Electrical fire early warning algorithm based on multi-source data collaborative awareness

Abstract

The invention discloses an electrical fire early warning algorithm based on multi-source data collaborative perception, and belongs to the technical field of fire early warning. The method comprises the following steps: firstly, synchronously acquiring multi-source time sequence data such as current, temperature and images, and extracting key characteristics such as harmonic waves, temperature rise and zero-sequence current spectrum; and performing dynamic weight fusion on the features by using an attention mechanism neural network, and inputting the fused features into an improved isolated forest algorithm to identify transient and steady state abnormal modes. And then, combining historical data and equipment archives, and dynamically calculating a fire risk level by using a time sequence risk assessment model. And finally, a grading early warning signal and a diagnosis suggestion are output according to the risk grade, and linkage control is automatically started when the risk is high. According to the method, the defects of high false alarm rate and early warning lagging of a traditional single-parameter threshold method are effectively overcome, and early-stage accurate early warning and intelligent active defense of electrical fire hazards are realized.

Description

Technical Field

[0001] This invention provides an electrical fire early warning algorithm based on multi-source data collaborative sensing, belonging to the field of fire early warning technology. Background Technology

[0002] Electrical fires, due to their suddenness, high concealment, and difficulty in extinguishing, have become one of the major threats to fire safety in modern buildings. Existing electrical fire early warning technologies largely rely on threshold monitoring of single physical quantities, such as overload warnings based on the effective value of current or leakage warnings based on residual current. These methods have significant shortcomings: First, single-parameter warnings have a high false alarm rate, making it difficult to distinguish between normal equipment start-up and shutdown, impact loads, and actual faults; second, they lack the ability to identify early, latent faults, typically only triggering alarms after obvious dangerous characteristics such as electric arcs or high temperatures appear, thus losing valuable early intervention time; finally, the system's intelligence level is low, unable to perform correlation analysis and status evolution assessment of multi-type, cross-regional data, making it difficult to achieve the leap from "single-point alarm" to "system diagnosis," resulting in severely insufficient accuracy and foresight in early warnings, failing to meet the urgent needs of modern smart fire protection for precise and proactive early warnings. Summary of the Invention

[0003] The technical problem to be solved by this invention is that electrical fire early warning technology suffers from high false alarm rate, insensitivity to early hidden faults, and lack of multi-source data correlation analysis capabilities due to its reliance on single parameter threshold monitoring, thus failing to achieve accurate and advanced early warning.

[0004] To address the aforementioned problems, the present invention proposes the following technical solution: an electrical fire early warning method based on multi-source data collaborative sensing, comprising the following steps:

[0005] S1. Real-time acquisition of multi-source electrical time-series data within the monitoring area, wherein the multi-source electrical time-series data includes at least: three-phase current of the line, zero-sequence current, cable temperature and ambient temperature;

[0006] S2. Preprocess and extract features from the multi-source electrical timing data to construct a dynamic feature vector containing time-domain, frequency-domain, and time-frequency-domain features;

[0007] S3. Input the dynamic feature vector into a pre-trained multi-level early warning decision model, the multi-level early warning decision model including a feature fusion layer, an anomaly detection layer and a risk assessment layer connected in sequence;

[0008] S4. The anomaly detection layer identifies transient and steady-state anomaly patterns in electrical circuits based on an improved isolated forest algorithm;

[0009] S5. The risk assessment layer calculates the current electrical fire risk level based on the anomaly detection results, combined with historical early warning data and equipment file information;

[0010] S6. Output the warning signal and diagnostic suggestions corresponding to the risk level.

[0011] Preferably, in step S1, the multi-source electrical timing data further includes: line voltage, ambient humidity, arc light intensity signal, and distribution box image data.

[0012] Preferably, the feature extraction in step S2 includes:

[0013] The harmonic distortion rate and three-phase imbalance of the current signal are calculated, the temperature rise rate and spatial temperature difference distribution of the temperature data are extracted, and the energy characteristics of a specific frequency band are extracted from the zero-sequence current.

[0014] Preferably, the feature fusion layer in step S3 adopts a neural network model based on an attention mechanism to perform weighted fusion of features from different sources and at different scales to generate a fused feature vector.

[0015] Preferably, step S4 specifically includes:

[0016] S41. Use the fused feature vectors to train multiple isolated forest models for different failure modes;

[0017] S42. Input the real-time fused feature vectors into each isolated forest model and calculate the comprehensive anomaly score;

[0018] S43. If the overall anomaly score exceeds the first threshold, it is determined that there is an anomaly, and the anomaly pattern category is located.

[0019] Preferably, the risk assessment layer in step S5 adopts a time-series risk assessment model, and the formula for calculating the risk level R is:

[0020] R = α*S_t + β*Σ(S_hist) + γ*C

[0021] Where S_t is the current comprehensive anomaly score, S_hist is the weighted sum of the historical anomaly score sequence, C is the risk coefficient based on equipment type, aging degree and environmental factors, and α, β and γ are dynamically adjusted weight parameters.

[0022] Preferably, the warning signal output in step S6 is divided into multiple levels, including:

[0023] Level 1 Warning: Equipment status monitoring provides alerts and maintenance recommendations;

[0024] Level 2 Warning: Warning of potential electrical fire hazards; on-site inspection recommended.

[0025] Level 3 warning: High risk of electrical fire alarm, and linked control device to execute power cut-off or audible and visual alarm.

[0026] An electrical fire early warning system based on multi-source data collaborative sensing includes:

[0027] A multi-source data acquisition module is used to perform step S1 as described in claim 1;

[0028] A data preprocessing and feature engineering module is used to perform step S2 as described in claim 1;

[0029] A multi-level early warning decision-making model module, comprising a feature fusion submodule, an anomaly detection submodule, and a risk assessment submodule, is used to execute steps S3-S5 as described in claim 1;

[0030] The early warning and linkage control module is used to execute step S6 as described in claim 1.

[0031] The beneficial effects of this invention are:

[0032] This invention overcomes the limitations of traditional single-parameter threshold methods by constructing dynamic feature vectors through the fusion of heterogeneous data from multiple sources, including current, temperature, leakage current, and environmental data. The core of this invention lies in employing a multi-level early warning decision-making model that incorporates feature fusion, anomaly detection, and risk assessment. In particular, it utilizes an improved isolated forest algorithm to accurately identify early, subtle anomaly patterns from complex data, such as intermittent arcing caused by poor contact. This model combines historical trends and equipment status for comprehensive risk assessment, achieving a leap from "anomaly alarm" to "risk prediction." Ultimately, the system can output tiered early warnings and specific diagnostic suggestions, with a low false alarm rate and significantly earlier warning times. This provides an intelligent and precise solution for electrical fire prevention, greatly enhancing the initiative and effectiveness of electrical safety management. Attached Figure Description

[0033] Figure 1 This is a schematic diagram of the process of the present invention. Detailed Implementation

[0034] The following embodiments further illustrate the present invention.

[0035] Example 1: Refined Online Early Warning and Diagnosis for Distribution Boxes

[0036] This embodiment uses a low-voltage distribution box in a commercial building that supplies power to a data center as the monitoring object to demonstrate the algorithm's ability to identify early potential problems in critical loads.

[0037] 1. System Deployment and Data Acquisition:

[0038] Inside the target distribution box, a multi-sensor fusion deployment is implemented. Specifically, this includes: high-precision open-type current transformers (CTs) connected to the A, B, and C phase busbars and the neutral line, with a sampling frequency of 10kHz to capture microsecond-level current transients; surface-mount digital temperature sensors installed at six key temperature measurement points, including circuit breaker inlet and outlet terminals and copper busbar connections; a zero-sequence current transformer connected to the total path of the three-phase cables; and a miniature wide-angle thermal imaging camera installed inside the distribution box door, capturing temperature field distribution maps of the entire box at a frequency of one frame per minute. All sensor data is synchronized and preprocessed by a built-in intelligent data acquisition unit before being uploaded to a locally deployed edge server via industrial Ethernet.

[0039] 2. Data Processing and Feature Engineering:

[0040] After receiving the raw data stream, the edge server performs the following feature extraction operations:

[0041] Current signal analysis: Perform FFT transformation on each phase current to calculate the content of harmonics up to the 31st harmonic, paying particular attention to the variation trend of the 3rd, 5th, and 7th harmonics; calculate the unbalance of the three-phase current; use wavelet transform to analyze whether there are transient pulse groups in the current waveform with characteristic frequency bands (such as 3-10kHz), which are usually characteristics of early arcs.

[0042] Temperature signal analysis: Calculate the real-time temperature rise (difference from ambient temperature) and its rate of change (dT / dt) at each temperature measurement point. Construct a two-dimensional temperature field model based on the temperature data from the six points and calculate the maximum spatial temperature difference (ΔT_max).

[0043] Thermal imaging analysis: Using image recognition algorithms, the outlines of equipment such as circuit breakers and terminals are automatically selected from thermal images, and their surface maximum temperature, average temperature, and temperature standard deviation are extracted.

[0044] Feature fusion: Combine all the above features (52 dimensions in total) with information such as timestamps and load rates (calculated based on current) to form a dynamically updated comprehensive feature vector with a time window of 5 minutes.

[0045] 3. Model Inference and Early Warning Generation:

[0046] The comprehensive feature vector of the distribution box is pushed to the multi-level early warning decision model on the cloud platform in real time. The model operation record is as follows:

[0047] During 72 hours of continuous monitoring, the model found that the third harmonic content of the C-phase current slowly increased from an initial 15% to 22%. Simultaneously, the temperature rise rate at the measuring point of the lower terminal of the C-phase remained stable at 0.8℃ / hour, while the temperature rise rate of other phase terminals was less than 0.2℃ / hour. Thermal imaging analysis also revealed a gradually expanding "hot spot" in this terminal area.

[0048] The attention mechanism network of the feature fusion layer automatically assigns high weights to "harmonic variation rate" and "local temperature rise rate".

[0049] The anomaly score calculated by the isolated forest model dedicated to the "poor contact" of the anomaly detection layer continues to accumulate and eventually exceeds the threshold.

[0050] The risk assessment layer retrieved the loop file (the load is a "precision air conditioning unit" that has been in operation for 8 years) and, combined with the continuous abnormal patterns, calculated that the fire risk index R had risen to the "high risk" level.

[0051] Warning Output: The system immediately generates a Level 2 warning: "Warning Location: 3F Distribution Box - Circuit C3; Warning Type: Abnormally increased contact resistance accompanied by increased harmonics; Risk Level: High; Possible Cause: Loose or oxidized wiring at the lower end of the C-phase circuit breaker; Handling Recommendation: Immediately shut off power, inspect and tighten the C3 circuit terminals, and test the filter device at the power input end of the air conditioning unit." This warning is sent to maintenance personnel via APP and SMS, and a maintenance work order is generated.

[0052] 4. Effect: Traditional methods typically only trigger an alarm when the temperature exceeds 70℃ or harmonics are severely exceeded. This embodiment identifies potential hazards approximately 48 hours in advance, even when the terminal temperature reaches only 55℃ and harmonics are within acceptable limits, thus preventing insulation carbonization or even fire that could be caused by overheating of the contact point.

[0053] Example 2: Multi-loop correlation analysis and regional collaborative early warning for building floors

[0054] This embodiment uses the entire office floor of a comprehensive office building as the monitoring scope to demonstrate the algorithm's ability to perform macro-situational awareness and latent fault location in complex power environments.

[0055] 1. System deployment and data synchronization:

[0056] Within the electrical shaft of this floor, comprehensive monitoring is conducted on the 12 outgoing circuits of the distribution cabinet (including lighting, sockets, air conditioning, emergency power, etc.). A monitoring unit is deployed on each circuit to simultaneously collect current, zero-sequence current, and cable sheath temperature. Additionally, four environmental monitoring points are installed in the common areas of the floor to collect ambient temperature and humidity data. All monitoring units are synchronized using a high-precision hardware clock to ensure time alignment of data across the entire floor, with an error of less than 1 millisecond.

[0057] 2. Cross-loop association feature extraction:

[0058] After receiving data from all floors, the cloud platform not only performs single-loop analysis but also focuses on extracting cross-loop correlation features:

[0059] Load correlation analysis: Analyze the temporal correlation of current changes in different circuits. For example, it was found that when the current of the "Northwest Area Lighting Circuit (L1)" decreases, a short-term current spike will appear in the "Adjacent Area Socket Circuit (L2)" after hundreds of milliseconds. This pattern occurs frequently after get off work every day.

[0060] Zero-sequence current spectrum analysis: The spectrum analysis of the zero-sequence current of all 12 loops was performed. It was found that the L1 and L2 loops both have intermittent energy pulses in a specific high-frequency range (such as around 5.8kHz), and the timing of the pulses highly overlaps, while other loops do not have this phenomenon.

[0061] Spatiotemporal pattern construction: By combining the above-mentioned correlation features with the physical location information of sensors, an "electrical state spatiotemporal matrix" of the floor is constructed to visualize the distribution and propagation of abnormal patterns.

[0062] 3. Collaborative sensing and early warning:

[0063] One night, the early warning decision model detected a high-frequency energy anomaly in the zero-sequence current of the L1 loop, while the current in the L2 loop showed an unexpected, minor fluctuation. The model activated its correlation analysis engine, confirming that the anomaly pattern was cross-loop and simultaneous.

[0064] The risk assessment team immediately queried the Building Information Modeling (BIM) database to obtain the piping layout drawings for L1 and L2 circuits, discovering that these two lines shared a section of cable tray within the suspended ceiling. Combined with the maintenance log indicating that "partial renovations were carried out on this floor three months ago," the model significantly increased the weighting of the possibility that "damaged insulation between wires caused intermittent discharge to the cable tray."

[0065] Instead of immediately triggering a high-level alarm, the system activated "deep diagnostic mode" and issued a command to temporarily increase the sampling frequency of the relevant circuit to 100kHz, continuously capturing the high-frequency arc current characteristic waveform with an amplitude of several hundred milliamperes for about 2 milliseconds.

[0066] Warning Output: The system generates a Level 3 warning with conclusive evidence: "Warning Area: 5F West Zone; Fault Location: Suspected insulation defect exists between lighting circuit L1 and socket circuit L2 in the No. 03 ceiling cable tray; Risk Level: Emergency (clear arc characteristics detected); Handling Recommendations: Immediately evacuate personnel from the area, remotely disconnect the power supply to circuits L1 and L2, and arrange maintenance personnel to conduct insulation testing and replacement of the designated cable tray section." The warning simultaneously triggers the activation of emergency lighting in the area and notifies the property management duty room.

[0067] 4. Results: In this case, looking at the data of any single circuit, the current and temperature were within the normal range, and the zero-sequence current value was far below the 30mA leakage current alarm threshold, making it completely undetectable by traditional systems. This invention, through multi-circuit data collaborative sensing and correlation analysis, successfully located a major hidden danger that could potentially develop into an arc fire, achieving a fundamental leap from "no-fault alarm" to "early warning of potential disasters."

[0068] The present invention and its embodiments have been described above. This description is not restrictive, and the accompanying drawings are only one embodiment of the present invention; the actual structure is not limited thereto. In conclusion, if those skilled in the art are inspired by this description and design similar structures and embodiments without departing from the spirit of the invention, such designs should fall within the protection scope of the present invention.

Claims

1. An electrical fire early warning method based on multi-source data collaborative sensing, characterized in that, Includes the following steps: S1. Real-time acquisition of multi-source electrical time-series data within the monitoring area, wherein the multi-source electrical time-series data includes at least: three-phase current of the line, zero-sequence current, cable temperature and ambient temperature; S2. Preprocess and extract features from the multi-source electrical timing data to construct a dynamic feature vector containing time-domain, frequency-domain, and time-frequency-domain features; S3. Input the dynamic feature vector into a pre-trained multi-level early warning decision model, the multi-level early warning decision model including a feature fusion layer, an anomaly detection layer and a risk assessment layer connected in sequence; S4. The anomaly detection layer identifies transient and steady-state anomaly patterns in electrical circuits based on an improved isolated forest algorithm; S5. The risk assessment layer calculates the current electrical fire risk level based on the anomaly detection results, combined with historical early warning data and equipment file information; S6. Output the warning signal and diagnostic suggestions corresponding to the risk level.

2. The method according to claim 1, characterized in that, In step S1, the multi-source electrical timing data also includes: line voltage, ambient humidity, arc light intensity signal and distribution box image data.

3. The method according to claim 1, characterized in that, The feature extraction in step S2 includes: The harmonic distortion rate and three-phase imbalance of the current signal are calculated, the temperature rise rate and spatial temperature difference distribution of the temperature data are extracted, and the energy characteristics of a specific frequency band are extracted from the zero-sequence current.

4. The method according to claim 1, characterized in that, The feature fusion layer in step S3 uses a neural network model based on an attention mechanism to weight and fuse features from different sources and at different scales to generate a fused feature vector.

5. The method according to claim 4, characterized in that, Step S4 specifically includes: S41. Use the fused feature vectors to train multiple isolated forest models for different failure modes; S42. Input the real-time fused feature vectors into each isolated forest model and calculate the comprehensive anomaly score; S43. If the overall anomaly score exceeds the first threshold, it is determined that there is an anomaly, and the anomaly pattern category is located.

6. The method according to claim 5, characterized in that, The risk assessment layer described in step S5 adopts a time-series risk assessment model, and the formula for calculating the risk level R is as follows: R = α*S_t + β*Σ(S_hist) + γ*C Where S_t is the current comprehensive anomaly score, S_hist is the weighted sum of the historical anomaly score sequence, C is the risk coefficient based on equipment type, aging degree and environmental factors, and α, β and γ are dynamically adjusted weight parameters.

7. The method according to claim 6, characterized in that, The warning signal output in step S6 is divided into multiple levels, including: Level 1 Warning: Equipment status monitoring provides alerts and maintenance recommendations; Level 2 Warning: Warning of potential electrical fire hazards; on-site inspection recommended. Level 3 warning: High risk of electrical fire alarm, and linked control device to execute power cut-off or audible and visual alarm.

8. An electrical fire early warning system based on multi-source data collaborative sensing, characterized in that, include: A multi-source data acquisition module is used to perform step S1 as described in claim 1; A data preprocessing and feature engineering module is used to perform step S2 as described in claim 1; A multi-level early warning decision-making model module, comprising a feature fusion submodule, an anomaly detection submodule, and a risk assessment submodule, is used to execute steps S3-S5 as described in claim 1; The early warning and linkage control module is used to execute step S6 as described in claim 1.

Images