechanism sets a fixed window length and step size, and updates the normalized parameters of the data entering the window online on the edge device to adapt to site or seasonal changes. In this embodiment, the window length and step size are determined by offline evaluation from site to site, prioritizing coverage of key dynamics in a single charge-discharge cycle. The online update of the normalized parameters adopts an exponential method, with a default half-life of 30-60 minutes. When site switching or seasonal changes cause abrupt changes in distribution, the half-life is temporarily shortened to 10-20 minutes to accelerate convergence. If batch delays occur, the statistics of the most recent window are updated first, and then the historical windows are recalculated.

[0106] In one embodiment, the deep learning model further includes an uncertainty estimation header for outputting an uncertainty estimate corresponding to the score, and suppressing alarms or delaying the determination when the uncertainty exceeds a set upper limit; optionally, the upper limit of uncertainty is 0.3 by default and can be adjusted between 0.2 and 0.4; when the uncertainty of multiple consecutive windows exceeds the upper limit, the corresponding score is only used for trend observation and does not trigger an alarm, while extending the minimum duration requirement; when the uncertainty rises briefly and the subsequent window returns to normal, the original determination is retained to avoid frequent flipping.

[0107] In one embodiment, the training process also includes a cross-site or cross-cabinet domain adaptation step, which reduces the impact of data distribution differences between different sites and cabinets on health status scores through distribution alignment constraints or adversarial domain discriminators. Furthermore, domain adaptation is periodically enabled during multi-site mixed training, usually with several regular training iterations as one cycle, and distribution alignment is performed at the end of each cycle. The source domain to target domain sample ratio is recommended to be between 1:1 and 3:1 to balance stability and generalization. The alignment effect is based on the score stability and detection rate on the target domain validation set. If the expected results are not achieved, the alignment strength is increased or the alignment cycle is extended. When the target domain samples are very few, weak alignment is used first to avoid overfitting.

[0108] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit them. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features therein. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of this application.

[0109] Furthermore, those skilled in the art will understand that although some embodiments herein include certain features included in other embodiments but not others, combinations of features from different embodiments are meant to be within the scope of this application and form different embodiments. For example, all the embodiments above can be used in any combination. The information disclosed in this background section is intended only to enhance the understanding of the general background of this application and should not be construed as an admission or in any way implying that such information constitutes prior art known to those skilled in the art.

Claims

1. A method for detecting the health status of a battery swapping cabinet based on deep learning, characterized in that, include: Step S1: Collect multi-source time-series data of batteries in the battery swapping cabinet. The multi-source time-series data includes at least the voltage, current, and temperature of the battery cells or modules, as well as the temperature and relative humidity of the cabinet environment. Step S2: Preprocess the multi-source time series data to obtain a standardized time series tensor. The preprocessing includes time alignment, missing value imputation, detrending and deseasonalization, and numerical scaling. Step S3: Input the standardized temporal tensor into a deep learning model, which includes a temporal feature extraction layer and an environment-driven dynamic weighting layer, and outputs a health status score. Step S4: Based on the health status score and the judgment threshold, determine whether the battery has an early failure. Step S5: Deploy the trained deep learning model on the edge computing device of the battery swapping cabinet, and update the health status score online using a sliding window mechanism.

2. The method for detecting the health status of a battery swapping cabinet based on deep learning as described in claim 1, characterized in that, The temporal feature extraction layer employs a gated recurrent unit network or a one-dimensional temporal convolutional network.

3. The method for detecting the health status of a battery swapping cabinet based on deep learning as described in claim 1, characterized in that, The dynamic weighting layer employs an attention mechanism and calculates segment-level attention weights for time segments based on an environmental encoding vector composed of ambient temperature and relative humidity. The calculation method for the fragment-level attention weights is as follows: The normalized temporal tensor is divided into segments using a fixed window. From each segment, we obtain segment-level temporal feature vectors. With the corresponding environment encoding vector This aggregation is accomplished jointly by a temporal feature extraction layer and environmental data statistics. Calculate the first in the public attention subspace The matching energy of each segment is used as an attention precursor: ， in, Indicates the first The matching score of each segment. Indicates the first Temporal feature vectors of each segment, This represents the context encoding vector of the segment. and These represent the mapping matrices to the common subspace, A constant representing the dimension of the common subspace. This represents the score bias scalar. This represents the fragment index, with a value range of 100. ; Introduce an amplification factor based on temperature and humidity thresholds to the matching score: ， in, Indicates the first The threshold modulation coefficients of each segment, and These represent the environmental temperature and relative humidity statistics for this segment, respectively. and These represent the temperature and relative humidity thresholds, respectively. and These represent the amplification coefficients of the two channels, and These represent the slope coefficients of the two channels, respectively. Represents the Sigmoid function; By combining invalid fragment masking and normalization, fragment-level attention weights are obtained: ， in, Indicates the first Attention weights for each segment, Indicates the matching score. Represents the threshold modulation coefficient. Indicates the masking indicator quantity, take This segment is not included in the normalization process. Indicates the total number of segments. This represents a nonnegative constant representing the stable term in the denominator. Indicates the summation index; When all segments are masked, a rollback strategy is used; adjustments are made according to the site strategy during seasonal changes. and .

4. The method for detecting the health status of a battery swapping cabinet based on deep learning as described in claim 1, characterized in that, The deep learning model employs adversarial training during the training phase, by adding data perturbations simulating high temperature and high humidity conditions to the input time series; The data perturbation simulating high temperature and high humidity conditions is generated in the following way: Based on historical site data, the empirical distribution and correlation of ambient temperature and relative humidity are estimated, and the amplitude range of the joint disturbance is given: Temperature disturbance Unit: °C; humidity disturbance is taken as... The unit is %RH; Constructing a two-dimensional AR(1) process: ， in, Indicates time The joint temperature and humidity perturbation vector, This represents the autocorrelation coefficient, used to control the strength of the time correlation of the disturbance. This represents a zero-mean Gaussian noise vector. This represents a two-dimensional positive definite covariance matrix, the values ​​of which are derived from historical operating condition estimates. Represents the zero vector. Indicates transpose; to satisfy amplitude constraints, after sampling... and Perform interval projection to and ; Environmental disturbances are injected into the electrical and thermal channels in a linear coupling manner to form adversarial examples used during training: ， in, This represents the three-channel sequence of voltage, current, and cell temperature after the disturbance. Represents the original sequence. This represents the coupling matrix, whose elements encode the temperature and humidity sensitivity of each channel. This represents a small-amplitude zero-mean noise vector, used to avoid completely deterministic linear drift; A moment-matching-based statistical consistency term is added to the training objective to keep the perturbed environmental trajectory close to the true distribution. ， in, This represents the statistical consistency loss. These represent the weighting coefficients for the mean and covariance terms, respectively. These represent the perturbation sequences respectively. The estimated empirical mean and covariance, Let these represent the empirical mean and covariance of historical operating conditions, respectively. and Let them represent the 2-norm and the Frobenius norm, respectively; During the injection phase, Apply engineering constraints: The voltage change rate and cell temperature rise rate are kept within the achievable range of the equipment; any occurrence exceeding this range is unacceptable. or The samples were regressed to boundary values. The symbols and relative magnitudes are derived from historical regression or prior knowledge and are fine-tuned during adversarial training.

5. The method for detecting the health status of a battery swapping cabinet based on deep learning as described in claim 1, characterized in that, The deep learning model is trained using a focus loss function for classification or scoring.

6. The method for detecting the health status of a battery swapping cabinet based on deep learning as described in claim 1, characterized in that, The missing value imputation adopts temporal interpolation based on the temporal neighborhood trend; The de-seasonalization process decomposes and removes environmental fluctuations that include daily or weekly cycles.

7. The method for detecting the health status of a battery swapping cabinet based on deep learning as described in claim 1, characterized in that, The health status score ranges from 0 to 1. When the score is below the threshold, a maintenance alarm is triggered. The threshold is determined based on the quantiles of the healthy sample set and is adaptively corrected according to environmental condition coding.

8. The method for detecting the health status of a battery swapping cabinet based on deep learning as described in claim 1, characterized in that, The sliding window mechanism sets a fixed window length and step size, and performs online updates of normalized parameters on the data entering the window on the edge device.

9. The method for detecting the health status of a battery swapping cabinet based on deep learning as described in claim 1, characterized in that, The deep learning model further includes an uncertainty estimation head, which is used to output the uncertainty estimate corresponding to the score, and to suppress alarms or delay the judgment when the uncertainty is higher than a set upper limit.

10. The method for detecting the health status of a battery swapping cabinet based on deep learning as described in claim 1, characterized in that, The training process also includes a cross-site or cross-cabinet domain adaptation step, which reduces the impact of data distribution differences between different sites and cabinets on health status scores through distribution alignment constraints or adversarial domain discriminators.

Images