A method and system for intelligent diagnosis and early warning of energy storage container faults

Abstract

This invention discloses an intelligent fault diagnosis and early warning method and system for energy storage containers, relating to the field of energy storage container fault diagnosis technology. The method includes: acquiring distributed temperature sensing data and dynamic internal resistance parameters to generate a multi-source monitoring data set; constructing a temperature field snapshot sequence and an internal resistance impedance spectrum evolution sequence; inputting a thermo-electric relaxation coupling mapping network to extract local thermal diffusion hysteresis correlation feature maps; performing time-frequency transformation to generate a fault relaxation feature spectrum; matching with a fault mode library, and outputting the fault location and type when dynamic early warning conditions are met; generating cooling control commands to drive the cooling system. This invention, by constructing a three-dimensional temperature field and impedance spectrum evolution sequence, models the relaxation hysteresis characteristics of thermal diffusion relative to electrochemical changes, separating the weak oscillations excited by irreversible thermal processes from normal fluctuations, improving the sensitivity of early fault identification, and forming a closed loop through cooling linkage and predictive early warning to ensure the safe operation of energy storage containers.

Description

Technical Field

[0001] This invention relates to the field of energy storage container fault diagnosis technology, specifically to an intelligent fault diagnosis and early warning method and system for energy storage containers. Background Technology

[0002] Energy storage containers are the core equipment of electrochemical energy storage power stations. They contain a large number of densely packed lithium-ion battery cells and are equipped with a battery management system, temperature sensor array, and liquid-cooled or air-cooled thermal management system. During long-term charge-discharge cycles, the battery cells gradually degrade due to side reactions such as thickening of the solid electrolyte interface film, lithium plating on the negative electrode, and electrolyte decomposition. This degradation process is accompanied by increased internal resistance, capacity decay, and abnormal heat generation. When the rate of local heat accumulation exceeds the heat dissipation capacity, thermal runaway may be triggered. Therefore, implementing online fault diagnosis and early warning for energy storage containers is a key technical requirement for ensuring the safe operation of energy storage power stations.

[0003] Existing technologies, such as those published in CN114441978A, CN121254107A, CN118782940B, and CN116299009A related to fault diagnosis of energy storage containers, typically rely on independent temperature monitoring or independent electrical parameter analysis. Temperature monitoring schemes collect multiple temperature points within the container using thermocouples or fiber optic sensors, and determine anomalies based on threshold comparisons or temperature rise rates. Electrical parameter analysis schemes extract voltage, current, and internal resistance data from the battery management system, and identify and assess battery health using internal resistance growth models or equivalent circuit parameters. Some schemes attempt to simply concatenate temperature and electrical parameters and input them into a neural network classifier to output the fault type; however, in these schemes, temperature and electrical parameters are mostly extracted as independent features, without establishing a relaxation coupling relationship between the two in the time dimension.

[0004] Existing technical solutions have the following drawbacks. The temperature threshold setting is significantly affected by ambient temperature and charge / discharge conditions. Fixed thresholds are prone to false alarms under high-temperature, high-rate conditions, while they may miss alarms under low-temperature, light-load conditions. Most internal resistance analysis schemes rely on offline internal resistance measurements or single-frequency impedance detection by the battery management system under specific charge states, lacking continuous tracking of the dynamic evolution of the broadband impedance spectrum, making it difficult to capture early fault characteristics caused by changes in charge transfer processes at the battery interface. While simply concatenating temperature and electrical parameters and feeding them into a neural network achieves information fusion, it ignores the inherent hysteresis of thermal diffusion relative to electrochemical changes, failing to distinguish between normal operating response and fault-induced relaxation oscillations from a time-series correlation perspective. This results in insufficient timeliness and accuracy in diagnosing progressive faults such as internal micro-short circuits and loose busbar connections. Summary of the Invention To address the aforementioned technical shortcomings, the present invention aims to provide an intelligent diagnosis and early warning method and system for energy storage container faults.

[0005] To solve the above technical problems, the present invention adopts the following technical solution: The first aspect of the present invention provides an intelligent diagnosis and early warning method for energy storage container faults, including: S1: acquiring distributed temperature sensing data and dynamic internal resistance parameters of battery cells in the energy storage container, and generating a multi-source monitoring data set; S2: Based on multi-source monitoring data sets, a temperature field snapshot sequence characterizing the continuous spatiotemporal changes of the three-dimensional temperature distribution inside the container is constructed, as well as an internal resistance impedance spectrum evolution sequence characterizing the dynamic changes of the internal resistance frequency response of the battery cell is constructed. S3: Simultaneously input the temperature field snapshot sequence and the internal resistance impedance spectrum evolution sequence into the thermo-electric relaxation coupling mapping network to extract the time lag correlation pattern of the temperature change rate feature of the local temperature field relative to the change of the internal resistance impedance spectrum, and generate a local thermal diffusion lag correlation feature map. S4: Perform time-frequency transformation on the local thermal diffusion hysteresis correlation feature map, extract the relaxation feature oscillation component excited by the irreversible thermal process, and generate the fault relaxation feature spectrum. S5: Calculate the coherence matching degree between the fault relaxation feature spectrum and the standard relaxation spectrum of each fault prototype in the fault mode library. When the matching degree meets the dynamic early warning condition generated based on the spectrum fluctuation statistics under historical normal operation, output a fault early warning signal containing the fault location and fault type. S6: Generates loop enhancement cooling target and cooling parameter adjustment commands, and drives the cooling control system; S7: Train the parameter evolution prediction network using temperature field snapshot sequences and internal resistance spectrum evolution sequences, and generate the predicted fault relaxation feature spectrum to achieve predictive early warning.

[0006] A second aspect of the present invention provides a system for performing the intelligent diagnosis and early warning method for faults in an energy storage container, comprising: a data preprocessing module for acquiring distributed temperature sensing data and dynamic internal resistance parameters of battery cells within the energy storage container, and generating a multi-source monitoring data set; The sequence construction module is used to construct a temperature field snapshot sequence that characterizes the continuous spatiotemporal changes of the three-dimensional temperature distribution inside the container based on multi-source monitoring data sets, and to construct an internal resistance impedance spectrum evolution sequence that characterizes the dynamic changes of the frequency response of the internal resistance of the battery cell. The relaxation mapping module is used to synchronously input the temperature field snapshot sequence and the internal resistance impedance spectrum evolution sequence into the thermo-electric relaxation coupling mapping network, extract the time lag correlation pattern of the temperature change rate feature of the local temperature field relative to the change of the internal resistance impedance spectrum, and generate a local thermal diffusion lag correlation feature map. The spectrum generation module is used to perform time-frequency transformation on the local thermal diffusion hysteresis correlation feature map, extract the relaxation feature oscillation component excited by the irreversible thermal process, and generate the fault relaxation feature spectrum. The matching and early warning module is used to calculate the coherence matching degree between the fault relaxation feature spectrum and the standard relaxation spectrum of each fault prototype in the fault mode library. When the matching degree meets the dynamic early warning condition generated based on the spectrum fluctuation statistics under historical normal operation, it outputs a fault early warning signal containing the fault location and fault type. The cooling control module is used to generate loop-enhanced cooling targets and cooling parameter adjustment commands, and to drive the cooling control system. The prediction module is used to train the parameter evolution prediction network using temperature field snapshot sequences and internal resistance impedance spectrum evolution sequences, and generate the predicted fault relaxation feature spectrum to achieve predictive early warning.

[0007] The beneficial effects of the present invention are as follows: (1) The present invention constructs a three-dimensional temperature field snapshot sequence inside the energy storage container and an internal impedance spectrum evolution sequence of the battery cell simultaneously, and characterizes the system operating status from two dimensions: spatial thermal distribution and electrochemical frequency response. This breaks through the limitation that monitoring a single physical quantity cannot fully reflect the thermal-electric coupling fault mechanism, and enables the fault diagnosis basis to cover the complete evolution chain from electrochemical decay to thermal runaway, laying a multi-source data foundation for achieving early and reliable early warning.

[0008] (2) This invention introduces a thermal-electric relaxation coupling mapping network to extract the time lag correlation mode of the temperature change rate in the local area of ​​the temperature field relative to the change of the internal resistance spectrum. It explicitly models the inherent relaxation lag characteristics of the thermal process to the electrochemical change, distinguishes the weak relaxation characteristic oscillations generated by irreversible side reactions or structural damage from the fluctuations of normal operation, and improves the identification sensitivity and positioning accuracy of early latent faults such as loose connections and internal micro short circuits.

[0009] (3) This invention utilizes the coherence matching of fault relaxation feature spectrum and fault mode library and combines it with the dynamic early warning threshold generated by historical normal operation fluctuation statistics to avoid false alarms or missed alarms caused by fixed thresholds when operating conditions change. At the same time, it directly outputs the fault type and fault battery cell location through fault prototype matching, providing a precise loop enhancement cooling target for the cooling control system, realizing closed-loop linkage of diagnosis, location and emergency heat dissipation, and effectively suppressing the spread of fault thermal state.

[0010] (4) This invention extrapolates the future evolution of temperature field and internal impedance spectrum through parameter evolution prediction network, sends the extrapolation results into the diagnostic link in advance to generate the predicted fault relaxation characteristic spectrum and perform predictive early warning, and moves the early warning time point from the stage where the fault can be diagnosed in real time to the stage of fault precursor, thus gaining sufficient time window for planned shutdown and proactive maintenance. Attached Figure Description

[0011] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0012] Figure 1 This is a schematic diagram of the implementation steps of the method of the present invention.

[0013] Figure 2 This is a schematic diagram of the system structure connection of the present invention. Detailed Implementation

[0014] 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. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0015] Reference Figure 1 As shown, the first aspect of the present invention provides an intelligent diagnosis and early warning method for faults in an energy storage container, comprising: S1: acquiring distributed temperature sensing data and dynamic internal resistance parameters of battery cells within the energy storage container, and generating a multi-source monitoring data set.

[0016] It should be noted that the distributed temperature sensing data is acquired in real time by an array of fiber Bragg grating temperature sensors deployed in the gaps between battery modules, busbars, and key locations on the casing within the energy storage container. Each sensor channel records the temperature time series at a fixed sampling frequency, carrying a timestamp and sensor spatial location identifier. The dynamic internal resistance parameters of the battery cells are obtained by periodically injecting small AC excitation signals into the battery cells through the battery management system and calculating them in conjunction with the response voltage and current; these parameters also carry independent timestamps.

[0017] In a specific embodiment of the present invention, generating a multi-source monitoring data set includes: performing wavelet threshold denoising on the distributed temperature sensing data to obtain denoised temperature data.

[0018] Specifically, the wavelet thresholding denoising process involves selecting the Daubechies 4 wavelet function, which possesses good time-frequency compact support characteristics, as the basis function. Multi-level discrete wavelet decomposition is performed on the original temperature time series of each temperature sensing channel. The number of decomposition levels is set to 5 based on the sampling frequency and the dominant frequency band of the battery thermal process, yielding wavelet coefficients at each scale. A general threshold rule is applied to determine the threshold value for the high-frequency detail coefficients obtained from the decomposition. This threshold value is determined by the formula: Calculation, where The noise standard deviation in the high-frequency detail coefficients is estimated by robustly dividing the median of the absolute values ​​of the wavelet coefficients at each level by 0.6745. This is the length of the detail coefficients for this layer. For absolute values ​​less than... The wavelet coefficients are set to zero, and for those with absolute values ​​higher than zero... The coefficients are shrunk using a soft thresholding function, the expression of which is:

[0019] In the formula These are the original wavelet coefficients. The coefficients after thresholding. The sign function is used. The temperature time series is reconstructed using the thresholded wavelet coefficients through inverse discrete wavelet transform, yielding denoised temperature data for each channel. The denoised temperature data retains the gradual trend related to the battery thermal process and the local temperature rise peaks during the thermal runaway incubation period, while suppressing high-frequency spikes introduced by electromagnetic interference.

[0020] The dynamic internal resistance parameters of the battery cells are timestamped and missing frames are interpolated according to a unified time base to generate a regular internal resistance parameter sequence.

[0021] Specifically, the dynamic internal resistance parameters of the battery cells are recorded at unequal time intervals, with the acquisition time determined by the measurement scheduling of the battery management system. Meanwhile, the sampling of the distributed temperature sensors has an independent and fixed-period time grid. To eliminate time misalignment between multi-source data, a unified time reference needs to be established. The unified time reference is selected by starting from the earliest common timestamp among all data, with the sampling interval of the distributed temperature sensors as the step size. Equal-interval time series ,in The original time series of the dynamic internal resistance parameters of the battery cell is denoted as... , For the first The timestamp of the internal resistance measurement. This is the internal resistance value obtained from this measurement. Timestamp alignment and missing frame interpolation are performed on this sequence to generate a regularized internal resistance parameter sequence. Specifically, for timestamps within a uniform time grid... If a matching measurement timestamp exists, the internal resistance value is used directly; otherwise... Falling on two measured timestamps and Between these points, linear interpolation is used to calculate the internal resistance value at that moment. The interpolation formula is as follows:

[0022] in To normalize the internal resistance parameter sequence in The internal resistance value at each time point is determined. For extrapolated times outside the measured time range, the nearest neighbor principle is used to fill in the gaps, i.e., the measured internal resistance value closest to that time point is taken. Through the above processing, a regularized sequence of internal resistance parameters that corresponds one-to-one with the denoised temperature data on the time grid is obtained, with its time resolution consistent with the temperature data, thus eliminating the time asynchrony problem.

[0023] The denoised temperature data and the regularized internal resistance parameter sequence are spliced ​​together in the fusion dimension to generate a multi-source monitoring data set.

[0024] It should be noted that the denoised temperature data reflects the multi-channel thermal state of the space, with the dimension of degrees Celsius; while the regularized internal resistance parameter sequence reflects the sum of the ohmic internal resistance and polarization internal resistance of the battery cell, with the dimension of milliohms. These two data have different physical meanings and significantly different numerical scales; directly concatenating them would cause the subsequent data-driven model to be biased towards the larger numerical dimension. Therefore, before concatenation, Z-score standardization was performed on both the denoised temperature data and the regularized internal resistance parameter sequence in the channel dimension to eliminate the influence of dimensions and unify the value range. The mean and standard deviation used for standardization were statistically obtained based on historical data collected from multiple sets of energy storage containers under normal cyclic operating conditions. This historical data covers the stable operating range under different ambient temperatures and charge / discharge rates.

[0025] S2: Based on multi-source monitoring data sets, a temperature field snapshot sequence characterizing the continuous spatiotemporal changes of the three-dimensional temperature distribution inside the container is constructed, as well as an internal resistance impedance spectrum evolution sequence characterizing the dynamic changes of the frequency response of the battery cell internal resistance.

[0026] In a specific embodiment of the present invention, the construction of the temperature field snapshot sequence includes: extracting distributed temperature sensing data from a multi-source monitoring data set, and constructing an initial three-dimensional temperature field for each time frame using Kriging space interpolation.

[0027] Specifically, the multi-source monitoring data set already contains standardized temperature and internal resistance data. Before proceeding to this step, these data are first restored to their original physical dimensions to support physical model calculations. The process of constructing a temperature field snapshot sequence begins with extracting distributed temperature sensing data from the multi-source monitoring data set. This distributed temperature sensing data includes… A fiber grating temperature sensor in a unified time grid On the noise reduction temperature value Each sensor has independent three-dimensional spatial coordinates. These coordinates were calibrated using a laser rangefinder and recorded in the configuration file during sensor installation. For each time frame... The initial three-dimensional temperature field at that moment was constructed using Kriging space interpolation.

[0028] It should be noted that the Kriging spatial interpolation is an optimal linear unbiased estimation method based on spatial statistics. Its core idea is to use the spatial correlation structure between known discrete sampling points to predict the attribute values ​​of unknown locations. The existing technology is relatively mature and will not be described in detail here.

[0029] The initial three-dimensional temperature field was physically calibrated using a heat conduction model of the shell structure of an energy storage container to obtain a calibrated three-dimensional temperature field.

[0030] It should be noted that directly using the initial three-dimensional temperature field as the basis for temperature field analysis carries risks, because Kriging interpolation is essentially a data-driven method and does not incorporate the heat transfer patterns between the energy storage container's shell structure and the ambient air. Near the container's shell boundary, where sensors are sparse, the interpolation results may exhibit anomalous temperature distributions that violate thermodynamic principles. Therefore, a thermal conduction model of the energy storage container's shell structure is used to perform physical consistency calibration on the initial three-dimensional temperature field.

[0031] The thermal conduction model of the energy storage container's shell structure is established based on Fourier's law of thermal conduction and the boundary convection heat transfer equation. The heat transfer process between the air domain inside the container and the solid battery module is described by a three-dimensional transient thermal conduction equation:

[0032] In the formula For the density of the medium, For the specific heat capacity of the medium, The thermal conductivity of the medium, This is the heat source for the battery cell. Within the space occupied by the battery cell, It is obtained by estimating real-time charge / discharge power and internal resistance parameters; in the air domain, Zero. The heat exchange between the inner surface of the container shell and the external environment is given by the third type of boundary condition.

[0033] in This represents the temperature gradient along the normal direction of the shell. The equivalent convective heat transfer coefficient of the shell, calculated based on the coupling of the container material's thermal resistance and natural air convection, is determined to be 8.5 W / m² / KV. The ambient temperature is provided in real time by weather sensors outside the container. The calibration process uses an initial three-dimensional temperature field. As an initial condition, the heat conduction model is applied to a three-dimensional mesh using the finite volume method for one time step. Transient solution, step size The time step is consistent with that of the multi-source monitoring data set. The obtained temperature field is denoted as... This step is equivalent to projecting and correcting the pure interpolation result towards a direction that simultaneously satisfies spatial smoothness and thermal conductivity physical constraints. Interpolation results near the shell boundary that exhibit excessively rapid temperature abrupt changes are smoothed by the diffusion effect of the thermal conductivity model and made compatible with the convective boundary conditions. This is the result of physical consistency calibration. This is to calibrate the three-dimensional temperature field.

[0034] The calibrated three-dimensional temperature fields are stacked in chronological order to form a sequence of temperature field snapshots describing the continuous spatiotemporal changes in temperature distribution.

[0035] Specifically, for all time frames Repeating the spatial interpolation and physical calibration operations above yields a time-sequential sequence of calibrated three-dimensional temperature fields. These calibrated three-dimensional temperature fields are then stacked along the time dimension as basic units to form a four-dimensional data structure consisting of three spatial dimensions and one temporal dimension; this is the temperature field snapshot sequence, denoted as... Each element in the temperature field snapshot sequence It fully describes the continuous spatiotemporal changes of the internal temperature distribution of an energy storage container over time.

[0036] In a specific embodiment of the present invention, the construction of the internal resistance impedance spectrum evolution sequence includes: extracting the voltage, current and surface temperature of the battery cell from the dynamic internal resistance parameters of the battery cell in the multi-source monitoring data set, performing online electrochemical impedance spectroscopy identification, and obtaining real-time impedance complex plane parameters.

[0037] From the dynamic internal resistance parameters of battery cells in the multi-source monitoring data set, the voltage, current, and surface temperature of each battery cell are extracted for online electrochemical impedance spectroscopy (EIS) identification. Online EIS identification does not require a dedicated shutdown for full-band frequency sweep excitation. Instead, it utilizes the wide-spectrum perturbation signal naturally present in the current of the battery cells under normal charge and discharge conditions, combined with low-amplitude multi-frequency superimposed excitation periodically injected by the battery management system to obtain wide-band impedance information.

[0038] Specifically, online electrochemical impedance spectroscopy (EIS) identification is achieved as follows: a low-amplitude multi-sinusoidal excitation signal is superimposed on the normal charge / discharge current of the battery cell, while the time-domain waveforms of the response voltage and excitation current are simultaneously acquired. A Fast Fourier Transform (FFT) is performed on the acquired signals to obtain the complex amplitudes of the voltage and current at each frequency point. The real-time impedance complex plane parameter is defined as follows: at each frequency point, the complex amplitude of the response voltage is divided by the complex amplitude of the excitation current; the resulting complex number is the impedance at that frequency. To improve noise immunity, the spectra of multiple consecutive time frames are averaged, and a window function is applied to the time-domain signal to reduce spectral leakage. The curve connecting the impedance values ​​at all frequency points on the complex plane is the real-time impedance complex plane parameter. This parameter contains the electrochemical impedance spectral information of the battery cell. After obtaining this parameter, subsequent frequency smoothing and outlier removal processing are performed.

[0039] Frequency smoothing and outlier removal are applied to the real-time impedance complex plane parameters to obtain denoised impedance complex plane data.

[0040] It should be noted that real-time impedance complex plane parameters are inevitably affected by electromagnetic interference and current sensor quantization noise in engineering measurement environments, manifesting as isolated jumps or irregular sawtooth fluctuations in the high-frequency band on the complex plane curve in the original identification results. Frequency smoothing and outlier removal are applied to the real-time impedance complex plane parameters. Frequency smoothing uses a sliding median filter along the frequency axis, with a sliding window width of 5 frequency points. The median of the real and imaginary parts of the impedance at each frequency point within the window is taken as the smoothed value at the window's center frequency. Outlier removal targets isolated points that still deviate from the complex plane trajectory beyond the criteria after smoothing: the mean Euclidean distance between the impedance vectors of three adjacent frequency points after smoothing is calculated. If the distance between the impedance vector of a frequency point and the locally fitted line formed by the preceding and following points exceeds three times this mean, it is identified as an outlier and replaced with linear interpolation of the preceding and following points. After these two steps, denoised impedance complex plane data is obtained, denoted as […]. .

[0041] The noise-reduced impedance complex plane data is compared with the battery's factory reference impedance spectrum. The impedance magnitude offset and phase angle offset are calculated and combined into an internal resistance offset feature vector.

[0042] Specifically, the battery's factory-standard impedance spectrum is a complex impedance spectrum obtained at the time of manufacture by performing a full-band frequency sweep test from 0.01 Hz to 10 kHz using a precision electrochemical workstation at an ambient temperature of 25 degrees Celsius and a state of 50% charge. It is denoted as... The reference spectrum is stored in the database as a data table and is bound to the battery serial number. For each time frame... and battery cells Calculate the comprehensive impedance offset index; specifically, calculate the denoised real-time complex impedance spectrum. and reference impedance spectrum Treating them as two vector sequences on the complex plane, the normalized Euclidean distance between them is calculated as a composite impedance offset index. , This indicates the calculation of the minimum angle between two angles in the complex plane. and These represent the frequency points of battery cell j under historical normal operating conditions. The standard deviation of impedance magnitude and phase angle at the point of intersection The total number of frequency points constitutes the internal resistance offset characteristic vector. .

[0043] Arrange the internal resistance offset eigenvectors in chronological order to generate an internal resistance impedance spectrum evolution sequence.

[0044] For each time frame The above online identification, denoising, and offset calculations are performed on all cells to obtain the internal resistance offset feature vectors of each cell at different times. Arranging the internal resistance offset feature vectors of all cells at all times in chronological order generates the internal resistance impedance spectrum evolution sequence, denoted as... In the evolution sequence of internal resistance impedance spectrum, the first... Each battery cell at time The evolutionary data is The study fully recorded the dynamic change trajectory of the internal resistance frequency response of each battery cell, laying the foundation for subsequent capture of the time-series correlation of electrochemical degradation-induced thermal anomalies.

[0045] This invention constructs a three-dimensional temperature field snapshot sequence inside the energy storage container and an internal impedance spectrum evolution sequence of the battery cells simultaneously. It jointly characterizes the system's operating status from two dimensions: spatial thermal distribution and electrochemical frequency response. This overcomes the limitation that monitoring a single physical quantity cannot fully reflect the thermal-electric coupling fault mechanism. It enables fault diagnosis to cover the complete evolution chain from electrochemical degradation to thermal runaway, laying a multi-source data foundation for achieving early and reliable early warning.

[0046] S3: Simultaneously input the temperature field snapshot sequence and the internal resistance impedance spectrum evolution sequence into the thermo-electric relaxation coupling mapping network to extract the time lag correlation pattern of the temperature change rate feature of the local temperature field relative to the change of the internal resistance impedance spectrum, and generate a local thermal diffusion lag correlation feature map.

[0047] It should be noted that S2 has constructed temperature field snapshot sequences and internal resistance impedance spectrum evolution sequences to describe the operating state of the energy storage container from the perspectives of spatial thermal distribution and electrochemical state, respectively. However, the correlation between the two is not instantaneous and synchronous, but rather includes an inherent relaxation lag between thermal processes and electrochemical changes. To extract this time-lag correlation pattern between the temperature change rate characteristics of local temperature fields and changes in the internal resistance impedance spectrum, the temperature field snapshot sequences and the internal resistance impedance spectrum evolution sequences need to be synchronously input into a specially designed thermal-electric relaxation coupling mapping network. This network consists of three parts: a parallel time encoder, a cross-scale relaxation attention module, and an embedding layer, which captures the latent representation of thermal-electric relaxation coupling in a data-driven manner.

[0048] In a specific embodiment of the present invention, the thermal-electric relaxation coupling mapping network includes a parallel temporal encoder, a cross-scale relaxation attention module, and an embedding layer. The generation of local thermal diffusion hysteresis correlation feature map includes: using the temporal encoder to perform temporal convolution encoding on the temperature field snapshot sequence and the internal resistance impedance spectrum evolution sequence, respectively, to generate a temperature field encoding sequence and an internal resistance impedance spectrum encoding sequence.

[0049] Specifically, the temperature field snapshot sequence is denoted as Its structure is a four-dimensional array, containing the time frame dimension. and spatial three-dimensional grid coordinates The evolution sequence of the internal resistance impedance spectrum is denoted as... Its structure includes a time frame dimension. and battery cell index dimension Each element is a two-dimensional internal resistance offset feature vector. Both are then fed into a timing encoder for processing.

[0050] The temporal encoder extracts high-level coding features with dynamic multi-timescale representation capabilities from the original sequence through temporal convolution operations.

[0051] A temporal encoder for temperature field snapshot sequences needs to simultaneously process the temporal evolution of every voxel point in the spatial dimension. This encoder employs three stacked layers of causal dilated convolutions. The causal dilated convolutions ensure that the encoded output at the current moment depends only on information from the current and past moments, preventing the introduction of future data leakage. Let a certain spatial location... Temperature time series The input is the first convolutional layer. The dilation rate of each convolutional layer doubles progressively, set to 1, 2, and 4. The kernel length is set to 7 time steps, and the number of output channels is 32, 64, and 128 respectively. The calculation of the dilated convolution can be expressed as: for the... Layer convolution, dilation rate The convolution kernel weights are , bias is The input features are Output features Depend on Given. In the formula This refers to the index of the weights within the convolution kernel. To increase the dilation rate, the receptive field is expanded by skipping intermediate sampling points. A linear rectified function with leakage is used as the activation function after each convolutional layer, with a negative slope of 0.1. The feature map output from the three convolutional layers maintains the same stride as the original sequence in the temporal dimension, but the number of feature channels at each time step becomes 128. This process iterates through all spatial locations. Performing the same convolution parameter processing, we obtain the temperature field encoded sequence, denoted as... Its spatial dimensions remain unchanged, and each spatial point is a 128-dimensional vector.

[0052] A time-series encoder for the evolution sequence of internal resistance impedance spectrum, processing each battery cell. Internal resistance offset eigenvector sequence The eigenvector was originally two-dimensional and included the impedance magnitude offset. and phase angle offset Since the current multiplier modulates the impedance offset amplitude, to eliminate the multiplier effect and make the encoding focus more on the trend of change, the two offset components are first divided by the normalized current multiplier factor at that moment. This factor is defined as the ratio of the absolute value of the current to the rated current and clamped between 0.01 and 1. The preprocessed input vector is: The timing encoder also employs a three-layer causal dilated convolution with dilation rates of 1, 2, and 4, a kernel length of 7, and output channels increasing to 16, 32, and 64 layer by layer. The sequence of each battery cell is independently passed through this encoder to generate a 64-dimensional internal impedance spectrum encoded sequence, denoted as... This encoding sequence encodes the slow trend and rapid fluctuations of internal resistance offset in multi-scale features.

[0053] The time slices of each spatial local region in the temperature field coding sequence are calculated by the cross-scale relaxation attention module to calculate the hysteresis correlation weights of each scale component in the internal impedance spectrum coding sequence, generating a relaxation hysteresis correlation matrix. This relaxation hysteresis correlation matrix characterizes the delayed response intensity of the temperature change rate relative to the change in the internal impedance spectrum.

[0054] The task of the cross-scale relaxation attention module is to explicitly model the hysteresis correlation weights of each scale component in the internal impedance spectrum encoding sequence for time slices of local temperature fields.

[0055] The specific definition of the local temperature field region borrows from the geometric projection relationship of the battery cell within the container: for each battery cell Based on its installation coordinates inside the container A cubic sub-region with a side length of 0.3 meters, centered at this coordinate, is defined within the temperature field grid. This sub-region contains the thermal states of the battery cell and its neighboring air domain. The temperature field encoding sequence is then used to... In this process, the encoded vectors of all grid points within the sub-region are extracted, and the vectors of these spatial points are averaged and pooled to obtain a time series vector representing the local temperature field state of the battery cell. The dimension is 128.

[0056] In order to capture the characteristics of the rate of temperature change, rather than just the absolute level of temperature, Calculate the first-order time difference to obtain the eigenvector of the temperature change rate. This operation highlights the dynamic process of heat diffusion.

[0057] Cross-scale attention calculations are performed between the local temperature change rate feature sequence and the internal resistance impedance spectrum encoding sequence, introducing a controllable time lag variable. Each dimension of the internal resistance impedance spectrum encoding sequence is treated as an independent scale component, with different scale components corresponding to different electrochemical relaxation time constants. For each cell and each lag time step, this module calculates the correlation strength of the local region under that lag using the following method: First, the current temperature change rate feature vector is mapped to a query vector with the same dimension as the internal resistance encoding vector through a fully connected network. Then, for each scale component (i.e., each dimension of the internal resistance encoding vector), historical values ​​of that component over multiple time steps are extracted, forming a key sequence where each key is the scalar value of that component at a specific historical time step. Next, the similarity between the query vector and the keys at each historical time step is calculated. Specifically, each element of the query vector is multiplied by the key value at that historical time step, and the sum is obtained to get a scalar similarity. Softmax normalization is applied to the similarities across all historical time steps to obtain the attention weight of that scale component at that lag time. This weight reflects the probability that the internal resistance scale component will respond after a lag of that many time steps when the temperature changes at the current rate.

[0058] Since different scale components have different physical meanings—for example, low-frequency components correspond to diffusion processes, and high-frequency components correspond to charge transfer—the above calculation process is performed independently for each scale component, thereby obtaining an attention weight associated with the scale component index. Finally, the attention weights of all scale components are averaged to obtain the comprehensive correlation strength of the battery cell at that hysteresis time. Repeating the above calculation for each hysteresis time from 0 to a preset maximum hysteresis, such as 300 seconds, yields a hysteresis correlation vector. This vector describes the distribution of the delayed response intensity of the temperature change rate relative to the change in the internal resistance spectrum.

[0059] The relaxation hysteresis correlation matrix is ​​input into the embedding layer for spatial-scale dimension compression to obtain the local thermal diffusion hysteresis correlation feature map.

[0060] It should be noted that the embedding layer consists of two fully connected network layers. The first layer transfers the hysteresis correlation vector of each battery cell from... The first layer compresses the 32-dimensional representations of all battery cells to 32 dimensions, using orthogonal linear units as the activation function. The second layer performs cross-cell interaction on the 32-dimensional representations of all battery cells, employing a multi-head self-attention mechanism with four heads, each focusing on other cells to fuse spatial context. The final output is a 16-dimensional feature representation compressed to each battery cell. This... The 16-dimensional features of each battery cell are reconstructed into a two-dimensional graph structure according to the spatial arrangement of the batteries within the container, forming a local thermal diffusion hysteresis correlation feature map. Each spatial node in this feature map corresponds to a battery cell, and the 16-dimensional vector on each node encapsulates the hysteresis coupling information of temperature changes at that location on the internal resistance spectrum changes of itself and neighboring cells.

[0061] This invention introduces a thermo-electric relaxation coupling mapping network to extract the time lag correlation pattern of the temperature change rate in a local temperature field relative to the change in the internal resistance spectrum. It explicitly models the inherent relaxation lag characteristics of the thermal process to the electrochemical change, distinguishing the weak relaxation characteristic oscillations generated by irreversible side reactions or structural damage from the fluctuations of normal operation, thereby improving the sensitivity and location accuracy of early latent faults such as loose connections and internal micro-short circuits.

[0062] S4: Perform time-frequency transformation on the local thermal diffusion hysteresis correlation feature map, extract the relaxation feature oscillation component generated by the irreversible thermal process, and generate the fault relaxation feature spectrum.

[0063] It should be noted that the local thermal diffusion hysteresis correlation feature map generated by S3 condenses the hysteresis coupling information between the local temperature change rate and the internal resistance spectrum change of each battery cell. However, this information still appears as a discrete sequence evolving with each sampling frame in the time dimension. The relaxation characteristic oscillation component excited by the irreversible thermal process caused by the fault is masked by normal operation fluctuations in the time domain. Therefore, the main task of S4 is to perform time-frequency transformation on the local thermal diffusion hysteresis correlation feature map to extract these relaxation oscillations hidden in the time domain signal, forming a structured fault relaxation characteristic spectrum to support subsequent matching and identification with the fault mode library.

[0064] In a specific embodiment of the present invention, generating the fault relaxation feature spectrum includes: performing a short-time Fourier transform on the local thermal diffusion hysteresis correlation feature map along the time dimension to obtain a relaxation feature time-frequency map.

[0065] It should be noted that the local thermal diffusion hysteresis correlation feature map has been stored as a 16-dimensional feature vector sequence for each battery cell node by the end of S3, with the time frame number being the same as the input sequence, denoted as . The feature map is denoted as... ,in This is the battery cell index, with a value of Each It is a 16-dimensional vector. To perform time-frequency analysis independently on each battery cell while preserving the differences in thermal diffusion coupling between cells, the feature map is decomposed along the battery cell dimension, forming... The data consists of several parallel multivariate time series, each corresponding to a 16-dimensional feature trajectory of a battery cell across all time frames.

[0066] Specifically, the short-time Fourier transform (SFT) of the local thermal diffusion hysteresis correlation feature map along the time dimension is performed separately for each battery cell feature sequence. The SFT is a commonly used time-frequency analysis method that maps non-stationary time-domain signals to the time-frequency plane. It is implemented by using a finite-length window function that slides along the time axis with a fixed step size, performing a Fourier transform on the signal segments within the window to obtain the signal spectrum's variation over time. For battery cells... The dimensional feature components, whose time series are denoted as ,in The total number of time frames is given, with a 1-second time interval between adjacent frames. A Hanning window of length 64 seconds is used. The window function at discrete time points is defined as: when the discrete time index... When the absolute value of the function does not exceed half the window length, i.e., 32, the window function value is Otherwise, the value is 0. The sliding step of the window function is set to 4 seconds, meaning the window moves once every 4 seconds, with a 60-second overlap between adjacent windows. For each window position, the 64 consecutive feature values ​​within the window are multiplied by the window function value, and then a Fast Fourier Transform is performed on the product sequence to obtain the spectrum corresponding to the center time of that window. The frequency range of the spectrum is from 0 to 0.5 Hz, which is determined by the sampling theorem; this is only an example, and the frequency resolution is approximately 0.0078 Hz. Repeating the above operation for all window positions yields the time-frequency plot of the feature component, where the horizontal axis represents time and the vertical axis represents frequency. The value of each time-frequency grid point represents the energy intensity of the frequency component near that time.

[0067] For battery cells The 16 characteristic components are subjected to the above short-time Fourier transform, resulting in 16 time-frequency spectra. The squared modulus of each time-frequency spectrum, i.e., the power spectral density, is averaged along the characteristic dimension to obtain the relaxation characteristic time-frequency diagram of battery cell j, denoted as... Its dimension is the original characteristic variance density. The relaxation characteristic time-frequency diagram intuitively shows the energy distribution of the thermal-electric relaxation coupling characteristics of the battery cell at different time and frequency scales. When the irreversible thermal process caused by the fault excites relaxation oscillations of a specific frequency, the energy will be concentrated in the corresponding time-frequency region.

[0068] The baseline component and high-frequency random fluctuation component caused by periodic charging and discharging are separated from the relaxation characteristic time-frequency diagram, and the coherent oscillation component within the battery thermal fault relaxation response frequency band is retained to form a filtered relaxation time-frequency representation.

[0069] It should be noted that the relaxation characteristic time-frequency graph contains signal components from multiple sources, and not all components are related to the fault. Periodic charging and discharging causes the battery cell temperature and internal resistance to fluctuate regularly with the charging and discharging cycle, which is represented on the time-frequency graph as a baseline component strictly synchronized with the charging and discharging cycle, its frequency position corresponding to the fundamental frequency of charging and discharging and its integer multiples of harmonics. High-frequency random fluctuation components originate from sensor electronic noise, micro-temperature disturbances caused by cooling fan turbulence, and random interference in the electromagnetic compatibility environment, appearing as irregularly scattered spots in the high-frequency band on the time-frequency graph. The true fault characterizer is the relaxation characteristic oscillation component excited by irreversible thermal processes. This type of oscillation has physical causal constraints: it only occurs after irreversible side reactions or structural damage occur inside the battery, and its oscillation frequency falls within a specific battery thermal fault relaxation response frequency band.

[0070] Specifically, the determination of the battery thermal fault relaxation response frequency band is based on the physical mechanism of the incubation period of lithium-ion battery thermal runaway. Statistical analysis of measured data from multiple energy storage containers covering normal cycling, overcharging, internal short circuits, and loose connections revealed that in the early stages of a fault, the dynamic imbalance between the heat release from side reactions and the thermal diffusion within the cell triggers temperature-impedance coupling oscillations in the frequency range of 0.005 Hz to 0.1 Hz. The lower limit of this frequency band, 0.005 Hz, corresponds to an ultra-long period of approximately 200 seconds, matching the thermal inertia time of the cell's thermal mass; the upper limit, 0.1 Hz, corresponds to a period of approximately 10 seconds, matching the response speed established by internal thermal diffusion within the cell. Therefore, the battery thermal fault relaxation response frequency band is set as follows: hertz.

[0071] In order to obtain the relaxation characteristic time-frequency plot The baseline component and high-frequency random fluctuation component caused by periodic charging and discharging are separated, while the coherent oscillation component within the battery thermal fault relaxation response frequency band is retained. A processing strategy combining frequency domain masking and time domain filtering is adopted. First, the base frequency corresponding to the charging and discharging cycle is determined: the charging and discharging cycle of the current operating condition is provided by the battery management system and is set as follows. Seconds, corresponding to the base frequency is Hertz. On a time-frequency graph, with... and its first three harmonics Centered on a specific frequency, narrowband notch masks with a width of 0.002 Hz are set to directly zero the time-frequency energy at these frequency locations to eliminate baseline components. For high-frequency random fluctuation components, a cutoff frequency is set. Hertz, a frequency higher than The time-frequency energy is set to zero because the fault-related thermal relaxation process does not involve rapid oscillations above 0.5 Hz. This threshold references the clustering region of the noise energy spectrum in multiple sets of measured data. After the above two-step frequency domain masking process, the remaining time-frequency energy is the coherent oscillation component within the battery thermal fault relaxation response frequency band, forming the filtered relaxation time-frequency representation, denoted as . .

[0072] The filtered relaxation time-frequency representation is normalized along the frequency axis and the frequency resolution is unified to generate the fault relaxation characteristic spectrum.

[0073] It should be noted that the filtered relaxation time-frequency representation Only in the frequency range While non-zero energy exists within the Hertzian sphere, the bias level and dynamic range of the eigenvectors of different battery cells may vary. This is due to inconsistent heat dissipation conditions caused by the different spatial positions of each cell within the container, as well as individual differences in manufacturing processes. To eliminate amplitude differences caused by these non-fault factors, the filtered relaxation time-frequency representation is normalized along the frequency axis, and the frequency resolution is standardized.

[0074] The normalization process employs global maximum normalization along the frequency axis, i.e., for each battery cell... Calculate the maximum energy value of its filtered relaxation time-frequency representation across all time and frequency ranges. .

[0075] Then, the normalized fault relaxation characteristic spectrum is defined as follows: .

[0076] In the formula To prevent small constants with denominators of zero, take Unified frequency resolution refers to adjusting the frequency sampling interval of the spectrum of different battery cells to be consistent. Since the frequency resolution of the short-time Fourier transform is determined by the window length and all cells use the same parameters, this condition is naturally satisfied. The final generated fault relaxation characteristic spectrum... Each battery cell The two-dimensional normalized time-frequency matrix is ​​defined, with the frequency axis limited to 0.005 Hz to 0.1 Hz and the time axis covering the entire analysis period. The value of each time-frequency grid point is dimensionless, ranging from 0 to 1. This fault relaxation characteristic spectrum directly characterizes the energy distribution of the relaxation characteristic oscillation component of the battery cell excited by the irreversible thermal process in the time-frequency domain.

[0077] S5: Calculate the coherence matching degree between the fault relaxation feature spectrum and the standard relaxation spectrum of each fault prototype in the fault mode library. When the matching degree meets the dynamic early warning condition generated based on the spectrum fluctuation statistics under historical normal operation, output a fault early warning signal containing the fault location and fault type.

[0078] It should be noted that the fault relaxation feature spectrum generated by S4 provides a normalized time-frequency energy distribution for each battery cell, with the frequency axis limited to the battery thermal fault relaxation response band from 0.005 Hz to 0.1 Hz. The main task of S5 is to match and compare these fault relaxation feature spectra with a pre-established fault mode library, and output the exact fault location and fault type when the dynamic warning conditions are met.

[0079] In a specific embodiment of the present invention, the output of the fault warning signal containing the fault location and fault type includes: calculating the correlation coefficient between the fault relaxation feature spectrum and the standard relaxation spectrum of each fault prototype in the fault mode library, as the coherence matching degree.

[0080] It should be noted that the correlation coefficient measures the degree of linear correlation between the energy distribution patterns of two two-dimensional time-frequency matrices. It is insensitive to the absolute value of the spectral amplitude but focuses on capturing the morphological similarity of the time-frequency structure. This perfectly matches the characteristic that the normalized spectrum of fault relaxation features has a uniform amplitude range but different time-frequency patterns. Calculating the correlation coefficient requires simultaneously expanding all elements of the two time-frequency matrices in both the time and frequency dimensions, flattening the two-dimensional matrices into a one-dimensional vector, and then calculating the Pearson product-moment correlation coefficient. Let the fault relaxation feature spectrum... The values ​​taken at discrete time-frequency grid points form a vector. , No. Standard relaxation spectrum of a fault prototype The values ​​taken at the corresponding discrete time-frequency grid points form a vector. The vector lengths are all This is equal to the number of time frames multiplied by the number of frequency points. The coherence matching degree is calculated by comparing the fault relaxation feature spectrum with the first fault mode in the fault mode library. The two-dimensional normalized cross-correlation between the standard relaxation feature spectra of each fault prototype is used as the coherence matching degree. .

[0081] The peak matching degree is selected from the coherence matching degree corresponding to all fault prototypes.

[0082] For each battery cell Traverse all fault modes in the fault mode library A fault prototype was calculated. Each coherence matching value The peak matching degree is selected from the coherence matching degrees corresponding to all fault prototypes, that is, the maximum value among all matching degrees corresponding to the battery cell is taken.

[0083] Simultaneously record the fault prototype index that achieved this peak matching degree. Peak matching degree Characterizing battery cells The matching strength between the current fault relaxation feature spectrum and the most similar fault mode, index This points to the fault prototype most relevant to it.

[0084] When the peak matching degree is greater than the dynamic warning threshold obtained by statistically analyzing the spectrum fluctuations under historical normal operating conditions, the fault type and faulty battery cell location associated with the fault prototype corresponding to the peak matching degree will be output as a fault warning signal.

[0085] It should be noted that the dynamic early warning threshold is not a fixed value, but an adaptive threshold obtained by statistically analyzing the spectral fluctuations under historical normal operating conditions. Historical normal operating conditions refer to the accumulated operating data of the energy storage container during the period when it is confirmed to be fault-free through regular maintenance and all cell performance tests are passed. This data covers the thermo-electric relaxation coupling response under different ambient temperatures, different charge / discharge rates, and different states of charge ranges. For each battery cell under historical normal operating conditions, the coherence matching degree between the fault relaxation characteristic spectrum and each fault prototype is continuously calculated in the background according to the process from S1 to S5, forming a historical sequence of normal operating matching degrees. Due to the existence of noise, operating condition fluctuations, and model approximation errors, even normal battery cells will produce non-zero matching degrees, but the values ​​are low and exhibit random fluctuations. In this embodiment, the dynamic early warning threshold is set as the mean of the historical sequence of normal operating matching degrees plus a statistical upper bound of 3 times the standard deviation.

[0086] The fault mode library is an offline-built knowledge base, where each fault prototype corresponds to a typical fault mode encountered by the energy storage battery in actual operation. These include five fault types: overcharging leading to solid electrolyte interface film decomposition, lithium plating at the negative electrode inducing dendrite growth, internal micro-short circuits within the cell, loose busbar connections, and capacity degradation of the battery cell. The standard relaxation spectrum for each fault prototype is obtained offline by applying the corresponding fault excitation condition to the same type of battery cell under controlled laboratory conditions, collecting complete temperature field snapshot sequences and internal resistance spectrum evolution sequences, and then extracting the fault relaxation feature spectrum offline according to the exact same processing procedures as S3 and S4. The standard relaxation spectrum is stored in the form of a normalized time-frequency matrix, possessing the same time resolution, frequency resolution, and frequency axis range as the real-time generated fault relaxation feature spectrum. Let the first... The standard relaxation spectrum of a fault prototype is: , , This represents the total number of faulty prototypes.

[0087] This invention utilizes the coherence matching of fault relaxation feature spectrum with fault mode library and combines it with the dynamic early warning threshold generated by historical normal operation fluctuation statistics. This avoids false alarms or missed alarms caused by fixed thresholds when operating conditions change. At the same time, it directly outputs the fault type and fault battery cell location through fault prototype matching, providing the cooling control system with a precise loop enhancement cooling target. This realizes the closed-loop linkage of diagnosis, location and emergency heat dissipation, effectively suppressing the spread of fault thermal state.

[0088] S6: Generates loop enhancement cooling target and cooling parameter adjustment commands, and drives the cooling control system.

[0089] It should be noted that while the fault warning signal output by S5 clearly identifies the location and type of the faulty battery cell, the diagnosis itself does not directly mitigate the fault's progression. S6 is triggered upon receiving the fault warning signal. Its primary task is to translate the diagnostic results into targeted cooling control actions. By adjusting the operating parameters of the thermal management system, it enhances heat dissipation in the faulty area, thereby suppressing heat accumulation in the faulty battery cell and delaying the thermal runaway process, thus buying time for maintenance personnel to respond. The fault warning signal is transmitted to S6 in the form of a structured data message, which contains at least two fields: a faulty battery cell location field storing the faulty battery cell index, and a fault type field storing the fault prototype's tag string.

[0090] In a specific embodiment of the present invention, the generation loop enhances the cooling target and cooling parameter adjustment command, and drives the cooling control system, including: parsing the location of the faulty battery cell in the fault warning signal, locating the thermal management loop where the faulty battery cell is located in the temperature field snapshot sequence, and generating the loop to enhance the cooling target.

[0091] Specifically, the first step in parsing the fault warning signal is to map the location of the faulty battery cell to the thermal management loop topology of the energy storage container. The cooling system of the energy storage container adopts a multi-loop parallel architecture design. Each thermal management loop covers a row of battery modules arranged along the length of the container, and each row contains eight battery cells. The division of the thermal management loops is fixed during the container design phase and stored as a configuration file in the programmable logic controller (PLC) of the cooling control system. This configuration file records the correspondence between the unique identifier of each loop and the set of battery cell indices it covers. The correspondence between loops and battery cells can be represented as a mapping function. Index the battery cells Mapping to loop identifier , , The total number of loops is 32. In this embodiment, 32 battery cells correspond to 4 loops, and each loop covers 8 consecutively numbered battery cells. During parsing, the index value of the faulty battery cell location field is read, the mapping function is queried, and the thermal management loop identifier of the battery cell is determined. This loop is then marked as the loop enhancement cooling target.

[0092] The enhanced cooling target for the loop includes a loop identifier and an assessment of the urgency of current heat dissipation needs based on a temperature field snapshot sequence. This temperature field snapshot sequence is constructed in S2 and continuously updated, with S6 directly reading the latest snapshot from memory. The spatial grid region of all battery cells covered by the loop identifier is located in the three-dimensional temperature field, and the peak value of all grid points within that region is calculated. and the maximum value of temperature time difference in this region. The latter reflects the local temperature rise rate. These two quantities are written into the data structure of the loop-enhanced cooling target as the basis for subsequently determining the cooling regulation intensity.

[0093] Based on the fault type corresponding to the fault warning signal, determine the adjustment strategy for the cooling medium flow rate and generate cooling parameter adjustment instructions.

[0094] It should be noted that different fault types result in significant differences in the risk level of thermal runaway and the heat generation mechanism within the battery, requiring correspondingly different cooling response strategies. The adjustment strategy is retrieved from a pre-set rule base using a lookup table approach. This rule base is based on open-loop experimental data of battery thermal behavior characteristics under different fault types.

[0095] For example, for a fault of "loose busbar connection," the thermal characteristic is increased contact resistance leading to localized Joule heat accumulation. The temperature rise is relatively slow but continuous, and it is unlikely to cause instantaneous thermal runaway. The corresponding adjustment strategy is to increase the cooling medium flow rate of the target circuit to 1.5 times the rated flow rate. For a fault of "internal micro-short circuit in the cell," the thermal characteristic is rapid heat release at the internal short circuit point. The temperature rise rate is high and it may develop into thermal runaway within minutes. The corresponding adjustment strategy is to immediately increase the cooling medium flow rate of the target circuit to 2.0 times the maximum design flow rate, while simultaneously increasing the flow rate of adjacent upstream and downstream circuits to 1.3 times to establish a thermal isolation barrier. For a fault of "overcharging causing solid electrolyte interface film decomposition," the thermal characteristic is exothermic side reactions leading to a gradual temperature rise, accompanied by gas production. The adjustment strategy is to increase the flow rate of the target circuit to 1.8 times. For a fault of "negative electrode lithium plating causing dendrite growth," the adjustment strategy is to increase the flow rate of the target circuit to 1.6 times and limit the charging power. For the "battery cell capacity degradation and differentiation" type of fault, the thermal risk is relatively low. The adjustment strategy is set to increase the target circuit flow to 1.3 times and trigger the equalization maintenance prompt at the same time.

[0096] The calculation process of converting the cooling medium flow regulation strategy into cooling parameter adjustment commands requires mapping the rate adjustment amount to the physical parameters of the actuators in the cooling control system. The cooling circuit of the energy storage container uses a variable frequency water pump to drive the coolant circulation, with the flow rate adjusted by the pump speed. Assume the circuit under rated operating conditions... The water pump speed is Its corresponding rated flow rate of the circuit The pump speed and flow rate have an approximately linear relationship within the normal operating range. The scaling factor given by the adjustment strategy is denoted as... The target traffic is The corresponding target pump speed is from calculate.

[0097] The cooling parameter adjustment command includes not only the target water pump speed but also a dynamic bias term. This bias term is fine-tuned based on the temperature peak and temperature rise rate carried by the loop-enhanced cooling target. The calibration formula is as follows: .

[0098] In the formula For reference, the upper limit of the safe operating temperature of the battery cell is taken as 45 degrees Celsius. The temperature deviation feedback coefficient is set to 15 revolutions per minute per degree Celsius. The temperature rise rate feedforward coefficient is set to 200 revolutions per minute per degree Celsius per second. The values ​​of the two coefficients are determined based on the system identification results of multiple sets of cooling circuit step response tests. For loop The maximum rate of temperature change of all grid points in the current frame relative to the previous frame. The limiter constrains the water pump's physical executable speed between the minimum speed of 0 and the maximum speed of 6000 rpm, thus forming the final cooling parameter adjustment command.

[0099] The loop enhancement cooling target and cooling parameter adjustment command are sent to the cooling control system of the energy storage container to drive the corresponding loop to perform enhanced heat dissipation.

[0100] S7: Train the parameter evolution prediction network using temperature field snapshot sequences and internal resistance spectrum evolution sequences, and generate the predicted fault relaxation feature spectrum to achieve predictive early warning.

[0101] It should be noted that S1 to S6 construct a complete closed loop from multi-source sensor data acquisition to fault diagnosis and early warning, and then to cooling linkage response. However, the diagnosis timing depends on the presence of identifiable abnormal patterns in the fault relaxation characteristic spectrum. S7 runs in parallel outside the above process, using a parameter evolution prediction network to extrapolate the thermal-electrical parameters for future time periods and sending the extrapolation results into the diagnostic link in advance. This allows for the output of a predictive early warning signal before the actual occurrence of the fault, shifting the early warning timing from the stage where the fault has already manifested to the stage of fault precursors.

[0102] In a specific embodiment of the present invention, the step of training a parameter evolution prediction network using a temperature field snapshot sequence and an internal resistance spectrum evolution sequence, and generating a predictive fault relaxation feature spectrum to achieve predictive early warning, includes: The parameter evolution prediction network is trained using the historically stored temperature field snapshot sequence and internal resistance spectrum evolution sequence to obtain the thermo-electric parameter evolution prediction model.

[0103] The recently acquired temperature field snapshot sequence and internal resistance spectrum evolution sequence are input into the thermo-electric parameter evolution prediction model, and extrapolation is used to generate the predicted temperature field snapshot sequence and predicted internal resistance spectrum evolution sequence for future time periods.

[0104] The predicted temperature field snapshot sequence and the predicted internal resistance spectrum evolution sequence are input into the thermo-electric relaxation coupling mapping network to obtain the predicted local thermal diffusion hysteresis correlation feature map. The predicted fault relaxation feature spectrum is obtained by performing time-frequency transformation on the predicted local thermal diffusion hysteresis correlation feature map.

[0105] By utilizing the predicted fault relaxation characteristic spectrum, coherence matching is performed in advance so as to output a predictive warning signal before the fault actually occurs.

[0106] It should be noted that the parameter evolution prediction network is a deep learning model based on a hybrid of temporal convolutional structures and gated recurrent units, specifically designed to simultaneously capture the spatial diffusion evolution of the temperature field and the asymptotic shift trend of the internal resistance spectrum. The network takes as input a sequence of temperature field snapshots and an internal resistance spectrum evolution sequence within a continuous historical time window, and outputs predicted values ​​for the temperature field snapshots and internal resistance spectrum evolution sequences for future consecutive time periods immediately following that window.

[0107] The internal structure of the parameter evolution prediction network consists of three parts: a shared feature extraction backbone, a temperature field prediction branch, and an internal resistance impedance spectrum prediction branch. The shared feature extraction backbone receives a multimodal input consisting of a concatenated temperature field snapshot sequence and an internal resistance impedance spectrum evolution sequence. It employs a four-layer stack of 3D causal dilated convolutions, with kernels of length 5 in the time dimension, 3 in the temperature field spatial dimension, and 1 in the internal resistance cell dimension, with dilation rates of 1, 2, 4, and 8 respectively, resulting in 32, 64, 128, and 256 output channels. Each convolutional layer is followed by a normalization layer and a leaky linear rectified activation function with a negative slope of 0.1. The output from the backbone is fed into two parallel branches: the temperature field prediction branch uses a stack of 3D transposed convolutions to progressively restore the spatial resolution to match the input temperature field, with its final layer activation function being a linear function to support continuous temperature value output; the internal resistance impedance spectrum prediction branch maps features to a two-dimensional space of the internal resistance offset feature vector through fully connected layers, also employing linear activation.

[0108] The loss function during training consists of a weighted sum of two parts: temperature field prediction loss and internal resistance prediction loss. The temperature field prediction loss is defined as the mean square error of all grid points across all prediction frames between the predicted temperature field snapshot sequence and the true temperature field snapshot sequence. The internal resistance prediction loss is defined as the mean square error of the predicted internal resistance offset feature vector across all battery cells and all prediction frames between the predicted internal resistance offset feature vector and the true internal resistance offset feature vector.

[0109] It should also be noted that even if the early warning signal is triggered during the prediction phase, the real-time diagnostic link will not be interrupted. After the real data covers the prediction period, the system will verify the consistency between the real-time generated fault relaxation feature spectrum and the prediction result, and feed the verification error back to the incremental learning link of the parameter evolution prediction network to continuously improve the extrapolation accuracy.

[0110] This invention extrapolates the future evolution of temperature field and internal impedance spectrum through parameter evolution prediction network, and sends the extrapolation results into the diagnostic link in advance to generate predicted fault relaxation characteristic spectrum and perform predictive early warning. This moves the early warning time point from the stage where the fault can be diagnosed in real time to the stage of fault precursor, thus gaining sufficient time window for planned shutdown and proactive maintenance.

[0111] Reference Figure 2 As shown, a second aspect of the present invention provides a system for performing the intelligent diagnosis and early warning method for faults in an energy storage container, comprising: a data preprocessing module for acquiring distributed temperature sensing data and dynamic internal resistance parameters of battery cells within the energy storage container, and generating a multi-source monitoring data set.

[0112] The sequence construction module is used to construct a temperature field snapshot sequence that characterizes the continuous spatiotemporal changes of the three-dimensional temperature distribution inside the container based on multi-source monitoring data sets, and to construct an internal resistance impedance spectrum evolution sequence that characterizes the dynamic changes of the frequency response of the battery cell internal resistance.

[0113] The relaxation mapping module is used to synchronously input the temperature field snapshot sequence and the internal resistance impedance spectrum evolution sequence into the thermo-electric relaxation coupling mapping network, extract the time lag correlation pattern of the temperature change rate feature of the local temperature field relative to the change of the internal resistance impedance spectrum, and generate a local thermal diffusion lag correlation feature map.

[0114] The spectrum generation module is used to perform time-frequency transformation on the local thermal diffusion hysteresis correlation feature map, extract the relaxation feature oscillation component excited by the irreversible thermal process, and generate the fault relaxation feature spectrum.

[0115] The matching and early warning module is used to calculate the coherence matching degree between the fault relaxation feature spectrum and the standard relaxation spectrum of each fault prototype in the fault mode library. When the matching degree meets the dynamic early warning conditions generated based on the spectrum fluctuation statistics under historical normal operation, it outputs a fault early warning signal containing the fault location and fault type.

[0116] The cooling control module is used to generate loop enhancement cooling targets and cooling parameter adjustment commands, and to drive the cooling control system.

[0117] The prediction module is used to train the parameter evolution prediction network using temperature field snapshot sequences and internal resistance impedance spectrum evolution sequences, and generate the predicted fault relaxation feature spectrum to achieve predictive early warning.

[0118] The above embodiments can be implemented, in whole or in part, by software, hardware, firmware, or any other combination thereof. When implemented in software, the above embodiments can be implemented, in whole or in part, as a computer program product. Those skilled in the art will recognize that the algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0119] In addition, the functional modules in the various embodiments of this application can be integrated into one processing module, or each module can exist physically separately, or two or more modules can be integrated into one module.

[0120] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of protection of the claims.

[0121] The above content is merely an example and illustration of the concept of the present invention. Those skilled in the art can make various modifications or additions to the specific embodiments described, or use similar methods to replace them, as long as they do not deviate from the concept of the invention or exceed the scope defined by the present invention, and all such modifications and additions should fall within the protection scope of the present invention.

Claims

1. A method for intelligent diagnosis and early warning of faults in energy storage containers, characterized in that, include: S1: Acquire distributed temperature sensing data and dynamic internal resistance parameters of battery cells within the energy storage container to generate a multi-source monitoring data set; S2: Based on multi-source monitoring data sets, a temperature field snapshot sequence characterizing the continuous spatiotemporal changes of the three-dimensional temperature distribution inside the container is constructed, as well as an internal resistance impedance spectrum evolution sequence characterizing the dynamic changes of the internal resistance frequency response of the battery cell is constructed. S3: Simultaneously input the temperature field snapshot sequence and the internal resistance impedance spectrum evolution sequence into the thermo-electric relaxation coupling mapping network to extract the time lag correlation pattern of the temperature change rate feature of the local temperature field relative to the change of the internal resistance impedance spectrum, and generate a local thermal diffusion lag correlation feature map. S4: Perform time-frequency transformation on the local thermal diffusion hysteresis correlation feature map, extract the relaxation feature oscillation component excited by the irreversible thermal process, and generate the fault relaxation feature spectrum. S5: Calculate the coherence matching degree between the fault relaxation feature spectrum and the standard relaxation spectrum of each fault prototype in the fault mode library. When the matching degree meets the dynamic early warning condition generated based on the spectrum fluctuation statistics under historical normal operation, output a fault early warning signal containing the fault location and fault type. S6: Generates loop enhancement cooling target and cooling parameter adjustment commands, and drives the cooling control system; S7: Train the parameter evolution prediction network using temperature field snapshot sequences and internal resistance spectrum evolution sequences, and generate the predicted fault relaxation feature spectrum to achieve predictive early warning.

2. The intelligent fault diagnosis and early warning method for energy storage containers according to claim 1, characterized in that, The generated multi-source monitoring data set includes: Wavelet threshold denoising is performed on the distributed temperature sensing data to obtain denoised temperature data; The dynamic internal resistance parameters of the battery cells are timestamped and missing frames are interpolated according to a unified time base to generate a regular internal resistance parameter sequence. The denoised temperature data and the regularized internal resistance parameter sequence are spliced ​​together in the fusion dimension to generate a multi-source monitoring data set.

3. The intelligent fault diagnosis and early warning method for energy storage containers according to claim 1, characterized in that, The constructed temperature field snapshot sequence includes: Distributed temperature sensing data are extracted from the multi-source monitoring data set, and the initial three-dimensional temperature field for each time frame is constructed using Kriging spatial interpolation. The initial three-dimensional temperature field was physically calibrated using a heat conduction model of the shell structure of an energy storage container to obtain a calibrated three-dimensional temperature field. The calibrated three-dimensional temperature fields are stacked in chronological order to form a sequence of temperature field snapshots describing the continuous spatiotemporal changes in temperature distribution.

4. The intelligent fault diagnosis and early warning method for energy storage containers according to claim 1, characterized in that, The constructed internal resistance spectrum evolution sequence includes: From the dynamic internal resistance parameters of the battery cells in the multi-source monitoring data set, the voltage, current and surface temperature of the battery cells are extracted, and online electrochemical impedance spectroscopy identification is performed to obtain the real-time impedance complex plane parameters. Frequency smoothing and outlier removal are applied to the real-time impedance complex plane parameters to obtain denoised impedance complex plane data. The noise-reduced impedance complex plane data is compared with the battery factory reference impedance spectrum, and the impedance magnitude offset and phase angle offset are calculated and combined into an internal resistance offset feature vector. Arrange the internal resistance offset eigenvectors in chronological order to generate an internal resistance impedance spectrum evolution sequence.

5. The intelligent fault diagnosis and early warning method for energy storage containers according to claim 1, characterized in that, The thermal-electric relaxation coupling mapping network includes a parallel temporal encoder, a cross-scale relaxation attention module, and an embedding layer. The generation of local thermal diffusion hysteresis correlation feature maps includes: A time-series encoder is used to perform time convolution encoding on the temperature field snapshot sequence and the internal resistance impedance spectrum evolution sequence to generate a temperature field encoded sequence and an internal resistance impedance spectrum encoded sequence. The time slices of each spatial local region in the temperature field coding sequence are calculated by the cross-scale relaxation attention module to the hysteresis correlation weights of each scale component in the internal impedance spectrum coding sequence, generating a relaxation hysteresis correlation matrix. This relaxation hysteresis correlation matrix characterizes the delayed response intensity of the temperature change rate relative to the change in the internal impedance spectrum. The relaxation hysteresis correlation matrix is ​​input into the embedding layer for spatial-scale dimension compression to obtain the local thermal diffusion hysteresis correlation feature map.

6. The intelligent fault diagnosis and early warning method for energy storage containers according to claim 5, characterized in that, The generated fault relaxation feature spectrum includes: A short-time Fourier transform is performed along the time dimension on the local thermal diffusion hysteresis correlation feature map to obtain the relaxation feature time-frequency map; The baseline component and high-frequency random fluctuation component caused by periodic charging and discharging are separated from the relaxation characteristic time-frequency diagram, and the coherent oscillation component within the battery thermal fault relaxation response frequency band is retained to form a filtered relaxation time-frequency representation. The filtered relaxation time-frequency representation is normalized along the frequency axis and the frequency resolution is unified to generate the fault relaxation characteristic spectrum.

7. The intelligent fault diagnosis and early warning method for energy storage containers according to claim 1, characterized in that, The output includes fault warning signals that specify the fault location and fault type, including: The correlation coefficient between the fault relaxation feature spectrum and the standard relaxation spectrum of each fault prototype in the fault mode library is calculated as the coherence matching degree. The peak matching degree is selected from the coherence matching degrees corresponding to all fault prototypes; When the peak matching degree is greater than the dynamic warning threshold obtained by statistically analyzing the spectrum fluctuations under historical normal operating conditions, the fault type and faulty battery cell location associated with the fault prototype corresponding to the peak matching degree will be output as a fault warning signal.

8. The intelligent fault diagnosis and early warning method for energy storage containers according to claim 1, characterized in that, The generation loop enhances the cooling target and cooling parameter adjustment commands, and drives the cooling control system, including: Analyze the location of the faulty battery cell in the fault warning signal, locate the thermal management loop where the faulty battery cell is located in the temperature field snapshot sequence, and generate a loop enhancement cooling target; Based on the fault type corresponding to the fault warning signal, determine the adjustment strategy for the cooling medium flow rate and generate cooling parameter adjustment instructions; The loop enhancement cooling target and cooling parameter adjustment command are sent to the cooling control system of the energy storage container to drive the corresponding loop to perform enhanced heat dissipation.

9. The intelligent fault diagnosis and early warning method for energy storage containers according to claim 1, characterized in that, The method of training a parameter evolution prediction network using temperature field snapshot sequences and internal resistance spectrum evolution sequences, and generating a predictive fault relaxation feature spectrum to achieve predictive early warning, includes: The parameter evolution prediction network is trained using the historically stored temperature field snapshot sequence and internal resistance spectrum evolution sequence to obtain the thermo-electric parameter evolution prediction model; The recently acquired temperature field snapshot sequence and internal resistance spectrum evolution sequence are input into the thermo-electric parameter evolution prediction model, and extrapolation is used to generate the predicted temperature field snapshot sequence and predicted internal resistance spectrum evolution sequence for future time periods. The predicted temperature field snapshot sequence and the predicted internal resistance spectrum evolution sequence are input into the thermo-electric relaxation coupling mapping network to obtain the predicted local thermal diffusion hysteresis correlation feature map. The predicted fault relaxation feature spectrum is obtained by performing time-frequency transformation on the predicted local thermal diffusion hysteresis correlation feature map. By utilizing the predicted fault relaxation characteristic spectrum, coherence matching is performed in advance so as to output a predictive warning signal before the fault actually occurs.

10. A system for executing the intelligent fault diagnosis and early warning method for energy storage containers according to any one of claims 1-9, characterized in that, include: The data preprocessing module is used to acquire distributed temperature sensing data and dynamic internal resistance parameters of battery cells inside the energy storage container, and generate multi-source monitoring data sets. The sequence construction module is used to construct a temperature field snapshot sequence that characterizes the continuous spatiotemporal changes of the three-dimensional temperature distribution inside the container based on multi-source monitoring data sets, and to construct an internal resistance impedance spectrum evolution sequence that characterizes the dynamic changes of the internal resistance frequency response of battery cells. The relaxation mapping module is used to synchronously input the temperature field snapshot sequence and the internal resistance impedance spectrum evolution sequence into the thermo-electric relaxation coupling mapping network, extract the time lag correlation pattern of the temperature change rate feature of the local temperature field relative to the change of the internal resistance impedance spectrum, and generate a local thermal diffusion lag correlation feature map. The spectrum generation module is used to perform time-frequency transformation on the local thermal diffusion hysteresis correlation feature map, extract the relaxation feature oscillation component excited by the irreversible thermal process, and generate the fault relaxation feature spectrum. The matching and early warning module is used to calculate the coherence matching degree between the fault relaxation feature spectrum and the standard relaxation spectrum of each fault prototype in the fault mode library. When the matching degree meets the dynamic early warning condition generated based on the spectrum fluctuation statistics under historical normal operation, it outputs a fault early warning signal containing the fault location and fault type. The cooling control module is used to generate loop-enhanced cooling targets and cooling parameter adjustment commands, and to drive the cooling control system. The prediction module is used to train the parameter evolution prediction network using temperature field snapshot sequences and internal resistance impedance spectrum evolution sequences, and generate the predicted fault relaxation feature spectrum to achieve predictive early warning.

Images