Battery Pack Anomaly Diagnosis System and Method Based on Visual Acquisition and Intelligent Language Model

Abstract

This invention discloses a battery pack anomaly diagnosis system and method based on a visual acquisition intelligent language model, belonging to the field of battery pack anomaly diagnosis technology. It includes a perception layer, an edge computing layer, a cloud layer, and an interaction layer. Compared with existing technologies, the advantages of this invention are: This application constructs a three-in-one diagnostic system of "vision-electrical-semantic": visual data is collected through multispectral cameras deployed within the battery pack, and BMS electrical parameters are simultaneously accessed. The visual acquisition intelligent language model achieves a complete closed loop of "anomaly feature recognition - risk level determination - natural language interaction," solving the pain points of existing technologies such as "incomplete perception, inaccurate decision-making, and poor interaction."

Description

Technical Field

[0001] This invention relates to the field of battery pack anomaly diagnosis technology, specifically to a battery pack anomaly diagnosis system and method based on a visual acquisition intelligent language model. Background Technology

[0002] Currently, mainstream battery management systems (BMS) can only collect electrical parameters (temperature, voltage, current) at a single point, which has three major drawbacks:

[0003] Limited perception dimensions: It cannot capture "non-electrical anomalies" inside the battery pack, such as visual features like faint smoke, electrolyte leakage, and cell bulging, as well as olfactory features like the irritating odor produced by electrolyte evaporation. These features are often the earliest precursors to thermal runaway (3-5 minutes earlier than temperature changes).

[0004] Rigid decision-making logic: Using fixed threshold judgments (such as alarms when the temperature is >65℃) without considering environmental dynamics (such as the threshold should be lowered by 15% in low-temperature environments), resulting in persistently high false alarm rates (>12%) and false negative rates (>8%).

[0005] Lack of human-computer interaction: It can only output "abnormal codes" and cannot describe abnormal scenarios to users in natural language (such as "Module 3 cell 2 has a 5mm diameter leakage liquid, accompanied by light blue smoke"), which makes it impossible for users and maintenance personnel to quickly judge the risk level.

[0006] Some existing technologies attempt to incorporate computer vision (such as the YOLO model), but they can only achieve single-target detection, lack linkage analysis with electrical parameters, and cannot handle the gradual risk of "accumulated small changes" (such as the smoke area increasing from 0.5cm² to 2cm² within 10 seconds), thus limiting their practicality.

[0007] Based on this, the present invention designs a battery pack anomaly diagnosis system and method based on a visual acquisition intelligent language model to solve the above problems. Summary of the Invention

[0008] To address the aforementioned shortcomings of existing technologies, this invention provides a battery pack anomaly diagnosis system and method based on a visual acquisition intelligent language model.

[0009] To achieve the above objectives, the present invention provides the following technical solution:

[0010] A battery pack anomaly diagnosis system based on a visual acquisition intelligent language model includes:

[0011] Sensing layer: Used for collecting battery pack data and ambient temperature and altitude data of the battery pack, and adjusting sampling standards according to risk level;

[0012] Edge computing layer: The safety threshold is determined by collecting ambient temperature and altitude data of the battery pack within a threshold pool. Based on the visual acquisition intelligent language model, feature extraction is performed on the collected battery pack data. If the extracted features exceed the safety threshold, the risk level of the features is calculated. The sampling standard of the perception layer is adjusted according to the risk level. Combined with the structured report of natural language interaction based on the risk level, it is deployed on the MCU.

[0013] Cloud layer: Connects to the cloud server for storing historical data, updating visual acquisition and intelligent language models, and updating the dynamic threshold pool;

[0014] Interaction layer: Used to receive structured reports of natural language interaction and to output information and execute instructions based on the structured reports of natural language interaction.

[0015] Furthermore, the sensing layer includes an infrared thermal imaging camera, a high-definition visible light camera, a gas sensor, a data acquisition unit, and a monitoring module;

[0016] Infrared thermal imaging camera: Located at the top center of the battery pack, used to collect infrared thermal images of the battery pack;

[0017] High-definition visible light camera: It uses side-view or front-view angles to capture photos of the battery module's shape;

[0018] Gas sensor: Installed at the bottom of the battery pack, it is used to detect the concentration of carbonate gases in the electrolyte at the bottom of the battery pack, reducing false judgments by infrared thermal imaging cameras and high-definition visible light cameras;

[0019] Data acquisition unit: Communicates with BMS to acquire real-time data on individual cell voltage, temperature, and charge / discharge rate of the battery pack, and calculates voltage standard deviation and temperature rise rate;

[0020] The monitoring module includes a temperature sensor and a barometric pressure sensor. Both the temperature sensor and the barometric pressure sensor are installed inside the battery pack installation compartment. The temperature sensor is used to monitor the external environment of the battery pack, and the barometric pressure sensor is used to monitor the external air pressure of the battery pack. The altitude is then determined by the barometric pressure sensor method.

[0021] The monitoring module samples once every 2 seconds.

[0022] Furthermore, the edge computing layer specifically includes a feature extraction submodule and a multimodal fusion decision module;

[0023] Feature extraction submodule: Based on the visual acquisition intelligent language model, features are extracted from the acquired battery pack data;

[0024] Multimodal fusion decision submodule: Determines a safety threshold within a threshold pool by collecting ambient temperature and altitude data of the battery pack. If the extracted features exceed the safety threshold, the risk level of the features is calculated, the sampling standard of the perception layer is adjusted according to the risk level, and a structured report with natural language interaction is generated in conjunction with the risk level.

[0025] Furthermore, the risk levels include warning, alert, and emergency.

[0026] Furthermore, the sampling standards include Level 1 early warning sampling standards, Level 2 alarm sampling standards, and Level 3 emergency sampling standards;

[0027] Level 1 Early Warning Sampling Standard:

[0028] The infrared thermal imaging camera operates at 15fps.

[0029] The infrared thermal imaging camera and the high-definition visible light camera sample once every 1 second, for a continuous 10 frames;

[0030] The data acquisition unit samples charge / discharge rate and temperature data every 500ms, for a total of 5 sets.

[0031] The monitoring module samples once every 2 seconds, for a total of 5 sets;

[0032] Level 2 alarm sampling criteria:

[0033] The infrared thermal imaging camera operates at 30fps.

[0034] The infrared thermal imaging camera and the high-definition visible light camera sample once every 1 second, for a continuous 20 frames;

[0035] The data acquisition unit samples charge / discharge rate and temperature data every 500ms, for a total of 5 sets.

[0036] Gas sensor 500ms / test, 20 consecutive sets;

[0037] Level 3 Emergency Sampling Standard:

[0038] The infrared thermal imaging camera operates at 30fps.

[0039] Infrared thermal imaging cameras and high-definition visible light cameras sample once per second, for 10 or 30 consecutive frames.

[0040] The data acquisition unit samples the charge / discharge rate every 100ms and samples the temperature data in real time, for 10 or 30 consecutive sets.

[0041] Gas sensor 50ms / time, 10 or 30 consecutive sets.

[0042] Furthermore, the interaction layer includes a voice module, a user app, and an operations and maintenance platform.

[0043] A diagnostic method, utilizing a battery pack anomaly diagnostic system based on a visual acquisition intelligent language model, includes the following steps:

[0044] Step 1: Determine the safety threshold within the threshold pool by collecting ambient temperature and altitude data of the battery pack.

[0045] Step 2: The perception layer performs first-level early warning sampling standard acquisition, collecting infrared thermal images of the battery pack, photos of the battery module's shape, and the concentration value of carbonate gases in the electrolyte at the bottom of the battery pack. Using YOLOv8-Small, it extracts the initial features of the battery pack smoke area, the initial features of the battery pack leakage contour, and the initial features of the battery pack bulge displacement from the infrared thermal images and photos of the battery module's shape. It also extracts the initial features of the carbonate gas concentration value in the electrolyte at the bottom of the battery pack. It reads the voltage, temperature, and charge / discharge rate data of the single battery module from the BMS. It calculates the initial features of the single battery module temperature and the initial features of the single battery module's charge / discharge rate data. It then determines whether the initial features of the battery pack smoke area, the initial features of the battery pack leakage contour, the initial features of the battery pack bulge displacement, the initial features of the carbonate gas concentration value in the electrolyte at the bottom of the battery pack, the initial features of the single battery module temperature, or the initial features of the single battery module's charge / discharge rate data exceed a single emergency threshold. If the determination is yes, proceed to step 7; if the determination is no, continue to step 3.

[0046] Step 3: Combine 3D CNN to calculate the initial features of battery pack smoke area, battery pack leakage contour, and battery pack bulge displacement within a continuous sequence of 10 frames. Obtain the change rate of the initial features of battery pack smoke area, battery pack leakage contour, and battery pack bulge displacement. Calculate the initial feature change rate of carbonate gas concentration in the electrolyte at the bottom of the battery pack, the initial feature change rate of single-cell module temperature, and the initial feature change rate of charge / discharge rate data of single-cell module. Identify those with positive growth rates among the initial feature change rates of battery pack smoke area, battery pack leakage contour, battery pack bulge displacement, carbonate gas concentration in the electrolyte at the bottom of the battery pack, single-cell module temperature, and single-cell module charge / discharge rate data. Calculate the first confidence level of the positive growth rate and determine whether the first confidence level of the positive growth rate is less than 70%. If the determination is yes, proceed to Step 2; if the determination is no, proceed to Step 4.

[0047] Step 4: Determine if the first confidence level for positive growth in the rate of change is greater than 80%. If yes, proceed to Step 5. If no, generate a structured report with natural language interaction at the warning level risk level, silently report it to the cloud layer, and then proceed to Step 2.

[0048] Step 5: The perception layer executes the secondary alarm sampling standard to collect data of the initial feature with a positive rate of change, and extracts the abnormal features of the data of the initial feature with a positive rate of change according to the processing logic of Step 2. It judges the abnormal features of the data of the initial feature with a positive rate of change and compares them with a single emergency threshold to determine whether the difference is greater than zero. If the judgment is yes, proceed to Step 7; if the judgment is no, continue to Step 6.

[0049] Step Six: Calculate the first confidence level of the data anomaly features of the initial features with positive rate of change for each feature. Then, use a weighted Bayesian fusion algorithm, input the dual-feature likelihood probability, and finally output the first intermediate confidence level through the visual acquisition intelligent language model. Then, use a time-series correlation analysis algorithm to calculate the first correlation between the first intermediate confidence levels. If the first correlation is greater than 0.7, add 0.1 to the first intermediate confidence level to generate the second confidence level. If the first correlation is less than 0.7, subtract 0.2 from the first intermediate confidence level to generate the second confidence level. Generate a structured report of natural language interaction for the alarm-level risk level. Determine if the second confidence level is greater than 80%. If yes, send the structured report of natural language interaction for the alarm-level risk level to the voice module, user APP, and cloud layer. The voice module plays the structured report of natural language interaction for the alarm-level risk level. Then, execute Step Seven. If no, send the structured report of natural language interaction for the alarm-level risk level to the user APP and cloud layer. Then, execute Step Two.

[0050] Step 7: The perception layer executes the three-level emergency sampling standard to collect initial features with positive growth rate of change, data where the initial features in Step 2 exceed a single emergency threshold, or data where the initial features in Step 5 exceed a single emergency threshold. The three-level emergency sampling standard sets the second intermediate confidence level to 95%. Then, it calculates the second correlation degree between the second intermediate confidence levels through a time series correlation analysis algorithm. The second intermediate confidence level and the second correlation degree are combined to generate the third intermediate confidence level. Then, it calculates the feature change rate through time series analysis. When the change rate increases exponentially, the third intermediate confidence level is increased by 5% to form the third confidence level. Alternatively, when the change rate does not increase exponentially, the third intermediate confidence level is the third confidence level. A structured report with natural language interaction is generated based on the emergency level risk level. An emergency operation plan is generated. It is determined whether the third confidence level is greater than 95%. If the determination is yes, proceed to Step 8. If the determination is no, proceed to Step 9.

[0051] Step 8: The BMS disconnects the high-voltage circuit of the battery pack and sends a structured report with natural language interaction of the emergency risk level to the voice module, user APP, cloud layer and operation and maintenance platform. The voice module then broadcasts the emergency operation plan and the structured report with natural language interaction of the emergency risk level in turn.

[0052] Step Nine: Determine if the third confidence level is greater than or equal to 90%. If the determination is yes, the BMS disconnects the high-voltage circuit of the battery pack. The structured report of the emergency level risk level in natural language interaction is sent to the voice module, the user APP, and the cloud layer. The voice module then broadcasts the emergency operation plan and the structured report of the emergency level risk level in natural language interaction. If the determination is no, the structured report of the emergency level risk level in natural language interaction is sent to the cloud layer, and then Step Five is executed.

[0053] Furthermore, the initial features of the battery pack infrared thermal image include the presence or absence of open flames and the area value of smoke on the outside of the battery pack;

[0054] Initial features of the battery pack's external appearance include the battery pack's shape, cell bulge displacement values, and traces of electrolyte leakage.

[0055] The initial characteristics of single-cell module voltage include the voltage mutation value and the voltage standard deviation value of single-cell module.

[0056] The initial characteristics of the carbonate gas concentration value of the electrolyte at the bottom of the battery pack include the carbonate gas concentration value of the electrolyte at the bottom of the battery pack;

[0057] The initial temperature characteristics of a single battery module include the temperature value of the single battery module, the temperature rise rate of the single battery module, the temperature rise duration of the single battery module, and the temperature difference between any two single battery modules.

[0058] Initial characteristics of charge / discharge rate data for a single battery module: whether there are fluctuations.

[0059] Furthermore, the safety threshold:

[0060] A. The temperature of a single battery module is >60℃ in normal temperature scenarios or >55℃ in high-rate charging scenarios (greater than 3C).

[0061] B. The temperature rise rate of a single battery module is >1℃ / s in normal temperature scenarios or >1℃ / s in high-rate charging scenarios greater than 3C, and the duration is ≥3s;

[0062] C. Temperature difference between multiple single-cell modules > 8℃;

[0063] D. Battery pack indicator shows open flame;

[0064] E. Battery pack display: Smoke concentration > 0.5 mg / m³; diffusion rate > 20% / s;

[0065] F. The displacement of the battery module's cell bulge is >1.5mm at normal temperature or >3mm below 0℃, and continues to increase.

[0066] G. The battery module's exterior photo shows traces of electrolyte leakage;

[0067] H. Single-cell module voltage surge value > 0.05V / 10s;

[0068] I. Standard deviation of single-cell module voltage > 0.03V;

[0069] J. Abnormal fluctuations in the charge / discharge rate of a single battery module;

[0070] K. The initial characteristic of the carbonate gas concentration in the electrolyte at the bottom of the battery pack is >50ppm, and the response time is <1s.

[0071] Compared with the prior art, the beneficial effects of this invention are as follows: This application constructs a three-in-one diagnostic system of "vision-electrical-semantic": visual data is collected by a multispectral camera deployed in the battery pack, and BMS electrical parameters are accessed simultaneously. The visual acquisition intelligent language model realizes a complete closed loop of "abnormal feature recognition-risk level judgment-natural language interaction", which solves the pain points of "incomplete perception, inaccurate decision-making and poor interaction" in the prior art. Attached Figure Description

[0072] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the accompanying drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are merely some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without any creative effort.

[0073] Figure 1 This is a flowchart of the diagnostic method of the present invention. Detailed Implementation

[0074] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.

[0075] Example 1: In some embodiments, a battery pack anomaly diagnosis system based on a visual acquisition intelligent language model includes:

[0076] Sensing layer: Used for collecting battery pack data and ambient temperature and altitude data of the battery pack, and adjusting sampling standards according to risk level;

[0077] Edge computing layer: The safety threshold is determined by collecting ambient temperature and altitude data of the battery pack within a threshold pool. Based on the visual acquisition intelligent language model, feature extraction is performed on the collected battery pack data. If the extracted features exceed the safety threshold, the risk level of the features is calculated. The sampling standard of the perception layer is adjusted according to the risk level. Combined with the structured report of natural language interaction based on the risk level, it is deployed on the MCU.

[0078] The structured report using natural language interaction indicates that the risk level is emergency. The battery pack cell temperature has reached 65°C, exceeding the safety threshold, and there is a risk of thermal runaway. Please slow down immediately!

[0079] Cloud layer: Connects to the cloud server for storing historical data, updating visual acquisition and intelligent language models, and updating the dynamic threshold pool;

[0080] Interaction layer: Used to receive structured reports of natural language interaction and to output information and execute instructions based on the structured reports of natural language interaction.

[0081] The sensing layer includes an infrared thermal imaging camera, a high-definition visible light camera, a gas sensor, a data acquisition unit, and a monitoring module;

[0082] Infrared thermal imaging camera: Located at the top center of the battery pack, used to collect infrared thermal images of the battery pack;

[0083] Infrared thermal imaging camera: Temperature measurement range -40℃~200℃, accuracy ±0.5℃;

[0084] High-definition visible light camera: It uses side-view or front-view angle to capture photos of the battery module's shape. The total number and placement of cameras are determined based on whether all battery cells can be seen.

[0085] The high-definition visible light camera has a resolution of 1080P and an aperture of F1.8.

[0086] Gas sensor: Installed at the bottom of the battery pack, it is used to detect the concentration of carbonate gases in the electrolyte at the bottom of the battery pack, reducing false judgments by infrared thermal imaging cameras and high-definition visible light cameras;

[0087] Gas sensor response time <1s, detection range 0-500ppm (for carbonate gases).

[0088] Data acquisition unit: Communicates with BMS to acquire real-time data on individual cell voltage, temperature, and charge / discharge rate of the battery pack, and calculates voltage standard deviation and temperature rise rate;

[0089] BMS: Battery Management System;

[0090] The monitoring module includes a temperature sensor and a barometric pressure sensor. Both the temperature sensor and the barometric pressure sensor are installed inside the battery pack installation compartment. The temperature sensor is used to monitor the external environment of the battery pack, and the barometric pressure sensor is used to monitor the external air pressure of the battery pack. The altitude is then determined by the barometric pressure sensor method.

[0091] The monitoring module samples once every 2 seconds.

[0092] The edge computing layer specifically includes a feature extraction submodule and a multimodal fusion decision module;

[0093] Feature extraction submodule: Based on the visual acquisition intelligent language model, features are extracted from the acquired battery pack data;

[0094] The computer vision model used is YOLOv8-Small;

[0095] Multimodal fusion decision submodule: Determines a safety threshold within a threshold pool by collecting ambient temperature and altitude data of the battery pack. If the extracted features exceed the safety threshold, the risk level of the features is calculated, the sampling standard of the perception layer is adjusted according to the risk level, and a structured report with natural language interaction is generated in conjunction with the risk level.

[0096] Risk levels include warning, alert, and emergency.

[0097] The sampling standards include Level 1 early warning sampling standards, Level 2 alarm sampling standards, and Level 3 emergency sampling standards;

[0098] Level 1 Early Warning Sampling Standard:

[0099] The infrared thermal imaging camera operates at 15fps.

[0100] The infrared thermal imaging camera and the high-definition visible light camera sample once every 1 second, for a continuous 10 frames;

[0101] The data acquisition unit samples charge / discharge rate and temperature data every 500ms, for a total of 5 sets.

[0102] The monitoring module samples once every 2 seconds, for a total of 5 sets;

[0103] Level 2 alarm sampling criteria:

[0104] The infrared thermal imaging camera operates at 30fps.

[0105] The infrared thermal imaging camera and the high-definition visible light camera sample once every 1 second, for a continuous 20 frames;

[0106] The data acquisition unit samples charge / discharge rate and temperature data every 500ms, for a total of 5 sets.

[0107] Gas sensor 500ms / test, 20 consecutive sets;

[0108] Level 3 Emergency Sampling Standard:

[0109] The infrared thermal imaging camera operates at 30fps.

[0110] Infrared thermal imaging cameras and high-definition visible light cameras sample once per second, for 10 or 30 consecutive frames.

[0111] The data acquisition unit samples the charge / discharge rate every 100ms and samples the temperature data in real time, for 10 or 30 consecutive sets.

[0112] Gas sensor 50ms / time, 10 or 30 consecutive sets.

[0113] The interaction layer includes a voice module, a user app, and an operations and maintenance platform.

[0114] The voice module is located within the voice device of the battery pack-using equipment.

[0115] Battery packs are used in devices including, but not limited to, automobiles and aircraft;

[0116] Please refer to the accompanying drawings in the instruction manual. Figure 1 A diagnostic method utilizing a battery pack anomaly diagnostic system based on a visual acquisition intelligent language model includes the following steps:

[0117] Step 1: Determine the safety threshold within the threshold pool by collecting ambient temperature and altitude data of the battery pack.

[0118] Step 2: The perception layer performs first-level early warning sampling standard acquisition, collecting infrared thermal images of the battery pack, photos of the battery module's shape, and the concentration value of carbonate gases in the electrolyte at the bottom of the battery pack. Using YOLOv8-Small, it extracts the initial features of the battery pack smoke area, the initial features of the battery pack leakage contour, and the initial features of the battery pack bulge displacement from the infrared thermal images and photos of the battery module's shape. It also extracts the initial features of the carbonate gas concentration value in the electrolyte at the bottom of the battery pack. It reads the voltage, temperature, and charge / discharge rate data of the single battery module from the BMS. It calculates the initial features of the single battery module temperature and the initial features of the single battery module's charge / discharge rate data. It then determines whether the initial features of the battery pack smoke area, the initial features of the battery pack leakage contour, the initial features of the battery pack bulge displacement, the initial features of the carbonate gas concentration value in the electrolyte at the bottom of the battery pack, the initial features of the single battery module temperature, or the initial features of the single battery module's charge / discharge rate data exceed a single emergency threshold. If the determination is yes, proceed to step 7; if the determination is no, continue to step 3.

[0119] Initial features of the infrared thermal image of the battery pack include the presence or absence of open flames and the area of ​​smoke on the outside of the battery pack.

[0120] Initial features of the battery pack's external appearance include the battery pack's shape, cell bulge displacement values, and traces of electrolyte leakage.

[0121] The initial characteristics of single-cell module voltage include the voltage mutation value and the voltage standard deviation value of single-cell module.

[0122] The initial characteristics of the carbonate gas concentration value of the electrolyte at the bottom of the battery pack include the carbonate gas concentration value of the electrolyte at the bottom of the battery pack;

[0123] The initial temperature characteristics of a single battery module include the temperature value of the single battery module, the temperature rise rate of the single battery module, the temperature rise duration of the single battery module, and the temperature difference between any two single battery modules.

[0124] Initial characteristics of charge / discharge rate data for a single battery module: whether there are fluctuations;

[0125] Safety threshold:

[0126] A. The temperature of a single battery module is >60℃ in normal temperature scenarios or >55℃ in high-rate charging scenarios (greater than 3C).

[0127] B. The temperature rise rate of a single battery module is >1℃ / s in normal temperature scenarios or >1℃ / s in high-rate charging scenarios greater than 3C, and the duration is ≥3s;

[0128] C. Temperature difference between multiple single-cell modules > 8℃ (reflects cell inconsistency failure, a precursor to thermal runaway);

[0129] D. Battery pack indicator shows open flame;

[0130] E. The battery pack displays a smoke concentration > 0.5 mg / m³; a diffusion rate > 20% / s;

[0131] F. The displacement of the battery module's cell bulge is >1.5mm at normal temperature or >3mm below 0℃, and continues to increase.

[0132] G. The battery module's exterior photo shows traces of electrolyte leakage (pixel-level identification, droplet area > 0.5cm²).

[0133] H. Single-cell module voltage sudden change value > 0.05V / 10s (sudden rise or sudden drop);

[0134] I. Standard deviation of single-cell module voltage > 0.03V (severe imbalance in cell consistency);

[0135] J. Abnormal fluctuation in charge / discharge rate of single battery module (±0.3C change under no-operation conditions);

[0136] K. The initial concentration of carbonate gases in the electrolyte at the bottom of the battery pack is >50ppm and the response time is <1s (confirming the authenticity of the leak).

[0137] Step 3: Combine 3D CNN to calculate the initial features of battery pack smoke area, battery pack leakage contour, and battery pack bulge displacement within a continuous sequence of 10 frames. Obtain the change rate of the initial features of battery pack smoke area, battery pack leakage contour, and battery pack bulge displacement. Calculate the initial feature change rate of carbonate gas concentration in the electrolyte at the bottom of the battery pack, the initial feature change rate of single-cell module temperature, and the initial feature change rate of charge / discharge rate data of single-cell module. Identify those with positive growth rates among the initial feature change rates of battery pack smoke area, battery pack leakage contour, battery pack bulge displacement, carbonate gas concentration in the electrolyte at the bottom of the battery pack, single-cell module temperature, and single-cell module charge / discharge rate data. Calculate the first confidence level of the positive growth rate and determine whether the first confidence level of the positive growth rate is less than 70%. If the determination is yes, proceed to Step 2; if the determination is no, proceed to Step 4.

[0138] First confidence level The specific calculations are as follows:

[0139]

[0140]

[0141]

[0142] The first confidence level; Calculate the coefficients for the confidence level; The confidence score is the base score for the smoke area of ​​the smoke battery pack. The confidence score of the bulge displacement is 0.95. It is 0.7; Add points for timing consistency; The variance of the rate of change over 10 consecutive frames.

[0143] Step 4: Determine if the first confidence level for positive growth in the rate of change is greater than 80%. If yes, proceed to Step 5. If no, generate a structured report with natural language interaction at the warning level risk level, silently report it to the cloud layer, and then proceed to Step 2.

[0144] Step 5: The perception layer executes the secondary alarm sampling standard to collect data of the initial feature with a positive rate of change, and extracts the abnormal features of the data of the initial feature with a positive rate of change according to the processing logic of Step 2. It judges the abnormal features of the data of the initial feature with a positive rate of change and compares them with a single emergency threshold to determine whether the difference is greater than zero. If the judgment is yes, proceed to Step 7; if the judgment is no, continue to Step 6.

[0145] Step Six: Calculate the first confidence level of the data anomaly features of the initial features with positive rate of change for each feature. Then, use a weighted Bayesian fusion algorithm, input the dual-feature likelihood probability, and finally output the first intermediate confidence level through the visual acquisition intelligent language model. Then, use a time-series correlation analysis algorithm to calculate the first correlation between the first intermediate confidence levels. If the first correlation is greater than 0.7, add 0.1 to the first intermediate confidence level to generate the second confidence level. If the first correlation is less than 0.7, subtract 0.2 from the first intermediate confidence level to generate the second confidence level. Generate a structured report of natural language interaction for the alarm-level risk level. Determine if the second confidence level is greater than 80%. If yes, send the structured report of natural language interaction for the alarm-level risk level to the voice module, user APP, and cloud layer. The voice module plays the structured report of natural language interaction for the alarm-level risk level. Then, execute Step Seven. If no, send the structured report of natural language interaction for the alarm-level risk level to the user APP and cloud layer. Then, execute Step Two.

[0146] The core of weight setting is to allocate weights based on the "risk contribution" and "detection reliability" of features, ensuring that high-risk, high-reliability features receive higher weights. The specific rules are as follows:

[0147]

[0148] Example of dual-feature fusion weight calculation:

[0149] Scene: Smoke characteristics (visual) + Temperature rise characteristics (electrical)

[0150] Smoke characteristics: YOLOv8 recognition confidence level 92%, rate of change 12% / s → weight = 0.4 + 0.1 + 0.05 = 0.55

[0151] Temperature characteristics: Acquisition accuracy 0.3%, exceeding threshold 30% → Weight = 0.4 + 0.1 + 0.05 = 0.55

[0152] Weight normalization: Smoke weight = 0.55 / (0.55+0.55) = 0.5; Temperature weight = 0.55 / (0.55+0.55) = 0.5;

[0153] Methods for obtaining likelihood probability

[0154] Likelihood probability represents "the probability of an anomaly occurring when a certain feature appears," and is obtained through historical case statistics + visual data acquisition and intelligent language model scene correction. Specific steps include:

[0155] Basic Likelihood Probability Statistics

[0156] From over 100,000 anomaly cases, the probability of association between a single feature and anomaly results was statistically analyzed:

[0157] Smoke area > 1cm 2 At that time, the probability of the precursor to thermal runaway is 85% → the likelihood probability of smoke characteristics P(smoke|anomaly) = 0.85

[0158] When the rate of temperature rise is >3℃ / min, the probability of thermal runaway precursors is 78% → temperature characteristic likelihood probability P(temperature|anomaly) = 0.78

[0159] When the gas concentration is >100 ppm, the probability of electrolyte leakage is 90% → the gas characteristic likelihood probability P(gas|anomaly) = 0.90

[0160] Visual acquisition, intelligent language model, scene correction

[0161] The likelihood probability is adjusted based on real-time scenario (ambient temperature, altitude, charging rate), and the formula is revised accordingly:

[0162]

[0163] This is a scene correction factor (values ​​range from ±0.1 to ±0.3).

[0164] S is the scene impact factor (e.g., S=0.2 for high-rate charging, S=-0.1 for low-temperature environment).

[0165] Example: In a high-rate charging scenario, the basic likelihood probability of the temperature feature is 0.78, the correction coefficient k=0.2, and the scenario factor S=0.2. Then the corrected likelihood probability Pcorrected=0.78×(1+0.2×0.2)=0.8112.

[0166] Combination of two feature likelihood probabilities

[0167] When fusing two features, the "joint likelihood probability" is used for calculation:

[0168]

[0169] The correlation coefficient represents the historical co-occurrence probability of two features (e.g., the correlation coefficient between smoke and temperature is 0.9).

[0170] First intermediate confidence level calculation process

[0171] Input weights of two features , Corrected likelihood probability , ;

[0172] Weighted fusion calculation:

[0173]

[0174] The visual acquisition intelligent language model, combined with temporal stability (number of consecutive frames of feature occurrence), outputs the first intermediate confidence score:

[0175] First intermediate confidence score = P fusion * temporal stability score (temporal stability score: 1.0 for features appearing consecutively for 10 frames, 0.8 for features appearing consecutively for 5 frames)

[0176] Temporal correlation analysis algorithm: First correlation degree calculation method

[0177] The first correlation is used to quantify the temporal synchronicity of the two features, determining whether the two features are caused by the same anomalous event (rather than accidental superposition). The core calculation logic is as follows:

[0178] 1. Data preprocessing: Feature temporal alignment

[0179] Align the sampled data with dual features according to the timestamp to ensure consistency in the time dimension:

[0180] Visual features: Sampled once every 1 second for 20 consecutive frames (time span 20 seconds).

[0181] Electrical features: Sampled once every 500ms, for 40 consecutive groups (time span 20s), interpolated to 1 data point per 1s, aligned with visual features.

[0182] 2. Core Calculation Steps (Based on an Improved Pearson Correlation Coefficient)

[0183] Extracting feature change sequences

[0184] Calculate the time-series variation sequences of the two features respectively. and , where n is the number of sampling points (n=20 for level 2 alarms).

[0185] Calculate the serial correlation coefficient

[0186] An improved Pearson correlation coefficient is used to quantify the synchronicity of two changing sequences:

[0187]

[0188] The value of r ranges from [-1, 1]. The closer r is to 1, the more synchronous the changes of the two features are; the closer r is to 0, the less obvious the synchronicity is.

[0189] Relevance determination rules and confidence adjustment

[0190] First correlation score > 0.7: Determine that the two features are caused by the same anomaly, and increase the first intermediate confidence score by 0.1 to generate the second confidence score;

[0191] 0.3 ≤ First correlation score ≤ 0.7: Determine a weak correlation between the two features, and the first intermediate confidence score remains unchanged;

[0192] First correlation < 0.3: Determine the two features as accidental superposition, and generate the second confidence level by subtracting 0.2 from the first intermediate confidence level.

[0193] Step 7: The perception layer executes the three-level emergency sampling standard to collect initial features with positive growth rate of change, data where the initial features in Step 2 exceed a single emergency threshold, or data where the initial features in Step 5 exceed a single emergency threshold. The three-level emergency sampling standard sets the second intermediate confidence level to 95%. Then, it calculates the second correlation degree between the second intermediate confidence levels through a time series correlation analysis algorithm. The second intermediate confidence level and the second correlation degree are combined to generate the third intermediate confidence level. Then, it calculates the feature change rate through time series analysis. When the change rate increases exponentially, the third intermediate confidence level is increased by 5% to form the third confidence level. Alternatively, when the change rate does not increase exponentially, the third intermediate confidence level is the third confidence level. A structured report with natural language interaction is generated based on the emergency level risk level. An emergency operation plan is generated. It is determined whether the third confidence level is greater than 95%. If the determination is yes, proceed to Step 8. If the determination is no, proceed to Step 9.

[0194] The second correlation essentially quantifies the temporal synchronicity and causal correlation of multiple high-confidence anomalous features (such as smoke + temperature + gas) under Level 3 emergency sampling, and determines whether these features are caused by the same thermal runaway event (rather than independent and sporadic anomalies).

[0195] Complete calculation process of the time-series correlation analysis algorithm for the second degree of correlation

[0196] Input data: Multi-feature second intermediate confidence sequence collected according to the Level 3 emergency sampling standard (denoted as...) , (Characteristic quantities, such as smoke, temperature, gas, voltage, etc.).

[0197] Step-by-step calculation process

[0198] Step 1: Data preprocessing (time-series alignment + standardization)

[0199] Time series alignment: Align the second intermediate confidence sequences of all features to the smallest sampling granularity (50ms for level 3 emergency sampling) according to the timestamp, forming an equal-length time series array:

[0200] Example:

[0201] Smoke feature confidence sequence [0.95, 0.95, 0.96, 0.98, 1.00] (50ms / time, 5 time points in total);

[0202] Temperature feature confidence sequence [0.95, 0.96, 0.97, 0.99, 1.00] (Real-time sampling interpolation is 50ms granularity);

[0203] Gas feature confidence sequence [0.95, 0.95, 0.97, 0.98, 0.99] (50ms / time).

[0204] Standardization: Eliminates differences in absolute confidence scores, focuses on trends, formula:

[0205]

[0206] in:

[0207] : No. The feature in the first The second median confidence level at a given time point;

[0208] Standardized confidence level (value 0~1);

[0209] : No. The minimum / maximum value of the confidence sequence of each feature.

[0210] Example (smoke characteristics):

[0211] → , .

[0212] Step 2: Calculate the temporal trend similarity of feature pairs (core)

[0213] For each pair of features (such as smoke-temperature, smoke-gas, temperature-gas), the "similarity of change trend" is calculated using an improved Dynamic Time Warping (DTW) algorithm (adapted to data of unequal length / variable rate, superior to the traditional Pearson coefficient):

[0214] Constructing the temporal variation sequence of feature pairs:

[0215]

[0216] ( , hour )

[0217] Example (smoke-temperature):

[0218] : [0, 0, 0.2, 0.4, 0.4];

[0219] [0, 0.2, 0.2, 0.4, 0.2].

[0220] Calculate the DTW distance (a measure of similarity between two sequences; the smaller the value, the more similar they are):

[0221] Construct the distance matrix :

[0222]

[0223] Features The point in time, Features The point in time;

[0224] Dynamic programming to find the minimum cumulative distance:

[0225]

[0226] Normalized DTW distance is the similarity:

[0227]

[0228] ( (1 indicates complete synchronization, 0 indicates complete unrelatedness)

[0229] Example: DTW distance of smoke-temperature = 0.2 → → .

[0230] Step 3: Calculate the comprehensive second correlation degree of multiple features;

[0231] The weighted similarity of all feature pairs is fused to obtain the final second correlation score, as shown in the formula:

[0232]

[0233] in: Weights of feature pairs (based on their contribution to thermal runaway, with core features having higher weights).

[0234] Weighting rules (for battery pack scenarios):

[0235]

[0236] Example calculation (3 features: smoke, temperature, gas):

[0237]

[0238]

[0239]

[0240] Weighted sum: ;

[0241] .

[0242] Step 4: Correlation Correction (Adapting to Exponentially Changing Features)

[0243] The second correlation degree has an added correction term:

[0244]

[0245] in:

[0246] Correction coefficient (values ​​range from 0.1 to 0.2, derived from the visual acquisition intelligent language model trained on over 100,000 cases).

[0247] Exponential change indicator ( This indicates that at least one feature changes at an exponential rate. (Then none);

[0248] Exponential change criterion: The rate of change of the characteristic satisfies And it holds true for three consecutive time points.

[0249] Example:

[0250] The smoke characteristics change exponentially ( ), → .

[0251] Computational Example (Complete Closed Loop)

[0252] Assume that 30 sets of data are collected under Level 3 emergency sampling (50ms / sample, total 1.5s), with characteristics of smoke, temperature, and gas:

[0253] After time series alignment, three confidence sequences of length 30 are obtained, and after standardization, the trend of change is focused.

[0254] Calculate feature similarity to DTW: , , ;

[0255] Weighted calculation of basic correlation degree: ;

[0256] The smoke characteristics were determined to be exponential changes. ), → Corrected ;

[0257] The final correlation coefficient was 0.941, reflecting a high degree of synchronization among the three characteristics, confirming that they were caused by the same thermal runaway event.

[0258] Step 8: The BMS disconnects the high-voltage circuit of the battery pack and sends a structured report with natural language interaction of the emergency risk level to the voice module, user APP, cloud layer and operation and maintenance platform. The voice module then broadcasts the emergency operation plan and the structured report with natural language interaction of the emergency risk level in turn.

[0259] Step Nine: Determine if the third confidence level is greater than or equal to 90%. If the determination is yes, the BMS disconnects the high-voltage circuit of the battery pack. The structured report of the emergency level risk level in natural language interaction is sent to the voice module, the user APP, and the cloud layer. The voice module then broadcasts the emergency operation plan and the structured report of the emergency level risk level in natural language interaction. If the determination is no, the structured report of the emergency level risk level in natural language interaction is sent to the cloud layer, and then Step Five is executed.

[0260] This application constructs a three-in-one diagnostic system of "vision-electricity-semantics": visual data is collected by a multispectral camera deployed in the battery pack, and electrical parameters of the BMS are accessed simultaneously. The visual acquisition intelligent language model realizes a closed loop of "abnormal feature recognition-risk level determination-natural language interaction", which solves the pain points of existing technologies such as "incomplete perception, inaccurate decision-making and poor interaction".

[0261] The above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions will not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A battery pack anomaly diagnosis system based on a visual acquisition intelligent language model, characterized in that: include: Sensing layer: Used for collecting battery pack data and ambient temperature and altitude data of the battery pack, and adjusting sampling standards according to risk level; Edge computing layer: The safety threshold is determined by collecting ambient temperature and altitude data of the battery pack within a threshold pool. Based on the visual acquisition intelligent language model, feature extraction is performed on the collected battery pack data. If the extracted features exceed the safety threshold, the risk level of the features is calculated. The sampling standard of the perception layer is adjusted according to the risk level. Combined with the structured report of natural language interaction based on the risk level, it is deployed on the MCU. Cloud layer: Connects to the cloud server for storing historical data, updating visual acquisition and intelligent language models, and updating the dynamic threshold pool; Interaction layer: Used to receive structured reports of natural language interaction and to output information and execute instructions based on the structured reports of natural language interaction.

2. The battery pack anomaly diagnosis system based on a visual acquisition intelligent language model according to claim 1, characterized in that, The sensing layer includes an infrared thermal imaging camera, a high-definition visible light camera, a gas sensor, a data acquisition unit, and a monitoring module; Infrared thermal imaging camera: Located at the top center of the battery pack, used to collect infrared thermal images of the battery pack; High-definition visible light camera: It uses side-view or front-view angles to capture photos of the battery module's shape; Gas sensor: Installed at the bottom of the battery pack, it is used to detect the concentration of carbonate gases in the electrolyte at the bottom of the battery pack, reducing false judgments by infrared thermal imaging cameras and high-definition visible light cameras; Data acquisition unit: Communicates with BMS to acquire real-time data on individual cell voltage, temperature, and charge / discharge rate of the battery pack, and calculates voltage standard deviation and temperature rise rate; The monitoring module includes a temperature sensor and a barometric pressure sensor. Both the temperature sensor and the barometric pressure sensor are installed inside the battery pack installation compartment. The temperature sensor is used to monitor the external environment of the battery pack, and the barometric pressure sensor is used to monitor the external air pressure of the battery pack. The altitude is then determined by the barometric pressure sensor method. The monitoring module samples once every 2 seconds.

3. The battery pack anomaly diagnosis system based on a visual acquisition intelligent language model according to claim 1, characterized in that, The edge computing layer specifically includes a feature extraction submodule and a multimodal fusion decision module; Feature extraction submodule: Based on the visual acquisition intelligent language model, features are extracted from the acquired battery pack data; Multimodal fusion decision submodule: Determines a safety threshold within a threshold pool by collecting ambient temperature and altitude data of the battery pack. If the extracted features exceed the safety threshold, the risk level of the features is calculated, the sampling standard of the perception layer is adjusted according to the risk level, and a structured report with natural language interaction is generated in conjunction with the risk level.

4. The battery pack anomaly diagnosis system based on a visual acquisition intelligent language model according to claim 3, characterized in that, Risk levels include warning, alert, and emergency.

5. A battery pack anomaly diagnosis system based on a visual acquisition intelligent language model according to claim 4, characterized in that, The sampling standards include Level 1 early warning sampling standards, Level 2 alarm sampling standards, and Level 3 emergency sampling standards; Level 1 Early Warning Sampling Standard: The infrared thermal imaging camera operates at 15fps. The infrared thermal imaging camera and the high-definition visible light camera sample once every 1 second, for a continuous 10 frames; The data acquisition unit samples charge / discharge rate and temperature data every 500ms, for a total of 5 sets. The monitoring module samples once every 2 seconds, for a total of 5 sets; Level 2 alarm sampling criteria: The infrared thermal imaging camera operates at 30fps. The infrared thermal imaging camera and the high-definition visible light camera sample once every 1 second, for a continuous 20 frames; The data acquisition unit samples charge / discharge rate and temperature data every 500ms, for a total of 5 sets. Gas sensor 500ms / test, 20 consecutive sets; Level 3 Emergency Sampling Standard: The infrared thermal imaging camera operates at 30fps. Infrared thermal imaging cameras and high-definition visible light cameras sample once per second, for 10 or 30 consecutive frames. The data acquisition unit samples the charge / discharge rate every 100ms and samples the temperature data in real time, for 10 or 30 consecutive sets. Gas sensor 50ms / time, 10 or 30 consecutive sets.

6. A battery pack anomaly diagnosis system based on a visual acquisition intelligent language model according to claim 5, characterized in that, The interaction layer includes a voice module, a user app, and an operations and maintenance platform.

7. A diagnostic method, utilizing the battery pack anomaly diagnostic system based on a visual acquisition intelligent language model as described in claim 6, characterized in that, Includes the following steps: Step 1: Determine the safety threshold within the threshold pool by collecting ambient temperature and altitude data of the battery pack. Step 2: The perception layer performs first-level early warning sampling standard acquisition, collecting infrared thermal images of the battery pack, photos of the battery module's shape, and the concentration value of carbonate gases in the electrolyte at the bottom of the battery pack. Using YOLOv8-Small, it extracts the initial features of the battery pack smoke area, the initial features of the battery pack leakage contour, and the initial features of the battery pack bulge displacement from the infrared thermal images and photos of the battery module's shape. It also extracts the initial features of the carbonate gas concentration value in the electrolyte at the bottom of the battery pack. It reads the voltage, temperature, and charge / discharge rate data of the single battery module from the BMS. It calculates the initial features of the single battery module temperature and the initial features of the single battery module's charge / discharge rate data. It then determines whether the initial features of the battery pack smoke area, the initial features of the battery pack leakage contour, the initial features of the battery pack bulge displacement, the initial features of the carbonate gas concentration value in the electrolyte at the bottom of the battery pack, the initial features of the single battery module temperature, or the initial features of the single battery module's charge / discharge rate data exceed a single emergency threshold. If the determination is yes, proceed to step 7; if the determination is no, continue to step 3. Step 3: Combine 3D CNN to calculate the initial features of battery pack smoke area, battery pack leakage contour, and battery pack bulge displacement within a continuous sequence of 10 frames. Obtain the change rate of the initial features of battery pack smoke area, battery pack leakage contour, and battery pack bulge displacement. Calculate the initial feature change rate of carbonate gas concentration in the electrolyte at the bottom of the battery pack, the initial feature change rate of single-cell module temperature, and the initial feature change rate of charge / discharge rate data of single-cell module. Identify those with positive growth rates among the initial feature change rates of battery pack smoke area, battery pack leakage contour, battery pack bulge displacement, carbonate gas concentration in the electrolyte at the bottom of the battery pack, single-cell module temperature, and single-cell module charge / discharge rate data. Calculate the first confidence level of the positive growth rate and determine whether the first confidence level of the positive growth rate is less than 70%. If the determination is yes, proceed to Step 2; if the determination is no, proceed to Step 4. Step 4: Determine if the first confidence level for positive growth in the rate of change is greater than 80%. If yes, proceed to Step 5. If no, generate a structured report with natural language interaction at the warning level risk level, silently report it to the cloud layer, and then proceed to Step 2. Step 5: The perception layer executes the secondary alarm sampling standard to collect data of the initial feature with a positive rate of change, and extracts the abnormal features of the data of the initial feature with a positive rate of change according to the processing logic of Step 2. It judges the abnormal features of the data of the initial feature with a positive rate of change and compares them with a single emergency threshold to determine whether the difference is greater than zero. If the judgment is yes, proceed to Step 7; if the judgment is no, continue to Step 6. Step Six: Calculate the first confidence level of the data anomaly features for each initial feature with a positive rate of change. Then, use a weighted Bayesian fusion algorithm, input the dual-feature likelihood probability, and finally output the first intermediate confidence level through the visual acquisition intelligent language model. Then, use a time-series correlation analysis algorithm to calculate the first correlation between the first intermediate confidence levels. If the first correlation is greater than 0.7, add 0.1 to the first intermediate confidence level to generate the second confidence level. If the first correlation is less than 0.7, subtract 0.2 from the first intermediate confidence level to generate the second confidence level. Generate a structured report of natural language interaction for the alarm-level risk level. Determine if the second confidence level is greater than 80%. If yes, send the structured report of natural language interaction for the alarm-level risk level to the voice module, user APP, and cloud layer. The voice module plays the structured report of natural language interaction for the alarm-level risk level. Then execute Step Seven. If no, send the structured report of natural language interaction for the alarm-level risk level to the user APP and cloud layer. Then execute Step Five. Step 7: The perception layer executes the three-level emergency sampling standard to collect initial features with positive growth rate of change, data where the initial features in Step 2 exceed a single emergency threshold, or data where the initial features in Step 5 exceed a single emergency threshold. The three-level emergency sampling standard sets the second intermediate confidence level to 95%. Then, it calculates the second correlation degree between the second intermediate confidence levels through a time series correlation analysis algorithm. The second intermediate confidence level and the second correlation degree are combined to generate the third intermediate confidence level. Then, it calculates the feature change rate through time series analysis. When the change rate increases exponentially, the third intermediate confidence level is increased by 5% to form the third confidence level. Alternatively, when the change rate does not increase exponentially, the third intermediate confidence level is the third confidence level. A structured report with natural language interaction is generated based on the emergency level risk level. An emergency operation plan is generated. It is determined whether the third confidence level is greater than 95%. If the determination is yes, proceed to Step 8. If the determination is no, proceed to Step 9. Step 8: The BMS disconnects the high-voltage circuit of the battery pack and sends a structured report with natural language interaction of the emergency risk level to the voice module, user APP, cloud layer and operation and maintenance platform. The voice module then broadcasts the emergency operation plan and the structured report with natural language interaction of the emergency risk level in turn. Step Nine: Determine if the third confidence level is greater than or equal to 90%. If the determination is yes, the BMS disconnects the high-voltage circuit of the battery pack. The structured report of the emergency level risk level in natural language interaction is sent to the voice module, the user APP, and the cloud layer. The voice module then broadcasts the emergency operation plan and the structured report of the emergency level risk level in natural language interaction. If the determination is no, the structured report of the emergency level risk level in natural language interaction is sent to the cloud layer, and then Step Seven is executed.

8. The diagnostic method according to claim 7, characterized in that, Initial features of the infrared thermal image of the battery pack include the presence or absence of open flames and the area of ​​smoke on the outside of the battery pack. Initial features of the battery pack's external appearance include the battery pack's shape, cell bulge displacement values, and traces of electrolyte leakage. The initial characteristics of single-cell module voltage include the voltage mutation value and the voltage standard deviation value of single-cell module. The initial characteristics of the carbonate gas concentration value of the electrolyte at the bottom of the battery pack include the carbonate gas concentration value of the electrolyte at the bottom of the battery pack; The initial temperature characteristics of a single battery module include the temperature value of the single battery module, the temperature rise rate of the single battery module, the temperature rise duration of the single battery module, and the temperature difference between any two single battery modules. Initial characteristics of charge / discharge rate data for a single battery module: whether there are fluctuations.

9. The diagnostic method according to claim 7, characterized in that, Safety threshold: A. The temperature of a single battery module is >60℃ in normal temperature scenarios or >55℃ in high-rate charging scenarios (greater than 3C). B. The temperature rise rate of a single battery module is >1℃ / s in normal temperature scenarios or >1℃ / s in high-rate charging scenarios greater than 3C, and the duration is ≥3s; C. Temperature difference between multiple single-cell modules > 8℃; D. Battery pack indicator shows open flame; E. Battery pack display: Smoke concentration > 0.5 mg / m³; diffusion rate > 20% / s; F. The displacement of the battery module's cell bulge is >1.5mm at normal temperature or >3mm below 0℃, and continues to increase. G. The battery module's exterior photo shows traces of electrolyte leakage; H. Single-cell module voltage surge value > 0.05V / 10s; I. Standard deviation of single-cell module voltage > 0.03V; J. Abnormal fluctuations in the charge / discharge rate of a single battery module; K. The initial characteristic of the carbonate gas concentration in the electrolyte at the bottom of the battery pack is >50ppm, and the response time is <1s.

Images