Battery thermal runaway early warning method and system based on multi-source data fusion

Abstract

The application discloses a battery thermal runaway early warning method and system based on multi-source data fusion, relates to the technical field of early warning analysis, and comprises the following steps: updating current total energy integral data according to waveform integral area, generating dynamic forgetting factor data according to macroscopic charging and discharging current sequence and tab surface temperature sequence, performing attenuation calculation on the current total energy integral data by using the dynamic forgetting factor data, and calculating a time dispersion variation coefficient according to a time interval from a previous wave peak in a cyclic memory queue; when the current total energy integral data is greater than a first preset threshold value and the time dispersion variation coefficient is less than a second preset threshold value, temperature gradient data in a corresponding time period of the tab surface temperature sequence is calculated; when the temperature gradient data is greater than a third preset threshold value and a half-height width is greater than a preset time threshold value, a battery thermal runaway early warning instruction is output. The application has the effect of improving the early warning efficiency of battery thermal runaway.

Description

Technical Field

[0001] This application relates to the field of early warning analysis technology, and in particular to a method and system for early warning of battery thermal runaway based on multi-source data fusion. Background Technology

[0002] Under long-term charging and discharging and harsh operating conditions, the welding interface between the cell tabs and the winding core of the power battery is subjected to repeated current impacts and thermomechanical stresses, which can easily lead to fatigue deterioration and internal microcracks. This is a major hidden danger that can cause abnormal heating inside the battery and even thermal runaway.

[0003] In related technologies, existing monitoring and early warning mechanisms reveal several deep-seated defects when facing the complex operating conditions of battery systems. On the one hand, even if attempts are made to increase the sampling frequency of voltage monitoring, the high-frequency switching action of the inverter inside the battery system and the electromagnetic coupling of the main circuit will generate strong ripples and background noise. Traditional methods based on fixed trigger thresholds or single-dimensional peak determination are completely unable to successfully separate the real electrode microcrack discharge signal from the complex interference pulses, which easily leads to the early warning link being frequently triggered by external environmental noise, resulting in a large number of false alarms. On the other hand, traditional technologies fail to establish a dynamic correlation between extremely small transient electrical events and the macroscopic physical environment of the battery, treating current disturbances, temperature states, and impedance changes in isolation. They ignore the structural expansion and stress relaxation characteristics of the electrode contact interface under different current loads and thermal states, and therefore cannot track and analyze the time-discrete clustering trend of high-frequency pulses. This prevents the system from forming a closed-loop monitoring system with multi-source cross-verification, seriously weakening the credibility and reliability of early safety warnings. Summary of the Invention

[0004] To address the shortcomings of existing technologies, this application provides a method and system for early warning of battery thermal runaway based on multi-source data fusion.

[0005] In a first aspect, this application provides a method for early warning of battery thermal runaway based on multi-source data fusion, including the following steps: The high-frequency transient voltage sequence of the battery tab connection circuit is obtained by a high-frequency transient voltage sensor, the high-frequency current derivative sequence of the same circuit is obtained by a high-frequency current sensor, and the macroscopic charge and discharge current sequence and tab surface temperature sequence transmitted by the battery management unit are received. Based on the sampling timestamp, the high-frequency transient voltage sequence, high-frequency current derivative sequence, macroscopic charge-discharge current sequence, and tab surface temperature sequence are time-aligned to form multi-source aligned data under the same time reference. An adaptive cancellation signal is generated based on the high-frequency current derivative sequence. The high-frequency transient voltage sequence and the adaptive cancellation signal are then subjected to point-by-point differential processing to obtain the target transient voltage mutation sequence. The dynamic trigger threshold is calculated based on the macroscopic charging and discharging current sequence and the electrode surface temperature sequence. When the voltage amplitude of the target transient voltage change sequence exceeds the dynamic trigger threshold, the voltage peak value, half width at half maximum, waveform integral area and time interval from the previous peak of the corresponding waveform segment are extracted to generate a transient event feature vector and written into a pre-defined circular memory queue. Enter the thermo-mechanical coupling determination loop, update the current total energy integral data based on the waveform integral area, generate dynamic forgetting factor data based on the macroscopic charge and discharge current sequence and the electrode surface temperature sequence, use the dynamic forgetting factor data to perform attenuation calculation on the current total energy integral data, and calculate the time discrete variation coefficient based on the time interval from the previous peak in the circular memory queue. When the current total energy integral data is greater than the first preset threshold and the time discrete variation coefficient is less than the second preset threshold, calculate the temperature gradient data within the time period corresponding to the electrode surface temperature sequence; when the temperature gradient data is greater than the third preset threshold and the half-width at half-maximum is greater than the preset time threshold, output the early warning command for battery thermal runaway.

[0006] Secondly, this application provides a battery thermal runaway early warning system based on multi-source data fusion, including: The acquisition module is used to acquire the high-frequency transient voltage sequence of the battery tab connection circuit through a high-frequency transient voltage sensor, acquire the high-frequency current derivative sequence of the same circuit through a high-frequency current sensor, and receive the macroscopic charge and discharge current sequence and tab surface temperature sequence transmitted by the battery management unit. The alignment module is used to perform time-series alignment of high-frequency transient voltage sequences, high-frequency current derivative sequences, macroscopic charge-discharge current sequences, and tab surface temperature sequences based on sampling timestamps, forming multi-source aligned data under the same time reference; The processing module is used to generate an adaptive cancellation signal based on the high-frequency current derivative sequence, and to perform point-by-point differential processing on the high-frequency transient voltage sequence and the adaptive cancellation signal to obtain the target transient voltage mutation sequence. The extraction module is used to calculate the dynamic trigger threshold based on the macroscopic charging and discharging current sequence and the electrode surface temperature sequence. When the voltage amplitude of the target transient voltage change sequence exceeds the dynamic trigger threshold, the voltage peak value, half width at half maximum, waveform integral area and time interval from the previous peak of the corresponding waveform segment are extracted to generate a transient event feature vector and write it into a pre-ordered circular memory queue. The determination module is used to enter the thermo-mechanical coupling determination loop, update the current total energy integral data based on the waveform integral area, generate dynamic forgetting factor data based on the macroscopic charge and discharge current sequence and the electrode surface temperature sequence, use the dynamic forgetting factor data to perform attenuation calculation on the current total energy integral data, and calculate the time discrete variation coefficient based on the time interval from the previous peak in the circular memory queue. The output module is used to calculate the temperature gradient data within the time period corresponding to the electrode surface temperature sequence when the current total energy integral data is greater than the first preset threshold and the time discrete variation coefficient is less than the second preset threshold; when the temperature gradient data is greater than the third preset threshold and the half-width is greater than the preset time threshold, it outputs an early warning command for battery thermal runaway.

[0007] In summary, this application includes at least one of the following beneficial technical effects: 1. This application provides an early warning method for battery thermal runaway based on multi-source data fusion. It captures high-frequency transient voltage sequences and high-frequency current derivative sequences of the tab circuit in real time through an electrical sensor array. The extracted high-frequency sequences are then time-aligned with the macroscopic current and temperature data fed back by the battery management unit under the sampling timestamp to construct multi-source aligned data with a unified time base. The voltage waveform is then processed point by point by point using an adaptive cancellation signal generated by the high-frequency current derivative to remove external noise such as inverter switching and obtain the target transient voltage change sequence. This effectively solves the perception blind zone of traditional monitoring methods in the high-frequency band. By combining the spatiotemporal coupling of multimodal data with hardware-level denoising technology, it can accurately capture weak damage signals hidden in the tab connection under strong electromagnetic background, which significantly improves the purity and real-time performance of the system in identifying early physical defects inside the battery. 2. By using a dynamic trigger threshold generated by correlating macroscopic charge-discharge current sequences with ambient temperature data as the basis for judgment, when the high-frequency voltage peak exceeds the threshold, a transient event feature vector containing energy and time distribution is extracted and pushed into a circular memory queue, thereby initiating a thermo-mechanical coupling judgment loop. The dynamic forgetting factor, which reflects the effect of thermal stress relaxation on the tabs, is used to perform nonlinear decay calculation on the total energy integral data, and cross-validation is performed using the time discrete variation coefficient and temperature gradient data, thereby outputting an early warning command for battery thermal runaway. The above dynamic judgment logic based on the physical damage evolution law not only overcomes the high false alarm rate defect of traditional fixed threshold judgment methods under complex and variable operating conditions, enabling the judgment process to be in real time aligned with the physical degradation state of the battery tab welding interface, but also provides highly valuable deterministic early warnings during the critical window period before thermal runaway occurs through in-depth tracking of the microcrack cluster arcing trend, providing more targeted and individualized high-precision monitoring and protection for the safe operation of the power battery system throughout its entire life cycle. Attached Figure Description

[0008] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the description of the embodiments 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.

[0009] Figure 1 This is a flowchart of a method for early warning of battery thermal runaway based on multi-source data fusion, as described in this application.

[0010] Figure 2 This is an experimental data graph of the dynamic trigger threshold provided in the embodiments of this application.

[0011] Figure 3 This is an experimental data graph of the dynamic forgetting factor provided in the embodiments of this application.

[0012] Figure 4 This is a schematic diagram of a battery thermal runaway early warning system based on multi-source data fusion, according to an embodiment of this application. Detailed Implementation

[0013] The following description, in conjunction with the implementation of the present invention, 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, and all such modifications and additions should fall within the protection scope of the present invention.

[0014] Example 1 This application discloses an early warning method for battery thermal runaway based on multi-source data fusion.

[0015] Reference Figure 1-3 A method for early warning of battery thermal runaway based on multi-source data fusion includes the following steps: The high-frequency transient voltage sequence of the battery tab connection circuit is obtained by a high-frequency transient voltage sensor, the high-frequency current derivative sequence of the same circuit is obtained by a high-frequency current sensor, and the macroscopic charge and discharge current sequence and tab surface temperature sequence transmitted by the battery management unit are received. In one specific embodiment, a high-frequency transient voltage sensor is arranged at both ends of the battery tab connection circuit. The input end of the high-frequency transient voltage sensor is connected in parallel to the detection point of the tab connection circuit, and the output end is connected to a high-speed sampling unit. The high-speed sampling unit can be an analog-to-digital converter with a sampling frequency of not less than 20 mega-samples per second to form a high-frequency transient voltage sequence. Each sampled data in the high-frequency transient voltage sequence carries a sampling timestamp. The data format can be that the timestamp and voltage amplitude are stored in pairs. For example, a nanosecond-level count value can be used as the sampling timestamp, and the voltage amplitude can be recorded with a microvolt-level resolution.

[0016] Meanwhile, a high-frequency current sensor is installed on the same battery tab connection circuit. The high-frequency current sensor can be a Rogowski coil, a high-frequency current transformer, or a current detection device with microsecond-level response capability. Its output is differentiated to form a high-frequency current derivative sequence. The high-frequency current derivative sequence also includes a sampling timestamp to characterize the rate of current change within the same circuit.

[0017] The battery management unit sends the macroscopic charge / discharge current sequence and the tab surface temperature sequence to the early warning processor according to its original sampling period. The macroscopic charge / discharge current sequence comes from the battery pack main circuit current sampling unit, which typically has a sampling period of ten milliseconds to one second, recording the battery pack's operating state during charging, discharging, and at different current amplitudes. The tab surface temperature sequence comes from temperature sensors located near the tab connection area, such as thermistors attached to the area adjacent to the busbar or at the corresponding tab location on the cell casing.

[0018] Based on the sampling timestamp, the high-frequency transient voltage sequence, high-frequency current derivative sequence, macroscopic charge-discharge current sequence, and tab surface temperature sequence are time-aligned to form multi-source aligned data under the same time reference. In one specific embodiment, since the sampling frequencies of the macroscopic charge / discharge current sequence and the tab surface temperature sequence are lower than those of the high-frequency transient voltage sequence and the high-frequency current derivative sequence, the warning processor, after receiving the above data, first establishes a unified time axis according to the sampling timestamps, and then supplements the low-frequency data into the time period where the high-frequency data is located. The supplementation method can be to maintain the sampling value of the previous sampling timestamp, or to use linear interpolation of adjacent sampling points. In this embodiment, to avoid unrealistic abrupt changes in the low-frequency data within a short period, the macroscopic charge / discharge current sequence uses the method of maintaining the sampling value of the previous sampling timestamp, and the tab surface temperature sequence uses linear interpolation of adjacent sampling points. After this processing, the high-frequency transient voltage sequence, the high-frequency current derivative sequence, the macroscopic charge / discharge current sequence, and the tab surface temperature sequence form multi-source aligned data under the same time reference.

[0019] An adaptive cancellation signal is generated based on the high-frequency current derivative sequence. The high-frequency transient voltage sequence and the adaptive cancellation signal are then subjected to point-by-point differential processing to obtain the target transient voltage mutation sequence. Furthermore, an adaptive cancellation signal is generated based on the high-frequency current derivative sequence. The high-frequency transient voltage sequence and the adaptive cancellation signal are then subjected to point-by-point differential processing to obtain the target transient voltage change sequence, including: Extracting current abrupt change data fragments from high-frequency current derivative sequences; The current mutation data fragments are converted with the equivalent inductive reactance data of the battery tab connection circuit to generate the predicted induced voltage sequence. The estimated induced voltage sequence is enveloped and smoothed to generate an adaptive cancellation signal. Under the same time reference, the high-frequency transient voltage sequence and the adaptive cancellation signal are differentially processed point by point to generate the residual voltage sequence; Data segments with amplitudes greater than the baseline noise amplitude are retained from the residual voltage sequence and arranged in chronological order to form the target transient voltage mutation sequence.

[0020] In one specific embodiment, after the multi-source aligned data is formed, the early warning processor does not directly determine the peaks of the high-frequency transient voltage sequence. Instead, it first generates an adaptive cancellation signal using the high-frequency current derivative sequence. This is because the operation of the inverter, DC-DC converter, and contactor in the battery system can cause strong electromagnetic coupling, which may also manifest as spikes in the high-frequency transient voltage sequence. If only the voltage amplitude is used as the triggering basis, external switching noise can easily be mistaken for micro-crack signals inside the tab connection circuit. Therefore, the early warning processor identifies the rapidly changing current portion from the high-frequency current derivative sequence and converts this portion into a cancellation amount with dimensions consistent with the high-frequency transient voltage sequence. After the adaptive cancellation signal is generated, the early warning processor, under the same time reference, maps the high-frequency transient voltage sequence to the adaptive cancellation signal point by point. The voltage amplitude at the same sampling timestamp is subtracted from the amplitude of the adaptive cancellation signal to obtain voltage data stripped of some switching interference. This differential result, after baseline screening, is used as the target transient voltage mutation sequence.

[0021] Based on the above scheme, the generation process of the adaptive cancellation signal can be implemented in a more detailed manner. The early warning processor first extracts current mutation data segments from the high-frequency current derivative sequence. The extraction of current mutation data segments can adopt a local peak search method, that is, finding sampling points in the high-frequency current derivative sequence that exceed the background derivative amplitude, and taking the continuous data before and after the sampling point as the current mutation data segment. The background derivative amplitude can be obtained from the static calibration data after startup, or it can be calculated from the stable window before the current determination cycle. The current mutation data segments extracted in this way usually correspond to inverter switching, current ripple edges, and electromagnetic coupling events in the main circuit.

[0022] The current surge data segment has a different dimension than voltage and cannot be directly subtracted from the high-frequency transient voltage sequence. Therefore, the warning processor reads the equivalent inductive reactance data of the battery tab connection circuit and converts the current surge data segment with the equivalent inductive reactance data to generate an estimated induced voltage sequence. The equivalent inductive reactance data can be written into the non-volatile memory of the warning processor during the battery pack's factory testing phase, or it can be obtained from a standard impulse response test during the vehicle's power-on self-test. The converted estimated induced voltage sequence has the same dimension as the high-frequency transient voltage sequence, and each sampling point still maintains the same sampling timestamp as the high-frequency current derivative sequence.

[0023] To avoid introducing new ringing during cancellation due to sharp edges in the estimated induced voltage sequence, the warning processor performs envelope extraction and smoothing on the estimated induced voltage sequence. Envelope extraction can be achieved by taking local absolute peak values ​​to form an envelope line, and smoothing can be performed using moving average, median filtering, or finite impulse response filtering. In this embodiment, moving average processing is used, with the window length set to five to twenty-one sampling points based on the sampling rate of the high-frequency transient voltage sequence, ensuring that envelope changes can follow the amplitude of switching noise without excessive lag. An adaptive cancellation signal is formed after envelope extraction and smoothing. The warning processor then performs point-by-point differential processing between the high-frequency transient voltage sequence and the adaptive cancellation signal at the same time reference to generate a residual voltage sequence. For data points where the sampling timestamps do not completely overlap, the adaptive cancellation signal is first time-interpolated before differential processing is performed. The residual voltage sequence may still contain small fluctuations near the baseline. The warning processor retains data segments from the residual voltage sequence with amplitudes greater than the baseline noise amplitude and arranges them in chronological order to form the target transient voltage mutation sequence. The baseline noise amplitude is derived from historical stable data from the same detection channel, so this screening will not introduce judgment criteria that contradict the subsequent dynamic trigger threshold.

[0024] The dynamic trigger threshold is calculated based on the macroscopic charging and discharging current sequence and the electrode surface temperature sequence. When the voltage amplitude of the target transient voltage change sequence exceeds the dynamic trigger threshold, the voltage peak value, half width at half maximum, waveform integral area and time interval from the previous peak of the corresponding waveform segment are extracted to generate a transient event feature vector and written into a pre-defined circular memory queue. Furthermore, the dynamic trigger threshold is calculated based on the macroscopic charge / discharge current sequence and the tab surface temperature sequence, including: Extract the noise observation window up to the current time from the target transient voltage change sequence, and calculate the baseline mean data and baseline discrete data within the noise observation window; Extract the macroscopic charge-discharge current sequence and tab surface temperature sequence corresponding to the noise observation window from the multi-source aligned data; Threshold correction data are generated based on the macroscopic charge-discharge current sequence and the tab surface temperature sequence. The baseline mean data, baseline discrete data, and threshold correction data are combined to obtain the dynamic trigger threshold.

[0025] In one specific embodiment, the target transient voltage mutation sequence is the effective voltage mutation data retained after being constrained by noise background. The warning processor extracts a noise observation window before the current processing time of the target transient voltage mutation sequence; for example, it extracts a historical window with a length of one to five milliseconds. The baseline mean data and baseline discrete data within this historical window are calculated. The baseline mean data is used to represent the zero-point offset of the current detection channel, and the baseline discrete data is used to represent the background noise amplitude.

[0026] Subsequently, the early warning processor extracts the macroscopic charge / discharge current sequence and the tab surface temperature sequence corresponding to the same noise observation window from the multi-source aligned data. The macroscopic charge / discharge current sequence reflects the current load of the battery tab connection circuit, and the tab surface temperature sequence reflects the thermal state of the tab connection area. Under high current and high temperature conditions, the background fluctuation of the detection channel usually increases. Therefore, in this embodiment, threshold correction data is generated based on the macroscopic charge / discharge current sequence and the tab surface temperature sequence. Then, the baseline mean data, baseline discrete data, and threshold correction data are synthesized to obtain the dynamic trigger threshold. The dynamic trigger threshold obtained by the above operation changes with the operating conditions, and will not be overly sensitive under high noise conditions, nor will it lose the ability to identify small changes under stable operating conditions.

[0027] Specifically, the larger the macroscopic charge / discharge current sequence, the stronger the loop ripple and electromagnetic coupling effects; the higher the tab surface temperature sequence, the greater the potential increase in detection channel thermal noise and contact thermal drift. Therefore, the warning processor generates threshold correction data based on the macroscopic charge / discharge current sequence and the tab surface temperature sequence. The threshold correction data can be obtained by looking up tables; for example, the macroscopic charge / discharge current sequence can be divided into low load, medium load, and high load, and the tab surface temperature sequence into normal temperature, rising temperature, and high temperature zones, with the corresponding correction values ​​read from a two-dimensional table. Finally, the warning processor synthesizes the baseline mean data, baseline discrete data, and threshold correction data to obtain the dynamic trigger threshold. The synthesis can be achieved by adding a multiple of the baseline discrete data to the baseline mean data, and then superimposing the threshold correction data. The multiple of the baseline discrete data can be determined during the factory calibration stage. This dynamic trigger threshold is updated in real-time with the noise observation window and is used to determine the peak trigger of the target transient voltage change sequence.

[0028] Furthermore, the transient event feature vector is generated, including: When the voltage amplitude of the target transient voltage change sequence exceeds the dynamic trigger threshold, the voltage waveform segment is truncated with the sampling point exceeding the dynamic trigger threshold as the center; Extract the maximum amplitude value from the voltage waveform segment as the voltage peak value; Extract the half-width at half-maximum (WHM) based on the half-amplitude position corresponding to the voltage peak. The integral area of ​​the waveform is obtained by accumulating the area of ​​the portion of the voltage waveform segment that is above the baseline mean. Calculate the difference between the current peak timestamp and the previous peak timestamp to obtain the time interval from the previous peak; The voltage peak value, half-width at half-maximum (FWHM), waveform integral area, and time interval from the previous peak are written in a fixed bit order to form a transient event feature vector.

[0029] In one specific embodiment, when a voltage amplitude in a target transient voltage mutation sequence exceeds a dynamic trigger threshold, the warning processor extracts a voltage waveform segment centered on the sampling point exceeding the dynamic trigger threshold. The extraction range can be determined according to the sampling frequency of the high-speed sampling module, for example, extracting twenty to one hundred sampling points forward and twenty to two hundred sampling points backward, so that the voltage waveform segment can cover the rising edge of the spike, the peak region, and the falling back region. The warning processor extracts the maximum amplitude value from the voltage waveform segment as the voltage peak value, and determines the half-width at half-maximum (WHM) based on the left and right intersection positions corresponding to half of the voltage peak value. For cases where the sampling point cannot fall exactly at the half-amplitude position, linear interpolation can be performed between adjacent sampling points to obtain a more stable WHM.

[0030] Subsequently, the early warning processor uses the baseline mean data as the integration benchmark and accumulates the area of ​​the portion of the voltage waveform segment that is higher than the baseline mean data to obtain the waveform integral area. If the event is not the first recorded event, the early warning processor further calculates the difference between the current peak timestamp and the previous peak timestamp to obtain the time interval from the previous peak; if the event is the first event after startup, the time interval from the previous peak can be set as a null flag, and this null flag is skipped during the calculation in the circular memory queue. The voltage peak value, half-width at half-maximum (WHM), waveform integral area, and time interval from the previous peak are written into the same transient event feature vector in a fixed bit order.

[0031] After the transient event feature vector is generated, it is written to a circular memory queue of a predetermined number of bits. The predetermined number of bits can be set according to the storage capacity of the warning processor and the response requirements of the battery system, for example, 256 bits, 512 bits, or 1024 bits. The circular memory queue saves data in the order of the generation time of the transient event feature vectors. When the number of writes has not reached the predetermined number of bits, new transient event feature vectors are appended sequentially; when the number of writes has reached the predetermined number of bits, new transient event feature vectors overwrite the oldest written transient event feature vector. The circular memory queue is maintained by a read / write pointer by the warning processor. The read / write pointer moves one bit after each write, ensuring that the queue always stores event information related to the abnormality of the electrode connection circuit within the most recent period. This circular memory queue subsequently provides both the waveform integral area for energy accumulation and the time interval from the previous peak for calculating the coefficient of variation of time discreteness.

[0032] Enter the thermo-mechanical coupling determination loop, update the current total energy integral data based on the waveform integral area, generate dynamic forgetting factor data based on the macroscopic charge and discharge current sequence and the electrode surface temperature sequence, use the dynamic forgetting factor data to perform attenuation calculation on the current total energy integral data, and calculate the time discrete variation coefficient based on the time interval from the previous peak in the circular memory queue. Furthermore, dynamic forgetting factor data is generated based on the macroscopic charge-discharge current sequence and the tab surface temperature sequence, including: Extract the macroscopic charge / discharge current sequence and the tab surface temperature sequence within the current determination cycle from multi-source aligned data; The current load level data of the tab connection circuit is determined based on the macroscopic charging and discharging current sequence; The thermal state level data of the electrode connection area is determined based on the electrode surface temperature sequence; By performing a two-dimensional mapping between current load level data and thermal state level data, contact surface stress relaxation level data can be obtained. Dynamic forgetting factor data is generated based on the stress relaxation level data of the contact surface.

[0033] In one specific embodiment, the thermo-mechanical coupling decision loop is executed within each decision loop cycle. The decision loop cycle can be set according to the battery system's computing power and real-time warning requirements, for example, to one second, five seconds, or ten seconds. Within a decision loop cycle, the warning processor extracts the feature vector of newly generated transient events in that cycle and updates the current total energy integral data based on the waveform integral area within it. The current total energy integral data is not equivalent to the instantaneous energy of a single spike, but rather is used to characterize the degree of accumulation of abnormal abrupt changes in the tab connection circuit over a period of time. Simultaneously, the warning processor extracts the macroscopic charge-discharge current sequence and tab surface temperature sequence within that decision loop cycle from the multi-source aligned data to generate dynamic forgetting factor data. The dynamic forgetting factor data reflects the impact of contact surface stress relaxation on abnormal accumulation in the tab connection region under different current loads and thermal states. Subsequently, the warning processor uses the dynamic forgetting factor data to perform attenuation calculations on the current total energy integral data, avoiding false alarms caused by occasional isolated spikes remaining in the energy memory space for extended periods.

[0034] Specifically, in the process of generating dynamic forgetting factor data, the early warning processor uses the current judgment cycle as the calculation range and extracts the macroscopic charge-discharge current sequence and the tab surface temperature sequence from the multi-source aligned data. The macroscopic charge-discharge current sequence can be the average value within the judgment cycle or the effective value; in this embodiment, the effective value is used when the battery is in a pulse charge-discharge condition, and the average value is used when the current change is relatively gradual. The early warning processor determines the current load level data of the tab connection circuit based on the macroscopic charge-discharge current sequence. The current load level data can be divided into several levels, such as zero load, low load, medium load, and high load, and the boundaries of each level are determined by the rated capacity of the cell and the allowable charge-discharge rate. The tab surface temperature sequence is used to determine the thermal state level data of the tab connection area, and the thermal state level data can be divided into low temperature zone, normal temperature zone, rising temperature zone, and high temperature zone according to the allowable temperature range of the cell material.

[0035] Current load level data and thermal state level data are fed into a two-dimensional mapping table to obtain contact surface stress relaxation level data. The two-dimensional mapping table can be determined during the experimental phase from the high-frequency response statistical results of battery samples under different current loads and temperature conditions. For example, under a combination of high load and low temperature, the tab connection area is more prone to contact instability; the contact surface stress relaxation level data can be set to a lower level, resulting in a smaller dynamic forgetting factor. Under a combination of zero load and rising temperature, the tab connection area may experience transient adhesion due to thermal expansion; the contact surface stress relaxation level data can be set to a higher level, resulting in a larger dynamic forgetting factor. The warning processor generates dynamic forgetting factor data based on the contact surface stress relaxation level data and provides it to the decay calculation of the current total energy integral data. Because the dynamic forgetting factor data is constrained by both the macroscopic charge / discharge current sequence and the tab surface temperature sequence, the accumulated amount in the energy memory space does not decay simply at a fixed rate, but rather matches the actual operating conditions of the tab connection area.

[0036] Furthermore, updates to the current total energy integral data include: Extract the waveform integral area from the feature vector of the latest transient event, and normalize the waveform integral area according to the voltage peak value to obtain the energy data of a single event; The energy data of a single event is added to the total energy integral data output from the previous judgment cycle to obtain the energy integral data to be decayed. Based on the dynamic forgetting factor data, the energy integral data to be decayed is periodically decayed to obtain the current total energy integral data; Save the current total energy integral data as the total energy integral data output by the previous judgment cycle in the next judgment cycle.

[0037] In one specific embodiment, the update of the current total energy integral data is closely linked to the transient event feature vector. The early warning processor extracts the waveform integral area from the latest transient event feature vector and normalizes the waveform integral area based on the voltage peak value in the same transient event feature vector to obtain the single event energy data. The purpose of the normalization is to avoid the incomparability of waveform integral areas caused by differences in the gain of different detection channels. In practice, the early warning processor can divide the waveform integral area by the product of the reference voltage amplitude and the reference time width, and then correct it by combining the ratio of the voltage peak value to the reference voltage amplitude to obtain dimensionless single event energy data. The reference voltage amplitude and reference time width can be written into the early warning processor during the system calibration stage.

[0038] After the energy data of a single event is generated, the early warning processor adds it to the total energy integral data output from the previous judgment cycle to obtain the energy integral data to be attenuated. If multiple transient event feature vectors exist within the current judgment cycle, multiple single event energy data are generated separately, and then the sum of these multiple single event energy data is added to the total energy integral data output from the previous judgment cycle. If no new transient event feature vectors exist within the current judgment cycle, the energy integral data to be attenuated is equal to the total energy integral data output from the previous judgment cycle.

[0039] Subsequently, the early warning processor performs periodic decay on the energy integral data to be decayed based on the dynamic forgetting factor data, obtaining the current total energy integral data. Periodic decay can be performed exponentially or using a decay ratio obtained from a lookup table; this embodiment uses exponential decay, and the decay intensity is determined jointly by the dynamic forgetting factor data and the judgment cycle. The current total energy integral data is saved at the end of the current judgment cycle and used as the total energy integral data output from the previous judgment cycle for the next judgment cycle.

[0040] Furthermore, the calculation of the time discrete coefficient of variation includes: Extract time intervals from the previous peak from a predetermined number of positions from the circular memory queue to form a time interval sequence; The time interval sequence is divided into the earlier time interval sequence and the more recent time interval sequence according to the time index; Calculate the average interval data of the preceding time interval sequence and the recent time interval sequence respectively, and calculate the dispersion data of the recent time interval sequence; The time discrete variation coefficient is generated based on the average interval data and the dispersion data of the recent time interval sequence.

[0041] In one specific embodiment, the early warning processor extracts a predetermined number of time intervals from the previous peak from the circular memory queue to form a time interval sequence. If the circular memory queue is not yet full, the early warning processor only extracts data that has been effectively written; if there is a null flag corresponding to the first event, that null flag is not included in the calculation. The time interval sequence is divided into an earlier time interval sequence and a more recent time interval sequence according to the time index. For example, when the number of available time intervals from the previous peak is one hundred, the first fifty can be classified as the earlier time interval sequence, and the last fifty as the more recent time interval sequence. For the startup phase with a small amount of data, the time discrete variation coefficient can be output after the minimum number of calculations is reached.

[0042] The early warning processor calculates the average interval data of the earlier time interval sequence and the recent time interval sequence, and also calculates the dispersion data of the recent time interval sequence. The average interval data of the earlier time interval sequence is used to represent the average interval of the earlier stage of abnormal peaks, and the average interval data of the recent time interval sequence is used to represent the average interval of the recent abnormal peaks. The dispersion data of the recent time interval sequence can be expressed using standard deviation.

[0043] Subsequently, the early warning processor generates a time discrete variation coefficient based on the average interval data and dispersion data of the recent time interval sequence. This involves normalizing the dispersion data relative to the average interval data of the recent time interval sequence to obtain an index used to judge the stability of event intervals. The smaller the time discrete variation coefficient, the more concentrated the time intervals of recent abnormal peaks. In this embodiment, the time discrete variation coefficient is fed into a thermo-mechanical coupling decision loop, where it, along with the current total energy integral data, determines whether to proceed to temperature gradient data calculation. In this way, the system will only further trigger thermal state verification when the abnormal energy reaches a certain level and the abnormal events exhibit clustered characteristics in time, reducing misjudgments caused by occasional electromagnetic interference.

[0044] When the current total energy integral data is greater than the first preset threshold and the time discrete variation coefficient is less than the second preset threshold, calculate the temperature gradient data within the time period corresponding to the electrode surface temperature sequence; when the temperature gradient data is greater than the third preset threshold and the half-width at half-maximum is greater than the preset time threshold, output the early warning command for battery thermal runaway.

[0045] In a specific embodiment, within the same thermomechanical coupling determination loop, the early warning processor also reads the time interval from the previous peak from the circular memory queue and calculates the time discrete variation coefficient. The time discrete variation coefficient is used to characterize whether the interval of recent abnormal peaks has changed from a discrete and sporadic state to a relatively dense cluster state. When the microcrack in the electrode connection circuit is in its initial stage, the target transient voltage mutation sequence may only have sporadic peaks, with a large difference in the time interval from the previous peak; as the contact interface deteriorates, the time interval of abnormal peaks tends to narrow, and the time discrete variation coefficient decreases. When the current total energy integral data is greater than a first preset threshold and the time discrete variation coefficient is less than a second preset threshold, the early warning processor considers that the high-frequency data has met the conditions for entering the thermal state verification, and then calculates the temperature gradient data within the corresponding time period of the electrode surface temperature sequence. The temperature gradient data can be obtained by dividing the difference between the end temperature and the start temperature of the electrode surface temperature sequence within the corresponding time period by the time length, or it can be represented by a multi-point fitting slope. In this embodiment, a multi-point fitting slope is used to reduce the impact of jitter at a single temperature sampling point. When the temperature gradient data exceeds a third preset threshold, and the half-width at half-maximum (WHM) of the feature vector of the most recent transient event is greater than a preset time threshold, the warning processor outputs an early warning command for battery thermal runaway. A WHM greater than the preset time threshold indicates that the voltage surge is not an isolated, extremely narrow disturbance, but rather more consistent with the waveform of micro-arc discharge or contact instability occurring in the tab connection area.

[0046] Furthermore, before outputting the early warning command for battery thermal runaway, a low-frequency state verification step is also included: Receive the cell voltage sequence and macroscopic charge / discharge current sequence from the battery management unit; Calculate DC internal resistance variation data based on individual cell voltage sequences and macroscopic charge / discharge current sequences; By correlating the DC internal resistance change data with the temperature gradient data within the same judgment cycle, low-frequency state verification data is obtained. When the low-frequency state verification data meets the preset verification conditions, the early warning command for battery thermal runaway is marked as a valid command. When the low-frequency state verification data does not meet the preset verification conditions, the current total energy integral data and the circular memory queue are retained, and the next judgment cycle is entered.

[0047] In one specific embodiment, a low-frequency state verification step may be introduced before outputting the early warning command for battery thermal runaway. In addition to sending the macroscopic charge / discharge current sequence and the tab surface temperature sequence, the battery management unit also sends a single-cell voltage sequence to the warning processor. The single-cell voltage sequence is the routine sampling result of the battery management unit on the terminal voltage of a single cell, with a sampling period of 100 milliseconds to 1 second. After receiving the single-cell voltage sequence and the macroscopic charge / discharge current sequence, the warning processor extracts data according to the same judgment cycle and calculates the DC internal resistance change data. During the calculation, within the time period when a usable current step occurs in the macroscopic charge / discharge current sequence, the ratio of the voltage change in the single-cell voltage sequence to the current change in the macroscopic charge / discharge current sequence is taken to obtain the estimated DC internal resistance value for that time period. This estimated value is then compared with the historical baseline DC internal resistance estimate to form the DC internal resistance change data. If there is insufficient current change within the judgment cycle, the warning processor can use the most recent valid DC internal resistance change data and mark its time validity.

[0048] DC internal resistance change data does not determine the early warning output independently. Instead, it is correlated with temperature gradient data within the same judgment cycle to obtain low-frequency state verification data. The correlation method can employ a joint scoring approach. For example, an internal resistance anomaly flag is given when the DC internal resistance change data exceeds its historical fluctuation range, and a temperature anomaly flag is given when the temperature gradient data exceeds a third preset threshold. When both flags are met simultaneously within the same judgment cycle, the low-frequency state verification data meets the preset verification conditions. More granular correlation weights can also be set based on battery type. However, regardless of the method used, both the DC internal resistance change data and temperature gradient data originate from the same judgment cycle to avoid time mismatch between high-frequency anomalies and low-frequency changes. When the low-frequency state verification data meets the preset verification conditions, the early warning processor marks the early warning command for battery thermal runaway as a valid command. If the low-frequency state verification data does not meet the preset verification conditions, the early warning processor does not clear the already formed current total energy integral data and the cyclic memory queue, but retains them and proceeds to the next judgment cycle. This processing method ensures that the system does not lose high-frequency anomaly memory due to a single low-frequency sampling delay, while also avoiding directly outputting a strong warning based solely on high-frequency spikes.

[0049] Furthermore, the output of the early warning command for battery thermal runaway includes: When the current total energy integral data is greater than the first preset threshold, the time discrete variation coefficient is less than the second preset threshold, and the low frequency state verification data meets the preset verification conditions, the first-level warning data is generated and sent to the battery management unit. When the current total energy integral data is greater than the energy threshold above the first preset threshold, the half-width is greater than the preset time threshold, and the temperature gradient data is greater than the third preset threshold, a second-level warning data is generated. Based on the second-level warning data, an early warning command for battery thermal runaway is generated and sent to the vehicle control unit, causing the vehicle control unit to perform current limiting and contactor disconnection control.

[0050] In one specific embodiment, the output of the early warning command for battery thermal runaway can be performed in a hierarchical manner. The warning processor generates first-level warning data and sends it to the battery management unit when the current total energy integral data is greater than a first preset threshold, the time discrete variation coefficient is less than a second preset threshold, and the low-frequency state verification data meets preset verification conditions. The first-level warning data may include the cell number, the cycle determination timestamp, the current total energy integral data, the time discrete variation coefficient, the temperature gradient data, and the low-frequency state verification data. After receiving the first-level warning data, the battery management unit can record fault codes, increase the sampling frequency of relevant channels, and appropriately limit the charging and discharging current according to the vehicle control strategy.

[0051] When the current total energy integral data exceeds the energy threshold above the first preset threshold, the half-width at half-maximum (HWHM) exceeds the preset time threshold, and the temperature gradient data exceeds the third preset threshold, the warning processor generates second-level warning data. The energy threshold exceeding the first preset threshold distinguishes between general early anomalies and anomalies closer to the failure boundary. The HWHM exceeding the preset time threshold indicates that the duration of the abnormal waveform in the target transient voltage mutation sequence has reached the characteristic range of micro-arc discharge or contact arcing. The temperature gradient data exceeding the third preset threshold indicates that a high-frequency anomaly has been accompanied by an observable thermal response. After the second-level warning data is generated, the warning processor generates an early warning command for battery thermal runaway based on the second-level warning data and sends the early warning command for battery thermal runaway to the vehicle control unit. Upon receiving the command, the vehicle control unit executes current limiting and contactor disconnection control, for example, first reducing the allowable charging and discharging current to a safe limit, and then disconnecting the high-voltage contactor according to the vehicle safety strategy. If the system is in a charging state, the vehicle control unit can also send a stop-charging request to the charging equipment to allow the battery tab connection circuit to exit the high-current state as quickly as possible.

[0052] In practical applications, the early warning processor can be integrated into the battery management unit (BMU) or connected to the BMU via a communication bus as an independent high-frequency early warning slave controller. Sampling data from the high-frequency transient voltage sensor and high-frequency current sensor are preprocessed in a local cache, while the macroscopic charge / discharge current sequence, tab surface temperature sequence, and cell voltage sequence are transmitted through the BMU's communication interface. To ensure time alignment accuracy, the high-speed sampling unit and the BMU can share the same system clock, or a unified timer-generated sampling timestamp can be included in the communication message. If there is a deviation between the high-frequency sampling clock and the BMU clock, the early warning processor can perform a time deviation correction at the beginning of each decision cycle, remapping the high-frequency transient voltage sequence, high-frequency current derivative sequence, macroscopic charge / discharge current sequence, and tab surface temperature sequence to the same time reference. This ensures a continuous data relationship between subsequent adaptive cancellation signals, dynamic trigger thresholds, current total energy integral data, time discrete variation coefficients, and low-frequency state verification data. This implementation method allows high-frequency electrical anomalies, macroscopic current loads, tab thermal response, and low-frequency internal resistance changes to be closed within the same processing link, enabling earlier identification of the process by which a microcrack in the tab connection circuit develops into a thermal runaway initiation.

[0053] Example 2 This application also discloses an early warning system for battery thermal runaway based on multi-source data fusion.

[0054] Reference Figure 4 A battery thermal runaway early warning system based on multi-source data fusion includes: The acquisition module is used to acquire the high-frequency transient voltage sequence of the battery tab connection circuit through a high-frequency transient voltage sensor, acquire the high-frequency current derivative sequence of the same circuit through a high-frequency current sensor, and receive the macroscopic charge and discharge current sequence and tab surface temperature sequence transmitted by the battery management unit. The alignment module is used to perform time-series alignment of high-frequency transient voltage sequences, high-frequency current derivative sequences, macroscopic charge-discharge current sequences, and tab surface temperature sequences based on sampling timestamps, forming multi-source aligned data under the same time reference; The processing module is used to generate an adaptive cancellation signal based on the high-frequency current derivative sequence, and to perform point-by-point differential processing on the high-frequency transient voltage sequence and the adaptive cancellation signal to obtain the target transient voltage mutation sequence. The extraction module is used to calculate the dynamic trigger threshold based on the macroscopic charging and discharging current sequence and the electrode surface temperature sequence. When the voltage amplitude of the target transient voltage change sequence exceeds the dynamic trigger threshold, the voltage peak value, half width at half maximum, waveform integral area and time interval from the previous peak of the corresponding waveform segment are extracted to generate a transient event feature vector and write it into a pre-ordered circular memory queue. The determination module is used to enter the thermo-mechanical coupling determination loop, update the current total energy integral data based on the waveform integral area, generate dynamic forgetting factor data based on the macroscopic charge and discharge current sequence and the electrode surface temperature sequence, use the dynamic forgetting factor data to perform attenuation calculation on the current total energy integral data, and calculate the time discrete variation coefficient based on the time interval from the previous peak in the circular memory queue. The output module is used to calculate the temperature gradient data within the time period corresponding to the electrode surface temperature sequence when the current total energy integral data is greater than the first preset threshold and the time discrete variation coefficient is less than the second preset threshold; when the temperature gradient data is greater than the third preset threshold and the half-width is greater than the preset time threshold, it outputs an early warning command for battery thermal runaway.

[0055] 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, they should all fall within the protection scope of the present invention.

[0056] In the description of this specification, references to terms such as "an embodiment," "example," "specific example," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of the invention. In this specification, illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples.

[0057] The preferred embodiments of the present invention disclosed above are merely illustrative of the invention. These preferred embodiments do not exhaustively describe all details, nor do they limit the invention to any specific implementation. Clearly, many modifications and variations can be made based on the content of this specification. This specification selects and specifically describes these embodiments to better explain the principles and practical applications of the invention, thereby enabling those skilled in the art to better understand and utilize the invention.

Claims

1. A method for early warning of battery thermal runaway based on multi-source data fusion, characterized in that, Includes the following steps: The high-frequency transient voltage sequence of the battery tab connection circuit is obtained by a high-frequency transient voltage sensor, the high-frequency current derivative sequence of the same circuit is obtained by a high-frequency current sensor, and the macroscopic charge and discharge current sequence and tab surface temperature sequence transmitted by the battery management unit are received. Based on the sampling timestamp, the high-frequency transient voltage sequence, high-frequency current derivative sequence, macroscopic charge-discharge current sequence, and tab surface temperature sequence are time-aligned to form multi-source aligned data under the same time reference. An adaptive cancellation signal is generated based on the high-frequency current derivative sequence. The high-frequency transient voltage sequence and the adaptive cancellation signal are then subjected to point-by-point differential processing to obtain the target transient voltage mutation sequence. The dynamic trigger threshold is calculated based on the macroscopic charging and discharging current sequence and the electrode surface temperature sequence. When the voltage amplitude of the target transient voltage change sequence exceeds the dynamic trigger threshold, the voltage peak value, half width at half maximum, waveform integral area and time interval from the previous peak of the corresponding waveform segment are extracted to generate a transient event feature vector and written into a pre-defined circular memory queue. Enter the thermo-mechanical coupling determination loop, update the current total energy integral data based on the waveform integral area, generate dynamic forgetting factor data based on the macroscopic charge and discharge current sequence and the electrode surface temperature sequence, use the dynamic forgetting factor data to perform attenuation calculation on the current total energy integral data, and calculate the time discrete variation coefficient based on the time interval from the previous peak in the circular memory queue. When the current total energy integral data is greater than the first preset threshold and the time discrete variation coefficient is less than the second preset threshold, calculate the temperature gradient data within the time period corresponding to the electrode surface temperature sequence; when the temperature gradient data is greater than the third preset threshold and the half-width at half-maximum is greater than the preset time threshold, output the early warning command for battery thermal runaway.

2. The method for early warning of battery thermal runaway based on multi-source data fusion according to claim 1, characterized in that, An adaptive cancellation signal is generated based on the high-frequency current derivative sequence. The high-frequency transient voltage sequence and the adaptive cancellation signal are then subjected to point-by-point differential processing to obtain the target transient voltage mutation sequence, including: Extracting current abrupt change data fragments from high-frequency current derivative sequences; The current mutation data fragments are converted with the equivalent inductive reactance data of the battery tab connection circuit to generate the predicted induced voltage sequence. The estimated induced voltage sequence is enveloped and smoothed to generate an adaptive cancellation signal. Under the same time reference, the high-frequency transient voltage sequence and the adaptive cancellation signal are differentially processed point by point to generate the residual voltage sequence; Data segments with amplitudes greater than the baseline noise amplitude are retained from the residual voltage sequence and arranged in chronological order to form the target transient voltage mutation sequence.

3. The method for early warning of battery thermal runaway based on multi-source data fusion according to claim 2, characterized in that, The dynamic trigger threshold is calculated based on the macroscopic charge / discharge current sequence and the tab surface temperature sequence, including: Extract the noise observation window up to the current time from the target transient voltage change sequence, and calculate the baseline mean data and baseline discrete data within the noise observation window; Extract the macroscopic charge-discharge current sequence and tab surface temperature sequence corresponding to the noise observation window from the multi-source aligned data; Threshold correction data are generated based on the macroscopic charge-discharge current sequence and the tab surface temperature sequence. The baseline mean data, baseline discrete data, and threshold correction data are combined to obtain the dynamic trigger threshold.

4. The method for early warning of battery thermal runaway based on multi-source data fusion according to claim 3, characterized in that, Generate transient event feature vectors, including: When the voltage amplitude of the target transient voltage change sequence exceeds the dynamic trigger threshold, the voltage waveform segment is truncated with the sampling point exceeding the dynamic trigger threshold as the center; Extract the maximum amplitude value from the voltage waveform segment as the voltage peak value; Extract the half-width at half-maximum (WHM) based on the half-amplitude position corresponding to the voltage peak. The integral area of ​​the waveform is obtained by accumulating the area of ​​the portion of the voltage waveform segment that is above the baseline mean. Calculate the difference between the current peak timestamp and the previous peak timestamp to obtain the time interval from the previous peak; The voltage peak value, half-width at half-maximum (FWHM), waveform integral area, and time interval from the previous peak are written in a fixed bit order to form a transient event feature vector.

5. The method for early warning of battery thermal runaway based on multi-source data fusion according to claim 1, characterized in that, Dynamic forgetting factor data is generated based on macroscopic charge / discharge current sequences and tab surface temperature sequences, including: Extract the macroscopic charge / discharge current sequence and the tab surface temperature sequence within the current determination cycle from multi-source aligned data; The current load level data of the tab connection circuit is determined based on the macroscopic charging and discharging current sequence; The thermal state level data of the electrode connection area is determined based on the electrode surface temperature sequence; By performing a two-dimensional mapping between current load level data and thermal state level data, contact surface stress relaxation level data can be obtained. Dynamic forgetting factor data is generated based on the stress relaxation level data of the contact surface.

6. The method for early warning of battery thermal runaway based on multi-source data fusion according to claim 4, characterized in that, The current update to the total energy integral data includes: Extract the waveform integral area from the feature vector of the latest transient event, and normalize the waveform integral area according to the voltage peak value to obtain the energy data of a single event; The energy data of a single event is added to the total energy integral data output from the previous judgment cycle to obtain the energy integral data to be decayed. Based on the dynamic forgetting factor data, the energy integral data to be decayed is periodically decayed to obtain the current total energy integral data; Save the current total energy integral data as the total energy integral data output by the previous judgment cycle in the next judgment cycle.

7. The method for early warning of battery thermal runaway based on multi-source data fusion according to claim 1, characterized in that, The calculation of the time discrete coefficient of variation includes: Extract time intervals from the previous peak from a predetermined number of positions from the circular memory queue to form a time interval sequence; The time interval sequence is divided into the earlier time interval sequence and the more recent time interval sequence according to the time index; Calculate the average interval data of the preceding time interval sequence and the recent time interval sequence respectively, and calculate the dispersion data of the recent time interval sequence; The time discrete variation coefficient is generated based on the average interval data and the dispersion data of the recent time interval sequence.

8. The method for early warning of battery thermal runaway based on multi-source data fusion according to claim 1, characterized in that, Before outputting the early warning command for battery thermal runaway, a low-frequency state verification step is also included: Receive the cell voltage sequence and macroscopic charge / discharge current sequence from the battery management unit; Calculate DC internal resistance variation data based on individual cell voltage sequences and macroscopic charge / discharge current sequences; By correlating the DC internal resistance change data with the temperature gradient data within the same judgment cycle, low-frequency state verification data is obtained. When the low-frequency state verification data meets the preset verification conditions, the early warning command for battery thermal runaway is marked as a valid command. When the low-frequency state verification data does not meet the preset verification conditions, the current total energy integral data and the circular memory queue are retained, and the next judgment cycle is entered.

9. The method for early warning of battery thermal runaway based on multi-source data fusion according to claim 8, characterized in that, The output of the early warning command for battery thermal runaway includes: When the current total energy integral data is greater than the first preset threshold, the time discrete variation coefficient is less than the second preset threshold, and the low frequency state verification data meets the preset verification conditions, the first-level warning data is generated and sent to the battery management unit. When the current total energy integral data is greater than the energy threshold above the first preset threshold, the half-width is greater than the preset time threshold, and the temperature gradient data is greater than the third preset threshold, a second-level warning data is generated. Based on the second-level warning data, an early warning command for battery thermal runaway is generated and sent to the vehicle control unit, causing the vehicle control unit to perform current limiting and contactor disconnection control.

10. A battery thermal runaway early warning system based on multi-source data fusion, applied to the battery thermal runaway early warning method based on multi-source data fusion as described in any one of claims 1-9, characterized in that, include: The acquisition module is used to acquire the high-frequency transient voltage sequence of the battery tab connection circuit through a high-frequency transient voltage sensor, acquire the high-frequency current derivative sequence of the same circuit through a high-frequency current sensor, and receive the macroscopic charge and discharge current sequence and tab surface temperature sequence transmitted by the battery management unit. The alignment module is used to perform time-series alignment of high-frequency transient voltage sequences, high-frequency current derivative sequences, macroscopic charge-discharge current sequences, and tab surface temperature sequences based on sampling timestamps, forming multi-source aligned data under the same time reference; The processing module is used to generate an adaptive cancellation signal based on the high-frequency current derivative sequence, and to perform point-by-point differential processing on the high-frequency transient voltage sequence and the adaptive cancellation signal to obtain the target transient voltage mutation sequence. The extraction module is used to calculate the dynamic trigger threshold based on the macroscopic charging and discharging current sequence and the electrode surface temperature sequence. When the voltage amplitude of the target transient voltage change sequence exceeds the dynamic trigger threshold, the voltage peak value, half width at half maximum, waveform integral area and time interval from the previous peak of the corresponding waveform segment are extracted to generate a transient event feature vector and write it into a pre-ordered circular memory queue. The determination module is used to enter the thermo-mechanical coupling determination loop, update the current total energy integral data based on the waveform integral area, generate dynamic forgetting factor data based on the macroscopic charge and discharge current sequence and the electrode surface temperature sequence, use the dynamic forgetting factor data to perform attenuation calculation on the current total energy integral data, and calculate the time discrete variation coefficient based on the time interval from the previous peak in the circular memory queue. The output module is used to calculate the temperature gradient data within the time period corresponding to the electrode surface temperature sequence when the current total energy integral data is greater than the first preset threshold and the time discrete variation coefficient is less than the second preset threshold; when the temperature gradient data is greater than the third preset threshold and the half-width is greater than the preset time threshold, it outputs an early warning command for battery thermal runaway.

Images