Artificial intelligence-based battery fault diagnosis methods, devices, and systems

Abstract

This invention provides a battery fault diagnosis method, device, and system based on artificial intelligence, relating to the field of artificial intelligence technology. Based on determining the operating condition patterns of monitoring data for each attribute, it performs cross-feature convolutional fusion to deeply mine the collaborative changes of multiple attributes. It superimposes and enhances early, weak fault features scattered across different channels, while filtering out invalid interference from operating condition fluctuations, preventing weak fault features from being overwhelmed by normal fluctuations and noise, thus significantly improving the ability to identify early faults. Furthermore, under the premise of perturbation suppression of the cross-feature enhancement vector, it constructs a globally perceptual feature representation that combines multi-attribute collaborative features and full-cycle evolution features. This enables quantitative assessment of the gradual evolution of battery faults, allowing the diagnostic model to predict faults in advance based on fault development trends, delaying or preventing fault deterioration at its source.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence technology, and in particular to a battery fault diagnosis method, apparatus and system based on artificial intelligence. Background Technology

[0002] With the acceleration of global energy transition and electrification, the large-scale application of battery systems, represented by electric vehicles and energy storage power stations, has placed unprecedented demands on safety and reliability. During long-term operation under complex conditions, batteries will undergo slow electrochemical decay and material degradation that are difficult to observe directly. In the early stages, these changes manifest as weak abnormal fluctuations in external monitoring parameters such as voltage, current, and temperature. If they cannot be detected and warned in time at their incipient stage, minor performance degradation may evolve into catastrophic failures such as thermal runaway.

[0003] In the feature extraction stage, existing technologies often process data from each feature channel independently, lacking in-depth exploration of the collaborative changes between multiple feature channels. Since the features of early battery faults are usually quite weak, without cross-channel collaborative feature mining, these weak fault features are easily obscured by normal fluctuations and interference information, making them difficult to capture accurately. This results in a significant shortcoming of existing technologies in early fault identification, failing to detect potential battery fault hazards in a timely manner.

[0004] Furthermore, existing technologies lack an effective quantitative assessment mechanism for the progressive evolution of battery failures. This deficiency severely limits the diagnostic models' ability to warn of early failure risks and makes it impossible to accurately predict failure trends. Consequently, it fails to meet the core requirement of taking early intervention measures through failure prediction in practical applications to ensure the safe and stable operation of battery systems, further restricting the applicability and application value of existing technologies in complex real-world scenarios. Summary of the Invention

[0005] The purpose of this invention is to provide a battery fault diagnosis method, device and system based on artificial intelligence, which can accurately predict the development trend of battery faults and overcome the limitations of the existing technology in terms of early warning capabilities.

[0006] In a first aspect, embodiments of the present invention provide a battery fault diagnosis method based on artificial intelligence. The method includes: acquiring operational monitoring data of a target battery; determining the monitoring operating condition mode of the attribute data at the corresponding monitoring time point based on the feature space distribution of each attribute data in the operational monitoring data at the monitoring time point; performing cross-feature convolutional fusion on each attribute data based on the monitoring operating condition mode to generate a cross-feature enhancement vector; suppressing the signal fluctuations of the cross-feature channel of the cross-feature enhancement vector and acquiring the temporal local details and temporal global context information of the cross-feature enhancement vector under the perturbation suppression to construct a globally perceived feature representation; and classifying and identifying the globally perceived feature representation using a pre-built battery fault diagnosis model to determine the fault category of the target battery based on the operational monitoring data.

[0007] In conjunction with the first aspect, this embodiment of the invention provides a first implementation of the first aspect, wherein the step of classifying and identifying the global perception feature representation using a pre-built battery fault diagnosis model to determine the fault category of the target battery based on operational monitoring data includes: constructing an auxiliary feature vector of the global perception feature representation based on the operating condition mode distribution and signal fluctuation pattern of the global perception feature representation; determining the auxiliary classification probability distribution of the auxiliary feature vector through a lightweight fully connected network; outputting the preliminary predicted probability corresponding to the global perception feature representation using the main fault classification layer of the diagnostic model; and adaptively fusing the auxiliary classification probability distribution and the preliminary predicted probability to determine the fault category corresponding to the target battery.

[0008] In conjunction with the first aspect, this embodiment of the invention provides a second implementation of the first aspect, wherein the step of constructing an auxiliary feature vector of the global sensing feature representation based on the operating condition mode distribution and signal fluctuation pattern of the global sensing feature representation includes: determining the operating condition mode distribution of the global sensing feature representation according to the time point proportion of the global sensing feature representation in each operating condition mode cluster; determining the signal fluctuation pattern of the global sensing feature representation according to the time series skewness of the global sensing feature representation; and concatenating the operating condition mode distribution, signal fluctuation pattern, and global sensing feature representation to form an auxiliary feature vector of the global sensing feature representation.

[0009] In conjunction with the first aspect, this embodiment of the invention provides a third implementation of the first aspect, wherein the total loss function of the battery fault diagnosis model includes multi-task supervision loss and fault evolution consistency loss; the multi-task supervision loss is used to jointly supervise the auxiliary classification probability distribution, the preliminary prediction probability, and the risk score of the temporal evolution of the global perception feature representation; the fault evolution consistency loss is used to measure the linear correlation between the risk score and the proportion of abnormal operating conditions in the global perception feature representation.

[0010] In conjunction with the first aspect, this embodiment of the invention provides a fourth implementation of the first aspect, wherein the step of determining the monitoring condition mode of the attribute data at the corresponding monitoring time point based on the feature space distribution of each attribute data of the operation monitoring data at the monitoring time point includes: performing frequency domain filtering on the operation monitoring data, and adaptively normalizing the filtered operation monitoring data according to the time-series statistics of the operation monitoring data to generate preprocessed data; dividing the monitoring time points of the attribute data into multiple operating condition mode clusters according to the distribution of each attribute data of the preprocessed data in the feature space; and determining the monitoring condition mode to which the current monitoring time point belongs based on the operating condition mode clusters.

[0011] In conjunction with the first aspect, this invention provides a fifth implementation of the first aspect, wherein the step of performing frequency domain filtering on the operation monitoring data includes: determining the wavelet coefficients of the operation monitoring data at multiple scales using discrete wavelet transform processing; the wavelet coefficients are used to characterize the energy components of the operation monitoring data at a preset frequency scale at a preset monitoring time point; determining an adaptive noise suppression threshold based on the neighborhood range data of the wavelet coefficients; performing noise suppression on the operation monitoring data using the adaptive noise suppression threshold to generate filtered data; and performing adaptive normalization on the filtered operation monitoring data based on the time-series statistics of the operation monitoring data to generate preprocessed data. The step further includes: using the monitoring time point of the operation monitoring data as a time window, performing adaptive sliding window normalization on the data of each feature channel of the operation monitoring data to generate preprocessed data.

[0012] In conjunction with the first aspect, this embodiment of the invention provides a sixth implementation of the first aspect, wherein the step of perturbation suppression of signal fluctuations across feature channels of the cross-feature enhancement vector includes: calculating an adaptive mask based on the variance of the cross-feature channel of the cross-feature enhancement vector at each monitoring time point; and perturbation suppression of signal fluctuations across feature channels based on the adaptive mask.

[0013] In conjunction with the first aspect, this invention provides a seventh implementation of the first aspect, wherein the step of acquiring temporal local details and temporal global context information of the cross-feature enhancement vector under perturbation suppression and constructing a globally perceptive feature representation includes: extracting temporal local details and temporal global context information of the cross-feature enhancement vector using two paths of a preset dual-path anti-perturbation module; wherein one path is used to capture the local temporal pattern of the cross-feature enhancement vector through one-dimensional convolution, and the other path is used to capture the global temporal dependency of the cross-feature enhancement vector using a multi-head self-attention mechanism to determine the corresponding global context information; adaptively weighting and fusing the temporal local details and temporal global context information to generate temporal recognition features; and temporal max pooling is performed on the temporal recognition features to generate a globally perceptive feature representation.

[0014] Secondly, embodiments of the present invention provide an artificial intelligence-based battery fault diagnosis device, which includes: a data processing module for acquiring operational monitoring data of a target battery, and determining the monitoring operating condition mode of the attribute data at the corresponding monitoring time point based on the feature space distribution of each attribute data in the operational monitoring data at the monitoring time point; a feature fusion module for performing cross-feature convolutional fusion on each attribute data based on the monitoring operating condition mode to generate a cross-feature enhancement vector; a feature extraction module for perturbation suppression of signal fluctuations in the cross-feature channels of the cross-feature enhancement vector, and acquiring the temporal local details and temporal global context information of the cross-feature enhancement vector under perturbation suppression to construct a globally perceived feature representation; and an execution module for classifying and identifying the globally perceived feature representation using a pre-built battery fault diagnosis model to determine the fault category of the target battery based on the operational monitoring data.

[0015] Thirdly, embodiments of the present invention provide an artificial intelligence-based battery fault diagnosis system, which is configured with the apparatus of the above embodiments for executing the methods of any of the above embodiments.

[0016] The embodiments of this invention bring the following beneficial effects: This invention provides a battery fault diagnosis method, device, and system based on artificial intelligence. It performs cross-feature convolutional fusion based on the monitoring operating condition mode of determined attribute monitoring data, deeply mining the collaborative change relationship between multiple attribute data under the same operating condition. It superimposes and enhances early weak fault features scattered in different channels, while filtering out invalid interference caused by operating condition fluctuations, preventing weak fault features from being submerged by normal fluctuations and noise. By breaking the limitation of existing technologies that independently process single feature channels, it significantly improves the identifiability of early fault features, solving the problems of weak early fault identification capability and inability to detect potential hidden dangers in a timely manner in existing technologies. Furthermore, by perturbation suppression of the cross-feature enhancement vector and simultaneous extraction of temporal local details and global context information, a globally perceptual feature representation that combines multi-attribute collaborative features and full-cycle evolution features is constructed. In summary, a quantitative assessment mechanism for the progressive evolution of battery faults can be established, enabling subsequent diagnostic models to not only accurately identify fault categories but also accurately predict fault development trends. This overcomes the limitations of existing technologies in early warning capabilities and meets the core needs of practical applications to take early intervention measures through fault prediction and ensure the safe and stable operation of battery systems. It significantly expands the applicability and application value of the technical solution in complex scenarios.

[0017] Other features and advantages of the invention will be set forth in the following description, or some features and advantages may be inferred from the description or determined without doubt, or may be learned by practicing the techniques described above.

[0018] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, preferred embodiments are described below in detail with reference to the accompanying drawings. Attached Figure Description

[0019] To more clearly illustrate the specific embodiments of the present invention or the technical solutions in the prior art, the drawings used in the description of the specific embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0020] Figure 1 A flowchart illustrating an artificial intelligence-based battery fault diagnosis method provided in an embodiment of the present invention;

[0021] Figure 2 A flowchart of another artificial intelligence-based battery fault diagnosis method provided in an embodiment of the present invention;

[0022] Figure 3 The time-domain waveform of the original signal (including noise);

[0023] Figure 4 The noise spectrum of the original signal;

[0024] Figure 5 The noise spectrum after Z-Score normalization;

[0025] Figure 6 This is a noise spectrum diagram after processing according to an embodiment of the present invention;

[0026] Figure 7 A comparison chart of normalized noise spectrum;

[0027] Figure 8 This is a schematic diagram showing the distribution of different battery fault types in the feature space according to an embodiment of the present invention.

[0028] Figure 9 A schematic diagram of the structure of a battery fault diagnosis device based on artificial intelligence provided in an embodiment of the present invention;

[0029] Figure 10 This is a schematic diagram of the structure of an electronic device provided in an embodiment of the present invention. Detailed Implementation

[0030] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. The components of the embodiments of the present invention described and shown in the accompanying drawings can generally be arranged and designed in various different configurations.

[0031] Therefore, the following detailed description of the embodiments of the invention provided in the accompanying drawings is not intended to limit the scope of the claimed invention, but merely to illustrate selected embodiments of the invention. All other embodiments obtained by those skilled in the art based on the embodiments of the invention without inventive effort are within the scope of protection of the invention.

[0032] This invention provides a battery fault diagnosis method, device, and system based on artificial intelligence, which can accurately predict the development trend of battery faults and overcome the limitations of existing technologies in early warning capabilities.

[0033] To facilitate understanding, a battery fault diagnosis method based on artificial intelligence provided in an embodiment of the present invention will first be described, referring to... Figure 1 The method includes the following steps:

[0034] Step S102: Obtain the operation monitoring data of the target battery, and determine the monitoring mode of the attribute data at the corresponding monitoring time point based on the characteristic spatial distribution of each attribute data at the monitoring time point.

[0035] The target battery's operational monitoring data refers to the set of multi-dimensional state parameters collected by sensors or a BMS (Battery Management System) during actual operation, reflecting the battery's working state. This data includes multi-dimensional attribute data during battery operation, such as voltage, current, temperature, SOC (State of Charge), internal resistance, and other parameters strongly correlated with the battery's state.

[0036] This invention calculates the spatial distribution characteristics of each attribute data at each monitoring time point, thereby determining the monitoring operating condition mode of the corresponding attribute data at that time point. The spatial distribution characteristics indicate the numerical distribution pattern of a certain attribute data within a specific dimensional space (such as the time domain, frequency domain, or statistical domain), characterizing the state characteristics of the attribute data. Examples include statistical distribution (mean, variance, skewness), frequency domain distribution (dominant frequency, harmonic proportion), or time-series distribution (slope of change, fluctuation period). Furthermore, the monitoring operating condition mode of the corresponding attribute data can be determined by combining a characteristic library of operating condition modes (such as feature spatial distribution templates containing typical modes like normal operating condition, high load operating condition, low temperature operating condition, and aging operating condition).

[0037] Step S104: Based on the monitoring working condition mode, perform cross-feature convolution fusion on each attribute data to generate a cross-feature enhancement vector.

[0038] Cross-feature convolutional fusion refers to a multi-attribute data fusion method with convolution operation as its core. Cross-feature enhancement vector refers to a high-dimensional feature vector generated after fusing collaborative features of multiple attribute channels. This vector not only contains local features of a single attribute, but also contains correlation features between multiple attributes.

[0039] Unlike the traditional approach of processing single-channel features independently, this invention uses multi-dimensional attribute data as parallel input channels. Based on the operating conditions of each attribute data, it mines the synergistic relationship between different attribute data (e.g., the synchronicity of voltage drop and temperature rise, the correlation between current fluctuation and internal resistance increase), and captures the implicit synergistic change patterns between different attribute data. This allows for the superposition and enhancement of weak fault features scattered in different attribute data (e.g., small voltage fluctuations + slow increase in internal resistance). Weak fault features that are difficult to identify under a single channel are transformed into significant cross-channel synergistic features, making the weak features of early faults stand out from background noise and greatly improving the identifiability of fault features.

[0040] In summary, the embodiments of this invention can perform adaptive baseline calibration based on the operating conditions of attribute data, eliminating normal fluctuations in the baseline of attribute data under different operating conditions, filtering out invalid interference caused by operating conditions, and preventing it from masking early weak fault characteristics. Furthermore, under unified operating condition constraints, convolutional operations are used to mine the physical coupling and collaborative change relationships between multi-attribute data, breaking the limitation of existing technologies that independently process single feature channels. Weak fault features scattered across different channels are superimposed and enhanced, generating cross-feature enhancement vectors that combine single-attribute local features and multi-attribute correlated features. Ultimately, this solves the core pain point of existing technologies, which lack cross-channel collaborative feature mining and cannot accurately capture early weak fault characteristics, thus achieving effective identification of potential battery fault hazards.

[0041] Step S106: Perturbation suppression is performed on the signal fluctuations across feature channels of the cross-feature enhancement vector, and the temporal local details and temporal global context information of the cross-feature enhancement vector under perturbation suppression are obtained to construct a globally perceived feature representation.

[0042] The cross-feature channels of the cross-feature enhancement vector correspond to different attribute data dimensions before fusion. Each channel carries the features of a single attribute and the correlation features between that attribute and other attributes. The signal fluctuations refer to the random fluctuation signals unrelated to battery faults superimposed in the attribute correlation channels of the cross-feature enhancement vector, such as sensor measurement noise, environmental electromagnetic interference, and residual instantaneous disturbances in the operating condition. By filtering out invalid interference fluctuations in the cross-feature channels, fault-related features in the cross-feature enhancement vector can be purified, especially avoiding the early weak fault features being masked by noise.

[0043] Temporal local details represent subtle changes in the cross-feature enhancement vector within a short time window, corresponding to the instantaneous and weak characteristics of early battery faults (e.g., small voltage spikes within a single charge-discharge cycle). Temporal global context information represents the evolutionary trend of the cross-feature enhancement vector over a long time series, corresponding to the gradual development process of battery faults (e.g., the slow upward trend of internal resistance over 50 consecutive charge-discharge cycles). This invention, through the fusion of instantaneous weak fault features and long-term gradual fault trends under interference suppression, constructs a globally perceptive feature representation. This representation integrates multi-channel collaborative features and fuses temporal features of local mutations and global evolution, forming a high-dimensional feature carrier with multi-channel collaboration and full temporal correlation. This not only accurately characterizes the current battery state but also quantifies the gradual evolution of faults, significantly improving the accuracy of early fault prediction.

[0044] Step S108: Use the pre-built battery fault diagnosis model to classify and identify the global perception feature representation to determine the fault category of the target battery based on the operation monitoring data.

[0045] The battery fault diagnosis model is an intelligent classification model trained using deep learning algorithms (such as CNN-LSTM hybrid model and Transformer model) based on labeled battery fault datasets. It can provide early warnings based on the feature trends of globally perceived feature representations, guiding maintenance personnel to take preventive intervention measures (such as optimizing charging and discharging strategies, strengthening thermal management, and reducing load rate). This invention upgrades battery fault prediction from passive post-event identification to proactive pre-event prediction and trend control, which can delay or prevent fault deterioration from the source.

[0046] In summary, this invention, based on determining the monitoring operating conditions of each attribute monitoring data, and using the corresponding operating conditions as constraints to perform cross-feature convolutional fusion, deeply mines the collaborative change relationships between multiple attribute data under the same operating condition. It superimposes and enhances early weak fault features scattered across different channels, while filtering out invalid interference caused by operating condition fluctuations, preventing weak fault features from being submerged by normal fluctuations and noise. By breaking the limitation of existing technologies that independently process single feature channels, it significantly improves the identifiability of early fault features, solving the problems of weak early fault identification capabilities and inability to promptly detect potential hazards in existing technologies. Furthermore, by perturbation suppression of the cross-feature enhancement vector and simultaneous extraction of temporal local details and global contextual information, a globally perceptual feature representation that combines multi-attribute collaborative features and full-cycle evolution features is constructed. In summary, a quantitative assessment mechanism for the progressive evolution of battery faults can be established, enabling subsequent diagnostic models to not only accurately identify fault categories but also accurately predict fault development trends. This overcomes the limitations of existing technologies in early warning capabilities and meets the core needs of practical applications to take early intervention measures through fault prediction and ensure the safe and stable operation of battery systems. It significantly expands the applicability and application value of the technical solution in complex scenarios.

[0047] Furthermore, based on the above embodiments, this invention provides another battery fault diagnosis method based on artificial intelligence, referring to... Figure 2 The method includes the following steps:

[0048] Step S202: Obtain the operational monitoring data of the target battery.

[0049] Step S204: Frequency domain filtering is performed on the operation monitoring data, and adaptive normalization is performed on the filtered operation monitoring data based on the time series statistics of the operation monitoring data to generate preprocessed data.

[0050] Regarding step S102 above, this embodiment of the invention uses data preprocessing and cluster analysis to accurately match the operating condition patterns of attribute data at each monitoring time point. This process can eliminate interference factors in the original data and achieve automated and precise division of operating conditions. The time-series data of current, voltage, temperature, and state of charge collected by sensors are high-dimensional and have temporal correlations. Current and voltage signals are easily affected by sensor noise and operating condition fluctuations, leading to significant differences in numerical scales between different features, and the existence of local outliers and high-frequency noise in the time series. Existing data preprocessing methods are relatively crude, typically employing standardization methods based on global statistics, which fail to fully consider the local fluctuation characteristics of the time series and are insufficient in suppressing local disturbances such as sensor noise and sudden changes in operating conditions, resulting in low-quality input data and affecting the performance of subsequent models.

[0051] In one implementation, conventional methods employ Z-Score normalization based on global statistics. While this can eliminate scale differences, it fails to consider local statistical characteristics in the time dimension. Consequently, the normalized data still contains high-frequency disturbances accumulated from local temporal fluctuations, thus reducing the robustness of subsequent diagnostic models. This invention, through a two-stage processing flow integrating local temporal statistical characteristics and frequency domain filtering, effectively suppresses local sudden disturbances and residual high-frequency noise while eliminating scale differences between features. Specifically, embodiments of this invention generate preprocessed data through the following steps:

[0052] 1) Use discrete wavelet transform to determine the wavelet coefficients of the operation monitoring data at multiple scales; determine the noise suppression adaptive threshold based on the neighborhood range data of the wavelet coefficients; use the noise suppression adaptive threshold to suppress noise in the operation monitoring data and generate filtered data.

[0053] To address wavelet threshold noise suppression, this embodiment of the invention performs discrete wavelet transform on the time-series data of each feature channel to obtain its wavelet coefficients at multiple scales, characterizing the energy components of the monitoring data at a preset frequency scale at a preset monitoring time point. Then, an adaptive threshold (i.e., an adaptive threshold for noise suppression) based on neighborhood coefficient statistics is constructed, and a soft thresholding function is applied to the wavelet coefficients to suppress noise. Furthermore, the signal is reconstructed through inverse discrete wavelet transform to obtain a noise-suppressed data matrix, effectively filtering out impulse noise while preserving abrupt signal edges, as shown below:

[0054]

[0055] In the formula, Indicates the first The first feature channel, the first The output values ​​at each time point after noise suppression are effectively filtered out by wavelet threshold denoising to remove impulse noise and residual high-frequency noise, while preserving the abrupt change edges of the signal. This represents the inverse discrete wavelet transform operator, used to reconstruct a time-domain signal from the processed wavelet coefficients; Indicates the first The wavelet coefficients obtained by discrete wavelet transforming the time-series data of the nth feature channel are the nth wavelet coefficients. Wavelet coefficients obtained by discrete wavelet transform of the time-series data of each characteristic channel characterize the energy components of the original signal at a specific time and frequency scale. This function represents the sign function; if the input is greater than 0, it outputs 1; if the input is equal to 0, it outputs 0; and if the input is less than 0, it outputs -1. Indicates the use of processing the first The first feature channel, the first The adaptive threshold of wavelet coefficients at each time point is used to determine whether the wavelet coefficients are mainly caused by noise. The calculation method is expressed as follows:

[0056]

[0057] This represents the threshold scaling factor, a hyperparameter used to globally adjust the strictness of the threshold; an example value is 0.7. The threshold calculation window radius is a hyperparameter used to define the size of the neighborhood wavelet coefficient range referenced when calculating the local threshold. An example value is 10. Indicates the first From the wavelet coefficient vector of each feature channel, from the index To index Slices; This represents the median function, used to robustly estimate the center level of the absolute values ​​of local wavelet coefficients, thus enabling the threshold to adapt to noise intensities of different time periods and characteristics. Wherein, wavelet coefficients... based on The entire time series is decomposed into multiple scales through discrete wavelet transform to obtain wavelet coefficients corresponding to specific time and frequency scales.

[0058] It should be noted that, The term represents the soft threshold function, which selects terms whose absolute values ​​are less than the adaptive threshold. The coefficients are shrunk to zero, thereby suppressing coefficients caused by noise, and coefficients with absolute values ​​greater than the threshold are shrunk. While filtering out noise, the effective components of the signal are preserved as much as possible, enhancing the robustness of denoising and avoiding over-smoothing.

[0059] 2) Using the monitoring time point of the operation monitoring data as the time window, adaptive sliding window normalization is performed on the data of each feature channel of the operation monitoring data to generate preprocessed data.

[0060] In practical implementation, for each feature channel and each time point of the input data matrix, a local time window is defined centered on that time point. The mean and standard deviation of the data within this window are calculated, and the original data at that center point is normalized accordingly. This allows for dynamic adjustment of the normalization scale based on the local fluctuation characteristics of the data on the time axis, as expressed in:

[0061]

[0062] In the formula, Indicates the first The first feature channel, the first The output values ​​at each time point are normalized by the mean and standard deviation of the local window. This allows for dynamic adjustment of the normalization scale based on local fluctuations in the data over time, eliminating scale differences between features while reducing high-frequency disturbances accumulated from local temporal fluctuations. This represents the feature channel index, with a value range of [value range missing]. These correspond to four time-series characteristics: current, voltage, temperature, and state of charge, respectively. This represents a point-in-time index, with a value range of [value range missing]. ; The total number of time points is determined by the sensor's sampling frequency and sampling duration. For example, if sampling is performed at a frequency of 1 Hz for 120 seconds, then... The example value is 120.

[0063] in, Represents the original input data matrix The Middle The first feature channel, the first The original sampled values ​​at each time point (or the filtered data mentioned above), the original input data matrix The dimension is It is obtained by collecting time-series data through sensors at a fixed sampling frequency and arranged according to characteristic channels and time points; Indicates the first The feature channel in the first The local window mean at each time point is used to estimate the local signal baseline near that time point, and the calculation method is expressed as follows:

[0064]

[0065] Indicates the first The feature channel in the first The local window standard deviation at each time point is used to measure the local fluctuation amplitude of the signal near that time point, and is calculated as follows:

[0066]

[0067] The smoothing factor is a very small positive constant used to ensure that the denominator is not zero, thus avoiding numerical errors. Examples of possible values ​​are given below. ; The window radius is a hyperparameter used to control the size of the calculation range of local statistics. It defines the radius of the sliding window, with an example value of 5. Indicates the index of a time point within a local sliding window; This function takes the larger of two arguments. This represents a function that takes the smaller of two parameters.

[0068] Furthermore, this embodiment of the invention compares the signal noise suppression effects of different preprocessing methods, verifying the superiority of the proposed data adaptive normalization and local noise suppression process over conventional preprocessing methods. The experiment compared three processing methods: the unprocessed original noisy signal, the signal after conventional global Z-score normalization, and the signal using the embodiment of this invention. The experiment used a 120-second battery voltage time-series signal, artificially injecting various composite noises including high-frequency random noise, sudden impulse noise, and low-frequency slowly varying interference to simulate the complex interference present in real sensor-acquired data. The figures illustrate this from both time-domain waveform and spectral analysis perspectives.

[0069] Reference Figures 3-7 The diagrams show the time-domain waveform of the original signal (including noise), the noise spectrum of the original signal, the noise spectrum after Z-Score normalization, the noise spectrum after processing according to the embodiment of the present invention, and a comparison of normalized noise spectra. The advantages of this technique are clearly evident from the experimental results. In the time-domain diagram, the original signal exhibits severe fluctuations and contains obvious spikes; after conventional Z-score normalization, the signal scale is unified, but glitches and abrupt changes in the time-domain waveform still exist. In the noise spectrum, the blue spectrum representing the original signal noise has a high energy distribution across the entire frequency band. After conventional processing, the green spectrum is somewhat suppressed in the mid-to-high frequency range, but significant residues remain in the frequency band corresponding to impulse noise. The red spectrum corresponding to the method of the present invention shows the lowest energy level across the entire frequency axis, especially in the high-frequency region, where its power spectral density value is significantly lower than the other two. The experimental results demonstrate that the adaptive local normalization of the present invention can better follow the local fluctuations of the signal, while the subsequent wavelet threshold denoising can effectively filter out impulses and high-frequency noise, thus providing cleaner and more robust model input data for subsequent diagnostic models.

[0070] Step S206: Based on the distribution of each attribute data in the feature space of the preprocessed data, the monitoring time points of the attribute data are divided into multiple working condition mode clusters.

[0071] Step S208: Determine the monitoring operating mode to which the current monitoring time point belongs based on the operating mode cluster.

[0072] In conjunction with the above embodiments, battery faults exhibit multi-scale correlation characteristics in their temporal manifestations. For example, a slow decrease in voltage may coexist with a sudden surge in current. This invention uses unsupervised clustering to dynamically segment time series to identify typical operating conditions, and further constructs a cross-feature fusion operation based on dynamic convolution kernels (i.e., step S210). This cross-feature fusion operation dynamically adjusts the convolution weights according to the operating condition at each time point, thereby fusing multi-feature interaction information and generating more discriminative enhanced features. Conventional feature extraction techniques typically process each feature channel independently, failing to effectively capture the collaborative changes between cross-feature channels and the dynamic characteristics of fault modes evolving over time. This leads to the easy neglect of weak correlation features in the early stages of faults, resulting in insufficient overall feature representation capabilities.

[0073] In practical implementation, each column of the noise-suppressed data matrix (i.e., the preprocessed data) can be treated as an independent sample, and unsupervised partitioning can be performed using the K-means clustering algorithm. Based on the distribution of the data in the feature space, The time points are divided into Clustering can be divided into different time periods or operating condition patterns. After clustering, a cluster label can be assigned to each time point to represent the typical operating condition pattern to which that time point belongs. This represents the preset number of clusters and is a hyperparameter. In one implementation, the optimal number of clusters can be determined using the elbow rule to balance segmentation granularity and computational complexity.

[0074] Step S210: Based on the monitoring operating condition mode, perform cross-feature convolution fusion on each attribute data to generate a cross-feature enhancement vector.

[0075] This method employs convolution operations to fuse interactive information between different feature channels and utilizes the results of dynamic segmentation. The convolution slides along the time dimension, but its kernel weights are dynamically determined by the clustering pattern of the current convolution window center and its neighboring time points, thereby achieving adaptive feature extraction for different working conditions. This can be represented as:

[0076]

[0077] In the formula, Represents the feature enhancement matrix The Middle The output channel, the first The values ​​at each time point are fused by cross-feature convolution to integrate the interaction information between different feature channels, and the convolution weights are adaptively adjusted according to the dynamic segmentation results, thereby enhancing the ability of features to express multi-scale temporal correlations and fault dynamic characteristics. Indicates the first The first feature channel, the first The output value after noise suppression at each time point; This represents the output feature channel index, with a value range of [value range missing]. ; This indicates the output feature dimension, a hyperparameter that defines the number of feature channels after this enhancement step; an example value is 12. This represents the offset of the convolution operation in the time dimension, used to define the time range covered by the convolution kernel relative to the center point. The offset; The width of the convolution kernel is a hyperparameter that defines the size of the time range covered by the kernel. An example value is 5. This represents the floor function; Denotes the cross-feature convolution weight tensor, where This tensor stores the index of the cluster centers. The convolutional kernel weights for different operating modes, with dimension 1. , is a trainable parameter, when its index is the dynamic kernel selection function. At that time, dynamic kernel selection function Used based on time point The cluster to which it belongs determines the 3D trainable weight tensor that should be used for the current computation. Which set of weights in the formula, and how is it calculated? ; Indicates a point in time The 4-dimensional input feature vector at that location, i.e. ; Indicates the first feature channel, the... The output value after noise suppression at each time point; Indicates the fourth feature channel, the The output value after noise suppression at each time point; Indicates the first The cluster center vectors are obtained using the K-means clustering algorithm; This represents the cluster index, with a value range of 100. ; Represents the trainable bias term, corresponding to the first... Each output channel is a trainable parameter.

[0078] Furthermore, embodiments of the present invention also analyze the distribution of different fault types in the feature space, referring to... Figure 8The visualization evaluation assesses the ability of features extracted by the "multi-scale temporal feature fusion and dynamic segmentation enhancement" module of this invention to distinguish different battery fault categories. The experiment included eight fault categories (normal, internal short circuit, lithium plating, overcharge, over-discharge, early thermal runaway, connection fault, and aging). Each sample had 12 dimensions of features, which simulated the enhanced features obtained after the complete process of this invention (including dynamic segmentation and cross-feature convolutional fusion). To display the spatial structure of high-dimensional features on a two-dimensional plane, the experiment employed a t-distributed random neighborhood embedding dimensionality reduction technique. From the generated scatter plot, it can be observed that samples of different fault types form relatively clustered yet separate clusters in the feature space, reflecting the high discriminative power of the features extracted by this invention. For example, the green dot cluster representing the "normal" state is concentrated in the left area of ​​the plot, while the purple dot cluster representing the "early thermal runaway" state is mainly located in the upper right, with a relatively clear interval between them. Furthermore, there is partial overlap between some categories (such as "lithium plating" and "over-discharge"), because the characterization of some battery faults may be similar. The light-colored contour lines surrounding certain point groups in the figure simulate the density distribution boundary of this type of feature, further outlining the "sphere of influence" of different failure modes. Overall, the good clustering and separation of the feature space demonstrates that the feature fusion method of this invention can effectively capture the essential differences between different failure modes, transforming high-dimensional time-series data into discriminative feature representations that are beneficial for classifier decision-making.

[0079] Step S212: Perturbation suppression is performed on the signal fluctuations across the feature channels of the cross-feature enhancement vector.

[0080] Battery faults are often subtle, requiring deep neural networks for high-order feature extraction and pattern differentiation. However, conventional deep neural networks are sensitive to perturbations in the input data, making it difficult to effectively model long-term time-series dependencies, and the propagation of perturbations within the network can easily degrade model performance. This invention uses an embedded adaptive masking mechanism to suppress the impact of high-noise periods, achieving precise targeted suppression of perturbations. Unlike the one-size-fits-all smoothing approach of traditional filtering, the adaptive masking dynamically adjusts the suppression strength according to the actual fluctuations of each channel. While filtering out invalid perturbations, it preserves the characteristic signals of early, weak faults to the greatest extent possible, avoiding over-smoothing of effective features.

[0081] In this embodiment of the invention, an adaptive mask is calculated based on the variance of the cross-feature channel of the cross-feature enhancement vector at each monitoring time point. Based on the adaptive mask, perturbation suppression is applied to signal fluctuations across the cross-feature channels. Specifically, the aforementioned feature enhancement matrix (i.e., the cross-feature enhancement vector) is first transformed in dimension through mapping using a fully connected layer. Simultaneously, an adaptive mask is calculated based on the variance of the feature enhancement matrix across all feature channels at each time point. The output of the fully connected layer is then multiplied element-wise with this mask to obtain the perturbation-suppressed output value, thereby automatically reducing the contribution of high-fluctuation or high-noise time points to the features of subsequent network layers. This is expressed as:

[0082]

[0083] In the formula, The first feature in the initial feature map after perturbation suppression is represented by the [missing information]. The first channel, the first The values ​​at each time point are used to suppress the feature contribution of high-noise time points through adaptive masking, thereby enhancing the robustness of the input layer to temporal perturbations. Indicates the feature channel index; This represents the weight matrix of the input layer, with dimension 1. , are trainable parameters. Represents the weight matrix of the input layer The Middle line, number The elements of the column, that is, starting from the first... The input feature channel to the th The weights of each output channel; This represents the number of output channels of the input layer; it is a hyperparameter with an example value of 32. This represents the bias term of the input layer, which is a trainable parameter. Represents the bias term of the input layer The Middle The bias value of each channel; Represents the linear rectification activation function; This represents the element-wise multiplication operator.

[0084] Indicates the first The adaptive perturbation mask value at each time point automatically reduces the feature weights at high-noise or high-fluctuation time points, thereby suppressing their interference with subsequent network layers and improving model robustness. The calculation method is expressed as follows: ; Represents the input matrix In the At any point in time, across all The variance of each characteristic channel is used to measure the signal fluctuation or noise level at that point in time, and is calculated as follows: ; Represents the feature enhancement matrix In the The mean across all feature channels at each time point is calculated as follows: ; Represents the natural exponential function; This represents the mask sensitivity hyperparameter, which controls the strength of the mask value's response to the noise variance. An example value is 2.0.

[0085] Step S214: Obtain temporal local details and temporal global context information of the cross-feature enhancement vector under perturbation suppression, and construct a globally perceptive feature representation.

[0086] 1) After perturbation suppression of the data, this embodiment of the invention employs a dual-path perturbation resistance module to fuse local details and global contextual temporal features in parallel, and dynamically balances the contributions of the two paths through adaptive weights. Specifically, the dual-path perturbation resistance module processes data through two parallel paths: one path uses one-dimensional convolution to capture local temporal patterns, and the other path uses a multi-head self-attention mechanism to capture global temporal dependencies to determine the corresponding global contextual information. The outputs of the two paths are weighted and fused using adaptive weights generated from learnable parameters, as shown below:

[0087]

[0088] In the formula, Indicates the first Output characteristic map of a dual-path anti-disturbance module The Middle The first channel, the first The values ​​at each time point are combined with local details and global context information. By dynamically balancing the contributions of the two paths through adaptive weights, the feature's ability to model temporal perturbations and long dependencies is enhanced. This represents the module layer index. For example, if a network can contain two dual-path disturbance rejection modules, then the corresponding... .in, Indicates the first In the module, the first The first channel, the first The fusion weights at time points are scalars between 0 and 1, used to dynamically balance the convolutional and attention paths at the 1st time point. The first channel, the first The contribution at each time point is calculated as follows:

[0089]

[0090] This represents a vector concatenation operation; Indicates the first The fusion weight vector of each module is a trainable parameter; Indicates the first The fusion bias term of each module is a trainable parameter; This represents the Sigmoid activation function. Indicates the first The output feature map of the convolutional path of the module is the first The first channel, the first The value at each time point is calculated as follows:

[0091]

[0092] Indicates difference from Feature channel index; Indicates the first The number of channels in the layer output feature map, i.e., the number of input channels to the first layer. The number of feature channels in each module; This represents the offset of the convolution operation in the time dimension, used to define the time point covered by the convolution kernel relative to the center point. The offset, when the convolution kernel width When, the range of values ​​is ; Indicates the first The 3D convolutional kernel weight tensor of each module's convolutional path is a trainable parameter. Indicates the first 3D convolution kernel weight tensor of each module convolution path The middle corresponds to the output channel Input Channel Time offset The weight value; Indicates the first Output feature map of each module's attention path The Middle The first channel, the first The value at each time point; Indicates the first Output characteristic map of a dual-path anti-disturbance module The Middle The first channel, the first The value at each time point.

[0093] In practical implementation, the dual-path disturbance rejection module uses the first... The output feature map of the layer is used as input and processed through two parallel paths. One path uses a one-dimensional convolutional layer (with a kernel width of 7, corresponding to...). One path captures local temporal patterns (from -3 to 3), while another path uses a multi-head self-attention mechanism (specifically a multi-head Transformer encoder structure) to capture global temporal dependencies. The outputs of the two paths are weighted and fused using adaptive weights generated from learnable parameters to obtain the output of this module. .

[0094] Among them, for the first The attention path of each dual-path anti-perturbation module outputs a feature map. It is calculated through forward propagation using the standard multi-head self-attention mechanism. First, the 1st... Output characteristic map of a dual-path anti-disturbance module Reconstructed into a sequence, with each time point corresponding to one The feature vectors are first processed in dimensionality. Then, a multi-head self-attention mechanism (specifically, a multi-head Transformer encoder structure) is applied. For each attention head, a trainable linear transformation is used to map the feature vector at each time point to a query vector, a key vector, and a value vector, respectively. Then, the attention weights between all time points are calculated by taking the dot product of the query and key vectors and dividing by a scaling factor (the scaling factor defaults to the square root of the feature dimension), followed by normalization using the Softmax function. This yields the attention distribution of each time point to other time points. Then, the value vectors are weighted and summed using the attention weights corresponding to the attention distribution to obtain the output vector of each head at each time point. Finally, the output vectors of all attention heads are concatenated along the feature dimension and fused and dimensionally adjusted using a trainable linear transformation to ensure that the number of channels in the output feature map is 1000. The resulting output feature map The shape is .

[0095] 2) A high-level feature representation is output through temporal pooling and a fully connected layer, which is the corresponding globally perceived feature representation. In the specific implementation, the feature map output by the dual-path anti-disturbance module is max-pooled along the time dimension to extract the maximum value of each feature channel at all time points, resulting in a compact feature vector independent of the time dimension. Then, the compact feature vector is fed into the fully connected layer for further compression and information fusion, outputting the final high-level feature representation used for fault diagnosis. In one implementation, the final dual-path anti-disturbance module is taken as the... Taking a single module as an example, the output of the last dual-path disturbance rejection module is subjected to time-series max pooling, as shown below:

[0096]

[0097] In the formula, This represents the output characteristic map of the second dual-path disturbance rejection module. The Middle The first channel, the first The value at each time point; This indicates the operation of taking the maximum value along the time dimension; Represents compact feature vectors The Middle The values ​​of each channel are used to retain the most significant feature information of each channel, forming a compact, globally-aware feature representation.

[0098] Step S216: Use the pre-built battery fault diagnosis model to classify and identify the global perception feature representation to determine the fault category of the target battery based on the operation monitoring data.

[0099] Battery faults evolve gradually and concurrently; for example, internal short circuits and lithium plating may coexist. Furthermore, early fault characteristics are weak, and conventional single-classification layers are insufficient to fully model their complex decision boundaries, resulting in inadequate sensitivity for early fault warnings. This invention achieves comprehensive diagnosis by constructing a classification framework that integrates a primary fault classification, auxiliary feature enhancement based on operating condition clustering, and progressively weighted attention. In specific implementation, fault prediction is performed through the following steps:

[0100] 1) Based on the operating condition pattern distribution represented by the global perception feature representation and the signal fluctuations corresponding to the operating condition pattern distribution, construct the auxiliary feature vector of the global perception feature representation.

[0101] In specific implementation, this embodiment of the invention determines the operating condition mode distribution of the globally perceived feature representation based on the proportion of time points in each operating condition mode cluster to which it belongs; and determines the signal fluctuation pattern of the globally perceived feature representation based on the time series skewness of the globally perceived feature representation. Further, the operating condition mode distribution, signal fluctuation pattern, and globally perceived feature representation are concatenated to form an auxiliary feature vector for the globally perceived feature representation. This embodiment of the invention enhances the model's perception of the correlation between fault dynamics and operating conditions by constructing an auxiliary feature vector and fusing it with the main feature. First, using the results of dynamic time segment clustering, the proportion of time points belonging to each operating condition mode cluster in the current data segment is statistically analyzed to form an operating condition distribution vector; second, the time series statistics of the local standard deviation of each feature channel after adaptive sliding window normalization are calculated to obtain the fluctuation pattern feature vector; further, the compact feature vector (i.e., the globally perceived feature representation), the operating condition distribution vector, and the fluctuation pattern vector are concatenated to form a fused comprehensive feature vector, represented as... ;in, This indicates that the integrated feature vector (i.e., auxiliary feature vector) integrates high-level features, operating condition distribution information, and signal fluctuation patterns, thereby enhancing the model's ability to perceive the dynamic evolution of faults and the correlation of operating conditions.

[0102] Represents the working condition distribution vector The The component, that is, the operating mode in this data segment is the first component. The percentage of time points in a cluster is calculated as follows: Furthermore, the working condition distribution vector The dimension is . Indicates the first The cluster label at the nth time point, i.e., after dynamic time segmentation clustering, the th... The cluster label assigned at each time point; It is an indicator function that takes the value 1 when the condition inside the parentheses is true, and 0 otherwise. Represents the characteristic vector of wave pattern The The component, i.e., the first component The skewness of the local standard deviation sequence of each characteristic channel is used to characterize the asymmetry of signal fluctuations, and is calculated as follows: Furthermore, the characteristic vector of the wave pattern The dimension is 4; Indicates the first Local standard deviation sequence of each feature channel The time series mean; Indicates the first Local standard deviation sequence of each feature channel The time series standard deviation.

[0103] 2) The auxiliary classification probability distribution of the auxiliary feature vector is determined by a lightweight fully connected network, and the corresponding preliminary prediction probability is represented by the output of the global perception feature representation of the main fault classification layer of the diagnostic model.

[0104] Furthermore, synthesize the feature vectors (i.e., auxiliary feature vectors) can generate auxiliary classification probability distributions through independent lightweight fully connected networks. Its network structure is similar to the main classification layer but has independent trainable parameters. In one implementation, the auxiliary classification network is a lightweight fully connected network with an input layer dimension of the fused comprehensive feature vector. The concatenation dimension is 64, the hidden layer dimension is 64, and the hidden layer activation function is used. Activation function, output layer dimension is The output layer activation function uses The activation function has all parameters independent of the main classification layer.

[0105] Furthermore, for the main fault classification layer, the compact feature vector can be mapped to the fault class space through a fully connected layer, and the Softmax function can be applied to obtain the initial main fault class probability distribution, expressed as:

[0106]

[0107] In the formula, The model initially judges it to be the first The probability of a class, along with the initial judgment of all classes, together constitute the main classification probability distribution. . This represents the category index, with a value range of 100. ; The total number of categories is indicated. In one embodiment, it includes: normal, internal short circuit, lithium plating, overcharge, over-discharge, early thermal runaway, connection failure, and aging, for a total of 8 categories. The table shows the number of channels in the output feature map of layer 2; Represents the weight matrix of the main classification layer The Middle line, number The elements of the column, that is, starting from the first... The input feature to the first The connection weights for each category are trainable parameters; Represents the weight matrix of the main classification layer The Middle line, number The elements of the column, that is, starting from the first... The input feature to the first The connection weights for each category are trainable parameters; Indicates difference from The category index, with a value range of 100. ; The bias vector of the main classification layer The Middle The bias values ​​corresponding to each category; The bias vector of the main classification layer The Middle The bias values ​​corresponding to each category; This represents the natural exponential function.

[0108] 3) Adaptively fuse the auxiliary classification probability distribution and the preliminary prediction probability to determine the fault category corresponding to the target battery.

[0109] Among them, it can be achieved by using compact feature vectors (i.e., dynamically generated adaptive weights based on globally perceived feature representation) Fusion of main classification probability distribution With auxiliary classification probability distribution The fusion method is represented as:

[0110]

[0111] in, The model indicates that the judgment is the first The final probability of each class, together with the final classification of all classes, constitutes the main classification probability distribution. . The adaptive fusion weights can be represented by compact feature vectors. Dynamically generated using the Sigmoid function, it is used to balance the contributions of the main classification result and the auxiliary classification result, and is expressed as:

[0112]

[0113] Indicates the weights used to generate adaptive fusion weights The parameter vector is a trainable parameter. Through linear transformation, the compact feature vector is mapped to a scalar, thereby dynamically balancing the contributions of the main classification and auxiliary classification results. Indicates the weights used to generate adaptive fusion weights The bias scalar is a trainable parameter.

[0114] Furthermore, this invention also designs a loss function for the model. Existing model training typically relies solely on data-driven supervised loss, lacking a mechanism to incorporate the inherent physical laws of battery failure as prior knowledge into the learning process. This can lead to the model learning features that violate actual evolutionary patterns, affecting its generalization ability and decision reliability in complex scenarios. Conventional cross-entropy loss only focuses on the final classification accuracy and cannot utilize the consistency laws and early risk information of the failure evolution process. This invention achieves multi-task joint optimization by constructing a composite loss function, including multi-task supervised loss and failure evolution consistency loss. The multi-task supervised loss simultaneously optimizes the three sub-tasks: main failure classification, auxiliary failure classification, and early risk scoring. The failure evolution consistency loss uses the failure evolution law as a regularization term, prompting the model to learn feature representations that conform to physical consistency.

[0115] In its implementation, the total loss function includes multi-task supervision loss and fault evolution consistency loss. The multi-task supervision loss is used to jointly supervise the classification probability distribution, the preliminary prediction probability, and the risk score of the temporal evolution of the globally perceived feature representation; the fault evolution consistency loss is used to measure the linear correlation between the risk score and the proportion of abnormal operating conditions in the globally perceived feature representation.

[0116] For the multi-task supervised loss: The multi-task supervised loss jointly optimizes the three tasks of primary fault classification, auxiliary fault classification, and early risk scoring. For the fault classification task, a labeled smoothed cross-entropy loss is used to reduce model overfitting. For the early risk scoring task, the objective value is set according to the actual fault stage of the sample, expressed as:

[0117]

[0118] In the formula, This represents the multi-task supervision loss, which improves the overall diagnostic performance of the model by using weighted and joint supervision of classification and risk scoring tasks. This represents the weighting coefficient of the main classification loss. It is a hyperparameter used to control the importance of the main classification task in the total loss. An example value is 1.0. The weighting coefficient for the auxiliary classification loss is a hyperparameter used to control the importance of the auxiliary classification task; an example value is 0.5. The weighting factor representing the risk scoring loss is a hyperparameter used to control the importance of the risk scoring task; an example value is 0.2. This represents the cross-entropy loss function, which measures the difference between the probability distribution of the model output and the true label distribution. It is the standard loss function in classification tasks. This represents the smoothed true label vector, used to mitigate model overfitting, for true labels in one-hot encoded format. , ; This is a smoothing factor to prevent the model from becoming overconfident in the training labels. Overfitting is mitigated through label smoothing techniques; an example value of 0.1 is used. This represents the mean squared error loss function, used to measure the squared difference between the risk score output by the model and the target value. The target value for the early risk score is set based on the actual failure stage of the sample, for example, for normal samples. Early failure samples Serious Fault Sample The specific values ​​can be adjusted according to the data annotation.

[0119] Regarding the consistency loss in fault evolution: the early risks assessed by the model should be positively correlated with the time proportion of abnormal operating conditions in the data segment. To ensure that the features and risk scores learned by the model conform to the physical evolution law of battery faults, a consistency regularization term based on early risk scores and dynamic segmentation results is adopted, expressed as:

[0120]

[0121] In the formula, This represents the consistency loss in fault evolution. The Pearson correlation coefficient is used to measure the linear correlation between the early risk score and the proportion of abnormal operating conditions, so that the features learned by the model conform to the physical laws of fault gradual evolution. This represents the weighting coefficient of this loss term. It is a hyperparameter used to control the importance of consistency loss in the total loss. An example value is 0.1. This represents the Pearson correlation coefficient, which measures the degree of linear correlation between two variable sequences.

[0122] in, Indicating training batch Early risk score sequence of each sample; Indicating training batch The sequence of abnormal operating conditions for each sample; Indicates the size of the training batch; This represents the sample index in the training batch, with a value range of 1. arrive ; Indicates the first The early risk score of each sample is calculated by the model based on the input data of that sample; Indicates the first Operating condition distribution vector of each sample Corresponding abnormal cluster The component, that is, the proportion of time points in the sample that belong to the abnormal operating condition mode; The index representing the abnormal operating condition pattern cluster is dynamically determined at the beginning of each training cycle. The selection criterion is the cluster with the greatest Euclidean distance to the mean cluster centers of all normal samples. It should be noted that the abnormal cluster... The clusters are dynamically determined at the beginning of each training cycle, using the following selection criteria: calculate the mean vector of all normal sample cluster centers, and then select the cluster with the greatest Euclidean distance from this mean vector as the outlier cluster. Furthermore, the total loss during model training... This is the sum of multi-task supervision loss and fault evolution consistency loss, i.e. .

[0123] Furthermore, to enhance the model's sensitivity to the fault evolution process, this embodiment of the invention also designs progressive attention and early risk scoring. Specifically, this embodiment employs a progressive attention mechanism based on the magnitude of feature changes, assigning higher attention weights to periods of drastic feature changes. Simultaneously, the masked weighted feature sequence is used to encode temporal evolution information through a recurrent neural network layer, and a scalar is output as the early risk score (i.e., the early risk score s in the aforementioned loss function), expressed as:

[0124]

[0125] In the formula, The early risk score indicates the degree of risk of early failure in the current data segment. It is a scalar value, and the closer the value is to 1, the higher the risk. This represents the weight vector that maps the hidden state of the last time step of the recurrent neural network to the risk score; these are trainable parameters. This represents a bias scalar that maps the hidden state of the last time step of the recurrent neural network to a risk score; it is a trainable parameter. This represents the hidden state vector of the gated recurrent unit network at the last time step, derived from the input sequence. The hidden state is obtained through computation using a gated recurrent unit network (GRU). The GRU iteratively updates the hidden state step-by-step, and the hidden state at the final time step is... . This represents the first feature map after progressive attention mask weighting. The feature vectors of all channels at each time point are calculated as follows: . This represents the first feature map after progressive attention mask weighting. Feature vectors of all channels at each time point This represents the first feature map after progressive attention mask weighting. Feature vectors of all channels at each time point; Indicates the first The progressive attention weight at each time point is calculated as follows: ; Represents the L2 norm; This represents the attention intensity coefficient, a hyperparameter that controls the sensitivity to the magnitude of feature changes; an example value is 1.5. The first feature in the initial feature map after perturbation suppression is represented by the [missing information]. Feature vectors of all channels at each time point; The first feature in the initial feature map after perturbation suppression is represented by the [missing information]. Feature vectors of all channels at each time point.

[0126] Furthermore, the battery fault diagnosis model training and parameter update process is based on the constructed training dataset. Iterative optimization enables the model to learn the mapping relationship between fault features and categories. In one implementation, this invention constructs a training dataset by collecting and precisely labeling time-series data covering various operating conditions and fault states. The data labeling method is determined by expert diagnosis or experimental verification based on the actual battery state. Each sample is assigned a fault category label, and the labeled category strictly corresponds to the corresponding model training task. For example, it includes eight categories: normal, internal short circuit, lithium plating, overcharge, over-discharge, early thermal runaway, connection fault, and aging. The data acquisition process involves real-time monitoring of four key parameters—current, voltage, temperature, and state of charge—using sensors installed on the battery system. Time-series signals are continuously collected at a fixed sampling frequency (e.g., 1 Hz), with the sampling duration for each sample set to 120 seconds, thus forming a 4×T original data matrix, where T is the total number of time points, for example, T=120. Data acquisition can be conducted in the laboratory or in real-world application scenarios, covering the entire process of battery evolution from normal state to various faults, including different charging and discharging rates, ambient temperature changes, and load fluctuations, to ensure the diversity and representativeness of the data.

[0127] Furthermore, to support the early risk scoring task, some samples are also labeled with early risk levels based on the fault development stage. For example, normal samples are labeled as low risk, early fault samples as medium risk, and severe fault samples as high risk. After labeling, the dataset is divided into training, validation, and test sets for model training, hyperparameter tuning, and performance evaluation, thus providing high-quality, supervised input for subsequent data preprocessing and model building. At the start of training, model parameters are randomly initialized. Subsequently, data is input to the training set in batches, with each batch containing B samples. Each sample undergoes adaptive data normalization, local noise suppression, multi-scale temporal feature fusion, and dynamic segmentation enhancement before being input into the perturbation-resistant deep neural network module and the fault diagnosis classification module. Forward propagation calculates the main fault classification probability distribution, auxiliary classification probability distribution, and early risk score.

[0128] The loss function described above is calculated as the sum of multi-task supervision loss and fault evolution consistency loss. The multi-task supervision loss jointly optimizes the main classification, auxiliary classification, and risk scoring tasks, while the fault evolution consistency loss constrains the positive correlation between early risk scores and the proportion of abnormal operating conditions through Pearson correlation coefficients, enhancing the model's feature learning to conform to physical laws. Parameter updates employ backpropagation to calculate gradients, combined with an adaptive optimizer (such as Adam) to adjust trainable parameters. The learning rate is set to 0.001 by default. The stopping condition for iteration can be based on validation set performance monitoring. Training terminates when the validation loss no longer significantly decreases over multiple consecutive training epochs or reaches the preset maximum number of iterations. Simultaneously, the model parameters with the best validation performance are saved to avoid overfitting and ensure model generalization ability. Optionally, regularization techniques such as dropout can be used during training to mitigate overfitting, ultimately resulting in a trained battery fault diagnosis model for subsequent applications.

[0129] Furthermore, after the battery fault diagnosis model is trained, it can be applied to online or offline fault diagnosis of actual battery systems, realizing automated, high-precision condition monitoring and early warning. Corresponding to the above embodiment, the newly acquired battery time-series data can first undergo preprocessing consistent with the training phase, including acquiring current, voltage, temperature, and state of charge time-series signals through sensors, and constructing an original data matrix with the same sampling frequency and duration. Then, the data undergoes adaptive normalization and local noise suppression steps. Adaptive sliding window normalization eliminates feature scale differences and suppresses local disturbances, and wavelet threshold noise suppression filters out high-frequency noise, resulting in a normalized data matrix. Next, multi-scale time-series feature fusion and dynamic segmentation enhancement are performed. Dynamic time-segment clustering identifies operating conditions, and cross-feature convolution fusion is used to generate an enhanced feature matrix.

[0130] The processed data is input into a pre-trained battery fault diagnosis model. The model's forward propagation sequentially extracts high-level features through an anti-perturbation deep neural network module, and then outputs the main fault category probability distribution, auxiliary classification probability distribution, and early risk score through a fault diagnosis classification module. The final diagnosis result is a comprehensive fault category prediction obtained by fusing the main and auxiliary classification probabilities; for example, the category with the highest probability is taken as the main fault type. Simultaneously, the early risk score is used to assess the risk of fault evolution; when the score exceeds a preset threshold, an early warning is triggered. The application system can be integrated into a battery management system platform to achieve real-time monitoring, fault reporting, and maintenance recommendations, improving the safety and reliability of the battery system.

[0131] In summary, the embodiments of the present invention are innovative compared to the prior art in the following aspects:

[0132] 1. A two-stage data preprocessing method integrating local statistical characteristics and frequency domain filtering is proposed. By combining adaptive sliding window normalization with wavelet threshold denoising, the method effectively suppresses temporal local disturbances and high-frequency noise while eliminating feature scale differences. This solves the problems of traditional global normalization methods being sensitive to temporal fluctuations and having serious noise residue.

[0133] 2. A multi-scale feature fusion mechanism based on dynamic working condition perception is proposed. The time series is dynamically segmented by unsupervised clustering, and cross-feature fusion is performed by dynamic convolution kernel driven by clustering results. This enables adaptive extraction of multi-feature interaction relationships under different working conditions, overcoming the limitations of traditional methods in capturing cross-feature collaborative changes and dynamic evolution characteristics of faults.

[0134] 3. A dual-path anti-disturbance deep network and multi-task classification framework are proposed. By suppressing noise periods through adaptive masking, fusing local and global features through dual paths, and combining adaptive fusion of primary and secondary classification with progressive attention mechanism, the robustness of the model to disturbances is improved, while enhancing the sensitivity to weak early fault features and fault evolution process.

[0135] 4. A composite loss function guided by physical laws is proposed. Based on the traditional supervised loss, a fault evolution consistency loss is added. The early risk score of the model learning is forced to be positively correlated with the proportion of abnormal working conditions. The combination of data-driven approach and fault physical evolution law improves the model's generalization ability and interpretability.

[0136] Furthermore, embodiments of the present invention also provide a battery fault diagnosis device based on artificial intelligence, referring to... Figure 9The device includes: a data processing module 10, used to acquire the operation monitoring data of the target battery, and determine the monitoring condition mode of the attribute data at the corresponding monitoring time point based on the feature space distribution of each attribute data at the monitoring time point; a feature fusion module 20, used to perform cross-feature convolution fusion on each attribute data based on the monitoring condition mode to generate a cross-feature enhancement vector; a feature extraction module 30, used to suppress the signal fluctuations of the cross-feature channel of the cross-feature enhancement vector, and acquire the temporal local details and temporal global context information of the cross-feature enhancement vector under the perturbation suppression to construct a globally perceived feature representation; and an execution module 40, used to classify and identify the globally perceived feature representation using a pre-built battery fault diagnosis model to determine the fault category of the target battery based on the operation monitoring data.

[0137] The battery fault diagnosis device based on artificial intelligence provided in this embodiment of the invention has the same implementation principle and technical effect as the aforementioned method embodiment. For the sake of brevity, any parts not mentioned in the device embodiment can be referred to the corresponding content in the aforementioned method embodiment.

[0138] Furthermore, embodiments of the present invention also provide an artificial intelligence-based battery fault diagnosis system, which is configured with the apparatus of the above embodiments for executing the methods of any of the above embodiments.

[0139] The battery fault diagnosis system based on artificial intelligence provided in this embodiment of the invention has the same implementation principle and technical effects as the aforementioned method embodiment. For the sake of brevity, any parts not mentioned in the system embodiment can be referred to the corresponding content in the aforementioned method embodiment.

[0140] This invention also provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements the above-described... Figures 1-2 The steps of any of the methods shown. Embodiments of the present invention also provide a computer-readable storage medium storing a computer program, which, when executed by a processor, performs the above-described steps. Figures 1-2 The steps of any of the methods shown. Embodiments of the present invention also provide a structural schematic diagram of an electronic device, such as... Figure 10 The diagram shows the structure of the electronic device, which includes a processor 101 and a memory 100. The memory 100 stores computer-executable instructions that can be executed by the processor 101. The processor 101 executes the computer-executable instructions to implement the above-mentioned... Figures 1-2 Any of the methods shown.

[0141] exist Figure 10In the illustrated embodiment, the electronic device further includes a bus 102 and a communication interface 103, wherein the processor 101, the communication interface 103, and the memory 100 are connected via the bus 102. The memory 100 may include high-speed random access memory (RAM) or non-volatile memory, such as at least one disk drive. Communication between this system network element and at least one other network element is achieved through at least one communication interface 103 (which can be wired or wireless), using the Internet, wide area network, local area network, metropolitan area network, etc. Bus 102 can be an ISA (Industry Standard Architecture) bus, a PCI (Peripheral Component Interconnect) bus, or an EISA (Extended Industry Standard Architecture) bus, or an AMBA (Advanced Microcontroller Bus Architecture) bus. AMBA defines three types of buses: APB (Advanced Peripheral Bus), AHB (Advanced High-performance Bus), and AXI (Advanced eXtensible Interface). Bus 102 can be divided into address bus, data bus, control bus, etc. For ease of representation, Figure 10The diagram uses only a single double-headed arrow, but this does not imply a single bus or a single type of bus. Processor 101 may be an integrated circuit chip with signal processing capabilities. In implementation, each step of the above method can be completed by the integrated logic circuitry in the hardware of processor 101 or by instructions in software form. Processor 101 can be a general-purpose processor, including a Central Processing Unit (CPU), a Network Processor (NP), etc.; it can also be a Digital Signal Processor (DSP), an Application Specific Integrated Circuit (ASIC), a Field-Programmable Gate Array (FPGA), or other programmable logic devices, discrete gate or transistor logic devices, or discrete hardware components. The general-purpose processor can be a microprocessor or any conventional processor. The steps of the method disclosed in the embodiments of this application can be directly manifested as execution by a hardware decoding processor, or execution by a combination of hardware and software modules in the decoding processor. The software module can reside in a mature storage medium in the field, such as random access memory, flash memory, read-only memory, programmable read-only memory, electrically erasable programmable memory, or registers. This storage medium is located in memory, and the processor 101 reads information from the memory and, in conjunction with its hardware, completes the aforementioned tasks. Figures 1-2 Any of the methods shown.

[0142] The computer program product of the battery fault diagnosis method, device, and system based on artificial intelligence provided in this invention includes a computer-readable storage medium storing program code. The instructions included in the program code can be used to execute the methods described in the preceding method embodiments. Specific implementations can be found in the method embodiments and will not be repeated here. Those skilled in the art will understand that, for convenience and brevity, the specific working process of the system described above can be referred to the corresponding process in the preceding method embodiments, and will not be repeated here. Furthermore, in the description of this invention, unless otherwise explicitly specified and limited, the terms "installation," "connection," and "linking" should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral connection; they can refer to a mechanical connection or an electrical connection; they can refer to a direct connection or an indirect connection through an intermediate medium; they can refer to the internal connection of two components. Those skilled in the art can understand the specific meaning of the above terms in this invention based on the specific circumstances. If the function is implemented in the form of a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0143] In the description of this invention, it should be noted that the terms "center," "upper," "lower," "left," "right," "vertical," "horizontal," "inner," and "outer," etc., indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings. They are used only for the convenience of describing the invention and for simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, they should not be construed as limitations on the invention. Furthermore, the terms "first," "second," and "third" are used for descriptive purposes only and should not be construed as indicating or implying relative importance. Finally, it should be noted that the above embodiments are merely specific implementations of the invention, used to illustrate the technical solutions of the invention, and not to limit it. The scope of protection of the invention is not limited thereto. Although the invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that any person skilled in the art can still modify or easily conceive of changes to the technical solutions described in the foregoing embodiments within the technical scope disclosed in this invention, or make equivalent substitutions for some of the technical features; and these modifications, changes, or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of this invention, and should all be covered within the scope of protection of this invention. Therefore, the scope of protection of this invention should be determined by the scope of the claims.

Claims

1. A battery fault diagnosis method based on artificial intelligence, characterized in that, The method includes: Acquire the operational monitoring data of the target battery, and determine the monitoring operating mode of the attribute data at the corresponding monitoring time point based on the feature spatial distribution of each attribute data of the operational monitoring data at the monitoring time point. Based on the monitoring operating mode, cross-feature convolutional fusion is performed on each of the attribute data to generate a cross-feature enhancement vector; The signal fluctuations of the cross-feature channel of the cross-feature enhancement vector are perturbed and suppressed, and the temporal local details and temporal global context information of the cross-feature enhancement vector under the perturbation suppression are obtained to construct a globally perceived feature representation; The global perception feature representation is classified and identified using a pre-built battery fault diagnosis model to determine the fault category of the target battery based on the operation monitoring data; The step of determining the monitoring condition mode of the attribute data at the corresponding monitoring time point based on the feature spatial distribution of each attribute data of the operational monitoring data at the monitoring time point includes: The operation monitoring data is filtered in the frequency domain, and the filtered operation monitoring data is adaptively normalized according to the time-series statistics of the operation monitoring data to generate preprocessed data. Based on the distribution of each attribute data in the feature space of the preprocessed data, the monitoring time points of the attribute data are divided into multiple working condition mode clusters; The monitoring operating condition mode to which the current monitoring time point belongs is determined based on the aforementioned operating condition mode cluster; The step of classifying and identifying the global sensing feature representation using a pre-built battery fault diagnosis model to determine the fault category of the target battery based on the operational monitoring data includes: Based on the operating condition pattern distribution and signal fluctuation pattern represented by the global sensing features, an auxiliary feature vector of the global sensing features is constructed. The auxiliary classification probability distribution of the auxiliary feature vector is determined by a lightweight fully connected network, and the preliminary prediction probability corresponding to the global perception feature representation is output by the main fault classification layer of the diagnostic model. The auxiliary classification probability distribution and the preliminary prediction probability are adaptively fused to determine the fault category corresponding to the target battery; The step of constructing an auxiliary feature vector for the global sensing feature representation based on the operating condition pattern distribution and signal fluctuation pattern of the global sensing feature representation includes: The distribution of the operating conditions represented by the global perception features is determined based on the proportion of the global perception features in each operating condition mode cluster at each time point. The signal fluctuation pattern represented by the global sensing features is determined based on the skewness of the time series represented by the global sensing features. The operating condition distribution, the signal fluctuation pattern, and the global perception feature representation are concatenated to form an auxiliary feature vector of the global perception feature representation.

2. The method according to claim 1, characterized in that, The total loss function of the battery fault diagnosis model includes multi-task supervision loss and fault evolution consistency loss. The multi-task supervision loss is used to jointly supervise the auxiliary classification probability distribution, the preliminary prediction probability, and the risk score of the temporal evolution of the globally perceived feature representation; The fault evolution consistency loss is used to measure the linear correlation between the risk score and the proportion of abnormal operating conditions represented by the global awareness features.

3. The method according to claim 1, characterized in that, The step of performing frequency domain filtering on the operation monitoring data includes: Discrete wavelet transform is used to determine the wavelet coefficients of the operational monitoring data at multiple scales; the wavelet coefficients are used to characterize the energy components of the operational monitoring data at a preset frequency scale at a preset monitoring time point. Based on the neighborhood range data of the wavelet coefficients, determine the adaptive threshold for noise suppression; The noise suppression adaptive threshold is used to suppress noise in the operation monitoring data to generate filtered data; The steps of adaptively normalizing the filtered operation monitoring data based on the time-series statistics of the operation monitoring data to generate preprocessed data include: Using the monitoring time point of the operation monitoring data as a time window, adaptive sliding window normalization is performed on the data of each feature channel of the operation monitoring data to generate preprocessed data.

4. The method according to claim 1, characterized in that, The step of perturbation suppression of signal fluctuations across feature channels of the cross-feature enhancement vector includes: Calculate the adaptive mask based on the variance of the cross-feature channel at each monitoring time point of the cross-feature enhancement vector; Based on the adaptive mask, disturbance suppression is performed on the signal fluctuations across the feature channels.

5. The method according to claim 1, characterized in that, The steps of acquiring the temporal local details and temporal global context information of the cross-feature enhancement vector under the perturbation suppression, and constructing a globally perceptive feature representation, include: The temporal local details and temporal global context information of the cross-feature enhancement vector are extracted by using two paths of the preset dual-path anti-perturbation module. One path is used to capture the local temporal pattern of the cross-feature enhancement vector through one-dimensional convolution, and the other path is used to capture the global temporal dependency of the cross-feature enhancement vector using a multi-head self-attention mechanism to determine the corresponding global context information. Adaptive weighted fusion of the temporal local details and the temporal global context information is performed to generate temporal recognition features; Temporal max pooling is performed on the temporal recognition features to generate a globally perceptual feature representation.

6. A battery fault diagnosis device based on artificial intelligence, characterized in that, The device includes: The data processing module is used to acquire the operation monitoring data of the target battery, and determine the monitoring condition mode of the attribute data at the corresponding monitoring time point based on the feature spatial distribution of each attribute data of the operation monitoring data at the monitoring time point. The feature fusion module is used to perform cross-feature convolutional fusion on each of the attribute data based on the monitoring working condition mode to generate a cross-feature enhancement vector; The feature extraction module is used to suppress the signal fluctuations across feature channels of the cross-feature enhancement vector, and to obtain the temporal local details and temporal global context information of the cross-feature enhancement vector under the perturbation suppression, and to construct a globally perceived feature representation. The execution module is used to classify and identify the global perception feature representation using a pre-built battery fault diagnosis model, and determine the fault category of the target battery based on the operation monitoring data. The data processing module is further configured to: perform frequency domain filtering on the operation monitoring data, and adaptively normalize the filtered operation monitoring data according to the time-series statistics of the operation monitoring data to generate preprocessed data; divide the monitoring time points of the attribute data into multiple operating condition mode clusters according to the distribution of each attribute data in the feature space of the preprocessed data; and determine the monitoring operating condition mode to which the current monitoring time point belongs based on the operating condition mode clusters. The execution module is further configured to: construct an auxiliary feature vector of the global sensing feature representation based on the operating condition mode distribution and signal fluctuation pattern of the global sensing feature representation; determine the auxiliary classification probability distribution of the auxiliary feature vector through a lightweight fully connected network; output the preliminary prediction probability corresponding to the global sensing feature representation using the main fault classification layer of the diagnostic model; and adaptively fuse the auxiliary classification probability distribution and the preliminary prediction probability to determine the fault category corresponding to the target battery. The execution module is further configured to: determine the operating mode distribution of the global perception feature representation based on the proportion of the global perception feature representation at each time point in its respective operating mode cluster; determine the signal fluctuation pattern of the global perception feature representation based on the time series skewness of the global perception feature representation; and perform data splicing on the operating mode distribution, the signal fluctuation pattern, and the global perception feature representation to form an auxiliary feature vector of the global perception feature representation.

7. A battery fault diagnosis system based on artificial intelligence, characterized in that, The system is equipped with the apparatus of claim 6.

Images