A Dynamic Safety Early Warning Method for New Energy Vehicles Integrating the NEV-AI Model

By combining multi-source heterogeneous signal processing and the NEV-AI model, a reliable time-series signal sequence is constructed and hierarchical early warning calculation is performed, which solves the problem of early fault identification and misjudgment in the safety early warning of new energy vehicles and improves the reliability and interpretability of the early warning.

CN123069148APending Publication Date: 2026-11-10CHINA AUTOMOTIVE INTELLIGENT INSPECTION (XIAMEN) NEW ENERGY CO LTD
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202611212256.8
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2026-08-11
Publication Date
2026-11-10

AI Technical Summary

Technical Problem

Existing safety early warning technologies for new energy vehicles cannot effectively identify early signs of gradual deterioration. They lack physical mechanisms such as electrochemical and electrical heat conduction constraints, have a high misjudgment rate, and the early warning results lack traceability and interpretability, making it difficult to support tiered operation and maintenance decisions.

Method used

By employing multi-source heterogeneous vehicle signal encapsulation, protocol parsing, and semantic mapping, a reliable time-series signal sequence is constructed. Combined with the NEV-AI model, hierarchical early warning calculation is performed. Through evidence chain integrity assessment and historical accuracy matching, the optimal early warning event is output.

Benefits of technology

It enables early identification of cascading failures, reduces the false negative rate, improves the reliability and interpretability of early warnings, and supports tiered operation and maintenance response.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN123069148A_ABST
    Figure CN123069148A_ABST
Patent Text Reader

Abstract

This invention belongs to the field of early warning technology for new energy vehicles, and discloses a dynamic safety early warning method for new energy vehicles that integrates the NEV-AI model. The method includes: collecting and processing multi-source heterogeneous vehicle signals to obtain a standardized signal frame set; constructing a reliable time-series signal sequence based on the standardized signal frame set; constructing a feature vector based on the reliable time-series signal sequence, calculating a fusion risk score, and obtaining an early warning level based on the fusion risk score; generating multiple candidate early warning events based on the early warning level, and selecting the optimal candidate early warning event for output. This invention achieves accurate quantitative differentiation of the reliability levels of different early warning events by integrating a comprehensive confidence quantification evaluation system that integrates real-time evidence chain completeness and historical fault mode performance. This provides a clear decision-making basis for tiered operation and maintenance response, effectively controlling the false alarm rate while ensuring the accuracy of early fault identification, and comprehensively improving the reliability and engineering practical value of dynamic safety early warning for new energy vehicles.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This invention relates to the field of early warning technology for new energy vehicles, and more specifically, to a dynamic safety early warning method for new energy vehicles that integrates the NEV-AI model. Background Technology

[0002] With the rapid development of the new energy vehicle industry, the operational safety of core systems such as power batteries, drive motors, and high-voltage electronic controls directly determines the reliability of the vehicle and the safety of passengers. Dynamic safety early warning technology has become one of the core technologies for the safety management and control of new energy vehicles.

[0003] Existing safety early warning technologies for new energy vehicles mainly fall into two categories: One is the traditional early warning scheme based on threshold judgment. This scheme sets fixed alarm thresholds for monitoring parameters such as battery cell voltage, temperature, and insulation resistance; exceeding these thresholds triggers an early warning. While simple in principle, this scheme can only identify obvious faults that have already occurred, failing to capture early signs of progressive degradation. Furthermore, it does not consider spatial relationships such as electrical connections and thermal coupling between cells, resulting in insufficient ability to predict cascading faults and a high false negative rate. The other category is intelligent early warning schemes based on machine learning. These schemes train neural network models by collecting historical operating data to output fault probabilities. However, existing AI early warning models generally lack constraints from physical mechanisms such as electrochemical, electrical, and thermal conduction, easily producing misjudgments that do not conform to objective operating laws. Additionally, the early warning outputs are mostly single probability values, lacking traceable and interpretable evidence chains, making it difficult for maintenance personnel to determine the source and credibility of the early warning conclusions. Furthermore, existing early warning technologies have not yet established a comprehensive quantitative evaluation system for early warning confidence. They have not verified the integrity of real-time early warning evidence from dimensions such as causal logic, temporal continuity, and multi-source mutual verification, nor have they conducted empirical verification based on historical performance of similar failure modes. As a result, the credibility level of early warning results cannot be quantitatively distinguished, making it difficult to support tiered response operation and maintenance decisions, and failing to ensure the sensitivity of early fault identification while controlling the false alarm rate.

[0004] In view of this, the present invention proposes a dynamic safety early warning method for new energy vehicles that integrates the NEV-AI model to solve the above problems. Summary of the Invention

[0005] To overcome the aforementioned deficiencies of the prior art and achieve the above objectives, the present invention provides the following technical solution: A dynamic safety early warning method for new energy vehicles integrating the NEV-AI model, the method includes: S1. Collect multi-source heterogeneous vehicle signals from new energy vehicles, encapsulate the multi-source heterogeneous vehicle signals into raw message frames, perform protocol parsing and semantic mapping processing on the raw message frames within a preset time window, and obtain a standardized signal frame set. S2. Perform multiple analysis on each standardized signal frame in the standardized signal frame set to obtain the comprehensive quality vector of each standardized signal frame; based on each comprehensive quality vector, select qualified signal frames from the standardized signal frame set, and construct a reliable time-series signal sequence based on all qualified signal frames; S3. Construct a feature vector based on a reliable time-series signal sequence, perform hierarchical early warning calculation on the feature vector to obtain a fusion risk score, and map the fusion risk score according to a preset hierarchical threshold to obtain the corresponding early warning level; S4. Generate multiple candidate warning events based on the warning level, and perform evidence chain integrity assessment and historical accuracy matching on each candidate warning event to obtain the assessment results and matching results. Based on the assessment results and matching results, select the optimal candidate warning event as the final warning event and output it.

[0006] Furthermore, the method for encapsulating multi-source heterogeneous vehicle signals into original message frames is as follows: Differential sampling is performed on multi-source heterogeneous vehicle signals. The multi-source heterogeneous vehicle signals after differential sampling are aggregated to the vehicle terminal through the vehicle bus. During the aggregation process, multi-source timestamp alignment and bus conflict resolution are performed to obtain the aggregated signal. The aggregated signal is then encapsulated to obtain the original message frame.

[0007] Furthermore, the method for obtaining the standardized signal frame set is as follows: The protocol type is identified by performing protocol type identification on the original message frame to obtain the protocol type of the original message frame. Based on the protocol type, the fields of the original message frame are extracted to obtain the source protocol field set. Perform GB / T semantic mapping on the source protocol field set to obtain an initial standardized field set. Perform cross-field physical constraint consistency verification on the initial standardized field set to obtain a verified standardized field set. The standardized field set after verification is recombined with the original message frame to obtain the corresponding standardized signal frame; a standardized signal frame set is constructed based on all standardized signal frames within a preset time window.

[0008] Furthermore, the method for obtaining the composite quality vector of each standardized signal frame is as follows: Perform authenticity verification on each standardized signal frame to obtain the authenticity quality component; perform missingness analysis on each standardized signal frame to obtain the missing quality component; perform transmission continuity analysis on each standardized signal frame to obtain the transmission quality component; perform sampling validity analysis on each standardized signal frame to obtain the sampling quality component. The true quality component, missing quality component, transmission quality component, and sampling quality component corresponding to the standardized signal frame are combined to obtain the comprehensive quality vector of each standardized signal frame.

[0009] Furthermore, the method for constructing a reliable time-series signal sequence is as follows: Perform hierarchical threshold judgment on each standardized signal frame, and form a candidate qualified signal frame set by forming standardized signal frames that pass the hierarchical threshold judgment. Perform quality degradation repair on standardized signal frames that fail the hierarchical threshold judgment and then form a repaired signal frame set. Perform timing quality continuity verification on the candidate qualified signal frame set, and combine the candidate qualified signal frames that pass the timing quality continuity verification into a timing verification signal frame set; The timing verification signal frame set and the repair signal frame set are sorted by timestamp to form an initial reliable timing signal sequence; A quality-weighted sliding smoothing process is performed on the comprehensive quality vector of each signal frame in the initial reliable time-series signal sequence to obtain the final reliable time-series signal sequence.

[0010] Furthermore, the method for constructing feature vectors is as follows: The reliable time-series signal sequence is divided into short-period subsequences, medium-period subsequences and long-period subsequences, and the working condition adaptive segmentation is performed on each periodic subsequence to obtain multiple working condition aligned subsequences for each periodic subsequence; hierarchical feature extraction is performed on each working condition aligned subsequence to obtain the basic statistical features, working condition alignment features and multidimensional coupling features corresponding to each working condition aligned subsequence. For each working condition under each periodic subsequence, perform intra-period aggregation on the features of the aligned subsequence to obtain the periodic feature vector corresponding to each periodic subsequence; then, horizontally concatenate the periodic feature vectors according to their dimensions to construct a feature vector; the feature vector includes short-period feature components, medium-period feature components, and long-period feature components.

[0011] Furthermore, the method for obtaining the fusion risk score is as follows: The short-cycle feature components are input into a preset black-box expert rule base to perform supplementary calculations and output supplementary risk scores. Input the mid-cycle feature components into the preset national standard rule engine to perform preliminary screening calculations and output the preliminary screening risk score; Long-period feature components are input into a pre-defined NEV-AI time-series deep learning model to perform predictive calculations and output progressive fault probability scores. Calculate the fusion weights of the supplementary risk score, the initial screening risk score, and the progressive failure probability score. Multiply the supplementary risk score, the initial screening risk score, and the progressive failure probability score by their respective fusion weights and sum them up to obtain the fusion risk score.

[0012] Furthermore, the method for obtaining the corresponding warning level is as follows: The system presets a first-level high-risk threshold, a second-level hidden danger threshold, and a third-level attention threshold, and constructs a tiered threshold range, while setting a boundary ambiguity zone between adjacent thresholds; If the fusion risk score falls within the corresponding graded threshold range, the corresponding graded threshold range will be directly mapped to the corresponding warning level. If the fusion risk score falls within the corresponding boundary ambiguity area, calculate the membership pairs of the fusion risk score to the two adjacent warning levels, and determine the corresponding warning level based on the membership pairs.

[0013] Furthermore, the method for selecting the optimal candidate early warning event is as follows: For each candidate early warning event, an evidence chain integrity assessment is performed to obtain an evidence chain integrity score for each candidate early warning event; for each candidate early warning event, a historical accuracy matching is performed to obtain a historical accuracy score for each candidate early warning event; the evidence chain integrity score and the historical accuracy score corresponding to each candidate early warning event are weighted and summed to obtain a comprehensive confidence score for each candidate early warning event. A multi-objective optimization function is constructed, which aims to maximize the comprehensive confidence score. Based on the multi-objective optimization function, Pareto front screening is performed on each candidate warning event, and the candidate warning event with the highest comprehensive confidence score is selected from the Pareto front as the final warning event.

[0014] Furthermore, the methods for obtaining the evidence chain completeness score and the historical accuracy score are as follows: Construct an evidence time-series causal graph, and calculate the causal structure score of each candidate early warning event based on the evidence time-series causal graph; Obtain the temporal evolution information of each candidate early warning event, and calculate the temporal coherence score of each candidate early warning event based on the temporal evolution information; Obtain multi-source evidence verification information for each candidate early warning event, and calculate the mutual verification reliability score for each candidate early warning event based on the multi-source evidence verification information; The causal structure score, temporal coherence score, and mutual evidence reliability score are weighted and accumulated to obtain the evidence chain integrity score. Construct candidate evolution feature sequences for each candidate early warning event, extract labeled evolution feature sequences of historical fault samples with the same early warning level as the candidate early warning event from a pre-set historical labeled sample library, calculate the similarity between the candidate evolution feature sequences and the labeled evolution feature sequences to obtain the historical pattern matching degree; retrieve the historical accuracy, historical false alarm rate and historical false alarm rate of the corresponding fault pattern based on the historical pattern matching degree. The historical accuracy score is obtained by weighted fusion of the inverse indices of historical accuracy, historical false positive rate, and historical false negative rate; where the inverse indices of historical false positive rate are 1 minus the historical false positive rate, and the inverse indices of historical false negative rate are 1 minus the historical false negative rate.

[0015] The technical effects and advantages of the new energy vehicle dynamic safety early warning method integrating the NEV-AI model of this invention are as follows: This invention covers all types of fault modes, including instantaneous sudden faults, mid-term cumulative faults, and long-term progressive faults, through multi-timescale subsequence partitioning and condition-adaptive feature construction. Combined with a time-series graph neural network architecture with the ability to perceive spatial correlation between electrical and thermal coupling, it can accurately capture early signs of cascading faults, significantly improve the sensitivity of progressively deteriorating fault identification, and effectively reduce the false alarm rate. At the same time, it adopts the NEV-AI early warning model with physical information neural network constraints, so that the early warning reasoning process conforms to objective physical laws such as electrochemistry and heat conduction, reducing false judgments that do not conform to the operating mechanism. With the construction of a multi-dimensional evidence chain and a causal logic verification mechanism, the early warning conclusion has traceable physical support, effectively making up for the shortcomings of traditional AI early warning in terms of interpretability. This invention achieves precise quantitative differentiation of the credibility levels of different early warning events by integrating real-time evidence chain integrity and historical failure mode performance into a comprehensive confidence quantification evaluation system. This provides a clear decision-making basis for hierarchical operation and maintenance response, and can ensure the accuracy of early fault identification while effectively controlling the false alarm rate, thus comprehensively improving the reliability and engineering practical value of dynamic safety early warning for new energy vehicles. Attached Figure Description

[0016] Figure 1 This is a flowchart of the method of the present invention. Detailed Implementation

[0017] 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.

[0018] In one embodiment, a dynamic safety early warning method for new energy vehicles integrating the NEV-AI model is disclosed, such as... Figure 1 As shown, the method includes: S1. Collect multi-source heterogeneous vehicle signals from new energy vehicles, encapsulate the multi-source heterogeneous vehicle signals into raw message frames, and perform protocol parsing and semantic mapping processing on the raw message frames within a preset time window to obtain a standardized signal frame set.

[0019] The method for encapsulating multi-source heterogeneous vehicle signals into original message frames is as follows: performing differential sampling on multi-source heterogeneous vehicle signals, converging the differentially sampled multi-source heterogeneous vehicle signals to the vehicle terminal through the vehicle bus, and performing multi-source timestamp alignment and bus conflict resolution during the convergence process to obtain the converged signal, and encapsulating the converged signal to obtain the original message frame. The method for obtaining the standardized signal frame set is as follows: perform protocol type identification on the original message frame to obtain the protocol type of the original message frame, and extract fields from the original message frame based on the protocol type to obtain the source protocol field set; Perform GB / T semantic mapping on the source protocol field set to obtain an initial standardized field set. Perform cross-field physical constraint consistency verification on the initial standardized field set to obtain a verified standardized field set. The standardized field set after verification is recombined with the original message frame to obtain the corresponding standardized signal frame; a standardized signal frame set is constructed based on all standardized signal frames within a preset time window.

[0020] Specifically, multi-source heterogeneous vehicle signals include raw measurement signals from five channels of new energy vehicles, covering core vehicle safety-related systems, including but not limited to: Battery Management System (BMS): Individual cell / module voltage sequence, total voltage, charge / discharge current, temperature field sequence, SOC, SOH, insulation resistance, relay status, charging status, alarm flags, etc. Drive motor and electronic control channel (MCU): motor speed, torque, controller temperature, DC bus voltage / current, motor over-temperature / over-current / stalled status, etc. High voltage circuit and accessory channels: High voltage interlock (HVIL) on / off status, DC / DC input / output voltage, current and temperature, fast and slow charging gun connection status, pre-charge complete indicator, etc.; Vehicle driving and operating condition channel (VCU / instrument / positioning): vehicle speed, mileage, gear, brake / accelerator pedal opening, latitude, longitude and altitude, ignition / ready status, etc.; Terminal and Link Status Channel (T-BOX): Terminal clock, message serial number, communication signal strength, packet loss statistics, firmware version, etc.

[0021] Based on the signal characteristics of each original measurement signal, differentiated sampling is performed on the original measurement signals of the five channels to balance bandwidth occupancy and anomaly capture accuracy. Specifically, those skilled in the art can classify the specific signals in the five channels into three categories: steady-state operation signals, sudden event signals, and operating condition switching point signals, based on the physical change characteristics and anomaly evolution laws of each signal, combined with engineering experience or expert calibration methods. For steady-state operation signals (such as total voltage and SOC in driving state), periodic sampling is performed according to the fixed period specified in GB / T32960 and the enterprise's extended protocol (e.g., 0s-30s by default in driving state, which can be extended to the minute level in static state). For sudden event signals (such as sudden drop in insulation strength, overvoltage of individual cells, sudden rise in temperature, and rapid acceleration or deceleration of vehicle speed), encrypted sampling is triggered, and the sampling period is shortened to the 1s-5s level. For operating condition switching point signals (such as charge / discharge switching, current step switching, etc.), segmentation marking is performed, and the switching timestamp is recorded to provide a benchmark for subsequent operating condition alignment feature calculation.

[0022] After differential sampling, the multi-source heterogeneous vehicle signals are aggregated to the vehicle terminal (T-BOX) via the vehicle CAN bus. Two processing steps are performed during the aggregation process: Multi-source timestamp alignment: Using the local clock of the vehicle terminal as a unified reference, a unified timestamp label is applied to the signals of each channel to eliminate clock drift deviation between different ECUs; Bus conflict resolution: When multiple channel signals request transmission in the same bus time slot, bus arbitration is performed according to the preset signal safety level priority to avoid high-priority safety signals being overridden. The preset signal safety level priority is based on the degree of impact of each channel signal on vehicle operation safety. High-voltage safety signals have the highest priority, followed by thermal safety signals, and then electrical status, driving conditions, and terminal link signals in descending order. For example, high-voltage safety signals (HVIL, insulation resistance) are higher than thermal safety signals (temperature, thermal runaway indicator) which are higher than electrical status signals (voltage, current, SOC) which are higher than driving conditions signals (vehicle speed, pedal) which are higher than terminal link signals (RSSI, serial number). After multi-source heterogeneous vehicle signals are aggregated to the vehicle terminal (T-BOX) via the vehicle CAN bus, a converged signal is obtained and encapsulated into a raw message frame. This raw message frame structure includes: a unique vehicle identifier field, a unified timestamp field, a protocol type field (identifying either a national standard protocol or a proprietary enterprise protocol), and a payload field set organized hierarchically according to five channel types. The payload field set is hierarchically encoded according to channel type, with each layer containing the physical quantity measurement value and status bit for that channel. The physical quantity measurement value refers to the physical quantity directly measured by sensors, such as power electrical... The voltage and current of the battery channel, the motor speed, torque, and controller temperature of the drive motor and electronic control channel, etc., and the status bits represent the discrete states of the corresponding channels, such as the relay status of the power battery channel (0=open, 1=closed), the charging status (0=not charged, 1=slow charging, 2=fast charging), the alarm flag (0=normal, 1=overvoltage alarm), the HVIL on / off status of the high voltage circuit and accessory channel (0=open, 1=connected), the pre-charge completion flag (0=not completed, 1=completed), etc., which facilitates subsequent protocol parsing by directly locating by layer without traversing all fields.

[0023] The original message frame is subjected to protocol type identification to obtain its protocol type. Based on the protocol type, fields are extracted from the original message frame to obtain the source protocol field set. Specifically: Matching is performed based on a pre-built protocol feature fingerprint database, which is pre-stored locally on the gateway. This database includes packet frame header structure features (including fixed header byte patterns for each protocol type), field length distribution features (including standard frame lengths and allowed frame length tolerance ranges for each protocol type), and timing interaction features (the handshake message timing patterns for each protocol type during session establishment). The header bytes, frame length, and preceding interaction timing of the original packet frame are compared with the protocol feature fingerprint database, and the protocol type confidence score is output. The protocol type with the highest confidence score is taken as the identification result. Specifically: The first N bytes of the original message frame (N is the preset frame header length, such as 4 bytes) are compared bit-by-bit with the fixed frame header byte patterns of each protocol type in the protocol feature fingerprint database. The number of successful byte matches is counted, and the frame header matching degree is obtained by dividing the number of successful byte matches by N bytes. The total frame length of the original message frame is compared with the standard frame length and the allowed frame length tolerance range of each protocol type. The difference between the total frame length and the standard frame length is recorded as the tolerance value. If the tolerance value is within the allowed frame length tolerance range, the frame length matching degree is 1. If the tolerance value is not within the allowed frame length tolerance range, the value exceeding the allowed frame length tolerance range is set to 1. The difference is recorded as the excess value. The historical maximum excess value is obtained from the protocol feature fingerprint database. The excess value is divided by the maximum excess value to obtain the excess rate. The excess rate is subtracted from 1 to obtain the corresponding frame length matching degree. The sequence similarity comparison is performed between the pre-processing interaction sequence (such as the pre-processing handshake message type sequence) of the original message frame and the handshake sequence pattern of each protocol type. The similarity comparison can use existing technologies such as Levinstein edit distance calculation to obtain the temporal matching degree. The frame header matching degree, frame length matching degree, and temporal matching degree are accumulated to obtain the confidence degree of each protocol type. The protocol type with the highest confidence degree is taken as the identification result.

[0024] Based on the field definition table corresponding to the matched protocol type, which is pre-stored in the protocol feature fingerprint database, the field definition table contains the field name, field order, field length, field data type and unit of measurement mapping rules of each field in the payload field set under the protocol type. Field decomposition is performed on the payload field set corresponding to the original message frame to extract the source protocol field set.

[0025] A national standard semantic mapping is performed on the source protocol field set to obtain an initial standardized field set, thereby achieving semantic unification of signals from multiple vehicle brands. National standard semantic mapping is an existing technology, and the specific mapping process will not be elaborated upon here. Cross-field physical constraint verification is performed on the initial standardized field set to filter out abnormal data that clearly violates physical laws. For example, the verification items include: Voltage-State of Charge Monotonicity Constraint Verification: Verify that the voltage increases monotonically with SOC during charging and decreases monotonically during discharging. Fields that do not conform to monotonicity are judged as conflicting fields. Current-power boundary constraint verification: Verify that the calculated power value (voltage × current) is within the rated power range of the corresponding device; values ​​exceeding the boundary are considered conflict fields. Temperature-thermodynamic feasible region constraint verification: Verify that the rate of temperature change conforms to the heat transfer limit, there is no abrupt abnormal temperature rise, and fields exceeding the feasible region are judged as conflict fields. When the cross-field physical constraint consistency check fails, it indicates that there are multiple conflicting fields for the same physical measurement value or that the mapping result violates physical laws. In this case, conflict resolution is performed on the conflicting fields according to the field confidence priority. The field confidence priority is preset based on the reliability of the acquisition link from which the field originates. For example, fields directly sampled by BMS have high confidence (direct sensor connection, no intermediate forwarding), fields forwarded by T-BOX have medium confidence (there may be delay or packet loss after bus aggregation), and fields estimated by VCU have low confidence (based on model inference rather than direct measurement). When there are conflicts for the same physical measurement value, the high confidence field is retained and the low confidence conflicting field is removed to obtain the standardized field set after verification.

[0026] Finally, the standardized field set after verification is recombined with the vehicle unique identifier field and unified timestamp field in the original message frame to generate the corresponding standardized signal frame; all standardized signal frames within the preset time window are collected in chronological order to form a standardized signal frame set.

[0027] By collecting signals from five channels of new energy vehicles across all dimensions, the system can comprehensively cover the operational status of core safety-related systems, breaking through the limitations of traditional solutions that only focus on monitoring individual battery parameters. Simultaneously, through differentiated sampling strategies, while controlling onboard transmission bandwidth and storage overhead, it effectively improves the ability to capture instantaneous anomalies and operational condition boundary points, avoiding missed anomalies caused by fixed-period sampling. Furthermore, through protocol type identification, national standard semantic mapping, and cross-field physical constraint consistency verification, it achieves unified standards and preliminary data purification for heterogeneous onboard signals from multiple brands and protocols. This not only solves the industry pain point of the inability to integrate and calculate data from different vehicle sources but also filters out outliers that violate physical laws from the data source, providing standardized and highly reliable basic data input for subsequent data quality control and hierarchical early warning calculations.

[0028] S2. Perform multiple analysis on each standardized signal frame in the standardized signal frame set to obtain the comprehensive quality vector of each standardized signal frame; based on each comprehensive quality vector, select qualified signal frames from the standardized signal frame set, and construct a reliable time-series signal sequence based on all qualified signal frames.

[0029] The method for obtaining the comprehensive quality vector of each standardized signal frame is as follows: perform authenticity verification on each standardized signal frame to obtain the authenticity quality component; perform missing quality analysis on each standardized signal frame to obtain the missing quality component; perform transmission continuity analysis on each standardized signal frame to obtain the transmission quality component; and perform sampling validity analysis on each standardized signal frame to obtain the sampling quality component. The true quality component, missing quality component, transmission quality component, and sampling quality component corresponding to the standardized signal frame are combined to obtain the comprehensive quality vector of each standardized signal frame. The method for constructing a reliable time-series signal sequence is as follows: perform hierarchical threshold judgment on each standardized signal frame, form a candidate qualified signal frame set by constructing each standardized signal frame that passes the hierarchical threshold judgment, and perform quality degradation repair on each standardized signal frame that fails the hierarchical threshold judgment to construct a repaired signal frame set. Perform timing quality continuity verification on the candidate qualified signal frame set, and combine the candidate qualified signal frames that pass the timing quality continuity verification into a timing verification signal frame set; The timing verification signal frame set and the repair signal frame set are sorted by timestamp to form an initial reliable timing signal sequence; A quality-weighted sliding smoothing process is performed on the comprehensive quality vector of each signal frame in the initial reliable time-series signal sequence to obtain the final reliable time-series signal sequence.

[0030] Specifically, for each standardized signal frame in the standardized signal frame set, authenticity verification, missing quality analysis, transmission continuity analysis, and sampling validity analysis are performed sequentially to generate a comprehensive quality vector for each standardized signal frame. Authenticity verification is performed on each standardized signal frame to obtain the authenticity quality component; missing quality analysis is performed on each standardized signal frame to obtain the missing quality component; transmission continuity analysis is performed on each standardized signal frame to obtain the transmission quality component; and sampling validity analysis is performed on each standardized signal frame to obtain the sampling quality component. Each component has a value range of 0-1, with higher scores indicating better quality in the corresponding dimension. The calculation method for each component is as follows: Cross-field coupled verification is performed on voltage-SOC consistency, current-power boundary consistency, and temperature-thermodynamic feasibility within the same standardized signal frame. Specifically, using a preset OCV-SOC lookup table model, the theoretical SOC value is inversely derived from the individual voltage sequence of the current standardized signal frame. The absolute value of the difference between the theoretical SOC value and the actual reported SOC value of the current standardized signal frame is calculated, and after normalization, the voltage-SOC consistency deviation rate is obtained. Subtracting the voltage-SOC consistency deviation rate from 1 yields the voltage-SOC consistency compliance. Instantaneous total power is calculated from total voltage and total current. The absolute value of the difference between the instantaneous total power and the maximum allowable charge / discharge power of the battery pack for this vehicle model is calculated, and after normalization, the power boundary deviation is obtained. Subtracting the power boundary deviation from 1 yields the power boundary compliance. The maximum allowable heat generation power and equivalent heat capacity of the battery for this vehicle model are obtained, as well as the current... The time interval of the previous standardized signal frame (the time difference between the timestamp of the current standardized signal frame and the timestamp of the previous standardized signal frame) is used to calculate the theoretical maximum temperature rise by multiplying the ratio of the battery's maximum allowable heat output to the battery's equivalent heat capacity by the time interval. The absolute value of the difference between the theoretical maximum temperature rise and the actual temperature rise (battery temperature minus ambient temperature) reported in the current standardized signal frame is calculated and normalized to obtain the thermodynamic feasible region deviation. Subtracting the thermodynamic feasible region deviation from 1 yields the thermodynamic feasible region compliance. The voltage-SOC consistency compliance, power boundary compliance, and thermodynamic feasible region compliance are weighted and summed to obtain the authenticity quality component. Here, each weight is preset based on expert experience, and the sum of the three weights is 1. The authenticity quality component represents the degree of conformity between the measured values ​​of each physical quantity in the current standardized signal frame and the inherent laws of battery electrochemistry. The larger the value, the higher the data authenticity.

[0031] Perform mandatory field null value detection on the current standardized signal frame, and calculate the ratio of the number of null values ​​in mandatory fields (such as unit voltage, total current, SOC, insulation resistance, etc.) to the total number of mandatory fields to obtain the intra-frame missing value. Then, trace back to previous standardized signal frames (all standardized signal frames in the set with timestamps earlier than the current standardized signal frame), and detect the timestamp interval (the difference between the timestamp of the current standardized signal frame and the timestamp of the previous standardized signal frame) of key channels (such as the BMS voltage channel and the MCU motor temperature channel). Use this timestamp interval as the inter-frame gap duration. Finally, correlate the inter-frame gap duration with the corresponding key channels. The ratio of the theoretical maximum allowable interval of the key channel (based on empirical presets, such as a theoretical interval of 10s for the BMS channel in driving mode) is denoted as the inter-frame missing value. The total missing value is obtained by weighting and fusing the intra-frame missing value and the inter-frame missing value. Here, the weights of the intra-frame missing value and the inter-frame missing value are preset according to expert experience. The total missing value is subtracted from 1 and normalized to obtain the missing quality component. The missing quality component represents the degree of data integrity, reflecting the completeness of the required fields and key channel data in the current standardized signal frame and the previous standardized signal frame. The larger the value (closer to 1), the less data is missing, the shorter the gap, and the better the integrity.

[0032] Let the sequence number of the current standardized signal frame be... The serial number of the previous standardized signal frame D1 is ,like If a serial number change occurs, it is determined that a change has occurred, and this is recorded as one change count; let the timestamp of the current standardized signal frame be... The timestamp of the previous standardized signal frame is ,like and If the difference is greater than the preset timeout threshold (set based on experience, such as 60s), a link interruption is determined to have occurred, and it is recorded as one timeout. The number of jumps and timeouts within the preset evaluation window are counted. The sum of the jumps and timeouts is divided by the total number of frames within the preset evaluation window to obtain the transmission anomaly rate. Subtracting the transmission anomaly rate from 1 gives the transmission quality component. The transmission quality component represents the degree of continuity of data transmission, reflecting the continuity of the message serial number and the stability of the link transmission. The larger the value (the closer to 1), the fewer the serial number jumps, the shorter the link interruption, and the more continuous and stable the transmission.

[0033] A time-series physical consistency check is performed based on the rate of change of physical measurements of adjacent standardized signal frames. For example, for the current standardized signal frame, the rate of change of single-cell voltage ΔU / Δt and the rate of change of temperature ΔT / Δt between it and the adjacent standardized signal frame (i.e., the previous standardized signal frame) are calculated and compared with preset battery physical limits (such as the maximum allowable rate of change of voltage and the maximum allowable rate of temperature rise set according to the characteristics of the battery itself). If there is a rate of change that exceeds the battery physical limits, the current standardized signal frame is determined to be an abnormal sampling point. The number of abnormal sampling points in the current standardized signal frame and the previous standardized signal frame is counted, and the ratio of the number of abnormal sampling points to the total number of sampling points is recorded as the sampling anomaly rate. The sampling quality component is obtained by subtracting the sampling anomaly rate from 1. The sampling quality component represents the degree of data sampling effectiveness and reflects the regularity of the sampling period. The larger the value (the closer to 1), the more stable the sampling period, the fewer repeated frames and noise spikes, and the higher the sampling effectiveness.

[0034] Based on the comprehensive quality vector of each standardized signal frame, qualified signal frames are selected from the set of standardized signal frames and a reliable time-series signal sequence is constructed. This process includes the following sub-steps: Based on historical data of new energy vehicles (i.e., similar models under the same brand and power battery technology) under different operating conditions, independent screening thresholds are set for each quality component according to the current operating conditions of new energy vehicles (such as charging, driving, and stationary conditions). For example, the comprehensive quality vector of all standardized signal frames in the past preset period (e.g., 30 days) of this model family is statistically analyzed in different dimensions to calculate the historical mean and standard deviation of each quality component. The independent screening threshold is obtained by subtracting sa times the standard deviation from the mean, where the value of sa is preset based on experience. A stratified threshold judgment is performed on each standardized signal frame: a standardized signal frame in which all four quality components are higher than the corresponding independent screening threshold is considered to have passed the stratified threshold judgment and is included in the candidate qualified signal frame set; a standardized signal frame in which any quality component is lower than the corresponding independent screening threshold is judged to have failed the stratified threshold judgment and is recorded as an unqualified signal frame, and enters the quality degradation repair process.

[0035] Quality degradation repair: Two repair strategies are used for non-qualified signal frames, namely constraint completion repair and timing interpolation repair, as detailed below: Constraint completion repair: For fields that are locally missing or abnormal within a non-qualified signal frame, the defective fields are completed by constraint based on the battery electrochemical and physical constraints; for example, the missing cell voltage is deduced from the SOC value using a preset voltage-SOC monotonic function and then completed. Timing interpolation repair: For missing fields in non-qualified signal frames, linear interpolation or cubic spline interpolation is used to fill in the missing fields based on the corresponding fields in the adjacent qualified signal frames. The signal frames obtained after quality degradation repair constitute the repaired signal frame set.

[0036] A time-series quality continuity check is performed on the candidate qualified signal frame set: the cosine similarity of the comprehensive quality vector of each candidate qualified signal frame with the preceding and following candidate qualified signal frames is calculated, and the smaller of the two values ​​is taken as the quality difference. If the quality difference exceeds the preset continuity threshold (determined by those skilled in the art based on the historical statistical distribution of the comprehensive quality vector difference of adjacent signal frames under normal operating conditions of the target vehicle model, combined with the system's requirements for the stability of time-series data), it indicates that the corresponding candidate qualified signal frame has a sudden quality anomaly or occasional interference, and is therefore removed. Finally, the candidate qualified signal frames that pass the check are combined into a time-series check signal frame set.

[0037] The timing verification signal frame set and the repair signal frame set are arranged in ascending order according to a unified timestamp to form an initial reliable timing signal sequence.

[0038] For the physical measurement values ​​of each signal frame in the initial reliable time-series signal sequence, a quality-weighted sliding smoothing process is performed: the magnitude of the comprehensive quality vector of each signal frame is used as the smoothing weight of that signal frame, and the higher the quality of the signal frame, the greater the weight; a fixed-size sliding window (the default window size is 5 frames) is used to perform a weighted average calculation on the physical measurement values ​​within the window; the specific calculation rule is: the smoothed physical measurement value at the current time is equal to the sum of the product of the smoothing weight of each signal frame within the sliding window and the physical measurement value of that signal frame, divided by the sum of the smoothing weights of all signal frames within the sliding window; the weighted smoothing suppresses the numerical noise caused by quality fluctuations, and the final reliable time-series signal sequence is output.

[0039] By employing a four-dimensional comprehensive quality vector—comprising authenticity, missing data, transmission continuity, and sampling validity—to perform refined quality quantification evaluation of standardized signal frames, this approach replaces the traditional single-dimensional dirty data removal logic, achieving multi-dimensional and precise control over data quality. Simultaneously, by combining hierarchical threshold screening and quality degradation repair mechanisms, it filters low-quality data while maximizing the retention of valid samples, avoiding temporal sequence breaks caused by data removal. Furthermore, through temporal quality continuity verification and quality-weighted sliding smoothing, it further filters out abnormal interference from quality abrupt changes and suppresses data fluctuation noise, ultimately outputting a highly stable and reliable temporal signal sequence. This significantly reduces the proportion of invalid dirty data in vehicle-reported data, providing a reliable data foundation from the data source for subsequent feature extraction and hierarchical early warning calculations, effectively avoiding the negative impact of dirty data on the accuracy of the early warning model.

[0040] S3. Construct a feature vector based on a reliable time-series signal sequence, perform hierarchical early warning calculation on the feature vector to obtain a fusion risk score, and map the fusion risk score according to a preset hierarchical threshold to obtain the corresponding early warning level.

[0041] The method for constructing feature vectors is as follows: the reliable time-series signal sequence is divided into short-period subsequences, medium-period subsequences and long-period subsequences, and the working condition adaptive segmentation is performed on each periodic subsequence to obtain multiple working condition aligned subsequences for each periodic subsequence; hierarchical feature extraction is performed on each working condition aligned subsequence to obtain the basic statistical features, working condition alignment features and multidimensional coupling features corresponding to each working condition aligned subsequence. For each working condition under each periodic subsequence, perform intra-period aggregation on the features of the aligned subsequence to obtain the periodic feature vector corresponding to each periodic subsequence. Then, horizontally concatenate the periodic feature vectors according to the dimensions to construct a feature vector. The feature vector includes short-period feature components, medium-period feature components, and long-period feature components. The method for obtaining the fusion risk score is as follows: input the short-cycle feature components into a preset black-box expert rule base to perform supplementary calculations and output the supplementary risk score; Input the mid-cycle feature components into the preset national standard rule engine to perform preliminary screening calculations and output the preliminary screening risk score; Long-period feature components are input into a pre-defined NEV-AI time-series deep learning model to perform predictive calculations and output progressive fault probability scores. Calculate the fusion weights of the supplementary risk score, the initial screening risk score, and the progressive failure probability score. Multiply the supplementary risk score, the initial screening risk score, and the progressive failure probability score by their respective fusion weights and sum them up to obtain the fusion risk score. The method for obtaining the corresponding warning level is as follows: preset a first-level high-risk threshold, a second-level hidden danger threshold, and a third-level attention threshold, and construct a graded threshold interval, while setting a boundary ambiguity zone between adjacent thresholds; If the fusion risk score falls within the corresponding graded threshold range, the corresponding graded threshold range will be directly mapped to the corresponding warning level. If the fusion risk score falls within the corresponding boundary ambiguity area, calculate the membership pairs of the fusion risk score to the two adjacent warning levels, and determine the corresponding warning level based on the membership pairs.

[0042] Specifically, a multi-timescale sliding window partitioning is performed on the reliable time-series signal sequence to obtain three types of subsequences with different time spans: short-cycle subsequences (time window of 1 hour to 1 day, focusing on feature extraction of instantaneous sudden anomalies), medium-cycle subsequences (time window of 1 day to 7 days, focusing on feature extraction of mid-term cumulative anomalies), and long-cycle subsequences (time window of 7 days to 30 days, focusing on feature extraction of long-term gradual degradation). For each cycle subsequence, adaptive segmentation based on operating conditions is performed. Operating condition boundary points are identified based on abrupt changes in charging / discharging state (current direction changes from positive (discharging) to negative (charging) or vice versa, and the absolute value of the current exceeds a preset threshold), current step switching (during charging or discharging, the absolute value of the current undergoes a step change, and the change exceeds a preset proportion), and resting time threshold (current remains at 0 for a duration exceeding a preset resting threshold). Each cycle subsequence is then divided into four operating condition segments: charging segment, discharging segment, current switching segment, and resting segment. Using the operating condition boundary points as alignment benchmarks, multiple operating condition aligned subsequences are extracted to eliminate the interference of operating condition differences on feature calculation.

[0043] Perform three-layer feature extraction on each aligned subsequence for each operating condition: Basic statistical features: Calculate the mean, variance, maximum, minimum, and average rate of change of each physical quantity in each working condition alignment subsequence (by dividing the difference of the physical quantity of adjacent sampling points in the working condition alignment subsequence by the corresponding time interval to obtain the rate of change of each sampling point, and then taking the arithmetic mean to obtain the average rate of change), and the cumulative duration of exceeding the national standard threshold (by counting the number of consecutive sampling points in the working condition alignment subsequence where the physical quantity exceeds the specified threshold, and multiplying by the sampling period to obtain the cumulative duration of exceeding the national standard threshold), forming a basic statistical feature set; Operating condition alignment features: Based on the charging segment and current switching segment, specific features are calculated, including the battery cell internal resistance consistency index (the cell internal resistance is first calculated by the voltage difference and current difference before and after the current switching, and then a cell internal resistance sequence is constructed. The mean internal resistance and internal resistance variance of the cell internal resistance sequence are calculated. The internal resistance consistency index is obtained by subtracting the ratio of the internal resistance variance to the internal resistance mean from 1), insulation resistance operating condition baseline deviation (the absolute value of the difference between the current insulation resistance and the historical insulation resistance baseline under the same operating condition is normalized to obtain the insulation resistance operating condition baseline deviation), etc. Multidimensional coupling features: Calculate cross-dimensional correlation features, including calculating the Pearson correlation coefficient (voltage-temperature coupling feature) between the single-cell voltage sequence and the battery temperature field sequence within the aligned sub-sequence under the same operating condition; obtain vehicle speed, current direction, and ambient temperature, and construct an operating condition context embedding vector (e.g., discretize vehicle speed, current direction (charging, discharging, etc.), and ambient temperature into category labels, map them to fixed-dimensional numerical vectors through one-hot encoding or low-dimensional embedding, and concatenate them to obtain the operating condition context embedding vector); calculate the conditional deviation between insulation resistance and the operating condition context embedding vector; specifically, substitute the operating condition context embedding vector into an insulation resistance-operating condition regression model pre-fitted based on historical data to obtain... The resistance reference value is calculated by taking the absolute value of the difference between the insulation resistance and the resistance reference value, and then normalized to obtain the conditional deviation (insulation-operating condition coupling characteristic). The standard deviation of the accelerator pedal opening change rate (by performing a first-order difference on the accelerator pedal opening at adjacent sampling times within the current operating condition alignment sub-sequence, and dividing the difference value by the time interval between adjacent sampling times) and the braking frequency are obtained. The two are added together to obtain the driving intensity index. The driving intensity index at each sampling time within the current operating condition alignment sub-sequence is linearly regressed with the battery temperature and controller temperature to obtain the regression coefficients of the driving intensity index on the battery temperature and controller temperature, which serve as the driving-thermal load coupling characteristic.

[0044] For each periodic subsequence, the features of multiple operating condition alignment subsequences are aggregated within the period. For example, the average or maximum value of the internal resistance consistency index of multiple current switching segments within the same medium periodic subsequence is taken, and the average value of the basic statistical features of multiple charging and discharging segments is taken. After the periodic aggregation, the periodic feature vector corresponding to each periodic subsequence is obtained. The periodic feature vector includes short periodic feature components, medium periodic feature components, and long periodic feature components. The short periodic feature components, medium periodic feature components, and long periodic feature components are horizontally concatenated according to the dimensions to construct the final feature vector.

[0045] Based on the short-cycle, medium-cycle, and long-cycle feature components in the feature vector, hierarchical early warning calculations are performed in parallel, and three types of risk scores are output respectively. The fused risk score is obtained through dynamic weighted fusion.

[0046] Short-period feature components are input into a pre-defined black-box expert rule base for supplementary calculations. The black-box expert rule base includes: Thermal runaway precursor rule set: Detect whether the following precursor modes are met: abnormal temperature rise rate (e.g., temperature change rate greater than 5), asynchronous voltage drop and temperature rise (e.g., the time difference between voltage drop rate and temperature rise rate is less than a preset threshold), abnormal gas production pressure correlation (pressure change rate greater than a preset threshold). Unconventional high-voltage anomaly rule set: detects high-voltage interlock disconnection timing anomalies (HVIL state and Ready state do not match), DC / DC output voltage oscillation mode anomalies (output voltage fluctuation frequency / amplitude exceeds the normal operating range), precharge circuit timeout anomalies (precharge completion flag is not set in the preset time), etc.

[0047] Based on the above rule set, pattern matching is performed on the short-cycle feature components. The calculation process adopts a pattern matching algorithm, which matches the short-cycle feature components with each abnormal pattern set in the black-box expert rule base rule by rule. The number of feature items in the short-cycle feature components that match the corresponding abnormal patterns is counted. The number of feature items that match the corresponding abnormal patterns is divided by the total number of feature items in the short-cycle feature components to obtain the pattern matching degree. After normalizing the pattern matching degree, a supplementary risk score is output.

[0048] The intermediate-cycle feature components are input into a preset national standard rule engine to perform initial screening calculations. The national standard rule engine performs initial risk screening based on the national standard threshold rule library (such as the insulation resistance threshold Rth=100Ω / V specified in GB / T45688-2025). The national standard threshold rule library can be dynamically adjusted in combination with the vehicle model adaptive correction threshold. The vehicle model adaptive correction threshold is based on the historical fault distribution characteristics of the vehicle model family to offset the national standard general threshold. For example, if historical data of a certain vehicle model family shows that its normal insulation resistance is generally higher than the national standard threshold by 20%, then the insulation resistance judgment threshold of the vehicle model is increased by 20% to reduce the false alarm rate of this type of vehicle. Each feature item in the intermediate-cycle feature components is compared with the corresponding national standard threshold item by item, and the number of feature items exceeding the international threshold is counted. The number of feature items exceeding the international threshold is divided by the total number of feature items in the intermediate-cycle feature components to obtain the initial screening risk score.

[0049] Long-period feature components are input into a pre-defined NEV-AI time-series deep learning model to perform predictive calculations and output progressive fault probability scores. The NEV-AI time-series deep learning model is a temporal graph neural network (GNN) constrained by a physical information neural network (PINN). The model structure includes: Graph structure layer: First, the functional units such as individual battery cells, motor controllers, and DC / DC converters in the battery pack are abstracted as nodes in the graph structure. The electrical connection relationships (such as the series relationship between battery cells) and thermal conduction relationships (such as the heat transfer relationship between adjacent cells) between the units are abstracted as connecting edges to construct a graph structure that reflects the physical topology of the high-voltage system of new energy vehicles.

[0050] Graph Attention Layer: Based on the graph structure, the model aggregates information about each node and its neighboring nodes through a graph attention mechanism. Specifically, when each node calculates its own feature representation, it adaptively learns the importance weights of its neighboring nodes, focusing on nodes with strong correlation to it. For example, when a battery cell experiences a voltage abnormality, the model can automatically identify the pattern of the abnormality propagating to adjacent cells, thereby capturing the fault correlation features in the spatial dimension and outputting the spatial embedding vector of each node.

[0051] Temporal Evolution Layer: The spatial embedding vectors output by the graph attention layer are arranged in chronological order to form a sequence, which is then input into the gated recurrent unit for processing. The gated recurrent unit selectively memorizes long-term historical information and forgets irrelevant information through its internal update gate and reset gate mechanism, thereby capturing the gradual degradation trend of physical quantities such as voltage and temperature over a time span of several days to tens of days, and outputting a temporal hidden state sequence.

[0052] Physical information constraint layer: During model training, the inherent physical laws of the battery electrochemical system are embedded into the loss function; for example, the sum of the voltages of the cells in series should equal the total voltage, the input and output power should satisfy the law of energy conservation, and the temperature change should conform to the law of thermal diffusion; when the model prediction results violate these physical constraints, an additional penalty term will be generated to guide the model to learn feature representations that conform to physical laws and avoid non-physical predictions caused by pure data-driven approaches.

[0053] Predictive output layer: Based on the temporal hidden state sequence, the progressive failure probability score of each node is mapped and output through a fully connected layer. These probability scores reflect the possibility of progressive failure of each functional unit in the future, providing a long-term predictive basis for subsequent fusion decision-making.

[0054] This model explicitly models the physical connections between components through a graph structure, adaptively learns fault propagation patterns through an attention mechanism, captures long-term temporal patterns through gated recurrent units, and ensures the physical consistency of prediction results through physical information constraints, thus achieving a deep integration of data-driven and physical prior knowledge. The temporal graph neural network architecture, graph attention mechanism, gated recurrent unit operation logic, and network construction and training methods for physical information neural network constraints are all mature technologies already existing in this field and will not be described in detail here.

[0055] Historical accuracy, false positive rate, and false negative rate of each engine (national standard rule engine, NEV-AI model, and black-box expert rule base) are obtained. An online Bayesian optimization algorithm is used to minimize the overall false positive rate as the optimization objective, dynamically determining the dynamic fusion weights α, β, and γ corresponding to the initial screening risk score, progressive failure probability score, and supplementary risk score. The online Bayesian optimization algorithm uses (α, β, γ) as optimization variables and the overall false positive rate on the corresponding historical validation set as the objective function. The posterior distribution of the objective function is constructed using Gaussian process regression, and the optimal weight combination is iteratively searched using a collection function (such as the desired improvement EI). The constraints are α + β + γ = 1 and α, β, γ ≥ 0, ultimately obtaining their respective fusion weights. The Gaussian process regression modeling, the desired improvement collection function design, and the iterative search process of online Bayesian optimization are all well-known technologies in the field and will not be elaborated here. The supplementary risk score, initial screening risk score, progressive failure probability score, and their respective fusion weights are multiplied and summed to obtain the fusion risk score. A higher fusion risk score indicates a higher safety risk for the new energy vehicle.

[0056] Based on the distribution characteristics of the fusion risk scores of historical fault samples of vehicle models (such as percentiles at the boundaries of each level), and combined with the impact of current operating conditions (such as charging conditions, driving conditions, and stationary conditions) on risk sensitivity, a first-level high-risk threshold θ1, a second-level hidden danger threshold θ2, and a third-level attention threshold θ3 are dynamically set, satisfying θ1>θ2>θ3; at the same time, a boundary ambiguity zone is set between adjacent thresholds, including the width of the first boundary ambiguity zone δ1 and the width of the second boundary ambiguity zone δ2. The width of the ambiguity zone is adaptively adjusted according to the historical misjudgment rate of the vehicle model at the corresponding boundary. If the historical misjudgment rate of the vehicle model at the θ1 boundary is high, δ1 is increased to strengthen the secondary verification near the boundary. Thus, the hierarchical threshold intervals and boundary ambiguity regions are constructed as follows: Level 1 High-Risk Zone: [θ1,1], corresponding to Level 1 Warning Level; First boundary ambiguity region: [θ1−δ1,θ1); Level II hazard range: [θ2, θ1−δ1), corresponding to Level II warning level; Second boundary ambiguity region: [θ2−δ2,θ2); The third-level attention interval is [θ3, θ2−δ2), corresponding to the third-level warning level; No warning interval: [0, θ3), corresponding to no warning level; If the fusion risk score falls within the corresponding graded threshold range, the corresponding graded threshold range is directly mapped to the corresponding warning level. For example, when the fusion risk score falls within the first-level high-risk range, the warning level is the first-level warning level. If the fusion risk score falls within the corresponding boundary fuzzy area, the membership degree pair of the fusion risk score to the two adjacent warning levels is calculated, and the corresponding warning level is determined based on the membership degree pair. For example, when the fusion risk score falls within the first boundary fuzzy area, the membership degree u1 of the fusion risk score to the first-level warning level and the membership degree u2 of the fusion risk score to the second-level warning level are calculated according to the fuzzy membership degree function (such as the trapezoidal membership degree function or the Gaussian membership degree function). If u1 is greater than or equal to u2, it is determined to be the first-level warning level; otherwise, it is determined to be the second-level warning level.

[0057] By employing a feature construction approach that combines multi-timescale subsequence partitioning with adaptive segmentation based on operating conditions, the system covers three types of fault modes: instantaneous sudden faults, mid-term cumulative faults, and long-term gradual faults. This effectively eliminates the interference of operating condition differences on feature stability and avoids missed fault detection caused by a single time window. Simultaneously, relying on a three-layer parallel early warning architecture comprised of a national standard rule engine, a NEV-AI time-series deep learning model, and a black-box expert rule base, the system balances the compliance and interpretability of early warning results, the early prediction capability of gradual faults, and the coverage of long-tail anomaly scenarios, significantly expanding the fault identification range compared to a single early warning model. Furthermore, based on a dynamic weight fusion mechanism of historical performance and a membership grading strategy for ambiguous boundary regions, the contribution ratio of each early warning engine can be adaptively optimized according to operating conditions, resolving the grading jitter problem at hard threshold boundaries. Ultimately, this significantly improves the fault identification accuracy under complex operating conditions, effectively reduces false alarm and false negative rates, and enables early prediction of latent faults such as power battery consistency degradation.

[0058] S4. Generate multiple candidate warning events based on the warning level, and perform evidence chain integrity assessment and historical accuracy matching on each candidate warning event to obtain the assessment results and matching results. Based on the assessment results and matching results, select the optimal candidate warning event as the final warning event and output it.

[0059] The method for selecting the optimal candidate early warning event is as follows: perform evidence chain integrity assessment on each candidate early warning event to obtain the evidence chain integrity score of each candidate early warning event; perform historical accuracy matching on each candidate early warning event to obtain the historical accuracy score of each candidate early warning event; and weight and sum the evidence chain integrity score and historical accuracy score of each candidate early warning event to obtain the comprehensive confidence score of each candidate early warning event. A multi-objective optimization function is constructed, which aims to maximize the comprehensive confidence score. Based on the multi-objective optimization function, Pareto front screening is performed on each candidate warning event, and the candidate warning event with the highest comprehensive confidence score is selected from the Pareto front as the final warning event. The methods for obtaining the evidence chain completeness score and historical accuracy score are as follows: construct an evidence time-series causal graph, and calculate the causal structure score of each candidate warning event based on the evidence time-series causal graph; Obtain the temporal evolution information of each candidate early warning event, and calculate the temporal coherence score of each candidate early warning event based on the temporal evolution information; Obtain multi-source evidence verification information for each candidate early warning event, and calculate the mutual verification reliability score for each candidate early warning event based on the multi-source evidence verification information; The causal structure score, temporal coherence score, and mutual corroboration reliability score are weighted and summed to obtain the evidence chain integrity score. Construct candidate evolution feature sequences for each candidate early warning event, extract labeled evolution feature sequences of historical fault samples with the same early warning level as the candidate early warning event from a pre-set historical labeled sample library, calculate the similarity between the candidate evolution feature sequences and the labeled evolution feature sequences to obtain the historical pattern matching degree; retrieve the historical accuracy, historical false alarm rate and historical false alarm rate of the corresponding fault pattern based on the historical pattern matching degree. The historical accuracy score is obtained by weighted fusion of the inverse indices of historical accuracy, historical false positive rate, and historical false negative rate; where the inverse indices of historical false positive rate are 1 minus the historical false positive rate, and the inverse indices of historical false negative rate are 1 minus the historical false negative rate.

[0060] Specifically, based on the determined warning level, multiple candidate warning events are generated. Each candidate warning event is generated based on the fault code set output by the national standard rule engine, the rule hit identifier list output by the black-box expert rule base, and the progressive fault probability score output by the NEV-AI time-series deep learning model. For example, for each fault code in the fault code set, each hit rule number in the rule hit identifier list, and each progressive fault probability score output by the NEV-AI time-series deep learning model that exceeds the corresponding preset probability threshold (constrained by expert experience), progressive potential faults of various components are further classified. Each fault code is independently generated as a candidate early warning event. The fault code set is a list of fault codes output by the national standard rule engine after performing initial screening calculations on the medium-cycle feature components and determining which national standard thresholds are triggered (such as low insulation resistance fault codes, excessive single-unit voltage range fault codes, etc.). Each fault code corresponds to a specific national standard judgment rule. The rule hit identifier list is a list of hit rule numbers output by the black-box expert rule base after performing pattern matching on the short-cycle feature components (such as rule R001 - abnormal temperature rise rate before thermal runaway, rule R003 - abnormal HVIL disconnection timing, etc.). The data structure for each candidate warning event includes a unique vehicle identifier, warning level, fusion risk score, evidence node set, and event occurrence timestamp. The evidence node set contains three types of nodes: abnormal physical measurement nodes (such as abnormal single-unit voltage nodes, abnormal temperature nodes, abnormal insulation resistance nodes, etc.), rule-hit nodes (such as national standard rule-hit nodes, expert rule-hit nodes, etc.), and model-predicted nodes (such as consistency degradation probability nodes and insulation degradation probability nodes output by NEV-AI, etc.).

[0061] For each candidate early warning event, an evidence chain integrity assessment is performed. An evidence chain integrity score (ranging from 0 to 1) is calculated based on three dimensions: causal structure score, temporal coherence score, and mutual corroboration reliability score. A higher score indicates a more complete and credible evidence chain. Details are as follows: An evidence time-series causal graph is constructed based on the set of evidence nodes for candidate early warning events. The nodes in the graph are of three types. The edge weights of the graph are determined according to the temporal sequence relationship and physical transmission path between the evidence nodes. For example, for nodes i and j, if the timestamp of the anomaly occurrence of node i is earlier than that of node j, a directed edge is established from node i to node j. The edge weight of the directed edge is determined according to the physical transmission strength between node i and node j. If there is a known physical transmission relationship between node i and node j (such as the transmission link of single-unit voltage anomaly → current anomaly → temperature anomaly), the corresponding edge weight is assigned according to the transmission strength. The specific values ​​of each edge weight can be determined by those skilled in the art in combination with the physical transmission mechanism of the fault and the engineering experience of experts, and will not be described in detail here. Based on the edge weight, the logical transmission strength between each evidence node is calculated. For example, the weighted path propagation algorithm is used to calculate the sum of logical transmission strength between all evidence nodes, and the causal structure score is obtained after normalization. The causal structure score measures whether the evidence conforms to the physical logic of fault transmission and reflects the causal logical self-consistency of the evidence chain. The larger the value, the more the transmission relationship between the evidence conforms to objective physical mechanisms such as electrochemical and electrical heat conduction, and the more sufficient the physical causal basis of the warning conclusion.

[0062] The temporal evolution information of each candidate warning event is obtained. This information includes: the number of consecutive frames of the key anomaly features corresponding to each candidate warning event on the time axis (the number of consecutive frames in which the key anomaly features appear). Key anomaly features refer to features directly related to the fault type indicated by the current candidate warning event. For example, if the candidate warning event is a potential battery cell consistency degradation problem, the corresponding key anomaly features would be the cell voltage deviation time series, cell internal resistance deviation series, cell temperature dispersion series, etc.; sampling interval uniformity (the uniformity of the temporal distribution of sampling points during the anomaly period, such as the reciprocal of the actual sampling interval variance); and anomaly amplitude volatility (the ratio of the standard deviation to the mean of the key anomaly features). The three indicators—consistency frame count, sampling interval uniformity, and anomaly amplitude volatility—are then analyzed separately. Normalization is performed to map all values ​​to a range of 0-1. A higher number of consecutive frames and a more uniform sampling interval result in a higher normalization score, as does a lower volatility of the anomaly (a reverse normalization is applied to the volatility index). Then, based on preset weighting coefficients (combined with empirical presets), the three normalized scores are weighted and summed to obtain a temporal coherence score ranging from 0-1. The temporal coherence score characterizes the consistency of the continuous evolution of key anomaly features corresponding to candidate warning events over time, reflecting the temporal stability of the anomaly. A higher score indicates a more continuous and stable development process of the anomaly features, suggesting a continuously evolving, genuine fault symptom rather than a momentary random fluctuation, thus increasing the temporal reliability of the warning conclusion.

[0063] Acquire multi-source evidence verification information for each candidate early warning event; the multi-source evidence verification information describes the cross-hit state of physical measurement evidence (sensor measurement anomalies), rule-hitting evidence (triggered by the international rule engine), and model prediction evidence (high-probability output of the NEV-AI time-series deep learning model); the cross-hit state includes: Single-source independent trigger: Only one type of evidence source is hit (e.g., only physical measurement values ​​are abnormal, but rule-matched evidence and model-predicted evidence are not hit). Dual-source cross-verification: Two types of evidence sources hit simultaneously (e.g., abnormal physical measurement values ​​and triggering of the international rules engine, but the model does not output a high probability). Three-source consistency confirmation: all three types of evidence sources are simultaneously present; Differentiated mutual verification weights are assigned based on the cross-hit status. For example, a single-source independent trigger has a mutual verification weight of 0.3, two-source cross-verification has a mutual verification weight of 0.7, and three-source consistent confirmation has a mutual verification weight of 1.0. The mutual verification reliability score is the mutual verification weight corresponding to the cross-hit status of the current candidate warning event. If there are multiple evidence combinations, the largest mutual verification weight is taken. The mutual verification reliability score is used to characterize the degree of cross-verification of the candidate warning event by the three types of evidence: physical measurement, rule hit, and model prediction, reflecting the consistency of multi-source evidence. The larger the value, the more dimensions of independent evidence jointly support the warning conclusion. The higher the strength of multi-source mutual verification, the stronger the comprehensive reliability of the warning result and the lower the probability of false alarm.

[0064] Finally, the causal structure score, temporal coherence score, and mutual corroboration reliability score are weighted and summed to obtain the evidence chain completeness score. The weights of each can be determined by those skilled in the art based on the physical transmission law of the fault, the contribution of each evidence dimension to the confidence of the warning, and expert engineering experience. The higher the evidence chain completeness score, the more the causal logic of the evidence of the candidate warning event fits the physical mechanism, the more continuous and stable the abnormal temporal evolution, the higher the cross-corroboration strength of multi-source evidence, the higher the completeness and credibility of the overall evidence chain, and the more sufficient the supporting basis for the warning conclusion.

[0065] For each candidate early warning event, historical accuracy matching is performed to obtain a historical accuracy score for each candidate early warning event, specifically including: Extract the fused risk score sequence and key anomaly feature sequence (such as voltage anomaly sequence, temperature anomaly sequence, insulation resistance anomaly sequence) of candidate early warning events within a preset backtracking time window (e.g., 72 hours before the early warning) to construct candidate evolution feature sequences; extract historical fault samples of the same early warning level as the candidate early warning events from a preset historical labeled sample library; for each historical fault sample, extract the labeled evolution feature sequence within a time window of the same length as the backtracking time window before the fault occurred to construct a standard evolution feature sequence set; perform dynamic time warping (DTW) alignment on the candidate evolution feature sequences and each standard evolution feature sequence in the standard evolution feature sequence set; the DTW algorithm finds the optimal warping path to minimize the cumulative distance between the two sequences to solve the problem of the difference in evolution rate between historical fault modes and current candidate events, which is an existing technology and will not be described in detail; normalize the cumulative distance to obtain the historical pattern matching degree: the closer the historical pattern matching degree is to 1, the more similar the evolution trajectory of the current candidate event is to the corresponding historical fault mode; Target fault modes are selected from the historical labeled sample library based on historical pattern matching degree (e.g., selecting the fault modes corresponding to the top-K historical samples with the highest historical pattern matching degree as target fault modes); historical statistical indicators corresponding to the target fault mode are queried, including historical accuracy (the proportion of fault modes that were correctly warned in history), historical false alarm rate (the proportion of fault modes that were falsely warned in history), and historical false negative rate (the proportion of fault modes that were not falsely warned in history); the inverse indicators of historical accuracy, historical false alarm rate, and historical false negative rate are weighted and fused to obtain the historical accuracy score; where the inverse indicator of historical false alarm rate is 1 minus the historical false alarm rate, and the inverse indicator of historical false negative rate is 1 minus the historical false negative rate, and the inverse indicator of historical false negative rate is 1 minus the historical false negative rate. Here, historical accuracy, historical false alarm rate, and historical false negative rate are weighted and fused together to obtain the historical accuracy score; The weights corresponding to the reverse indicators of historical false alarm rate and historical missed alarm rate can be determined by those skilled in the art in conjunction with the system's business priorities for early warning accuracy, false alarm control, and missed alarm control, along with expert engineering experience. These will not be described in detail here, but a weight ratio of 0.4:0.3:0.3 is acceptable. The historical accuracy score characterizes the overall performance level of the target fault mode matched by the current candidate early warning event in the historically labeled sample library, reflecting the historical reliability of the early warning results for this type of fault mode. A higher value indicates a higher early warning accuracy rate, lower false alarm and missed alarm probabilities for the corresponding fault mode in historical practice, a higher degree of fit between the current candidate early warning event and historically validated fault evolution modes, and more sufficient historical experience support for the early warning conclusion.

[0066] Finally, the evidence chain completeness score and historical accuracy score corresponding to each candidate early warning event are weighted and summed to obtain the comprehensive confidence score of each candidate early warning event. The weight of each can be determined by those skilled in the art in combination with the business priority of real-time evidence support and historical experience verification in confidence assessment, and in conjunction with expert engineering experience. The higher the comprehensive confidence score, the higher the comprehensive support of the real-time evidence chain completeness and historical failure mode reliability of the candidate early warning event, the stronger the overall credibility of the early warning conclusion, and the higher the actual probability of the corresponding failure.

[0067] A multi-objective optimization function is constructed, aiming to maximize the comprehensive confidence score Sconf, while simultaneously constraining the response time Tr (e.g., the expected time from warning generation to offline detection completion) and the offline detection cost C (e.g., the number of detection items, the distance between detection sites, and detection man-hours). The multi-objective optimization function can be expressed as:

[0068] in, The upper limit for the set response time limit, To set the upper limit of the offline detection cost budget, Pareto frontier screening is performed on each candidate early warning event according to the multi-objective optimization function. In the three-dimensional objective optimization space composed of (Sconf,Tr,C), the Pareto optimal solution set (i.e., the non-dominated solution set where no other candidate event is better than the event in all three objectives) is selected. Finally, the candidate early warning event with the highest comprehensive confidence score is selected from the Pareto optimal solution set as the final early warning event and is output first. If there are multiple candidate early warning events with the same comprehensive confidence score in the Pareto optimal solution set, the one with the shortest response time Tr is further selected as the final early warning event and is output first. The final early warning event is output to the business processing end to trigger subsequent hierarchical early warning push, offline detection work order generation, and closed-loop feedback.

[0069] By constructing an evidence time-series causal graph for each candidate early warning event, the completeness of the evidence chain is quantified from three dimensions: causal logical consistency, abnormal temporal continuity, and multi-source evidence mutual corroboration. Then, by integrating the historical performance of the corresponding fault modes, a comprehensive confidence score is calculated. This approach strengthens the physical mechanism support and evolutionary rationality of the early warning conclusions through multi-dimensional cross-validation, effectively avoiding the risk of misjudgment caused by a single evaluation dimension. Furthermore, the reliability and interpretability of the early warning results are further improved by relying on empirical verification of historical data. Simultaneously, the quantified confidence results can intuitively distinguish the credibility levels of different fault early warnings, providing a clear quantitative basis for subsequent early warning tiered response and handling decisions, helping to ensure the accuracy of identifying real faults while controlling the false alarm rate.

[0070] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

[0071] All formulas in this manual are dimensionless and calculated numerically. The formulas are derived from software simulations based on a large amount of collected data to obtain the most recent real-world results. The preset parameters and thresholds in the formulas are set by those skilled in the art according to the actual situation.

[0072] Although embodiments of the invention have been shown and described, those skilled in the art will understand that various changes, modifications, substitutions and variations can be made to these embodiments without departing from the principles and spirit of the invention.

Claims

1. A dynamic safety early warning method for new energy vehicles integrating the NEV-AI model, characterized in that the method... include: S1. Collect multi-source heterogeneous vehicle signals from new energy vehicles, encapsulate the multi-source heterogeneous vehicle signals into raw message frames, perform protocol parsing and semantic mapping processing on the raw message frames within a preset time window, and obtain a standardized signal frame set. S2. Perform multiple analysis on each standardized signal frame in the standardized signal frame set to obtain the comprehensive quality vector of each standardized signal frame; based on each comprehensive quality vector, select qualified signal frames from the standardized signal frame set, and construct a reliable time-series signal sequence based on all qualified signal frames; S3. Construct a feature vector based on a reliable time-series signal sequence, perform hierarchical early warning calculation on the feature vector to obtain a fusion risk score, and map the fusion risk score according to a preset hierarchical threshold to obtain the corresponding early warning level; S4. Generate multiple candidate warning events based on the warning level, and perform evidence chain integrity assessment and historical accuracy matching on each candidate warning event to obtain the assessment results and matching results. Based on the evaluation and matching results, the optimal candidate warning event is selected as the final warning event and output.

2. The method according to claim 1, characterized in that, The method for encapsulating multi-source heterogeneous vehicle signals into raw message frames is as follows: Differential sampling is performed on multi-source heterogeneous vehicle signals. The multi-source heterogeneous vehicle signals after differential sampling are aggregated to the vehicle terminal through the vehicle bus. During the aggregation process, multi-source timestamp alignment and bus conflict resolution are performed to obtain the aggregated signal. The aggregated signal is then encapsulated to obtain the original message frame.

3. The method according to claim 1, characterized in that, The method for obtaining the standardized signal frame set is as follows: The protocol type is identified by performing protocol type identification on the original message frame to obtain the protocol type of the original message frame. Based on the protocol type, the fields of the original message frame are extracted to obtain the source protocol field set. Perform GB / T semantic mapping on the source protocol field set to obtain an initial standardized field set. Perform cross-field physical constraint consistency verification on the initial standardized field set to obtain a verified standardized field set. The standardized field set after verification is reassembled with the original message frame to obtain the corresponding standardized signal frame; A set of standardized signal frames is constructed based on all standardized signal frames within a preset time window.

4. The method according to claim 1, characterized in that, The method for obtaining the composite quality vector of each standardized signal frame is as follows: Perform authenticity verification on each standardized signal frame to obtain the authenticity quality component; perform missingness analysis on each standardized signal frame to obtain the missing quality component; perform transmission continuity analysis on each standardized signal frame to obtain the transmission quality component; perform sampling validity analysis on each standardized signal frame to obtain the sampling quality component. The true quality component, missing quality component, transmission quality component, and sampling quality component corresponding to the standardized signal frame are combined to obtain the comprehensive quality vector of each standardized signal frame.

5. The method according to claim 1, characterized in that, The method for constructing a reliable time-series signal sequence is as follows: Perform hierarchical threshold judgment on each standardized signal frame, and form a candidate qualified signal frame set by forming standardized signal frames that pass the hierarchical threshold judgment. Perform quality degradation repair on standardized signal frames that fail the hierarchical threshold judgment and then form a repaired signal frame set. Perform timing quality continuity verification on the candidate qualified signal frame set, and combine the candidate qualified signal frames that pass the timing quality continuity verification into a timing verification signal frame set; The timing verification signal frame set and the repair signal frame set are sorted by timestamp to form an initial reliable timing signal sequence; A quality-weighted sliding smoothing process is performed on the comprehensive quality vector of each signal frame in the initial reliable time-series signal sequence to obtain the final reliable time-series signal sequence.

6. The method according to claim 1, characterized in that, The method for constructing feature vectors is as follows: The reliable time-series signal sequence is divided into short-period subsequences, medium-period subsequences and long-period subsequences, and the working condition adaptive segmentation is performed on each periodic subsequence to obtain multiple working condition aligned subsequences for each periodic subsequence. Hierarchical feature extraction is performed on each working condition aligned subsequence to obtain the basic statistical features, working condition alignment features and multidimensional coupling features corresponding to each working condition aligned subsequence; For each working condition under each periodic subsequence, perform intra-period aggregation on the features of the aligned subsequence to obtain the periodic feature vector corresponding to each periodic subsequence; then, horizontally concatenate the periodic feature vectors according to their dimensions to construct a feature vector; the feature vector includes short-period feature components, medium-period feature components, and long-period feature components.

7. The method according to claim 6, characterized in that, The method for obtaining the fusion risk score is as follows: The short-cycle feature components are input into a preset black-box expert rule base to perform supplementary calculations and output supplementary risk scores. Input the mid-cycle feature components into the preset national standard rule engine to perform preliminary screening calculations and output the preliminary screening risk score; Long-period feature components are input into a pre-defined NEV-AI time-series deep learning model to perform predictive calculations and output progressive fault probability scores. Calculate the fusion weights of the supplementary risk score, the initial screening risk score, and the progressive failure probability score. Multiply the supplementary risk score, the initial screening risk score, and the progressive failure probability score by their respective fusion weights and sum them up to obtain the fusion risk score.

8. The method according to claim 7, characterized in that, The method for obtaining the corresponding warning level is as follows: The system presets a first-level high-risk threshold, a second-level hidden danger threshold, and a third-level attention threshold, and constructs a tiered threshold range, while setting a boundary ambiguity zone between adjacent thresholds; If the fusion risk score falls within the corresponding graded threshold range, the corresponding graded threshold range will be directly mapped to the corresponding warning level. If the fusion risk score falls within the corresponding boundary ambiguity area, calculate the membership pairs of the fusion risk score to the two adjacent warning levels, and determine the corresponding warning level based on the membership pairs.

9. The method according to claim 1, characterized in that, The method for selecting the optimal candidate warning event is as follows: For each candidate early warning event, an evidence chain integrity assessment is performed to obtain an evidence chain integrity score for each candidate early warning event. For each candidate early warning event, historical accuracy matching is performed to obtain the historical accuracy score of each candidate early warning event. The evidence chain completeness score and the historical accuracy score corresponding to each candidate early warning event are weighted and accumulated to obtain the comprehensive confidence score of each candidate early warning event. A multi-objective optimization function is constructed, which aims to maximize the comprehensive confidence score. Based on the multi-objective optimization function, Pareto front screening is performed on each candidate warning event, and the candidate warning event with the highest comprehensive confidence score is selected from the Pareto front as the final warning event.

10. The method according to claim 9, characterized in that, The methods for obtaining the evidence chain completeness score and historical accuracy score are as follows: Construct an evidence time-series causal graph, and calculate the causal structure score of each candidate early warning event based on the evidence time-series causal graph; Obtain the temporal evolution information of each candidate early warning event, and calculate the temporal coherence score of each candidate early warning event based on the temporal evolution information; Obtain multi-source evidence verification information for each candidate early warning event, and calculate the mutual verification reliability score for each candidate early warning event based on the multi-source evidence verification information; The causal structure score, temporal coherence score, and mutual evidence reliability score are weighted and accumulated to obtain the evidence chain integrity score. Construct candidate evolution feature sequences for each candidate early warning event, extract labeled evolution feature sequences of historical fault samples with the same early warning level as the candidate early warning event from a pre-set historical labeled sample library, calculate the similarity between the candidate evolution feature sequences and the labeled evolution feature sequences to obtain the historical pattern matching degree; retrieve the historical accuracy, historical false alarm rate and historical false alarm rate of the corresponding fault pattern based on the historical pattern matching degree. The historical accuracy score is obtained by weighted fusion of the inverse indices of historical accuracy, historical false positive rate, and historical false negative rate; where the inverse indices of historical false positive rate are 1 minus the historical false positive rate, and the inverse indices of historical false negative rate are 1 minus the historical false negative rate.