Mine vehicle early fire warning method and system based on multi-source information fusion

Abstract

The application discloses a mine vehicle early fire warning method and system based on multi-source information fusion, and belongs to the technical field of fire alarm, wherein the method comprises the following steps: multi-source monitoring of the mine vehicle is performed to obtain vehicle environment, working condition and visual monitoring data; fire sign sensitive feature detection is performed on the vehicle environment monitoring data; fire sign sensitive feature detection is performed on the vehicle working condition data; fire sign sensitive feature detection is performed on the vehicle visual monitoring data; single-factor fire deduction is performed on the environment, working condition and visual fire sensitive feature vectors according to a vehicle fire history event set; multi-factor coupled fire deduction is performed according to a first fire deduction graph, and early fire warning is performed according to a second fire deduction graph. The application solves the technical problems that the existing mine vehicle fire warning relies on single-dimensional monitoring data, the information is one-sided, early multi-factor collaborative risks are difficult to capture, and thus fire misses, misjudgments and untimely warnings are caused.

Description

Technical Field

[0001] This invention relates to the field of fire alarm technology, specifically to a method and system for early fire warning of mining vehicles based on multi-source information fusion. Background Technology

[0002] Mining vehicles are core equipment for underground transportation in mines. The working environment is enclosed and dusty, and the equipment is under high load. Fires can spread rapidly, causing explosions and casualties. Early warning is a key link in ensuring safe production in mines.

[0003] Current early warning systems for fires in mining vehicles largely rely on single-dimensional monitoring technologies: they may only use temperature / smoke sensors to monitor environmental parameters, making it difficult to capture visual features such as localized high temperatures and burn marks; or they may only rely on operating condition sensors to monitor current and oil pressure, failing to correlate these with the accelerated aging effect of high environmental temperatures on equipment; and they lack collaborative analysis of multi-dimensional data, making it difficult to identify early coupling risks between the environment, operating conditions, and vision, resulting in high rates of missed and false alarms. Often, alarms are only triggered after the fire has spread, missing the opportunity for early response. Summary of the Invention

[0004] This application provides a method and system for early fire warning of mining vehicles based on multi-source information fusion. It aims to solve the technical problems in the existing technology of early fire warning of mining vehicles that rely on single-dimensional monitoring data, which results in one-sided information and difficulty in capturing early multi-factor collaborative risks, thus leading to missed detection, misjudgment and untimely warning of fire.

[0005] In view of the above problems, this application provides a method and system for early fire warning of mining vehicles based on multi-source information fusion.

[0006] The first aspect disclosed in this application provides a method for early fire warning of mining vehicles based on multi-source information fusion. This method includes: performing multi-source monitoring of mining vehicles to obtain vehicle environmental monitoring data, vehicle operating condition data, and vehicle visual monitoring data; performing fire sign sensitive feature detection on the vehicle environmental monitoring data based on a historical fire event set to obtain an environmental fire sensitive feature vector; performing fire sign sensitive feature detection on the vehicle operating condition data based on the historical fire event set to obtain an operating condition fire sensitive feature vector; performing fire sign sensitive feature detection on the vehicle visual monitoring data based on the historical fire event set to obtain a visual fire sensitive feature vector; performing single-factor fire prediction based on the environmental fire sensitive feature vector, the operating condition fire sensitive feature vector, and the visual fire sensitive feature vector to obtain a first fire prediction map; performing multi-factor coupled fire prediction based on the first fire prediction map to obtain a second fire prediction map; and performing early fire warning based on the second fire prediction map.

[0007] Furthermore, based on the vehicle fire history event set, fire sign sensitive feature detection is performed on the vehicle environmental monitoring data to obtain an environmental fire sensitive feature vector. This includes: collecting sensor feature data corresponding to each environmental monitoring parameter within the vehicle environmental monitoring data to obtain multiple monitoring sensor feature sequences; performing anomaly detection based on the multiple monitoring sensor feature sequences to obtain multiple sensor anomaly detection results; performing adaptive interference correction on the vehicle environmental monitoring data based on the multiple sensor anomaly detection results to obtain an environmental monitoring feature sequence; evaluating the fire indicativeness of the environmental monitoring feature sequence based on the vehicle fire history event set to obtain an environmental monitoring fire indicativeness sequence; and performing fire sign sensitive attention allocation on the environmental monitoring feature sequence based on the environmental monitoring fire indicativeness sequence to generate the environmental fire sensitive feature vector.

[0008] Further, the environmental monitoring feature sequence is evaluated for fire indicativeness based on the historical vehicle fire event set to obtain an environmental monitoring fire indicativeness sequence. This includes: identifying fire sign features based on the historical vehicle fire event set to obtain a fire sign feature library; evaluating the confidence level of each fire sign feature in the fire sign feature library based on the historical vehicle fire event set to obtain a fire feature confidence sequence; cleaning the fire sign feature library based on the fire feature confidence sequence and a confidence fire sign feature space; extracting the Rth environmental monitoring feature from the environmental monitoring feature sequence, where R is a positive integer; evaluating the Rth environmental monitoring feature for each confidence fire sign feature in the confidence fire sign feature space using twin comparisons to obtain an Rth twin comparison evaluation sequence; optimizing the Rth twin comparison evaluation sequence for maximum value to generate an Rth environmental monitoring fire indicativeness, and adding the Rth environmental monitoring fire indicativeness to the environmental monitoring fire indicativeness sequence.

[0009] Furthermore, based on the vehicle fire history event set, single-factor fire prediction is performed on the environmental fire sensitive feature vector, the working condition fire sensitive feature vector, and the visual fire sensitive feature vector to obtain a first fire prediction map. This includes: performing single-factor classification based on the vehicle fire history event set to obtain environmental fire history event set, working condition fire history event set, and visual fire history event set; performing environmental factor fire prediction on the environmental fire sensitive feature vector based on the environmental fire history event set to obtain a first fire prediction result; performing working condition factor fire prediction on the working condition fire sensitive feature vector based on the working condition fire history event set to obtain a second fire prediction result; performing visual factor fire prediction on the visual fire sensitive feature vector based on the visual fire history event set to obtain a third fire prediction result; and organizing the environmental fire sensitive feature vector, the working condition fire sensitive feature vector, the visual fire sensitive feature vector, the first fire prediction result, the second fire prediction result, and the third fire prediction result to generate the first fire prediction map.

[0010] Furthermore, based on the historical environmental fire event set, environmental fire sensitivity feature vectors are used to perform environmental factor fire inference to obtain a first fire inference result, including: activating a fire path tracing architecture, which includes a top event node, intermediate event nodes, and bottom event nodes, connected by logical relationship nodes; performing fire path tracing on the historical environmental fire event set based on the fire path tracing architecture to obtain an environmental fire tracing path set; injecting perturbations into the environmental fire tracing path set to obtain an environmental fire perturbation path set; performing adversarial training on the environmental fire tracing path set and the environmental fire perturbation path set to generate an environmental factor fire inference model; and inputting the environmental fire sensitivity feature vectors into the environmental factor fire inference model to obtain the first fire inference result.

[0011] Furthermore, based on the first fire scenario projection map, a multi-factor coupled fire scenario projection is performed to obtain a second fire scenario projection map, including: performing an environmental and working condition coupled fire scenario projection based on the first fire scenario projection map to obtain a fourth fire scenario projection result; performing an environmental and visual coupled fire scenario projection based on the first fire scenario projection map to obtain a fifth fire scenario projection result; performing a working condition and visual coupled fire scenario projection based on the first fire scenario projection map to obtain a sixth fire scenario projection result; performing a three-factor coupled fire scenario projection based on the first fire scenario projection map to obtain a seventh fire scenario projection result; and optimizing the first fire scenario projection map based on the fourth, fifth, sixth, and seventh fire scenario projection results to generate the second fire scenario projection map.

[0012] Furthermore, based on the first fire scenario projection map, environmental and working condition coupled fire scenario projection is performed to obtain a fourth fire scenario projection result, including: extracting environmental and working condition coupled factor events from the vehicle fire history event set to obtain an environmental and working condition coupled fire scenario event set; performing path tracing on the environmental and working condition coupled fire scenario event set according to the fire scenario path tracing architecture to obtain an environmental and working condition coupled fire scenario path set; training an environmental and working condition coupled fire scenario projection model based on the environmental and working condition coupled fire scenario path set; and outputting the fourth fire scenario projection result based on the environmental fire sensitive feature vector and the working condition fire sensitive feature vector, according to the environmental and working condition coupled fire scenario projection model.

[0013] Furthermore, the vehicle visual monitoring data includes infrared thermal image data and visible light image data of the mining vehicle.

[0014] Furthermore, early fire suppression management is implemented for the mining vehicles based on the second fire scenario projection map.

[0015] Another aspect of this application discloses a multi-source information fusion early fire warning system for mining vehicles, used to execute the multi-source information fusion early fire warning method for mining vehicles described in this application. The system includes: a multi-source monitoring acquisition module for performing multi-source monitoring of mining vehicles to acquire vehicle environmental monitoring data, vehicle operating condition data, and vehicle visual monitoring data; an environmental fire sensitive feature vector acquisition module for performing fire sign sensitive feature detection on the vehicle environmental monitoring data based on a set of historical vehicle fire events to obtain an environmental fire sensitive feature vector; and an operating condition fire sensitive feature vector acquisition module for performing fire sign sensitive feature detection on the vehicle operating condition data based on the set of historical vehicle fire events. The system employs a three-tiered approach: a fire-sensitive feature vector acquisition module, a visual fire-sensitive feature vector acquisition module, and an early fire warning execution module. The first fire scenario projection module performs single-factor fire scenario projection on the environmental fire-sensitive feature vector, the working condition fire-sensitive feature vector, and the visual fire-sensitive feature vector based on the vehicle fire history event set.

[0016] One or more technical solutions provided in this application have at least the following technical effects or advantages:

[0017] By integrating multi-source monitoring data from mining vehicles, including environmental, operational, and visual data, and extracting fire-sensitive feature vectors from various dimensions, a first fire map is generated through single-factor fire scenario deduction. A second fire map is then generated through multi-factor coupled deduction. Finally, early warning and suppression management are implemented based on the second fire map. This technical solution solves the problems of existing technologies that rely on single-dimensional data, provide incomplete information, and struggle to capture early multi-factor collaborative risks, leading to missed fire detection, misjudgments, and untimely warnings. The solution achieves the technical effects of comprehensively capturing early fire signs, improving the accuracy and timeliness of warnings, effectively curbing fire spread, and enhancing the operational safety of mining vehicles.

[0018] The above description is merely an overview of the technical solution of this application. In order to better understand the technical means of this application and to implement it in accordance with the contents of the specification, and to make the above and other objects, features and advantages of this application more obvious and understandable, specific embodiments of this application are given below. Attached Figure Description

[0019] Figure 1 A flowchart illustrating the method for early fire warning of mining vehicles based on multi-source information fusion is provided for embodiments of this application.

[0020] Figure 2 This application provides a flowchart illustrating the process of obtaining the first fire scenario map in a multi-source information fusion method for early fire warning of mining vehicles.

[0021] Figure 3 A schematic diagram of the structure of a multi-source information fusion early fire warning system for mining vehicles is provided for the embodiments of this application.

[0022] Figure labeling: Multi-source monitoring acquisition module 11, Environmental fire sensitive feature vector acquisition module 12, Working condition fire sensitive feature vector acquisition module 13, Visual fire sensitive feature vector acquisition module 14, Fire situation simulation first map acquisition module 15, Early fire warning execution module 16. Detailed Implementation

[0023] The overall concept of the technical solution provided in this application is as follows:

[0024] This application provides a method and system for early fire warning of mining vehicles based on multi-source information fusion. It integrates multi-source monitoring data from the mining vehicle's environment, operating conditions, and vision, extracting fire-sensitive feature vectors from various dimensions. First, a first fire map is generated through single-factor fire scenario deduction, and then a second map is generated through multi-factor coupled deduction. Finally, early fire warning and suppression management are implemented based on the second map, achieving comprehensive and precise early fire prevention and control.

[0025] After introducing the basic principles of this application, various non-limiting embodiments of this application will be described in detail below with reference to the accompanying drawings.

[0026] Example 1

[0027] like Figure 1 As shown in the embodiment of this application, an early fire warning method for mining vehicles based on multi-source information fusion is provided. The method includes:

[0028] Step S100: Perform multi-source monitoring on mining vehicles to obtain vehicle environmental monitoring data, vehicle operating condition data, and vehicle visual monitoring data.

[0029] Specifically, vehicle environmental monitoring data refers to the environmental physical parameters of the space where the mining vehicle is located, reflecting the environmental conditions around and inside the vehicle. Vehicle operating condition data refers to the core state parameters of the vehicle itself during operation, reflecting the working status of various systems of the vehicle, including power, transmission, and electrical systems. Vehicle visual monitoring data refers to key parts of the vehicle acquired through imaging equipment.

[0030] By deploying various types of monitoring equipment, mining vehicles are comprehensively monitored. Specifically, temperature sensors, smoke sensors, and combustible gas sensors are installed in key areas such as the engine compartment and driver's cab to collect environmental monitoring data in real time. Operating condition data is read through the vehicle's built-in OBD interface or additional operating condition sensors. High-definition visible light cameras are installed in fire-prone areas such as the vehicle's cable trays and braking system to periodically capture images and obtain visual monitoring data. Simultaneously, the data acquisition module aggregates and stores the data from various sensors in a time-series format, providing raw data support for subsequent feature extraction.

[0031] This step achieves comprehensive coverage of information related to fires in mining vehicles through multi-source monitoring. It includes potential risk signals at the environmental level, abnormal operating conditions of equipment, and intuitive visual signs, avoiding the one-sidedness of a single data source. This lays a multi-dimensional and high-fidelity data foundation for subsequent extraction of fire-sensitive features and accurate early warning.

[0032] Step S200: Based on the vehicle fire history event set, perform fire sign sensitive feature detection on the vehicle environmental monitoring data to obtain an environmental fire sensitive feature vector.

[0033] Specifically, the vehicle fire history event set refers to a complete collection of records of past fire events involving mining vehicles, including time-series data for each event, such as changes in environmental parameters before and during the fire, fire triggers (e.g., short circuits, oil leaks), fire development stages, and final results. Fire sign sensitive feature detection refers to identifying features highly correlated with fire and capable of indicating early fire risk from vehicle environmental monitoring data, such as abnormal temperature rise rates and sudden changes in specific gas concentrations. The core is to distinguish between "normal environmental fluctuations" and "fire precursor signals."

[0034] Specifically, feature mining is performed on environmental data from historical event sets to calculate the importance weights of various environmental parameters, such as temperature, smoke, and methane concentration, before a fire occurs, thus constructing a fire-sensitive feature library. Next, real-time vehicle environmental monitoring data is segmented temporally using the sliding window method, and sensor noise is eliminated using the Kalman filter algorithm. Subsequently, the isolated forest algorithm is used to detect whether there are anomalies in the segmented data, and the abnormal data is compared with the fire-sensitive feature library using cosine similarity. Finally, features with similarity exceeding a threshold are selected, sorted by weight, and quantized into vectors to obtain the environmental fire-sensitive feature vector.

[0035] This step guides the feature selection of real-time data with prior knowledge of historical event sets. It can accurately capture subtle signals in the early stages of a fire, such as the abnormal rate of temperature rise rather than just high temperature, and eliminate environmental interference. The generated sensitive feature vector focuses on core parameters that are highly correlated with the fire, providing a high signal-to-noise ratio input for subsequent fire simulation, and significantly improving the targeting and accuracy of early fire identification.

[0036] Step S300: Based on the vehicle fire history event set, perform fire sign sensitive feature detection on the vehicle operating condition data to obtain the operating condition fire sensitive feature vector.

[0037] Specifically, fire-sensitive feature detection identifies abnormal state features strongly correlated with fire from vehicle operating data, such as electrical system overload, sudden drop in oil pressure, and overheating of mechanical components. The core principle is to distinguish between "normal operating condition fluctuations" and "fire precursor anomalies" by analyzing historical fire patterns. The operating condition fire-sensitive feature vector refers to a multi-dimensional array formed by quantifying the detected operating condition sensitive features. It centrally reflects key information indicating fire risk in vehicle operation, such as: duration of motor current exceeding limits, magnitude of battery voltage drop, transmission oil temperature rise rate, and percentage deviation of engine oil pressure from the threshold.

[0038] First, the operating condition data from historical event sets is mined to extract abnormal state features that lead to fires due to operating conditions. The fire contribution weight of each operating condition parameter is calculated using the random forest algorithm to construct a fire-sensitive feature library for operating conditions. Next, the real-time collected operating condition data is processed using a wavelet transform algorithm to filter high-frequency noise and time-series segmentation using a sliding window. The DBSCAN clustering algorithm is then used to identify abnormal fluctuations within the segments. Finally, the abnormal fluctuation data is matched with the fire-sensitive feature library for operating conditions using Jaccard similarity. Features with similarity exceeding a threshold are selected, sorted by weight, and quantized into vectors to obtain the fire-sensitive feature vector for operating conditions.

[0039] This step extracts risk correlation patterns from operating data through historical fire events, enabling precise capture of abnormal equipment signals before a fire. Simultaneously, noise filtering and cluster analysis eliminate fluctuations in normal operating conditions, generating sensitive feature vectors that focus on core risk points in equipment operation status. This provides highly relevant and low-redundancy input for subsequent fire simulations, effectively improving the early identification capability of fires caused by equipment failures in mining vehicles and reducing early warning deviations caused by misjudgments of single parameters.

[0040] Step S400: Based on the vehicle fire history event set, perform fire sign sensitive feature detection on the vehicle visual monitoring data to obtain a visual fire sensitive feature vector.

[0041] Specifically, vehicle visual monitoring data refers to image information collected by imaging equipment installed in key parts of mining vehicles, such as engine compartments, cable trays, and braking systems, including visible light images and infrared thermal images. Fire sign sensitive feature detection, targeting visual data, identifies visual features directly related to the fire situation from the visual monitoring data, with the core being the distinction between "normal visual features" and "precursor visual signals of a fire." The visual fire sensitive feature vector refers to a multi-dimensional array formed by quantifying the detected visual sensitive features, centrally representing the key information in the visual data indicating fire risk.

[0042] Specifically, the OpenCV library in Python was used to extract features from visual data in historical event sets. A CNN model was used to identify fire signs and construct a visual fire-sensitive feature library, with the association weights of each feature with the fire situation labeled. The CNN model was constructed by collecting historical infrared thermal images and visible light images of mining vehicle fires, labeling three types of features—"circular high-temperature area," "grayish-white smoke texture," and "black burn spots"—and their corresponding parameters, and dividing the training and validation sets in a 7:3 ratio. A pre-trained ResNet-50 was used, with the first 15 layers frozen and the top layer replaced with the output nodes of the three types of features, and a new parameter regression branch was added. The Adam optimizer was used to train the model for multiple rounds with cross-entropy and MSE loss as targets until the classification accuracy on the validation set met the target. The feature vectors of the model's convolutional layers were extracted and combined with the labeled parameters to construct the visual fire-sensitive feature library. Next, the real-time acquired visual monitoring data is preprocessed. Noise is removed by Gaussian filtering, and areas with excessively high temperatures are extracted from the infrared image using a threshold segmentation method. Edge detection algorithms are used to identify changes in the surface morphology of objects from the visible light image. Subsequently, the Faster R-CNN model is used to extract features from the preprocessed image to obtain real-time visual features. These features are then compared with features in the visual fire sensitivity feature library using a Siamese network. Finally, features with similarity exceeding a threshold are selected, quantized into vectors according to their association weights, and the visual fire sensitivity feature vector is obtained.

[0043] This step, through multimodal analysis of visual data, enables the accurate capture of early visual signals of a fire, providing an intuitive and highly recognizable visual dimension input for subsequent multi-factor fire simulation, and significantly reducing the risk of missing early fire detection due to the limitations of a single data dimension.

[0044] Step S500: Based on the vehicle fire history event set, perform single-factor fire prediction on the environmental fire sensitive feature vector, the working condition fire sensitive feature vector, and the visual fire sensitive feature vector to obtain the first fire prediction map.

[0045] Specifically, single-factor fire prediction refers to conducting independent fire risk analysis on environmental fire-sensitive feature vectors, operational fire-sensitive feature vectors, and visual fire-sensitive feature vectors based on historical vehicle fire events. The first fire prediction map refers to a visualized and structured dataset that integrates the results of the three single-factor predictions, including the original values ​​of each feature vector, the corresponding single-factor fire probability, and the prediction confidence level.

[0046] Specifically, the historical event set of vehicle fires is classified into single-factor categories, and decomposed into environmental fire history subsets, working condition fire history subsets, and visual fire history subsets according to the dominant factors that cause the fires. Corresponding models are constructed, and the corresponding real-time data is input to obtain the fire probability and confidence level for each of the three dimensions. Finally, the Matplotlib library is used to integrate the original feature values ​​of the three vectors, the corresponding single-factor fire probability, and the inferred confidence level into a structured map to generate the first fire inference map.

[0047] This step, through independent single-factor extrapolation, achieves a precise breakdown of fire risk across three dimensions, eliminating cross-interference from features of different dimensions and clearly defining the independent risk intensity of each dimension. Training the model based on the categorized historical subsets makes the extrapolation results more closely match the actual fire patterns in each dimension. The generated first fire extrapolation map not only provides a single-dimensional risk benchmark for subsequent multi-factor coupling but also quickly identifies high-risk dimensions, improving the risk tracing capability of early fire warnings while reducing extrapolation errors caused by the mixing of multiple factors.

[0048] Step S600: Perform multi-factor coupled fire simulation based on the first fire simulation map to obtain the second fire simulation map, and execute early fire warning based on the second fire simulation map.

[0049] Specifically, multi-factor coupled fire simulation refers to the process of integrating the results of single-factor simulations from three dimensions—environment, operating conditions, and visual perception—to analyze the synergistic effects between these factors, rather than simply summing them up, and ultimately calculating the overall fire risk. The second fire simulation map refers to a structured map that integrates the results of multi-factor coupling, including the overall fire probability, the coupling weights of each factor, and the risk level.

[0050] Specifically, key data are extracted from the first fire scenario simulation map. An attention mechanism fusion model is constructed using Python's PyTorch library. The probability values ​​and feature vectors of three dimensions are input, and the coupling weights between factors are calculated through a self-attention layer. At the same time, a Bayesian network is introduced to model the coupling patterns in historical data. Subsequently, the single-factor probabilities and coupling weights are weighted and fused to obtain the overall fire probability. The second fire scenario simulation map is then generated by combining the calculated probability with a preset threshold.

[0051] This step overcomes the limitations of single-factor inference through multi-factor coupling analysis, captures the synergistic risks of "1+1>2", and improves the overall accuracy of fire identification; the risk level and coupling weight of the second map provide a precise basis for early warning measures and significantly improve the safety of mine transportation.

[0052] Furthermore, based on the vehicle fire history event set, fire sign sensitive feature detection is performed on the vehicle environmental monitoring data to obtain an environmental fire sensitive feature vector. This includes: collecting sensor feature data corresponding to each environmental monitoring parameter within the vehicle environmental monitoring data to obtain multiple monitoring sensor feature sequences; performing anomaly detection based on the multiple monitoring sensor feature sequences to obtain multiple sensor anomaly detection results; performing adaptive interference correction on the vehicle environmental monitoring data based on the multiple sensor anomaly detection results to obtain an environmental monitoring feature sequence; evaluating the fire indicativeness of the environmental monitoring feature sequence based on the vehicle fire history event set to obtain an environmental monitoring fire indicativeness sequence; and performing fire sign sensitive attention allocation on the environmental monitoring feature sequence based on the environmental monitoring fire indicativeness sequence to generate the environmental fire sensitive feature vector.

[0053] Specifically, fire indicative assessment refers to calculating the correlation between each data point in the environmental monitoring feature sequence and a fire, based on a historical set of vehicle fire events. The environmental monitoring fire indicative sequence is the set of fire indicative values ​​corresponding to each environmental monitoring feature sequence arranged chronologically. Fire sign sensitivity attention allocation refers to assigning weights to environmental monitoring feature sequences based on their fire indicative value, with high-indicative-value features receiving higher attention weights and low-indicative-value features receiving lower weights.

[0054] Specifically, mining-specific sensors, such as temperature sensors, smoke sensors, and catalytic combustion methane sensors, are used to collect environmental parameters from key areas of the vehicle, including the engine compartment, driver's cab, and cargo compartment. The raw data for each individual parameter is then organized into monitoring sensor feature sequences by timestamp using Python's Pandas library. An isolation forest algorithm is then employed to detect anomalies in each monitoring sensor feature sequence; for example, it identifies a sudden increase in smoke concentration from 0.02 mg / m³ to 0.1 mg / m³. The system identifies outliers and outputs sensor anomaly detection results including anomaly timestamps and deviation values. Based on these results, an adaptive interference correction is performed using a Kalman filter algorithm. If the anomaly is caused by temporary dust obstruction of the sensor, the value is corrected from 0.1 mg / m³ to 0.03 mg / m³, resulting in a noise-removed environmental monitoring feature sequence. Next, the XGBoost algorithm is used, with a set of historical vehicle fire events as training samples, to calculate the fire indicative intensity of each data point in the environmental monitoring feature sequence, forming an environmental monitoring fire indicative intensity sequence. Finally, a Softmax attention mechanism is used to assign weights to the environmental monitoring feature sequence, and the weighted values ​​of each feature are extracted and integrated into an environmental fire-sensitive feature vector.

[0055] This process involves three steps: First, by using isolated forest anomaly detection and Kalman filtering correction, it effectively eliminates sensor interference caused by dust and vibration in the mining environment, ensuring the authenticity of environmental data. Second, based on the fire indicativeness evaluation of historical fire event sets, it accurately identifies features strongly correlated with fires, avoiding interference from unrelated parameters. Third, by focusing on high-risk features through sensitive attention allocation, the generated environmental fire sensitive feature vector can condense core risk information, providing a high signal-to-noise ratio input for subsequent single-factor fire simulations, significantly improving the accuracy of early fire signs identification, and reducing false alarms and missed alarms caused by data noise or irrelevant parameters.

[0056] Further, the environmental monitoring feature sequence is evaluated for fire indicativeness based on the historical vehicle fire event set to obtain an environmental monitoring fire indicativeness sequence. This includes: identifying fire sign features based on the historical vehicle fire event set to obtain a fire sign feature library; evaluating the confidence level of each fire sign feature in the fire sign feature library based on the historical vehicle fire event set to obtain a fire feature confidence sequence; cleaning the fire sign feature library based on the fire feature confidence sequence and a confidence fire sign feature space; extracting the Rth environmental monitoring feature from the environmental monitoring feature sequence, where R is a positive integer; evaluating the Rth environmental monitoring feature for each confidence fire sign feature in the confidence fire sign feature space using twin comparisons to obtain an Rth twin comparison evaluation sequence; optimizing the Rth twin comparison evaluation sequence for maximum value to generate an Rth environmental monitoring fire indicativeness, and adding the Rth environmental monitoring fire indicativeness to the environmental monitoring fire indicativeness sequence.

[0057] Specifically, the fire sign feature library is a collection storing identified fire sign features, each feature containing specific parameter thresholds and time-series patterns. The fire feature confidence sequence refers to the sequence formed by arranging the confidence levels of each fire sign feature according to the feature library order. The predetermined fire feature confidence level is a manually set screening threshold used to retain high-confidence features and remove low-correlation features. The confidence fire sign feature space is a feature set composed of high-confidence fire sign features after being cleaned with the predetermined confidence level. The R-th feature of environmental monitoring refers to the R-th specific feature extracted from the environmental monitoring feature sequence, where R is a positive integer, including time-series patterns of single parameters or features of parameter combinations.

[0058] Specifically, the Scikit-learn tool in Python is used to perform feature mining on the historical event set of vehicle fires. Fire sign features are identified using association rule algorithms, such as Apriori, and these features are compiled into a fire sign feature library. Logistic regression is then used to calculate the confidence level of each feature in the library: the ratio of the number of actual fires occurring after the feature appears to the total number of times the feature appears. Subsequently, a predetermined fire feature confidence level is set, the feature library is cleaned, and a confidence fire sign feature space containing multiple features such as temperature and smoke is established. Next, the Rth feature is extracted from the environmental monitoring feature sequence, and a Siamese neural network built using PyTorch is used to compare this feature with features in the confidence space to calculate the Rth Siamese comparison evaluation sequence. Finally, the argmax function is used to optimize the maximum value of this sequence to obtain the Rth environmental monitoring fire indicator, which is then added to the environmental monitoring fire indicator sequence.

[0059] This step improves the reliability of fire indication through three layers of optimization: First, feature identification and confidence evaluation based on historical data ensure that the analyzed objects are features strongly correlated with the fire situation; second, the construction of a confidence feature space focuses on highly reliable features to improve matching efficiency; and third, twin comparison uses deep learning to capture subtle temporal correlations between features. The resulting indication sequence can accurately quantify the fire risk of real-time environmental features, improve the accuracy of early fire signs identification, provide high-value input for subsequent sensitive feature extraction, and effectively reduce early warning deviations caused by feature misjudgment.

[0060] Furthermore, such as Figure 2 As shown, based on the vehicle fire history event set, single-factor fire prediction is performed on the environmental fire sensitive feature vector, the working condition fire sensitive feature vector, and the visual fire sensitive feature vector to obtain a first fire prediction map. This includes: performing single-factor classification based on the vehicle fire history event set to obtain environmental fire history event sets, working condition fire history event sets, and visual fire history event sets; performing environmental factor fire prediction on the environmental fire sensitive feature vector based on the environmental fire history event set to obtain a first fire prediction result; performing working condition factor fire prediction on the working condition fire sensitive feature vector based on the working condition fire history event set to obtain a second fire prediction result; performing visual factor fire prediction on the visual fire sensitive feature vector based on the visual fire history event set to obtain a third fire prediction result; and organizing the environmental fire sensitive feature vector, the working condition fire sensitive feature vector, the visual fire sensitive feature vector, the first fire prediction result, the second fire prediction result, and the third fire prediction result to generate the first fire prediction map.

[0061] Specifically, the Pandas library in Python is used to perform single-factor classification on the historical event set of vehicle fires. By selecting the parameter dimension with the only anomaly before the fire, the historical event set of environmental fires, the historical event set of working condition fires, and the historical event set of visual fires are split into three categories. For the environmental fire-sensitive feature vector, an environmental factor fire prediction model is constructed using a subset of historical environmental fires as training samples. Real-time environmental vectors are input into the model to obtain the fire probability and confidence score for the environmental dimension. For the working condition fire-sensitive feature vector, a subset of historical working condition fires is used for training. The XGBoost model obtains a fire prediction model based on working conditions. After inputting real-time working condition vectors, it obtains the fire probability and confidence score for the working condition dimension. For visual fire-sensitive feature vectors, a CNN-LSTM hybrid model is trained using a subset of historical visual fire data to obtain a fire prediction model based on visual factors. After inputting real-time visual vectors, it obtains the fire probability and confidence score for the visual dimension. Finally, the Matplotlib library is used to integrate the original feature values ​​of the three vectors, the corresponding single-factor fire probabilities, and the prediction confidence scores into a structured map, generating the first fire prediction map.

[0062] This step eliminates cross-interference from features across different dimensions through independent single-factor extrapolation, making the fire risk more accurate across dimensions. Simultaneously, the categorized historical subsets allow the extrapolation model to better align with single-dimensional fire patterns, enhancing the extrapolation confidence. The generated first map clearly presents the risk differences across dimensions, providing a clear risk benchmark for subsequent multi-factor coupled extrapolation and enabling rapid identification of high-risk dimensions, effectively reducing extrapolation errors caused by the mixing of multiple factors.

[0063] Furthermore, based on the historical environmental fire event set, environmental fire sensitivity feature vectors are used to perform environmental factor fire inference to obtain a first fire inference result, including: activating a fire path tracing architecture, which includes a top event node, intermediate event nodes, and bottom event nodes, connected by logical relationship nodes; performing fire path tracing on the historical environmental fire event set based on the fire path tracing architecture to obtain an environmental fire tracing path set; injecting perturbations into the environmental fire tracing path set to obtain an environmental fire perturbation path set; performing adversarial training on the environmental fire tracing path set and the environmental fire perturbation path set to generate an environmental factor fire inference model; and inputting the environmental fire sensitivity feature vectors into the environmental factor fire inference model to obtain the first fire inference result.

[0064] Specifically, the fire path tracing architecture is a hierarchical event tree model that simulates the development logic of a fire, containing top event nodes, intermediate event nodes, and bottom event nodes, connected by logical relationships such as "AND" and "OR". The environmental fire tracing path set is a collection of specific fire development paths extracted from historical environmental fire events based on the fire path tracing architecture. Disturbance injection refers to artificially adding interfering factors, such as sensor measurement errors and sudden environmental fluctuations, to the environmental fire tracing path to simulate uncertainties in real-world scenarios. The environmental fire disturbance path set refers to the collection of fire paths containing interfering factors generated after disturbance injection. Adversarial training refers to using the environmental fire tracing path set and the environmental fire disturbance path set as positive and negative samples to train the model to identify the differences between real fire paths and interfering paths, improving the model's anti-interference ability. The environmental factor fire prediction model refers to a model that, after adversarial training, can predict the probability and confidence level of fire occurrence based on environmental fire-sensitive feature vectors.

[0065] Specifically, a fire path tracing architecture was constructed using Python's NetworkX library, defining top event nodes, intermediate event nodes, and bottom event nodes, connected by logical relationship nodes. Based on this architecture, the Pandas library was used to trace the paths of historical environmental fire events, extracting the bottom event → intermediate event → top event chain for each record to form an environmental fire tracing path set, for example: methane leak → concentration 1.8% LEL → temperature 15℃ / min → fire. Subsequently, perturbations were injected to generate an environmental fire perturbation path set. A bidirectional LSTM model was then built, using the traced path set as positive samples and the perturbation path set as negative samples for adversarial training. The training was iteratively continued until the model achieved satisfactory accuracy in recognizing real paths and satisfactory anti-interference rate against perturbation paths, generating an environmental factor fire inference model. Finally, the environmental fire sensitive feature vector was input into the model to obtain the first fire inference result. The bidirectional LSTM model is constructed as follows: the input layer receives environmental fire-sensitive feature vectors, with an input size of 3; the intermediate layer is a bidirectional LSTM, with forward and backward LSTMs capturing the forward and backward temporal dependencies, respectively; a dropout layer is added to prevent overfitting; a fully connected layer is then added; and the output layer uses sigmoid activation to output the fire probability. The model uses cross-entropy as the loss function and is trained with the Adam optimizer to adapt to the temporal feature inference of environmental fires.

[0066] Similarly, the working condition fire path tracing architecture is activated; based on the architecture, the historical event set of working condition fires is traced to obtain the working condition fire tracing path set; parameter fluctuation perturbations are injected to generate a perturbation path set; the working condition factor inference model is trained adversarially using both; and the working condition fire sensitive feature vector is input to obtain the second fire inference result. Furthermore, the visual fire path tracing architecture is activated; based on the architecture, the historical event set of visual fires is traced to obtain the visual fire tracing path set; image noise perturbations are injected to generate a perturbation path set; the visual factor inference model is trained adversarially; and the visual fire sensitive feature vector is input to obtain the third fire inference result.

[0067] This step reconstructs the logical chain of fires caused by environmental, working conditions, and visual factors through a fire path tracing architecture, making the simulation interpretable. Perturbation injection and adversarial training enable the model to resist common sensor errors and parameter fluctuations in mining environments, improving its anti-interference ability. The final fire simulation results not only output the fire probability but also trace the source of risk, providing accurate and interpretable environmental dimension basis for subsequent early warnings and reducing the false alarm rate of fires caused by single environmental factors.

[0068] Furthermore, based on the first fire scenario projection map, a multi-factor coupled fire scenario projection is performed to obtain a second fire scenario projection map, including: performing an environmental and working condition coupled fire scenario projection based on the first fire scenario projection map to obtain a fourth fire scenario projection result; performing an environmental and visual coupled fire scenario projection based on the first fire scenario projection map to obtain a fifth fire scenario projection result; performing a working condition and visual coupled fire scenario projection based on the first fire scenario projection map to obtain a sixth fire scenario projection result; performing a three-factor coupled fire scenario projection based on the first fire scenario projection map to obtain a seventh fire scenario projection result; and optimizing the first fire scenario projection map based on the fourth, fifth, sixth, and seventh fire scenario projection results to generate the second fire scenario projection map.

[0069] Specifically, environmental probability, operating condition probability, and visual probability are extracted from the first fire scenario projection map. A coupling model based on an attention mechanism is built to perform pairwise coupling projections. When environment and operating condition are coupled, the model calculates the synergistic correlation between environmental factors and operating condition factors such as "decrease in motor insulation performance under high temperature conditions" through a self-attention layer. For example, the model calculates the degree of synergy between environmental factors and operating condition factors such as "decrease in motor insulation performance under high temperature conditions." The fourth fire scenario projection result is output based on the data from the first map. When environment and vision are coupled, the correlation between environmental factors and visual factors is calculated, such as the overlap between the smoke diffusion range and the high-temperature zone, resulting in the fifth fire scenario projection result. When operating condition and vision are coupled, the correlation between operating condition factors and visual factors is calculated, such as the duration of abnormal current and the rate of burn mark expansion, resulting in the sixth fire scenario projection result. Subsequently, a Bayesian network is used to model the interaction relationship of the three factors, outputting the seventh fire scenario projection result. Finally, a weighted fusion algorithm is used to optimize the single-factor risk assessment of the first map, integrating single-factor risks with four types of coupled risks to generate a second fire scenario projection map that includes "single-factor, pairwise coupling, and three-factor risks and the degree of coupling influence."

[0070] This step captures collaborative risks that are overlooked in single-factor extrapolation through multi-dimensional coupling analysis, making fire risk assessment more comprehensive; the four types of coupling results and weights provide fine-grained basis for risk tracing; the optimized second map will improve the overall accuracy of fire identification, and at the same time reduce the uncertainty of early warning through coupling logic interpretation, providing more reliable decision support for subsequent graded early warning.

[0071] Furthermore, based on the first fire scenario projection map, environmental and working condition coupled fire scenario projection is performed to obtain a fourth fire scenario projection result, including: extracting environmental and working condition coupled factor events from the vehicle fire history event set to obtain an environmental and working condition coupled fire scenario event set; performing path tracing on the environmental and working condition coupled fire scenario event set according to the fire scenario path tracing architecture to obtain an environmental and working condition coupled fire scenario path set; training an environmental and working condition coupled fire scenario projection model based on the environmental and working condition coupled fire scenario path set; and outputting the fourth fire scenario projection result based on the environmental fire sensitive feature vector and the working condition fire sensitive feature vector, according to the environmental and working condition coupled fire scenario projection model.

[0072] Specifically, the environmental and operating condition coupling factor event extraction refers to screening and extracting fire events involving both abnormal environmental parameters and abnormal operating condition parameters from a set of historical vehicle fire events, focusing on cases where the combined effect of these two factors causes a fire. Similarly, it also includes fire events involving both abnormal environmental and visual parameters, and fire events involving both abnormal operating condition and visual parameters. The environmental and operating condition coupled fire prediction model refers to a model that uses a set of coupled paths as training data to learn the synergistic patterns of environmental and operating condition characteristics, outputting the fire risk after their coupling, and providing the probability and synergistic weights of coupled fires. Similarly, it also includes environmental-visual coupled fire prediction models and operating condition-visual coupled fire prediction models.

[0073] From the historical vehicle fire event set, events containing both environmental and operational anomalies were selected using Python's Pandas library. Environmental features, operational features, and fire results were extracted to form an environmental-operational condition coupled fire event set. Based on the previously constructed fire path tracing architecture, the NetworkX tool was used to trace the paths of the coupled event set, identifying the collaborative development chain of environmental and operational factors in each event, generating an environmental-operational condition coupled fire path set. After standardizing the feature data in the path set, a bidirectional LSTM model with an attention mechanism was built using PyTorch. The input layer receives environmental and operational feature vectors, the attention layer calculates the collaborative weights of the two, the LSTM layer captures temporal correlations, and the output layer outputs the coupled fire probability. Using the coupled path set as training samples, the model was trained until the accuracy on the validation set was satisfactory, resulting in an environmental-operational condition coupled fire inference model. Finally, the real-time environmental fire-sensitive feature vectors were input into the model to output the fourth fire inference result.

[0074] From the historical fire event set, Pandas was used to filter events where both environmental and visual parameters were abnormal, extracting environmental features, visual features, and fire results to form an environmental-visual coupled event set. Based on the fire path tracing architecture, NetworkX was used to generate a coupled path set. A CNN-LSTM fusion model was built using PyTorch, where CNN extracts visual image features, LSTM processes environmental temporal data, and the attention layer calculates the spatial correlation weights between the two. Using cross-entropy as the loss, the Adam optimizer was trained until the validation set accuracy was satisfactory, completing the model construction. Furthermore, from the historical event set, Pandas was used to extract fire events where the working conditions and visual parameters were abnormally correlated, labeling working condition features, visual features, and results to form a working condition-visual coupled event set. Based on the path tracing architecture, NetworkX was used to generate a coupled path set. A bidirectional LSTM attention model was built using PyTorch, inputting working condition temporal data and visual features, with the attention layer quantifying the temporal correlation between the two. Using MSE and cross-entropy as losses, the Adam optimizer was trained until the validation set accuracy was satisfactory, completing the model construction.

[0075] This step precisely captures the collaborative risks of the environment and working conditions, environment and vision, and working conditions and vision by specifically extracting coupled events. The introduction of the attention mechanism enables the model to quantify the weights of the collaborative impact of the two, improving the interpretability of the results. The coupled model after training improves the accuracy of identifying collaborative risks of the environment and working conditions, environment and vision, and working conditions and vision, effectively making up for the underestimation of collaborative risks by single-factor analysis, and significantly improving the comprehensiveness and accuracy of early fire warning.

[0076] Furthermore, the vehicle visual monitoring data includes infrared thermal image data and visible light image data of the mining vehicle.

[0077] Specifically, vehicle visual monitoring data includes infrared thermal imaging data and visible light image data, used to capture signs of fire from different dimensions. Infrared thermal imaging data, acquired by an infrared thermal imager, represents the surface temperature distribution of an object in grayscale or pseudo-color, visually showing temperature differences within the equipment; for example, high-temperature areas are displayed in red, and low-temperature areas in blue. Visible light image data, acquired by a visible light camera, reflects the visual characteristics of an object, such as color, shape, and texture, and is consistent with human vision.

[0078] Specifically, infrared thermal imagers and visible light cameras are deployed in key areas of mining vehicles, such as the engine compartment, cable trays, and braking system, to collect infrared thermal image data and visible light image data, respectively. The raw data is preprocessed using Python's OpenCV library: non-uniformity correction is applied to the infrared thermal image data to eliminate sensor noise and temperature calibration, converting grayscale values ​​to actual temperatures; and dehazing and distortion correction are performed on the visible light image data. Next, a dual-branch CNN model is built using PyTorch. The infrared branch extracts features such as temperature gradients and area of ​​high-temperature areas using ResNet, while the visible light branch identifies morphological features such as smoke textures and burn marks using YOLOv5. Finally, the features from the two branches are fused for visual fire-sensitive feature detection.

[0079] Infrared thermal imaging data can identify abnormally high temperatures in equipment at the early stages of a fire, compensating for the blind spots of environmental sensors in sensing local temperatures. Visible light image data can capture morphological features such as smoke and burn marks, complementing temperature features. The combination of the two enables visual monitoring to cover the entire process of "temperature anomalies and morphological changes." Compared with visual data alone, this improves the coverage of fire sign identification and effectively reduces missed detections caused by the limitations of a single data dimension, such as the potential for missing high-temperature hazards without smoke when relying solely on visible light.

[0080] Furthermore, early fire suppression management is implemented for the mining vehicles based on the second fire scenario projection map.

[0081] Specifically, early fire suppression management refers to the proactive intervention process that takes targeted measures at the nascent stage of a fire, based on the risk level and coupling characteristics of the second fire scenario projection map, to prevent the fire from developing and expanding. The core principle is "early detection and early response." The suppression strategy library refers to a set of graded response plans pre-set based on historical fire handling experience and the characteristics of the second scenario projection map. For example, high risk corresponds to "automatic activation of fire extinguishing devices + power cut-off," while medium risk corresponds to "localized cooling + alarm prompts."

[0082] Specifically, the system reads core information from the second fire scenario map in real time via the CAN bus, including the overall risk level, dominant coupling factors, and key risk points. A Python-based decision tree algorithm analyzes the map features and matches corresponding solutions from the suppression strategy library. If the risk is high and environmental conditions are the dominant factors, the system invokes the strategy of "activating the automatic fire extinguishing device in the engine compartment + forcibly reducing the motor load + cutting off unnecessary circuits." If the risk is medium and the visually high-temperature zone is dominant, the system executes "directional spraying of the cooling system + audible and visual alarm in the driver's cabin." Simultaneously, the system uses a PLC to link vehicle actuators, such as the fire extinguishing device solenoid valve, motor controller, and cooling water pump, and utilizes an edge computing module to collect real-time environmental, operational, and visual data after suppression, feeding this data back to the second map for dynamic adjustments. For example, when the second map shows a high risk due to three coupled factors, focusing on the braking system, the system automatically activates brake disc spraying for cooling, cuts off the brake motor power, and confirms the cooling effect via a camera.

[0083] This step directly links risk assessment with proactive suppression, avoiding the lag of merely issuing warnings without taking action. The hierarchical strategy and coupling characteristics based on the second fire map make suppression measures more precise. A real-time feedback and adjustment mechanism ensures the verifiability of suppression effects and avoids over-intervention. In practical applications, this can shorten early fire response time, reduce fire spread rates, and significantly reduce equipment damage and operational interruptions to mining vehicles caused by fires.

[0084] In summary, the multi-source information fusion method for early fire warning of mining vehicles provided in this application has the following technical effects:

[0085] 1. By fusing multi-source information and integrating data from multiple dimensions such as environment, operating conditions, and vision, the limitations of single monitoring dimensions are broken. From feature extraction to multi-factor coupling and inference, a complete early warning link is constructed to comprehensively capture subtle early signs of fire, avoiding missed or misjudgments due to incomplete information. This provides comprehensive and three-dimensional data support for early fire identification of mining vehicles, improving the integrity and reliability of the early warning system.

[0086] 2. Through single-factor classification and independent extrapolation, historical events are broken down into environmental, operational, and visual subsets, eliminating cross-interference from features of different dimensions. Each dimension model focuses on learning its own correlation with the fire situation, making single-factor risk assessment more accurate, clearly presenting the independent risk intensity of each dimension, providing a clear benchmark for subsequent multi-factor coupled analysis, and enhancing the ability to trace risk sources and the interpretability of extrapolation results.

[0087] 3. By using multi-factor coupling simulation, the system captures the synergistic risks among environment, operating conditions, and vision, compensating for the neglect of the "1+1>2" superposition effect in single-factor analysis. By integrating the results of pairwise and three-factor coupling, the system optimizes the generated second map, comprehensively reflecting the overall risk situation and coupling weights, providing a more realistic basis for early warning decisions, and improving the comprehensiveness and accuracy of early fire assessment.

[0088] Example 2

[0089] Based on the same inventive concept as the early fire warning method for mining vehicles using multi-source information fusion in the aforementioned embodiments, such as Figure 3 As shown in the embodiment of this application, an early fire warning system for mining vehicles based on multi-source information fusion is provided. The system includes:

[0090] Multi-source monitoring acquisition module 11 performs multi-source monitoring on mining vehicles to obtain vehicle environmental monitoring data, vehicle operating condition data, and vehicle visual monitoring data; environmental fire sensitive feature vector acquisition module 12 performs fire sign sensitive feature detection on the vehicle environmental monitoring data based on the vehicle fire historical event set to obtain environmental fire sensitive feature vector; operating condition fire sensitive feature vector acquisition module 13 performs fire sign sensitive feature detection on the vehicle operating condition data based on the vehicle fire historical event set to obtain operating condition fire sensitive feature vector; visual fire sensitive feature vector acquisition module 14, based on... The system uses a set of historical vehicle fire events to detect fire-sign sensitive features in the vehicle visual monitoring data, obtaining a visual fire sensitive feature vector. A fire scenario first map acquisition module 15 performs single-factor fire scenario deduction based on the historical vehicle fire events, the environmental fire sensitive feature vector, the operational condition fire sensitive feature vector, and the visual fire sensitive feature vector, obtaining a fire scenario first map. An early fire warning execution module 16 performs multi-factor coupled fire scenario deduction based on the first fire scenario map, obtaining a fire scenario second map, and executes an early fire warning based on the second fire scenario map.

[0091] Furthermore, the environmental fire sensitive feature vector acquisition module 12 is also used to perform the following steps: collecting sensor feature data corresponding to each environmental monitoring parameter in the vehicle environmental monitoring data to obtain multiple monitoring sensor feature sequences; performing anomaly detection based on the multiple monitoring sensor feature sequences to obtain multiple sensor anomaly detection results; performing adaptive interference correction on the vehicle environmental monitoring data based on the multiple sensor anomaly detection results to obtain environmental monitoring feature sequences; evaluating the fire indicativeness of the environmental monitoring feature sequences based on the vehicle fire historical event set to obtain environmental monitoring fire indicativeness sequences; and performing fire sign sensitive attention allocation on the environmental monitoring feature sequences based on the environmental monitoring fire indicativeness sequences to generate the environmental fire sensitive feature vector.

[0092] Furthermore, the environmental fire sensitive feature vector acquisition module 12 is also used to perform the following steps: identifying fire sign features based on the vehicle fire history event set to obtain a fire sign feature library; evaluating the confidence level of each fire sign feature in the fire sign feature library based on the vehicle fire history event set to obtain a fire feature confidence sequence; cleaning the fire sign feature library based on the fire feature confidence sequence and a predetermined fire feature confidence level to establish a confidence fire sign feature space; extracting the Rth environmental monitoring feature from the environmental monitoring feature sequence, where R is a positive integer; evaluating the Rth environmental monitoring feature based on each confidence fire sign feature in the confidence fire sign feature space using twin comparisons to obtain an Rth twin comparison evaluation sequence; optimizing the Rth twin comparison evaluation sequence for maximum value to generate an Rth environmental monitoring fire indicator, and adding the Rth environmental monitoring fire indicator to the environmental monitoring fire indicator sequence.

[0093] Furthermore, the fire scenario simulation first atlas acquisition module 15 is also used to perform the following steps: perform single-factor classification based on the vehicle fire history event set to obtain an environmental fire history event set, an operational fire history event set, and a visual fire history event set; perform environmental factor fire scenario simulation on the environmental fire sensitive feature vector based on the environmental fire history event set to obtain a first fire scenario simulation result; perform operational factor fire scenario simulation on the operational fire sensitive feature vector based on the operational fire history event set to obtain a second fire scenario simulation result; perform visual factor fire scenario simulation on the visual fire sensitive feature vector based on the visual fire history event set to obtain a third fire scenario simulation result; and organize the environmental fire sensitive feature vector, the operational fire sensitive feature vector, the visual fire sensitive feature vector, the first fire scenario simulation result, the second fire scenario simulation result, and the third fire scenario simulation result to generate the fire scenario simulation first atlas.

[0094] Furthermore, the fire scenario simulation first map acquisition module 15 is also used to perform the following steps: activating the fire path tracing architecture, which includes a top event node, intermediate event nodes, and bottom event nodes, and the top event node, intermediate event node, and bottom event node are connected through logical relationship nodes; performing fire path tracing on the environmental fire historical event set according to the fire path tracing architecture to obtain an environmental fire tracing path set; injecting perturbations into the environmental fire tracing path set to obtain an environmental fire perturbation path set; performing adversarial training on the environmental fire tracing path set and the environmental fire perturbation path set to generate an environmental factor fire scenario simulation model; and inputting the environmental fire sensitive feature vector into the environmental factor fire scenario simulation model to obtain the first fire scenario simulation result.

[0095] Furthermore, the early fire warning execution module 16 is also used to perform the following steps: performing environmental and working condition coupled fire simulation based on the first fire simulation map to obtain a fourth fire simulation result; performing environmental and visual coupled fire simulation based on the first fire simulation map to obtain a fifth fire simulation result; performing working condition and visual coupled fire simulation based on the first fire simulation map to obtain a sixth fire simulation result; performing three-factor coupled fire simulation based on the first fire simulation map to obtain a seventh fire simulation result; optimizing the first fire simulation map based on the fourth, fifth, sixth, and seventh fire simulation results to generate the second fire simulation map.

[0096] Furthermore, the early fire warning execution module 16 is also used to perform the following steps: extracting environmental and working condition coupling factor events based on the vehicle fire history event set to obtain an environmental and working condition coupled fire event set; performing path tracing on the environmental and working condition coupled fire event set according to the fire path tracing architecture to obtain an environmental and working condition coupled fire path set; training an environmental and working condition coupled fire inference model based on the environmental and working condition coupled fire path set; and outputting the fourth fire inference result based on the environmental fire sensitive feature vector and the working condition fire sensitive feature vector, according to the environmental and working condition coupled fire inference model.

[0097] Furthermore, the system is also used to perform the following steps: the vehicle visual monitoring data includes infrared thermal image data and visible light image data of the mining vehicle.

[0098] Furthermore, the system is also used to perform the following steps: to perform early fire suppression management on the mining vehicles based on the second fire scenario projection map.

[0099] In summary, any step of the method described above can be stored as a computer instruction or program in an unrestricted computer memory, and can be called and identified by an unrestricted computer processor to implement any method in the embodiments of this application, without any additional restrictions.

[0100] Furthermore, the "first" or "second" mentioned above may not only represent a sequential relationship, but may also represent a specific concept, and / or refer to the individual or collective selection of multiple elements. Clearly, those skilled in the art can make various modifications and variations to this application without departing from its scope. Therefore, if such modifications and variations fall within the scope of this application and its equivalents, this application intends to include such modifications and variations.

Claims

1. A method for early fire warning of mining vehicles based on multi-source information fusion, characterized in that, include: Multi-source monitoring of mining vehicles was conducted to obtain vehicle environmental monitoring data, vehicle operating condition data, and vehicle visual monitoring data. Based on the historical event set of vehicle fires, fire-sensitive feature detection is performed on the vehicle environmental monitoring data to obtain an environmental fire-sensitive feature vector. Based on the vehicle fire history event set, fire sign sensitive feature detection is performed on the vehicle operating condition data to obtain the operating condition fire sensitive feature vector. Based on the vehicle fire history event set, fire sign sensitive feature detection is performed on the vehicle visual monitoring data to obtain a visual fire sensitive feature vector. Based on the vehicle fire history event set, single-factor fire prediction is performed on the environmental fire sensitive feature vector, the working condition fire sensitive feature vector and the visual fire sensitive feature vector to obtain the first fire prediction map. Based on the first fire situation simulation map, a multi-factor coupled fire situation simulation is performed to obtain a second fire situation simulation map, and an early fire warning is executed based on the second fire situation simulation map. Specifically, based on the vehicle fire history event set, single-factor fire prediction is performed on the environmental fire-sensitive feature vector, the working condition fire-sensitive feature vector, and the visual fire-sensitive feature vector to obtain a first fire prediction map, including: Based on the vehicle fire history event set, single-factor classification was performed to obtain environmental fire history event set, working condition fire history event set and visual fire history event set; Based on the set of historical environmental fire events, environmental fire sensitive feature vectors are used to perform environmental factor fire inference to obtain the first fire inference result. Based on the historical event set of fires under the working conditions, the fire sensitive feature vector of the working conditions is used to extrapolate the fire situation of the working conditions and obtain the second fire situation extrapolation result. Based on the set of historical visual fire events, visual fire sensitivity feature vectors are used to perform visual factor fire inference to obtain the third fire inference result. The environmental fire-sensitive feature vector, the working condition fire-sensitive feature vector, the visual fire-sensitive feature vector, the first fire scenario prediction result, the second fire scenario prediction result, and the third fire scenario prediction result are organized to generate the first fire scenario prediction map; The process includes performing multi-factor coupled fire situation simulation based on the first fire situation simulation map to obtain a second fire situation simulation map, including: Based on the first fire scenario simulation map, environmental conditions coupled with fire scenario simulation are performed to obtain the fourth fire scenario simulation result. Based on the first fire scenario simulation map, environmental visual coupling fire scenario simulation is performed to obtain the fifth fire scenario simulation result; Based on the first fire scenario simulation map, a visual coupling fire scenario simulation under working conditions is performed to obtain the sixth fire scenario simulation result. Based on the first fire scenario simulation map, a three-factor coupled fire scenario simulation was performed to obtain the seventh fire scenario simulation result. Based on the results of the fourth, fifth, sixth, and seventh fire scenario simulations, the first fire scenario simulation map is optimized to generate the second fire scenario simulation map.

2. The method for early fire warning of mining vehicles based on multi-source information fusion as described in claim 1, characterized in that, Based on the vehicle fire history event set, fire-sensitive feature detection is performed on the vehicle environmental monitoring data to obtain an environmental fire-sensitive feature vector, including: Collect sensor feature data corresponding to each environmental monitoring parameter in the vehicle environmental monitoring data to obtain multiple monitoring sensor feature sequences; Anomaly detection is performed based on the multiple monitoring sensor feature sequences to obtain multiple sensor anomaly detection results; Based on the multiple sensor anomaly detection results, adaptive interference correction is performed on the vehicle environmental monitoring data to obtain an environmental monitoring feature sequence. The environmental monitoring feature sequence is evaluated for fire indicativeness based on the vehicle fire history event set to obtain the environmental monitoring fire indicative sequence. Based on the environmental monitoring fire indicator sequence, fire sign sensitive attention is assigned to the environmental monitoring feature sequence to generate the environmental fire sensitive feature vector.

3. The method for early fire warning of mining vehicles based on multi-source information fusion as described in claim 2, characterized in that, Based on the historical event set of vehicle fires, the environmental monitoring feature sequence is evaluated for fire indicative intensity to obtain an environmental monitoring fire indicative intensity sequence, including: Fire sign feature identification is performed based on the vehicle fire history event set to obtain a fire sign feature library; Based on the historical event set of vehicle fires, the confidence level of each fire sign feature in the fire sign feature database is evaluated to obtain a fire feature confidence sequence. Based on the fire feature confidence sequence, the fire sign feature library is cleaned according to the predetermined fire feature confidence level to establish a confidence fire sign feature space; Based on the environmental monitoring feature sequence, extract the R-th feature of environmental monitoring, where R is a positive integer; Based on each confidence fire indication feature in the confidence fire indication feature space, the Rth feature of the environmental monitoring is evaluated by twin comparison to obtain the Rth twin comparison evaluation sequence; The maximum value of the Rth twin ratio evaluation sequence is optimized to generate the Rth environmental monitoring fire indicator, and the Rth environmental monitoring fire indicator is added to the environmental monitoring fire indicator sequence.

4. The method for early fire warning of mining vehicles based on multi-source information fusion as described in claim 1, characterized in that, Based on the historical environmental fire event set, environmental fire sensitivity feature vectors are used to perform environmental factor fire scenario estimation to obtain the first fire scenario estimation result, including: Activate the fire path tracing architecture, which includes a top event node, intermediate event nodes, and a bottom event node, and the top event node, intermediate event nodes, and bottom event nodes are connected by logical relationship nodes; Based on the fire path tracing architecture, the environmental fire historical event set is traced to obtain the environmental fire tracing path set. Based on the environmental fire tracing path set, a disturbance is injected to obtain the environmental fire disturbance path set; Based on the environmental fire tracing path set and the environmental fire disturbance path set, adversarial training is performed to generate an environmental factor fire inference model. The environmental fire-sensitive feature vector is input into the environmental factor fire prediction model to obtain the first fire prediction result.

5. The method for early fire warning of mining vehicles based on multi-source information fusion as described in claim 1, characterized in that, Based on the first fire scenario simulation map, an environmental condition coupled fire scenario simulation is performed to obtain the fourth fire scenario simulation result, including: Based on the historical event set of vehicle fires, environmental and working condition coupling factor events are extracted to obtain an environmental and working condition coupled fire event set. Based on the fire path tracing architecture, the environmental condition coupled fire event set is traced to obtain the environmental condition coupled fire path set. Based on the set of fire path coupled with environmental conditions, train the fire simulation model coupled with environmental conditions. Based on the environmental fire-sensitive feature vector and the working condition fire-sensitive feature vector, and according to the environmental working condition coupled fire simulation model, the fourth fire simulation result is output.

6. The method for early fire warning of mining vehicles based on multi-source information fusion as described in claim 1, characterized in that, The vehicle visual monitoring data includes infrared thermal image data and visible light image data of the mining vehicle.

7. The method for early fire warning of mining vehicles based on multi-source information fusion as described in claim 1, characterized in that, Early fire suppression management is carried out on the mining vehicles based on the second fire scenario projection map.

8. A multi-source information fusion early fire warning system for mining vehicles, characterized in that, The system is used to perform the multi-source information fusion method for early fire warning of mining vehicles according to any one of claims 1 to 7, the system comprising: The multi-source monitoring module performs multi-source monitoring on mining vehicles to obtain vehicle environmental monitoring data, vehicle operating condition data, and vehicle visual monitoring data. The environmental fire sensitive feature vector acquisition module performs fire sign sensitive feature detection on the vehicle environmental monitoring data based on the vehicle fire history event set to obtain the environmental fire sensitive feature vector. The fire-sensitive feature vector acquisition module for operating conditions detects fire signs based on the vehicle's historical fire event set and the vehicle's operating condition data to obtain the fire-sensitive feature vector for operating conditions. The visual fire sensitive feature vector acquisition module performs fire sign sensitive feature detection on the vehicle visual monitoring data based on the vehicle fire history event set to obtain the visual fire sensitive feature vector. The fire simulation first map acquisition module performs single-factor fire simulation based on the vehicle fire history event set, the environmental fire sensitive feature vector, the working condition fire sensitive feature vector, and the visual fire sensitive feature vector to obtain the fire simulation first map. The early fire warning execution module performs multi-factor coupled fire prediction based on the first fire prediction map to obtain the second fire prediction map, and executes early fire warning based on the second fire prediction map.

Images