Multifunctional integrated monitoring and fire control method and system based on BMS

Abstract

The invention discloses a BMS-based multifunctional integrated monitoring and fire control method and system, and the method comprises the steps: synchronously collecting the temperature, combustible gas concentration and dynamic environment monitoring data in a container through a BMU battery box management unit of a BMS, and generating a space-time correlated container environment state matrix; based on the environment state matrix, extracting temperature anomaly features, gas concentration gradient features and dynamic environment fluctuation features through a multi-modal feature fusion model, and constructing a multi-dimensional risk feature vector; inputting the multi-dimensional risk feature vector into a fire-fighting early warning optimization algorithm to generate a graded fire-fighting early warning signal; and according to the graded fire-fighting early warning signal, an active fire-fighting intervention strategy is triggered through a BMU unit of the BMS, a fire extinguishing device, a ventilation system and a power cut-off module are dynamically controlled, and a fire-fighting response path is synchronously optimized to be matched with the current risk grade. According to the embodiment of the invention, graded fire-fighting early warning can be realized, and the accuracy and timeliness of fire early warning are improved.

Description

Technical Field

[0001] The present invention belongs to the field of fire protection technology, and particularly relates to a multi-functional integrated monitoring and fire control method and system based on BMS. Background Art

[0002] With the continuous improvement of industrial automation level and the development of intelligent manufacturing, the Battery Management System (BMS) is becoming increasingly popular in the fields of energy storage, power drive, etc. As the core device to ensure battery safety and improve system reliability, BMS is not only responsible for the state monitoring and management of batteries, but also gradually develops into an important platform for multi-functional integrated monitoring and intelligent control. In practical applications, as an important energy storage unit, the containerized battery pack has a complex and changeable operating environment, which may be affected by factors such as temperature rise, gas accumulation, and environmental fluctuations. If the monitoring means are single or there are information islands, it is difficult to comprehensively understand the internal state in a timely manner, and there are potential fire hazards. Once a fire accident occurs, it will not only cause property losses, but also pose a threat to personnel safety and the stable operation of the system. Traditional fire monitoring methods mainly rely on single temperature or gas sensors, which have defects such as high false alarm rate and slow response, and it is difficult to dynamically track the characteristics of environmental changes. Summary of the Invention

[0003] The purpose of the present invention is to provide a multi-functional integrated monitoring and fire control method and system based on BMS to solve the deficiencies in the prior art, and to achieve hierarchical fire early warning, improving the accuracy and timeliness of fire early warning.

[0004] An embodiment of the present application provides a multi-functional integrated monitoring and fire control method based on BMS, and the method includes: Synchronously collect the temperature, combustible gas concentration and dynamic environment monitoring data in the container through the BMU unit of BMS, and use the spatio-temporal alignment algorithm of heterogeneous sensor networks to fuse and process multi-source data to generate a spatio-temporally correlated container environment state matrix; Based on the environment state matrix, extract temperature anomaly features, gas concentration gradient features and dynamic environment fluctuation features through a multi-modal feature fusion model to construct a multi-dimensional risk feature vector, wherein the multi-modal feature fusion model dynamically weights the contribution values of different monitoring dimensions through an attention mechanism; Input the multi-dimensional risk feature vector into the fire early warning optimization algorithm, and combine the dynamic threshold adjustment rule trained with historical accident data to calculate the fire risk index and early warning level, and generate a hierarchical fire early warning signal; According to the hierarchical fire early warning signal, trigger an active fire intervention strategy through the BMU unit of BMS, dynamically control the fire extinguishing device, ventilation system and power cut-off module, and synchronously optimize the fire response path to match the current risk level.

[0005] Optionally, the temperature, combustible gas concentration, and dynamic environment monitoring data inside the container are synchronously collected by the BMU unit of the BMS, and a heterogeneous sensor network spatio-temporal alignment algorithm is used to fuse and process the multi-source data to generate a spatio-temporally correlated container environment state matrix, including: Based on the original data streams of the temperature sensor, combustible gas sensor, and dynamic environment monitoring device collected by the BMU unit of the BMS, perform timestamp synchronization calibration, and eliminate the sampling frequency differences of each sensor through sliding window mean compensation to generate a time-aligned original data set; Perform spatial position calibration on the original data set, establish a spatial coordinate mapping table based on the sensor deployment topology structure, and use the covariance matrix matching algorithm to eliminate cross-regional monitoring errors to generate a spatially aligned multi-dimensional monitoring matrix; According to the multi-dimensional monitoring matrix, use the spatio-temporal alignment algorithm based on the heterogeneous sensor network, combine the Gaussian mixture model to perform probability distribution modeling on the temperature, gas concentration, and dynamic environment parameters, calculate the cross-modal data correlation degree, and generate a spatio-temporal correlation confidence map; According to the spatio-temporal correlation confidence map, use Bayesian inference to fuse multi-source data, perform weighted correction on the conflicting monitoring values, output the spatio-temporally correlated container environment state matrix, and store it in the edge computing node.

[0006] Optionally, based on the environment state matrix, extract temperature anomaly features, gas concentration gradient features, and dynamic environment fluctuation features through a multi-modal feature fusion model to construct a multi-dimensional risk feature vector, where the multi-modal feature fusion model dynamically weights the contribution values of different monitoring dimensions through an attention mechanism, including: Extract the temperature time series data from the environment state matrix, calculate the local fluctuation amplitude through moving standard deviation, and perform differential analysis in combination with the historical normal temperature range to generate a temperature anomaly feature vector; Analyze the spatial distribution gradient of the gas concentration data, use Fourier transform to separate the steady-state component and the transient component, calculate the concentration change rate and diffusion direction of each monitoring point, and generate a gas concentration gradient feature vector; Perform wavelet packet decomposition on the dynamic environment monitoring data, extract the high-frequency sub-band energy ratio as the dynamic environment fluctuation index, and generate a dynamic environment fluctuation feature vector; Construct a multi-modal feature fusion model, parallel input the temperature anomaly feature vector, gas concentration gradient feature vector, and dynamic environment fluctuation feature vector into the attention mechanism layer, and dynamically calculate the contribution values of each feature dimension through a trainable weight matrix; According to the dynamically weighted feature contribution values, perform feature concatenation and normalization processing to generate a risk feature vector containing multi-dimensional risk correlation relationships, and append a timestamp index to store it in the feature database.

[0007] Optionally, inputting the multi-dimensional risk feature vector into the fire warning optimization algorithm, combining the dynamic threshold adjustment rule trained with historical accident data, calculating the fire risk index and warning level, and generating a graded fire warning signal, including: Inputting the multi-dimensional risk feature vector into a pre-trained long short-term memory network to predict the risk evolution trend in the next three monitoring cycles and generate a risk time series prediction curve; Based on the fire cases in the historical accident database, extracting the threshold boundaries of key feature dimensions, and dynamically adjusting the threshold sensitivity coefficient of the current environmental state through a sliding window matching algorithm; Calculating the Mahalanobis distance between the real-time multi-dimensional risk feature vector and the dynamic risk threshold, and comprehensively evaluating the fire risk index in combination with the slope change of the risk time series prediction curve; Dividing into four warning levels according to the size of the fire risk index. When the fire risk index exceeds the preset level threshold, trigger a cross-regional risk linkage verification mechanism; Fusing the real-time evaluation result and the linkage verification feedback, generating a graded fire warning signal with confidence level annotation, and encrypting and transmitting it to the BMS control center.

[0008] Optionally, according to the graded fire warning signal, triggering an active fire intervention strategy through the BMU unit of the BMS, dynamically controlling the fire extinguishing device, ventilation system and power cut-off module, and synchronously optimizing the fire response path to match the current risk level, including: Analyzing the risk level of the graded fire warning signal, matching the preset fire strategy template library, and planning an initial intervention instruction set including the start priority of the fire extinguishing device, the ventilation rate gear and the power cut-off delay; According to the internal topology map of the container and the device deployment location, using the Dijkstra algorithm to calculate the optimal response path of each fire-fighting device, eliminating action conflicts and optimizing the response time sequence of the initial intervention instruction set to obtain a fire-fighting instruction set; Simulating and executing the fire-fighting instruction set at the edge computing node, verifying the change trend of the environmental parameters after the instruction execution through a digital twin model, and triggering instruction backtracking and replanning if the simulation result does not meet the expectation; Issuing the finally coordinated scheduling instruction passed the verification to the BMU unit, synchronously controlling the spraying angle of the fire extinguishing device, the variable frequency fan of the ventilation system and the intelligent circuit breaker, and real-time transmitting the execution status back to the central monitoring platform.

[0009] Another embodiment of the present application provides a multi-functional integrated monitoring and fire control system based on the BMS. The system includes: A fusion module is used to synchronously collect the temperature, concentration of combustible gas, and dynamic environment monitoring data inside the container through the BMU unit of the BMS, and adopt a spatio-temporal alignment algorithm for heterogeneous sensor networks to fuse and process multi-source data, generating a spatio-temporally correlated container environment state matrix; A construction module is used to extract temperature anomaly features, gas concentration gradient features, and dynamic environment fluctuation features based on the environment state matrix through a multi-modal feature fusion model, and construct a multi-dimensional risk feature vector. Among them, the multi-modal feature fusion model dynamically weights the contribution values of different monitoring dimensions through an attention mechanism; A grading module is used to input the multi-dimensional risk feature vector into a fire warning optimization algorithm, and calculate the fire risk index and warning level in combination with the dynamic threshold adjustment rule trained with historical accident data, generating a graded fire warning signal; A control module is used to trigger an active fire intervention strategy through the BMU unit of the BMS according to the graded fire warning signal, dynamically control the fire extinguishing device, ventilation system, and power cut-off module, and synchronously optimize the fire response path to match the current risk level.

[0010] Another embodiment of the present application provides a storage medium in which a computer program is stored. Among them, the computer program is set to execute the method described in any one of the above when running.

[0011] Another embodiment of the present application provides an electronic device, including a memory and a processor. A computer program is stored in the memory, and the processor is set to run the computer program to execute the method described in any one of the above.

[0012] Compared with the prior art, a multi-functional integrated monitoring and fire control method based on BMS provided by the present invention synchronously collects the temperature, concentration of combustible gas, and dynamic environment monitoring data inside the container through the BMU unit of the BMS, generating a spatio-temporally correlated container environment state matrix; based on the environment state matrix, temperature anomaly features, gas concentration gradient features, and dynamic environment fluctuation features are extracted through a multi-modal feature fusion model, constructing a multi-dimensional risk feature vector; the multi-dimensional risk feature vector is input into a fire warning optimization algorithm, generating a graded fire warning signal; according to the graded fire warning signal, an active fire intervention strategy is triggered through the BMU unit of the BMS, dynamically controlling the fire extinguishing device, ventilation system, and power cut-off module, and synchronously optimizing the fire response path to match the current risk level, so as to be able to achieve graded fire warning and improve the accuracy and timeliness of fire warning. Description of the Drawings

[0013] Figure 1 It is a hardware structure block diagram of a computer terminal for a multi-functional integrated monitoring and fire control method based on BMS provided by an embodiment of the present invention; Figure 2 Schematic flow diagram of a multi-functional integrated monitoring and fire control method based on BMS provided by an embodiment of the present invention; Figure 3 Schematic structural diagram of a multi-functional integrated monitoring and fire control system based on BMS provided by an embodiment of the present invention. Detailed implementation manners

[0014] The embodiments described below by referring to the accompanying drawings are exemplary and are only used to explain the present invention, and should not be construed as limiting the present invention.

[0015] An embodiment of the present invention first provides a multi-functional integrated monitoring and fire control method based on BMS. This method can be applied to electronic devices, such as computer terminals, specifically, ordinary computers, etc. Hereinafter, it will be described in detail by taking the operation on a computer terminal as an example. Figure 1 Hardware structure block diagram of a computer terminal for a multi-functional integrated monitoring and fire control method based on BMS provided by an embodiment of the present invention. As Figure 1 shown, this computer device includes a processor, a memory, and a network interface connected through a system bus. Among them, the memory may include a non-volatile storage medium and an internal memory.

[0016] The non-volatile storage medium can store an operating system and a computer program. This computer program includes program instructions. When the program instructions are executed, the processor can execute any multi-functional integrated monitoring and fire control method based on BMS.

[0017] The processor is used to provide computing and control capabilities to support the operation of the entire computer device.

[0018] The internal memory provides an environment for the operation of the computer program in the non-volatile storage medium. When the computer program is executed by the processor, the processor can execute any multi-functional integrated monitoring and fire control method based on BMS.

[0019] This network interface is used for network communication, such as sending assigned tasks, etc. Those skilled in the art can understand that Figure 1 the structure shown in

[0020] It should be understood that the processor may be a central processing unit (CPU), and the processor may also be other general-purpose processors, digital signal processors (DSPs), application specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs) or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. Among them, the general-purpose processor may be a microprocessor or the processor may also be any conventional processor, etc.

[0021] See Figure 2 , an embodiment of the present invention provides a multi-functional integrated monitoring and fire control method based on BMS, which may include the following steps: S201, synchronously collect the temperature, combustible gas concentration and dynamic environment monitoring data in the container through the BMU unit of the BMS, and use the heterogeneous sensor network spatio-temporal alignment algorithm to fuse and process the multi-source data to generate a spatio-temporally associated container environment state matrix; specifically, it may include: According to the original data streams of the temperature sensor, combustible gas sensor and dynamic environment monitoring device collected by the BMU unit of the BMS, perform timestamp synchronization calibration, and eliminate the sampling frequency differences of each sensor through sliding window mean compensation to generate a time-aligned original data set; The core objective of this step is to unify the time bases of multi-source sensors and solve the problem of inconsistent sampling frequencies, ensuring that the data for subsequent analysis is time-aligned. The specific implementation needs to solve timestamp calibration, data interpolation and sliding window compensation technologies.

[0022] Timestamp synchronization calibration: Hardware clock alignment: The BMU (Battery Management Unit) synchronizes all sensor clocks through CAN communication or Ethernet communication, and the error is controlled within ±10 milliseconds. For example, the temperature sensor (model DS18B20), combustible gas sensor (MQ-5 module) and dynamic environment monitoring device (temperature and humidity + vibration sensor) uniformly adopt the UTC time format YYYY-MM-DD HH:MM:SS.sss.

[0023] Data Stream Alignment: Sort the data collected per second according to timestamps. For example, if the sampling frequency of the temperature sensor is 2 Hz (once every 0.5 seconds), that of the gas sensor is 0.5 Hz (once every 2 seconds), and that of the dynamic environment monitoring device is 1 Hz (once every second), the data stream needs to be sliced by the minimum time unit of 0.5 seconds, and the missing data is marked as NaN.

[0024] Sliding Window Mean Compensation: Window Parameter Setting: The size of the sliding window is 5 seconds (covering 10 temperature sampling points, 2.5 gas sampling points, and 5 dynamic environment sampling points), and the window sliding step is 1 second.

[0025] Interpolation Strategy: For the low-frequency data of the gas sensor, linear interpolation is used to fill in the missing values. For example, if the gas concentration data within a certain window is [0.5, NaN, NaN, NaN, 0.6], after interpolation, it becomes [0.5, 0.525, 0.55, 0.575, 0.6].

[0026] Mean Calculation: Calculate the mean of the data of the same sensor within each window. For example, the mean of the temperature data within the window from 10:00:00 to 10:00:05 is 25.3 °C, replacing the original fluctuating data.

[0027] Application Example: Input: The original data stream of a certain container from 10:00:00 to 10:00:05 (10 sampling points of the temperature sensor [24.8, 25.1, …, 25.5], 3 sampling points of the gas sensor [0.5, 0.55, 0.6]); Processing: Align the time axis at 0.5-second intervals, interpolate the gas data to obtain 10 points [0.5, 0.525, …, 0.6], and directly take the mean of the dynamic environment data; Output: The original data set with time alignment, in the format of [timestamp, temperature, gas concentration, dynamic environment parameter], such as 10:00:00.0, 24.8, 0.5, 60%RH.

[0028] Perform spatial position calibration on the original data set, establish a spatial coordinate mapping table based on the sensor deployment topology structure, and use the covariance matrix matching algorithm to eliminate the monitoring error across regions, generating a spatially aligned multi-dimensional monitoring matrix; This step needs to solve the monitoring error caused by the physical position differences of the sensors, and achieve cross-region data calibration through spatial mapping and covariance analysis.

[0029] Sensor Deployment Topology Modeling: Coordinate mapping table construction: Based on the 3D CAD drawing of the container (dimensions 12m × 2.4m × 2.6m), mark the installation positions of each sensor. For example, the temperature sensor T1 is located at coordinates (x = 1.2m, y = 0.6m, z = 1.8m), and the gas sensor G1 is located at (x = 3.6m, y = 0.6m, z = 1.5m).

[0030] Area division: Divide the container into 6 monitoring areas (each section is 2 meters long), and each area contains 1 temperature sensor, 1 gas sensor, and 1 dynamic environment sensor.

[0031] Covariance matrix matching algorithm: Cross-region data association: Calculate the covariance matrices of sensors in different regions at the same time. For example, the covariance of the temperature in region 1 and the temperature in region 2 is 0.85, and the covariance of the gas concentration in region 1 and the gas concentration in region 3 is -0.2 (a negative correlation indicates the direction of leakage diffusion).

[0032] Error correction: If the data of the temperature sensor T2 in region 2 is abnormal due to position deviation (such as 5°C higher than the adjacent region), through the eigenvalue decomposition of the covariance matrix, adjust the data of T2 to T2_corrected = (T1 × 0.7 + T3 × 0.3).

[0033] Application example: Input: The temperature in region 1 is 25°C, the temperature in region 2 is 30°C (an outlier), and the temperature in region 3 is 26°C; Processing: According to the covariance matrix (the covariance of the temperature between region 1 and region 2 is 0.8), calculate the corrected value of region 2 as 25 × 0.8 + 26 × 0.2 = 25.2°C; Output: A spatially aligned multi-dimensional monitoring matrix in the format of [region ID, temperature, gas concentration, dynamic environment parameter], such as Zone1, 25.0, 0.5, 60%RH.

[0034] Based on the multi-dimensional monitoring matrix, use the spatio-temporal alignment algorithm based on the heterogeneous sensor network, combine the Gaussian mixture model to model the probability distributions of temperature, gas concentration, and dynamic environment parameters, calculate the cross-modal data association degree, and generate a spatio-temporal association confidence map; This step quantifies the spatio-temporal correlation of multi-source data through a probability model, providing a confidence basis for subsequent data fusion.

[0035] Gaussian mixture model (GMM) modeling: Data distribution fitting: Three Gaussian distribution components are established for temperature, gas concentration, and dynamic environment parameters respectively. For example, temperature data is divided into low temperature (μ = 20°C, σ = 2°C), normal temperature (μ = 25°C, σ = 3°C), and high temperature (μ = 30°C, σ = 4°C).

[0036] Joint probability calculation: Calculate the joint probability of multi-modal data at the same moment. For example, if the temperature is 25°C (probability of belonging to the normal temperature distribution is 0.7), the gas concentration is 0.5 ppm (probability of belonging to the safe distribution is 0.9), and the dynamic environment parameter is normal (probability is 0.8), then the joint probability is 0.7×0.9×0.8 = 0.504.

[0037] Cross-modal correlation analysis: Correlation matrix: Construct a 6×6 matrix (6 regions), and the elements represent the data correlation strength between region i and region j. For example, the temperature correlation degree between region 1 and region 2 is 0.8, and the gas correlation degree is 0.6.

[0038] Confidence calculation: If the temperature in region 1 and the gas concentration in region 2 rise synchronously within the time window, the spatio-temporal correlation confidence is 0.8×0.6×time overlap rate 0.9 = 0.432.

[0039] Application example: Input: Temperature in region 1 is 25°C (normal temperature distribution), gas is 0.5 ppm (safe distribution), temperature in region 2 is 25.2°C (normal temperature distribution), gas is 0.6 ppm (warning distribution); Processing: Calculate the joint confidence of region 1 - 2 as 0.7×0.9×0.7×0.5 = 0.2205; Output: Spatio-temporal correlation confidence map, in the format of [timestamp, region pair, confidence], such as 10:00:00,Zone1-Zone2, 0.432.

[0040] According to the spatio-temporal correlation confidence map, use Bayesian inference to fuse multi-source data, weight and correct the conflict monitoring values, and output the spatio-temporal correlated container environment status matrix and store it in the edge computing node.

[0041] This step solves data conflicts through probability reasoning and generates a highly reliable environment status matrix.

[0042] Bayesian inference data fusion: Prior probability setting: According to historical data, the prior probability of temperature anomaly is 5%, and the prior probability of gas leakage is 2%.

[0043] Likelihood Probability Calculation: When the temperature in Area 1 reaches 28°C (high-temperature distribution probability 0.6), and the gas concentration in Area 2 is 0.8 ppm (hazardous distribution probability 0.7), the likelihood probability is 0.6 × 0.7 = 0.42.

[0044] Posterior Probability Update: If the prior probability is 5%, the posterior probability is (0.42 × 0.05) / (0.42 × 0.05 + 0.58 × 0.95) = 3.6%, which is determined to be a low risk.

[0045] Conflict Data Correction: Weighting Strategy: For conflicting sensor data (e.g., the temperature sensor in Area 1 shows 30°C, but the adjacent areas are all 25°C), it is corrected to 30 × 0.3 + 25 × 0.7 = 26.5°C according to the confidence weights (confidence in Area 1 is 0.3, confidence in Areas 2 - 3 is 0.7).

[0046] Application Example: Input: Temperature in Area 1 is 30°C (abnormal), gas is 0.5 ppm (normal), temperature in Area 2 is 25°C (normal), gas is 0.8 ppm (abnormal); Processing: According to the confidence map (correlation between Areas 1 - 2 is 0.4), the temperature in Area 1 is corrected to 30 × 0.4 + 25 × 0.6 = 27°C; Output: A spatio-temporal correlated environmental state matrix in the format of [timestamp, area ID, temperature, gas concentration, dynamic environment parameter, confidence], e.g., 10:00:00, Zone1, 27.0, 0.5, 60%RH, 0.85.

[0047] In this step, the battery management unit (BMU) of the BMS is used to collect real-time heterogeneous data streams from temperature sensors, combustible gas sensors, and dynamic environment monitoring devices (such as current and voltage) inside the container. The sliding window mean compensation is used to eliminate the sampling frequency differences of sensors (such as time alignment between the 1Hz sampling of the temperature sensor and the 0.5Hz sampling of the gas sensor), and the spatial positions of the monitoring data are calibrated through a spatial coordinate mapping table (based on the physical deployment topology of the sensors). The Gaussian mixture model is used to model the probability distribution of multi-source data, and Bayesian inference is combined to weight and correct the conflicting data (such as abnormal temperature but normal gas concentration in a certain area). Finally, a spatio-temporal correlated container environmental state matrix (a three-dimensional tensor containing timestamps, position coordinates, and normalized parameter values) is generated to achieve precise fusion of cross-modal data, solve the problems of spatio-temporal misalignment and fragmentation of multi-source data in traditional monitoring systems, improve data consistency through spatio-temporal alignment algorithms, provide high-precision input for subsequent risk analysis, avoid misjudgment caused by sensor sampling differences or spatial errors, and provide a global and real-time digital mapping of the container environmental state.

[0048] S202. Based on the environmental state matrix, extract the temperature anomaly feature, gas concentration gradient feature, and dynamic environmental fluctuation feature through a multi-modal feature fusion model to construct a multi-dimensional risk feature vector. Among them, the multi-modal feature fusion model dynamically weights the contribution values of different monitoring dimensions through an attention mechanism. Specifically, it may include: Extract the temperature time series data from the environmental state matrix, calculate the local fluctuation amplitude through moving standard deviation, and perform differential analysis in combination with the historical normal temperature range to generate a temperature anomaly feature vector. In this step, by dynamically capturing the fluctuation anomalies of the temperature sequence, quantify the amplitude of the temperature deviation from the historical normal range, and construct a temperature anomaly feature vector.

[0049] Moving standard deviation calculation: Window parameter setting: The sliding window size is set to 30 seconds (corresponding to 60 sampling points, assuming the sampling frequency is 2Hz), and the window slides 5 seconds each time. For example, calculate the standard deviation of the temperature data [25.0, 25.2, …, 26.1] from timestamp 10:00:00 to 10:00:30, and the temperature fluctuation value within the window is σ = 0.8°C.

[0050] Local fluctuation amplitude extraction: The standard deviation of each window is used as the fluctuation index at that time point. For example, if the standard deviation of a certain window suddenly increases to 2.5°C (the historical average fluctuation is 0.6°C), it is marked as a potential anomaly.

[0051] Definition of historical normal temperature range: Source of reference data: Based on the temperature data of the past 30 days, statistically calculate the temperature mean ± 3σ range for the same time period every day. For example, the historical normal range at 10:00:00 is 24.5 - 26.5°C (mean 25.5°C, σ = 0.33°C).

[0052] Differential analysis: The absolute value of the difference between the real-time temperature data and the historical mean is used as the deviation degree. For example, the deviation degree between the current temperature of 26.8°C and the historical mean of 25.5°C is 1.3°C. If it exceeds 3σ (1.0°C), it is determined to be abnormal.

[0053] Feature vector generation: Feature dimension: Each time point contains three features: the standard deviation of the current window (local_std = 2.5), the deviation degree (deviation = 1.3), and the deviation duration (duration = 120 seconds).

[0054] Application example: The temperature data within a certain window has been continuously higher than the upper limit of the historical range for 3 minutes, generating a feature vector [2.5, 1.3, 180], which is normalized to [0.8, 0.65, 1.0] (the maximum values are 3.0 °C, 2.0 °C, and 300 seconds respectively).

[0055] Analyze the spatial distribution gradient of gas concentration data, use Fourier transform to separate the steady-state component and the transient component, calculate the concentration change rate and diffusion direction at each monitoring point, and generate a gas concentration gradient feature vector; In this step, the location of the leakage source and the diffusion trend are identified by analyzing the spatial gradient and dynamic changes of gas concentration.

[0056] Calculation of spatial distribution gradient: Grid processing: Divide the container into grids of 1m × 1m, and associate the sensor data closest to the center point of each grid. For example, the coordinates of grid G1 are (1.5m, 0.5m), and it is associated with the gas sensor data of area 1, which is 0.5 ppm.

[0057] Gradient calculation: Use the Sobel operator to calculate the concentration gradient direction. For example, the concentration in area 1 is 0.5 ppm, the concentration in area 2 is 0.8 ppm, and the concentration in area 3 is 0.6 ppm. The horizontal gradient is (0.8 - 0.5) / 1m = 0.3 ppm / m, the vertical gradient is (0.6 - 0.5) / 1m = 0.1 ppm / m, and the combined gradient direction is 26.6° north of east.

[0058] Separation of Fourier transform components: Frequency domain analysis parameters: Perform FFT (Fast Fourier Transform) on the time series data of gas concentration (sampling rate 1 Hz), and intercept the components with frequencies lower than 0.1 Hz as the steady-state component (such as the environmental background concentration), and the components with frequencies higher than 0.1 Hz as the transient component (such as sudden leakage).

[0059] Application example: A certain gas concentration sequence shows periodic fluctuations (0.05 Hz) superimposed with sudden spikes (0.5 Hz) within 10 minutes. After FFT separation, the steady-state component is 0.6 ppm, and the transient component is 0.2 ppm.

[0060] Calculation of diffusion direction and change rate: Calculation of change rate: The slope of the transient component is used as the concentration change rate. For example, the transient component rises from 0.2 ppm to 0.5 ppm within 5 seconds, and the change rate is (0.5 - 0.2) / 5 = 0.06 ppm / s.

[0061] Diffusion Direction Judgment: Combining the gradient direction and the rate of change, if the gradient from Region 1 to Region 2 is 0.3 ppm / m and the rate of change is positive, it is determined that the leakage source is in Region 2 and diffuses eastward.

[0062] Feature Vector Generation: Each monitoring point contains four features: steady-state concentration (steady = 0.6), transient concentration (transient = 0.2), rate of change (rate = 0.06), and diffusion direction encoding (e.g., east = 1, north = 2, synthetic direction 1.26). After normalization, the feature vector is [0.6, 0.2, 0.06, 1.26].

[0063] Perform wavelet packet decomposition on the dynamic ring monitoring data, extract the proportion of high-frequency subband energy as the dynamic ring fluctuation index, and generate a dynamic environment fluctuation feature vector; This step captures high-frequency abnormal fluctuations of dynamic ring parameters (such as vibration, humidity) through wavelet packet decomposition to quantify the operating state of the device.

[0064] Wavelet Packet Decomposition Parameter Setting: Wavelet Basis Selection: Use db4 (Daubechies 4th-order wavelet) as the basis function, the decomposition level is 3 layers, and 8 subbands are generated (from low-frequency LLL to high-frequency HHH).

[0065] Decomposition Process: Perform three-layer decomposition on the dynamic ring vibration signal (sampling rate 100 Hz). The 3rd-layer high-frequency subband HHH corresponds to the frequency band of 12.5 - 25 Hz, which can characterize mechanical vibration anomalies.

[0066] Calculation of High-Frequency Energy Proportion: Energy Calculation: Calculate the sum of squares of each subband signal as the energy. For example, if the energy of the HHH subband is E_high = 1200 and the total energy E_total = 5000, then the proportion of high-frequency energy is 1200 / 5000 = 24%.

[0067] Fluctuation Index Generation: If the historical normal high-frequency proportion is 10% - 15% and the current value of 24% exceeds the threshold, it is determined as abnormal vibration.

[0068] Application Example: Input Data: A certain dynamic ring vibration signal [0.1, 0.3, …, 0.8] g (g is the unit of gravitational acceleration). After wavelet packet decomposition, the proportion of energy in the HHH subband is 24%; Feature Vector Generation: It contains three features: proportion of high-frequency energy (high_energy = 24%), proportion of medium-frequency energy (mid_energy = 50%), and proportion of low-frequency energy (low_energy = 26%). After normalization, the vector is [0.24, 0.50, 0.26].

[0069] Construct a multi-modal feature fusion model, and parallelly input the temperature anomaly feature vector, gas concentration gradient feature vector, and dynamic environment fluctuation feature vector into the attention mechanism layer, and dynamically calculate the contribution values of each feature dimension through a trainable weight matrix; In this step, the weights of different sensor features are dynamically allocated through the attention mechanism to achieve effective fusion of multi-modal features.

[0070] Model architecture design: Input layer: Temperature anomaly feature vector (3D), gas gradient feature vector (4D), dynamic environment fluctuation feature vector (3D), with a total input dimension of 10D.

[0071] Attention layer: The size of the trainable weight matrix W_attn is 10×3, corresponding to the attention scores of the three modalities of temperature, gas, and dynamic environment respectively.

[0072] Calculation of attention scores: Dynamic weight allocation: The input features pass through a fully connected layer to generate query vector Q, key vector K, and value vector V. For example, the query vector Q_temp of the temperature feature = [0.8, 0.2], the key vector K_gas of the gas feature = [0.3, 0.7], and the similarity score score = Q·K = 0.8×0.3 + 0.2×0.7 = 0.38.

[0073] Softmax normalization: The scores [0.38, 0.45, 0.17] of the three modalities are processed by Softmax to obtain the weights [0.36, 0.42, 0.22], indicating that the gas feature has the highest contribution degree.

[0074] Application example: Input features: Temperature anomaly vector [0.8, 0.65, 1.0], gas gradient vector [0.6, 0.2, 0.06, 1.26], dynamic environment fluctuation vector [0.24, 0.50, 0.26]; Weight calculation: After model training, the gas feature weight is 0.42, the temperature is 0.36, and the dynamic environment is 0.22; Weighted output: The fused feature is 0.36×temperature feature + 0.42×gas feature + 0.22×dynamic environment feature.

[0075] According to the dynamically weighted feature contribution values, feature concatenation and normalization processing are adopted to generate a risk feature vector containing multi-dimensional risk correlation relationships, and a timestamp index is attached and stored in the feature database.

[0076] This step integrates the weighted multi-modal features to generate a standardized risk feature vector and persists the storage.

[0077] Feature Cascade and Normalization Cascade Operation: Concatenate the weighted temperature, gas, and dynamic ring features in sequence. For example, after weighting, the temperature is 3-dimensional [0.29, 0.23, 0.36], the gas is 4-dimensional [0.25, 0.08, 0.02, 0.42], and the dynamic ring is 3-dimensional [0.05, 0.11, 0.06], which are combined into a 10-dimensional vector.

[0078] Min-Max Normalization: Normalize each feature dimension to [0, 1] respectively. For example, the original range of the temperature deviation is [0, 3.0], and the normalized value 0.65 → 0.65 / 3.0 ≈ 0.217.

[0079] Database Storage Design Timestamp Index: Use the Unix timestamp format (e.g., 1625097600 represents 2021-07-01 00:00:00) as the primary key.

[0080] Data Structure: Each record contains a timestamp, a 10-dimensional risk feature vector, and the hash value of the original data (such as the SHA-256 checksum).

[0081] Application Example Input: The weighted feature vector [0.29, 0.23, …, 0.06]; Normalization: After normalizing according to the maximum value of each dimension [3.0, 2.0, …, 1.0], we get [0.097, 0.115, …, 0.06]; Storage: Write to the MySQL database table risk_feature, with fields including timestamp BIGINT, feature_vector JSON, and hash CHAR(64).

[0082] Extract the temperature time series data from the environmental state matrix, calculate the local fluctuation amplitude through moving standard deviation (such as the temperature change rate exceeds 2°C / s within a 10-second window), and generate temperature anomaly features by comparing with the historical normal range; for gas concentration data, use Fourier transform to separate the steady-state component (such as background concentration) and transient component (such as sudden leakage), calculate the spatial gradient change rate and diffusion direction; perform wavelet packet decomposition on the dynamic environment parameters (such as current harmonics), and extract the proportion of high-frequency subband energy as the device anomaly fluctuation index. Dynamically allocate weights through the attention mechanism (such as temperature anomaly weight 0.6, gas gradient weight 0.3), fuse multi-modal features to generate a risk feature vector, achieve adaptive focusing on key risk dimensions, break through the limitations of single-dimensional threshold alarms, and accurately identify compound risks (such as sudden temperature rise accompanied by gas leakage) through multi-modal feature fusion and dynamic weighting, improve the sensitivity of early warning, and provide multi-dimensional quantitative basis for fire risk assessment.

[0083] S203, input the multi-dimensional risk feature vector into the fire warning optimization algorithm, and calculate the fire risk index and warning level in combination with the dynamic threshold adjustment rule trained by historical accident data, and generate a graded fire warning signal; specifically, it may include: Input the multi-dimensional risk feature vector into the pre-trained long short-term memory network to predict the risk evolution trend in the next three monitoring cycles and generate a risk time series prediction curve; The long short-term memory network (LSTM) is the core model for time series prediction. Through the time series dependence relationship of historical risk feature vectors, it predicts the fire risk evolution trend in the next three monitoring cycles (each cycle is 5 seconds).

[0084] ‌LSTM Model Architecture and Training‌: ‌Input and Output Design‌: The input layer receives a 10-dimensional risk feature vector (generated by fusing temperature, gas, and dynamic environment features), the hidden layer contains 64 LSTM units, and the output layer is a fully connected layer, predicting the risk index sequence for the next 15 seconds (3×5 seconds). When training the model, a dataset of the past 3 months (about 1.5 million records) is used, with the mean squared error (MSE) as the loss function and the Adam optimizer (learning rate 0.001) for parameter update.

[0085] ‌Application Example‌: Input the current risk feature vector [0.21, 0.35, 0.18, …, 0.29], and the LSTM outputs the predicted values [0.45, 0.62, 0.78] for the next three cycles, indicating that the risk index shows an upward trend.

[0086] ‌Prediction Curve Generation‌: Time alignment processing: The prediction results are aligned according to the timestamps. For example, if the current time is 10:00:00, the prediction curve covers three time points: 10:00:05, 10:00:10, and 10:00:15.

[0087] Confidence interval annotation: Based on the uncertainty of the model prediction, a confidence interval of ±10% is superimposed. For example, the confidence interval for a predicted value of 0.62 is 0.56 - 0.68, which is used for the robustness judgment of subsequent threshold adjustment.

[0088] Based on the fire cases in the historical accident database, extract the threshold boundaries of the key feature dimensions, and dynamically adjust the threshold sensitivity coefficient of the current environmental state through the sliding window matching algorithm; By analyzing the distribution law of the risk feature vectors in historical fire events, dynamically adjust the threshold sensitivity of the current monitoring data to adapt to the early warning requirements in different scenarios.

[0089] Historical case feature extraction: Key feature dimension screening: Extract three core features, namely temperature deviation, gas transient change rate, and the proportion of high-frequency energy in the dynamic environment, from the historical database (storing 100,000 accident records), and statistically calculate their threshold boundaries. For example, in fire cases, the average temperature deviation exceeds 2.5°C, and the gas transient change rate exceeds 0.1 ppm / s.

[0090] Sliding window matching parameters: Set the time window to 30 days, and perform similarity matching between the current environmental state and the feature distribution of historical cases within the window. For example, if the current temperature deviation is 2.8°C and the proportion of historical similar cases within the window is 15%, the sensitivity coefficient is adjusted from the baseline value of 1.0 to 1.2.

[0091] Dynamic threshold adjustment mechanism: Sensitivity coefficient calculation: According to the matching result, adjust the threshold through linear interpolation. For example, if the matching similarity is 60%, the sensitivity coefficient k = 1.0 + 0.2×(60 / 100) = 1.12, and the temperature deviation threshold is adjusted from 2.5°C to 2.5×1.12 = 2.8°C.

[0092] Application example: On a certain day, the environmental humidity is high (the dynamic environment humidity sensor shows 85%). Historical data shows that the gas diffusion speed increases in a high-humidity environment. The system lowers the gas transient change rate threshold from 0.1 ppm / s to 0.08 ppm / s to improve the early warning sensitivity.

[0093] Calculate the Mahalanobis distance between the real-time multi-dimensional risk feature vector and the dynamic risk threshold, and combine the slope change of the risk time series prediction curve to comprehensively evaluate the fire risk index; The Mahalanobis Distance is used to measure the degree of deviation between the real-time feature vector and the historical fire feature distribution. Combining with the trend slope of the prediction curve, it quantifies the comprehensive risk level.

[0094] ‌Mahalanobis Distance calculation process: ‌Covariance matrix construction: Based on the feature vectors (10-dimensional) of historical fire cases, calculate the covariance matrix between each dimension (size 10×10) to characterize the correlation between features. For example, the covariance between the temperature deviation and the gas transient change rate is 0.35, indicating a positive correlation between the two.

[0095] ‌Distance calculation example: The real-time feature vector is [0.21, 0.35, …, 0.29], the mean of historical fire features is [0.50, 0.60, …, 0.75], and the Mahalanobis Distance is calculated after inverting the covariance matrix, with the result being 3.8 (the average distance of historical fire cases is 2.5, and the threshold is set to 3.0).

[0096] ‌Analysis of the slope of the prediction curve: ‌Slope calculation method: Perform a linear fit on the predicted risk indices for three cycles [0.45, 0.62, 0.78] to obtain the slope k = (0.78 - 0.45) / 15s = 0.022 / s.

[0097] ‌Weight allocation: The Mahalanobis Distance accounts for 70% of the comprehensive risk index, and the slope accounts for 30%. For example, the Mahalanobis Distance score is 3.8 / 5.0 = 0.76, the slope score is 0.022 / 0.03 = 0.73, and the comprehensive risk index is 0.76×0.7 + 0.73×0.3 = 0.75.

[0098] Four warning levels are defined according to the magnitude of the fire risk index. When the fire risk index exceeds the preset level threshold, a cross-regional risk linkage verification mechanism is triggered; The four warning levels correspond to different emergency response strategies, and false alarms are avoided through cross-regional data verification.

[0099] ‌Warning level division rules: ‌Definition of level thresholds: Level 1 (red, risk index ≥ 0.8), Level 2 (orange, 0.6 ≤ index < 0.8), Level 3 (yellow, 0.4 ≤ index < 0.6), Level 4 (blue, index < 0.4). For example, a comprehensive risk index of 0.75 triggers a Level 2 warning.

[0100] ‌Trigger conditions for linkage verification: When the risk index ≥ 0.6, send a verification request to the BMS system of adjacent containers to obtain their temperature and gas concentration data, and confirm whether there is risk diffusion between regions.

[0101] ‌Cross-regional verification process: Data Request and Response: Send real-time data query instructions to adjacent containers via the Modbus TCP protocol, with the timeout set to 500 ms. For example, if the risk index of the current container A is 0.75 and it requests the temperature data of container B, and the temperature deviation of B is 2.1 °C (lower than 2.8 °C of A), it is determined as a local risk; otherwise, the regional linkage plan is activated.

[0102] Verification Result Fusion: If 2 out of 3 adjacent containers return confirmation signals, the early warning level is raised. For example, container A triggers a level-two early warning, and the adjacent containers B and C return risk indexes of 0.68 and 0.71, and the system raises the early warning level from level two to level one.

[0103] Fuse the real-time assessment results and the linkage verification feedback, generate a graded fire early warning signal with confidence level annotation, and encrypt and transmit it to the BMS control center.

[0104] The final early warning signal integrates the real-time analysis results and the cross-regional verification conclusions, attaches a confidence level assessment, and ensures transmission security through an encrypted channel.

[0105] Signal Generation and Annotation: Confidence Level Calculation: Calculate the confidence level of the current risk index according to the statistical distribution of the Mahalanobis distance (assuming it conforms to the chi-square distribution). For example, the Mahalanobis distance of 3.8 corresponds to the chi-square distribution with 10 degrees of freedom, and the confidence level P = 92%.

[0106] Signal Data Structure: The early warning signal contains fields: timestamp (1625097600), risk level (level two), confidence level (92%), and linkage verification result (2 / 3 confirmations).

[0107] Encryption Transmission Mechanism: Encryption Algorithm Selection: Use AES-256 (Advanced Encryption Standard, 256-bit key) symmetric encryption, and the key is rotated every 24 hours.

[0108] Transmission Protocol Example: The early warning signal is encapsulated in JSON format and sent to the BMS control center via the HTTPS protocol, with a response time less than 200 ms. For example, the sent content is {"timestamp":1625097600,"level":"orange","confidence":0.92,"validation":"2 / 3"}, and the encrypted length is 128 bytes.

[0109] Predict the future risk trend (such as the temperature rise slope within 30 seconds) based on the long short-term memory network (LSTM), combine the threshold boundaries of similar scenarios in the historical accident database (such as the critical gas concentration corresponding to the ignition point of a certain type of cargo), and dynamically adjust the current threshold sensitivity (such as reducing the gas concentration alarm threshold in a high-temperature environment). By calculating the Mahalanobis distance between the real-time feature vector and the dynamic threshold (considering multi-dimensional covariance), comprehensively evaluate the fire risk index (scaled from 0 to 100), and divide it into four levels of early warning (such as level one: risk index ≥ 80, triggering full-area fire extinguishing). Introduce cross-regional linkage verification (such as cross-verification of adjacent container data) to improve the confidence of early warning, and finally generate an encrypted fire warning signal, realizing threshold adaptive optimization and risk dynamic classification, avoiding false alarms caused by fixed thresholds (such as normal fluctuations in environmental temperature), supporting a smooth transition from "low-risk monitoring" to "emergency disposal", and providing a decision-making basis for precise fire intervention.

[0110] S204, according to the classified fire warning signal, trigger an active fire intervention strategy through the BMU unit of the BMS, dynamically control the fire extinguishing device, ventilation system, and power cut-off module, and synchronously optimize the fire response path to match the current risk level. Specifically, it may include: Analyze the risk level of the classified fire warning signal, match the preset fire strategy template library, and plan an initial intervention instruction set including the start priority of the fire extinguishing device, ventilation rate gear, and power cut-off delay; The fire strategy template library is a set of instructions predefined based on the risk level and container scenario characteristics, used to quickly generate an initial fire intervention plan.

[0111] ‌Construction and matching of the fire strategy template library‌ ‌Template structure‌: The template library is stored in JSON format. Each template contains three core fields: risk_level (risk level, 1-4 levels), fire_extinguisher_priority (list of start priorities of the fire extinguishing device), ventilation_rate (target gear of the ventilation system, 1-5 gears), power_cut_delay (power cut-off delay, in milliseconds). For example, the template corresponding to the first-level early warning (red) is: { "risk_level": 1, "fire_extinguisher_priority": ["A1-nozzle", "B2-powder", "C3-foam"], "ventilation_rate": 5, "power_cut_delay": 200 }。

[0112] Matching logic: Parse the risk_level field in the warning signal (e.g., risk_level = 1), traverse all entries in the template library that match the risk_level. If there are multiple matching templates (e.g., different container types), further filter according to the container ID (e.g., container_type = ESS-01).

[0113] Generation of the initial intervention instruction set Dynamic adjustment of priority: According to the internal fire source location result of the container (such as infrared thermal imaging data), adjust the start order of the fire extinguishing devices. For example, if the fire source is in area A, the A1 nozzle is started first, and ["A1-nozzle", "B2-powder", "C3-foam"] in the original template remains unchanged; if the fire source spreads to area B, it is adjusted to ["B2-powder", "A1-nozzle", "C3-foam"].

[0114] Calculation of delay parameters: The power cut delay power_cut_delay is dynamically adjusted according to the remaining battery capacity. For example, if the battery capacity is lower than 20% (reported by the BMS), the delay is shortened from 200 ms to 100 ms to avoid secondary risks caused by electric sparks.

[0115] According to the internal topology map of the container and the device deployment location, use the Dijkstra algorithm to calculate the optimal response path of each fire-fighting device, eliminate action conflicts and optimize the response timing of the initial intervention instruction set to obtain the fire-fighting instruction set; The internal topology map of the container is represented in a graph structure (Graph), where the nodes (Nodes) are fire-fighting devices or key locations, and the edge (Edge) weights are jointly determined by the physical distance and the device response time.

[0116] Topology map modeling and application of the Dijkstra algorithm Node definition: Each fire-fighting device (such as fire nozzle A1, vent V3) and power supply node (such as P1 circuit breaker) is used as a node in the graph, with a total of N nodes (e.g., N = 15).

[0117] Edge weight calculation: The weight of the edge weight = distance × 0.3 + response_time, where distance is the physical distance between two nodes (in meters) and response_time is the device startup time (in milliseconds). For example, the distance from A1 to V3 is 2 meters, the response time of A1 is 50 ms, and that of V3 is 30 ms, then the edge weight is 2×0.3 + (50 + 30) = 80.6.

[0118] Path planning example: With the goal of "starting nozzle A1 and vent V3 simultaneously", the Dijkstra algorithm calculates the shortest paths from the control center to A1 and V3. If the shortest path weight of A1 is 80.6 and that of V3 is 75.2, an instruction sequence [V3_start@t=0ms, A1_start@t=30ms] is generated to ensure that the ventilation system starts before the fire extinguishing device, avoiding airflow interference with the spread of the fire extinguishing agent.

[0119] Conflict detection and timing optimization Resource competition detection: Detect conflicts in the occupation of the same physical resource (such as circuit channels, air pump pressure) by multiple instructions. For example, if nozzle A1 and powder device B2 share the same air pump pipeline, the system calculates the maximum parallelism based on the air pump flow rate (such as 200L / min). If A1 requires 150L / min and B2 requires 100L / min, a serial instruction A1_start → B2_start (with a 50ms interval) is generated.

[0120] Response timing rearrangement: Based on the device response time and path weight, the initial instruction set is optimized by bubble sorting. For example, the original instruction set is [A1@80ms, V3@75ms, P1@200ms], and after sorting, it becomes [V3@75ms, A1@80ms, P1@200ms].

[0121] Simulate the execution of the fire fighting instruction set on the edge computing node, verify the change trend of the environmental parameters after the instruction execution through the digital twin model, and trigger instruction backtracking and replanning if the simulation results do not meet the expectations; The digital twin model is a simulation engine based on physical laws and data-driven, which can predict the temperature, gas concentration, and smoke diffusion effect after the execution of the fire fighting instructions.

[0122] Digital twin modeling and parameter initialization Physical field modeling: The container is divided into 1,000 grid cells using the finite element method (FEM), and the initial values of temperature, gas concentration, and air flow velocity are defined for each cell (obtained from the environmental state matrix). For example, the initial temperature of the grid where the fire source is located is 300°C, and the rest of the area is 25°C.

[0123] Device action mapping: Convert the fire fighting instructions into simulation parameters. For example, the start instruction of nozzle A1 corresponds to injecting the fire extinguishing agent into grids G5 - G8 in the simulation (flow rate 150L / min), and ventilation gear 5 corresponds to a wind speed of 3m / s.

[0124] Simulation execution and result verification Key Index Prediction: The simulation running time is 10 seconds (simulating the real response process), and the temperature drop rate (e.g., from 300°C → 250°C → 200°C) and gas concentration dilution rate (e.g., from 500 ppm → 300 ppm → 100 ppm) are output once per second.

[0125] Compliance Condition Judgment: If the temperature of the fire source grid is lower than 80°C and the smoke concentration is lower than 50 ppm after the simulation ends, the instruction is determined to be valid; otherwise, re-planning is triggered. For example, if the first simulation result shows that the temperature only drops to 120°C, the system automatically returns to the instruction generation stage, increases the ventilation gear to 5, and extends the fire extinguishing agent injection time to 8 seconds.

[0126] Instruction Backtracking Mechanism Re-planning Strategy: If the simulation fails, the system selects the optimization direction according to the type of unmet indicators. For example: Temperature Not Meeting the Standard: Prioritize increasing the number of activated fire extinguishing devices (such as adding B2 powder injection); Gas Concentration Not Meeting the Standard: Increase the ventilation gear or extend the ventilation time.

[0127] Iteration Limit: A maximum of 3 re-planning iterations are allowed. If the standard is still not met, an artificial intervention warning is triggered (notifying the operation and maintenance personnel through the BMS central unit).

[0128] The final coordinated scheduling instruction that passes the verification is sent to the BMU unit, synchronously controlling the spraying angle of the fire extinguishing device, the variable-frequency fan of the ventilation system, and the intelligent circuit breaker, and real-time transmitting the execution status back to the central monitoring platform.

[0129] The final instruction is sent to the BMU unit through the BMS communication protocol (such as CAN bus or Modbus TCP) to achieve multi-device collaborative control and status monitoring.

[0130] Instruction Encoding and Sending Instruction Encapsulation Format: The instruction is encoded using the TLV (Tag-Length-Value) structure. For example, the Tag of the fire extinguishing device control instruction is 0x01, the Length is 4 bytes, and the Value is the spraying angle (0 - 180°) and the duration (ms), specifically 0104 0000B4 07D0 (angle 180°, duration 2000 ms).

[0131] Synchronous Control Mechanism: The BMU unit is built with a high-precision timer (error ±1 ms), and multiple device actions are triggered simultaneously at the specified timestamp (such as Unix timestamp 1625097600123). For example, at t = 1625097600123, the ventilation fan switches to gear 5 (instruction VENT = 5), and the fire extinguishing nozzle A1 opens for 180° spraying.

[0132] Detailed Explanation of Equipment Control Parameters Spraying Angle of Fire Extinguishing Device: Dynamically adjusted according to the fire source location. For example, when the fire source is in Area A (coordinates x = 2.3m, y = 1.5m), the Euler angles of Nozzle A1 are calculated as azimuth angle 45° and pitch angle 30°.

[0133] Parameters of Ventilation Variable Frequency Fan: Gear 5 corresponds to a fan speed of 2,000 RPM (Revolutions Per Minute), an air volume of 800m³ / h, and the motor current is adjusted in real time by a PID controller (target value 4.5A).

[0134] Intelligent Circuit Breaker Delay: power_cut_delay = 200ms means that the BMU cuts off Circuit Breaker P1 200ms after receiving the command. During this period, if a sudden drop in current is detected (such as the battery pack being disconnected), the delay is immediately terminated and an emergency power-off is executed.

[0135] Status Transmission and Monitoring Transmission Protocol and Frequency: The device status is reported every 500ms through the MQTT protocol. The data packet contains fields: device ID, current status, parameter values, and timestamp. For example: { "device_id": "A1-nozzle", "status": "active", "angle": 180, "flow_rate": 150, "timestamp": 1625097600500 }

[0136] Exception Handling Mechanism: If the device does not respond within the expected time (such as the status of Nozzle A1 not becoming "active"), the BMU starts the redundant device (such as switching to Powder Device B2) within 3 seconds and generates a fault log.

[0137] After analyzing the warning levels, match the preset policy library (for example, the ventilation system is preferentially activated for secondary warnings, and the fire extinguishing device is linked for tertiary warnings). Based on the internal topological map of the container (such as the location of fire extinguishers and the layout of ventilation ducts), use the Dijkstra algorithm to optimize the device response path (such as activating the nearest fire sprinkler at the shortest path). Simulate the execution effect of the instruction through the digital twin model (such as whether the coverage area of the fire extinguishing agent includes the fire source point), dynamically adjust the spraying angle, fan speed, and power-off timing, and eliminate device action conflicts (such as interference between ventilation and fire extinguishing airflows). Finally, issue collaborative control instructions, and transmit the execution status (such as the pressure value of the fire extinguishing device and the status of the circuit breaker) to the central platform in real time to form a closed-loop control, solving the problems of rigid response and low device collaboration efficiency in traditional fire protection systems, and minimizing device loss and secondary risks while ensuring safety.

[0138] It can be seen that by synchronously collecting the temperature, combustible gas concentration, and dynamic environment monitoring data inside the container through the BMU unit of the BMS, a spatio-temporally correlated container environmental state matrix is generated; based on the environmental state matrix, temperature anomaly features, gas concentration gradient features, and dynamic environmental fluctuation features are extracted through a multi-modal feature fusion model, and a multi-dimensional risk feature vector is constructed; the multi-dimensional risk feature vector is input into the fire warning optimization algorithm to generate a hierarchical fire warning signal; according to the hierarchical fire warning signal, the BMU unit of the BMS triggers an active fire intervention strategy, dynamically controls the fire extinguishing device, ventilation system, and power cut-off module, and synchronously optimizes the fire response path to match the current risk level, so as to achieve hierarchical fire warning and improve the accuracy and timeliness of fire warning.

[0139] Another embodiment of the present invention provides a multi-functional integrated monitoring and fire control system based on BMS. Refer to Figure 3 , the system may include: A fusion module 301, configured to synchronously collect the temperature, combustible gas concentration, and dynamic environment monitoring data inside the container through the BMU unit of the BMS, and perform fusion processing on the multi-source data by using the spatio-temporal alignment algorithm of the heterogeneous sensor network to generate a spatio-temporally correlated container environmental state matrix; A construction module 302, configured to extract temperature anomaly features, gas concentration gradient features, and dynamic environmental fluctuation features based on the environmental state matrix through a multi-modal feature fusion model, and construct a multi-dimensional risk feature vector, wherein the multi-modal feature fusion model dynamically weights the contribution values of different monitoring dimensions through an attention mechanism; A grading module 303, configured to input the multi-dimensional risk feature vector into the fire warning optimization algorithm, calculate the fire risk index and warning level in combination with the dynamic threshold adjustment rule trained by historical accident data, and generate a hierarchical fire warning signal; The control module 304 is configured to trigger an active fire intervention strategy through the BMU unit of the BMS according to the hierarchical fire warning signal, dynamically control the fire extinguishing device, the ventilation system and the power cut-off module, and synchronously optimize the fire response path to match the current risk level.

[0140] It can be seen that the BMU unit of the BMS is used to synchronously collect the temperature, combustible gas concentration and dynamic environment monitoring data in the container to generate a spatio-temporally correlated container environment state matrix. Based on the environment state matrix, a multi-dimensional risk feature vector is constructed by extracting temperature anomaly features, gas concentration gradient features and dynamic environment fluctuation features through a multi-modal feature fusion model. The multi-dimensional risk feature vector is input into the fire warning optimization algorithm to generate a hierarchical fire warning signal. According to the hierarchical fire warning signal, the BMU unit of the BMS triggers an active fire intervention strategy, dynamically controls the fire extinguishing device, the ventilation system and the power cut-off module, and synchronously optimizes the fire response path to match the current risk level, so as to realize hierarchical fire warning and improve the accuracy and timeliness of fire warning.

[0141] An embodiment of the present invention also provides a storage medium, in which a computer program is stored, and the computer program is configured to execute the steps in any one of the above method embodiments when running.

[0142] Specifically, in this embodiment, the above storage medium can be configured to store a computer program for executing the following steps: S201, synchronously collect the temperature, combustible gas concentration and dynamic environment monitoring data in the container through the BMU unit of the BMS, and use the spatio-temporal alignment algorithm of the heterogeneous sensor network to fuse and process the multi-source data to generate a spatio-temporally correlated container environment state matrix; S202, based on the environment state matrix, extract temperature anomaly features, gas concentration gradient features and dynamic environment fluctuation features through a multi-modal feature fusion model to construct a multi-dimensional risk feature vector, wherein the multi-modal feature fusion model dynamically weights the contribution values of different monitoring dimensions through an attention mechanism; S203, input the multi-dimensional risk feature vector into the fire warning optimization algorithm, combine the dynamic threshold adjustment rule trained with historical accident data, calculate the fire risk index and warning level, and generate a hierarchical fire warning signal; S204, according to the hierarchical fire warning signal, trigger an active fire intervention strategy through the BMU unit of the BMS, dynamically control the fire extinguishing device, the ventilation system and the power cut-off module, and synchronously optimize the fire response path to match the current risk level.

[0143] It can be seen that the BMU unit of the BMS synchronously collects the temperature, combustible gas concentration and dynamic environment monitoring data in the container to generate a spatio-temporal associated container environmental state matrix; based on the environmental state matrix, a multi-modal feature fusion model is used to extract temperature anomaly features, gas concentration gradient features and dynamic environmental fluctuation features to construct a multi-dimensional risk feature vector; the multi-dimensional risk feature vector is input into the fire warning optimization algorithm to generate a hierarchical fire warning signal; according to the hierarchical fire warning signal, the BMU unit of the BMS triggers an active fire intervention strategy to dynamically control the fire extinguishing device, ventilation system and power cut-off module, and synchronously optimize the fire response path to match the current risk level, so as to achieve hierarchical fire warning and improve the accuracy and timeliness of fire warning.

[0144] An embodiment of the present invention further provides an electronic device, including a memory and a processor, wherein a computer program is stored in the memory, and the processor is configured to run the computer program to execute the steps in any one of the above method embodiments.

[0145] Specifically, the above electronic device may further include a transmission device and an input / output device, wherein the transmission device is connected to the above processor, and the input / output device is connected to the above processor.

[0146] Specifically, in this embodiment, the above processor may be configured to execute the following steps through a computer program: S201, synchronously collect the temperature, combustible gas concentration and dynamic environment monitoring data in the container through the BMU unit of the BMS, and use the spatio-temporal alignment algorithm of the heterogeneous sensor network to fuse and process the multi-source data to generate a spatio-temporal associated container environmental state matrix; S202, based on the environmental state matrix, use a multi-modal feature fusion model to extract temperature anomaly features, gas concentration gradient features and dynamic environmental fluctuation features to construct a multi-dimensional risk feature vector, wherein the multi-modal feature fusion model dynamically weights the contribution values of different monitoring dimensions through an attention mechanism; S203, input the multi-dimensional risk feature vector into the fire warning optimization algorithm, combine the dynamic threshold adjustment rule trained with historical accident data, calculate the fire risk index and warning level, and generate a hierarchical fire warning signal; S204, according to the hierarchical fire warning signal, trigger an active fire intervention strategy through the BMU unit of the BMS to dynamically control the fire extinguishing device, ventilation system and power cut-off module, and synchronously optimize the fire response path to match the current risk level.

[0147] It can be seen that the BMU unit of the BMS synchronously collects the temperature, combustible gas concentration and dynamic environment monitoring data in the container to generate a spatio-temporal correlated container environmental state matrix; based on the environmental state matrix, a multi-modal feature fusion model is used to extract temperature anomaly features, gas concentration gradient features and dynamic environmental fluctuation features to construct a multi-dimensional risk feature vector; the multi-dimensional risk feature vector is input into the fire warning optimization algorithm to generate a hierarchical fire warning signal; according to the hierarchical fire warning signal, the BMU unit of the BMS triggers an active fire intervention strategy to dynamically control the fire extinguishing device, ventilation system and power cut-off module, and synchronously optimize the fire response path to match the current risk level, so as to achieve hierarchical fire warning and improve the accuracy and timeliness of fire warning.

[0148] The structure, features and effects of the present invention have been described in detail based on the embodiments shown in the drawings. The above is only the preferred embodiment of the present invention, but the present invention is not limited to the scope shown in the drawings. Any changes made according to the concept of the present invention, or equivalent embodiments modified into equivalent changes, should still be within the protection scope of the present invention as long as they do not exceed the spirit covered by the specification and the drawings.

Claims

1. A multi-functional integrated monitoring and fire control method based on BMS, characterized in that, The method includes: Synchronously collecting the temperature, concentration of combustible gas, and dynamic environment monitoring data inside the container through the BMU unit of the BMS, and using the spatio-temporal alignment algorithm of heterogeneous sensor networks to fuse and process multi-source data to generate a spatio-temporally correlated container environmental state matrix; Based on the environmental state matrix, extracting temperature anomaly features, gas concentration gradient features, and dynamic environment fluctuation features through a multi-modal feature fusion model to construct a multi-dimensional risk feature vector, where the multi-modal feature fusion model dynamically weights the contribution values of different monitoring dimensions through an attention mechanism; Inputting the multi-dimensional risk feature vector into a fire warning optimization algorithm, and combining the dynamic threshold adjustment rule trained with historical accident data to calculate the fire risk index and warning level, and generating a graded fire warning signal; According to the graded fire warning signal, triggering an active fire intervention strategy through the BMU unit of the BMS, dynamically controlling the fire extinguishing device, ventilation system, and power cut-off module, and synchronously optimizing the fire response path to match the current risk level.

2. The method according to claim 1, wherein The step of synchronously collecting the temperature, concentration of combustible gas, and dynamic environment monitoring data inside the container through the BMU unit of the BMS, and using the spatio-temporal alignment algorithm of heterogeneous sensor networks to fuse and process multi-source data to generate a spatio-temporally correlated container environmental state matrix includes: According to the original data streams of the temperature sensor, combustible gas sensor, and dynamic environment monitoring device collected by the BMU unit of the BMS, performing timestamp synchronization calibration, and eliminating the sampling frequency differences of each sensor through sliding window mean compensation to generate a time-aligned original data set; Performing spatial position calibration on the original data set, establishing a spatial coordinate mapping table based on the sensor deployment topology structure, and using the covariance matrix matching algorithm to eliminate cross-regional monitoring errors to generate a spatially aligned multi-dimensional monitoring matrix; According to the multi-dimensional monitoring matrix, using the spatio-temporal alignment algorithm based on heterogeneous sensor networks, and combining the Gaussian mixture model to perform probability distribution modeling on the temperature, gas concentration, and dynamic environment parameters, calculating the cross-modal data correlation degree, and generating a spatio-temporal correlation confidence map; According to the spatio-temporal correlation confidence map, using Bayesian inference to fuse multi-source data, weighted correction of conflicting monitoring values, and outputting a spatio-temporally correlated container environmental state matrix and storing it in the edge computing node.

3. The method according to claim 2, wherein The step of based on the environmental state matrix, extracting temperature anomaly features, gas concentration gradient features, and dynamic environment fluctuation features through a multi-modal feature fusion model to construct a multi-dimensional risk feature vector, where the multi-modal feature fusion model dynamically weights the contribution values of different monitoring dimensions through an attention mechanism includes: Extracting temperature time series data from the environmental state matrix, calculating the local fluctuation amplitude through moving standard deviation, and performing differential analysis in combination with the historical normal temperature range to generate a temperature anomaly feature vector; Analyzing the spatial distribution gradient of the gas concentration data, separating the steady-state component and transient component by Fourier transform, and calculating the concentration change rate and diffusion direction of each monitoring point to generate a gas concentration gradient feature vector; Performing wavelet packet decomposition on the dynamic environment monitoring data, extracting the high-frequency sub-band energy ratio as the dynamic environment fluctuation index, and generating a dynamic environment fluctuation feature vector; Build a multi-modal feature fusion model, and parallelly input the temperature anomaly feature vector, gas concentration gradient feature vector, and dynamic environment fluctuation feature vector into the attention mechanism layer, and dynamically calculate the contribution values of each feature dimension through a trainable weight matrix; According to the dynamically weighted feature contribution values, perform feature concatenation and normalization processing to generate a risk feature vector containing multi-dimensional risk association relationships, and attach a timestamp index and store it in the feature database.

4. The method according to claim 3, wherein Input the multi-dimensional risk feature vector into the fire warning optimization algorithm, and combine the dynamic threshold adjustment rule trained with historical accident data to calculate the fire risk index and warning level, and generate a graded fire warning signal, including: Input the multi-dimensional risk feature vector into a pre-trained long short-term memory network to predict the risk evolution trend in the next three monitoring cycles and generate a risk time series prediction curve; Based on the fire cases in the historical accident database, extract the threshold boundaries of the key feature dimensions, and dynamically adjust the threshold sensitivity coefficient of the current environmental state through a sliding window matching algorithm; Calculate the Mahalanobis distance between the real-time multi-dimensional risk feature vector and the dynamic risk threshold, and combine the slope change of the risk time series prediction curve to comprehensively evaluate the fire risk index; Divide the four-level warning level according to the size of the fire risk index. When the fire risk index exceeds the preset level threshold, trigger the cross-regional risk linkage verification mechanism; Fuse the real-time evaluation results and the linkage verification feedback, generate a graded fire warning signal with confidence annotation, and encrypt and transmit it to the BMS control center.

5. The method according to claim 4, wherein According to the graded fire warning signal, trigger an active fire intervention strategy through the BMU unit of the BMS, dynamically control the fire extinguishing device, ventilation system, and power cut-off module, and synchronously optimize the fire response path to match the current risk level, including: Analyze the risk level of the graded fire warning signal, match the preset fire strategy template library, and plan an initial intervention instruction set including the start priority of the fire extinguishing device, ventilation rate gear, and power cut-off delay; According to the internal topology map of the container and the deployment location of the equipment, use the Dijkstra algorithm to calculate the optimal response path of each fire equipment, eliminate action conflicts and optimize the response timing of the initial intervention instruction set to obtain a fire instruction set; Simulate the execution of the fire instruction set at the edge computing node, and verify the change trend of the environmental parameters after the instruction execution through the digital twin model. If the simulation result does not meet the expectation, trigger the instruction backtracking and replanning; Send the finally verified collaborative scheduling instruction to the BMU unit, synchronously control the spraying angle of the fire extinguishing device, the variable frequency fan of the ventilation system, and the intelligent circuit breaker, and real-time transmit the execution status back to the central monitoring platform.

6. A multi-functional integrated monitoring and fire control system based on BMS, characterized in that, The system includes: A fusion module, which is used to synchronously collect the temperature, combustible gas concentration, and dynamic environment monitoring data in the container through the BMU unit of the BMS, and use the heterogeneous sensor network spatio-temporal alignment algorithm to perform fusion processing on the multi-source data to generate a spatio-temporally correlated container environmental state matrix; A building block for extracting temperature anomaly features, gas concentration gradient features, and dynamic environment fluctuation features based on the environmental state matrix through a multi-modal feature fusion model, and constructing a multi-dimensional risk feature vector, where the multi-modal feature fusion model dynamically weights the contribution values of different monitoring dimensions through an attention mechanism; A grading module for inputting the multi-dimensional risk feature vector into a fire warning optimization algorithm, calculating a fire risk index and a warning level in combination with a dynamic threshold adjustment rule trained with historical accident data, and generating a graded fire warning signal; A control module for triggering an active fire intervention strategy through the BMU unit of the BMS according to the graded fire warning signal, dynamically controlling a fire extinguishing device, a ventilation system, and a power cut-off module, and synchronously optimizing the fire response path to match the current risk level.

7. The system according to claim 6, wherein The fusion module is specifically used for: Performing timestamp synchronization calibration on the original data streams of temperature sensors, combustible gas sensors, and dynamic environment monitoring devices collected by the BMU unit of the BMS, and eliminating the sampling frequency differences of each sensor through sliding window mean compensation to generate a time-aligned original data set; Performing spatial position calibration on the original data set, establishing a spatial coordinate mapping table based on the sensor deployment topology structure, and eliminating cross-regional monitoring errors by using a covariance matrix matching algorithm to generate a spatially aligned multi-dimensional monitoring matrix; According to the multi-dimensional monitoring matrix, using a spatio-temporal alignment algorithm based on a heterogeneous sensor network, and combining a Gaussian mixture model to perform probability distribution modeling on temperature, gas concentration, and dynamic environment parameters, calculating the cross-modal data correlation degree, and generating a spatio-temporal correlation confidence map; According to the spatio-temporal correlation confidence map, using Bayesian inference to fuse multi-source data, weighted correction of conflicting monitoring values, and outputting a spatio-temporally correlated container environmental state matrix and storing it in an edge computing node.

8. The system according to claim 7, wherein The building block is specifically used for: Extracting temperature time-series data from the environmental state matrix, calculating the local fluctuation amplitude through moving standard deviation, and performing differential analysis in combination with the historical normal temperature range to generate a temperature anomaly feature vector; Analyzing the spatial distribution gradient of gas concentration data, separating the steady-state component and the transient component by using Fourier transform, and calculating the concentration change rate and diffusion direction of each monitoring point to generate a gas concentration gradient feature vector; Performing wavelet packet decomposition on the dynamic environment monitoring data, extracting the high-frequency sub-band energy ratio as the dynamic environment fluctuation index, and generating a dynamic environment fluctuation feature vector; Constructing a multi-modal feature fusion model, parallelly inputting the temperature anomaly feature vector, the gas concentration gradient feature vector, and the dynamic environment fluctuation feature vector into the attention mechanism layer, and dynamically calculating the contribution values of each feature dimension through a trainable weight matrix; According to the dynamically weighted feature contribution values, using feature concatenation and normalization processing to generate a risk feature vector containing multi-dimensional risk correlation relationships, and attaching a timestamp index to store it in a feature database.

9. A storage medium, characterized in that, A computer program is stored in the storage medium, where the computer program is set to execute the method according to any one of claims 1-5 when running.

10. An electronic device, comprising a memory and a processor, characterized in that, A computer program is stored in the memory, and the processor is configured to run the computer program to execute the method according to any one of claims 1-5.

Images