Industrial park-oriented electrical fire comprehensive monitoring and early warning method and corresponding product

Abstract

The invention relates to the field of fire early warning, and provides an industrial park-oriented electrical fire comprehensive monitoring and early warning method and a corresponding product. The method comprises the following steps: synchronously acquiring multi-source data of the industrial park power supply network; processing the collected multi-source data of the industrial park power supply network so as to respectively extract electrical transient characteristics and thermal hidden danger space characteristics, fusing the characteristics with the multi-gas concentration data based on an attention mechanism, and outputting a high-dimensional hidden danger characteristic vector; carrying out parallel processing on the high-dimensional hidden danger feature vectors to respectively obtain a physical inconsistency score and a data-driven hidden danger probability score; dynamically adjusting the weight of the physical inconsistency score and the weight of the data-driven hidden danger probability score according to the signal-to-noise ratio and the feature uncertainty of the real-time data through an adaptive weighted fusion mechanism, and performing fusion to generate a dynamic electrical fire risk index; and according to the threshold interval in which the dynamic electrical fire risk index is located, generating early warning information of a corresponding level.

Description

Technical Field

[0001] This application relates to the field of fire early warning, and in particular to a comprehensive monitoring and early warning method and corresponding products for electrical fires in industrial parks. Background Technology

[0002] With the acceleration of industrialization, the power supply networks in industrial parks are becoming increasingly complex, characterized by large power loads, diverse equipment, and dense wiring, significantly increasing the risk of electrical fires. Electrical fire hazards are typically characterized by their concealment and tendency to spread; once they occur, they often cause significant loss of life and property. Therefore, early and accurate monitoring and warning of electrical fires are crucial to ensuring the safe operation of industrial parks.

[0003] Currently, monitoring and early warning technologies for electrical fires have evolved from traditional single-threshold alarms to comprehensive early warning systems that integrate multiple sensors and intelligent algorithms. Existing solutions typically deploy various sensors at key nodes of the power distribution system in industrial parks to collect signals such as residual current, temperature, electric arc, and smoke, and then analyze and process these signals using rule engines, statistical models, or machine learning algorithms. For example, some solutions simply summarize or weightedly average data from multiple sensors and compare it with a preset threshold to trigger an alarm; others utilize historical data to train classification or regression models and predict risks by identifying abnormal data patterns.

[0004] However, these existing technical solutions still have the following obvious defects or deficiencies in practical applications: 1) At the feature extraction and fusion level, most solutions have relatively simple processing methods for multi-source heterogeneous data such as electrical, thermal, and chemical data, making it difficult to capture early and weak fault characteristics induced by the coupling of multiple hidden dangers; 2) At the risk assessment level, existing solutions mostly rely on a single data-driven model or fixed physical rule thresholds, resulting in insufficient accuracy of early warning and a tendency to miss or false alarms; 3) There is a lack of an adaptive mechanism that can dynamically adjust the weights of different assessment criteria based on real-time operating conditions and data quality, making the early warning system less robust in complex and time-varying industrial environments. Summary of the Invention

[0005] This application provides a comprehensive monitoring and early warning method and corresponding products for electrical fires in industrial parks. Through intelligent fusion of multi-source data and adaptive hybrid evaluation, it significantly improves the accuracy and reliability of electrical fire early warning.

[0006] On the one hand, this application provides a comprehensive monitoring and early warning method for electrical fires in industrial parks, the method comprising:

[0007] Simultaneously collect multi-source data of the power supply network of the industrial park. The multi-source data includes three-phase current / voltage time sequence data of key nodes of the power supply network, infrared thermal image sequences of cable joints and key locations of distribution cabinets, and multi-gas concentration data inside the distribution box.

[0008] The multi-source data is processed to extract electrical transient features and thermal hazard spatial features, and the electrical transient features, thermal hazard spatial features and multi-gas concentration data are fused with the data based on an attention mechanism to output a high-dimensional hazard feature vector that comprehensively reflects the electrical, thermal and chemical states.

[0009] The high-dimensional hazard feature vector is input in parallel into two evaluation branches of the hybrid risk assessment model to obtain a physical inconsistency score and a data-driven hazard probability score, respectively. The two evaluation branches include an evaluation branch based on physical law verification and an evaluation branch based on historical data.

[0010] Through an adaptive weighted fusion mechanism, the weights of the physical inconsistency score and the data-driven hazard probability score are dynamically adjusted based on the signal-to-noise ratio and feature uncertainty of real-time data, and then fused to generate a dynamic electrical fire risk index.

[0011] Based on the threshold range of the dynamic electrical fire risk index, a corresponding level of early warning information is generated.

[0012] On the other hand, this application provides a comprehensive electrical fire monitoring and early warning device for industrial parks, the device comprising:

[0013] The acquisition module is used to synchronously acquire multi-source data of the power supply network of the industrial park. The multi-source data includes three-phase current / voltage time sequence data of key nodes of the power supply network, infrared thermal image sequences of cable joints and key locations of distribution cabinets, and multi-gas concentration data inside the distribution box.

[0014] The feature extraction module is used to process the multi-source data to extract electrical transient features and thermal hazard spatial features respectively, and to fuse the electrical transient features and thermal hazard spatial features with the multi-gas concentration data based on an attention mechanism to output a high-dimensional hazard feature vector that comprehensively reflects the electrical, thermal and chemical states.

[0015] The acquisition module is used to input the high-dimensional hazard feature vector in parallel into two evaluation branches of the hybrid risk assessment model to obtain a physical inconsistency score and a data-driven hazard probability score, respectively. The two evaluation branches include an evaluation branch based on physical law verification and an evaluation branch based on historical data.

[0016] The fusion module is used to dynamically adjust the weights of the physical inconsistency score and the data-driven hazard probability score based on the signal-to-noise ratio and feature uncertainty of real-time data through an adaptive weighted fusion mechanism, and then fuse them to generate a dynamic electrical fire risk index.

[0017] The generation module is used to generate early warning information of the corresponding level based on the threshold range of the dynamic electrical fire risk index.

[0018] Thirdly, this application provides an apparatus comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the steps of the technical solution of the above-described integrated monitoring and early warning method for electrical fires in industrial parks.

[0019] Fourthly, this application provides a storage medium storing a computer program, which, when executed by a processor, implements the steps of the technical solution described above for the integrated monitoring and early warning method for electrical fires in industrial parks.

[0020] As can be seen from the technical solution provided in this application, on the one hand, by simultaneously collecting heterogeneous data from multiple sources such as electrical, temperature, and gas sources, and processing it using a fusion method based on an attention mechanism, data from different physical sources can be effectively integrated. This allows the high-dimensional hazard feature vector generated by the fusion to not only contain various independent abnormal information, but also dynamically capture the intrinsic correlation and spatiotemporal coupling relationship between electrical, thermal, and chemical states through attention weights. This constructs a comprehensive feature that is more sensitive to complex coupled faults and has a stronger characterization ability, laying the foundation for subsequent accurate assessment. On the other hand, by adopting a hybrid risk assessment model that includes both physical law verification and historical data-driven branches, physical laws (e.g., circuit thermodynamic equations) are used to verify the consistency of observed data and assess the degree to which it violates known physical laws. This enhances the interpretability of the assessment results and the inference ability in areas lacking data. Furthermore, by mining complex patterns in historical data through a data-driven model, this parallel processing method can obtain physical inconsistency scores and data-driven hazard probability scores. The current state is cross-validated and comprehensively judged from two dimensions: physical laws and historical experience. This approach is more accurate than a single model or rule in identifying hazards. First, it avoids real-world hazards, reducing false alarms and missed alarms. Second, by introducing an adaptive weighted fusion mechanism, the weights of physical inconsistency scores and data-driven hazard probability scores in the final decision are dynamically adjusted based on the signal-to-noise ratio and feature uncertainty of real-time data. This enables the system to intelligently cope with various situations in complex industrial sites. Its dynamic adjustment strategy allows the entire early warning system to maintain stable early warning performance when facing challenges such as sensor errors, environmental interference, and new working conditions, exhibiting stronger environmental adaptability and robustness. Third, starting from multi-source data acquisition, through feature intelligent fusion, hybrid model risk assessment, and adaptive decision-making, it ultimately generates graded early warning information, forming a complete technical closed loop. This not only outputs risk levels but also inherently combines physical laws, data intelligence, and dynamic decision-making throughout the entire technical path. This makes early warning decisions no longer a simple threshold comparison but a complex decision-making process based on multi-dimensional and adaptive analysis. This helps park safety management personnel obtain earlier and more reliable early warning signals before a fire occurs and understand the possible causes behind the risk, thereby gaining valuable time to take precise and effective preventive measures and improving the initiative and intelligence level of safety management. In summary, the technical solution of this application significantly improves the accuracy and reliability of electrical fire early warning through intelligent fusion of multi-source data and adaptive hybrid evaluation. Attached Figure Description

[0021] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0022] Figure 1 This is a flowchart of a comprehensive monitoring and early warning method for electrical fires in industrial parks, provided in an embodiment of this application.

[0023] Figure 2 This is a schematic diagram of the structure of the integrated electrical fire monitoring and early warning device for industrial parks provided in the embodiments of this application;

[0024] Figure 3 This is a schematic diagram of the device provided in the embodiments of this application. Detailed Implementation

[0025] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0026] In this specification, adjectives such as "first" and "second" are used only to distinguish one element or action from another, without necessarily requiring or implying any actual such relationship or order. Where circumstances permit, reference to an element or component or step (etc.) should not be construed as being limited to only one of the elements, components, or steps, but may be one or more of the elements, components, or steps, etc.

[0027] For ease of description, the dimensions of the various parts shown in the accompanying drawings are not drawn to actual scale.

[0028] Currently, monitoring and early warning technologies for electrical fires have evolved from traditional single-threshold alarms to comprehensive early warning systems that integrate multiple sensors and intelligent algorithms. Existing solutions typically deploy various sensors at key nodes of the power distribution system in industrial parks to collect signals such as residual current, temperature, electric arc, and smoke, and then analyze and process these signals using rule engines, statistical models, or machine learning algorithms. For example, some solutions simply summarize or weightedly average data from multiple sensors and compare it with a preset threshold to trigger an alarm; others utilize historical data to train classification or regression models and predict risks by identifying abnormal data patterns. However, these existing technical solutions still have the following obvious defects or deficiencies in practical applications: 1) At the feature extraction and fusion level, most solutions handle multi-source heterogeneous data such as electrical, thermal, and chemical data in a relatively simple way. They usually extract features from each source data independently and then directly splice or perform shallow weighting. They fail to fully consider the deep correlation and spatiotemporal coupling between different data patterns, resulting in limited representation capabilities of the fused feature vectors, making it difficult to capture early and weak fault features induced by the coupling of multiple hidden dangers; 2) At the risk assessment level, existing solutions mostly rely on a single data-driven model (e.g., neural network, support vector machine) or fixed physical rule thresholds. The former has weak generalization ability and poor interpretability when the data quality is poor or the operating conditions are complex and changeable, while the latter is not sensitive to latent and developing faults, resulting in insufficient accuracy of early warning and easy to miss or false alarms; 3) There is a lack of an adaptive mechanism that can dynamically adjust the weights of different assessment criteria according to real-time operating conditions and data quality, making the early warning system less robust in complex and time-varying industrial environments.

[0029] To address the aforementioned problems in existing technologies, this application proposes a comprehensive monitoring and early warning method for electrical fires in industrial parks, the flowchart of which is attached. Figure 1 As shown, the main steps include S101 to S105, which are detailed below:

[0030] Step S101: Synchronously collect multi-source data of the power supply network of the industrial park, including three-phase current / voltage time sequence data of key nodes of the power supply network, infrared thermal image sequences of key locations of cable joints and distribution cabinets, and multi-gas concentration data inside the distribution box.

[0031] Early signs of electrical fire hazards may be scattered across changes in various physical quantities. To achieve comprehensive detection, this application abandons the approach of monitoring a single signal and simultaneously collects multi-source data from the industrial park's power supply network. This multi-source data includes three-phase current / voltage time-series data of key nodes in the power supply network, infrared thermal image sequences of cable joints and key locations in distribution cabinets, and multi-gas concentration data inside distribution boxes. These data provide an information foundation for comprehensive assessment from three key dimensions: electrical operating status, localized overheating phenomena, and insulation material decomposition. Specifically, as an embodiment of this application, the simultaneous collection of multi-source data from the industrial park's power supply network can be achieved through steps S1011 to S1014, as detailed below:

[0032] Step S1011: Configure a high-frequency current sensor and a voltage sensor for each phase of the key nodes of the power supply network, and synchronously collect three-phase current / voltage timing data at a preset first sampling frequency.

[0033] Key nodes include, but are not limited to, transformer outgoing terminals, low-voltage main incoming switchgear, and important feeder circuits. The first sampling frequency (which can be set to 10kHz) must be sufficient to capture high-frequency fault characteristics such as current harmonics and transient inrush currents. The sensor must have high precision and synchronous sampling capabilities to ensure the accurate phase relationship of the three-phase data, which is a prerequisite for subsequent analysis of potential electrical imbalances, harmonics, and other hazards.

[0034] Step S1012: Deploy infrared thermal imagers at key locations such as cable joints and distribution cabinets, periodically acquire infrared thermal image sequences at a second sampling frequency lower than the first sampling frequency, and timestamp each frame of thermal image.

[0035] Since temperature changes relatively slowly compared to electrical signal changes, a second sampling frequency (which can be set to 1Hz) is sufficient for monitoring requirements. Infrared thermal imagers can non-contactly measure the surface temperature field distribution of equipment. Timestamping each frame of the thermal image is done to align it with the electrical data in the time dimension, laying the foundation for subsequent analysis of the electro-thermal coupling relationship.

[0036] Step S1013: Deploy a gas sensor array in the distribution box, wherein the gas sensor array includes sensors for detecting hydrogen (H2), carbon monoxide (CO) and total volatile organic compounds (TVOC), and acquires the multi-gas concentration data at a third sampling frequency.

[0037] Insulating materials release characteristic gases at different stages of overheating decomposition. H2 and CO are typical products of electric arcing and severe overheating, while TVOC encompasses a variety of organic gases released during the initial thermal decomposition of insulating materials (e.g., polyethylene, epoxy resin). A third sampling frequency (which can be set to 0.1 Hz) is typically lower than the temperature sampling frequency. The gas sensor array constitutes a direct monitoring layer for these "invisible" chemical changes.

[0038] Step S1014: Provide a time reference for all sensors through a unified clock source to ensure that the three-phase current / voltage timing data, infrared thermal imaging sequence and multi-gas concentration data are synchronized in time.

[0039] This is the cornerstone for achieving effective fusion of multi-source data. By providing all sensing units with a unified nanosecond or microsecond-level timestamp through GPS or IEEE 1588 precision clock protocol, data misalignment caused by asynchronous acquisition systems can be eliminated, ensuring that subsequent feature fusion and correlation analysis are based on precise spatiotemporal alignment.

[0040] After synchronously collecting multi-source data, in order to improve data quality, these multi-source data can be preprocessed, specifically including steps S1015 to S1017, which are explained in detail below:

[0041] Step S1015: Denoise and normalize the three-phase current / voltage time series data, and interpolate missing values.

[0042] Denoising can be achieved using wavelet thresholding or digital filters to eliminate high-frequency electromagnetic interference. Normalization can scale the current and voltage values ​​of each phase to the [0,1] or [-1,1] range, eliminating the influence of dimensions and accelerating model convergence. For missing values ​​caused by brief communication interruptions, linear interpolation or interpolation algorithms based on historical patterns can be used to fill in the missing values.

[0043] Step S1016: Perform distortion correction and non-uniformity correction on the infrared thermal image sequence, and convert the gray values ​​of the thermal image into temperature values ​​based on the emissivity calibration results of the device surface to form a temperature field sequence.

[0044] The infrared thermal imager lens exhibits distortion, and the responses of individual pixels are inconsistent. Distortion correction is performed geometrically using the intrinsic parameter matrix obtained from the calibration board. Non-uniformity correction utilizes blackbody calibration data to establish a grayscale-temperature response curve for each pixel. Emissivity calibration based on the device surface is crucial because the emissivity (ε) varies significantly between different materials (e.g., copper busbars, insulating ceramic insulators, sprayed metal). The conversion formula is: (Under certain approximate conditions), where, The radiation temperature is calculated from the original grayscale value using the sensor response function. This represents the corrected surface temperature of the actual object. This yields a precise two-dimensional temperature field matrix sequence.

[0045] Step S1017: Unify the units and normalize the dimensions of the multi-gas concentration data.

[0046] The concentration values ​​output by different gas sensors are uniformly converted to standard units, such as mg / m³, and normalized to make them close to the numerical range of electrical and thermal characteristics, which facilitates subsequent fusion model processing.

[0047] Step S102: Process the multi-source data of the power supply network of the industrial park to extract electrical transient features and thermal hazard spatial features respectively, and fuse the electrical transient features, thermal hazard spatial features and multi-gas concentration data based on attention mechanism to output a high-dimensional hazard feature vector that comprehensively reflects the electrical, thermal and chemical states.

[0048] For multi-source data of industrial park power supply networks collected synchronously, one existing processing method is to simply concatenate the extracted electrical, thermal, and gas feature vectors into a long vector. However, this method treats features with different physical meanings, dimensions, and importance levels equally, making it difficult for the model to automatically learn the intrinsic relationships between them. It is also susceptible to interference from redundant and noisy features, resulting in low learning efficiency and weak generalization ability of subsequent models. Another existing processing method is to assign fixed or empirical weights to features from different sources and then add them together. The drawback of this method is that the weights are fixed and cannot be dynamically adjusted according to the specific context of the current data (e.g., a certain hidden danger pattern is manifested as strong thermal features while electrical features are weak). This results in poor flexibility and insufficient adaptability in complex and ever-changing industrial scenarios. Therefore, in order to effectively extract and fuse a comprehensive feature representation that can fully and deeply reflect the early signs of complex coupled faults from multi-source heterogeneous data such as electrical, temperature, and gas data, and to avoid information "dilution" and loss of correlation caused by simple feature splicing, the technical solution adopted in this application is to process the multi-source data to extract electrical transient features and thermal hazard spatial features separately, and then fuse the electrical transient features, thermal hazard spatial features and multi-gas concentration data based on an attention mechanism to output a high-dimensional hazard feature vector that comprehensively reflects the electrical, thermal, and chemical states. The following is a detailed explanation through steps S1021 to S1023:

[0049] Step S1021: Extract electrical transient features.

[0050] Electrical transient characteristics are crucial for capturing early fault information (e.g., harmonic distortion, arcing characteristics, imbalance, etc.) implicit in current and voltage signals. Conventional time-domain statistical characteristics (e.g., RMS value, peak value) are insensitive to such transient changes. This application performs wavelet packet transform on three-phase current / voltage time-series data, decomposing the signal layer by layer according to a preset number of decomposition levels to obtain multiple sub-band signals of different frequency bands. Specifically, the time-series signal of each phase current or voltage is preprocessed... The signal undergoes N-level wavelet packet decomposition. The number of decomposition levels N needs to be determined based on the sampling frequency and the frequency band of the fault characteristics of interest. For example, to capture high-frequency arc characteristics, N can be set to 5 or 6. After N-level decomposition, the original signal is decomposed into... Each sub-band signal , Each sub-band covers a specific frequency range. This process can be visualized as a fine time-frequency filter bank that unfolds the energy distribution of the signal from the time domain to a two-dimensional time-frequency plane.

[0051] Then, the wavelet packet energy of each sub-band signal is calculated, and its energy entropy is calculated based on the wavelet packet energy to form a set of electrical transient feature vectors characterizing the energy distribution state of current / voltage signals in different frequency bands.

[0052] First, calculate the energy of the i-th sub-band signal: Where M is the number of discrete points of the sub-band signal, then the total energy of the signal is... Then, calculate the proportion of energy in the total energy for each sub-band. Finally, the wavelet packet energy entropy H is calculated: Energy entropy H is a scalar that quantifies the uniformity of signal energy distribution in the time-frequency domain. When the signal is normal and stable, the energy distribution is relatively concentrated, and the entropy value is low; when a fault occurs, causing energy to be dispersed across multiple frequency bands (e.g., generating abundant harmonics), the entropy value increases significantly. However, using only the entropy value loses the pattern information of the energy distribution. Therefore, this application uses the energy proportion of each sub-band. The vector formed Alternatively, to reduce dimensionality, the feature vector composed of the energy of each layer and node can be used as the electrical transient feature vector. This vector not only contains information about the energy magnitude, but more importantly, it contains the energy distribution pattern at different frequency scales, making it extremely sensitive to early and weak electrical faults.

[0053] Step S1022: Extract spatial features of thermal hazards.

[0054] Infrared thermography provides the spatial distribution of temperature, but a single temperature value cannot reflect the dynamic process and spatial gradient of heat diffusion, which are crucial for locating hotspots and assessing the severity of overheating. Existing methods typically focus only on the highest temperature point, are susceptible to environmental reflection interference, and cannot distinguish between uniform heating and localized dangerous hotspots. The technical solution adopted in this application is to calculate the dense optical flow field between two consecutive frames of temperature field images in an infrared thermographic sequence, obtaining the temperature diffusion velocity vector for each pixel.

[0055] Let time t and time t be two consecutive frames. The temperature field images are respectively and Based on the assumption of constant brightness (in this case, the assumption of constant temperature for material points), dense optical flow methods (such as the Farneback algorithm) can calculate the brightness of each pixel. exist Displacement vector within time This vector is the temperature diffusion rate vector (unit: pixels per second). It indicates the direction and speed of temperature (or isotherm) movement on the image plane, directly reflecting the trend of heat transfer.

[0056] Subsequently, based on the temperature diffusion velocity vector, the magnitude and direction of the spatial gradient of the temperature field are calculated, generating a thermal diffusion gradient map. Spatial gradient of temperature field T. It inherently reflects the direction and rate of the fastest temperature change. Combining optical flow information, this application can calculate a weighted or enhanced gradient. One implementation is to directly use the magnitude of the optical flow vector. The intensity and direction of thermal diffusion are approximated or corrected using optical flow and temperature gradient. A more refined approach is to consider the synergy between optical flow and temperature gradient. Ultimately, this yields a gradient magnitude for each pixel. and direction A thermal diffusion gradient map. This map highlights areas where heat is rapidly accumulating or flowing, revealing potential and developing thermal hazards more effectively than a static temperature map.

[0057] Furthermore, by combining the three-dimensional geometric model of the target device, the heat diffusion gradient map is mapped onto the surface of the three-dimensional geometric model, and the normal and tangential heat flux densities at key connection points of the model are extracted as the spatial features of the thermal hazard. Two-dimensional thermal images lose depth information and the true geometry of the device, making it impossible to accurately calculate the actual heat flux density. This application uses a pre-established three-dimensional geometric model of the target device, such as a power distribution cabinet or cable joint (e.g., a CAD model), and utilizes camera calibration parameters to back-project each pixel in the two-dimensional heat diffusion gradient map onto the corresponding point on the surface of the three-dimensional model. At point P on the surface of the 3D model, there exists a normal vector. From two-dimensional gradient The heat flow vector within the tangent plane of a three-dimensional surface can be derived. Normal heat flux density (This needs to be estimated in conjunction with the laws of thermal conduction.) It reflects the ability of heat to dissipate into the environment; a value that is too small indicates poor heat dissipation. Tangential heat flux density. This reflects the diffusion of heat along the equipment surface. At connection points, abnormal tangential heat flow may indicate increased contact resistance. This is achieved by extracting data from key connection points in the model (e.g., bolted joints, contacts). and We obtained a set of spatial feature vectors for thermal hazards that have clear physical meaning, are strongly correlated with equipment structure, and are crucial for locating overheating faults.

[0058] Step S1023: Integrate electrical transient features, spatial features of thermal hazards, and multi-gas concentration data based on an attention mechanism.

[0059] Thus, three types of heterogeneous features are now available: electrical transient feature vectors characterizing power quality anomalies. Spatial feature vectors characterizing thermal hazards caused by abnormal heat accumulation and diffusion and multi-gas concentration data vectors characterizing the decomposition of insulating materials Simple concatenation The assumption that all feature dimensions contribute statically and equally to the final risk assessment is clearly unrealistic. For example, in some cases, the overheating feature ( ) may be dominant, while in other cases the arc characteristics ( Specific frequency bands in) and acetylene gas ( The combination of components is even more crucial.

[0060] To address this issue, this application employs an attention-based fusion method, specifically implemented through the following steps S1023a to S1023c:

[0061] Step S1023a: Map electrical transient features, thermal hazard spatial features, and multi-gas concentration data to the same high-dimensional feature space to form corresponding feature sequences.

[0062] First, three independent linear transformation layers (or small fully connected networks) are used to transform... , and Mapped to a common high-dimensional space of the same dimension: , as well as .in, (include , and ) is the weight matrix, (include , and () is the bias vector. This is how we obtain... , and Having the same semantic depth and dimension This facilitates subsequent interactive computation. They are considered as a feature sequence of length 3, where each element is a... A dimensional vector.

[0063] Step S1023b: Input the feature sequence into a multi-head attention module and calculate the cross-attention weights among the electrical transient feature sequence, the thermal hazard spatial feature sequence, and the multi-gas concentration data feature sequence.

[0064] This application uses a multi-head attention module to compute cross-attention weights. Specifically, the sequence... As input. For each head, first generate a query vector, key vector, and value vector for each element in the sequence. For example, for the i-th feature (e.g., ): , and .

[0065] Then, calculate the attention score of the i-th feature to all features in the sequence (including itself) (using scaled dot product attention): ,in, It is the dimension of the key vector. This represents how much attention should be given to the j-th feature during the fusion process for the i-th feature, i.e., the cross-attention weight. For example, This indicates how much weight should be given to thermal features during fusion of electrical features. Multi-head attention performs this calculation multiple times in parallel, then concatenates and linearly transforms the outputs of the different heads, allowing the model to focus on information from different feature sources in different representation subspaces.

[0066] Step S1023c: Based on the cross-attention weights, the three feature sequences are weighted and summed and interacted with to generate a context-aware fusion feature vector as a high-dimensional hidden danger feature vector.

[0067] After obtaining the attention weight matrix, for each feature, the output after weighted aggregation through the attention mechanism is: .

[0068] This output is no longer the original feature. It is not a new representation that incorporates other feature information. Finally, all of them Aggregation (e.g., summing or concatenating again before passing through a feedforward network) is performed to form the final context-aware fused feature vector. It's called "context-aware" because during the fusion process, each feature dynamically adjusts its "expression" based on the strength of its association with other features (attention weights), ultimately... It contains profound information about how heat and gas behave under the current electrical conditions, and vice versa. This is a high-dimensional hazard feature vector, which is a unified and powerful feature representation that can comprehensively reflect the electrical, thermal, and chemical states and their internal relationships, providing optimal information input for subsequent risk assessment.

[0069] As can be seen from the above steps S102 and their specific implementation, context-aware intelligent fusion is achieved by introducing an attention mechanism. This intelligent fusion mechanism can dynamically and selectively focus on features from different data sources. During the fusion process, it automatically learns and strengthens the feature combinations most relevant to the current risk state, while weakening irrelevant or noisy information. This makes the generated high-dimensional hidden danger feature vector not only contain the original information, but also contain cross-modal related semantics. Its representation ability is stronger, and it can more keenly capture the early, weak composite fault features caused by the combined effects of multiple factors (e.g., poor contact leading to overheating and accompanied by gas production from the decomposition of insulation materials), providing high-quality information input for subsequent accurate risk assessment.

[0070] Step S103: Input the high-dimensional hazard feature vector in parallel into the two evaluation branches of the hybrid risk assessment model to obtain the physical inconsistency score and the data-driven hazard probability score, respectively. The two evaluation branches include an evaluation branch based on physical law verification and an evaluation branch based on historical data.

[0071] Overcoming the limitations of single risk assessment models, such as the "black box" nature of data-driven models, their dependence on training data distribution, and the insensitivity of physical models to complex latent faults, and constructing a robust assessment framework that combines interpretability, generalization ability, and latent fault detection capability, has always been a focus of industry attention. One related technology uses pure data-driven models (e.g., deep neural networks) that rely on historical data for training and prediction. This approach heavily depends on the quality and coverage of training data, and is prone to false alarms or missed alarms in industrial scenarios with variable operating conditions and few rare fault samples. Furthermore, the model's decision-making process lacks interpretability, making it difficult for maintenance personnel to understand and trust the warning results. Another existing technology uses pure physical / rule-based models, which judge based solely on known physical laws (e.g., Ohm's law, thermodynamic formulas) and set fixed thresholds or rules. However, such pure physical or rule-based models struggle to model complex, nonlinear system behaviors, are insensitive to latent faults without significant physical quantity changes, such as insulation aging and intermittent arcing, and face difficulties in setting warning thresholds—too conservative leads to missed alarms, while too sensitive leads to false alarms. Therefore, this application inputs high-dimensional hazard feature vectors in parallel into two evaluation branches of a hybrid risk assessment model, obtaining a physical inconsistency score and a data-driven hazard probability score, respectively. The two evaluation branches include one based on physical law verification and the other based on historical data. By constructing a parallel dual-branch evaluation architecture of physical verification and data-driven approaches, the evaluation paradigms can be complemented and enhanced. In other words, the physical branch provides interpretable consistency checks based on first principles, independent of data volume, and excels at identifying anomalies that violate basic physical laws. The data-driven branch, on the other hand, mines complex failure modes from historical data that are difficult to describe with explicit equations. Both branches work in parallel, independently and concurrently evaluating the same hazard feature vector, providing risk assessment basis from both theoretical and empirical dimensions. This not only improves the reliability of the assessment (dual verification) but also enhances the interpretability of the results (the physical inconsistency score provides an intuitive physical explanation) and avoids the direct transmission of errors from one model, forming a robust evaluation structure with mutual verification and complementary information.

[0072] In the above embodiments, the objective of the evaluation branch based on physical law verification is to assess the degree of consistency between the current observation data and known physical laws. Its basic assumption is that severe electrical fault precursors often cause the system's operating state to deviate from fundamental physical laws, such as energy conservation and heat conduction equations. The physical inconsistency score obtained by the evaluation branch based on physical law verification is mainly achieved through steps Sa1031 to Sa1034, as detailed below:

[0073] Step Sa1031: Construct a physical information neural network that describes the relationship between the temperature rise at the circuit connection point and the current, contact resistance, and heat dissipation conditions. The loss function of the network includes the residual constraints of the circuit overheating differential equation.

[0074] Specifically, first, a physical model is established to describe the temperature rise at electrical connection points (e.g., cable joints, switch contacts). Based on the principle of thermal equilibrium, the temperature rise process can be described by differential equations. A simplified form is: ,in, It is relative to the ambient temperature. The temperature rise, C is heat capacity, G is thermal conductivity. It is the current flowing through the contact point, and R is the contact resistance. This represents Joule thermal power. The core idea of ​​the Physics-Informed Neural Network (PINN) is to embed this differential equation as a constraint into the training process of the neural network. A system is constructed using time t and associated state features (which can be obtained from...) Extracting data (such as equivalent current characteristics, current temperature characteristics, etc.) as input to predict temperature rise. The neural network for output ,in, These are network parameters. The loss function includes residual constraints from the circuit overheating differential equation, specifically in the form:

[0075]

[0076]

[0077]

[0078] in, It is a data fitting loss, which uses historical normal data during training. It is the physical residual loss, which forces the network to approximately satisfy the differential equation at any input point (including points that have not seen the data). It involves balancing hyperparameters. After the network is trained, its parameters... It encodes a mapping relationship that both fits the data and obeys the laws of physics.

[0079] Step Sa1032: Input the feature subset containing the thermally related part of the electrical transient features and the high-dimensional hidden danger feature vector into the physical information neural network.

[0080] Step Sa1033: The physical information neural network performs forward calculations to obtain the predicted temperature rise curve, and calculates the residual between the predicted result and the physical law described by the circuit overheating differential equation.

[0081] Step Sa1034: Normalize the norm of the residual between the prediction result and the physical law described by the circuit overheating differential equation and use it as the physical inconsistency score.

[0082] Steps Sa1032 to Sa1034 are the actions performed during the network inference (early warning) phase. In this phase, a subset of features, including electrical transient features and the thermally relevant portion of the high-dimensional hazard feature vector, is input into the physical information neural network. The network performs forward computation to obtain the predicted temperature rise curve. Simultaneously, substituting the same input into the right-hand side of the physical equation (heat source term) and utilizing the network's automatic differentiation capability to calculate the left-hand side of the equation, we obtain... Then, the residual between the predicted result and the physical law described by the circuit overheating differential equation is calculated. This residual This measures the extent to which the neural network's predictions "violate" the physical laws that it was trained to follow. Finally, the norm of the residuals is normalized and used as a score for physical inconsistency. ,For example: .

[0083] in, It is the average residual norm calculated on the training set and used for normalization. The larger the residual, the greater the degree to which physical laws are violated, indicating the possible existence of abnormal thermal processes that cannot be explained by normal physical models, thus indicating a high risk.

[0084] It should be noted that physical information neural networks are more reliable in regions with dense training data distribution, but exhibit high uncertainty in regions outside this distribution. To quantify this, this application, when obtaining the physical inconsistency score, also includes the following evaluation process for the confidence of the physical model: monitoring the distribution density of the feature subset input to the physical information neural network within the domain of the circuit overheating differential equation; determining the confidence of the physical model when the feature subset is located in a high-density region of the historical training data distribution. If the score is high, it is considered high; otherwise, it is considered low. The confidence level of the physical model is used as a confidence correction factor for the physical inconsistency score, and is applied to subsequent adaptive weighted fusion. For example, the final physical score can be corrected to... When the confidence level is low, the score will be appropriately discounted, making the decision more prudent. In the above embodiments, the distribution density of the feature subset of the neural network monitoring the input physical information within the domain of the circuit overheating differential equation can be achieved by calculating the distance between the feature subset and the features of the training dataset in a high-dimensional space (e.g., K-nearest neighbor distance).

[0085] In the above embodiments, the goal of the historical data-driven evaluation branch is to learn complex, potential failure mode mapping relationships from a large amount of historical data (including normal and failure cases). The historical data-driven evaluation branch obtains the data-driven hazard probability score mainly through steps Sb1031 to Sb1033, as detailed below:

[0086] Step Sb1031: Input the high-dimensional hidden danger feature vector into the temporal convolutional network model, where the temporal convolutional network model extracts long-term dependencies in the feature sequence through dilated causal convolution.

[0087] Temporal Convolutional Networks (TCNs) offer advantages over Recurrent Neural Networks (RNNs) in terms of training stability, parallel efficiency, and the ability to capture longer historical dependencies. Their core component is the dilated causal convolution. For a one-dimensional sequence input... (or its time window sequence), in a certain layer of the TCN, for an element at position t in the output sequence, its value is determined by the position in the input sequence. The element-wise convolution is obtained, where k is the index of the convolution kernel element and d is the dilation factor. The dilation factor increases exponentially with the number of network layers l (e.g., ...). This allows high-level neurons to perceive a very wide range of input history with fewer layers, thereby extracting long-term dependencies in feature sequences.

[0088] Step Sb1032: The temporal convolutional network model outputs a matching score representing the degree of matching between the current multi-source data state and the features of historical electrical fire cases.

[0089] The TCN model is trained end-to-end using a large amount of historical data (including multidimensional feature sequences and corresponding labels for "whether a fire occurred" or "how long before the fire") during the training phase. During the inference phase, the temporal convolutional network model outputs a matching score representing the degree of matching between the current multi-source data state and the features of historical electrical fire cases. .

[0090] Step Sb1033: Map the matching score between the current multi-source data status and the characteristics of historical electrical fire cases to a probability value between 0 and 1 using the Sigmoid function, and use it as a data-driven hazard probability score.

[0091] Matching score of current multi-source data status with historical electrical fire case features Essentially, this reflects the similarity between the current feature pattern and known hazard patterns in the model's "memory." Finally, the matching score is mapped to a probability value between 0 and 1 via a sigmoid function, serving as the data-driven hazard probability score. :

[0092]

[0093] It is a scalar between 0 and 1, which can be intuitively understood as the probability of an electrical fire occurring at the current moment, inferred from historical experience.

[0094] Step S104: Through an adaptive weighted fusion mechanism, the weights of the physical inconsistency score and the data-driven hazard probability score are dynamically adjusted based on the signal-to-noise ratio and feature uncertainty of the real-time data, and then fused to generate a dynamic electrical fire risk index.

[0095] Scores were obtained from both physics and data-driven perspectives. and Subsequently, how to make intelligent and dynamic final decisions based on real-time and specific operating conditions (e.g., data quality, environmental interference, newness of operating conditions, etc.) to improve the overall robustness and scientific nature of the system in complex real-world environments is a key challenge. One existing solution is fixed-weight fusion, such as simple averaging or pre-setting fixed weights based on experience. This solution completely ignores changes in real-time data quality and specific contexts. For example, when new equipment or sensors are subjected to strong interference, the output of the data-driven model may be unreliable, but the fixed weights will still give it a large weight in the decision-making process, leading to misjudgments. Another existing solution is a hard selection based on results, such as always using the model with the highest score or setting a fixed threshold for selection. This solution results in an either / or choice and fails to fully utilize the complementary information provided by the two models. When one model temporarily outputs an extremely high score due to occasional noise, it may incorrectly dominate the decision-making process. To enhance the adaptability, stability, and overall robustness of the early warning system in the face of various uncertainties such as data noise, sensor anomalies, new equipment commissioning, and extreme loads, and to transform the method from a laboratory model into an engineering system that can reliably operate in complex industrial sites, the technical solution adopted in this application is to dynamically adjust the weights of physical inconsistency scores and data-driven hazard probability scores based on the signal-to-noise ratio and characteristic uncertainties of real-time data, and then fuse them to generate a dynamic electrical fire risk index. Specifically, this can be achieved through steps S1041 to S1044, as detailed below:

[0096] Step S1041: Calculate the signal-to-noise ratio of the three-phase current / voltage time series data in real time, and estimate the uncertainty of each component in the high-dimensional hidden danger feature vector by using the model prediction variance or Monte Carlo Dropout method.

[0097] Signal-to-noise ratio (SNR) calculation: SNR is a core metric for measuring the quality of raw data. For each phase current / voltage time series data... A short-time window of signal containing the current moment can be selected. First, the noise power is estimated. This can be obtained by calculating the variance after high-pass filtering, or by statistically analyzing the background noise during periods of no load abrupt changes. Signal power The signal-to-noise ratio (dB) of this phase data can be estimated by subtracting the noise power from the total power: .

[0098] Final system-level signal-to-noise ratio The minimum value among the three phases can be taken, representing the quality of the worst channel. When the signal-to-noise ratio is below a first threshold (e.g., 20 dB), it indicates that the original electrical data is severely disturbed, and the reliability of all features extracted based on this and the model predictions they drive will decrease.

[0099] Feature uncertainty estimation: measuring the eigenvector The reliability of the model prediction variance is assessed. This application employs one of two methods: model prediction variance and Monte Carlo Dropout.

[0100] 1) Model Prediction Variance: If the Data-Driven Branch (TCN) is constructed with a probabilistic model (e.g., a Bayesian neural network), its output itself carries a variance estimate. It can be used directly as Uncertainty in rating The physical branch can also be obtained similarly through multiple inference variances of PINN. .

[0101] 2) Monte Carlo Dropout (MC Dropout): This is a practical uncertainty estimation method. During inference, Dropout is enabled for both TCN and PINN, and T forward propagations are performed (e.g., T=50). T outputs are obtained. Its prediction variance is the uncertainty: Similarly, we can obtain Overall uncertainty of eigenvectors Can be taken and This is a combination of factors, such as the maximum or mean. If the feature uncertainty is higher than the second threshold, it indicates that the model is unsure about the current input, and the prediction results are unreliable.

[0102] Step S1042: Set a physical score weighting factor and a data score weighting factor, the sum of which is 1.

[0103] Suppose a physical inconsistency score The weight is Data-driven hazard probability scoring The weight is And it satisfies the following constraints: , .

[0104] Step S1043: When the signal-to-noise ratio is lower than the first threshold or the feature uncertainty is higher than the second threshold, increase the value of the physical scoring weight factor and decrease the value of the data scoring weight factor.

[0105] This is the core of adaptive logic. The weight adjustment function can be designed as a dynamic mapping. For example, a data reliability factor R can be defined: ,in, and It is a monotonically increasing function that maps SNR and the reciprocal of uncertainty to the interval [0,1]. When SNR is high and uncertainty is low, When SNR is low or uncertainty is high, Then, dynamically set the weights: , This means that when data quality is high and predictions are certain, the system places more trust in the data-driven model. Large); when data is disturbed or the model is uncertain, the system becomes more reliant on the testing of physical laws ( (Large). This is a decision-making approach that aligns with the thinking of human experts: relying on experience when the evidence is conclusive, and reverting to basic principles when the situation is unclear.

[0106] Step S1044: Based on the adjusted weights, the physical inconsistency score and the data-driven hazard probability score are weighted and summed to generate the dynamic electrical fire risk index.

[0107] The final Risk Index (RI) is calculated as follows: .

[0108] because and All values ​​have been normalized to similar dimensions (e.g., between 0 and 1), and the RI is also a scalar between 0 and 1. It is not a simple probability or physical bias, but a dynamic, comprehensive risk measure that integrates physical verification and data experience, intelligently weighing current data quality. A higher RI value indicates a greater overall fire risk assessed by the system. This dynamic characteristic allows the risk index to adapt to the complex and ever-changing environment of industrial sites, significantly improving the overall robustness and reliability of the early warning system.

[0109] As can be seen from step S104 and its specific implementation, by introducing an adaptive weighted fusion mechanism based on real-time data signal-to-noise ratio and feature uncertainty, the system is endowed with dynamic environmental perception and intelligent decision-making capabilities. This mechanism can assess the credibility status of the current data in real time: when the signal-to-noise ratio is high and the features are certain, it indicates that the data quality is good and the operating conditions are common, and the system tends to trust the data-driven model that learns complex patterns from massive data; when the signal-to-noise ratio is low and the features are uncertain, it indicates that the data may be disturbed or is in an operating condition outside the model's cognition, and the system tends to trust the model with universal physical laws. This strategy of dynamically and smoothly adjusting the weights according to the environment enables the final dynamic electrical fire risk index to flexibly and optimally integrate the advantages of the two assessment perspectives, significantly improving the adaptability, stability, and overall robustness of the early warning system in the face of various uncertainties such as data noise, sensor anomalies, new equipment commissioning, and extreme loads, transforming the method from a laboratory model into an engineering system that can reliably operate in complex industrial sites.

[0110] Step S105: Generate early warning information of the corresponding level based on the threshold range of the dynamic electrical fire risk index.

[0111] After generating the dynamic electrical fire risk index RI, it can be converted into actionable early warning instructions. To provide differentiated emergency response guidance, this application generates corresponding levels of early warning information based on the threshold range of the dynamic electrical fire risk index. Tiered early warning helps avoid the "crying wolf" effect, prioritizing limited operational resources for the highest-risk hazards. Specifically, generating corresponding levels of early warning information based on the threshold range of the dynamic electrical fire risk index can be achieved through steps S1051 to S1053, as detailed below:

[0112] Step S1051: Preset at least three risk thresholds to divide the risk index range into four levels: low risk, medium risk, high risk, and emergency risk.

[0113] For example, setting three thresholds , and ,and The risk levels are classified as follows:

[0114] Low risk (Level 0): Status monitoring and logging;

[0115] Medium risk (Level 1): Early warning and notification, and arrangement of inspections;

[0116] High risk (Level 2): An on-site alarm has been triggered; immediate verification is required.

[0117] Emergency Risk (Level 3): In case of an emergency alarm, consider remote power outage or initiating emergency procedures.

[0118] The above threshold , and It can be set through historical data analysis, expert experience, or false alarm rate tolerance.

[0119] Step S1052: When the risk index exceeds a certain level threshold, generate an early warning message containing the risk level, risk index value, and timestamp.

[0120] Once the RI value enters or crosses a certain threshold range, the system immediately triggers the warning generation process. The generated warning message is a structured data packet that contains at least the following fields: alert_level: risk level (Level 0, Level 1, Level 2, Level 3); risk_index: the specific RI value; timestamp: warning generation time; alert_id: unique warning identifier.

[0121] Step S1053: Push the early warning message to the terminal of the corresponding safety officer through the park's IoT platform and trigger different levels of audible and visual alarm signals.

[0122] Early warning messages are pushed to the park's IoT platform in real time via message queues (such as MQTT) or industrial Ethernet. The platform then distributes these messages to different responsible terminals based on the warning level, including:

[0123] Medium risk: Push notification to the mobile app of the area's maintenance personnel;

[0124] High risk: Additional notifications will be sent to the control room screen and the shift supervisor's terminal, and audible and visual alarms will be triggered in the hazardous area;

[0125] Emergency Risk: The alert will be sent to the safety manager and emergency command center, and will trigger the highest level of alarm signal in the entire plant area.

[0126] Through step S105, this application completes the conversion from a quantitative risk index to a tiered early warning action instruction, achieving a closed-loop output for monitoring and early warning. However, an advanced early warning system should not only inform users of "there is a risk," but also, as far as possible, inform them "where the risk is" and "what to do," and provide further credibility verification for early warning decisions. To this end, this application can activate a series of enhanced functions simultaneously with or after generating basic early warning information, as detailed below.

[0127] The steps S101 to S105 described above constitute the core process of the comprehensive electrical fire monitoring and early warning method of this application. To further improve the accuracy, operability, and intelligence of the early warning system, this application can integrate a series of extended functions simultaneously with or after generating basic early warning information (step S105). These functions, as enhancements to the core process, together constitute a complete and advanced early warning solution.

[0128] Simply knowing that a risk exists is insufficient; quickly locating the potential hazard is a prerequisite for taking effective countermeasures. Therefore, this application, while generating early warning information, performs the following steps Sa1051 to Sa1053 for hazard location, detailed below:

[0129] Step Sa1051: Backtrack the high-dimensional hidden danger feature vector during the generation process and calculate the cross-attention weights through the attention mechanism.

[0130] In step S1023b, the multi-head attention module calculates the cross-attention weights among the electrical transient feature sequence, the thermal hazard spatial feature sequence, and the multi-gas concentration data feature sequence. These weights form a valuable contribution map, indicating the contribution factors in generating the final result. This involves understanding the correlation of importance between different types of features and between different dimensions within a feature. When the system determines a risk level to be high, tracing back these weights allows us to identify which original features led to the high-risk assessment.

[0131] Step Sa1052: Identify the top K feature dimensions that contribute the most to the current risk index and their corresponding original data sources and sensor locations.

[0132] Specifically, high-dimensional hazard feature vectors can be calculated using gradient backpropagation or feature importance analysis algorithms (e.g., Integrated Gradients). The contribution of each dimension to the final risk index RI Then, attention weights are combined. This allows us to reverse-map out the top K feature dimensions that contribute the most to R. For example, assuming the contributing dimension mainly comes from the fusion feature strongly correlated with the dimension representing the 5th harmonic energy proportion in the electrical feature sequence, and this dimension assigns a high attention weight to the dimension of "normal heat flux density at a connection point in the thermal feature sequence" during fusion, then we can identify the key feature combination: "5th harmonic anomaly" and "poor heat dissipation at the A-phase connector of XX cabinet". Furthermore, through the metadata (sensor-feature-geographic location mapping table) deployed by the system, we can pinpoint the original data source and sensor location corresponding to these feature dimensions, such as "A-phase current sensor on the low-voltage side of transformer No. 1" and "infrared temperature measurement point at the upper left connector of distribution cabinet No. 3".

[0133] Step Sa1053: On the digital twin map of the park, highlight the location of the sensor and its associated electrical lines or equipment to complete the initial location of the potential hazard.

[0134] The campus digital twin map is a 3D visualization model that maps to the physical campus at a 1:1 scale, integrating all electrical equipment, wiring topology, and sensor information. The system highlights the sensor locations identified in step Sa1052 on the digital twin map (e.g., as flashing red icons). Simultaneously, based on electrical connections, it automatically associates and highlights electrical lines or equipment directly connected to the sensor monitoring point (such as a cable or a circuit breaker). This allows maintenance personnel to intuitively and quickly see the precise physical location of suspected potential hazards in a 3D visualization environment, completing the initial hazard localization and significantly shortening the troubleshooting path.

[0135] After locating the hazard, guidance on how to handle it is required. Following the initial hazard identification, this application further provides differentiated handling decision-making suggestions, specifically including the following steps Sb1051 to Sb1053:

[0136] Step Sb1051: Based on the risk level and the feature type that contributes the most, match the corresponding disposal plan from the pre-set disposal knowledge base.

[0137] A pre-defined knowledge base stores rules for risk scenarios and corresponding actions. Scenarios are defined by a combination of risk level and dominant characteristic type. For example:

[0138] Scenario: [High risk, characteristic type: mainly electrical transient characteristics];

[0139] Scenario: [Medium risk, characteristic type: mainly spatial characteristics of thermal hazards];

[0140] Scenario: [Emergency Risk, Feature Type: Combination of Electrical and Gas Features]

[0141] The system matches the most contributing feature type identified by step Sa1052 in the knowledge base based on the current risk level of the warning, and retrieves the corresponding response plan.

[0142] Step Sb1052: Retrieve the corresponding emergency response plan, including: for hazards mainly characterized by electrical transients, it is recommended to conduct harmonic detection and mitigation; for hazards mainly characterized by thermal spatial features, it is recommended to immediately conduct infrared retesting and tightening checks; for hazards mainly characterized by multi-gas concentration features, it is recommended to disconnect the power and conduct insulation testing.

[0143] A contingency plan is a specific, actionable set of instructions. For example, a contingency plan matching "primarily focusing on the spatial characteristics of thermal hazards" might contain the following:

[0144] Immediate Notification: Instruct the nearest security officer to bring a portable infrared thermal imager to location XX.

[0145] Verification check: Compare the temperature measurement results from online monitoring and portable devices, and check the bolt tightening torque.

[0146] Temporary measures: If overheating is confirmed, it is recommended to reduce the load on the circuit and install temporary ventilation.

[0147] Fundamental solution: Plan to shut down the power supply to grind, replace, or tighten the contact surfaces.

[0148] Step Sb1053: Push the matching contingency plan text and illustrated operation guide along with the warning information to the mobile terminal of the operation and maintenance personnel.

[0149] The system packages structured early warning messages, digital twin map location views, and matched emergency response plan texts along with illustrated operation guidelines into a comprehensive alarm work order, which is then pushed to the mobile terminals of relevant maintenance personnel via an app. This allows on-site personnel to obtain all information—"what happened, where, and what to do"—in one stop, significantly improving the efficiency and standardization of emergency response.

[0150] The early warning system needs to have self-verification and continuous learning capabilities to reduce false alarms and optimize performance. That is, after generating early warning information, it also includes early warning verification and model update steps, as detailed in steps Sc1051 to Sc1053 below:

[0151] Step Sc1051: After the warning is issued, start the high-frequency data verification and collection cycle for the warning location.

[0152] Once an early warning is generated, the system automatically adjusts the acquisition strategy of the relevant sensors at the location of the hazard, and enters a high-frequency data verification and acquisition cycle (such as increasing the infrared thermal imaging sampling frequency from 1Hz to 10Hz for 2 minutes) to obtain more dense and detailed subsequent status data.

[0153] Step Sc1052: Input the data collected for review into the hybrid risk assessment model. If the risk index output by the hybrid risk assessment model continuously exceeds the threshold in multiple consecutive review cycles, it is confirmed as a true warning, and the case is recorded in the historical fire case database.

[0154] The reviewed data is input into the model to obtain a series of new RI values. If the risk index output by the hybrid risk assessment model consistently exceeds a threshold across multiple review periods (e.g., RI > 1 for five consecutive periods), the risk assessment will be affected. If the data is negative, the system will determine it as a genuine warning. All multi-source data sequences, features, intermediate model results, and final handling records for this event will be anonymized and recorded as a valuable positive sample in the historical fire case database to enrich the training set for future data-driven models.

[0155] Step Sc1053: If the risk index quickly falls below the threshold, it is determined to be a false alarm. Analyze the cause of the false alarm and use the data to perform incremental learning on the temporal convolutional network model to optimize its discrimination boundary.

[0156] If the risk index rapidly falls below the threshold (for example, if the RI drops to a certain level in the next review period), If the following occurs, it is very likely a false alarm. The system will trigger a false alarm analysis process, automatically checking data quality (e.g., a sudden drop in SNR), external events (e.g., the start-up or shutdown of large equipment), etc. More importantly, this false alarm data will be labeled and used for incremental learning of the temporal convolutional network model. Through online learning or periodic retraining, the model can correct its judgment of such "similar anomalies but actually safe" patterns, optimize its discrimination boundaries, thereby reducing the probability of similar false alarms in the future and achieving system self-optimization.

[0157] Extended Function 4: In-depth Verification of Early Warning Based on Physical Simulation

[0158] For warnings of extremely high risks, in order to further enhance the credibility of decision-making, this application introduces an independent verification step based on a high-fidelity physical model, namely, the process of physically simulating and verifying the warning information in steps Sd1051 to Sd1056, which is described in detail below:

[0159] Step Sd1051: After generating the early warning information, immediately obtain the electrical circuit topology, equipment parameters and real-time load conditions associated with the current early warning location from the digital twin model of the park's power supply network.

[0160] The digital twin model of the park's power supply network is a virtual system that includes precise electrical connections (electrical circuit topology), equipment nameplate data and material properties (equipment parameters), and real-time measurement point data (real-time load conditions). When a high-risk warning is generated, the system automatically retrieves all relevant information associated with the current warning location from this model to prepare input conditions for physical simulation.

[0161] Step Sd1052: Construct a finite element simulation model based on the electrical circuit topology and equipment parameters, and use the real-time load conditions and high-dimensional hidden danger feature vectors as the boundary conditions and initial conditions of the finite element simulation model.

[0162] Based on the acquired topology and parameters, the system automatically invokes or generates a targeted finite element simulation model (e.g., 3D thermo-electric coupling modeling of the connection point and its surrounding structure for early warning). Real-time load conditions (e.g., current values) serve as the boundary conditions (heat sources) for the simulation. The current temperature, heat flow characteristics, and other parameters extracted are used as the initial conditions for the simulation. This allows the simulation to begin from a "snapshot" state that is highly consistent with the real world.

[0163] Step Sd1053: Run the finite element simulation model to calculate the evolution of the temperature field and thermal stress field at the warning location within a set time period under the boundary and initial conditions.

[0164] On the cloud or high-performance computing nodes, a finite element simulation model is run to solve a system of partial differential equations, calculating how the temperature at the warning point will change with time and space (temperature field evolution) over a future period (e.g., the next 30 minutes), and how the resulting thermal expansion stress will be distributed (thermal stress field evolution). This is an independent physical prediction based on first principles.

[0165] Step Sd1054: Extract the maximum values ​​of the temperature field and thermal stress field during the evolution process.

[0166] Step Sd1055: If any maximum value exceeds its corresponding material safety threshold, a physical simulation confirmation mark is added to the warning information, and an enhanced warning report containing the simulation prediction trend is generated.

[0167] Step Sd1055: If any maximum value does not exceed its corresponding material safety threshold, then add a "requires on-site verification" prompt mark to the warning information.

[0168] The highest temperature during the entire evolution process was extracted from the simulation results. and maximum thermal stress The system compares this value with safety thresholds for materials (e.g., the softening temperature of copper, the heat resistance rating of insulation materials). If any maximum value exceeds the threshold, it indicates that the risk is real and imminent, even from a rigorous physical simulation perspective. The system adds a physical simulation confirmation flag to the original warning information and generates an enhanced warning report containing simulation curves and predicted deterioration time points, greatly enhancing the authority and urgency of the warning. If the threshold is not exceeded, it indicates that although the current state is judged as abnormal by the data model, it has not yet reached the critical point of physical failure. The system adds a "requires on-site verification" flag, prompting maintenance personnel to prioritize manual inspection rather than immediately taking extreme measures. This constitutes another line of defense for safety and reliability on top of the algorithmic warning system.

[0169] For high-level warnings, it is necessary to activate a cross-departmental and cross-system emergency response, namely... Figure 1The example method can be described in detail in steps Se1051 to Se1054 below.

[0170] Step Se1051: When generating early warning information of medium risk or above for the first time, automatically trigger the data verification process for associated sensors and redundant backup sensors to confirm the validity of the data.

[0171] To prevent false alarms caused by a single sensor failure, the system does not immediately activate a system-wide alarm when a medium-risk or higher warning is triggered for the first time. Instead, it automatically triggers a data verification process. For example, it retrieves data from adjacent or similar redundant sensors for comparison, checks the sensor self-test status codes, and verifies the physical validity of the data (such as whether the temperature is within the possible range).

[0172] Step Se1052: If the data verification passes, a primary alarm is immediately sent to the security personnel at the first response level, and the video surveillance device closest to the alarm location is automatically activated to push the real-time image to the monitoring center.

[0173] If the data verification passes, the data is confirmed to be reliable. The system immediately sends a primary alarm to the terminal of the first-response level security personnel (such as area inspectors). At the same time, it automatically activates the video surveillance equipment (PTZ cameras) near the potential hazard point (based on the location on the digital twin map), controls them to turn and focus on the warning location, and pushes the real-time image to the monitoring center to provide visual evidence for remote confirmation.

[0174] Step Se1053: After a preset short delay, the comprehensive alarm information, which includes early warning information, physical simulation verification results and real-time monitoring images, as well as the emergency resource demand list generated based on hazard location information and response plan, will be pushed to the emergency command center.

[0175] After a short delay (e.g., 30 seconds to complete initial verification and video retrieval), the system integrates all information into a comprehensive alarm message and, based on the contingency plan, automatically analyzes and generates an emergency resource requirement list (e.g., requiring 2 electrical repairmen, 1 infrared thermometer, and possibly 1 insulated bucket truck). This information is then pushed to the decision-making screen in the emergency command center.

[0176] Step Se1054: The emergency command center automatically generates and issues the optimal dispatch instructions based on the emergency resource demand list and the real-time status of emergency resources (including personnel, equipment, and fire lanes) within the park.

[0177] The intelligent dispatch system of the emergency command center, based on the emergency resource demand list and combined with the real-time status of emergency resources in the park (such as personnel GPS location, equipment availability, and fire lane access), automatically generates and issues the optimal dispatch instructions through path planning and resource allocation algorithms, directly dispatching them to relevant personnel and vehicle terminals, thereby achieving rapid, accurate, and automated collaborative emergency response.

[0178] From the above appendix Figure 1The example of a comprehensive monitoring and early warning method for electrical fires in industrial parks demonstrates that, on the one hand, by simultaneously collecting heterogeneous data from multiple sources such as electrical, temperature, and gas sources, and processing it using an attention-based fusion method, data from different physical sources can be effectively integrated. This results in a high-dimensional hazard feature vector that not only includes various independent anomaly information but also dynamically captures the intrinsic correlations and spatiotemporal coupling relationships between electrical, thermal, and chemical states through attention weights. This constructs a comprehensive feature vector that is more sensitive to complex coupled faults and has stronger characterization capabilities, laying the foundation for subsequent accurate assessment. On the other hand, by employing a hybrid risk assessment model that includes both physical law verification and historical data-driven approaches, the model uses physical laws (e.g., circuit thermodynamic equations) to verify the consistency of observed data and assess the degree to which it violates known physical laws. This enhances the interpretability of the assessment results and the ability to infer in areas lacking data. Furthermore, by mining complex patterns in historical data through a data-driven model, this parallel processing approach can obtain physical inconsistency scores and data-driven hazard probability scores. This allows for cross-validation and comprehensive judgment of the current state from both physical laws and historical experience dimensions, making it more effective than a single model or rule. The system can accurately identify real hidden dangers, reducing false alarms and missed alarms. Thirdly, by introducing an adaptive weighted fusion mechanism, the weights of physical inconsistency scores and data-driven hazard probability scores in the final decision are dynamically adjusted based on the signal-to-noise ratio and feature uncertainty of real-time data. This enables the system to intelligently cope with various situations in complex industrial sites. Its dynamic adjustment strategy allows the entire early warning system to maintain stable early warning performance when facing challenges such as sensor errors, environmental interference, and new working conditions, exhibiting stronger environmental adaptability and robustness. Fourthly, starting from multi-source data acquisition, through feature intelligent fusion, hybrid model risk assessment, and adaptive decision-making, a graded early warning information is finally generated, forming a complete technical closed loop. This not only outputs risk levels but also inherently combines physical laws, data intelligence, and dynamic decision-making throughout the entire technical path. This makes early warning decisions no longer simple threshold comparisons but a complex decision-making process based on multi-dimensional, adaptive analysis. This helps park safety management personnel obtain earlier and more reliable early warning signals before a fire occurs and understand the possible causes behind the risk, thus gaining valuable time to take precise and effective preventative measures and improving the initiative and intelligence level of safety management. In summary, the technical solution of this application significantly improves the accuracy and reliability of electrical fire early warning through intelligent fusion of multi-source data and adaptive hybrid evaluation.

[0179] Please see the appendix Figure 2 This application provides a comprehensive electrical fire monitoring and early warning device for industrial parks. The device may include a data acquisition module 201, a feature extraction module 202, an acquisition module 203, a fusion module 204, and a generation module 205, as detailed below:

[0180] The acquisition module 201 is used to synchronously acquire multi-source data of the power supply network of the industrial park. The multi-source data includes three-phase current / voltage time sequence data of key nodes of the power supply network, infrared thermal image sequences of cable joints and key locations of distribution cabinets, and multi-gas concentration data inside the distribution box.

[0181] The feature extraction module 202 is used to process the multi-source data of the power supply network of the industrial park to extract electrical transient features and thermal hazard spatial features respectively. The electrical transient features, thermal hazard spatial features and multi-gas concentration data are fused with attention mechanism to output a high-dimensional hazard feature vector that comprehensively reflects the electrical, thermal and chemical states.

[0182] The acquisition module 203 is used to input the high-dimensional hazard feature vector in parallel into two evaluation branches of the hybrid risk assessment model to obtain the physical inconsistency score and the data-driven hazard probability score, respectively. The two evaluation branches include an evaluation branch based on physical law verification and an evaluation branch based on historical data.

[0183] The fusion module 204 is used to dynamically adjust the weights of the physical inconsistency score and the data-driven hazard probability score based on the signal-to-noise ratio and feature uncertainty of the real-time data through an adaptive weighted fusion mechanism, and then fuse them to generate a dynamic electrical fire risk index.

[0184] The generation module 205 is used to generate early warning information of the corresponding level based on the threshold range of the dynamic electrical fire risk index.

[0185] From the above appendix Figure 2The example of an integrated monitoring and early warning device for electrical fires in industrial parks demonstrates that, on the one hand, by simultaneously collecting heterogeneous data from multiple sources such as electrical, temperature, and gas sources, and processing it using an attention-based fusion method, data from different physical sources can be effectively integrated. This results in a high-dimensional hazard feature vector that not only includes various independent anomaly information but also dynamically captures the intrinsic correlations and spatiotemporal coupling relationships between electrical, thermal, and chemical states through attention weights. This constructs a comprehensive feature vector that is more sensitive to complex coupled faults and has stronger characterization capabilities, laying the foundation for subsequent accurate assessment. On the other hand, by employing a hybrid risk assessment model that includes both physical law verification and historical data-driven approaches, the model uses physical laws (e.g., circuit thermodynamic equations) to verify the consistency of observed data and assess the degree to which it violates known physical laws. This enhances the interpretability of the assessment results and the ability to infer in areas lacking data. Furthermore, by mining complex patterns in historical data through a data-driven model, this parallel processing approach can obtain physical inconsistency scores and data-driven hazard probability scores. This allows for cross-validation and comprehensive judgment of the current state from both physical laws and historical experience dimensions, making it more effective than a single model or rule. The system can accurately identify real hidden dangers, reducing false alarms and missed alarms. Thirdly, by introducing an adaptive weighted fusion mechanism, the weights of physical inconsistency scores and data-driven hazard probability scores in the final decision are dynamically adjusted based on the signal-to-noise ratio and feature uncertainty of real-time data. This enables the system to intelligently cope with various situations in complex industrial sites. Its dynamic adjustment strategy allows the entire early warning system to maintain stable early warning performance when facing challenges such as sensor errors, environmental interference, and new working conditions, exhibiting stronger environmental adaptability and robustness. Fourthly, starting from multi-source data acquisition, through feature intelligent fusion, hybrid model risk assessment, and adaptive decision-making, a graded early warning information is finally generated, forming a complete technical closed loop. This not only outputs risk levels but also inherently combines physical laws, data intelligence, and dynamic decision-making throughout the entire technical path. This makes early warning decisions no longer simple threshold comparisons but a complex decision-making process based on multi-dimensional, adaptive analysis. This helps park safety management personnel obtain earlier and more reliable early warning signals before a fire occurs and understand the possible causes behind the risk, thus gaining valuable time to take precise and effective preventative measures and improving the initiative and intelligence level of safety management. In summary, the technical solution of this application significantly improves the accuracy and reliability of electrical fire early warning through intelligent fusion of multi-source data and adaptive hybrid evaluation.

[0186] Figure 3 This is a schematic diagram of the structure of a device provided in one embodiment of this application. For example... Figure 3As shown, the device 3 in this embodiment mainly includes: a processor 30, a memory 31, and a computer program 32 stored in the memory 31 and executable on the processor 30, such as a program for a comprehensive electrical fire monitoring and early warning method for industrial parks. When the processor 30 executes the computer program 32, it implements the steps described in the above embodiment of the comprehensive electrical fire monitoring and early warning method for industrial parks, for example... Figure 1 The steps S101 to S105 are shown. Alternatively, when the processor 30 executes the computer program 32, it implements the functions of each module / unit in the above-described device embodiments, for example... Figure 2 The functions of the acquisition module 201, feature extraction module 202, acquisition module 203, fusion module 204, and generation module 205 are shown.

[0187] For example, the computer program 32 of the comprehensive monitoring and early warning method for electrical fires in industrial parks mainly includes: synchronously collecting multi-source data from the power supply network of the industrial park, wherein the multi-source data includes three-phase current / voltage time-series data of key nodes of the power supply network, infrared thermal image sequences of cable joints and key locations of distribution cabinets, and multi-gas concentration data inside distribution boxes; processing the collected multi-source data of the power supply network of the industrial park to extract electrical transient features and thermal hazard spatial features respectively, and fusing the electrical transient features, thermal hazard spatial features and multi-gas concentration data with attention mechanism to output a high-resolution image that comprehensively reflects the electrical, thermal and chemical states. A high-dimensional hazard feature vector is generated; this high-dimensional hazard feature vector is input in parallel into two evaluation branches of the hybrid risk assessment model to obtain a physical inconsistency score and a data-driven hazard probability score, respectively. The two evaluation branches include one evaluation branch based on physical law verification and one evaluation branch based on historical data. Through an adaptive weighted fusion mechanism, the weights of the physical inconsistency score and the data-driven hazard probability score are dynamically adjusted based on the signal-to-noise ratio and feature uncertainty of the real-time data, and then fused to generate a dynamic electrical fire risk index. Based on the threshold range of the dynamic electrical fire risk index, a corresponding level of early warning information is generated. The computer program 32 can be divided into one or more modules / units, which are stored in memory 31 and executed by processor 30 to complete this application. One or more modules / units can be a series of computer program instruction segments capable of performing specific functions, which describe the execution process of the computer program 32 in the device 3.For example, computer program 32 can be divided into the functions of acquisition module 201, feature extraction module 202, acquisition module 203, fusion module 204, and generation module 205 (a module in the virtual device). The specific functions of each module are as follows: Acquisition module 201 is used to synchronously acquire multi-source data of the power supply network of the industrial park. The multi-source data includes three-phase current / voltage time series data of key nodes of the power supply network, infrared thermal image sequences of cable joints and key locations of distribution cabinets, and multi-gas concentration data inside the distribution box; Feature extraction module 202 is used to process the acquired multi-source data of the power supply network of the industrial park to extract electrical transient features and thermal hazard spatial features respectively, and to fuse the electrical transient features, thermal hazard spatial features, and multi-gas concentration data with attention mechanism. The system outputs a high-dimensional hazard feature vector that comprehensively reflects the electrical, thermal, and chemical conditions. The acquisition module 203 inputs the high-dimensional hazard feature vector in parallel into two evaluation branches of the hybrid risk assessment model, obtaining a physical inconsistency score and a data-driven hazard probability score, respectively. Each evaluation branch includes one based on physical law verification and one based on historical data. The fusion module 204 dynamically adjusts the weights of the physical inconsistency score and the data-driven hazard probability score based on the signal-to-noise ratio and feature uncertainty of the real-time data using an adaptive weighted fusion mechanism, and then fuses them to generate a dynamic electrical fire risk index. The generation module 205 generates corresponding warning information based on the threshold range of the dynamic electrical fire risk index.

[0188] Device 3 may include, but is not limited to, processor 30 and memory 31. Those skilled in the art will understand that... Figure 3 This is merely an example of device 3 and does not constitute a limitation on device 3. It may include more or fewer components than shown, or combine certain components, or different components. For example, the device may also include input / output devices, network access devices, buses, etc.

[0189] The processor 30 may be a Central Processing Unit (CPU), or other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. A general-purpose processor may be a microprocessor or any conventional processor.

[0190] The memory 31 can be an internal storage unit of the device 3, such as a hard disk or RAM of the device 3. The memory 31 can also be an external storage device of the device 3, such as a plug-in hard disk, Smart MediaCard (SMC), Secure Digital (SD) card, or Flash Card equipped on the device 3. Furthermore, the memory 31 can include both internal and external storage units of the device 3. The memory 31 is used to store computer programs and other programs and data required by the device. The memory 31 can also be used to temporarily store data that has been output or will be output.

[0191] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the above-described division of functional units and modules is merely an example. In practical applications, the above functions can be assigned to different functional units and modules as needed. That is, the internal structure of the device can be divided into different functional units or modules to complete all or part of the functions described above. The functional units and modules in the embodiments can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit. Furthermore, the specific names of the functional units and modules are only for easy differentiation and are not intended to limit the scope of protection of this application. The specific working process of the units and modules in the above-described device can be referred to the corresponding process in the foregoing method embodiments, and will not be repeated here.

[0192] In the above embodiments, the descriptions of each embodiment have different focuses. For parts that are not described in detail or recorded in a certain embodiment, please refer to the relevant descriptions of other embodiments.

[0193] Those skilled in the art will recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0194] In the embodiments provided in this application, it should be understood that the disclosed apparatus / device and method can be implemented in other ways. For example, the apparatus / device embodiments described above are merely illustrative. For instance, the division of modules or units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another device, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be through some interfaces; the indirect coupling or communication connection between devices or units may be electrical, mechanical, or other forms.

[0195] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0196] Furthermore, the functional units in the various embodiments of this application can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit.

[0197] If integrated modules / units are implemented as software functional units and sold or used as independent products, they can be stored in a storage medium. Based on this understanding, all or part of the processes in the methods of the above embodiments can also be implemented by a computer program instructing related hardware. The computer program for the comprehensive monitoring and early warning method for electrical fires in industrial parks can be stored in a storage medium. When the computer program is executed by a processor, it can implement the steps of the above method embodiments, namely, synchronously collecting multi-source data of the power supply network of the industrial park, wherein the multi-source data includes three-phase current / voltage time series data of key nodes of the power supply network, infrared thermal image sequences of cable joints and key locations of distribution cabinets, and multi-gas concentration data inside the distribution box; processing the collected multi-source data of the power supply network of the industrial park to extract electrical transient features and thermal hazard spatial features respectively, and combining the electrical transient features and thermal hazard spatial features. Spatial features and multi-gas concentration data are fused using an attention mechanism to output a high-dimensional hazard feature vector that comprehensively reflects electrical, thermal, and chemical states. This high-dimensional hazard feature vector is then input in parallel into two evaluation branches of a hybrid risk assessment model, yielding a physical inconsistency score and a data-driven hazard probability score. Each evaluation branch includes one based on physical law verification and one based on historical data. An adaptive weighted fusion mechanism dynamically adjusts the weights of the physical inconsistency score and the data-driven hazard probability score based on the signal-to-noise ratio and feature uncertainty of the real-time data, and then fuses them to generate a dynamic electrical fire risk index. Based on the threshold range of the dynamic electrical fire risk index, corresponding warning information is generated. The computer program includes computer program code, which can be in the form of source code, object code, executable files, or some intermediate form. The storage medium can include any entity or device capable of carrying computer program code, recording media, USB flash drives, external hard drives, magnetic disks, optical disks, computer memory, read-only memory (ROM), random access memory (RAM), electrical carrier signals, telecommunication signals, and software distribution media, etc. It should be noted that the contents of the storage medium may be appropriately added to or subtracted from the contents according to the requirements of legislation and patent practice in the jurisdiction. For example, in some jurisdictions, according to legislation and patent practice, the storage medium may not include electrical carrier signals and telecommunication signals.

[0198] The above embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit them. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. These modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of this application, and should all be included within the protection scope of this application. The specific embodiments described above further illustrate the purpose, technical solutions, and beneficial effects of this application. It should be understood that the above descriptions are merely specific embodiments of this application and are not intended to limit the protection scope of this application. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the protection scope of this application.

Claims

1. A comprehensive monitoring and early warning method for electrical fires in industrial parks, characterized in that, The method includes: Simultaneously collect multi-source data of the power supply network of the industrial park. The multi-source data includes three-phase current / voltage time sequence data of key nodes of the power supply network, infrared thermal image sequences of cable joints and key locations of distribution cabinets, and multi-gas concentration data inside the distribution box. The multi-source data is processed to extract electrical transient features and thermal hazard spatial features, and the electrical transient features, thermal hazard spatial features and multi-gas concentration data are fused with the data based on an attention mechanism to output a high-dimensional hazard feature vector that comprehensively reflects the electrical, thermal and chemical states. The high-dimensional hazard feature vector is input in parallel into two evaluation branches of the hybrid risk assessment model to obtain a physical inconsistency score and a data-driven hazard probability score, respectively. The two evaluation branches include an evaluation branch based on physical law verification and an evaluation branch based on historical data. Through an adaptive weighted fusion mechanism, the weights of the physical inconsistency score and the data-driven hazard probability score are dynamically adjusted based on the signal-to-noise ratio and feature uncertainty of real-time data, and then fused to generate a dynamic electrical fire risk index. Based on the threshold range of the dynamic electrical fire risk index, a corresponding level of early warning information is generated.

2. The comprehensive monitoring and early warning method for electrical fires in industrial parks according to claim 1, characterized in that, The process of fusing the electrical transient features, thermal hazard spatial features, and multi-gas concentration data based on an attention mechanism to output a high-dimensional hazard feature vector that comprehensively reflects the electrical, thermal, and chemical states includes: The electrical transient features, thermal hazard spatial features, and multi-gas concentration data are respectively mapped to the same high-dimensional feature space to form corresponding feature sequences; The feature sequence is input into the multi-head attention module to calculate the cross-attention weights among the electrical transient feature sequence, the thermal hazard spatial feature sequence, and the multi-gas concentration data feature sequence. Based on the cross-attention weights, the three feature sequences are weighted and summed and interacted with to generate a context-aware fusion feature vector as the high-dimensional hidden danger feature vector.

3. The comprehensive monitoring and early warning method for electrical fires in industrial parks according to claim 1, characterized in that, The evaluation branch based on physical law verification obtains its physical inconsistency score by including: A physical information neural network is constructed to describe the relationship between temperature rise at circuit connection points and current, contact resistance, and heat dissipation conditions. The loss function of the network includes the residual constraints of the circuit overheating differential equation. The feature subset containing the thermally related part of the electrical transient features and the high-dimensional hidden danger feature vector is input into the physical information neural network; The physical information neural network performs forward calculations to obtain the predicted temperature rise curve, and calculates the residual between the predicted result and the physical law described by the circuit overheating differential equation. The norm of the residual is normalized and used as the physical inconsistency score.

4. The comprehensive monitoring and early warning method for electrical fires in industrial parks according to claim 3, characterized in that, The process of obtaining the physical inconsistency score also includes the following evaluation process for the confidence of the physical model: The distribution density of the feature subset of the neural network that monitors the input physical information within the domain of the circuit overheating differential equation; When the feature subset is located in a high-density region of the historical training data distribution, the confidence level of the physical model is determined to be high; otherwise, it is determined to be low. The confidence level of the physical model is used as a confidence correction factor for the physical inconsistency score, and is then used for subsequent adaptive weighted fusion.

5. The comprehensive monitoring and early warning method for electrical fires in industrial parks according to claim 1, characterized in that, The historical data-driven assessment branch obtains a data-driven hazard probability score including: The high-dimensional hidden danger feature vector is input into a temporal convolutional network model, which extracts long-term dependencies in the feature sequence through dilated causal convolution. The temporal convolutional network model outputs a matching score that represents the degree of matching between the current multi-source data state and the features of historical electrical fire cases; The matching score is mapped to a probability value between 0 and 1 via a Sigmoid function and used as the data-driven hazard probability score.

6. The comprehensive monitoring and early warning method for electrical fires in industrial parks according to claim 1, characterized in that, The adaptive weighted fusion mechanism dynamically adjusts the weights of the physical inconsistency score and the data-driven hazard probability score based on the signal-to-noise ratio and feature uncertainty of real-time data, and then fuses them to generate a dynamic electrical fire risk index, including: The signal-to-noise ratio of the three-phase current / voltage time-series data is calculated in real time, and the uncertainty of each component in the high-dimensional hidden danger feature vector is monitored. The uncertainty is estimated by the variance of the model prediction or by Dropout technique. Set a weighting factor for physical scoring and a weighting factor for data scoring, with the sum of the two being 1; When the signal-to-noise ratio is lower than the first threshold, or the feature uncertainty is higher than the second threshold, the value of the physical scoring weight factor is increased, and the value of the data scoring weight factor is decreased. Based on the adjusted weights, the physical inconsistency score and the data-driven hazard probability score are weighted and summed to generate the dynamic electrical fire risk index.

7. The comprehensive monitoring and early warning method for electrical fires in industrial parks according to claim 1, characterized in that, After generating the warning information, the process also includes the following physical simulation verification of the warning information: After generating the early warning information, the electrical circuit topology, equipment parameters and real-time load conditions associated with the current early warning location are immediately obtained from the digital twin model of the park's power supply network. A finite element simulation model is constructed based on the electrical circuit topology and the equipment parameters, and the real-time load conditions and the high-dimensional hidden danger feature vector are used as the boundary conditions and initial conditions of the finite element simulation model. Run the finite element simulation model to calculate the evolution of the temperature field and thermal stress field at the warning location within a set time period under the given boundary and initial conditions. Extract the maximum values ​​of the temperature field and thermal stress field during the evolution process; If any of the aforementioned maximum values ​​exceeds its corresponding material safety threshold, a physical simulation confirmation mark is added to the warning information, and an enhanced warning report containing the simulation prediction trend is generated. If any of the maximum values ​​does not exceed its corresponding material safety threshold, a "requires on-site verification" prompt mark will be added to the warning information.

8. A comprehensive electrical fire monitoring and early warning device for industrial parks, characterized in that, The device includes: The acquisition module is used to synchronously acquire multi-source data of the power supply network of the industrial park. The multi-source data includes three-phase current / voltage time sequence data of key nodes of the power supply network, infrared thermal image sequences of cable joints and key locations of distribution cabinets, and multi-gas concentration data inside the distribution box. The feature extraction module is used to process the multi-source data to extract electrical transient features and thermal hazard spatial features respectively, and to fuse the electrical transient features and thermal hazard spatial features with the multi-gas concentration data based on an attention mechanism to output a high-dimensional hazard feature vector that comprehensively reflects the electrical, thermal and chemical states. The acquisition module is used to input the high-dimensional hazard feature vector in parallel into two evaluation branches of the hybrid risk assessment model to obtain a physical inconsistency score and a data-driven hazard probability score, respectively. The two evaluation branches include an evaluation branch based on physical law verification and an evaluation branch based on historical data. The fusion module is used to dynamically adjust the weights of the physical inconsistency score and the data-driven hazard probability score based on the signal-to-noise ratio and feature uncertainty of real-time data through an adaptive weighted fusion mechanism, and then fuse them to generate a dynamic electrical fire risk index. The generation module is used to generate early warning information of the corresponding level based on the threshold range of the dynamic electrical fire risk index.

9. An apparatus comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the computer program, it implements the steps of the method as described in any one of claims 1 to 7.

10. A storage medium storing a computer program, characterized in that, When the computer program is executed by a processor, it implements the steps of the method as described in any one of claims 1 to 7.

Images