An industrial environment fire detection method, device, electronic equipment, storage medium and product

Abstract

The application discloses an industrial environment fire detection method and device, electronic equipment, storage medium and product. The method comprises the following steps: acquiring original environment data collected by each sensor in multiple sensors deployed in an industrial environment; obtaining target data feature information corresponding to each sensor, and performing fire detection based on the network model corresponding to each sensor and the target data feature information to obtain fire detection results of each model; converting the fire detection results to obtain a first belief quality set; determining the conflict degree between each two first belief quality sets to adjust and process the first belief quality set to obtain a second belief quality set; performing fusion processing on the second belief quality set to obtain a target belief quality set; determining a first trust degree and a first likelihood, a second trust degree and a second likelihood based on the target belief quality set, which are used to determine a target fire detection result. The industrial environment fire detection precision and reliability are improved.

Description

Technical Field

[0001] This invention relates to the field of industrial environmental fire detection technology, and in particular to an industrial environmental fire detection method, apparatus, electronic device, storage medium and product. Background Technology

[0002] In the field of industrial fire detection, traditional fire detection systems mainly rely on single-type sensors, such as smoke sensors and temperature sensors, to determine whether a fire has occurred by setting fixed thresholds. This type of fire detection scheme suffers from a single detection dimension, relying solely on changes in a single physical quantity to determine a fire, making it prone to false alarms or missed alarms due to environmental interference. Therefore, attempts have been made to use multiple sensors for fire detection, such as simultaneously using temperature and smoke sensors, transmitting the detection data from both sensors separately to a control center. The control center performs simple logical judgments on the data, issuing an alarm only when both temperature and smoke concentration exceed their respective thresholds. However, this fusion method is relatively simple and does not consider the uncertainty and conflicts in sensor data. When data from different sensors contradict each other, an accurate judgment cannot be made. This results in problems of low detection accuracy and low reliability. Summary of the Invention

[0003] This invention provides a method, apparatus, electronic device, storage medium, and product for detecting fires in industrial environments, in order to solve the problems of low accuracy and poor reliability in detecting fires in industrial environments.

[0004] According to one aspect of the present invention, an industrial environment fire detection method is provided, comprising: Acquire multi-dimensional environmental data collected by various sensors deployed in an industrial environment. The multi-dimensional environmental data includes the raw environmental data collected by each sensor. Data processing is performed on each type of raw environmental data to obtain target data feature information corresponding to each sensor. Fire detection is then performed based on the network model corresponding to each sensor and the target data feature information to obtain the fire detection results output by each network model. The fire detection results output by each network model are transformed into belief quality to obtain the first belief quality set corresponding to each network model. The first belief quality set includes: belief quality for the existence of fire, belief quality for the non-existence of fire, and belief quality for uncertainty. Determine the degree of conflict between any two sets of first belief quality, and adjust the sets of first belief quality based on the degree of conflict to obtain the adjusted sets of second belief quality. The second belief quality set corresponding to each network model is fused to obtain the fused target belief quality set. Based on the target belief quality set, the first confidence level and first likelihood corresponding to the existence of a fire, and the second confidence level and second likelihood corresponding to the absence of a fire are determined. Based on the first confidence level, first likelihood, second confidence level and second likelihood, the target fire detection result is determined.

[0005] Optionally, multiple sensors are included, such as: camera sensors, infrared sensors, gas sensors, smoke sensors, and flame sensors; multi-dimensional environmental data includes: color image data, ambient temperature data, gas concentration data, smoke concentration data, and infrared radiation intensity data; the network model corresponding to the camera sensor is: a convolutional network model based on a depthwise separable convolutional structure; the network model corresponding to the infrared sensor is: a convolutional network model based on a composite coefficient optimization strategy; the network model corresponding to the gas sensor is: a random forest model; the network model corresponding to the smoke sensor is: a support vector machine model; and the network model corresponding to the flame sensor is: a decision tree model.

[0006] Optionally, data processing is performed on each type of raw environmental data to obtain target data feature information corresponding to each sensor, including: for each sensor, preprocessing the raw environmental data collected by the sensor to obtain preprocessed target environmental data; extracting statistical features from the target environmental data to obtain data statistical feature information corresponding to the sensor; identifying the data status based on the target environmental data to obtain data status information corresponding to the sensor, wherein the data status information is used to characterize whether there is abnormal fire data in the target environmental data; and obtaining target data feature information corresponding to the sensor based on the target environmental data, data statistical feature information, and data status information.

[0007] Optionally, the fire detection results include a first predicted probability of the existence of a fire and a second predicted probability of the absence of a fire. The fire detection results output by each network model are transformed into a belief quality set to obtain a first belief quality set corresponding to each network model. This includes: for each network model's output fire detection results, determining the first predicted probability in the fire detection results as the belief quality of the existence of a fire in the first belief quality set corresponding to that network model; determining the second predicted probability in the fire detection results as the belief quality of the absence of a fire in the first belief quality set corresponding to that network model; and subtracting the belief quality of the existence of a fire and the belief quality of the absence of a fire from 1, and determining the difference as the belief quality of an uncertain fire in the first belief quality set corresponding to that network model.

[0008] Optionally, determining the degree of conflict between any two sets of first belief quality includes: determining the Hellinger distance between any two sets of first belief quality and defining the Hellinger distance as the corresponding degree of conflict.

[0009] Optionally, the first set of belief quality is adjusted based on the degree of conflict to obtain an adjusted second set of belief quality. This includes: for every two sets of first belief quality, if the degree of conflict between the two sets of first belief quality is greater than a preset degree of conflict, then the uncertainty corresponding to each set of first belief quality is determined, and the first set of belief quality is adjusted based on the uncertainty to obtain an adjusted second set of belief quality; if the degree of conflict between the two sets of first belief quality is less than or equal to the preset degree of conflict, then each set of first belief quality is directly determined as the second set of belief quality.

[0010] Optionally, the uncertainty corresponding to each of the two sets of first belief quality is determined, and the first belief quality sets are adjusted based on the uncertainty to obtain the adjusted second belief quality set. This includes: determining the Deng entropy corresponding to each of the two sets of first belief quality and determining the Deng entropy as the corresponding uncertainty; determining the target first belief quality set whose uncertainty is greater than a preset uncertainty among the two sets of first belief quality; determining the adjustment weight corresponding to the target first belief quality set based on the uncertainty corresponding to the target first belief quality set; and adjusting the target first belief quality set based on the adjustment weight to obtain the adjusted second belief quality set.

[0011] Optionally, based on the uncertainty corresponding to the target first belief quality set, the adjustment weight corresponding to the target first belief quality set is determined, including: dividing the uncertainty corresponding to the target first belief quality set by the maximum uncertainty, and subtracting 1 from the division result, and using the difference as the adjustment weight corresponding to the target first belief quality set; wherein, the maximum uncertainty refers to the maximum value among the uncertainties corresponding to all first belief quality sets.

[0012] Optionally, based on the adjustment weights, the target first belief quality set is adjusted to obtain an adjusted second belief quality set, including: multiplying the belief quality of the existence of fire in the target first belief quality set by the adjustment weights, and using the multiplication result as the belief quality of the existence of fire in the adjusted second belief quality set; multiplying the belief quality of the non-existence of fire in the target first belief quality set by the adjustment weights, and using the multiplication result as the belief quality of the non-existence of fire in the adjusted second belief quality set; and subtracting 1 from the adjusted belief quality of the existence of fire and the belief quality of the non-existence of fire, and determining the difference as the belief quality of uncertain fire in the adjusted second belief quality set.

[0013] Optionally, the second belief quality set corresponding to each network model is fused to obtain the fused target belief quality set. This includes: following the fusion order of all second belief quality sets, first fusing the first two second belief quality sets, then fusing the fused belief quality set with the third second belief quality set, until the target belief quality set is obtained after fusing with the last second belief quality set. The fusion process for each pair of second belief quality sets includes: performing cross-multiplication of all combinations in the identification framework based on the two second belief quality sets, and classifying and adding each multiplication result to determine the added belief quality for the existence of fire, the belief quality for the absence of fire, the belief quality for uncertain fire, and the conflict coefficient; based on the conflict coefficient, normalizing the added belief quality for the existence of fire, the belief quality for the absence of fire, and the belief quality for uncertain fire, and using the normalized belief quality for the existence of fire, the belief quality for the absence of fire, and the belief quality for uncertain fire as the fused belief quality set.

[0014] Optionally, the method further includes: if the conflict coefficient is 1, then the second belief quality set with the smallest uncertainty among the two second belief quality sets is determined as the fused belief quality set.

[0015] Optionally, based on the target belief quality set, determining the first confidence level and first likelihood corresponding to the existence of a fire, and the second confidence level and second likelihood corresponding to the absence of a fire, includes: determining the belief quality of the existence of a fire in the target belief quality set as the first confidence level corresponding to the existence of a fire; adding the belief quality of the existence of a fire and the belief quality of an uncertain fire in the target belief quality set, and using the sum as the first likelihood corresponding to the existence of a fire; determining the belief quality of the absence of a fire in the target belief quality set as the second confidence level corresponding to the absence of a fire; and adding the belief quality of the absence of a fire and the belief quality of an uncertain fire in the target belief quality set, and using the sum as the second likelihood corresponding to the absence of a fire.

[0016] Optionally, the target fire detection result is determined based on a first confidence level, a first likelihood level, a second confidence level, and a second likelihood level, including: if the first confidence level is greater than a preset confidence level, the target fire detection result is determined to be a fire; if the second confidence level is greater than the preset confidence level, the target fire detection result is determined to be a fire; if the first confidence level is less than or equal to the preset confidence level and the second confidence level is less than or equal to the preset confidence level, the target fire detection result is determined to be in an uncertain state, the first likelihood level and the second likelihood level are output, and when the first likelihood level is greater than the preset likelihood level, a suspected fire alarm is triggered.

[0017] Optionally, each sensor is deployed at key risk locations in the industrial environment to form a complementary and seamless environmental perception network. Each sensor continuously collects multi-dimensional environmental data based on a preset acquisition frequency and transmits the continuously collected multi-dimensional environmental data to the edge server through a publish-subscribe message transmission protocol.

[0018] According to another aspect of the present invention, an industrial environment fire detection device is provided, comprising: The raw environmental data acquisition module is used to acquire multi-dimensional environmental data collected by various sensors deployed in the industrial environment. The multi-dimensional environmental data includes the raw environmental data collected by each sensor. The fire detection result acquisition module is used to process each type of raw environmental data, obtain the target data feature information corresponding to each sensor, and perform fire detection based on the network model and target data feature information corresponding to each sensor, and obtain the fire detection result output by each network model. The first belief quality set acquisition module is used to convert the belief quality of the fire detection results output by each network model to obtain the first belief quality set corresponding to each network model. The first belief quality set includes: belief quality for the existence of fire, belief quality for the absence of fire, and belief quality for uncertainty. The second belief quality set acquisition module is used to determine the degree of conflict between every two first belief quality sets, and adjust the first belief quality sets based on the degree of conflict to obtain the adjusted second belief quality sets. The target belief quality set acquisition module is used to fuse the second belief quality set corresponding to each network model to obtain the fused target belief quality set. The target fire detection result determination module is used to determine the first confidence level and first likelihood corresponding to the existence of a fire and the second confidence level and second likelihood corresponding to the absence of a fire based on the target belief quality set, and to determine the target fire detection result based on the first confidence level, first likelihood, second confidence level and second likelihood.

[0019] According to another aspect of the present invention, an electronic device is provided, the electronic device comprising: At least one processor; and A memory that is communicatively connected to at least one processor; wherein, The memory stores a computer program that can be executed by at least one processor, such that the at least one processor is able to perform the industrial environment fire detection method according to any embodiment of the present invention.

[0020] According to another aspect of the present invention, a computer-readable storage medium is provided, which stores computer instructions for causing a processor to execute and implement the industrial environment fire detection method of any embodiment of the present invention.

[0021] According to another aspect of the present invention, a computer program product is provided, comprising a computer program that, when executed by a processor, implements the industrial environment fire detection method of any embodiment of the present invention.

[0022] The technical solution of this invention acquires multi-dimensional environmental data collected by various sensors deployed in an industrial environment. This multi-dimensional environmental data includes raw environmental data collected by each sensor, comprehensively covering key environmental information in the industrial scenario and avoiding blind spots in monitoring from a single data source. Data processing is performed on each type of raw environmental data to obtain target data feature information corresponding to each sensor. Fire detection is then performed based on the network model corresponding to each sensor and the target data feature information, obtaining fire detection results output by each network model. This allows for targeted adaptation to the data characteristics of different sensors, improving the accuracy and adaptability of single-source detection. The fire detection results output by each network model are then transformed into belief quality to obtain a first belief quality set corresponding to each network model. This first belief quality set includes: belief quality regarding the existence of a fire, belief quality regarding the absence of a fire, and belief quality regarding uncertainty. The first belief quality set quantifies the uncertainty of detection results and adapts to complex noise interference in industrial environments. It determines the degree of conflict between any two first belief quality sets and adjusts these sets based on the conflict level to obtain an adjusted second belief quality set. This effectively reduces interference from high-conflict anomaly detection results and improves information reliability. The second belief quality sets corresponding to each network model are fused to obtain a fused target belief quality set, which fully integrates multi-source complementary information and enhances the robustness of overall judgment. Based on the target belief quality set, it determines the first confidence level and first likelihood corresponding to the presence of a fire, and the second confidence level and second likelihood corresponding to the absence of a fire. Based on the first confidence level, first likelihood, second confidence level, and second likelihood, it determines the target fire detection result, reducing the probability of false positives and false negatives based on rigorous uncertainty reasoning. This solution fully leverages the complementarity of multi-source sensor information and reduces the risk of false detection and missed detection by a single sensor or model through belief quality conversion and conflict resolution. It combines multi-dimensional perception with intelligent analysis, completely solving the false and missed detection problems of single / simple fusion systems. It effectively handles data uncertainty and information conflict in complex industrial environments, addressing the issues of low accuracy and poor reliability in industrial fire detection, and significantly improving the reliability, robustness, and detection accuracy of fire judgment.

[0023] It should be understood that the description in this section is not intended to identify key or essential features of the embodiments of the present invention, nor is it intended to limit the scope of the invention. Other features of the invention will become readily apparent from the following description. Attached Figure Description

[0024] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0025] Figure 1 This is a flowchart of an industrial environment fire detection method provided in Embodiment 1 of the present invention; Figure 2 This is a flowchart of an industrial environment fire detection method provided according to Embodiment 2 of the present invention; Figure 3 This is a schematic diagram of the structure of an industrial environment fire detection device according to Embodiment 3 of the present invention; Figure 4 This is a schematic diagram of the structure of an electronic device that implements the industrial environment fire detection method of this invention. Detailed Implementation

[0026] To enable those skilled in the art to better understand the present invention, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of the present invention.

[0027] It should be noted that the terms "first," "second," etc., in the specification, claims, and accompanying drawings of this invention are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments of the invention described herein can be implemented in orders other than those illustrated or described herein. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover a non-exclusive inclusion; for example, a process, method, system, product, or apparatus that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.

[0028] Example 1 Figure 1 This is a flowchart illustrating an industrial environment fire detection method according to Embodiment 1 of the present invention. This embodiment is applicable to fire detection in industrial environments. The method can be executed by an industrial environment fire detection device, which can be implemented in hardware and / or software and can be configured in electronic devices such as computers and servers. Figure 1 As shown, the method includes: S110. Acquire multi-dimensional environmental data collected by various sensors deployed in an industrial environment. The multi-dimensional environmental data includes the raw environmental data collected by each sensor.

[0029] Among these, multiple sensors refer to different types of monitoring equipment deployed at key risk points in the industrial environment, including but not limited to camera sensors, infrared sensors, gas sensors, smoke sensors, and flame sensors. Each sensor is responsible for sensing environmental information in its corresponding dimension, providing comprehensive input for subsequent processing. Multi-dimensional environmental data refers to the various types of raw information collected simultaneously by multiple deployed sensors, reflecting the existence of fire hazards in the environment from different perspectives, providing comprehensive and diverse input for subsequent model judgments and belief fusion.

[0030] Specifically, multiple types of sensors are deployed at key risk locations in the industrial environment to build a comprehensive and complementary environmental sensing network. This network continuously acquires multi-dimensional raw environmental data at a preset acquisition frequency, enabling real-time collection of multi-source monitoring information.

[0031] In this embodiment, multi-dimensional environmental data collected by multiple sensors can comprehensively cover fire characteristic signals in industrial scenarios. Multi-dimensional and multi-location data collection effectively avoids blind spots in single-sensor monitoring, providing rich and reliable raw information support for subsequent fire detection, improving the perception integrity and anti-interference capability of the entire detection system, effectively supporting intelligent decision-making and refined management, and reducing manual inspection costs and safety hazards.

[0032] Optionally, multiple sensors are included, such as: camera sensors, infrared sensors, gas sensors, smoke sensors, and flame sensors; multi-dimensional environmental data includes: color image data, ambient temperature data, gas concentration data, smoke concentration data, and infrared radiation intensity data; the network model corresponding to the camera sensor is: a convolutional network model based on a depthwise separable convolutional structure; the network model corresponding to the infrared sensor is: a convolutional network model based on a composite coefficient optimization strategy; the network model corresponding to the gas sensor is: a random forest model; the network model corresponding to the smoke sensor is: a support vector machine model; and the network model corresponding to the flame sensor is: a decision tree model.

[0033] In this embodiment, in an industrial environmental monitoring scenario, multiple sensing devices such as camera sensors, infrared sensors, gas sensors, smoke sensors, and flame sensors are deployed to collect multi-dimensional raw environmental data, including color image data, ambient temperature data, gas concentration data, smoke concentration data, and infrared radiation intensity. Then, intelligent models are matched and adapted to the characteristics of different sensor data for processing. Specifically, the camera sensor uses a convolutional network model based on a depthwise separable convolutional structure to process image information, the infrared sensor uses a convolutional network model based on a composite coefficient optimization strategy to analyze infrared radiation data, and the gas sensor, smoke sensor, and flame sensor use a random forest model, a support vector machine model, and a decision tree model, respectively, to complete the identification and judgment of corresponding concentrations and features, thereby achieving efficient parsing and fusion of multi-source heterogeneous environmental data.

[0034] In this embodiment, the advantages of adapting different sensors and corresponding models can be fully utilized, ensuring the comprehensiveness of multi-dimensional information collection such as images, temperature, gas, smoke, and flames. At the same time, the data processing accuracy and inference efficiency can be improved through dedicated models. It takes into account both complex visual feature extraction and rapid classification of traditional sensor data, effectively enhancing the accuracy and real-time performance of environmental anomaly identification and safety hazard early warning, and providing stable and reliable multi-dimensional data support for industrial safety monitoring and intelligent management.

[0035] S120. Perform data processing on each type of raw environmental data to obtain target data feature information corresponding to each sensor, and perform fire detection based on the network model corresponding to each sensor and the target data feature information to obtain the fire detection results output by each network model.

[0036] Specifically, target data feature information can be understood as key features that effectively characterize the fire state. This feature information can be extracted from the raw environmental data of the sensors after processing such as cleaning, noise reduction, normalization, and feature extraction. It includes, but is not limited to, flame texture features in images, temperature abrupt changes in infrared data, and abnormal concentration changes in gas and smoke data. This serves as the core input for the network model to make fire judgments. Fire detection results can be understood as initially reflecting the existence of fire hazards under the corresponding monitoring dimension, providing a basis for subsequent belief conversion and multi-source fusion. The judgment results can be inferred and output by the network models corresponding to each sensor based on the target data feature information.

[0037] Specifically, targeted data processing such as denoising, normalization, and feature extraction can be carried out on the raw environmental data collected by various sensors to extract target data feature information related to fire monitoring. Then, the processed target data feature information is input into the respective matched dedicated network models for inference calculation, thereby obtaining the fire detection results output independently by each model.

[0038] For example, the entire preprocessing process of data cleaning, normalization, and feature extraction is completed at the edge node: sensor noise is removed by Kalman filtering (processing time < 0.1 seconds), data range is unified by min-max normalization (< 0.05 seconds), key features (such as flame outline and gas concentration change rate in RGB images, < 0.2 seconds) are extracted, and only 10% of the feature data is transmitted to subsequent modules, reducing the data volume by 90%. In a rapid fire scenario in a metallurgical workshop (fire spread speed > 1 m / s), it only takes 0.35 seconds from sensor acquisition to edge preprocessing completion, while traditional cloud platform systems require 2-3 seconds for data transmission alone, improving response efficiency by more than 7 times and saving critical time for personnel evacuation and firefighting.

[0039] In this embodiment, data quality is improved and fire characteristics are highlighted through data processing. Combined with a dedicated network model, accurate identification is achieved. This not only fully leverages the data value of different sensors but also improves the accuracy of single-source detection using the model. This provides a reliable foundation for subsequent multi-source evidence fusion and effectively enhances the overall fire detection's anti-interference capability and judgment accuracy.

[0040] Optionally, data processing is performed on each type of raw environmental data to obtain target data feature information corresponding to each sensor, including: for each sensor, preprocessing the raw environmental data collected by the sensor to obtain preprocessed target environmental data; extracting statistical features from the target environmental data to obtain data statistical feature information corresponding to the sensor; identifying the data status based on the target environmental data to obtain data status information corresponding to the sensor, wherein the data status information is used to characterize whether there is abnormal fire data in the target environmental data; and obtaining target data feature information corresponding to the sensor based on the target environmental data, data statistical feature information, and data status information.

[0041] Specifically, statistical data features can be understood as quantitative indicators extracted from the preprocessed target environmental data from various sensors, reflecting the data distribution and variation patterns. These indicators include mean, variance, extreme values, rate of change, and fluctuation amplitude, objectively describing the overall trend and dispersion of environmental parameters. Data status information can be understood as the identification results obtained after anomaly detection based on the target environmental data. It is mainly used to mark whether the current data shows abnormal states related to fire, such as sudden temperature increases, excessive gas concentrations, abnormal smoke concentrations, or sudden changes in infrared radiation. This intuitively characterizes whether there are signs of fire hazards in the monitored environment. Together, these two aspects constitute more discriminative target data feature information.

[0042] Specifically, for the raw environmental data collected by various sensors, standardized preprocessing is carried out in sequence to remove noise, fill in missing values ​​and obtain standardized target environmental data. Then, statistical feature extraction is performed to obtain corresponding data statistical feature information. At the same time, the data status is identified to determine whether there is fire-related abnormal data, including but not limited to sudden temperature rise, excessive gas concentration, abnormal smoke or flame, etc. Finally, the target environmental data, data statistical feature information and data status information are integrated to form the target data feature information specific to each sensor.

[0043] In this embodiment, a complete transformation from raw data to effective features is achieved through hierarchical processing. This not only ensures data quality and reduces the impact of interference factors, but also comprehensively portrays fire-related features from multiple levels, including numerical values, trends, and abnormal states. This enables subsequent models to conduct detection based on more complete and identifiable information, significantly improving the reliability and anti-interference capability of fire identification.

[0044] S130. Perform belief quality transformation on the fire detection results output by each network model to obtain the first belief quality set corresponding to each network model. The first belief quality set includes: belief quality for the existence of fire, belief quality for the non-existence of fire, and belief quality for uncertainty.

[0045] The conversion of belief quality refers to mapping the fire detection results output by the network model into quantified probabilistic values ​​according to rules, transforming a single judgment into a computable and fusionable belief distribution. The first set of belief quality refers to a set of structured confidence scores obtained after the conversion, representing the belief quality of "a fire exists," the belief quality of "a fire does not exist," and the "uncertainty" belief quality that cannot be clearly judged due to data ambiguity or insufficient information. This provides a unified quantitative basis for the fusion decision of multi-sensor fire detection results.

[0046] Specifically, the fire detection results output by the network models corresponding to each sensor are first standardized by performing a belief quality transformation. The fire detection results output by each network model are transformed into a first belief quality set containing three categories of belief quality: "fire exists", "fire does not exist" and "uncertainty". The discrete discrimination results are then transformed into a quantifiable and fusionable probability distribution.

[0047] In this embodiment, the belief transformation method can retain the fuzzy information and uncertain components in the model judgment, and is no longer limited to simple "yes / no" binary results. It is more in line with the complex and ever-changing real situation of industrial environment, and provides a unified and flexible expression form for subsequent conflict calculation, evidence adjustment and multi-source fusion, effectively improving the detail of fire detection and the reliability of subsequent decision-making.

[0048] Optionally, the fire detection results include a first predicted probability of the existence of a fire and a second predicted probability of the absence of a fire. The fire detection results output by each network model are transformed into a belief quality set to obtain a first belief quality set corresponding to each network model. This includes: for each network model's output fire detection results, determining the first predicted probability in the fire detection results as the belief quality of the existence of a fire in the first belief quality set corresponding to that network model; determining the second predicted probability in the fire detection results as the belief quality of the absence of a fire in the first belief quality set corresponding to that network model; and subtracting the belief quality of the existence of a fire and the belief quality of the absence of a fire from 1, and determining the difference as the belief quality of an uncertain fire in the first belief quality set corresponding to that network model.

[0049] The first predicted probability refers to the numerical value output by the network model corresponding to various sensors after completing fire detection inference, representing the likelihood of a fire in the current monitoring environment. This value can be obtained by the network model based on the target data feature information of the corresponding sensors, predicting the probability of a fire in the current environment. The second predicted probability refers to the numerical value output by the same model simultaneously, representing the likelihood of no fire in the current environment. Both are probability values ​​between 0 and 1, jointly reflecting the model's confidence in the fire state and serving as the direct basis for subsequent conversion into belief quality and multi-source information fusion.

[0050] Specifically, based on the fire detection results output by each network model, which include a first predicted probability of the existence of a fire and a second predicted probability of the absence of a fire, when performing belief quality conversion, the first predicted probability is directly assigned to the belief quality of "the existence of a fire" in the first belief quality set of the corresponding model, and the second predicted probability is assigned to the belief quality of "the absence of a fire". The difference is then obtained by subtracting the sum of the first two belief qualities from 1, which is taken as the belief quality of "uncertainty". This completes the construction of the first belief quality set corresponding to each model.

[0051] In this embodiment, the method of converting belief quality has the advantages of being logically intuitive and computationally simple. It can quickly convert the model probability output into standardized belief quality, fully retain the original prediction credibility, and objectively reflect the data ambiguity and model discrimination hesitation by explicitly calculating the uncertainty belief quality. This avoids the one-sidedness of single probability judgment and provides structurally unified and quantifiable evidence information for subsequent multi-sensor information fusion, effectively improving the robustness and decision rationality of fire detection.

[0052] S140. Determine the degree of conflict between every two sets of first belief quality, and adjust the sets of first belief quality based on the degree of conflict to obtain the adjusted sets of second belief quality.

[0053] The degree of conflict is used to measure the degree of contradiction and inconsistency among the first belief quality sets output by different sensor models in terms of "the existence of a fire", "the absence of a fire", and uncertainty judgments, reflecting the deviation and conflict between multi-source detection evidence. The second belief quality set is a new belief quality set obtained by weighting and modifying the original first belief quality set after calculating the above-mentioned degree of conflict. It is used to weaken the adverse effects of highly conflicting evidence, make multi-source evidence more coordinated and reliable, and provide a more reasonable basis for subsequent fire decision-making.

[0054] Specifically, a preset conflict degree calculation method is invoked to calculate the conflict degree between any two sets of first belief quality sets. This measures the degree of contradiction and inconsistency in the fire discrimination results of different sensor models. Then, the original belief quality is adjusted by weighting, modifying or reconstructing according to the conflict degree to obtain a more reasonable and more consistent second belief quality set. For example, when the conflict is too high, the belief quality is adaptively adjusted by uncertainty and weight. If the conflict is within a reasonable range, the original set is directly retained, and finally a more consistent second belief quality set is obtained.

[0055] In this embodiment, by first determining the degree of conflict and then dynamically adjusting, conflict information between multi-source detection results can be effectively identified and quantified, avoiding belief distortion caused by sensor anomalies, model biases, or environmental interference. Dynamic adjustment reduces the negative impact of highly conflicting evidence on decision-making, improves the reliability and coordination of belief distribution, and provides a more stable and credible evidentiary basis for subsequent fusion decision-making.

[0056] Optionally, determining the degree of conflict between any two sets of first belief quality includes: determining the Hellinger distance between any two sets of first belief quality and defining the Hellinger distance as the corresponding degree of conflict.

[0057] Hellinger distance is a metric used to measure the similarity between two probability distributions or belief quality distributions. It reflects the degree of deviation by calculating the difference between the two sets of data distributions. For example, the value can be set between 0 and 1. The smaller the distance, the closer the two sets of belief quality sets are and the lower the conflict. The larger the distance, the more significant the difference between the two sets and the higher the degree of conflict. It has the characteristics of numerical stability, moderate sensitivity and no influence from extreme values. It can objectively and accurately quantify the degree of contradiction between multi-sensor fire detection results.

[0058] Specifically, by calculating the Hellinger distance between every two sets of first belief masses and using this distance value directly as the degree of conflict between the corresponding two sets of belief masses, the differences and contradictions in fire judgment results of different sensor models can be quantified.

[0059] In this embodiment, the Hellinger distance is used to measure the degree of conflict. It has the advantages of stable calculation, standardized values, sensitivity to differences in probability distribution, and no excessive interference from outliers. It can accurately reflect the degree of deviation between multiple sources of evidence, while ensuring that the conflict measurement results are smooth, reasonable, and have clear physical meaning. It provides an objective and reliable quantitative basis for subsequent belief quality adjustment and effectively improves the accuracy and robustness of multi-model fire detection result fusion.

[0060] S150. Perform fusion processing on the second belief quality set corresponding to each network model to obtain the fused target belief quality set.

[0061] Among them, the target belief quality set refers to the final belief distribution result obtained after fusing the second belief quality sets corresponding to each sensor. It also includes three types of belief quality: "fire exists", "fire does not exist" and "uncertainty". It integrates the detection information of all sensors, corrects the conflict differences between evidence, and can comprehensively and objectively reflect the true fire status of the overall monitoring environment, providing the most reliable and unified quantitative basis for the final fire judgment and early warning decision.

[0062] Specifically, the second belief quality sets obtained from conflict-corrected sensor models are comprehensively fused according to preset fusion rules. This unifies and aggregates multi-source, heterogeneous fire judgment evidence, ultimately generating a set of target belief quality sets that takes into account information from various sensors and is globally consistent. The evidence fusion rules can be a standardized method for synthesizing multiple conflict-corrected second belief quality sets. Through mathematical operations, fire judgment information from multiple sensors and models is weighted, combined, and normalized, unifying scattered and locally contradictory belief distributions into a consistent global belief. This allows for reasonable allocation of confidence levels, weakening conflicting information, and strengthening consistent evidence, ultimately yielding robust and reliable fusion results and providing a scientific basis for final fire judgment.

[0063] In this embodiment, by fusing multiple sources of belief, the errors and interference of a single sensor can be effectively offset, and complementary information can be fully utilized to improve the reliability of judgment, making the final fire decision more robust and accurate, and significantly reducing the false alarm and false alarm rates in complex industrial environments.

[0064] Optionally, the second belief quality set corresponding to each network model is fused to obtain the fused target belief quality set. This includes: following the fusion order of all second belief quality sets, first fusing the first two second belief quality sets, then fusing the fused belief quality set with the third second belief quality set, until the target belief quality set is obtained after fusing with the last second belief quality set. The fusion process for each pair of second belief quality sets includes: performing cross-multiplication of all combinations in the identification framework based on the two second belief quality sets, and classifying and adding each multiplication result to determine the added belief quality for the existence of fire, the belief quality for the absence of fire, the belief quality for uncertain fire, and the conflict coefficient; based on the conflict coefficient, normalizing the added belief quality for the existence of fire, the belief quality for the absence of fire, and the belief quality for uncertain fire, and using the normalized belief quality for the existence of fire, the belief quality for the absence of fire, and the belief quality for uncertain fire as the fused belief quality set.

[0065] Specifically, the second belief quality sets of each sensor are iteratively fused in a preset fusion order. First, the first two sets of belief quality sets are cross-multiplied according to the recognition framework for all combinations. The product results of similar propositions are classified and summed to obtain three preliminary belief quality sets and conflict coefficients: fire, non-fire, and uncertainty. Then, the above results are normalized using the conflict coefficients to generate intermediate fusion results. Subsequently, the above cross-multiplication, classification and addition, conflict calculation and normalization steps are repeated with the next set of second belief quality sets until all sets are fused and the final target belief quality set is obtained.

[0066] For example, the specific steps for fusing the second belief quality set corresponding to each network model to obtain the fused target belief quality set are as follows: Step 1: Define the evidence (mass function) for the two sensors. Temperature sensor (Evidence 1): m1(abnormal) = 0.6 (60% certainty of abnormality), m1(normal) = 0.3 (30% certainty of normality), m1(uncertain) = 0.1 (10% unknown), Uncertain = could be either normal or abnormal, indistinguishable. RGB sensor (Evidence 2): m2(abnormal) = 0.5 (50% certainty of abnormality), m2(normal) = 0.2 (20% certainty of normality), m2(uncertain) = 0.3 (30% unknown).

[0067] Step 2: Multiply all evidence in pairs (core operation): DS involves cross-multiplying all combinations, resulting in a total of 3×3=9 combinations: Abnormal × Abnormal = 0.6×0.5 = 0.30 → assigned to Abnormal; Abnormal × Normal = 0.6×0.2 = 0.12 → conflict (one side is abnormal, the other is normal); Abnormal × Uncertain = 0.6×0.3 = 0.18 → assigned to Abnormal; Normal × Abnormal = 0.3×0.5 = 0.15 → conflict; Normal × Normal = 0.3×0.2 = 0.06 → assigned to Normal; Normal × Uncertain = 0.3×0.3 = 0.09 → assigned to Normal; Uncertain × Abnormal = 0.1×0.5 = 0.05 → assigned to Abnormal; Uncertain × Normal = 0.1×0.2 = 0.02 → assigned to Normal; Uncertain × Uncertain = 0.1×0.3 = 0.03 → assigned to Uncertain.

[0068] Step 3: Isolate the conflicting parts: Conflict = Completely opposite opinions. Abnormal × Normal = 0.12, Normal × Abnormal = 0.15, Total Conflict K = 0.12 + 0.15 = 0.27. DS will discard the conflicting parts and only calculate the remaining ones. The remaining usable total weight = 1 - K = 1 - 0.27 = 0.73.

[0069] Step 4: Sum the results of the same type: the sum of those attributed to "abnormal" is 0.30 + 0.18 + 0.05 = 0.53, the sum of those attributed to "normal" is 0.06 + 0.09 + 0.02 = 0.17, and the sum of those attributed to "uncertain" is 0.03 = 0.03. The unnormalized results are: Abnormal: 0.53, Normal: 0.17, Uncertain: 0.03, and the sum is 0.53 + 0.17 + 0.03 = 0.73 (which is exactly equal to 1-K).

[0070] Step 5: Normalization (divided by 0.73): Final trust level m (final): m (abnormal) = 0.53 / 0.73 ≈ 0.726, m (normal) = 0.17 / 0.73 ≈ 0.233, m (uncertain) = 0.03 / 0.73 ≈ 0.041.

[0071] In this embodiment, by combining iterative fusion with cross-combination calculation and conflict normalization, the evidence combination rules can be strictly followed to fully aggregate multi-source heterogeneous detection information, accurately quantify and resolve judgment conflicts between sensors, fully preserve the belief distribution of various propositions, effectively suppress interference caused by single sensor anomalies or model biases, make the fusion results more globally consistent and reliable, and significantly improve the accuracy and anti-interference ability of fire detection decisions.

[0072] Optionally, the method further includes: if the conflict coefficient is 1, then the second belief quality set with the smallest uncertainty among the two second belief quality sets is determined as the fused belief quality set.

[0073] Specifically, when fusing two sets of second belief quality, if the calculated conflict coefficient is 1, it indicates that the two sets of evidence are completely contradictory and cannot be effectively fused using conventional rules. In this case, the set of second belief quality with the smallest uncertainty is directly selected as the result of this fusion, thereby completing the robust fusion process under abnormal conflict.

[0074] In this embodiment, by providing a clear and feasible processing strategy for the extreme case of complete conflict, the fusion failure or calculation abnormality caused by the conflict coefficient being 1 is avoided. At the same time, by selecting the belief set with the least uncertainty, more reliable discrimination information can be retained when the evidence is completely contradictory. This ensures the continuous and stable fusion process and improves the rationality and usability of fire detection results in extreme conflict scenarios.

[0075] S160. Based on the target belief quality set, determine the first confidence level and first likelihood corresponding to the existence of a fire, and the second confidence level and second likelihood corresponding to the absence of a fire. Based on the first confidence level, first likelihood, second confidence level and second likelihood, determine the target fire detection result.

[0076] The first confidence level refers to the minimum level of confidence calculated from the target belief quality set, used to clearly support the existence of a fire. The first likelihood level refers to the maximum probability range that does not rule out the existence of a fire. The second confidence level refers to the minimum level of confidence that clearly supports the absence of a fire. The second likelihood level refers to the maximum probability range that does not rule out the absence of a fire. The target fire detection result refers to the final judgment conclusion derived by combining the above confidence and likelihood levels, used to clearly determine whether a fire has occurred in the current environment, providing the final basis for fire early warning and safe handling.

[0077] Specifically, from the fused set of target belief quality, the first confidence level and first likelihood level corresponding to "fire exists" and the second confidence level and second likelihood level corresponding to "fire does not exist" are calculated respectively. Then, the decision is made by combining these four indicators, and the target fire detection result is finally determined.

[0078] In this embodiment, by combining both trust and likelihood dimensions for decision-making, the supporting information and probability range in the fused belief can be fully utilized to effectively distinguish between certain and uncertain intervals, reduce the risk of misjudgment caused by a single judgment, and make the final fire detection result more rigorous, objective and reliable, thereby improving the accuracy and credibility of fire identification in complex industrial environments.

[0079] Optionally, based on the target belief quality set, determining the first confidence level and first likelihood corresponding to the existence of a fire, and the second confidence level and second likelihood corresponding to the absence of a fire, includes: determining the belief quality of the existence of a fire in the target belief quality set as the first confidence level corresponding to the existence of a fire; adding the belief quality of the existence of a fire and the belief quality of an uncertain fire in the target belief quality set, and using the sum as the first likelihood corresponding to the existence of a fire; determining the belief quality of the absence of a fire in the target belief quality set as the second confidence level corresponding to the absence of a fire; and adding the belief quality of the absence of a fire and the belief quality of an uncertain fire in the target belief quality set, and using the sum as the second likelihood corresponding to the absence of a fire.

[0080] Specifically, the confidence and likelihood are calculated based on the fused set of target belief qualities. The confidence quality of "there is a fire" is taken as the first confidence quality, and it is added to the confidence quality of uncertain beliefs to obtain the first likelihood. The confidence quality of "there is no fire" is taken as the second confidence quality, and it is added to the confidence quality of uncertain beliefs to obtain the second likelihood.

[0081] In this embodiment, based on the target belief quality set, the calculation method for determining the first confidence level and first likelihood corresponding to the existence of a fire, and the second confidence level and second likelihood corresponding to the absence of a fire, is characterized by clear rules, high computational efficiency, and the ability to intuitively reflect the degree of certainty and the maximum possible range of fire and non-fire propositions. It retains deterministic judgment information while fully incorporating uncertain factors, making the decision-making basis more complete and reasonable, and effectively improving the rigor and anti-interference ability of the final fire detection results.

[0082] Optionally, the target fire detection result is determined based on a first confidence level, a first likelihood level, a second confidence level, and a second likelihood level, including: if the first confidence level is greater than a preset confidence level, the target fire detection result is determined to be a fire; if the second confidence level is greater than the preset confidence level, the target fire detection result is determined to be a fire; if the first confidence level is less than or equal to the preset confidence level and the second confidence level is less than or equal to the preset confidence level, the target fire detection result is determined to be in an uncertain state, the first likelihood level and the second likelihood level are output, and when the first likelihood level is greater than the preset likelihood level, a suspected fire alarm is triggered.

[0083] Among them, the preset confidence level refers to a judgment threshold set in advance according to the actual fire detection needs. It is used to measure whether the belief of "a fire exists" or "a fire does not exist" reaches a sufficiently reliable decision-making standard. When a certain confidence level exceeds the threshold, the system can make a clear and definite judgment. Otherwise, it is considered that the evidence is insufficient and cannot be directly judged. This provides a unified and standardized decision-making basis for fire detection results and ensures the rigor and consistency of early warning and judgment.

[0084] Specifically, a tiered decision-making process is implemented based on the first confidence level, the first likelihood level, the second confidence level, and the second likelihood level: when the first confidence level corresponding to a fire is greater than the preset confidence level, the target fire detection result is directly determined to be a fire; when the second confidence level corresponding to a non-fire is greater than the preset confidence level, it is determined to be a fire; if both confidence levels are not greater than the threshold, it is determined to be an uncertain state and the corresponding likelihood level is output. At the same time, a suspected fire alarm is triggered when the fire likelihood level exceeds the preset value.

[0085] In this embodiment, the decision-making method for determining the target fire detection result based on the first confidence level, the first likelihood level, the second confidence level, and the second likelihood level has clear logic and clear judgment criteria. It relies on the confidence level to achieve highly reliable deterministic judgment, and uses the likelihood level to refine the handling of ambiguous scenarios. It takes into account both the accuracy and safety of fire detection, effectively avoids misjudgment or omission due to insufficient evidence, and improves the rationality and practicality of fire early warning in complex environments.

[0086] Based on the above embodiments, a complementary and comprehensive environmental sensing network with no blind spots is formed for each type of key risk location prone to fire. Each sensor continuously collects multi-dimensional environmental data at a preset collection frequency and transmits the data to the edge server in real time through a publish-subscribe messaging protocol.

[0087] In this embodiment, the deployment and transmission method can achieve comprehensive perception of key areas, avoid monitoring blind spots, improve data reliability through multi-sensor complementarity, ensure efficient, stable, low-latency, and easily scalable transmission through the publish-subscribe model, and reduce cloud pressure through edge processing, thereby improving the real-time performance, completeness, and robustness of industrial environment fire detection as a whole.

[0088] In specific embodiments, a modular and scalable layered architecture can be adopted, divided from bottom to top into a device sensor layer, a communication layer, an edge field processing layer, a machine learning model layer, a Dempster-Shafer prediction layer, and a central system / cloud layer. Each layer forms a closed loop through data flow and command interaction to ensure real-time performance and reliability. The architecture is centered on edge computing to achieve localized data processing, while relying on the cloud for long-term data storage and model optimization. Specific architecture details are as follows: The device sensor layer deploys five types of dedicated devices: RGB sensors, IR sensors, gas sensors, smoke sensors, and flame sensors, continuously collecting raw fire-related data; the communication layer uses the MQTT (Message Queuing Telemetry Transport) protocol to achieve low-latency data transmission between sensors and edge servers, employing a publish-subscribe model to ensure efficient data distribution; the edge field processing layer is built on a low-cost single-board computer, integrating modules for data acquisition, security verification, data filtering, normalization, feature extraction, and anomaly detection to complete the preprocessing and local storage of raw data; the machine learning model layer runs TensorFlow. The Lite lightweight environment deploys five types of models: convolutional network models based on composite coefficient optimization strategies, random forest models, support vector machine models, and decision tree models. Examples of these models include MobileNet, EfficientNet, RandomForest, SVM, and DecisionTree. Each model processes preprocessed data from its corresponding sensor and outputs a fire probability prediction. The Dempster-Shafer prediction layer fuses the outputs of multiple models into a final fire decision through steps such as belief quality transformation, conflict assessment, uncertainty adjustment, and evidence fusion. The central system / cloud layer is responsible for long-term data archiving, federated learning model updates, and NFPA (National Fire Protection Association) standard compliance verification, optimizing edge model performance through a feedback loop. The detailed workflow of each layer is shown below: 1. Equipment sensor layer: Multi-dimensional fire feature data acquisition equipment sensor layer is the data source of fire detection method. It can realize the real-time acquisition of multi-dimensional fire features through five types of special sensors, ensuring coverage of the full stage features of fire occurrence, and providing comprehensive input for subsequent processing. It should be noted that (1) Sensor selection and function RGB sensor: adopts camera module, has image acquisition and preliminary processing capabilities, mainly captures the visual features of visible flames, and reflects the color, texture and shape of flames through the three-channel pixel values ​​of red (R), green (G) and blue (B). For example, flames usually appear red and orange, corresponding to the characteristics of high R channel value (>200) and low G and B channel value (<100); Infrared (IR) sensor: based on infrared thermal imaging technology, detects the temperature distribution in the environment, outputs temperature data in degrees Celsius (°C), is not affected by smoke obstruction, light changes, etc., can identify early thermal anomalies of fire, and the temperature detection range Covering temperatures from 0-500℃, suitable for both normal oil and gas environments and high-temperature fire scenarios; Gas sensor: Employing a tin dioxide (SnO2) semiconductor sensing element, when it detects characteristic combustion gases such as carbon monoxide (CO), carbon dioxide (CO2), benzene (C6H6), and alcohol, the element's conductivity changes, outputting gas concentration data in ppm (parts per million by volume), enabling the identification of gas leaks or early combustion before a fire develops into an open flame; Smoke sensor: Also based on the SnO2 semiconductor principle, specifically detecting smoke particles and combustible gases (such as methane CH4 and liquefied petroleum gas LPG), outputting smoke concentration (unit: μg / m³). 3 (2) Data Acquisition Method and Timing All sensors continuously acquire data at a frequency of 10Hz, that is, generate a set of raw data every 100ms to ensure that the weak features in the early stage of the fire are not missed. During the data acquisition process, each sensor outputs a set of raw data in a specific format according to its own principle: RGB sensor outputs a set of pixel points. ,in, , ; IR sensor output temperature data set , Gas sensor output concentration data set Smoke sensor output concentration data set Flame sensor output intensity data set , The timing of data collection is not limited by environmental conditions. Whether in normal operation or in an emergency, the sensor continues to work to ensure the continuity and integrity of the data. (3) Deployment scenario and environment adaptation: The sensor is deployed at key risk points in the oil and gas industry, including the area around the oil storage tank, the oil pipeline joint, the refueling / gas filling operation area, and the pump station room. In order to adapt to the high temperature, high humidity and dust characteristics of the industrial environment, the sensor adopts industrial-grade protection design, such as IP65 dustproof and waterproof rating. At the same time, the detection accuracy is ensured by regular calibration (once a month) to avoid drift caused by environmental factors. For example, gas sensors need to avoid direct contact with oil stains, and IR sensors need to have their lens dust cleaned regularly to ensure the accuracy of data acquisition.

[0089] 2. Communication Layer: After collecting multi-dimensional environmental data, communication is required through the low-latency data transmission and distribution communication layer. The low-latency data transmission and distribution communication layer is the key link connecting the sensor and the edge processing layer. Its core objective is to achieve low-latency and highly reliable data transmission, avoid the high overhead problem of the traditional TCP / IP protocol in the Internet of Things scenario, and ensure that the raw data is delivered to the edge server quickly. It should also be noted that (1) the communication protocol and architecture adopt the MQTT protocol as the core communication protocol. This protocol is a lightweight publish-subscribe message transmission protocol designed specifically for the Internet of Things. It has advantages such as small message format (minimum header is only 2 bytes), low bandwidth consumption, and support for QoS (Quality of Service) levels, and is suitable for the limited network resources in the industrial environment. The communication architecture consists of an MQTT broker and an IoT gateway: the MQTT broker is responsible for managing topics and message routing, and assigns independent topics to each type of sensor, such as "oil-gas / sensor / rgb" for RGB sensors and "oil-gas / sensor / ir" for IR sensors; the IoT gateway (built on Raspberry Pi) is deployed in the field control room as an intermediate node between the sensors and the edge server. On the one hand, it receives the "publish" messages from the sensors, and on the other hand, it forwards the messages to the "subscribe" ports of the edge server to achieve centralized data distribution. (2) Data transmission process and timing After each set of data collection is completed (100ms interval), the sensor immediately pushes the raw data to the corresponding MQTT topic as a "publisher"; the MQTT broker receives the messages in real time and stores them in the message queue, and forwards the messages to all "subscribers" (i.e. edge servers) who subscribe to the topic; the edge server subscribes to all sensor topics through the gateway and receives multi-sensor data in parallel to ensure that the delay from data collection to delivery to the edge server is controlled within 50ms. To ensure reliability, QoS level 1 configuration is adopted to ensure that messages are delivered at least once and to avoid data loss. At the same time, data integrity is detected by a message verification mechanism (such as CRC32). If the verification fails, the sensor is requested to resend. (3) Network adaptation and fault handling In response to the network instability that may exist in the oil and gas industry, the communication layer supports a local caching mechanism: if the edge server temporarily disconnects from the MQTT agent, the gateway will cache the sensor data to local storage (maximum cache capacity 1GB) and upload it in batches after the network is restored to avoid data interruption; if a single sensor interrupts communication with the gateway, the gateway will immediately send a fault alarm to the edge server. The edge processing layer marks the sensor data as "missing" and uses other sensor data to complement it to ensure that the overall process is not interrupted. In addition, the communication layer adopts encrypted transmission (TLS / SSL) and encrypts the MQTT message with a shared key to prevent the data from being tampered with or stolen during transmission and to ensure the security of the Internet of Things system.

[0090] 3. Edge Field Processing Layer: Used for data preprocessing and localized management. This layer performs data preprocessing and feature extraction, serving as the solution's "data cleaning and feature extraction center." For example, it uses a Raspberry Pi edge server to preprocess raw data, removing noise, standardizing formats, and extracting key features to provide high-quality input for machine learning models. Simultaneously, it enables localized data storage and secure management, avoiding latency caused by cloud reliance. The edge field processing layer includes: 1. Core Preprocessing Flow: Data Acquisition and Reception: The edge server's "data acquisition module" receives raw data from various sensors through the MQ gateway, categorizes and stores it according to sensor type, forming a raw data acquisition set, expressed as: (RGB sensor) (IR sensor) (Gas sensor) (Smoke sensor) (Flame sensor), where "Dac" represents Data Acquisition. This step is completed within 100ms after data reception to ensure rapid initiation of subsequent processing. Security verification and integrity check: The "security verification module" uses Tcryptographic technology and parity check to ensure data reliability. First, HMAC-SHA256 hash verification is performed on the data from each sensor using a shared key. The hash value of the data is calculated and compared with the check value sent by the sensor. If they do not match, the data is considered to have been tampered with, and the data set is discarded. Second, transmission errors are detected through parity check, expressed as follows: , , ..., (CP stands for Cryptographic Validation and Parity Check). If the validation fails, the sensor is requested to resend the data. This step aims to filter unreliable data and ensure the integrity of the input data, with an execution time of no more than 50ms. Data Filtering and Noise Processing: The "Data Filtering Module" processes random noise (such as sensor circuit interference and environmental fluctuations) in the raw data using a third-order moving average filtering algorithm. Taking IR sensor temperature data as an example, the filtering formula is: Weighted averaging reduces the impact of sudden noise on the data; for missing values ​​(such as data gaps caused by temporary sensor malfunctions), the mean of the first three valid data points is used to fill the gaps. If missing, then Ensure data continuity. This step effectively improves data smoothness and takes approximately 100ms to execute. Data normalization: The "data normalization module" standardizes sensor data of different magnitudes and units to the [0,1] range, eliminating the impact of data magnitude differences on machine learning models. Dedicated normalization formulas are designed for different sensor data: the R, G, and B channels of the RGB sensor are normalized separately using the following formulas: ;in, , These represent the minimum and maximum values ​​of the R channel, and the same applies to the G and BT channels; where, The normalization formula for IR sensors is: ; The normalization formula for gas sensors is: ; The normalization formula for smoke sensors is: ; The normalization formula for flame sensors is: ; Here, "DN" represents Data Normalization. Normalized data is more conducive to model training and inference; this step takes approximately 50ms. Feature Extraction: The "Feature Extraction Module" T calculates the statistical characteristics of each sensor's data, including the mean and standard deviation (Std), to reflect the overall trend and dispersion of the data, providing crucial information for anomaly detection and model input. For mean calculation, the RGB sensor mean is the set of the three channel means, expressed as: ; The average value of the IR sensor is: ; The mean values ​​for the gas, smoke, and flame sensors are as follows: ; ; ; Regarding standard deviation calculation, the formula for the standard deviation of an IR sensor is: ; The standard deviation calculation method for other sensors is similar. Abnormal fluctuations in data can be identified through standard deviation; for example, the temperature standard deviation will significantly increase during a fire. This step takes approximately 100ms. Anomaly Detection and Data Conversion: The "Anomaly Detection Module" identifies potential fire anomalies based on preset thresholds and marks the normalized data as abnormal (1 indicates abnormal, 0 indicates normal). Specific thresholds are set according to the characteristics of the oil and gas industry environment and NFPA standards: For RGB sensors, if Ri > 200 and Gi < 100 and Bi < 100 (corresponding to the red-orange hue of a flame), then... If the IR sensor's Ti > 80℃ (critical fire temperature), then If the CO concentration Gi of the gas sensor is greater than 500 ppm, then... If the smoke sensor concentration Si > 250 μg / m³, then If the flame sensor intensity Fi > 50, then Here, "ADC" stands for Anomaly and Abnormal Detection. Finally, the "Data Transformation Module" T serializes the preprocessing results into a structured format ("DTS" stands for Data Transformation and Serialization), facilitating reading and processing by machine learning models. This step takes approximately 100ms, and the total preprocessing time is controlled within 500ms, meeting real-time requirements.

[0091] In addition, localized storage and security management are employed. The edge server is equipped with a local storage module (using an SD card or SSD, with a capacity of no less than 32GB) to store pre-processed sensor data and system logs. The default storage period is 30 days, after which old data is automatically deleted or uploaded to the cloud for archiving. During storage, data is stored in an encrypted format (AES-256) to prevent local data leakage. Simultaneously, the edge server's "security module" restricts unauthorized access through access control (such as username / password authentication and SSH key login) to ensure data security. Furthermore, the "device management module" monitors the operating status of the sensors and edge server in real time, including sensor battery level, communication connection, and data acquisition frequency. If any anomalies are detected (such as sensor battery level below 20% or data acquisition interruption exceeding 10 seconds), an alarm is immediately sent to on-site maintenance personnel to ensure stable system operation.

[0092] 4. Machine Learning Model Layer: Sensor-Specific Model Inference and Probability Prediction. The machine learning model layer is the "core of feature recognition and probability output" of the solution. Relying on the lightweight TensorFlow Lite runtime environment of the edge server, it matches a dedicated machine learning model to each type of sensor data. Through model inference, it extracts fire features and quantifies fire probabilities, providing accurate quantitative basis for subsequent Dempster-Shafer (DST) fusion decisions. The core design idea of ​​this layer is "sensor-model feature adaptation," that is, selecting a model that balances computational efficiency and accuracy based on data type (image, numerical, time series) and feature complexity, and ensuring real-time operation on edge devices through lightweight optimization. (1) Model Selection and Lightweight Optimization Based on the differentiated characteristics of the five types of sensor data, the model selection follows the principle of "functional matching and efficiency priority". The specific selection and optimization measures are as follows: RGB Sensor (Image Data) - MobileNetCNN: The output of the RGB sensor is flame visual image data, which requires the extraction of image features such as color, texture, and edge. Therefore, MobileNet Convolutional Neural Network (CNN) is selected. This model adopts the "depth-wise separable convolution" technology, which splits the traditional convolution into depthwise convolution and pointwise convolution. The computational cost is only 1 / 8 to 1 / 9 of that of the traditional CNN, and the number of parameters is optimized to 1.4M, which can be used for fast inference on edge devices. To further adapt to the edge environment, TensorFlow Lite's INT8 quantization technology is used to convert the model weights from 32-bit floating-point numbers to 8-bit integers, compressing the model size by 75% and increasing the inference speed by 3 times. At the same time, the accuracy loss after quantization is controlled within 2% by calibrating the dataset (containing 100,000 flame / non-flame images of oil and gas scenes). IR Sensor (Thermal Distribution Data) - EfficientNetCNN: IR sensor data is essentially a "thermal image" of temperature distribution. To identify the shape and diffusion trend of abnormal temperature regions, EfficientNetCNN is chosen. This model uses "compound coefficients" to collaboratively optimize network depth, width, and resolution, achieving better accuracy than MobileNet with the same computational cost. For edge devices, a "model pruning" technique is used to remove redundant convolutional kernels (pruning rate 30%), retaining key temperature feature extraction layers. Simultaneously, TensorFlowLite's dynamic shape inference function is used to adapt to thermal image inputs of different resolutions from IR sensors, ensuring inference time remains stable within 200ms. Gas Sensor (Time-Series Concentration Data) - RandomForest: Gas sensor data is time-series gas concentration data. To identify concentration change trends (such as a sudden increase in CO concentration), the Random Forest model is chosen. This model consists of 50 decision trees, using ensemble learning to reduce the risk of overfitting in individual trees, while also possessing strong noise resistance. In terms of optimization, the maximum depth of each tree is limited to 10 and the minimum number of samples per leaf node is 5 to reduce the amount of computation; model parameters (such as tree structure and split threshold) are stored in a lightweight binary format, reducing the loading time to 50ms and ensuring real-time response to time series data.Smoke Sensor (Multi-dimensional Concentration Data) - SVM (RBF Kernel): The smoke sensor needs to process multi-dimensional data on smoke particle concentration and combustible gas content simultaneously, and it needs to distinguish between "normal smoke (such as equipment emissions)" and "fire smoke." Support Vector Machine (SVM) is selected, and an RBF (Radial Basis Function) kernel is configured. The RBF kernel can map low-dimensional data to a high-dimensional space, solving the non-linear classification problem of smoke features. During optimization, the optimal parameters (kernel parameter γ=0.1, penalty coefficient C=10) are determined through grid search to balance classification accuracy and generalization ability. Simultaneously, a "sample sampling" technique is used to undersample massive historical data (50% of non-fire samples are retained), reducing the amount of data for model training and inference, and controlling the inference time to within 150ms. Flame Sensor (Intensity Time Series Data) - Decision Tree: The flame sensor outputs time series values ​​of flame intensity. The features are simple, but a quick determination of "flame presence or absence" is required. The Decision Tree model is selected. This model splits nodes through "information gain," has clear decision logic, low computational cost, and can directly output classification results. In terms of optimization, the minimum information gain of the split threshold is set to 0.1 to avoid overfitting; the decision rules are transformed into logical judgment statements that can be directly executed by the hardware (such as "if Fi>50, then it is judged as abnormal"), and the inference time can be compressed to 50ms, which is the fastest response among the five types of models.

[0093] (2) Model Training and Inference Process: Model training adopts an "offline training + online fine-tuning" mode. Offline training is completed in the cloud, and online fine-tuning is performed on the edge server to ensure that the model adapts to the special characteristics of the oil and gas industry scenario: Offline training stage: Collect historical sensor data of the oil and gas industry scenario (including real fire case data and normal operation data) to build a training dataset (total of 1 million records, with fire samples accounting for 15%). For RGB and IR image data, data augmentation techniques (rotation, flipping, brightness adjustment) are used to expand the samples; for numerical data of gas, smoke, and flame, a "sliding window" is used to generate time-series samples (window size 10, step size 1). During training, MobileNet and EfficientNet use the Adam optimizer (learning rate 1e-4) and the loss function is cross-entropy loss; RandomForest, SVM, and DecisionTree use default optimization parameters, and the hyperparameters are adjusted through 5-fold cross-validation. After training, the model is converted to TensorFlow Lite format (.tflite) and deployed to the edge server. Online fine-tuning phase: The edge server periodically (every 7 days) collects newly generated preprocessed data locally. If new data contains features not seen by the model (such as abnormal concentrations caused by new combustible gas leaks), online fine-tuning is initiated. During fine-tuning, the bottom feature extraction layer of the model is fixed, and only the top fully connected layer (MobileNet / EfficientNet) or decision nodes (RandomForest / DecisionTree) are updated. Mini-batch gradient descent (BatchSize=32, learning rate 1e-5) is used, and the fine-tuning cycle is controlled within 1 hour to avoid affecting the real-time detection function. The inference process follows the "sensor-model one-to-one" principle. The preprocessed data of each type of sensor is only input into the corresponding model. The specific steps are as follows: Data input: The structured data SiDTS (including normalized data, mean, standard deviation, and outlier markers) output by the edge field processing layer is converted according to the model requirements. For example, RGB data needs to have its normalized pixel values ​​reshaped into an image tensor of (224, 224, 3) (MobileNet input size), and gas data needs to have its concentration mean, standard deviation, and time series segments combined into a feature vector (12 dimensions). Feature extraction: The model extracts features from the input data.MobileNet extracts flame edge and color features from RGB images through a stack of convolutional layers, batch normalization layers, and ReLU activation layers; EfficientNet extracts temperature anomaly region features from IR images through composite convolutional blocks; RandomForest extracts trend features of gas concentration (e.g., "CO concentration increases >100ppm for 3 consecutive moments") by splitting nodes in multiple decision trees; SVM transforms smoke data into a classification hyperplane in a high-dimensional feature space through RBF kernel mapping; DecisionTree extracts key judgment features by splitting nodes based on the feature with the greatest information gain (e.g., mean flame intensity). Probability output: The model outputs the probability prediction results of "fire exists" and "nofire does not exist," satisfying Pi(fire) + Pi(nofire) = 1. The specific expressions are as follows: MobileNet inference result: MCNNCPP=MobileNet(S1DTS)=[P1(fire),P1(nofire)], where MCNNCPP represents MobileNet class probability prediction; EfficientNet inference result: ECNNCPP=EfficientNet(S2DTS)=[P2(fire),P2(nofire)]; RandomForest inference result: RFCPP=RandomForest(S3DTS)=[P3(fire),P3(nofire)]; SVM inference result: SVMCPP=SVM(S4DTS)=[P4(fire),P4(nofire)]; DecisionTree inference result: DTCPP=DecisionTree(S5DTS)=[P5(fire),P5(nofire)]. During inference, all models execute in parallel, with the total inference time kept under 1 second (MobileNet: 200ms, EfficientNet: 250ms, RandomForest: 150ms, SVM: 150ms, DecisionTree: 50ms), meeting the real-time requirements of edge environments. Inference results are transmitted to the Dempster-Shafer prediction layer in real time and simultaneously stored in the inference log on the edge server for subsequent model fine-tuning and fault tracing.

[0094] (3) Model monitoring and fault handling To ensure the reliability of model inference, this layer is equipped with a "model monitoring module" to track model performance indicators and operating status in real time: Performance monitoring: Calculate the real-time accuracy of the model (based on manually labeled verification data), false positive rate (the proportion of non-fire samples that are mistakenly identified as fires), and false negative rate (the proportion of fire samples that are missed) every hour. If the accuracy is lower than 95%, the false positive rate is higher than 5%, or the false negative rate is higher than 3%, a model alarm will be triggered immediately, and the best historical model version will be loaded automatically (the edge server stores the three most recent versions of the model) to ensure that the detection function is not interrupted; Operation monitoring: Monitor the resource consumption (CPU usage and memory usage) during model inference. If the CPU usage exceeds 80% or the memory usage exceeds 1GB, the "resource scheduling mechanism" will be started to prioritize the operation of key sensor models such as flame and IR, and temporarily reduce the inference frequency of the RGB model (from 10Hz to 5Hz) and restore it after the resources are released; Fault recovery: If the model file is damaged or the inference is faulty, the edge server will automatically download the latest model from the local backup (stored in an independent partition) or the cloud. The recovery time is controlled within 30 seconds to avoid fire detection interruption due to model failure.

[0095] 5. Dempster-Shafer Prediction Layer: Multi-Evidence Fusion Decision Making and Fire Assessment. The Dempster-Shafer prediction layer is the "decision core" of the solution. It fuses the probability outputs of five types of machine learning models using Dempster-Shafer Theory (DST) to handle data uncertainty and model conflicts, generating the final fire decision result. The core value of this layer lies in its ability to quantify "uncertainty" (i.e., the degree to which the model cannot determine the existence of a fire) compared to traditional "majority voting" or "weighted average" fusion methods. Even in scenarios with conflicting or missing sensor data, or severe environmental interference, it can still output reliable decisions.

[0096] (1) Core Concepts and Preprocessing of DST Fusion Before evidence fusion, the core concepts and data transformation rules of DST need to be clarified: Frame of Discretion (θ): Defines all possible outcomes of the decision. In this scheme, θ = {fire, nofire}, that is, only two mutually exclusive outcomes are considered; Basic Probability Assignment (BPA): Converts the probability output of the machine learning model into a "belief assignment" for each hypothesis, including the belief m(fire) for "fire", the belief m(nofire) for "nofire", and the belief m(θ) for "uncertainty", satisfying m(fire) + m(nofire) + m(θ) = 1; Conflict Coefficient (K): Measures the degree of conflict between different model evidence. The larger the K value, the more serious the evidence conflict. The weight of highly conflicting evidence needs to be reduced through a conflict handling mechanism. First, "belief quality conversion" is performed to convert the probability prediction [Pi(fire), Pi(nofire)] of the machine learning model into BPA. The conversion rule is based on "direct probability mapping + uncertainty completion", and the formula is as follows: ; ; After the conversion, we obtain the BPA set for five types of models: ,in, Corresponding to MobileNet, Corresponding to EfficientNet, Corresponding to RandomForest Corresponding to SVM, Corresponding to DecisionTree.

[0097] (2) Conflict Assessment and Uncertainty Adjustment Due to issues such as sensor interference and model errors in the oil and gas industry environment, the BPA of different models may conflict (e.g., the output of the flame sensor model). The smoke sensor model outputs data due to smoke obstruction. First, conflict assessment and uncertainty adjustment are necessary to avoid highly conflicting evidence misleading decision-making. Conflict assessment (Hellinger distance): The Hellinger distance measures the degree of conflict between any two model BPAs. This distance quantifies the similarity between two probability distributions, with a value ranging from [0,1]. A closer distance indicates more consistent evidence, while a greater distance indicates more severe conflict. The calculation formula is as follows: ; in, The HD represents the belief quality of the i-th model regarding the k-th hypothesis (k=1: fire, k=2: nofire, k=3: θ). If HD > 0.5, it is considered "high-conflict evidence," and the weight of the corresponding model needs to be reduced; if HD ≤ 0.5, it is considered "low-conflict evidence," and the original weight is retained. Uncertainty Adjustment (Deng Entropy): Deng entropy is used to quantify the uncertainty of each model's BPA. The larger the entropy value, the higher the uncertainty of the model, and its weight in the fusion needs to be reduced. The formula for calculating Deng entropy is as follows: ; The model weights are calculated based on Deng entropy, using the following formula: ; in, It is the maximum value of the entropy of all models (in this scheme). Weight ∈[0,1], a larger weight indicates lower model uncertainty and a higher proportion during fusion. Applying the weights to adjust the BPA yields the adjusted BPA: , , The adjustment still meets the requirements. To ensure the effectiveness of BPA. (3) Evidence fusion and final decision evidence fusion adopts Dempster's Rule of Combination. The core of this rule is to merge the adjusted BPA of different models through "intersection calculation" and at the same time eliminate the interference of contradictory evidence through "conflict coefficient" to finally obtain the belief quality after global fusion. The fusion process employs an iterative approach. First, the BPAs of the first two models are merged. Then, the result is merged with the BPA of the third model, and so on, until the fusion of all five models is complete. The specific steps and formulas are as follows: First, the core formula of the Dempster rule is defined. For two adjusted BPAs... and The quality of beliefs after integration (A is any subset of the recognition frame θ, i.e., fire, nofire, or θ) Calculated as follows: ; The numerator represents the sum of the product of the qualities of all beliefs that satisfy "the intersection of subset B and subset C is A", reflecting the accumulation of consistent evidence; the denominator represents... The conflict coefficient between two BPAs is calculated using the following formula: , This represents the sum of the products of contradictory evidence from two models (e.g., when B=fire and C=nofire), with the denominator expressed as " "Normalize consistent evidence to ensure that the sum of the qualities of all beliefs after fusion is 1. If..." If there is a complete conflict, fusion is not performed, and the model BPA, with lower uncertainty (lower entropy), is used as a temporary result. Otherwise, fusion is performed. After fusion, the reliability of the decision needs to be further quantified using the "belief function" and "likelihood function," and then the final fire judgment is output based on preset rules. The belief function represents the "minimum belief quality that supports a hypothesis," which is the sum of the belief qualities of all subsets of that hypothesis; the likelihood function represents the "maximum belief quality that does not oppose a hypothesis," which is the sum of the belief qualities of all subsets whose intersection with that hypothesis is not empty. The calculation formula is as follows: Belief function for fire hypothesis: (Since 'fire' is a singlet subset with no other subsets contained within it), the likelihood function of the fire hypothesis is: (Since the intersection of fire and nofire is empty, only the intersection of fire and θ with fire is not empty) Belief function for the no-fire assumption: The likelihood function without the fire assumption: Based on the belief function and likelihood function, and considering the stringent requirements for "missed detection" in the oil and gas industry environment (prioritizing the avoidance of missed fire detections), the final decision rules are set as follows: If Belief(fire) > 0.5: a "fire exists" condition is determined, and a real-time early warning is immediately triggered (e.g., sending audible and visual alarms to the on-site safety team, pushing SMS notifications, and simultaneously activating on-site fire-fighting equipment); if Belief(nofire) > 0.5: no fire is determined, and monitoring of sensor data and model output continues; if Belief(fire) ≤ 0.5 and Belief(nofire) ≤ 0.5: an "uncertain state" condition is determined, and Plausibility(fire) and Plausibility(nofire) are output. If Plausibility(fire) > 0.7, a "suspected fire" alarm is triggered (requiring manual verification of the on-site situation); otherwise, normal monitoring is maintained. After the decision is made, this layer will... The belief function, likelihood function, and final decision results are stored in the "decision log" on the edge server. At the same time, the log and preprocessed data are uploaded to the central system / cloud layer at a frequency of minutes for subsequent model optimization and historical data analysis.

[0098] 6. Central System / Cloud Layer: Long-term Data Archiving, Model Optimization and Compliance Management. The central system / cloud layer is the "long-term support and optimization center" of the solution. It mainly undertakes three major functions: long-term data storage, global model updates and NFPA standard compliance verification. Through collaboration with the edge layer, it enables continuous iteration and industrialization of the solution, solving the problems of "limited storage and insufficient global optimization capabilities" of the edge layer.

[0099] (1) Long-term data archiving and management The cloud layer uses PostgreSQL open-source relational database as the core storage, and HDFS (Hadoop Distributed File System) is used to store massive amounts of unstructured data (such as raw images and video clips of RGB / IR sensors) to build a hybrid storage architecture of "structured + unstructured". The specific archiving process and management strategy are as follows: In terms of data uploading, the edge server transmits data to the cloud in the mode of "minute-level incremental upload + daily full backup": Every minute, the preprocessed data (SiDTS), model inference results (MCNNCPP, etc.), and DST decision results (mcombined, etc.) of the edge layer are uploaded to PostgreSQL in JSON format. Each data includes key fields such as timestamp, sensor ID, edge node ID, and data content, which facilitates subsequent retrieval. At 2:00 am every day (the off-peak period of the industrial site), the unstructured data such as raw images and video clips stored in the edge layer are compressed and uploaded to HDFS. The directory structure of "edge node ID-date-sensor type" (such as "node01-20250311-rgb") is adopted to ensure that the data is organized in an orderly manner. In terms of data storage strategy, tiered storage is implemented based on data value and access frequency: Hot data (decision logs and inference results from the past 30 days): stored on high-performance disks (SSDs) of PostgreSQL to ensure real-time query response (query latency <1 second) for daily operation and maintenance monitoring; Warm data (preprocessed data from 30 days to 1 year, unstructured data): stored on ordinary disks of HDFS for monthly / quarterly performance analysis; Cold data (historical data from more than 1 year): stored on low-cost object storage (such as AWS S3 compatible storage) for compliance auditing and long-term trend analysis, with a default storage period of 5 years, in line with NFPA's requirement that industrial fire data be "retained for at least 3 years". In terms of data security, the cloud layer ensures data reliability through triple protection: First, transmission encryption: the edge layer and the cloud use the HTTPS protocol to transmit data, and key fields (such as sensor IDs and decision results) are encrypted using AES-256; second, storage encryption: the PostgreSQL database uses Transparent Data Encryption (TDE), and HDFS files use block-level encryption; third, access control: permissions are assigned based on the RBAC (Role-Based Access Control) model, allowing only operations and maintenance personnel and security management personnel to access the corresponding data, and operation logs are audited throughout the process to prevent unauthorized data tampering or leakage.

[0100] (2) Federated Learning and Global Model Optimization: To address the issues of "data silos" and "insufficient model generalization ability" at the edge layer, the cloud layer employs federated learning technology. This achieves collaborative optimization of the global model without collecting raw data from the edge layer, avoiding privacy leaks and high data transmission costs associated with traditional centralized training. The specific process is as follows: First, the cloud layer initializes the global machine learning model (with the same structure as the edge layer model, such as MobileNet or RandomForest), and distributes the initial model parameters (such as the convolutional kernel weights of MobileNet and the tree structure of RandomForest) to all edge nodes (such as the edge servers of multiple gas stations in an oil and gas field). After the edge nodes complete model fine-tuning locally (every 7 days, based on new local data), they calculate the update increment of the model parameters (such as weight differences). ,in, These are the parameters after local fine-tuning. The updated global parameters (from the previous round) are encrypted and uploaded to the cloud layer. The cloud layer's "Federated Learning Aggregation Module" uses a "weighted average" strategy to aggregate the parameter increments of all edge nodes. The weights are determined by the proportion of data volume of each edge node (nodes with larger data volumes have higher weights to avoid interference from small sample nodes on the global model). The calculation formula is as follows: ; Where K is the number of edge nodes. Let be the local data volume of the k-th node. This represents the parameter increment for the k-th node. Aggregation yields... Then, the cloud layer updates the global model parameters: ,in, This is the learning rate (default 0.01), used to control the update step size and avoid model oscillations. After the update is complete, the cloud layer will... The new global parameters are distributed to all edge nodes, which then replace their local model parameters with the new global parameters, completing one round of federated learning iteration. The iteration cycle is set to once a month to balance model optimization frequency with cloud resource consumption. After three rounds of federated learning in simulated real-world production scenarios, the global model's average accuracy improved by 3.5% across different oil and gas scenarios, demonstrating significantly enhanced generalization ability. Furthermore, the cloud layer periodically (quarterly) performs a "performance evaluation" of the global model, calculating its accuracy, precision, and recall using a cross-edge node validation dataset (including fire / non-fire samples from each node). If performance degrades by more than 5%, a "global model retraining" is triggered, retraining the model based on historical data archived in the cloud (after de-identification processing) to ensure the global model remains in optimal condition.

[0101] (3) NFPA Standard Compliance Verification and Safety Management The oil and gas industry is a high-risk fire sector and must strictly adhere to the series of standards issued by NFPA (National Fire Protection Association). The cloud layer ensures that the solution meets the requirements of key standards such as NFPA30 (Flammable Liquid Storage), NFPA70 (Electrical Safety), NFPA72 (Fire Alarm Systems), and NFPA13 (Automatic Sprinkler Systems) through the "Compliance Verification Module". The specific verification process and measures are as follows: In terms of compliance data collection, the cloud layer extracts key parameters of concern to the NFPA standard from the edge decision log, such as: NFPA30 related... : Temperature of flammable liquid storage area (IR sensor data), combustible gas concentration (MQ-135 data), fire alarm response time (time from sensor acquisition to decision output); NFPA70 related: Number of temperature anomalies around electrical equipment (IR sensor anomaly detection results), false alarm rate of fire related to electrical system; NFPA72 related: Accuracy of fire alarm, trigger delay of audible and visual alarms, time for alarm information to be transmitted to the fire department; NFPA13 related: Response time of automatic sprinkler system linkage (time from DST decision of "fire exists" to activation of fire-fighting equipment). Regarding compliance verification, the cloud layer operates on a "daily automatic verification + monthly manual review" model: daily automatic comparison of key parameters with NFPA standard thresholds (e.g., NFPA30 requires immediate warning if the temperature in the flammable liquid storage area exceeds 80°C, and NFPA72 requires alarm triggering delay <10 seconds), generating a "compliance verification report". If a parameter fails to meet the standard for three consecutive days (e.g., alarm delay >12 seconds), a "compliance risk alarm" is sent to the operations team, and the cause is automatically analyzed (e.g., excessive network latency at edge nodes, or failure of fire equipment linkage). Monthly, safety management personnel manually review the compliance verification report, combine it with on-site inspection records, and form a "monthly compliance assessment report", which is submitted to the enterprise's safety management department and the local fire regulatory agency to meet compliance audit requirements. In terms of security management, in addition to data security measures, the cloud layer also has a "global alarm monitoring" function: it receives real-time device fault alarms (such as sensor offline, edge server CPU overload), model performance alarms (such as decreased inference accuracy), and compliance risk alarms from all edge nodes, and displays them in a visual way on the cloud dashboard (such as alarm heatmaps and fault statistics charts). It supports maintenance personnel to locate faulty nodes with one click and remotely issue repair commands (such as restarting edge servers and reloading models). The simulated actual production cloud alarm response time is <30 seconds, and the average fault repair time is <10 minutes, which significantly improves the system operation and maintenance efficiency.

[0102] The technical solution of this embodiment collects multi-dimensional raw environmental data by deploying multiple sensors in an industrial environment. After preprocessing the data to extract target features, a detection result containing the predicted probabilities of fire and non-fire is obtained through a dedicated network model for each sensor. The predicted probabilities are then converted into a first belief quality set containing three terms: fire, non-fire, and uncertainty. Next, the degree of conflict between each pair of belief sets is calculated, and the high-conflict sets are adaptively adjusted to obtain a second belief quality set. Subsequently, the sets are iteratively fused in pairs in sequence. Through cross-multiplication, classification accumulation, and conflict normalization, a global target belief quality set is obtained. Finally, the confidence and likelihood corresponding to fire and non-fire are calculated based on this set, and the final fire detection result is determined comprehensively. The overall solution fully utilizes the complementary advantages of multi-source sensor information, completely preserves uncertainty information through belief transformation, effectively suppresses interference from abnormal and contradictory data through conflict detection and adaptive adjustment, and improves the consistency of judgment through rigorous evidence fusion. It has outstanding advantages such as strong anti-interference, high detection accuracy, low false alarm and false negative rates, and adaptability to complex industrial scenarios.

[0103] Example 2 Figure 2 This is a flowchart of an industrial environment fire detection method provided in Embodiment 2 of the present invention. The method in this embodiment is a further optimization of the method in the above embodiments. Optionally, for every two first belief quality sets, if the degree of conflict between these two first belief quality sets is greater than a preset conflict degree, then the uncertainty corresponding to each of the two first belief quality sets is determined, and the first belief quality sets are adjusted based on the uncertainty to obtain an adjusted second belief quality set; if the degree of conflict between these two first belief quality sets is less than or equal to the preset conflict degree, then each of the two first belief quality sets is directly determined as the second belief quality set. Figure 2 As shown, the method includes: S210. Acquire multi-dimensional environmental data collected by various sensors deployed in an industrial environment, including the raw environmental data collected by each sensor.

[0104] S220. Perform data processing on each type of raw environmental data to obtain target data feature information corresponding to each sensor, and perform fire detection based on the network model corresponding to each sensor and the target data feature information to obtain the fire detection results output by each network model.

[0105] S230. Perform belief quality transformation on the fire detection results output by each network model to obtain the first belief quality set corresponding to each network model. The first belief quality set includes: belief quality for the existence of fire, belief quality for the non-existence of fire, and belief quality for uncertainty.

[0106] S240, Determine the degree of conflict between each pair of first belief quality sets.

[0107] S250. For every two sets of first belief quality, if the degree of conflict between the two sets of first belief quality is greater than the preset degree of conflict, then determine the uncertainty corresponding to each set of first belief quality in the two sets of first belief quality, and adjust the first belief quality set based on the uncertainty to obtain the adjusted second belief quality set.

[0108] The preset conflict level refers to the degree of contradiction between two sets of first belief quality sets that is used to determine whether the degree of contradiction exceeds the acceptable range. It can be a conflict threshold that is preset according to the actual fire detection scenario and the needs of multi-sensor fusion. When the calculated conflict level is greater than the preset conflict level, the system determines that the evidence conflict is too high and initiates adaptive adjustment. This distinguishes between normal detection differences and abnormal conflicts, provides a unified and clear triggering standard for the correction of belief quality, and ensures the stability of multi-source information fusion and the reliability of decision-making.

[0109] Specifically, the pre-built conflict degree calculation method is called to calculate the conflict degree between each pair of first belief quality sets. When the conflict degree is greater than the preset conflict degree, the uncertainty corresponding to the two belief quality sets is calculated respectively. Then, the original first belief quality sets are adaptively adjusted according to the uncertainty to obtain a more coordinated second belief quality set with less conflict.

[0110] In this embodiment, an adaptive adjustment strategy is constructed to adjust the original first belief quality set to obtain a more coordinated and less conflicting second belief quality set. This effectively identifies and processes highly conflicting multi-source detection information. By introducing uncertainty to reasonably correct the original belief quality, the distortion problem caused by direct fusion of highly conflicting evidence is avoided, while retaining the effective detection information of each sensor. This reduces the interference of abnormal or contradictory data on the final fire judgment, significantly improving the stability and reliability of evidence fusion. The resulting target belief quality set is more in line with the real environmental state, effectively improving the adaptability and judgment accuracy of the fire detection system in complex interference environments.

[0111] Optionally, the uncertainty corresponding to each of the two sets of first belief quality is determined, and the first belief quality sets are adjusted based on the uncertainty to obtain the adjusted second belief quality set. This includes: determining the Deng entropy corresponding to each of the two sets of first belief quality and determining the Deng entropy as the corresponding uncertainty; determining the target first belief quality set whose uncertainty is greater than a preset uncertainty among the two sets of first belief quality; determining the adjustment weight corresponding to the target first belief quality set based on the uncertainty corresponding to the target first belief quality set; and adjusting the target first belief quality set based on the adjustment weight to obtain the adjusted second belief quality set.

[0112] Specifically, Deng's entropy can be understood as an index used to measure the disorder, ambiguity, and uncertainty of information in the belief quality set, objectively reflecting the reliability of sensor detection results. The uncertainty volume can be understood as a value determined by Deng's entropy, representing the level of information ambiguity and conflict in the corresponding belief quality set. The target first belief quality set can be understood as the belief quality set whose uncertainty exceeds a preset uncertainty, has lower information reliability, and higher conflict among two sets of evidence. The adjusted weight volume can be understood as a correction coefficient assigned based on the magnitude of the uncertainty of this target set, used to reasonably weaken or correct it, ultimately obtaining a more reliable second belief quality set.

[0113] Specifically, by calculating the Deng entropy of each of the two first belief quality sets and using it as the corresponding uncertainty, a target first belief quality set with an uncertainty greater than a preset uncertainty is selected. Then, corresponding adjustment weights are assigned according to the magnitude of their uncertainty, and finally, the target set is corrected according to these weights, thereby obtaining a second belief quality set with less conflict and stronger consistency.

[0114] It should be noted that if the uncertainty of only one of the two first belief quality sets is greater than the preset uncertainty, then only the first belief quality set with the uncertainty greater than the preset uncertainty is adjusted to obtain the corresponding second belief quality set. The other first belief quality set is not adjusted, and the unadjusted first belief quality set is directly determined as the second belief quality set.

[0115] In this embodiment, using Deng entropy to characterize uncertainty can more accurately measure the ambiguity and disorder in the belief distribution. Combined with preset thresholds and dynamic weight adjustments, the impact of high uncertainty and high conflict evidence can be specifically weakened. This retains the effective information of reliable evidence with low uncertainty, while making the adjustment process objective, quantitative, and logically rigorous, effectively improving the stability and accuracy of the multi-model fire detection fusion results.

[0116] Optionally, based on the uncertainty corresponding to the target first belief quality set, the adjustment weight corresponding to the target first belief quality set is determined, including: dividing the uncertainty corresponding to the target first belief quality set by the maximum uncertainty, and subtracting 1 from the division result, and using the difference as the adjustment weight corresponding to the target first belief quality set; wherein, the maximum uncertainty refers to the maximum value among the uncertainties corresponding to all first belief quality sets.

[0117] Specifically, the maximum value is found among the uncertainties of all first belief quality sets as the maximum uncertainty. Then, the uncertainty of the target first belief quality set itself is divided by this maximum value. Finally, this ratio is subtracted from 1, and the difference is used as its corresponding adjustment weight.

[0118] In this embodiment, the method of normalizing based on the global maximum uncertainty and constructing weights in reverse is simple and intuitive to calculate, with clear physical meaning. It allows unreliable evidence with higher uncertainty to receive smaller adjustment weights, achieving adaptive weakening of highly conflicting and highly ambiguous information. This not only ensures the fairness and rationality of weight allocation, but also effectively improves the stability of subsequent fusion results and the reliability of fire detection.

[0119] Optionally, based on the adjustment weights, the target first belief quality set is adjusted to obtain an adjusted second belief quality set, including: multiplying the belief quality of the existence of fire in the target first belief quality set by the adjustment weights, and using the multiplication result as the belief quality of the existence of fire in the adjusted second belief quality set; multiplying the belief quality of the non-existence of fire in the target first belief quality set by the adjustment weights, and using the multiplication result as the belief quality of the non-existence of fire in the adjusted second belief quality set; and subtracting 1 from the adjusted belief quality of the existence of fire and the belief quality of the non-existence of fire, and determining the difference as the belief quality of uncertain fire in the adjusted second belief quality set.

[0120] Specifically, the first set of target belief quality is adjusted item by item according to the adjustment weights: First, the belief quality of "there is a fire" and "there is no fire" is multiplied by the weights respectively to obtain the corresponding adjusted values. Then, 1 is subtracted from these two adjusted values, and the difference is used as the adjusted uncertainty belief quality, thereby generating the second set of belief quality.

[0121] In this embodiment, the fire and non-fire belief quality are weighted and corrected by uniform weights, and then normalized uncertainty is automatically generated. This not only accurately weakens the interference of high-conflict and high-uncertainty information based on the reliability of evidence, but also strictly ensures that the sum of belief quality is 1, meeting the requirements of mathematical norms and subsequent fusion. At the same time, the calculation is simple and efficient, which greatly improves the consistency of multi-source evidence while retaining effective detection information, making the adjusted second belief quality set more reliable and laying a stable foundation for subsequent fusion and fire determination.

[0122] S260. If the degree of conflict between the two first belief quality sets is less than or equal to the preset degree of conflict, then each of the two first belief quality sets is directly determined as the second belief quality set.

[0123] Specifically, when the degree of conflict between the two sets of first belief quality is no greater than the preset degree of conflict, it indicates that the consistency of the sensor detection results is good and the contradiction of evidence is within an acceptable range. In this case, there is no need to make additional corrections to the belief quality, and the two sets of original first belief quality can be directly used as the adjusted second belief quality sets.

[0124] In this embodiment, the processing method is logically simple and efficient. In low-conflict scenarios, it eliminates unnecessary calculation and adjustment steps, retains the original reliable information of the sensor, improves the efficiency of the overall fusion process, and ensures the accuracy and real-time performance of subsequent fire detection results.

[0125] S270. Perform fusion processing on the second belief quality set corresponding to each network model to obtain the fused target belief quality set.

[0126] S280. Based on the target belief quality set, determine the first confidence level and first likelihood corresponding to the existence of a fire, and the second confidence level and second likelihood corresponding to the absence of a fire. Based on the first confidence level, first likelihood, second confidence level and second likelihood, determine the target fire detection result.

[0127] The technical solution of this embodiment collects multi-dimensional raw environmental data from multiple sensors in an industrial environment, processes the data to obtain target data features, and then uses the network model corresponding to each sensor to perform fire detection. The model output is transformed into a first set of belief quality containing three categories of belief quality: fire, non-fire, and uncertainty. Then, the degree of conflict between any two sets of first belief quality is calculated. If the conflict is greater than a preset threshold, the weight is calculated using uncertainty (Dun entropy) and adjusted to obtain a second set of belief quality. If the conflict is within the threshold, the original set is directly retained. Then, all second set of belief quality are iteratively fused to obtain the target set of belief quality. Finally, the confidence and likelihood are calculated based on this set, and the final fire detection result is determined according to the threshold rule. The overall solution achieves multi-sensor data complementarity, conflict adaptive correction, and evidence-based hierarchical decision-making. It can solve the problems of "missed and false alarms" through multi-dimensional perception, "real-time" through edge intelligence, and "reliability" through uncertainty handling. It can eliminate monitoring blind spots and suppress interference from abnormal sensors, and improve the reliability of judgment through belief fusion. It has high accuracy, strong robustness, and low false alarm rate, achieving a generational leap in the accuracy, efficiency, and adaptability of industrial fire monitoring. It can effectively adapt to the needs of accurate and real-time fire early warning in complex industrial environments, and is especially suitable for high-risk scenarios such as petroleum, chemical, and metallurgy. Its technological advantages are irreplaceable.

[0128] Example 3 Figure 3 This is a schematic diagram of an industrial environment fire detection device provided in Embodiment 3 of the present invention. Figure 3 As shown, the device includes: The raw environmental data acquisition module 310 is used to acquire multi-dimensional environmental data collected by various sensors deployed in an industrial environment. The multi-dimensional environmental data includes the raw environmental data collected by each sensor. The fire detection result acquisition module 320 is used to process each type of raw environmental data, obtain the target data feature information corresponding to each sensor, and perform fire detection based on the network model and target data feature information corresponding to each sensor, and obtain the fire detection result output by each network model. The first belief quality set acquisition module 330 is used to convert the belief quality of the fire detection results output by each network model to obtain the first belief quality set corresponding to each network model. The first belief quality set includes: belief quality for the existence of fire, belief quality for the absence of fire, and belief quality for uncertainty. The second belief quality set acquisition module 340 is used to determine the degree of conflict between every two first belief quality sets, and adjust the first belief quality sets based on the degree of conflict to obtain the adjusted second belief quality set. The target belief quality set acquisition module 350 is used to fuse the second belief quality set corresponding to each network model to obtain the fused target belief quality set. The target fire detection result determination module 360 ​​is used to determine the first confidence level and first likelihood corresponding to the existence of a fire and the second confidence level and second likelihood corresponding to the absence of a fire based on the target belief quality set, and to determine the target fire detection result based on the first confidence level, first likelihood, second confidence level and second likelihood.

[0129] The technical solution of this embodiment acquires multi-dimensional environmental data collected by various sensors deployed in an industrial environment through a raw environmental data acquisition module. This multi-dimensional environmental data includes raw environmental data collected by each sensor. A fire detection result acquisition module processes each type of raw environmental data to obtain target data feature information corresponding to each sensor. Based on the network model corresponding to each sensor and the target data feature information, fire detection is performed to obtain the fire detection result output by each network model. A first belief quality set acquisition module converts the belief quality of the fire detection result output by each network model to obtain a first belief quality set corresponding to each network model. The first belief quality set includes: the belief quality regarding the existence of a fire. The first belief quality set acquisition module determines the degree of conflict between each pair of first belief quality sets and adjusts the first belief quality sets based on the degree of conflict to obtain the adjusted second belief quality set. The target belief quality set acquisition module fuses the second belief quality sets corresponding to each network model to obtain the fused target belief quality set. The target fire detection result determination module determines the first confidence level and first likelihood corresponding to the existence of a fire, and the second confidence level and second likelihood corresponding to the absence of a fire, based on the target belief quality set, and determines the target fire detection result based on the first confidence level, first likelihood, second confidence level, and second likelihood. Intelligent industrial fire detection is achieved through a modular architecture. The system sequentially collects multi-sensor, multi-dimensional data through a raw environmental data acquisition module. A fire detection result acquisition module processes features and performs model inference and discrimination. A first belief quality set acquisition module transforms the prediction results into a quantitative belief distribution including fire, non-fire, and uncertainty metrics. A second belief quality set acquisition module adaptively corrects the data based on the degree of evidence conflict to obtain a reliable set. A target belief quality set acquisition module iteratively fuses multi-source beliefs to generate a globally unified result. Finally, a target fire detection result determination module outputs the final judgment based on trust and likelihood. The modules within the overall device have clear division of labor and rigorous logic. The modular design facilitates maintenance and expansion. Through multi-sensor information complementarity, conflict adaptive adjustment, and evidence inference fusion, it fully retains and optimizes various detection information, effectively suppressing abnormal interference and data contradictions, significantly improving the accuracy, stability, and robustness of fire detection, and adapting to the real-time, reliable early warning requirements of complex industrial environments.

[0130] Based on the above embodiments, optionally, multiple sensors include: a camera sensor, an infrared sensor, a gas sensor, a smoke sensor, and a flame sensor; multi-dimensional environmental data includes: color image data, ambient temperature data, gas concentration data, smoke concentration data, and infrared radiation intensity data; the network model corresponding to the camera sensor is: a convolutional network model based on a depthwise separable convolutional structure; the network model corresponding to the infrared sensor is: a convolutional network model based on a composite coefficient optimization strategy; the network model corresponding to the gas sensor is: a random forest model; the network model corresponding to the smoke sensor is: a support vector machine model; and the network model corresponding to the flame sensor is: a decision tree model.

[0131] Optionally, the fire detection result acquisition module 320 is specifically used to preprocess the raw environmental data collected by each sensor to obtain preprocessed target environmental data; extract statistical features from the target environmental data to obtain the data statistical feature information corresponding to the sensor; identify the data status based on the target environmental data to obtain the data status information corresponding to the sensor, the data status information being used to characterize whether there is abnormal fire data in the target environmental data; and obtain the target data feature information corresponding to the sensor based on the target environmental data, data statistical feature information, and data status information.

[0132] Optionally, the fire detection result includes a first predicted probability of the existence of a fire and a second predicted probability of the absence of a fire; the first belief quality set acquisition module 330 is specifically used to, for each network model output fire detection result, determine the first predicted probability in the fire detection result as the belief quality of the existence of a fire in the first belief quality set corresponding to the network model; determine the second predicted probability in the fire detection result as the belief quality of the absence of a fire in the first belief quality set corresponding to the network model; and determine the difference between 1 and the belief quality of the existence of a fire and the belief quality of the absence of a fire as the belief quality of the uncertain fire in the first belief quality set corresponding to the network model.

[0133] Optionally, the second belief quality set acquisition module 340 includes a conflict degree determination unit and a second belief quality set determination unit. The conflict degree determination unit is used to determine the Hellinger distance between every two first belief quality sets and to determine the Hellinger distance as the corresponding conflict degree. The second belief quality set determination unit is used to, for every two first belief quality sets, if the conflict degree between the two first belief quality sets is greater than a preset conflict degree, determine the uncertainty corresponding to each of the two first belief quality sets, and adjust the first belief quality sets based on the uncertainty to obtain an adjusted second belief quality set; if the conflict degree between the two first belief quality sets is less than or equal to the preset conflict degree, then directly determine each of the two first belief quality sets as the second belief quality set.

[0134] Optionally, the second belief quality set determination unit is specifically used to determine the Deng entropy corresponding to each of the two first belief quality sets, and to determine the Deng entropy as the corresponding uncertainty; to determine the target first belief quality set whose uncertainty is greater than the preset uncertainty among the two first belief quality sets; to determine the adjustment weight corresponding to the target first belief quality set based on the uncertainty corresponding to the target first belief quality set; and to adjust the target first belief quality set based on the adjustment weight to obtain the adjusted second belief quality set.

[0135] Optionally, the second belief quality set determination unit is specifically used to divide the uncertainty corresponding to the target first belief quality set by the maximum uncertainty, and subtract the division result from 1, and use the difference as the adjustment weight corresponding to the target first belief quality set; wherein, the maximum uncertainty refers to the maximum value among the uncertainties corresponding to all first belief quality sets.

[0136] Optionally, the second belief quality set determination unit is specifically used to multiply the belief quality of the existence of fire in the target first belief quality set by an adjustment weight, and the resulting multiplication is used as the belief quality of the existence of fire in the adjusted second belief quality set; multiply the belief quality of the non-existence of fire in the target first belief quality set by an adjustment weight, and the resulting multiplication is used as the belief quality of the non-existence of fire in the adjusted second belief quality set; and subtract 1 from the adjusted belief quality of the existence of fire and the belief quality of the non-existence of fire, and the difference is determined as the belief quality of uncertain fire in the adjusted second belief quality set.

[0137] Optionally, the target belief quality set acquisition module 350 is specifically used to, according to the fusion order of all second belief quality sets, first fuse the first two second belief quality sets, then fuse the fused belief quality set with the third second belief quality set, until the target belief quality set after fusion with the last second belief quality set is obtained. It is also specifically used to perform cross-multiplication of all combinations in the identification framework based on two second belief quality sets, and to classify and add each multiplication result to determine the added belief quality for the existence of fire, the belief quality for the absence of fire, the belief quality for uncertain fire, and the conflict coefficient; based on the conflict coefficient, the added belief quality for the existence of fire, the belief quality for the absence of fire, and the belief quality for uncertain fire are normalized, and the normalized belief quality for the existence of fire, the belief quality for the absence of fire, and the belief quality for uncertain fire are used as the fused belief quality set.

[0138] Optionally, the target belief quality set acquisition module 350 is also specifically used to determine the second belief quality set with the smallest uncertainty among the two second belief quality sets as the fused belief quality set if the conflict coefficient is 1.

[0139] Optionally, the target fire detection result determination module 360 ​​is specifically used to determine the belief quality of the existence of a fire in the target belief quality set as the first confidence level corresponding to the existence of a fire; to add the belief quality of the existence of a fire and the belief quality of uncertain fire in the target belief quality set, and use the sum as the first likelihood corresponding to the existence of a fire; to determine the belief quality of the non-existence of a fire in the target belief quality set as the second confidence level corresponding to the non-existence of a fire; and to add the belief quality of the non-existence of a fire and the belief quality of uncertain fire in the target belief quality set, and use the sum as the second likelihood corresponding to the non-existence of a fire.

[0140] Optionally, the target fire detection result determination module 360 ​​is specifically used to determine that the target fire detection result is that a fire exists if the first confidence level is greater than the preset confidence level; to determine that the target fire detection result is that a fire does not exist if the second confidence level is greater than the preset confidence level; and to determine that the target fire detection result is in an uncertain state if the first confidence level is less than or equal to the preset confidence level and the second confidence level is less than or equal to the preset confidence level, outputting the first likelihood and the second likelihood, and triggering a suspected fire alarm when the first likelihood is greater than the preset likelihood.

[0141] Optionally, each sensor is deployed at key risk locations in the industrial environment to form a complementary and seamless environmental perception network. Each sensor continuously collects multi-dimensional environmental data based on a preset acquisition frequency and transmits the continuously collected multi-dimensional environmental data to the edge server through a publish-subscribe message transmission protocol.

[0142] The industrial environment fire detection device provided in this embodiment of the invention can execute the industrial environment fire detection method provided in any embodiment of the invention, and has the corresponding functional modules and beneficial effects of the method.

[0143] Example 4 Figure 4 A schematic diagram of an electronic device 10, which can be used to implement embodiments of the present invention, is shown. The electronic device is intended to represent various forms of digital computers, such as laptop computers, desktop computers, workstations, personal digital assistants, servers, blade servers, mainframe computers, and other suitable computers. The electronic device can also represent various forms of mobile devices, such as personal digital processors, cellular phones, smartphones, wearable devices (e.g., helmets, glasses, watches, etc.), and other similar computing devices. The components shown herein, their connections and relationships, and their functions are merely illustrative and are not intended to limit the implementation of the invention described and / or claimed herein.

[0144] like Figure 4 As shown, the electronic device 10 includes at least one processor 11 and a memory, such as a read-only memory (ROM) 12 or a random access memory (RAM) 13, communicatively connected to the at least one processor 11. The memory stores computer programs executable by the at least one processor. The processor 11 can perform various appropriate actions and processes based on the computer program stored in the ROM 12 or loaded from storage unit 18 into the RAM 13. The RAM 13 can also store various programs and data required for the operation of the electronic device 10. The processor 11, ROM 12, and RAM 13 are interconnected via a bus 14. An input / output (I / O) interface 15 is also connected to the bus 14.

[0145] Multiple components in electronic device 10 are connected to I / O interface 15, including: input unit 16, such as keyboard, mouse, etc.; output unit 17, such as various types of displays, speakers, etc.; storage unit 18, such as disk, optical disk, etc.; and communication unit 19, such as network card, modem, wireless transceiver, etc. Communication unit 19 allows electronic device 10 to exchange information / data with other devices through computer networks such as the Internet and / or various telecommunications networks.

[0146] Processor 11 can be a variety of general-purpose and / or special-purpose processing components with processing and computing capabilities. Some examples of processor 11 include, but are not limited to, central processing unit (CPU), graphics processing unit (GPU), various special-purpose artificial intelligence (AI) computing chips, various processors running machine learning model algorithms, digital signal processors (DSPs), and any suitable processor, controller, microcontroller, etc. Processor 11 performs the various methods and processes described above, such as industrial environment fire detection methods.

[0147] In some embodiments, the industrial environment fire detection method may be implemented as a computer program tangibly contained in a computer-readable storage medium, such as storage unit 18. In some embodiments, part or all of the computer program may be loaded and / or installed on electronic device 10 via ROM 12 and / or communication unit 19. When the computer program is loaded into RAM 13 and executed by processor 11, one or more steps of the industrial environment fire detection method described above may be performed. Alternatively, in other embodiments, processor 11 may be configured to perform the industrial environment fire detection method by any other suitable means (e.g., by means of firmware).

[0148] Various embodiments of the systems and techniques described above herein can be implemented in digital electronic circuit systems, integrated circuit systems, field-programmable gate arrays (FPGAs), application-specific integrated circuits (ASICs), application-specific standard products (ASSPs), systems-on-a-chip (SoCs), payload-programmable logic devices (CPLDs), computer hardware, firmware, software, and / or combinations thereof. These various embodiments may include implementations in one or more computer programs that can be executed and / or interpreted on a programmable system including at least one programmable processor, which may be a dedicated or general-purpose programmable processor, capable of receiving data and instructions from a storage system, at least one input device, and at least one output device, and transmitting data and instructions to the storage system, the at least one input device, and the at least one output device.

[0149] Computer programs used to implement the methods of the present invention may be written in any combination of one or more programming languages. These computer programs may be provided to a processor of a general-purpose computer, a special-purpose computer, or other programmable data processing device, such that when executed by the processor, the computer programs cause the functions / operations specified in the flowcharts and / or block diagrams to be performed. The computer programs may be executed entirely on a machine, partially on a machine, or as a standalone software package, partially on a machine and partially on a remote machine, or entirely on a remote machine or server.

[0150] In the context of this invention, a computer-readable storage medium can be a tangible medium that may contain or store a computer program for use by or in conjunction with an instruction execution system, apparatus, or device. A computer-readable storage medium may include, but is not limited to, electronic, magnetic, optical, electromagnetic, infrared, or semiconductor systems, apparatus, or devices, or any suitable combination thereof. Alternatively, a computer-readable storage medium may be a machine-readable signal medium. More specific examples of machine-readable storage media include electrical connections based on one or more wires, portable computer disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fibers, portable compact disk read-only memory (CD-ROM), optical storage devices, magnetic storage devices, or any suitable combination thereof.

[0151] To provide interaction with a user, the systems and techniques described herein can be implemented on an electronic device having: a display device (e.g., a CRT (cathode ray tube) or LCD (liquid crystal display) monitor) for displaying information to the user; and a keyboard and pointing device (e.g., a mouse or trackball) through which the user provides input to the electronic device. Other types of devices can also be used to provide interaction with the user; for example, feedback provided to the user can be any form of sensory feedback (e.g., visual feedback, auditory feedback, or tactile feedback); and input from the user can be received in any form (including sound input, voice input, or tactile input).

[0152] The systems and technologies described herein can be implemented in computing systems that include backend components (e.g., as data servers), or middleware components (e.g., application servers), or frontend components (e.g., user computers with graphical user interfaces or web browsers through which users can interact with implementations of the systems and technologies described herein), or any combination of such backend, middleware, or frontend components. The components of the system can be interconnected via digital data communication of any form or medium (e.g., communication networks). Examples of communication networks include local area networks (LANs), wide area networks (WANs), blockchain networks, and the Internet.

[0153] A computing system can include clients and servers. Clients and servers are generally located far apart and typically interact through communication networks. The client-server relationship is created by computer programs running on the respective computers and having a client-server relationship with each other. The server can be a cloud server, also known as a cloud computing server or cloud host, which is a hosting product within the cloud computing service system to address the shortcomings of traditional physical hosts and VPS services, such as high management difficulty and weak business scalability.

[0154] In the context of this invention, a computer program product includes a computer program that, when executed by a processor, implements the industrial environment fire detection method of any embodiment of this invention.

[0155] In the implementation of a computer program product, computer program code for performing the operations of this invention can be written in one or more programming languages ​​or a combination thereof. Programming languages ​​include object-oriented programming languages ​​as well as conventional procedural programming languages. The program code can be executed entirely on the user's computer, partially on the user's computer, as a standalone software package, partially on the user's computer and partially on a remote computer, or entirely on a remote computer or server. In cases involving remote computers, the remote computer can be connected to the user's computer via any type of network—including a local area network (LAN) or a wide area network (WAN)—or can be connected to an external computer (e.g., via the Internet using an Internet service provider).

[0156] It should be understood that the various forms of processes shown above can be used, with steps reordered, added, or deleted. For example, the steps described in this invention can be executed in parallel, sequentially, or in different orders, as long as the desired result of the technical solution of this invention can be achieved, and this is not limited herein.

[0157] The specific embodiments described above do not constitute a limitation on the scope of protection of this invention. Those skilled in the art should understand that various modifications, combinations, sub-combinations, and substitutions can be made according to design requirements and other factors. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of this invention should be included within the scope of protection of this invention.

Claims

1. A method for detecting fires in an industrial environment, characterized in that, include: Acquire multi-dimensional environmental data collected by various sensors deployed in an industrial environment, including raw environmental data collected by each sensor. Data processing is performed on each type of raw environmental data to obtain target data feature information corresponding to each sensor. Fire detection is then performed based on the network model corresponding to each sensor and the target data feature information to obtain the fire detection results output by each network model. The fire detection results output by each network model are transformed into belief quality to obtain the first belief quality set corresponding to each network model. The first belief quality set includes: belief quality for the existence of fire, belief quality for the non-existence of fire, and belief quality for uncertainty. Determine the degree of conflict between every two first belief quality sets, and adjust the first belief quality sets based on the degree of conflict to obtain the adjusted second belief quality sets; The second belief quality set corresponding to each network model is fused to obtain the fused target belief quality set. Based on the target belief quality set, a first confidence level and a first likelihood level corresponding to the presence of a fire are determined, and a second confidence level and a second likelihood level corresponding to the absence of a fire are determined. Based on the first confidence level, the first likelihood level, the second confidence level, and the second likelihood level, the target fire detection result is determined.

2. The method according to claim 1, characterized in that, The various sensors include: camera sensors, infrared sensors, gas sensors, smoke sensors, and flame sensors; The multi-dimensional environmental data includes: color image data, ambient temperature data, gas concentration data, smoke concentration data, and infrared radiation intensity data; The network model corresponding to the camera sensor is a convolutional network model based on a depth-separable convolutional structure. The network model corresponding to the infrared sensor is a convolutional network model based on a composite coefficient optimization strategy. The network model corresponding to the gas sensor is a random forest model. The network model corresponding to the smoke sensor is a support vector machine model. The network model corresponding to the flame sensor is a decision tree model.

3. The method according to claim 1, characterized in that, The process of processing each type of raw environmental data to obtain target data feature information corresponding to each sensor includes: For each type of sensor, the raw environmental data collected by the sensor is preprocessed to obtain preprocessed target environmental data; Statistical feature extraction is performed on the target environment data to obtain the data statistical feature information corresponding to the sensor; Based on the target environment data, the data status is identified to obtain the data status information corresponding to the sensor. The data status information is used to characterize whether there is abnormal fire data in the target environment data. Based on the target environment data, the data statistical feature information, and the data status information, the target data feature information corresponding to the sensor is obtained.

4. The method according to claim 1, characterized in that, The fire detection results include a first predicted probability of the presence of a fire and a second predicted probability of the absence of a fire. The process of converting the fire detection results output by each network model into belief quality to obtain a first belief quality set corresponding to each network model includes: For each network model's output fire detection result, the first predicted probability in the fire detection result is determined as the belief quality of the existence of a fire in the first belief quality set corresponding to that network model. The second predicted probability in the fire detection result is determined as the belief quality of the absence of fire in the first belief quality set corresponding to the network model; Subtracting the belief quality of the existence of a fire and the belief quality of the non-existence of a fire from 1, the difference is determined as the belief quality of uncertain fires in the first belief quality set corresponding to the network model.

5. The method according to claim 1, characterized in that, Determining the degree of conflict between every two sets of first belief quality includes: Determine the Hellinger distance between every two sets of first belief quality, and define the Hellinger distance as the corresponding degree of conflict.

6. The method according to claim 1, characterized in that, The step of adjusting the first set of belief quality based on the degree of conflict to obtain an adjusted second set of belief quality includes: For every two sets of first belief quality, if the degree of conflict between the two sets of first belief quality is greater than a preset degree of conflict, then the uncertainty corresponding to each set of first belief quality is determined, and the first belief quality set is adjusted based on the uncertainty to obtain the adjusted second belief quality set. If the degree of conflict between these two sets of first belief quality is less than or equal to the preset degree of conflict, then each of these two sets of first belief quality is directly determined as the second set of belief quality.

7. The method according to claim 6, characterized in that, The step of determining the uncertainty corresponding to each of the two first belief quality sets, and adjusting the first belief quality set based on the uncertainty to obtain the adjusted second belief quality set, includes: Determine the Deng entropy corresponding to each of the two first belief quality sets, and define the Deng entropy as the corresponding uncertainty; Determine the target first belief quality set whose uncertainty is greater than a preset uncertainty from these two first belief quality sets; Based on the uncertainty corresponding to the target first belief quality set, determine the adjustment weight corresponding to the target first belief quality set; Based on the adjusted weights, the target first belief quality set is adjusted to obtain the adjusted second belief quality set.

8. The method according to claim 7, characterized in that, The step of determining the adjustment weight corresponding to the target first belief quality set based on the uncertainty corresponding to the target first belief quality set includes: Divide the uncertainty corresponding to the first set of target belief quality by the maximum uncertainty, and subtract 1 from the division result. The difference is used as the adjustment weight corresponding to the first set of target belief quality. The maximum uncertainty refers to the maximum value among the uncertainties corresponding to all first belief quality sets.

9. The method according to claim 7, characterized in that, The step of adjusting the target first belief quality set based on the adjustment weights to obtain the adjusted second belief quality set includes: The belief quality of the existence of fire in the first set of target belief quality is multiplied by the adjustment weight, and the resulting multiplication is used as the belief quality of the existence of fire in the adjusted second set of belief quality. The belief quality of the absence of fire in the first set of target belief quality is multiplied by the adjustment weight, and the resulting multiplication is used as the belief quality of the absence of fire in the adjusted second set of belief quality. Subtracting 1 from the adjusted belief quality for the existence of a fire and the belief quality for the non-existence of a fire, the difference is determined as the belief quality for uncertain fires in the adjusted second set of belief qualities.

10. The method according to claim 1, characterized in that, The process of fusing the second belief quality set corresponding to each network model to obtain the fused target belief quality set includes: Following the fusion order of all second belief quality sets, the first two second belief quality sets are fused first, and then the fused belief quality set is fused with the third second belief quality set, until the target belief quality set is obtained after fusion with the last second belief quality set. The fusion process corresponding to each pair of second belief quality sets includes: Based on two sets of second belief quality, cross-multiplication of all combinations in the identification framework is performed, and each multiplication result is categorized and added together to determine the belief quality for the existence of fire, the belief quality for the non-existence of fire, the belief quality for uncertain fire, and the conflict coefficient. Based on the aforementioned conflict coefficient, the summed belief quality regarding the existence of a fire, the belief quality regarding the absence of a fire, and the belief quality regarding an uncertain fire are normalized, and the normalized belief quality regarding the existence of a fire, the belief quality regarding the absence of a fire, and the belief quality regarding an uncertain fire are used as the fused belief quality set.

11. The method according to claim 10, characterized in that, The method further includes: If the conflict coefficient is 1, then the second belief quality set with the smallest uncertainty among the two second belief quality sets is determined as the fused belief quality set.

12. The method according to claim 1, characterized in that, The step of determining the first confidence level and first likelihood corresponding to the existence of a fire, and the second confidence level and second likelihood corresponding to the absence of a fire, based on the target belief quality set, includes: The belief quality regarding the existence of a fire in the target belief quality set is determined as the first level of trust corresponding to the existence of a fire; The belief quality for the existence of a fire and the belief quality for an uncertain fire are added together in the target belief quality set, and the sum is used as the first likelihood corresponding to the existence of a fire. The belief quality of the absence of fire in the target belief quality set is determined as the second level of trust corresponding to the absence of fire; The belief quality for the absence of fire and the belief quality for uncertain fire are added together in the target belief quality set, and the sum is used as the second likelihood corresponding to the absence of fire.

13. The method according to claim 1, characterized in that, The determination of the target fire detection result based on the first confidence level, the first likelihood level, the second confidence level, and the second likelihood level includes: If the first level of trust is greater than the preset level of trust, then the target fire detection result is determined to be that a fire exists; If the second level of confidence is greater than the preset level of confidence, then the target fire detection result is determined to be that there is no fire. If the first confidence level is less than or equal to the preset confidence level and the second confidence level is less than or equal to the preset confidence level, then the target fire detection result is determined to be in an uncertain state, the first likelihood and the second likelihood are output, and when the first likelihood is greater than the preset likelihood, a suspected fire alarm is triggered.

14. The method according to any one of claims 1-13, characterized in that, Each sensor is deployed at key risk locations in the industrial environment to form a complementary and comprehensive environmental sensing network. Each sensor continuously collects multi-dimensional environmental data based on a preset acquisition frequency, and transmits the continuously collected multi-dimensional environmental data to the edge server through a publish-subscribe message transmission protocol.

15. An industrial environment fire detection device, characterized in that, include: The raw environmental data acquisition module is used to acquire multi-dimensional environmental data collected by various sensors deployed in an industrial environment, including raw environmental data collected by each sensor. The fire detection result acquisition module is used to process each type of raw environmental data, obtain target data feature information corresponding to each sensor, and perform fire detection based on the network model corresponding to each sensor and the target data feature information to obtain the fire detection result output by each network model. The first belief quality set acquisition module is used to convert the belief quality of the fire detection results output by each network model to obtain the first belief quality set corresponding to each network model. The first belief quality set includes: belief quality for the existence of fire, belief quality for the absence of fire, and belief quality for uncertainty. The second belief quality set acquisition module is used to determine the degree of conflict between every two first belief quality sets, and adjust the first belief quality sets based on the degree of conflict to obtain the adjusted second belief quality set. The target belief quality set acquisition module is used to fuse the second belief quality set corresponding to each network model to obtain the fused target belief quality set. The target fire detection result determination module is used to determine, based on the target belief quality set, a first confidence level and a first likelihood corresponding to the presence of a fire, and a second confidence level and a second likelihood corresponding to the absence of a fire, and to determine the target fire detection result based on the first confidence level, the first likelihood, the second confidence level, and the second likelihood.

16. An electronic device, characterized in that, The electronic device includes: At least one processor; and A memory communicatively connected to the at least one processor; wherein, The memory stores a computer program that can be executed by the at least one processor, the computer program being executed by the at least one processor to enable the at least one processor to perform the industrial environment fire detection method according to any one of claims 1-14.

17. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores computer instructions that, when executed by a processor, implement the industrial environment fire detection method according to any one of claims 1-14.

18. A computer program product, characterized in that, The computer program product includes a computer program that, when executed by a processor, implements the industrial environment fire detection method according to any one of claims 1-14.

Images