An electrical fire hazard detection and early warning method based on intelligent sensor

Abstract

The application discloses an electrical fire hazard detection and early warning method based on an intelligent sensor, relates to the technical field of intelligent sensors, and comprises the following steps: collecting data, generating a sensing data set and a loop structure table; mapping channels and arranging multi-source hazard time sequence samples; extracting an electrical fire hazard time-frequency feature set based on an improved WFTNet model; separating a zero sequence sine trajectory by adopting an SGMD algorithm, and recalibrating an electrical heating sine characteristic value; constructing a component-loop matrix and generating a hazard index; calculating component responses and generating a hazard discrimination matrix; and mapping early warning levels and disposal objects to output results. By introducing the improved WFTNet model and the SGMD algorithm, the electrical fire hazard time-frequency identification, mode decomposition, loop positioning and hierarchical early warning of multi-source sensing data of an electrical loop under complex electrical working conditions are realized.

Description

Technical Field

[0001] This invention relates to the field of intelligent sensor technology, and in particular to a method for detecting and warning of electrical fire hazards based on intelligent sensors. Background Technology

[0002] With the development of intelligent sensors, IoT communication, edge computing, and electrical safety monitoring technologies, the detection of electrical fire hazards in building power distribution systems, industrial distribution cabinets, low-voltage power lines, and electrical equipment is gradually becoming online. Existing electrical fire monitoring methods typically involve installing current sensors, residual current sensors, temperature sensors, arc detectors, and ambient temperature and humidity sensors at distribution boxes, switch circuits, terminal connection points, cable laying areas, and critical electrical equipment. These sensors collect data such as phase current, neutral current, residual current, voltage, load power, terminal temperature rise, cable temperature, arc disturbance, ambient temperature and humidity, and switch status. Then, through threshold comparison, trend analysis, rule-based judgment, or machine learning models, they identify risks such as leakage current, overload, poor contact, insulation aging, arc faults, and localized overheating. These methods enable continuous data acquisition, remote alarm, and multi-point coverage, making them highly valuable for the safety management of power distribution facilities.

[0003] Existing methods for detecting and warning of electrical fire hazards still have shortcomings. Some methods rely on single sensor parameters or fixed threshold alarms, making it difficult to distinguish between normal load cycles, equipment start-up and shutdown disturbances, and actual fire hazards. They are prone to misjudging short-term load fluctuations, changes in ambient temperature and humidity, or equipment start-up and shutdown processes as abnormalities. Arc disturbances are short-term, intermittent, and discontinuous. Ordinary time-domain analysis or simple filtering can easily weaken arc extinguishing segments, leading to missed detection of early arc hazards. Existing methods do not adequately utilize the correlation between multi-source sensor data and electrical circuit structure, making it difficult to map zero-sequence current changes, temperature rise lag, arc pulses, and load reference deviations to specific circuits, branches, terminals, and electrical equipment, resulting in insufficient accuracy in locating the source of hazards. Existing deep learning methods typically classify or predict sensor sequences directly, lacking features such as normal circuit spectral library memory, arc extinguishing segment cache fidelity, prior generation of hazard components, zero-sequence current symplectic trajectory rotation separation, and electrothermal residual symplectic eigenvalue recalibration. This makes it difficult to form a continuous technical processing chain from multi-source sensor data to early warning results, affecting the accuracy of early hazard identification, anti-interference capability under complex working conditions, hazard object location capability, and the manageability of early warning results.

[0004] Therefore, how to provide a method for detecting and warning of electrical fire hazards based on intelligent sensors is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0005] One objective of this invention is to propose a method for detecting and warning of electrical fire hazards based on intelligent sensors. This invention collects multi-source sensor data and electrical circuit configuration data related to electrical fires, and combines an improved WFTNet model and SGMD algorithm to analyze and process hazard characteristics such as leakage current changes, arc disturbances, contact overheating, and overload temperature rise during electrical circuit operation. This forms a complete processing flow from multi-source sensor data preprocessing, construction of multi-source hazard time-series samples, extraction of time-frequency features of electrical fire hazards, decomposition of hazard modal components, to the generation of electrical fire hazard detection and warning results. Through this method, stable hazard characteristics can be extracted under the influence of normal load fluctuations, equipment start-up and shutdown disturbances, and environmental changes, improving the accuracy of early electrical fire hazard identification, short-term arc disturbance capture capability, zero-sequence current anomaly separation capability, contact thermal instability identification capability, and hazard object location accuracy.

[0006] An electrical fire hazard detection and early warning method based on intelligent sensors according to an embodiment of the present invention includes: Collect multi-source sensor data and electrical circuit configuration data for electrical fires, preprocess the multi-source sensor data for electrical fires to generate a standardized electrical fire sensor dataset, and generate an electrical circuit structure table based on the electrical circuit configuration data; Based on the electrical circuit structure table, the standardized electrical fire sensor dataset is mapped to circuit objects, associated with sensor channels, and arranged with time windows to generate time-series samples of multi-source hazards. Based on the improved WFTNet model, wavelet Fourier time-frequency transform is performed on the time series samples of multi-source hidden dangers. Load spectrum residuals are extracted by combining the memory reconstruction of the normal circuit spectrum library. Local pulse preservation is achieved by using the arc extinguishing segment buffer fidelity. The time-frequency features of zero sequence, electrothermal, arc and load reference direction are generated and sorted through the prior of hidden danger components to obtain the time-frequency feature set of electrical fire hidden dangers. The SGMD algorithm is used to perform symplectic geometric mode decomposition on the time-frequency feature set of electrical fire hazards. Based on the rotational separation process of zero-sequence current symplectic trajectory, the zero-sequence related symplectic trajectory is split. The electrothermal residual symplectic eigenvalue recalibration process is introduced to correct the electrothermal related symplectic eigenvalue, and the set of modal components of electrical fire hazards is obtained. Based on the set of electrical fire hazard modal components and the electrical circuit structure table, a component-circuit correspondence matrix is ​​constructed and time-series alignment and adjacency transfer are performed to obtain the circuit hazard index matrix; The hazard component response of each loop object is calculated based on the loop hazard index matrix, and a hazard object discrimination matrix is ​​generated according to the combination relationship of the hazard component responses. Based on the hazard object identification matrix, the warning level mapping and disposal object matching are performed to generate electrical fire hazard detection and warning results.

[0007] Optionally, the multi-source sensing data for electrical fires specifically includes phase current data, neutral current data, residual current data, voltage data, load power data, power factor data, harmonic data, arc disturbance data, current spike data, voltage drop data, terminal temperature rise data, cable temperature data, distribution box internal temperature data, ambient temperature data, ambient humidity data, switch status data, and electrical equipment start / stop status data.

[0008] Optionally, the electrical circuit configuration data specifically includes circuit number, branch number, terminal number, switch number, circuit breaker number, electrical equipment number, circuit topology connection relationship, branch power supply relationship, terminal connection relationship, phase line connection relationship, neutral line connection relationship, grounding connection relationship, rated voltage, rated current, rated power, wire specifications, protection device parameters, sensor installation location, and sensor corresponding circuit relationship.

[0009] Optionally, generating the standardized electrical fire sensor dataset and electrical circuit structure table includes: Collect multi-source sensor data on electrical fires and electrical circuit configuration data, and establish original data association records according to sensor number, acquisition time, sensor type and circuit number; The original data association records of electrical fire multi-source sensor data are processed by deleting duplicate records, removing invalid data, filling missing samples, and correcting abnormal jumps to generate cleaned sensor data. The cleaned sensor data is time-aligned according to a unified sampling time axis, and the data of different sensor types are dimensionally unified and numerically standardized to generate a standardized electrical fire sensor dataset. Based on the object number, connection relationship and sensor installation location in the electrical circuit configuration data, establish structured associations between circuit objects, branch objects, terminal objects, switch objects, electrical equipment objects and sensor objects, and generate an electrical circuit structure table.

[0010] Optionally, generating time-series samples of multi-source hidden dangers includes: Based on the electrical circuit structure table, a circuit object index table is established, and the standardized electrical fire sensing dataset is mapped to the corresponding circuit object according to the sensor number and the corresponding circuit relationship of the sensor, thus generating a circuit object sensing data table. According to the sensing type, the sensing data table of the loop object is associated with sensing channels to generate current channel, voltage channel, residual current channel, temperature channel, arc disturbance channel, environment channel and status channel; Arrange the time windows of each sensing channel according to a unified sampling time axis to generate a loop sensing time window sequence. Based on the loop object index table, loop sensing time window sequence, and electrical loop structure table, establish sample association relationships for adjacent loops, the same branch, the same terminal, and the same electrical equipment, and generate multi-source hidden danger time sequence samples.

[0011] Optionally, the obtained time-frequency feature set of electrical fire hazards includes: An improved WFTNet model is constructed by setting the normal circuit spectrum library memory reconstruction layer on the Fourier periodic feature branch, setting the arc extinguishing segment cache fidelity layer on the wavelet local frequency band feature branch, and setting the hidden danger component prior generation layer after the Fourier periodic feature branch and the wavelet local frequency band feature branch. The normal loop spectrum library memory reconstruction layer archives the Fourier periodic features of multi-source hidden danger time series samples according to the loop object and normal operation window identifier, generates the normal loop spectrum library, aligns the Fourier periodic features of the current monitoring window with the normal loop spectrum library, extracts the frequency segments and durations that deviate from the normal operation spectrum of the current monitoring window, and performs residual backfilling according to the periodic position to generate the load spectrum residual representation. The arc extinction segment cache fidelity layer performs short-time pulse search on the wavelet local frequency band features of the multi-source hidden danger time series sample, extracts arc candidate wave packets, and caches and groups the arc candidate wave packets according to the arc initiation segment, arc extinction interval segment, and arc re-ignition segment. According to the equipment start-up and shutdown status, the start-up and shutdown disturbance segments are written into the bypass channel and separated from the cache group. The arc candidate wave packets in the cache group are continuously filled to generate an arc extinction fidelity representation. The prior generation layer for hidden danger components establishes candidate component channels in the zero-sequence direction, electrothermal direction, arc direction, and load reference direction based on the load spectrum residual representation, arc extinguishing fidelity representation, and the zero-sequence current correspondence and current-temperature rise correspondence under the same circuit object. It performs time-frequency position marking, channel normalization, and prior backfilling on each candidate component channel to generate a prior representation of hidden danger components and form a time-frequency feature set of electrical fire hazards. The improved WFTNet model was trained using a combination of time-frequency reconstruction error, load spectrum residual backfilling error, arc extinction buffer fidelity error, and hazard component prior discrimination error as joint optimization objectives. The network parameters of the normal circuit spectrum library memory reconstruction layer, the arc extinction segment buffer fidelity layer, and the hazard component prior generation layer were continuously optimized. When the change in the joint loss value in five consecutive training rounds was less than 0.001, the improved WFTNet model was considered to have completed convergence training.

[0012] Optionally, the obtained set of electrical fire hazard modal components includes: The time-frequency feature set of electrical fire hazards is serialized into components, and zero-sequence trajectory vector, electrothermal residual sequence, electric arc candidate sequence and load reference sequence are generated according to the zero-sequence direction, electrothermal direction, electric arc direction and load reference direction, respectively. Perform zero-sequence current symplectic trajectory rotation separation processing, write the zero-sequence trajectory vector into the symplectic trajectory matrix, establish the common-mode load direction, differential current leakage direction and sudden disturbance direction in the symplectic trajectory matrix, rotate the direction of the zero-sequence trajectory vector, and separate the load disturbance trajectory, leakage current change trajectory and sudden disturbance trajectory. Hamiltonian matrices are constructed and symplectic eigenvalues ​​are performed on leakage current change trajectories, sudden disturbance trajectories and arc candidate sequences. Symplectic feature pairs are arranged according to the continuous enhancement state of leakage current change trajectories, the short-term concentrated state of sudden disturbance trajectories and the distribution of quenching segments of arc candidate sequences, and the corresponding zero-sequence leakage current symplectic geometric components and arc disturbance symplectic geometric components are reconstructed. Perform electrothermal residual symmetric eigenvalue recalibration processing, write the electrothermal residual sequence into the electrothermal Hamilton matrix and perform symmetric eigenvalue decomposition, recalibrate the electrothermal related symmetric eigenvalues ​​according to the synchronous deviation state of the load spectrum residual segment and the temperature rise time frequency segment in the same time window, and reconstruct the corresponding electrothermal instability symmetric geometric components. The load reference symmetric geometric components are reconstructed based on the load disturbance trajectory and load reference sequence, and the zero-sequence leakage symmetric geometric components, arc disturbance symmetric geometric components, electrothermal instability symmetric geometric components and load reference symmetric geometric components are combined to form a set of electrical fire hazard mode components.

[0013] Optionally, obtaining the loop hazard index matrix includes: Establish a sequence of loop objects and a table of loop adjacency relationships based on the electrical loop structure table; Based on the loop object sequence and the acquisition time window, the zero-sequence leakage current symplectic geometric component, arc disturbance symplectic geometric component, electrothermal instability symplectic geometric component and load reference symplectic geometric component in the electrical fire hazard modal component set are mapped to the corresponding matrix positions to construct the component-loop correspondence matrix. Based on the loop adjacency table, the adjacent loops, upstream and downstream branches and component responses corresponding to the same terminal in the component-loop correspondence matrix are passed on adjacency to generate the component-loop correspondence matrix after adjacency passing. Based on the component-loop correspondence matrix after adjacency propagation, a loop hazard index matrix is ​​generated according to the loop object number, hazard component type, acquisition time window, and component response.

[0014] Optionally, the generation of the hazard object discrimination matrix includes: Based on the loop object number and the acquisition time window, the zero-sequence leakage response, arc disturbance response, electrothermal instability response and load reference deviation response are extracted from the loop hidden danger index matrix to generate a hidden danger component response matrix; Based on the combination relationship between zero-sequence leakage response, arc disturbance response, electrothermal instability response and load reference deviation response, the objects of each circuit are identified as leakage hazards, arc hazards, contact overheating hazards and overload temperature rise hazards, and a candidate matrix of hazard objects is generated. Based on the electrical circuit structure table, determine the branch relationships, terminal relationships, and adjacent circuit relationships among candidate objects. Merge and verify the candidate objects in the candidate matrix of potential hazards, retain the candidate objects with continuous component responses and consistent circuit positions, and generate the potential hazard discrimination matrix.

[0015] Optionally, the generation of electrical fire hazard detection and early warning results includes: Based on the loop object number and the collection time window, extract the hidden danger object, hidden danger component type, component response and loop location from the hidden danger object discrimination matrix, and match them with the preset warning level boundary table to generate a warning level identifier; Based on the electrical circuit structure table, the potential hazards are mapped to circuits, branches, terminals, and electrical equipment, generating hazard location identifiers and hazard object identifiers; By associating the hazard object, hazard type, warning level identifier, hazard location identifier, and disposal object identifier, electrical fire hazard detection and warning results are generated.

[0016] The beneficial effects of this invention are: This invention proposes an electrical fire hazard detection and early warning method based on intelligent sensors. By constructing a multi-source sensor data processing system for electrical fires, associating electrical circuit structures, arranging time-series samples of multi-source hazards, improving the time-frequency feature extraction of the WFTNet model, and implementing the SGMD algorithm for hazard mode decomposition, this method effectively improves the accuracy of early hazard identification and early warning judgment for electrical fires. Compared to traditional monitoring methods that rely on single sensor parameters or fixed thresholds, this invention can uniformly correlate data such as phase current, neutral current, residual current, voltage, load power, terminal temperature rise, cable temperature, arc disturbance, and ambient temperature and humidity. This reduces false alarms caused by normal load fluctuations, equipment start-up and shutdown disturbances, and environmental changes, and improves the stability of hazard identification under complex electrical operating conditions.

[0017] This invention improves the loop normal spectrum library memory reconstruction layer, arc extinguishing segment cache fidelity layer, and hazard component prior generation layer in the WFTNet model. It performs wavelet Fourier time-frequency transform, load spectrum residual extraction, arc extinguishing segment preservation, and hazard component prior processing on multi-source hazard time-series samples. This enables the simultaneous capture of normal load cycle deviation, short-duration arc pulses, zero-sequence current anomalies, and electrothermal anomalies. This processing method avoids the weakening of short-duration arc segments during ordinary filtering or sequence smoothing, improving the ability to identify intermittent arcs, early temperature rise due to poor contact, and overload temperature rise trends. This transforms electrical fire hazard detection from single-point anomaly judgment to comprehensive analysis of multi-source time-frequency features.

[0018] This invention utilizes the SGMD algorithm to perform zero-sequence current symmetric trajectory rotation separation processing and electrothermal residual symmetric eigenvalue recalibration processing. This further decomposes the time-frequency characteristics of electrical fire hazards into zero-sequence leakage current symmetric geometric components, arc disturbance symmetric geometric components, electrothermal instability symmetric geometric components, and load reference symmetric geometric components. Combined with the electrical circuit structure table, a circuit hazard index matrix and a hazard object discrimination matrix are established, enabling the correlated output of hazard category, hazard location, warning level, and disposal object. This method improves the positioning accuracy of leakage current hazards, arc hazards, contact overheating hazards, and overload temperature rise hazards, enhances the interpretability and scalability of warning results, and is applicable to online safety monitoring scenarios for building power distribution systems, industrial distribution cabinets, low-voltage power lines, and critical electrical equipment. Attached Figure Description

[0019] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings: Figure 1 This is a flowchart of an electrical fire hazard detection and early warning method based on intelligent sensors proposed in this invention; Figure 2 This is a schematic diagram of the structure of the improved WFTNet model for an electrical fire hazard detection and early warning method based on intelligent sensors proposed in this invention. Figure 3 This is a flowchart illustrating the SGMD algorithm used in the proposed method for detecting and warning electrical fire hazards based on intelligent sensors to generate a set of modal components for electrical fire hazards. Detailed Implementation

[0020] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0021] refer to Figure 1 , Figure 2 and Figure 3A method for detecting and warning of electrical fire hazards based on intelligent sensors, comprising: Collect multi-source sensor data and electrical circuit configuration data for electrical fires, preprocess the multi-source sensor data for electrical fires to generate a standardized electrical fire sensor dataset, and generate an electrical circuit structure table based on the electrical circuit configuration data; Based on the electrical circuit structure table, the standardized electrical fire sensor dataset is mapped to circuit objects, associated with sensor channels, and arranged with time windows to generate time-series samples of multi-source hazards. Based on the improved WFTNet model, wavelet Fourier time-frequency transform is performed on the time series samples of multi-source hidden dangers. Load spectrum residuals are extracted by combining the memory reconstruction of the normal circuit spectrum library. Local pulse preservation is achieved by using the arc extinguishing segment buffer fidelity. The time-frequency features of zero sequence, electrothermal, arc and load reference direction are generated and sorted through the prior of hidden danger components to obtain the time-frequency feature set of electrical fire hidden dangers. The SGMD algorithm is used to perform symplectic geometric mode decomposition on the time-frequency feature set of electrical fire hazards. Based on the rotational separation process of zero-sequence current symplectic trajectory, the zero-sequence related symplectic trajectory is split. The electrothermal residual symplectic eigenvalue recalibration process is introduced to correct the electrothermal related symplectic eigenvalue, and the set of modal components of electrical fire hazards is obtained. Based on the set of electrical fire hazard modal components and the electrical circuit structure table, a component-circuit correspondence matrix is ​​constructed and time-series alignment and adjacency transfer are performed to obtain the circuit hazard index matrix; The hazard component response of each loop object is calculated based on the loop hazard index matrix, and a hazard object discrimination matrix is ​​generated according to the combination relationship of the hazard component responses. Based on the hazard object identification matrix, the warning level mapping and disposal object matching are performed to generate electrical fire hazard detection and warning results.

[0022] In this embodiment, the multi-source sensing data for electrical fires specifically includes phase current data, neutral current data, residual current data, voltage data, load power data, power factor data, harmonic data, arc disturbance data, current spike data, voltage drop data, terminal temperature rise data, cable temperature data, distribution box internal temperature data, ambient temperature data, ambient humidity data, switch status data, and electrical equipment start / stop status data.

[0023] In this embodiment, the electrical circuit configuration data specifically includes circuit number, branch number, terminal number, switch number, circuit breaker number, electrical equipment number, circuit topology connection relationship, branch power supply relationship, terminal connection relationship, phase line connection relationship, neutral line connection relationship, grounding connection relationship, rated voltage, rated current, rated power, wire specifications, protection device parameters, sensor installation location, and sensor corresponding circuit relationship.

[0024] In this embodiment, generating the standardized electrical fire sensor dataset and electrical circuit structure table includes: Collect multi-source sensor data on electrical fires and electrical circuit configuration data, and establish original data association records according to sensor number, acquisition time, sensor type and circuit number; The original data association records of electrical fire multi-source sensor data are processed by deleting duplicate records, removing invalid data, filling in missing samples, and correcting abnormal jumps to generate cleaned sensor data, including: Missing sample completion and anomalous transition correction are handled as follows: The raw data association records are sorted according to sensor number, loop number, sensor type and acquisition time to determine the continuous sampling sequence corresponding to each sensor; The missing sampling positions are marked according to the sampling time interval in the continuous sampling sequence, and the missing sampling positions are filled in by generating a filler value and writing it into the missing sampling positions based on the change trend of valid sampling points before and after the missing sampling positions, or the same type of sensor under the same loop object in adjacent time windows. According to the continuous sampling sequence of the same sensor, calculate the numerical difference between the current sampling point and the previous sampling point, and match the jump judgment threshold table according to the sensor type, sensor range boundary, loop rated parameters and historical normal fluctuation range. When the value of the current sampling point exceeds the range boundary of the corresponding sensor, the current sampling point is marked as an invalid sampling point, removed from the continuous sampling sequence, and a replacement value is generated based on the adjacent valid sampling points before and after the current sampling point. When the current sampling point value does not exceed the corresponding sensor range boundary and the value difference reaches the jump judgment threshold of the corresponding sensor type in the jump judgment threshold table, the current sampling point is marked as a jump candidate point. When a candidate point of a change occurs within the same time window as the start / stop status of electrical equipment, arc disturbance data, or voltage drop data, or when two or more adjacent sampling points before and after a candidate point of a change occur in the same direction, the candidate point of a change is retained and marked as a valid change point. When a candidate point for a voltage jump is not located in the same acquisition time window as the start / stop status of electrical equipment, arc disturbance data, or voltage drop data, and no change in the same direction occurs between two or more adjacent sampling points before and after the candidate point, the candidate point for a voltage jump is marked as an abnormal voltage jump point. A correction value is generated based on the adjacent valid sampling points before and after the abnormal voltage jump point, and the abnormal voltage jump point is replaced with the correction value. Write the supplementary value, replacement value, effective change point and correction value into the corresponding continuous sampling sequence to generate cleaned sensor data; The cleaned sensor data is time-aligned according to a unified sampling time axis, and the data of different sensor types are dimensionally unified and numerically standardized to generate a standardized electrical fire sensor dataset. Based on the object numbers, connection relationships, and sensor installation locations in the electrical circuit configuration data, a structured association is established between circuit objects, branch objects, terminal objects, switch objects, electrical equipment objects, and sensor objects, generating an electrical circuit structure table, in which: Generate the electrical circuit structure table, specifically as follows: Generate loop object records, branch object records, terminal object records, switch object records, electrical equipment object records, and sensor object records according to the object number, and assign an object type identifier to each object record; Based on the loop topology connection relationship and branch power supply relationship, establish the power supply hierarchy relationship between loop objects, branch objects, switch objects and electrical equipment objects; Based on the terminal connection relationship, phase line connection relationship, neutral line connection relationship and grounding connection relationship, establish the electrical connection relationship between terminal objects and circuit objects, branch objects, switch objects and electrical equipment objects; Based on the sensor installation location and the corresponding circuit relationship, the sensor object is bound to the corresponding circuit object, branch object, terminal object or electrical equipment object to generate a sensor binding relationship; The power supply hierarchy, electrical connection, and sensor binding relationships are associated according to the circuit number, branch number, terminal number, electrical equipment number, and sensor number to generate an electrical circuit structure table.

[0025] In this embodiment, generating multi-source hidden danger time series samples includes: Based on the electrical circuit structure table, a circuit object index table is established, and the standardized electrical fire sensing dataset is mapped to the corresponding circuit object according to the sensor number and the corresponding circuit relationship of the sensor, thus generating a circuit object sensing data table. According to the sensing type, the sensing data table of the loop object is associated with sensing channels to generate current channel, voltage channel, residual current channel, temperature channel, arc disturbance channel, environment channel, and status channel, among which: The sensor channel association is performed as follows: Group the various types of sensor data in the loop object sensor data table according to the loop object number and the acquisition time; Combine phase current data, neutral current data, load power data, and power factor data into a current path; Combine voltage data and voltage drop data into a voltage channel; Combine the residual current data, phase current data, and neutral current data into a residual current channel; Combine terminal temperature rise data, cable temperature data, and distribution box internal temperature data into a temperature channel; The arc disturbance data, current spike data, and voltage drop data are combined into an arc disturbance channel; Combine ambient temperature data and ambient humidity data into an environmental channel; Combine switch status data and electrical equipment start / stop status data into a status channel; Arrange the time windows of each sensing channel according to a unified sampling time axis to generate a loop sensing time window sequence. Based on the loop object index table, loop sensing time window sequence, and electrical loop structure table, establish sample association relationships for adjacent loops, the same branch, the same terminal, and the same electrical equipment to generate multi-source hidden danger time sequence samples, including: Generate time-series samples of multi-source hidden dangers, specifically as follows: According to the loop object index table, bind each time window sample in the loop sensing time window sequence with the corresponding loop object, branch object, terminal object and electrical equipment object to generate loop window samples; Based on the circuit topology connection relationship and branch power supply relationship in the electrical circuit structure table, establish the association between the circuit window samples corresponding to adjacent circuits, the same branch, and upstream and downstream branches to generate circuit adjacency sample relationship; Based on the terminal connection relationship and electrical equipment number in the electrical circuit structure table, establish an association between the circuit window samples corresponding to the same terminal and the same electrical equipment to generate object peer sample relationship; The loop window samples, loop adjacent sample relationships, and object co-location sample relationships are combined according to the collection time window to generate multi-source hidden danger time series samples.

[0026] In this embodiment, obtaining the time-frequency feature set of electrical fire hazards includes: An improved WFTNet model is constructed by placing the normal circuit spectral library memory reconstruction layer on the Fourier periodic feature branch, the arc extinction segment buffer fidelity layer on the wavelet local frequency band feature branch, and the hazard component prior generation layer after the Fourier periodic feature branch and the wavelet local frequency band feature branch, wherein: The improved WFTNet model is constructed as follows: In the traditional WFTNet model, the input time series samples are transformed into Fourier periodic feature branches and wavelet local frequency band feature branches after wavelet Fourier time-frequency transform. A loop normal spectral library memory reconstruction layer is connected to the Fourier periodic feature branch of the traditional WFTNet model, and the output of the loop normal spectral library memory reconstruction layer is connected to the input of the hidden danger component prior generation layer. An arc extinguishing fragment buffer fidelity layer is added to the wavelet local frequency band feature branch of the traditional WFTNet model, and the output of the arc extinguishing fragment buffer fidelity layer is connected to the input of the hazard component prior generation layer. After the Fourier periodic feature branch and the wavelet local frequency band feature branch, the hazard component prior generation layer is connected. The input of the hazard component prior generation layer receives the output results of the Fourier periodic feature branch, the output results of the wavelet local frequency band feature branch, the output results of the loop normal spectrum library memory reconstruction layer, and the output results of the arc extinguishing segment buffer fidelity layer. The output of the prior generation layer for hidden danger components is connected to the time-frequency feature set generation structure for electrical fire hazards, resulting in an improved WFTNet model. The normal loop spectrum library memory reconstruction layer archives the Fourier periodic features of multi-source hidden danger time series samples according to the loop object and normal operation window identifier, generating a normal loop spectrum library. It then aligns the Fourier periodic features of the current monitoring window with the normal loop spectrum library, extracts the frequency segments and durations that deviate from the normal operation spectrum of the current monitoring window, and performs residual backfilling according to the periodic position to generate a load spectrum residual representation, where: The normal spectral library memory reconstruction layer includes: Loop object index register: stores the loop object number and associated object number; Normal operation window identifier register: stores the normal operation window identifier and the acquisition time window identifier; Fourier spectrum fragment cache queue: caches Fourier periodic features in time series samples with multiple potential risks; Normal Spectrum Library Table for Loops: Stores the normal operating frequency segments, period positions, and spectral response ranges for each loop object; Period position aligner: Aligns the current monitoring window with the period position in the loop normal spectral library table; Spectral residual comparator: marks frequency segments that deviate from the normal operating spectrum; Duration period counter: Counts the duration of deviations from the frequency segment; Residual backfill buffer: Writes frequency segment deviation and duration according to periodic position; Load spectrum residual output register: collects residual backfill results and generates a load spectrum residual representation; In the normal loop spectrum library memory reconstruction layer, the loop object index register and the normal operation window identifier register pass the loop object number, associated object number, normal operation window identifier, and acquisition time window identifier to the Fourier spectrum segment buffer queue. The Fourier spectrum segment buffer queue writes the Fourier periodic features corresponding to the normal operation window into the normal loop spectrum library table according to the loop object and acquisition time window, and passes the Fourier periodic features corresponding to the current monitoring window to the period position aligner. The normal loop spectrum library table and the Fourier periodic features of the current monitoring window are input into the period position aligner to obtain the period aligned spectrum segment. The period aligned spectrum segment generates the frequency segment deviation through the spectrum residual comparator, generates the deviation duration through the duration counter, and enters the residual backfill buffer. The load spectrum residual representation is generated by the load spectrum residual output register. Generate a normal spectral library of the loop, specifically: Based on the normal operation window identifier, Fourier periodic features corresponding to the normal operation window are selected from the multi-source hidden danger time series samples. The normal operation window identifier is generated based on the acquisition time window in the sensor data after cleaning that is not marked as an abnormal jump point, not marked as a valid change point, and has complete sensor channel data. Based on the loop object, acquisition time window, and sensing channel, the Fourier periodic features corresponding to the normal operation window are grouped to generate a set of normal operation spectrum segments; Under the same loop object, perform spectral response statistics on the normal operating frequency segment at the same period position to generate the normal frequency segment, period position, and spectral response range consisting of the lower and upper limits of the response; Associate the normal frequency segment, period position, spectral response range, loop object number, branch object number, terminal object number and sensor channel identifier, and write them into the loop normal spectrum library table to generate the loop normal spectrum library; Perform periodic position alignment, specifically as follows: Based on the loop object number corresponding to the current monitoring window, match the normal frequency segment, period position and spectral response range of the same loop object in the normal spectrum library; The period length is determined based on the frequency segments and acquisition time window in the Fourier periodic characteristics, and the current period position corresponding to each frequency segment in the current monitoring window is determined. The current period position is mapped to the period position in the normal spectral library to generate a period position mapping relationship; When the current period position is offset from the period position in the normal spectral library, the period position corresponding to the current frequency segment is moved according to the acquisition time window order. Arrange the current frequency segment after the movement, the normal frequency segment in the loop normal spectrum library, and the spectral response range according to the period position correspondence to generate a period-aligned spectral segment; The load spectrum residual representation is generated as follows: Compare the current spectral response value in the periodically aligned spectral segment with the spectral response range in the loop-normal spectral library; When the current spectral response value is outside the spectral response range, the corresponding frequency segment is marked as an off-frequency segment; Based on the data collection time window, the consecutive occurrence periods of deviation frequency segments are counted to generate deviation duration periods; Generate a spectral residual record based on the off-frequency segment, the current spectral response value, the spectral response range, and the duration of the off-frequency segment; The spectral residual records are backfilled to the corresponding time-frequency positions according to the periodic positions to generate a load spectrum residual representation; The arc extinction segment cache fidelity layer performs short-time pulse search on the wavelet local frequency band features of the multi-source hidden danger time series samples, extracts arc candidate wave packets, and caches and groups these candidate wave packets according to the arc initiation segment, arc extinction interval segment, and arc re-ignition segment. Based on the equipment start-up and shutdown status, start-up and shutdown disturbance segments are written into the bypass channel and separated from the cache groups. Continuity completion is performed on the arc candidate wave packets in the cache groups to generate a high-fidelity representation of arc extinction, where: The arc extinguishing fragment buffer fidelity layer includes: Wavelet band segment cache queue: caches short-time high-frequency segments in local wavelet band features; Pulse trigger register: stores trigger flags corresponding to current spikes, voltage dips, and arcing disturbances; Arc candidate packet extractor: Extracts arc candidate packets from short-time high-frequency segments; Extinguishing segment marking table: records the time and location of the arc initiation segment, the arc extinction interval segment, and the reignition segment; Extinguishing group buffer: buffers candidate arc packets according to the order of arc initiation, arc extinction interval, and arc re-ignition; Start / Stop Disturbance Bypass Register: Identifier of the disturbance segment corresponding to the start / stop state of the storage device; Bypass shunt comparator: Separates start-stop disturbance segments from arc candidate wave packets; Continuity completion buffer: Perform segment continuity completion on the candidate arc wave packets in the extinguishing group buffer; Arc fidelity output register: Collects the continuously padded candidate arc packets and generates a fidelity representation of arc extinguishing; In the arc extinction segment buffer fidelity layer, the wavelet frequency band segment buffer queue transmits short-time high-frequency segments from the local frequency band features of the wavelet to the arc candidate wave packet extractor. The pulse trigger register transmits the trigger identifiers corresponding to current spikes, voltage drops, and arc disturbances to the arc candidate wave packet extractor. The arc candidate wave packets generated by the arc candidate wave packet extractor enter the extinction segment marking table for time and position recording. The extinction segment marking table transmits the arc start segment, arc extinction interval segment, and re-arc segment to the extinction group buffer. The start-stop disturbance bypass register transmits the disturbance segment identifiers corresponding to the equipment start-stop state to the bypass shunt comparator. The bypass shunt comparator separates the start-stop disturbance segments from the arc candidate wave packets. The arc candidate wave packets in the extinction group buffer enter the continuity completion buffer for segment continuity completion. The continuous completion arc candidate wave packets are transmitted to the arc fidelity output register to generate an arc extinction fidelity representation. Extracting candidate wave packets of the electric arc, specifically: The wavelet local frequency band features are arranged according to the loop object, sensing channel, and acquisition time window to form a local frequency band response sequence; Based on the current spike trigger flag, voltage drop trigger flag, and arc disturbance trigger flag in the pulse trigger register, mark the candidate pulse time position in the local frequency band response sequence; Local wavelet frequency band segments are extracted within a truncation window determined by the sampling interval and the duration range of the arc pulse before and after the candidate pulse time position to generate candidate local wave packets; The candidate local wave packets are matched with the current spike data, voltage drop data and arc disturbance data within the same acquisition time window, and isolated high-frequency segments that do not form a sensing correspondence are screened out. The retained candidate local wave packets are recorded according to the loop object number, acquisition time window and sensor channel identifier to generate arc candidate wave packets; The extinguishing fragment cache is grouped as follows: Arrange the candidate arc wave packets in the order of acquisition time to generate an arc candidate wave packet sequence; The candidate wave packet boundaries are defined based on the response start position, response peak position, and response termination position of each candidate wave packet in the arc candidate wave packet sequence. The segments in the candidate wave packet boundary whose local frequency band response is higher than the adjacent background response are marked as arc initiation segments; The segment whose local frequency band response between two adjacent candidate arc wave packets falls back to the range of the adjacent background response is marked as the arc extinction interval segment; The segment whose local frequency response is higher than the adjacent background response again after the arc extinguishing interval segment is marked as the re-ignition segment; According to the acquisition time sequence of the arc initiation segment, arc extinction interval segment, and arc re-ignition segment, the arc candidate wave packets are written into the corresponding buffer groups to form the arc extinction segment buffer groups; The start / stop disturbance fragments are written to the bypass channel and separated from the cached group, specifically as follows: According to the acquisition time window, the candidate arc wave packets are matched with the start-stop status, switching status and load power change segments of the electrical equipment; When the candidate arc packet is in the same acquisition time window as the start-stop or switch state of the electrical equipment, and the load power change segment and the current spike segment correspond in time position, the corresponding candidate arc packet is marked as the start-stop disturbance segment. The start-stop disturbance segment is separated from the burnout segment buffer group and written into the bypass channel according to the loop object number, acquisition time window, sensor channel identifier and response intensity; The arc initiation segment, arc extinction interval segment, and arc reignition segment after the separation start-stop disturbance segment are rearranged according to the acquisition time order to generate a buffer group of the separated arc extinction segment; Perform continuity completion as follows: Arrange the arc initiation segment, arc extinction interval segment, and re-arc segment in the separated arc extinction segment cache group according to the acquisition time sequence; The discontinuity interval between arc segments is determined based on the time interval between adjacent arc initiation segments and arc re-ignition segments; The local frequency band response within the discontinuous interval is brought back to the range of the adjacent background response, and the wavelet local frequency band segment that is located in the same acquisition time window as the voltage drop data or short-time change data of the residual current and the adjacent arcing segment is marked as the arc extinguishing interval completion segment. Insert the arc extinguishing interval completion segment between adjacent arc ignition start segment and arc re-ignition segment to generate a continuous arc extinguishing segment sequence; According to the loop object number, acquisition time window and sensor channel identifier, the continuous arc extinction segment sequence is archived to generate a high-fidelity representation of arc extinction; The prior generation layer for potential hazards establishes candidate component channels in the zero-sequence direction, electrothermal direction, arc direction, and load reference direction based on the load spectrum residual representation, arc extinction fidelity representation, and the zero-sequence current and current-temperature rise correspondence under the same circuit object. Time-frequency position marking, channel normalization, and prior backfilling are performed on each candidate component channel to generate a prior representation of the potential hazards, forming a time-frequency feature set of electrical fire hazards, where: The prior generation layer for potential hazards includes: Residual input register: stores the load spectrum residual representation and the arc extinguishing fidelity representation; Zero-sequence correspondence table: stores the correspondence between phase current, neutral current, and residual current; Electrothermal Correspondence Table: Correspondence between changes in storage current, load spectrum residuals, and temperature rise; Component channel distributor: Establishes candidate component channels for zero-sequence direction, electrothermal direction, arc direction, and load reference direction; Time-frequency location marking table: records the time window, frequency segment, and loop object number corresponding to the candidate component channel; Channel normalization processor: Normalizes the time-frequency response of candidate component channels; Prior backfill buffer: Write the normalized candidate component channel response according to the time-frequency position; Prior output register: collects prior backfill results and generates prior representations of potential hazard components; Time-frequency feature set buffer: Associate the prior representation of the hidden danger components, the load spectrum residual representation, and the arc extinction fidelity representation to form a time-frequency feature set of electrical fire hazards; In the prior generation layer of the hidden danger components, the data from the residual input register, the zero-sequence correspondence table, and the electrothermal correspondence table are input into the component channel distributor to generate candidate component channels in the zero-sequence direction, electrothermal direction, arc direction, and load reference direction. The candidate component channels are marked by the time-frequency position marking table to form channel marking results. The channel normalization processor generates normalized candidate component channel responses. The normalized candidate component channel responses enter the prior backfill buffer and are passed to the prior output register to generate the prior representation of the hidden danger components. The time-frequency feature set buffer associates the prior representation of the hidden danger components, the load spectrum residual representation, and the arc extinguishing fidelity representation to form the time-frequency feature set of electrical fire hazards. Establish candidate component channels, specifically as follows: Align the load spectrum residual representation, arc extinguishing fidelity representation, zero-sequence current correspondence, and current-temperature rise correspondence according to the loop object number and acquisition time window; Based on the correspondence between phase current, neutral current and residual current of the same loop object, the zero-sequence deviation response is extracted and zero-sequence direction candidate component channels are generated. Based on the synchronous deviation relationship between the load spectrum residual, terminal temperature rise and cable temperature under the same circuit object, the electrothermal deviation response is extracted to generate candidate component channels for the electrothermal direction. Based on the arc initiation segment, arc extinction interval segment, and arc reignition segment in the arc extinguishing fidelity representation, the arc extinguishing response is extracted, and candidate component channels for the arc direction are generated. Based on the load cycle segments in the load spectrum residual representation, as well as the load power variation and power factor variation under the same loop object, the load reference response is extracted, and candidate component channels in the load reference direction are generated. The candidate component channels in the zero-sequence direction, the candidate component channels in the electrothermal direction, the candidate component channels in the arc direction, and the candidate component channels in the load reference direction are combined according to the loop object number and the acquisition time window to form candidate component channels; The execution of time-frequency location marking, channel normalization, and prior backfilling is as follows: Based on the loop object number, acquisition time window, sensor channel identifier and frequency segment position, time-frequency position markers are configured for the zero-sequence direction candidate component channel, the electrothermal direction candidate component channel, the arc direction candidate component channel and the load reference direction candidate component channel, and candidate component position records are generated. Within the same loop object and the same acquisition time window, the maximum response value, minimum response value, and current response value of each candidate component channel are calculated respectively. Using the minimum response value as the lower bound of normalization and the maximum response value as the upper bound of normalization, the current response value is mapped to the interval between 0 and 1 to generate normalized candidate component responses. When the maximum response value is equal to the minimum response value, the normalized candidate component response of the corresponding candidate component channel is set to 0. According to the candidate component position record, the normalized zero-sequence directional response, electrothermal directional response, electric arc directional response and load reference directional response are written into the corresponding time-frequency position. When there are two or more candidate component channel responses at the same time-frequency position, they are backfilled according to the candidate component channel type, and channel response merging is not performed. The zero-sequence directional response, electrothermal directional response, electric arc directional response and load reference directional response after backfilling are archived to generate a priori representation of the hidden danger components. The prior representation of the hidden danger component, the load spectrum residual representation, and the arc extinguishing fidelity representation are associated according to the circuit object number, the acquisition time window, and the sensor channel identifier to form a time-frequency feature set of electrical fire hazards; The improved WFTNet model was trained using a combination of time-frequency reconstruction error, load spectrum residual backfilling error, arc extinction buffer fidelity error, and hazard component prior discrimination error as joint optimization objectives. The network parameters of the loop normal spectrum library memory reconstruction layer, the arc extinction segment buffer fidelity layer, and the hazard component prior generation layer were continuously optimized. When the change in the joint loss value over five consecutive training rounds was less than 0.001, the improved WFTNet model was considered to have completed convergence training. The improved WFTNet model is trained as follows: A training sample set is composed of multi-source hidden danger time series samples and corresponding electrical fire hidden danger time-frequency feature reference records, load spectrum residual reference records, arc extinguishing segment reference records and hidden danger component reference records. The training sample set is divided into training set, validation set and test set according to 7:2:1. Each training batch contains 32 samples. The multi-source hidden danger time series samples in the training set are input into the improved WFTNet model, and the output is the load spectrum residual representation, arc extinguishing fidelity representation, hidden danger component prior representation and electrical fire hidden danger time-frequency feature set. The time-frequency reconstruction error is obtained by averaging the squared differences between the time-frequency feature set of electrical fire hazards and the reference record of electrical fire hazards. The load spectrum residual backfilling error is obtained by averaging the squared differences between the load spectrum residual representation and the reference record of load spectrum residuals. The arc extinguishing fidelity representation and the arc extinguishing segment reference record are averaged to obtain the arc extinguishing buffer fidelity error. The hazard component prior discrimination error is obtained by averaging the squared differences between the response in each direction of the hazard component prior representation and the hazard component reference record. The time-frequency reconstruction error, load spectrum residual backfilling error, arc extinguishing buffer fidelity error, and hazard component prior discrimination error are each multiplied by 0.25 and then summed to generate a joint loss value. The gradient value corresponding to each network parameter is calculated based on the joint loss value. The gradient value is multiplied by 0.001 and then subtracted from the current network parameter value to obtain the updated network parameter value. The updated network parameter value is written into the loop normal spectrum library memory reconstruction layer, the arc extinguishing segment buffer fidelity layer, and the hazard component prior generation layer. When the change of the joint loss value in 5 consecutive training rounds is less than 0.001, the improved WFTNet model is considered to have completed convergence training.

[0027] In this embodiment, obtaining the set of electrical fire hazard modal components includes: The time-frequency feature set of electrical fire hazards is component-serialized, generating zero-sequence trajectory vectors, electrothermal residual sequences, arc candidate sequences, and load reference sequences according to the zero-sequence direction, electrothermal direction, arc direction, and load reference direction, respectively. The zero-order trajectory vector, electrothermal residual sequence, arc candidate sequence, and load reference sequence are generated, specifically as follows: Based on the loop object number, acquisition time window, frequency segment position, and candidate component channel type, the prior representation, load spectrum residual representation, and arc extinction fidelity representation of the time-frequency characteristics of electrical fire hazards are grouped. The zero-sequence directional response is spliced ​​with the phase current time-frequency segment, neutral current time-frequency segment and residual current time-frequency segment under the same loop object according to the same acquisition time window to generate a zero-sequence trajectory vector. The electrothermal direction response, load spectrum residual segment, terminal temperature rise time-frequency segment, and cable temperature time-frequency segment are arranged in the order of the acquisition time window to generate an electrothermal residual sequence. Arrange the arc direction response and the arc extinguishing fidelity representation of the arc ignition start segment, arc extinguishing interval segment and arc reignition segment in the order of acquisition time to generate an arc candidate sequence; The load reference direction response, load cycle segment, load power change segment, and power factor change segment are arranged according to the same loop object and the same acquisition time window to generate a load reference sequence. Perform zero-sequence current symplectic trajectory rotation separation processing, write the zero-sequence trajectory vector into the symplectic trajectory matrix, establish the common-mode load direction, differential current leakage direction, and sudden disturbance direction in the symplectic trajectory matrix, and perform directional rotation on the zero-sequence trajectory vector to separate the load disturbance trajectory, leakage current change trajectory, and sudden disturbance trajectory, where: Perform zero-sequence current symplectic trajectory rotation separation processing, specifically as follows: According to the acquisition time window order, the phase current time-frequency segment, neutral current time-frequency segment, residual current time-frequency segment and zero-sequence direction response in the zero-sequence trajectory vector are written into the symplectic trajectory matrix; Based on the same-direction relationship between the time-frequency segments of the phase current and the time-frequency segment of the neutral current, the common-mode load direction is generated in the symplectic trajectory matrix; Based on the difference in the time-frequency segments of phase current, neutral current and residual current, the differential current leakage direction is generated in the symplectic trajectory matrix. Based on the time correspondence between the short-time surge segment in the zero-sequence directional response and the current spike segment and voltage drop segment within the same acquisition time window, the sudden disturbance direction is generated in the symplectic trajectory matrix. The common-mode load direction, differential current leakage direction, and sudden disturbance direction are respectively normalized to generate the common-mode load unit direction, differential current leakage unit direction, and sudden disturbance unit direction; The zero-sequence trajectory vector is inner-producted with the common-mode load unit direction, the differential leakage unit direction, and the sudden disturbance unit direction to obtain the common-mode load projection coefficient, the differential leakage projection coefficient, and the sudden disturbance projection coefficient. The load disturbance trajectory is reconstructed along the common-mode load unit direction based on the common-mode load projection coefficient; the leakage current change trajectory is reconstructed along the differential current leakage unit direction based on the differential current leakage projection coefficient; and the sudden disturbance trajectory is reconstructed along the sudden disturbance unit direction based on the sudden disturbance projection coefficient. Hamiltonian matrices are constructed and symplectic eigenvalue decomposition is performed on the leakage current change trajectory, sudden disturbance trajectory, and arc candidate sequence. Symplectic feature pairs are arranged according to the continuous enhancement state of the leakage current change trajectory, the short-time concentrated state of the sudden disturbance trajectory, and the distribution of the extinguishing segments of the arc candidate sequence. The corresponding zero-sequence leakage current symplectic geometric components and arc disturbance symplectic geometric components are reconstructed, where: Construct the Hamiltonian matrix and perform symplectic eigenvalue decomposition, specifically as follows: According to the acquisition time window order, the leakage current change trajectory, sudden disturbance trajectory and arc candidate sequence are respectively divided into state component and accompanying component. The state component consists of the trajectory response value within the current acquisition time window, and the accompanying component consists of the trajectory response change between adjacent acquisition time windows. The state components are arranged in the first block of the matrix according to the acquisition time window order, and the accompanying components are arranged in the second block of the matrix according to the acquisition time window order. The corresponding values ​​of the state components and accompanying components within the same acquisition time window are written into the cross block position to generate the leakage current change Hamilton matrix, the sudden disturbance Hamilton matrix, and the arc candidate Hamilton matrix. Perform symplectic eigendecomposition on the leakage current change Hamiltonian matrix, the sudden disturbance Hamiltonian matrix, and the arc candidate Hamiltonian matrix respectively, and calculate the symplectic eigenvalues ​​and symplectic eigenvectors corresponding to each Hamiltonian matrix; Based on the pairing relationship of positive and negative symmetric eigenvalues, the phase correspondence of symmetric eigenvectors, and the order of acquisition time windows, symmetric eigenvalues ​​and symmetric eigenvectors are paired and sorted to generate leakage current change symmetric eigenvalue pairs, sudden disturbance symmetric eigenvalue pairs, and arc candidate symmetric eigenvalue pairs. The reconstructed zero-sequence leakage current symplectic geometric components and arc disturbance symplectic geometric components are as follows: Arrange the response values ​​corresponding to the symplectic characteristic pairs of leakage current change according to the acquisition time window order, count the number of acquisition time windows in which the response values ​​increase continuously in the same direction to obtain the enhancement duration, and subtract the initial response value from the end response value of the continuous enhancement interval to obtain the response growth rate. Arrange the symplectic feature pairs of leakage current changes according to the enhanced duration and response growth rate, and generate a zero-sequence leakage current component reconstruction sequence; Arrange the response values ​​corresponding to the symplectic feature pairs of sudden disturbances according to the sampling point order. Divide the continuous sampling points whose response values ​​are higher than the background response range adjacent to the sudden disturbance trajectory into concentrated response intervals. Count the number of sampling points in the concentrated response intervals to obtain the number of continuous sampling points. Then, average the response values ​​in the concentrated response intervals to obtain the concentrated response intensity. Arrange the burst disturbance symplectic feature pairs according to the concentrated response intensity and the number of continuous sampling points, and generate the burst disturbance component reconstruction sequence; Based on the temporal distribution of the arc initiation segment, arc extinction interval segment, and arc reignition segment in the arc candidate sequence, the arc candidate symplectic feature pairs are classified and sorted by time to generate the arc extinguishing component reconstruction sequence. Based on the symplectic eigenvalues ​​and symplectic eigenvectors of leakage current change in the reconstructed sequence of zero-sequence leakage current components, the symplectic eigenvalues ​​of leakage current change are used as weight coefficients and the symplectic eigenvectors of leakage current change are used as reconstruction directions to perform weighted reconstruction of the leakage current change trajectory. The zero-sequence leakage current symplectic geometric components are generated by mapping according to the acquisition time window and frequency segment position. Based on the sudden disturbance symmetric eigenvalues ​​and sudden disturbance symmetric eigenvectors in the sudden disturbance component reconstruction sequence, and the arc candidate symmetric eigenvalues ​​and arc candidate symmetric eigenvectors in the arc extinguishing component reconstruction sequence, the sudden disturbance trajectory and arc candidate sequence are combined and reconstructed, and mapped according to the time order of the arc initiation segment, the arc extinguishing interval segment and the arc re-ignition segment to generate the arc disturbance symmetric geometric component. Perform electrothermal residual symmetric eigenvalue recalibration processing, write the electrothermal residual sequence into the electrothermal Hamiltonian matrix and perform symmetric eigenvalue decomposition. Recalibrate the electrothermal related symmetric eigenvalues ​​based on the synchronous deviation state of the load spectrum residual segment and the temperature rise time-frequency segment within the same time window, and reconstruct the corresponding electrothermal instability symmetric geometric components, where: The electrothermal residual symplectic eigenvalue recalibration process is performed as follows: According to the acquisition time window order, the load spectrum residual segment, terminal temperature rise time-frequency segment, and cable temperature time-frequency segment in the electrothermal residual sequence are written into the electrothermal residual matrix; The response value of the current acquisition time window in the electrothermal residual matrix is ​​taken as the state component, and the response change between adjacent acquisition time windows is taken as the adjoint component. The electrothermal Hamilton matrix is ​​constructed according to the correspondence between the state component and the adjoint component. Perform symplectic eigenvalue decomposition on the electrothermal Hamiltonian matrix, calculate the electrothermal symplectic eigenvalues ​​and symplectic eigenvectors corresponding to the electrothermal Hamiltonian matrix, and generate electrothermal symplectic feature pairs according to the correspondence between the electrothermal symplectic eigenvalues ​​and symplectic eigenvectors. According to the same acquisition time window, the residual response change in the load spectrum residual segment, the terminal temperature rise response change in the terminal temperature rise time-frequency segment, and the cable temperature response change in the cable temperature time-frequency segment are correlated. The product of the residual response change and the terminal temperature rise response change is calculated to generate the terminal synchronization direction quantity. The product of the residual response change and the cable temperature response change is calculated to generate the cable synchronization direction quantity. When the terminal synchronization direction is positive and the terminal temperature rise response change is positive, or when the cable synchronization direction is positive and the cable temperature response change is positive, the corresponding acquisition time window is marked as a synchronization deviation window, and the number of consecutive synchronization deviation windows is counted to obtain the duration of synchronization deviation. The absolute values ​​of the residual response change, terminal temperature rise response change, and cable temperature response change within the synchronization deviation window are averaged to obtain the synchronization deviation intensity. Multiply the duration of the synchronization deviation by the intensity of the synchronization deviation to generate the electrothermal recalibration coefficient. Multiply the electrothermal correlation symplectic eigenvalue by the electrothermal recalibration coefficient to obtain the recalibrated electrothermal symplectic eigenvalue. Rearrange the electrothermal symplectic eigenvalues ​​according to the recalibrated electrothermal symplectic eigenvalues ​​to generate the recalibrated electrothermal symplectic eigenvalue pairs. Based on the recalibrated electrothermal symmetric eigenvalues ​​and symmetric eigenvectors in the recalibrated electrothermal symmetric eigenvalues, the recalibrated electrothermal symmetric eigenvalues ​​are used as weighting coefficients and the symmetric eigenvectors are used as reconstruction directions to perform weighted reconstruction of the electrothermal residual sequence. The reconstruction is then performed according to the acquisition time window and frequency segment position to generate electrothermal instability symmetric geometric components. The load reference symmetric geometric components are reconstructed based on the load disturbance trajectory and load reference sequence, and the zero-sequence leakage symmetric geometric components, arc disturbance symmetric geometric components, electrothermal instability symmetric geometric components and load reference symmetric geometric components are combined to form a set of electrical fire hazard mode components.

[0028] In this embodiment, obtaining the loop hazard index matrix includes: Based on the electrical circuit structure table, establish the circuit object sequence and circuit adjacency table, specifically as follows: Based on the circuit objects, branch objects, terminal objects, and electrical equipment objects in the electrical circuit structure table, arrange them in hierarchical order according to the circuit object number, branch object number, terminal object number, and electrical equipment object number to generate a circuit object sequence; Based on the circuit topology and branch power supply relationships in the electrical circuit structure table, determine the relationships between adjacent circuits and the relationships between upstream and downstream branches; Based on the terminal connection relationships in the electrical circuit structure table, determine the terminal association relationships between circuit objects, branch objects, and electrical equipment objects connected to the same terminal; Associat the relationships between adjacent circuits, upstream and downstream branches, and terminals according to the circuit object number, branch object number, terminal object number, and electrical equipment object number to generate a circuit adjacency table; Based on the loop object sequence and acquisition time window, the zero-sequence leakage current symplectic geometric component, arc disturbance symplectic geometric component, electrothermal instability symplectic geometric component, and load reference symplectic geometric component in the electrical fire hazard modal component set are mapped to their corresponding matrix positions, constructing a component-loop correspondence matrix, where: Construct the component-loop correspondence matrix as follows: The matrix row index is determined according to the sequence of loop objects. The matrix row index includes the loop object number, branch object number, terminal object number, and electrical equipment object number. The matrix column index is determined according to the collection time window and the type of hidden danger component. The types of hidden danger components include zero-sequence leakage current component, arc disturbance component, electrothermal instability component and load reference component. Write the zero-sequence leakage current symmetric geometric components, arc disturbance symmetric geometric components, electrothermal instability symmetric geometric components, and load reference symmetric geometric components into the matrix position according to the loop object number, acquisition time window, and corresponding hidden danger component type. When the same loop object has component responses at multiple frequency segment locations within the same acquisition time window, the component responses at multiple frequency segment locations are averaged to generate a window component response, which is then written to the corresponding matrix location. When there is no component response at a matrix position, the component response at that matrix position is recorded as 0; The matrix row index, matrix column index, and component responses in the matrix position are correlated to generate a component-loop correspondence matrix; Based on the loop adjacency table, adjacency propagation is performed on adjacent loops, upstream and downstream branches, and component responses corresponding to the same terminal in the component-loop correspondence matrix to generate the component-loop correspondence matrix after adjacency propagation, where: Generate the component-loop correspondence matrix after adjacency transit, specifically as follows: Based on the loop adjacency table, determine the adjacent loop objects, upstream branch objects, downstream branch objects, and associated objects with the same terminal for each loop object in the component-loop correspondence matrix; Under the same acquisition time window and the same hidden danger component type, extract the component response of the current matrix row object, the component response of the adjacent loop object, the component response of the upstream branch object, the component response of the downstream branch object, and the component response of the object associated with the same terminal; Write the component responses of adjacent loop objects to the adjacent loop transmission positions corresponding to the current loop object according to the loop topology connection relationship; Write the component responses of the upstream branch object and the downstream branch object into the branch transmission position corresponding to the current loop object according to the branch power supply direction; Write the component responses of the same terminal-associated object to the terminal transmission position corresponding to the current loop object according to the terminal object number; Write the component responses of the current matrix row object, the component responses of the adjacent loop transmission positions, the component responses of the branch transmission positions, and the component responses of the terminal transmission positions into the corresponding matrix positions according to the response source, and generate the adjacent transmission component responses. Write the adjacency-transfer component response into the corresponding matrix position in the component-loop correspondence matrix to generate the adjacency-transfer component-loop correspondence matrix; Based on the component-loop correspondence matrix after adjacency propagation, a loop hazard index matrix is ​​generated according to the loop object number, hazard component type, acquisition time window, and component response.

[0029] In this embodiment, generating the hidden danger object discrimination matrix includes: Based on the loop object number and the acquisition time window, the zero-sequence leakage response, arc disturbance response, electrothermal instability response and load reference deviation response are extracted from the loop hidden danger index matrix to generate a hidden danger component response matrix; Based on the combined relationships between zero-sequence leakage response, arc disturbance response, electrothermal instability response, and load reference deviation response, leakage potential, arc potential, contact overheating potential, and overload temperature rise potential are identified for each circuit object, generating a candidate matrix of potential objects, where: The candidate matrix of potential hazards is generated as follows: According to the loop object number and the acquisition time window, the zero-sequence leakage response, arc disturbance response, electrothermal instability response and load reference deviation response are written into the same discrimination record; When the zero-sequence leakage response increases in the same direction within two or more consecutive acquisition time windows, and the zero-sequence leakage response value is greater than the load reference deviation response value, the corresponding circuit object is marked as a candidate object for leakage potential. When the value of the arc disturbance response is greater than 0, and the arc disturbance response and the zero-sequence leakage current response occur simultaneously within the same acquisition time window, the corresponding circuit object is marked as a candidate object for arc hazard. When the value of the electrothermal instability response is greater than 0, and the circuit location corresponding to the electrothermal instability response is connected to the terminal object or branch object, the corresponding terminal object or branch object is marked as a candidate object for contact overheating hazard. When the load reference deviation response value is greater than 0, and the load reference deviation response and the electrothermal instability response occur simultaneously within the same acquisition time window, the corresponding circuit object or electrical equipment object will be marked as a candidate object for overload temperature rise hazard. The candidate objects for leakage current hazards, electric arc hazards, contact overheating hazards, and overload temperature rise hazards are written into the matrix position according to the object number, circuit object number, hazard type, and collection time window to generate a hazard object candidate matrix; Based on the electrical circuit structure table, the branch relationships, terminal relationships, and adjacent circuit relationships among candidate objects are determined. Candidate objects in the potential hazard object candidate matrix are merged and verified, retaining those with continuous component responses and consistent circuit locations. A potential hazard object discrimination matrix is ​​then generated, in which: The hazard object discrimination matrix is ​​generated as follows: Candidate objects in the hazard object candidate matrix are grouped according to hazard type, collection time window, object number, and loop object number; Based on the branch relationships, terminal relationships, and adjacent circuit relationships in the electrical circuit structure table, determine the positional relationships between candidate objects within the same group; When candidate objects of the same type of hazard are located in the same branch, the same terminal or adjacent circuit, and the component responses change in the same direction within the continuous acquisition time window, the corresponding candidate objects will be merged into the same hazard object. When there is no branch relationship, terminal relationship or adjacent circuit relationship between candidate objects of the same hidden danger type, or when the component response does not change in the same direction within the continuous acquisition time window, the corresponding candidate objects will be retained as independent hidden danger objects. The merged hazard objects and independent hazard objects are written into the matrix positions according to the hazard type, loop object number, acquisition time window, component response and loop location to generate a hazard object discrimination matrix.

[0030] In this embodiment, generating electrical fire hazard detection and early warning results includes: Based on the loop object number and data collection time window, the hazard object, hazard component type, component response, and loop location are extracted from the hazard object discrimination matrix and matched with the preset warning level boundary table to generate a warning level identifier, wherein: Generate an early warning level identifier, specifically as follows: According to the type of hazard component, the corresponding level boundary record is matched from the preset warning level boundary table. The level boundary record includes the first level boundary value, the second level boundary value, and the third level boundary value. The component response is compared with the first-level boundary value, the second-level boundary value, and the third-level boundary value; When the component response is less than the first-level boundary value, a low-level warning level indicator is generated; When the component response is greater than or equal to the first-level boundary value and less than the second-level boundary value, an intermediate warning level identifier is generated; When the component response is greater than or equal to the second-level boundary value and less than the third-level boundary value, a high-level warning level identifier is generated; When the component response is greater than or equal to the third-level boundary value, an emergency warning level identifier is generated; When the same potential hazard has more than two warning level labels within the same data collection time window, the warning level label with the highest level will be used as the warning level label corresponding to the potential hazard. Based on the electrical circuit structure table, the potential hazards are mapped to circuits, branches, terminals, and electrical equipment, generating hazard location identifiers and hazard object identifiers, among which: Generate hazard location markers and hazard identification markers, specifically: Match the corresponding circuit objects, branch objects, terminal objects, switch objects, protection devices, and electrical equipment objects in the electrical circuit structure table according to the hazard object number, hazard type, and circuit location; When the type of hazard is leakage hazard or overload temperature rise hazard, mark the corresponding circuit object, branch object or electrical equipment object as the hazard location object, and mark the corresponding protection device, switch object or electrical equipment object as the disposal object; When the hazard type is electric arc hazard or contact overheating hazard, mark the corresponding circuit object, branch object or terminal object as the hazard location object, and mark the corresponding switch object, terminal object or electrical equipment object as the treatment object; Generate hazard location identifiers for hazard locations based on object number and object type; generate disposal object identifiers for disposal objects based on object number, object type, and the object number of the circuit to which they belong. By associating the hazard object, hazard type, warning level identifier, hazard location identifier, and disposal object identifier, electrical fire hazard detection and warning results are generated.

[0031] Example 1: To verify the feasibility of this invention in practice, it was applied to a low-voltage power distribution monitoring scenario. The distribution cabinet includes 12 outgoing circuits, 36 branch circuits, 94 terminal connection points, and 28 main electrical devices. Sensors are arranged on the incoming side, branch circuit side, terminal connection points, and in the cabinet environment. During one continuous monitoring cycle, a total of 172,800 multi-source sensor data points related to electrical fires were collected, with a sampling interval of 10 seconds. The collected data included phase current, neutral current, residual current, voltage, load power, power factor, harmonics, arc disturbance, current spikes, voltage drops, terminal temperature rise, cable temperature, cabinet temperature, ambient temperature, ambient humidity, switch status, and the start / stop status of electrical devices. The collected electrical circuit configuration data included 12 circuit numbers, 36 branch circuit numbers, 94 terminal circuit numbers, 12 circuit breaker numbers, 28 electrical device numbers, circuit topology connections, and sensor-corresponding circuit relationships.

[0032] After the data enters the processing flow, preprocessing is performed on the raw data. Before processing, the raw sensor data contained 214 duplicate records, 386 missing sampling points, and 129 abnormal jump points, with abnormal jumps mainly concentrated at the moment of equipment start-up and shutdown and in the arc disturbance channel. After deleting duplicate records, filling in missing samples, correcting abnormal jumps, and aligning the unified sampling time axis, a standardized electrical fire sensor dataset was formed. The data integrity rate increased from 99.68% to 99.96%, and the time offset of different sensor channels decreased from a maximum of 3.2 seconds to less than 0.4 seconds. After generating an electrical circuit structure table based on the electrical circuit configuration data, each sensor corresponds to a specific circuit object, branch object, terminal object, or electrical equipment object. Among them, the R03 circuit corresponds to 3 branches, 8 terminals, and 2 main load devices.

[0033] Based on the electrical circuit structure table, circuit object mapping was performed on the standardized electrical fire sensor dataset. Taking circuit R03 as an example, this circuit formed 360 sampling points within a 60-minute monitoring window, with each sampling point containing 17 types of sensor channels. After the sensor channels were associated, the average value of the phase current channel of circuit R03 was 9.8A, the average value of the neutral current channel was 9.5A, the average value of the residual current channel was 14.6mA, the average value of the terminal temperature rise channel was 8.4℃, and the arc disturbance channel showed a short pulse from the 37th to the 46th minute. After being arranged according to the 60-minute time window, 4800 multi-source hidden danger time sequence samples were obtained, including 3100 normal operation samples, 620 equipment start-up and shutdown disturbance samples, 410 leakage hazard samples, 260 arc disturbance samples, 280 contact overheating samples, and 130 overload temperature rise samples.

[0034] After the multi-source hidden danger time series samples are fed into the improved WFTNet model, wavelet Fourier time-frequency transform is first performed. Taking a sample of the R03 loop as an example, the Fourier periodicity feature shows that the main load period of this loop during the normal operation window is concentrated around 12 minutes and 30 minutes, and the average response of the corresponding frequency segment in the loop's normal spectral library is 0.22. In the current monitoring window, a continuous load increase occurs from the 18th minute to the 34th minute, with the load power increasing from 1.42kW to 2.86kW. After the current Fourier periodicity feature is aligned with the periodic position of the loop's normal spectral library, the load spectrum residual response increases from 0.18 to 0.49, lasting for 16 minutes. The loop's normal spectral library memory reconstruction layer backfills the residual of this deviated segment according to the periodic position, generating a load spectrum residual representation.

[0035] In the wavelet local frequency band feature branch, the arc extinction segment buffer fidelity layer performs short-time pulse search on the R03 sample. Arc candidate packets appear at 39 minutes 12 seconds, 41 minutes 30 seconds, 43 minutes 20 seconds, and 45 minutes 10 seconds, with each packet lasting 0.5 to 0.9 seconds. The current spike amplitude is 2.6 to 3.4 times the normal fluctuation average, accompanied by a voltage drop of 3.2V to 4.9V. At 36 minutes, a device startup disturbance occurs, lasting 2.4 seconds, with a synchronous increase in load power of 0.58kW. This segment is written into the bypass channel and separated from the arc buffer group. After buffering and continuous completion of the arc initiation segment, arc extinction interval segment, and re-arrival segment, the mean local response of the arc in the arc extinction fidelity representation reaches 0.76, and the response of the device start-stop segment is reduced to 0.19.

[0036] The prior generation layer for potential hazards establishes candidate component channels based on the load spectrum residual representation, arc extinction fidelity representation, zero-sequence current correspondence under the same circuit object, and current-temperature rise correspondence. In sample R03, the difference between phase current and neutral current increases from 0.3A to 0.8A after 42 minutes, and the residual current increases from 11mA to 24mA, with a zero-sequence direction candidate channel response of 0.63. After the load power drops to 1.92kW at 35 minutes, the terminal temperature rise still increases from 10.8℃ to 16.2℃, with an electrothermal direction candidate channel response of 0.81. The arc direction candidate channel response is 0.76, and the load reference direction candidate channel response is 0.44. After time-frequency location marking, channel normalization, and prior backfilling, a time-frequency feature set of electrical fire hazards is formed.

[0037] After the time-frequency feature set of electrical fire hazards enters the SGMD algorithm, it is first serialized into components. In sample R03, the time-frequency segments of phase current, neutral current, and residual current in the zero-sequence direction form the zero-sequence trajectory vector; the load spectrum residual segments and terminal temperature rise time-frequency segments in the electrothermal direction form the electrothermal residual sequence; the arc extinguishing fidelity segments in the arc direction form the arc candidate sequence; and the load cycle segments in the load reference direction form the load reference sequence. After performing zero-sequence current symplectic trajectory rotation separation processing, the zero-sequence trajectory vector is separated into load disturbance trajectory, leakage current change trajectory, and sudden disturbance trajectory, with a response of 0.58 for the leakage current change trajectory, 0.46 for the sudden disturbance trajectory, and 0.39 for the load disturbance trajectory.

[0038] Hamiltonian matrices were constructed and symplectic eigenvalue decomposition was performed on leakage current change trajectories, sudden disturbance trajectories, and candidate arc sequences. The leakage current change trajectory exhibited a continuous enhancement state, with the top three feature pairs contributing 68.4% of the symplectic feature pair ranking results. The sudden disturbance trajectory and the arc candidate sequence showed consistent extinguishing segment distributions, with the top two feature pairs contributing 61.7%. After reconstruction using the corresponding feature pairs, the symplectic geometric component response of the zero-sequence leakage current was 0.62, and the response of the arc disturbance was 0.57. After the electrothermal residual sequence was written into the electrothermal Hamiltonian matrix, the load spectrum residual segment and the temperature rise time-frequency segment showed synchronous deviation within the same time window, and the ranking of the electrothermal-related symplectic eigenvalues ​​was adjusted from 5th to 2nd, resulting in the reconstructed symplectic geometric component of electrothermal instability with a response value of 0.84. Combining the load disturbance trajectory and the load reference sequence, the load reference symplectic geometric component was reconstructed, with a response value of 0.41. The combination of these four types of symplectic geometric components formed a set of electrical fire hazard modal components.

[0039] Based on the electrical circuit structure table, the zero-sequence leakage current symmetric geometric components, arc disturbance symmetric geometric components, electrothermal instability symmetric geometric components, and load reference symmetric geometric components of sample R03 are mapped to the component-circuit correspondence matrix. Branch B03-2, terminal T03-6, and electrical equipment E03-1 of R03 show a concentrated response within the same acquisition time window. The electrothermal instability response of terminal T03-6 is 0.84, the zero-sequence leakage current response is 0.62, the arc disturbance response is 0.57, and the load reference deviation response is 0.41. After transmission through adjacent circuits, upstream and downstream branches, and the same terminal, the response of the main circuit of R03 is 0.66, the response of branch B03-2 is 0.78, and the response of terminal T03-6 is 0.86, generating a circuit hazard index matrix. After calculating the response of the hidden danger component based on the circuit hidden danger index matrix, the hidden danger object discrimination matrix determines that terminal T03-6 is dominated by the contact overheating hidden danger, accompanied by zero-sequence leakage current change and short-term arc disturbance. The warning level mapping result is medium risk, and the treatment object matching result is branch B03-2 of R03 and terminal T03-6.

[0040] In the comparative verification, the fixed threshold method set the terminal temperature rise alarm boundary to 20℃, the residual current alarm boundary to 30mA, and the arc pulse count alarm boundary to 8 times. In this sample, the highest terminal temperature rise was 16.2℃, the highest residual current was 24mA, and there were 4 arc pulse candidate packets. The fixed threshold method did not output a valid warning, only providing a normal load fluctuation indication. The ordinary time series classification method output an abnormal probability of 0.64, classifying this sample as an abnormal load fluctuation, but failing to locate the terminal object. In the risk warning output of this invention, the hidden danger location is R03-B03-2-T03-6, and the objects to be dealt with are the terminal connection point and the corresponding branch. The warning trigger time is approximately 13.6 minutes earlier than the estimated time when the terminal temperature rise reaches 20℃.

[0041] In a comparative experiment with 1,000 test samples, the overall accuracy of the traditional fixed threshold method was 79.1%, the false alarm rate was 15.2%, the false negative rate was 18.7%, the average early warning lead time was 4.8 minutes, and the terminal-level positioning accuracy was 53.4%.

[0042] The overall accuracy of the ordinary time series classification method is 86.9%, the false alarm rate is 9.4%, the missed detection rate is 11.6%, the average early warning lead time is 7.9 minutes, and the terminal-level positioning accuracy is 65.1%.

[0043] The overall accuracy of this invention is 94.2%, the false alarm rate is 4.3%, the missed detection rate is 5.8%, the average early warning lead time is 14.1 minutes, and the terminal-level positioning accuracy is 86.4%.

[0044] In the arc disturbance samples, the recall rate of the fixed threshold method was 62.0%, the recall rate of the ordinary time-series classification method was 77.5%, and the recall rate of this invention was 88.7%.

[0045] In the case of overheated contact samples, the recall rate of the fixed threshold method was 58.1%, the recall rate of the ordinary time-series classification method was 74.6%, and the recall rate of this invention was 92.3%.

[0046] As can be seen from Example 1, the present invention forms a traceable data change process in each step of the process, which can extract stable hidden danger characteristics under the influence of normal load fluctuations, equipment start-up and shutdown disturbances and environmental changes, and improve the identification accuracy, early warning amount and hidden danger location accuracy in early leakage, arc disturbance, contact overheating and overload temperature rise scenarios.

[0047] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A method for detecting and warning of electrical fire hazards based on intelligent sensors, characterized in that, include: Collect multi-source sensor data and electrical circuit configuration data for electrical fires, preprocess the multi-source sensor data for electrical fires to generate a standardized electrical fire sensor dataset, and generate an electrical circuit structure table based on the electrical circuit configuration data; Based on the electrical circuit structure table, the standardized electrical fire sensor dataset is mapped to circuit objects, associated with sensor channels, and arranged with time windows to generate time-series samples of multi-source hazards. Based on the improved WFTNet model, wavelet Fourier time-frequency transform is performed on the time series samples of multi-source hidden dangers. Load spectrum residuals are extracted by combining the memory reconstruction of the normal circuit spectrum library. Local pulse preservation is achieved by using the arc extinguishing segment buffer fidelity. The time-frequency features of zero sequence, electrothermal, arc and load reference direction are generated and sorted through the prior of hidden danger components to obtain the time-frequency feature set of electrical fire hidden dangers. The SGMD algorithm is used to perform symplectic geometric mode decomposition on the time-frequency feature set of electrical fire hazards. Based on the rotational separation process of zero-sequence current symplectic trajectory, the zero-sequence related symplectic trajectory is split. The electrothermal residual symplectic eigenvalue recalibration process is introduced to correct the electrothermal related symplectic eigenvalue, and the set of modal components of electrical fire hazards is obtained. Based on the set of electrical fire hazard modal components and the electrical circuit structure table, a component-circuit correspondence matrix is ​​constructed and time-series alignment and adjacency transfer are performed to obtain the circuit hazard index matrix; The hazard component response of each loop object is calculated based on the loop hazard index matrix, and a hazard object discrimination matrix is ​​generated according to the combination relationship of the hazard component responses. Based on the hazard object identification matrix, the warning level mapping and disposal object matching are performed to generate electrical fire hazard detection and warning results.

2. The method for detecting and warning of electrical fire hazards based on intelligent sensors according to claim 1, characterized in that, The specific multi-source sensing data for electrical fires includes phase current data, neutral current data, residual current data, voltage data, load power data, power factor data, harmonic data, arc disturbance data, current spike data, voltage drop data, terminal temperature rise data, cable temperature data, distribution box internal temperature data, ambient temperature data, ambient humidity data, switch status data, and electrical equipment start / stop status data.

3. The method for detecting and warning of electrical fire hazards based on intelligent sensors according to claim 1, characterized in that, The electrical circuit configuration data specifically includes circuit number, branch number, terminal number, switch number, circuit breaker number, electrical equipment number, circuit topology connection relationship, branch power supply relationship, terminal connection relationship, phase line connection relationship, neutral line connection relationship, grounding connection relationship, rated voltage, rated current, rated power, wire specifications, protection device parameters, sensor installation location, and sensor corresponding circuit relationship.

4. The method for detecting and warning of electrical fire hazards based on intelligent sensors according to claim 1, characterized in that, The generation of the standardized electrical fire sensor dataset and electrical circuit structure table includes: Collect multi-source sensor data on electrical fires and electrical circuit configuration data, and establish original data association records according to sensor number, acquisition time, sensor type and circuit number; The original data association records of electrical fire multi-source sensor data are processed by deleting duplicate records, removing invalid data, filling missing samples, and correcting abnormal jumps to generate cleaned sensor data. The cleaned sensor data is time-aligned according to a unified sampling time axis, and the data of different sensor types are dimensionally unified and numerically standardized to generate a standardized electrical fire sensor dataset. Based on the object number, connection relationship and sensor installation location in the electrical circuit configuration data, establish structured associations between circuit objects, branch objects, terminal objects, switch objects, electrical equipment objects and sensor objects, and generate an electrical circuit structure table.

5. The method for detecting and warning of electrical fire hazards based on intelligent sensors according to claim 1, characterized in that, The generation of multi-source hidden danger time series samples includes: Based on the electrical circuit structure table, a circuit object index table is established, and the standardized electrical fire sensing dataset is mapped to the corresponding circuit object according to the sensor number and the corresponding circuit relationship of the sensor, thus generating a circuit object sensing data table. According to the sensing type, the sensing data table of the loop object is associated with sensing channels to generate current channel, voltage channel, residual current channel, temperature channel, arc disturbance channel, environment channel and status channel; Arrange the time windows of each sensing channel according to a unified sampling time axis to generate a loop sensing time window sequence. Based on the loop object index table, loop sensing time window sequence, and electrical loop structure table, establish sample association relationships for adjacent loops, the same branch, the same terminal, and the same electrical equipment, and generate multi-source hidden danger time sequence samples.

6. The method for detecting and warning of electrical fire hazards based on intelligent sensors according to claim 1, characterized in that, The obtained time-frequency feature set of electrical fire hazards includes: An improved WFTNet model is constructed by setting the normal circuit spectrum library memory reconstruction layer on the Fourier periodic feature branch, setting the arc extinguishing segment cache fidelity layer on the wavelet local frequency band feature branch, and setting the hidden danger component prior generation layer after the Fourier periodic feature branch and the wavelet local frequency band feature branch. The normal loop spectrum library memory reconstruction layer archives the Fourier periodic features of multi-source hidden danger time series samples according to the loop object and normal operation window identifier, generates the normal loop spectrum library, aligns the Fourier periodic features of the current monitoring window with the normal loop spectrum library, extracts the frequency segments and durations that deviate from the normal operation spectrum of the current monitoring window, and performs residual backfilling according to the periodic position to generate the load spectrum residual representation. The arc extinction segment cache fidelity layer performs short-time pulse search on the wavelet local frequency band features of the multi-source hidden danger time series sample, extracts arc candidate wave packets, and caches and groups the arc candidate wave packets according to the arc initiation segment, arc extinction interval segment, and arc re-ignition segment. According to the equipment start-up and shutdown status, the start-up and shutdown disturbance segments are written into the bypass channel and separated from the cache group. The arc candidate wave packets in the cache group are continuously filled to generate an arc extinction fidelity representation. The prior generation layer for hidden danger components establishes candidate component channels in the zero-sequence direction, electrothermal direction, arc direction, and load reference direction based on the load spectrum residual representation, arc extinguishing fidelity representation, and the zero-sequence current correspondence and current-temperature rise correspondence under the same circuit object. It performs time-frequency position marking, channel normalization, and prior backfilling on each candidate component channel to generate a prior representation of hidden danger components and form a time-frequency feature set of electrical fire hazards. The improved WFTNet model was trained using a combination of time-frequency reconstruction error, load spectrum residual backfilling error, arc extinction buffer fidelity error, and hazard component prior discrimination error as joint optimization objectives. The network parameters of the normal circuit spectrum library memory reconstruction layer, the arc extinction segment buffer fidelity layer, and the hazard component prior generation layer were continuously optimized. When the change in the joint loss value in five consecutive training rounds was less than 0.001, the improved WFTNet model was considered to have completed convergence training.

7. The method for detecting and warning of electrical fire hazards based on intelligent sensors according to claim 1, characterized in that, The obtained set of electrical fire hazard modal components includes: The time-frequency feature set of electrical fire hazards is serialized into components, and zero-sequence trajectory vector, electrothermal residual sequence, electric arc candidate sequence and load reference sequence are generated according to the zero-sequence direction, electrothermal direction, electric arc direction and load reference direction, respectively. Perform zero-sequence current symplectic trajectory rotation separation processing, write the zero-sequence trajectory vector into the symplectic trajectory matrix, establish the common-mode load direction, differential current leakage direction and sudden disturbance direction in the symplectic trajectory matrix, rotate the direction of the zero-sequence trajectory vector, and separate the load disturbance trajectory, leakage current change trajectory and sudden disturbance trajectory. Hamiltonian matrices are constructed and symplectic eigenvalues ​​are performed on leakage current change trajectories, sudden disturbance trajectories and arc candidate sequences. Symplectic feature pairs are arranged according to the continuous enhancement state of leakage current change trajectories, the short-term concentrated state of sudden disturbance trajectories and the distribution of quenching segments of arc candidate sequences, and the corresponding zero-sequence leakage current symplectic geometric components and arc disturbance symplectic geometric components are reconstructed. Perform electrothermal residual symmetric eigenvalue recalibration processing, write the electrothermal residual sequence into the electrothermal Hamilton matrix and perform symmetric eigenvalue decomposition, recalibrate the electrothermal related symmetric eigenvalues ​​according to the synchronous deviation state of the load spectrum residual segment and the temperature rise time frequency segment in the same time window, and reconstruct the corresponding electrothermal instability symmetric geometric components. The load reference symmetric geometric components are reconstructed based on the load disturbance trajectory and load reference sequence, and the zero-sequence leakage symmetric geometric components, arc disturbance symmetric geometric components, electrothermal instability symmetric geometric components and load reference symmetric geometric components are combined to form a set of electrical fire hazard mode components.

8. The method for detecting and warning of electrical fire hazards based on intelligent sensors according to claim 1, characterized in that, The obtained loop hazard index matrix includes: Establish a sequence of loop objects and a table of loop adjacency relationships based on the electrical loop structure table; Based on the loop object sequence and the acquisition time window, the zero-sequence leakage current symplectic geometric component, arc disturbance symplectic geometric component, electrothermal instability symplectic geometric component and load reference symplectic geometric component in the electrical fire hazard modal component set are mapped to the corresponding matrix positions to construct the component-loop correspondence matrix. Based on the loop adjacency table, the adjacent loops, upstream and downstream branches and component responses corresponding to the same terminal in the component-loop correspondence matrix are passed on adjacency to generate the component-loop correspondence matrix after adjacency passing. Based on the component-loop correspondence matrix after adjacency propagation, a loop hazard index matrix is ​​generated according to the loop object number, hazard component type, acquisition time window, and component response.

9. The method for detecting and warning of electrical fire hazards based on intelligent sensors according to claim 1, characterized in that, The generated hazard object discrimination matrix includes: Based on the loop object number and the acquisition time window, the zero-sequence leakage response, arc disturbance response, electrothermal instability response and load reference deviation response are extracted from the loop hidden danger index matrix to generate a hidden danger component response matrix; Based on the combination relationship between zero-sequence leakage response, arc disturbance response, electrothermal instability response and load reference deviation response, the objects of each circuit are identified as leakage hazards, arc hazards, contact overheating hazards and overload temperature rise hazards, and a candidate matrix of hazard objects is generated. Based on the electrical circuit structure table, determine the branch relationships, terminal relationships, and adjacent circuit relationships among candidate objects. Merge and verify the candidate objects in the candidate matrix of potential hazards, retain the candidate objects with continuous component responses and consistent circuit positions, and generate the potential hazard discrimination matrix.

10. The method for detecting and warning of electrical fire hazards based on intelligent sensors according to claim 1, characterized in that, The generated electrical fire hazard detection and early warning results include: Based on the loop object number and the collection time window, extract the hidden danger object, hidden danger component type, component response and loop location from the hidden danger object discrimination matrix, and match them with the preset warning level boundary table to generate a warning level identifier; Based on the electrical circuit structure table, the potential hazards are mapped to circuits, branches, terminals, and electrical equipment, generating hazard location identifiers and hazard object identifiers; By associating the hazard object, hazard type, warning level identifier, hazard location identifier, and disposal object identifier, electrical fire hazard detection and warning results are generated.

Images