A multi-modal flame-retardant detection method for an electric bicycle battery compartment material

Abstract

The application discloses a kind of electric bicycle battery compartment material multimode flame-retardant detection methods, it is related to electric bicycle battery safety detection technical field.By deploying multidimensional sensor network inside battery compartment, real-time collection multidimensional physical state parameters;After pre-processing by edge computing unit, extract expansion force change rate, temperature change rate, voltage inconsistency coefficient, hydrogen concentration change rate and pressure-temperature coupling characteristic ratio;Adopt double-path parallel discrimination strategy, through adaptive dynamic threshold value to carry out rapid screening, combined with mahalanobis distance and random forest classifier to carry out accurate discrimination, and after time consistency check, execute grading alarm decision;While uploading event data to cloud, realize benchmark update and model iteration.Can realize high sensitivity early warning before heat runaway early several minutes, simultaneously through multi-feature fusion and time consistency check effectively reduce false alarm rate, can be widely applied in electric bicycle battery safety monitoring field.

Description

Technical Field

[0001] This invention relates to the field of electric bicycle battery safety testing technology, and in particular to a multimodal flame retardant testing method for electric bicycle battery compartment materials. Background Technology

[0002] With the rapid increase in the number of electric bicycles, fires caused by battery thermal runaway are on the rise, causing significant losses to public life and property. The flame-retardant properties of the battery compartment materials are a key factor determining the speed of fire spread. The current mandatory national standard GB 17761-2024 requires that the flammability rating of non-metallic materials in the battery compartment must reach V-0.

[0003] However, traditional flame retardancy testing of battery compartment materials mainly relies on vertical combustion tests based on the GB / T 5169.16 standard, which are conducted under static conditions using standard specimens. In actual use, batteries undergo multiple dynamic conditions such as charging and discharging heat generation, mechanical vibration, and external impact. The flame retardancy performance of materials under dynamic conditions may differ significantly from the static test results. Furthermore, the physical signals generated in the early stages of lithium battery thermal runaway (before the safety valve ruptures), such as changes in cell expansion force and the release of trace characteristic gases, are very weak. Traditional smoke detection and temperature threshold alarm methods are difficult to effectively identify these signals in the early stages and often only respond after an open flame is generated, thus missing the best opportunity for intervention.

[0004] To address the aforementioned issues, existing research has attempted to introduce multi-source sensor fusion technology for fire monitoring. However, in the specific application scenario of electric bicycle battery compartments, a multimodal flame-retardant detection scheme that can balance high sensitivity and low false alarm rate while adapting to dynamic operating conditions is still lacking. Therefore, this paper proposes a multimodal flame-retardant detection method for electric bicycle battery compartment materials. Summary of the Invention

[0005] This invention provides a multimodal flame retardant detection method and system for electric bicycle battery compartment materials, aiming to solve the technical problems in the prior art of static detection being disconnected from dynamic working conditions, difficulty in identifying early weak signs, and lack of dynamic evolution analysis capabilities.

[0006] To achieve the above objectives, the technical solution adopted by the present invention is as follows: In a first aspect, the present invention provides a multimodal flame retardancy testing method for battery compartment materials of electric bicycles, comprising the following steps: Step S1: Deployment and Data Acquisition of Multidimensional Sensor Network A multi-dimensional sensor network is deployed inside the battery compartment to collect various physical state parameters of the battery in real time at a preset sampling frequency.

[0007] The physical state parameters include at least the cell expansion force, cell surface temperature, single cell voltage, and hydrogen concentration inside the battery compartment.

[0008] The expansion force signal is collected by a flexible thin-film pressure sensor arranged between adjacent cells and between the cell and the battery compartment wall, with a sampling frequency of not less than 100Hz, to capture rapid changes in the cell expansion force.

[0009] Step S2: Edge Data Preprocessing The edge computing unit performs time synchronization, filtering and noise reduction, outlier removal and normalization preprocessing on the collected raw physical state parameters to obtain standardized time series data.

[0010] Preprocessing specifically includes: Kalman filtering is used to suppress noise in the signals from each sensor, eliminating high-frequency measurement noise. Based on the physical limit range of each physical quantity, obviously abnormal bad values ​​are eliminated; Normalize all physical quantities to a unified dimension to eliminate the impact of differences in the dimensions of different sensors on subsequent fusion algorithms.

[0011] Step S3: Extraction of multidimensional feature parameters Multidimensional feature parameters are extracted in real time based on standardized time-series data.

[0012] The multidimensional feature parameters include at least: Expansion force change rate: obtained by taking the first derivative of the expansion force signal, reflecting the gas generation rate inside the battery cell; Temperature change rate: obtained by taking the first derivative of the temperature signal, reflecting the rate of heat accumulation; Voltage inconsistency coefficient: obtained by calculating the standard deviation of the voltage of all individual cells, it characterizes the degree of inconsistency degradation between cells and is an indirect sign of internal short circuits; Hydrogen concentration change rate: obtained by taking the first derivative of the hydrogen concentration signal, reflecting the rate of electrolyte decomposition and gas production after the diaphragm ruptures; Pressure-temperature coupling characteristic ratio: This ratio is calculated as the ratio of the rate of change of expansion force to the rate of change of temperature, used to characterize the degree to which internal gas generation precedes heat accumulation. If this ratio increases abnormally while the rate of change of temperature does not change significantly, it indicates that internal gas generation (mechanical / chemical causes) precedes heat accumulation, which is a typical early mechanical precursor fingerprint of thermal runaway.

[0013] Step S4: Parallel Anomaly Detection via Dual Paths A dual-path parallel discrimination strategy is employed to determine anomalies in multi-dimensional feature parameters, including: Rapid screening path: Adaptive dynamic thresholds are calculated for the rate of change in expansion force and the rate of change in hydrogen concentration, respectively. The adaptive dynamic threshold is calculated as follows: using historical data within a preset time period as a sliding window, the mean and standard deviation of the target signal within the sliding window are calculated. The sum of the mean and a preset multiple of the standard deviation is used as the dynamic threshold at the current moment. This threshold is adaptively adjusted according to individual battery differences and changes in operating conditions. If any rate of change exceeds the corresponding dynamic threshold, a suspected anomaly marker is triggered.

[0014] Precise path identification: Construct a feature vector from multi-dimensional feature parameters and perform dual verification: First verification: Calculate the Mahalanobis distance of the feature vector relative to the normal state benchmark. The normal state benchmark is estimated from the historical feature vector data of the battery during its normal operation period. The Mahalanobis distance is obtained by calculating the square root of the weighted sum of squares of the difference vector between the feature vector and the mean vector of the normal state benchmark, and the inverse matrix of the covariance matrix of the normal state benchmark. If the Mahalanobis distance exceeds a first threshold, an anomaly is determined. The first threshold is determined by a preset quantile of the chi-square distribution.

[0015] The second verification step involves inputting the feature vector into a pre-trained random forest classifier to obtain the anomaly probability. The random forest classifier is an ensemble of multiple decision trees, each independently performing a binary classification of normal or abnormal. The anomaly probability is the proportion of decision trees classified as abnormal out of the total number of trees. If the anomaly probability exceeds a second threshold, the tree is classified as abnormal.

[0016] When the Mahalanobis distance exceeds the first threshold or the anomaly probability exceeds the second threshold, precise anomaly labeling is triggered.

[0017] The advantages of the dual-path parallel design are: rapid path screening ensures high sensitivity to weak signals, and accurate path identification reduces the false alarm rate through multi-feature fusion recognition.

[0018] Step S5: Time Consistency Verification Perform time consistency verification on precise anomaly markers. The verification conditions include two aspects: Precise anomaly markers are triggered every consecutive preset number of sampling periods; Second, both the rate of change of expansion force and the rate of change of hydrogen concentration showed a trend of increasing period by period during this continuous sampling period.

[0019] When both of the above conditions are met simultaneously, the abnormal state is confirmed to be valid; otherwise, it is determined to be occasional signal interference. This effectively filters out false alarms caused by transient noise and external disturbances.

[0020] Step S6: Tiered Alarm Decision Based on the combination of suspected anomaly markers and valid anomaly states, combined with the current absolute value of hydrogen concentration, a tiered alarm decision is executed. The specific logic is as follows: If neither the suspected anomaly marker nor the anomaly status is triggered, the system is classified as normal and continues routine monitoring. When a suspected anomaly marker is triggered but the anomaly status is valid and not triggered, it is determined to be a Level 1 warning. The system increases the data sampling frequency and records relevant data, but does not output alarms to avoid disturbing users. When both the suspected anomaly marker and the anomaly status are triggered and the current hydrogen concentration does not exceed the safety threshold, it is determined to be a level two alarm. The system will issue an audible and visual alarm and push alarm information to the user terminal. When both the suspected anomaly marker and the anomaly status are triggered and the current hydrogen concentration exceeds the safety threshold, it is determined to be a level three alarm. The system immediately cuts off the battery charging and discharging circuit and issues a strong alarm signal.

[0021] Step S7: Cloud-based iterative updates Event data triggering Level 1 alerts and above are uploaded to the cloud. The cloud uses a sliding window approach, selecting only feature vector data from periods deemed normal, and recalculates the mean vector and covariance matrix of the normal state baseline to ensure it tracks the slow changes caused by battery aging. Simultaneously, the uploaded trigger event data, after manual review and annotation, is used as new training samples to periodically retrain the random forest model in the cloud. The updated model parameters are then distributed to edge computing units, enabling the system to self-evolve.

[0022] Secondly, the present invention provides a multimodal flame retardant testing system for electric bicycle battery compartment materials, comprising: Multi-dimensional sensing module: Deployed inside the battery compartment, including at least a flexible thin-film pressure sensor, a temperature sensor, a voltage detection line, and a MEMS hydrogen sensor; Edge computing module: Communicates with the multi-dimensional sensing module to perform preprocessing, feature extraction, dual-path parallel discrimination, time consistency verification, and hierarchical alarm decision-making; Alarm Execution Module: Communicates with the edge computing module and is used to execute alarm actions corresponding to alarm decisions at all levels; Wireless communication module: Communicates with the edge computing module to upload triggered event data to the cloud; Cloud Iteration Module: Used to receive trigger event data, update the normal state baseline and iteratively train the random forest classifier, and send the updated parameters to the edge computing module.

[0023] Furthermore, the flexible thin-film pressure sensor is arranged between adjacent cells and between the cell and the battery compartment wall, with a sampling frequency of not less than 100 Hz, to capture rapid changes in the cell expansion force; The MEMS hydrogen sensor is located at the top of the battery compartment and uses the characteristic that hydrogen is less dense than air to monitor hydrogen accumulation. The sampling frequency is not less than 1 Hz. The temperature sensor is an NTC thermistor, and one is attached to the surface of every 2 to 3 battery cells. The system also includes an optional MEMS accelerometer, which is mounted on the battery compartment housing to collect vibration or impact signals and help distinguish between mechanical disturbances and precursors of thermal runaway. Beneficial effects

[0024] Compared with the prior art, the present invention has the following beneficial effects: This solution introduces two physical signals, expansion force and hydrogen concentration, which respond earlier than temperature and voltage. Combined with the pressure-temperature coupling characteristic ratio, an early mechanical precursor fingerprint, it can identify abnormal signs in the very early stages of thermal runaway (hundreds of seconds before the safety valve opens), thus buying valuable time for emergency response.

[0025] This scheme adopts a dual-path parallel discrimination strategy. Path A is responsible for high-sensitivity and rapid screening, while Path B performs secondary verification through multi-feature fusion and time consistency check, effectively eliminating occasional interference such as external environmental changes and sensor noise, and keeping the false alarm rate at a low level while ensuring high sensitivity.

[0026] This solution employs an adaptive dynamic threshold and a sliding window update mechanism based on normal state benchmarks, enabling the detection model to adapt to individual differences in batteries, different usage environments, and battery aging processes. This avoids false alarms or missed alarms caused by fixed thresholds when operating conditions change.

[0027] This approach differs from traditional single-dimensional vertical combustion tests. It comprehensively characterizes the flame-retardant evolution of battery compartment materials from multiple dimensions, including expansion force, temperature, voltage, and hydrogen concentration, thus achieving a shift from "static single-point detection" to "dynamic process monitoring."

[0028] This solution employs a cloud-based iterative update mechanism, enabling the detection model to continuously learn from real-world operational data. As usage time increases, the system's detection accuracy gradually improves, reducing reliance on manual calibration and regular maintenance. Attached Figure Description

[0029] Figure 1 This is a schematic diagram of the technical process for the multimodal flame retardant testing method for electric bicycle battery compartment materials provided by the present invention.

[0030] Figure 2This is a schematic diagram of the deployment structure of the multi-dimensional sensor network provided in the embodiment of the present invention within the battery compartment (14 groups of cells connected in series in the battery pack).

[0031] Figure 3 This is a schematic diagram of the dual-path parallel anomaly detection process provided in an embodiment of the present invention.

[0032] Figure 4 This is a schematic diagram of the hierarchical alarm decision logic provided in an embodiment of the present invention.

[0033] Figure 5 This is a structural block diagram of a multimodal flame-retardant detection system for electric bicycle battery compartment materials provided in an embodiment of the present invention. Detailed Implementation

[0034] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0035] Example 1: Detailed Implementation of the Detection Method: This embodiment provides a multimodal flame retardancy testing method for electric bicycle battery compartment materials, including the following steps: Step 1, Deployment of Multidimensional Sensor Network Taking a typical 48V / 20Ah lithium-ion battery pack (composed of 13 series 18650 cells) as an example, the following sensors are deployed in the battery compartment: 1) Flexible thin-film pressure sensor (model: FSR-406, range 0-20N): arranged between adjacent cell strings and between cell strings and battery compartment wall, a total of 6, sampling frequency set to 100Hz.

[0036] 2) MEMS hydrogen sensor (model: H2-BF-1000, range 0-1000ppm): located at the top center inside the battery compartment, with a sampling frequency set to 1Hz.

[0037] 3) NTC thermistor temperature sensor: one is attached to the surface of every two battery cells, for a total of 7, with a sampling frequency set to 1Hz.

[0038] 4) Voltage detection line: connected in parallel to the positive and negative terminals of each battery cell, with the sampling frequency set to 1Hz.

[0039] All sensors are connected to the edge computing unit (using an STM32F407 microcontroller) via a CAN bus.

[0040] Step 2, Data Preprocessing The edge computing unit performs the following preprocessing on the collected raw data: 1) Time synchronization: Align all sensor data with a unified timestamp, with a clock synchronization accuracy of ±1ms.

[0041] 2) Kalman Filtering: Perform one-dimensional Kalman filtering on the pressure signal. Assume the pressure measurement noise covariance R = 0.01 and the process noise covariance Q = 0.001. The initial estimate is taken from the measurement value of the first frame. The temperature and hydrogen concentration signals use the same parameters.

[0042] 3) Outlier removal: If any physical quantity exceeds the physical limit range (such as temperature <-20℃ or >150℃, voltage <2.0V or >4.3V), the data point is determined to be a bad value and removed, and linear interpolation is used to complete it.

[0043] 4) Normalization: The Z-score standardization method is used, and the mean and standard deviation of each physical quantity are taken from the normal operation data of the battery in the first 24 hours after the battery is first powered on.

[0044] Step 3, Multidimensional Feature Extraction Based on the preprocessed standardized data, five core features are extracted using the following formula: 1) Rate of change of expansion force r P (t k Using the central difference method: , where Δt = 0.01s.

[0045] 2) Rate of temperature change r T (t k ): , where Δt=1s.

[0046] 3) Voltage inconsistency coefficient σ V (t k ): .

[0047] 4) Rate of change of hydrogen concentration r C (t k ): , where Δt=1s.

[0048] 5) Pressure-temperature coupling characteristic ratio λ(t) k ): In this case, a minimum constant is added to the denominator to prevent division by zero.

[0049] Related research indicates that changes in expansion force can occur hundreds to over a thousand seconds before the battery safety valve opens, making it one of the earliest physical precursors to thermal runaway. Hydrogen concentration signals are collected by a MEMS hydrogen sensor positioned at the top of the battery compartment, utilizing the fact that hydrogen's density is less than air to monitor hydrogen accumulation. Hydrogen is a characteristic gas produced by electrolyte decomposition in the early stages of lithium-ion battery thermal runaway, and its concentration change rate is more specific than temperature changes.

[0050] Step 4, Parallel decision-making for dual paths Quick screening path: retrieve data from the past 24 hours. P and r C The data is used as a sliding window (N=86400 points, sampling frequency 1Hz), and the mean μ within the window is calculated respectively. P μ C and standard deviation σ P σ C Set the multiplier constant α=3, and calculate the dynamic threshold τ. P =μ P +3σ P , τ C =μ C +3σ C If r P >τ P or r C >τ C Then the suspected anomaly marker S is triggered. A =1.

[0051] Precise discrimination path: Combine the five features into a 5-dimensional feature vector f(t) k )=[r P ,r T ,σ V ,r C ,λ] T .

[0052] The mean vector μ0 and covariance matrix Σ0 of the normal state baseline are taken from the data of the first 7 days of normal charge-discharge cycles after the battery was delivered for use. The Mahalanobis distance calculation formula is: The first threshold is determined by... The 95th percentile is determined and taken as 3.31.

[0053] The random forest classifier contains 100 decision trees, and the training set includes 5000 normal data samples and 500 known thermal runaway data samples. The anomaly probability p... RF To determine the percentage of decision trees that are abnormal, the second threshold is set to 0.6.

[0054] When D M >3.31 or p RF When the value is greater than 0.6, the precise anomaly marker S is triggered.B =1.

[0055] Step 5, Time Consistency Verification Set the number of consecutive sampling periods N c =5 (corresponding to 5 seconds). If satisfied: ,and ,and Then the abnormal state is confirmed to be valid. ;otherwise .

[0056] Step 6, tiered alarm decision Safety threshold C safe The setting is 50 ppm. The decision logic is as follows:

[0057] Step 7, Cloud-based iterative updates Each time an event of Level 1 warning or above occurs, the edge computing module uploads complete waveform data (including raw sensor data, feature values, and intermediate judgment results) for 30 seconds before and after the event to the cloud server via 4G wireless communication.

[0058] Every 24 hours, the cloud recalculates μ0 and Σ0 of the normal state baseline, selecting only the data from the time period that is judged to be at the normal level on that day. The random forest model is retrained in the cloud once a week, with training data including all historical samples that have been manually verified and labeled. The updated model parameters are delivered to the edge computing module via OTA.

[0059] Example 2: Detailed Implementation of the Detection System: This embodiment provides a multimodal flame retardant testing system for electric bicycle battery compartment materials, such as... Figure 5 As shown, it includes the following modules: Multi-dimensional sensing module: 6 flexible thin film pressure sensors, arranged between adjacent cells (4) and between cells and the chamber wall (2), adopt FSR-406 type, with a range of 0-20N and an accuracy of ±0.1N. The output analog voltage signal is converted by ADC and then connected to the CAN bus.

[0060] One MEMS hydrogen sensor is located at the top center of the battery compartment. It is an H2-BF-1000 model with a range of 0-1000ppm and a resolution of 1ppm. The output digital signal is connected to the CAN bus via the I²C bus.

[0061] Seven NTC thermistors, one for every two battery cells, are used. They are of type MF52, with a resistance of 10kΩ at 25℃ and an accuracy of ±1%. They are connected to the ADC via a voltage divider circuit.

[0062] There are 13 voltage detection lines, which are directly connected in parallel to the positive and negative terminals of each battery cell and then connected to the ADC via a differential amplifier.

[0063] One MEMS accelerometer (optional) is placed at the bottom of the battery compartment housing to collect vibration / impact signals and help distinguish between mechanical disturbances and precursors of thermal runaway.

[0064] Edge computing module: Employs an STM32H743 microcontroller with a 400MHz clock speed, 2MB of built-in Flash memory, and 1MB of RAM. Running the FreeRTOS real-time operating system, it is responsible for executing data acquisition drivers, preprocessing algorithms, feature extraction, dual-path parallel discrimination, and time consistency verification. It communicates with the sensing module via a CAN bus (500kbps baud rate), with the alarm execution module via GPIO, and with the wireless communication module via UART.

[0065] Alarm execution module: includes a buzzer (85dB), a red LED indicator, and a MOSFET switch (used to cut off the charging and discharging circuit, withstand voltage 100V, current 50A), all controlled by the edge computing module GPIO.

[0066] Wireless communication module: Employs the EC200N-CN 4G CAT1 communication module, enabling data interaction with the edge computing module via AT commands. Event data is encapsulated in JSON format, containing information such as timestamps, raw sensor data, feature values, and alarm levels.

[0067] Cloud-based iteration module: Deployed on Alibaba Cloud ECS server, running Ubuntu 20.04 operating system. It contains three sub-modules: 1) Data receiving submodule: Receives event data uploaded from the edge terminal based on the MQTT protocol and stores it in the time-series database InfluxDB; 2) Baseline Update Submodule: Extracts normal time period data from InfluxDB daily and recalculates μ0 and Σ0. 3) Model training submodule: The random forest model is retrained weekly based on the accumulated labeled data (implemented by Python scikit-learn), and the model parameters are distributed to the edge via the differential OTA protocol.

[0068] Example 3: Application Case: The above system was installed in the battery compartment of a certain brand of electric bicycle (battery pack specifications: 48V / 20Ah lithium-ion battery), and the following verification tests were conducted: Test 1: Simulation of thermal runaway triggered by needle puncture A needle penetration test was performed on a cell located in the middle of the battery pack (needle diameter 3mm, penetration speed 10mm / s), and the system response time was recorded. Test results: Approximately 120 seconds after acupuncture, the rate of change of expansion force r P First time exceeding the dynamic threshold (triggered S) A =1), at this time the surface temperature of the cell only rises by 2.3℃, and the voltage does not change significantly; Approximately 180 seconds after acupuncture, the rate of change of hydrogen concentration r C Exceeding the dynamic threshold (enhanced S) A =1), and Mahalanobis distance D M Exceeding 3.31 (triggered S) B =1); Approximately 185 seconds after acupuncture (after 5 consecutive seconds), S valid B = 1, the system issues a level two alarm. At this time, the cell surface temperature is approximately 52°C, and the hydrogen concentration is approximately 35ppm. This alarm time is approximately 135 seconds before the cell surface temperature reaches 80°C (approximately 320 seconds after needle penetration) and approximately 95 seconds before visible smoke appears (approximately 280 seconds after needle penetration), verifying the invention's high sensitivity detection capability for early, weak signs.

[0069] Test 2: Simulating External Environmental Disturbances The battery compartment of the system was placed in a 40°C constant temperature chamber and subjected to simulated vibration (frequency 50Hz, amplitude 0.5mm) for 72 hours of continuous operation. During this process, the system did not trigger any alarms, verifying the anti-interference capability of the adaptive dynamic threshold and time consistency verification against external environmental disturbances.

[0070] Test 3: Long-term operational stability The system was connected to the battery compartments of three electric bicycles of different brands and with different service lives (1-3 years) and operated continuously for 6 months. During this period, a total of 17 Level 1 warnings, 3 Level 2 alarms, and 0 Level 3 alarms were triggered. Manual verification revealed that 15 of the 17 Level 1 warnings corresponded to abnormal individual battery cell voltages that did not develop into thermal runaway, and 2 were due to brief sensor signal drift. All 3 Level 2 alarms corresponded to actual precursors to thermal runaway (both caused by worsening inconsistencies in cell voltages within the battery pack leading to increased pressure). The overall system sensitivity (detection rate of actual precursors to thermal runaway) was 100%, and the false alarm rate was 1.87% (based on the total number of triggered events), verifying the effectiveness of this invention in balancing high sensitivity and low false alarm rate in practical applications.

[0071] The multimodal flame-retardant detection method and system for electric bicycle battery compartment materials provided by this invention can be widely applied to the field of battery safety monitoring for two-wheeled / three-wheeled electric vehicles such as electric bicycles, electric motorcycles, and electric tricycles. In charging cabinet scenarios, it can also be used in conjunction with existing charging cabinet fire suppression systems as a signal source for early warning of thermal runaway, triggering a graded fire suppression response.

[0072] Each module of this invention uses mature components and standard communication protocols, has a mature manufacturing process, controllable costs, and is industrially practical for large-scale application.

[0073] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely illustrative of the principles of the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the present invention as claimed. The scope of protection of this invention is defined by the appended claims and their equivalents.

Claims

1. A multimodal flame retardancy testing method for battery compartment materials of electric bicycles, characterized in that, Includes the following steps: S1: Deploy a multi-dimensional sensor network inside the battery compartment to collect various physical state parameters of the battery in real time at a preset sampling frequency; S2: The edge computing unit preprocesses the collected raw physical state parameters to obtain standardized time-series data; S3: Based on the standardized time-series data, extract multi-dimensional feature parameters in real time. The multi-dimensional feature parameters include at least the expansion force change rate, temperature change rate, voltage inconsistency coefficient, hydrogen concentration change rate, and pressure-temperature coupling feature ratio. S4: Use a dual-path parallel discrimination strategy to determine anomalies in the multi-dimensional feature parameters and trigger suspected anomaly marking or precise anomaly marking. S5: Perform time consistency verification on the precise anomaly marker. Only when the precise anomaly marker is valid for a continuous preset number of sampling periods and the multidimensional feature parameters show a monotonically increasing trend is the anomaly status confirmed to be valid. S6: Based on the combination of the suspected anomaly marker and the valid anomaly state, and combined with the current absolute value of hydrogen concentration, execute a graded alarm decision.

2. The multimodal flame retardancy testing method for electric bicycle battery compartment materials according to claim 1, characterized in that, The voltage inconsistency coefficient is obtained by calculating the standard deviation of the voltage of all individual cells in the battery pack; the pressure-temperature coupling characteristic ratio is obtained by calculating the ratio of the rate of change of expansion force to the rate of change of temperature.

3. The multimodal flame retardancy testing method for electric bicycle battery compartment materials according to claim 1, characterized in that, The dual-path parallel discrimination strategy includes: Rapid screening path: Calculate adaptive dynamic thresholds for the expansion force change rate and the hydrogen concentration change rate respectively. If either change rate exceeds the corresponding dynamic threshold, a suspected anomaly marker is triggered. Precise identification path: Construct the multidimensional feature parameters into a feature vector, calculate the Mahalanobis distance of the feature vector relative to the normal state benchmark, and input it into a pre-trained random forest classifier to obtain the anomaly probability. If the Mahalanobis distance exceeds a first threshold or the anomaly probability exceeds a second threshold, precise anomaly labeling is triggered.

4. The multimodal flame retardancy testing method for electric bicycle battery compartment materials according to claim 3, characterized in that, The adaptive dynamic threshold is adaptively adjusted according to individual battery differences and changes in operating conditions. The calculation method is as follows: using historical data within a preset time period as a sliding window, the mean and standard deviation of the expansion force change rate or the hydrogen concentration change rate within the sliding window are calculated respectively, and the sum of the mean and the preset multiple standard deviation is used as the dynamic threshold at the current moment.

5. The multimodal flame retardancy testing method for electric bicycle battery compartment materials according to claim 3, characterized in that, The Mahalanobis distance is calculated as follows: based on the difference vector between the feature vector and the mean vector of the normal state benchmark, and the inverse matrix of the covariance matrix of the normal state benchmark, the weighted sum of squares of the difference vector is calculated and then the square root is taken; the normal state benchmark is estimated from the historical feature vector data of the battery during normal operation, and the first threshold is determined by the preset quantile of the chi-square distribution.

6. The multimodal flame retardancy testing method for electric bicycle battery compartment materials according to claim 3, characterized in that, The random forest classifier is an ensemble of multiple decision trees. Each decision tree independently performs a binary classification of the feature vector as normal or abnormal. The abnormality probability is the proportion of the number of decision trees that are judged as abnormal out of the total number of decision trees. The random forest classifier is pre-trained using normal charge and discharge data samples and known thermal runaway data samples. The second threshold is a preset probability threshold value.

7. The multimodal flame retardancy testing method for electric bicycle battery compartment materials according to claim 1, characterized in that, The time consistency verification criteria are as follows: when the precise anomaly marker is triggered within a consecutive preset number of sampling periods, and the rate of change of expansion force and the rate of change of hydrogen concentration both show a trend of increasing period by period within the consecutive preset number of sampling periods, the anomaly state is confirmed to be valid; otherwise, it is determined to be occasional signal interference.

8. The multimodal flame retardancy testing method for electric bicycle battery compartment materials according to claim 1, characterized in that, The tiered alarm decision-making system includes at least four levels: normal, first-level warning, second-level alarm, and third-level alarm.

9. A multimodal flame retardancy testing method for electric bicycle battery compartment materials according to claim 8, characterized in that, The logic for the tiered alarm decision is as follows: When neither the suspected anomaly marker nor the anomaly status is triggered, the condition is determined to be at the normal level. When the suspected anomaly marker is triggered but the anomaly status is valid but not triggered, it is determined to be a Level 1 warning. When both the suspected anomaly marker and the anomaly status are triggered and the current hydrogen concentration does not exceed the safety threshold, it is determined to be a level two alarm. When both the suspected anomaly marker and the anomaly status are triggered and the current hydrogen concentration exceeds the safety threshold, it is determined to be a Level 3 alarm.

10. A multimodal flame retardancy testing method for battery compartment materials of electric bicycles according to any one of claims 1-9, characterized in that, The method further includes: S7: Upload event data that triggers a Level 1 warning or higher to the cloud for updating the normal state baseline and iteratively training the random forest classifier; The update method for the normal state benchmark is as follows: using a sliding window approach, only the feature vector data within the time period judged as normal level are selected, and the mean vector and covariance matrix of the normal state benchmark are recalculated so that the normal state benchmark can track the slow changes caused by battery aging; the iterative training method for the random forest classifier is as follows: the uploaded trigger event data is manually reviewed and labeled as new training samples, the random forest model is periodically retrained in the cloud, and the updated model parameters are sent to the edge computing unit.

Images