Method for multi-dimensionally and effectively studying and judging false alarm of fire alarm system in transformation power station

Abstract

The invention discloses a method for effectively studying and judging false alarm of a fire alarm system in a transformation power station in a multi-dimensional mode, and the method comprises the steps: collecting and fusing multi-channel sensor data, such as smoke concentration, temperature, humidity, electromagnetic field intensity, grounding state and the like, and constructing a scenarized data set in combination with an environment label and an operation and maintenance log; through multi-level feature processing such as normalization, wavelet denoising and dimension reduction, the data quality and discriminability are improved; according to the method, the space-time diagram neural network is used for achieving multi-period environment subarea and typical signal mode division, multi-period fusion confidence judgment is conducted on alarm signals based on the fuzzy set theory and Bayesian reasoning, dynamic judgment of false alarms and real alarms is achieved, and the accuracy and robustness of alarm judgment in the complex environment are effectively improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of substation fire monitoring and false alarm discrimination, and particularly relates to a method for multi-dimensional effective research and judgment of false alarms of a fire alarm system in a substation. BACKGROUND

[0002] At present, substation fire monitoring and intelligent false alarm discrimination technology is developing towards high sensitivity, multi-dimensional coordination and intelligence. The industry generally uses a multi-channel sensor network (such as smoke, temperature, humidity, electromagnetic field, grounding state, etc.) to realize real-time perception of the substation environment, and combines certain signal processing and discrimination algorithms to identify potential fire risks. In terms of technical implementation, the mainstream scheme is based on threshold discrimination, feature fusion, rule base, high-frequency data backtracking, machine learning classification, and some advanced work attempts to introduce neural networks and other AI technologies to improve the adaptive ability of abnormal identification. In addition, the system generally designs alarm grading, manual review mechanism and interface with the substation dispatching system to realize preliminary intelligent early warning.

[0003] However, in the actual operating environment, the complexity of the scene and the interference factors faced by substation fire monitoring are extremely significant. Especially under the interference of strong electromagnetic pulse, large equipment start, electrical switching and extreme climate, the original signals collected by the sensor are easily covered by background noise or distorted, causing the characteristics of the real fire signal to be masked, and false alarms (such as electromagnetic interference, abnormality during equipment operation) to occur frequently. These false alarms not only reduce the efficiency of operation and maintenance, but also may mask the real fire, causing great safety hazards.

[0004] The current technical system has the following outstanding problems:

[0005] (1) There is still a lack of effective solution to the problem of complex background interference and abnormal characteristics being easily masked. Traditional discrimination methods are mostly based on static analysis of single period and single type of signal, and it is difficult to break through the interference limit in scene recognition and signal resolution. The response capability to cross-cycle weak or superimposed abnormalities is limited;

[0006] (2) There is a lack of intelligent modeling using multi-dimensional environmental context. The existing algorithms fail to fully integrate time, weather, load, operation and maintenance, and other dynamic multi-source labels, and cannot fully reconstruct the environmental scene, resulting in that the discrimination of false alarms is easily affected by confidence fluctuations and environmental drift;

[0007] (3) The anti-interference feature extraction and adaptive ability are insufficient. It is difficult to accurately model and identify complex noise patterns under electromagnetic and other field interference. The model lacks self-calibration and dynamic parameter adjustment capability after long-term operation, and is prone to failure when encountering environmental changes and sensor aging;

[0008] (4) Multi-period, multi-scene cross-cycle trend judgment is insufficient, most systems focus on single-point or short-term characteristics, and lack an integrated reasoning mechanism for effectively fusing historical distribution, adjacent period trends, scene evolution and other information for false alarm investigation. SUMMARY

[0009] The present application provides a multi-dimensional effective research and judgment method for transforming false alarms of a fire alarm system in a power station.

[0010] The technical solution of the present application is as follows: A multi-dimensional effective research and judgment method for transforming false alarms of a fire alarm system in a power station, comprising:

[0011] S1: Collecting multi-channel sensor data in the substation, the sensor data including smoke concentration, temperature change, humidity, electromagnetic field intensity, line grounding state signal, and combining time stamp, weather information, load type, operation and maintenance activity log to form a scenario-based environment data set;

[0012] S2: Normalizing and denoising the collected sensor data to eliminate the influence of different dimensions and random noise on subsequent analysis;

[0013] S3: Clustering analysis of the normalized multi-dimensional sensor data based on a spatio-temporal graph neural network, dividing multiple time period environment sub-regions with similar scene semantics, and learning typical feature patterns of fire signals and interference noise in each sub-region;

[0014] S4: Constructing a fuzzy confidence interval model, expressing the alarm signal confidence of each feature detection channel in a fuzzy set, and combining historical scene distribution to establish an abnormal signal fuzzy confidence weighted table and a typical noise set for each time period;

[0015] S5: When the alarm signal is triggered, the system calls the historical data of the same scene according to the current scene label, makes a weighted judgment of the alarm signal under the multi-period fuzzy confidence model, and performs Bayesian inference on the abnormal change trend by fusing the detection data before and after the period;

[0016] S6: Based on the inference result, execute false alarm discrimination decision, if the signal shows high confidence anomaly in multiple periods and scenes and is out of the noise fuzzy set, mark it as high-confidence real alarm; if there is a significant intersection between the confidence interval and the noise set, mark it as a high-probability false alarm and push it to the artificial review queue;

[0017] S7: Periodically annotating and updating the fuzzy set of false alarm and real alarm cases, optimizing the fuzzy confidence model parameters using incremental learning mechanism to realize long-term adaptive optimization to the complex environmental changes in the substation;

[0018] S8: Dynamically adjust the threshold parameter of the discrimination model based on historical false alarm events and environmental change cycles to improve the model's adaptability to feature shifts caused by environmental drift and sensor aging.

[0019] The method for effectively judging and transforming false alarms of a fire alarm system in a power station provided by the application has the following beneficial effects:

[0020] (1) The application realizes holographic modeling of alarm features and environmental context (including weather, clock, load, operation and maintenance, etc.) by multi-period, multi-channel environmental perception and discrimination, combined with dynamic clustering modeling of scene semantics by a space-time graph neural network. Compared with traditional methods based only on single-channel or static threshold discrimination, the system can distinguish the evolution patterns of fire signals and noise under different interference backgrounds, significantly improving the discrimination ability for extreme / rapidly changing scene interference (such as large equipment startup, electromagnetic pulse, etc.);

[0021] (2) The application introduces fuzzy set theory and membership function interval modeling driven by historical scene samples, expresses multi-channel alarm confidence as a multi-dimensional fuzzy interval, and realizes adaptive judgment, intersection quantization and trend verification of false alarm patterns through multi-period context weighting and noise clustering interval fusion. Compared with traditional solutions that rely on a single threshold, have large errors and require heavy manual intervention, the application significantly improves the accuracy of false alarm automatic identification in high interference backgrounds, greatly reducing the burden of manual review and resource waste;

[0022] (3) The application dynamically fuses the abnormal evolution trends of the previous and subsequent scenes using fuzzy confidence weighting and Bayesian multi-period reasoning, sensitively distinguishes sudden, slow evolution, persistent or emergent fire alarms, and accurately captures atypical fire conditions. Compared with conventional methods based on single-point static discrimination, the alarm effectiveness is significantly enhanced. BRIEF DESCRIPTION OF DRAWINGS

[0023] Fig. 1 A flowchart of the method for effectively judging and transforming false alarms of a fire alarm system in a power station according to the application;

[0024] Fig. 2 A sub-flowchart of the method for effectively judging and transforming false alarms of a fire alarm system in a power station according to the application;

[0025] Fig. 3 Another sub-flowchart of the method for effectively judging and transforming false alarms of a fire alarm system in a power station according to the application. DETAILED DESCRIPTION

[0026] In order to make the objectives and advantages of the application clearer and more apparent, the application will be further described below with reference to examples; it should be understood that the specific examples described herein are only intended to explain the application and not to limit the application.

[0027] The preferred embodiments of the present application will be described herein below with reference to the accompanying drawings. Those skilled in the art will appreciate that the embodiments are only used to explain the technical principles of the present application, and are not intended to limit the protection scope of the present application.

[0028] As used herein, the singular forms "a," "an," and "the" include plural referents unless the context clearly dictates otherwise. It will be further understood that the terms "comprises," "comprising," "includes," "including," or "has," "having," and the like, when used herein, specify the presence of stated features, integers, steps, operations, elements, components, or combinations thereof, but do not preclude the presence or addition of one or more other features, integers, steps, operations, elements, components, or combinations thereof.

[0029] Please refer to Figs. 1-3 As shown in the drawings, a method for effectively analyzing and transforming false alarms of a fire alarm system in a substation includes:

[0030] S1: Collecting multi-channel sensor data in the substation, the sensor data including smoke concentration, temperature change, humidity, electromagnetic field intensity, line grounding state signal, and forming a scenario-based environment data set in combination with a time stamp, weather information, load type, and operation and maintenance activity log;

[0031] S2: Normalizing and denoising the collected sensor data to eliminate the influence of different dimensions and random noise on subsequent analysis;

[0032] S3: Clustering analysis of the normalized multi-dimensional sensor data based on a spatio-temporal graph neural network, dividing multiple time period environment sub-regions with similar scene semantics, and learning typical feature patterns of fire signals and interference noise in each sub-region;

[0033] S4: Constructing a fuzzy confidence interval model, expressing the alarm signal confidence of each feature detection channel in a fuzzy set, and establishing an abnormal signal fuzzy confidence weighted table and a typical noise set in each time period in combination with historical scene distribution;

[0034] S5: When the alarm signal is triggered, the system calls historical data of the same scene according to the current scene label, makes a weighted decision on the alarm signal under the multi-time period fuzzy confidence model, and performs Bayesian inference on the abnormal change trend by fusing the detection data of the previous and subsequent time periods;

[0035] S6: Based on the inference result, a false alarm discrimination decision is made. If the signal behaves as a high-confidence anomaly in multiple time periods and scenarios and deviates from the noise fuzzy set, it is marked as a high-confidence real alarm. If there is a significant intersection between the confidence interval and the noise set, it is marked as a high-probability false alarm and pushed to the manual review queue;

[0036] S7: Periodically update the scene annotation and fuzzy set of false alarms and real alarms, and optimize the fuzzy confidence model parameters using incremental learning mechanism to realize long-term adaptive optimization of the complex environment changes in the substation;

[0037] S8: Based on the historical false alarm events and environmental change period, dynamically adjust the threshold parameters of the discrimination model to improve the adaptability of the model to feature drift caused by environmental drift and sensor aging.

[0038] The step S1: Collecting multi-channel sensor data in the substation, the sensor data includes smoke concentration, temperature change, humidity, electromagnetic field intensity, line grounding state signal, and combining time stamp, weather information, load type, operation and maintenance activity log to form a scene-based environment data set. Specifically, it includes:

[0039] S1.1: Based on the multi-modal sensor network deployed on site in the substation, the analog voltage signal output by the smoke concentration sensor is obtained, which is converted into a digital signal sequence through an ADC analog-to-digital conversion module to obtain standardized smoke concentration feature data;

[0040] Based on the multi-modal sensor network deployed on site in the substation, the analog voltage signal output by the smoke concentration sensor is obtained, and the signal integrity and accuracy during the collection process are ensured;

[0041] A high-precision ADC analog-to-digital conversion module (sampling accuracy ≥ 16 bits, sampling frequency ≥ 1 kHz) is used, and the voltage range V min to V max is converted to a corresponding digital signal sequence, realizing high-resolution data capture;

[0042] Further, through the sensor calibration coefficient K cal and the zero point offset V offset , the digital signal is linearly converted once, and the calculation formula is:

[0043] C smoke =K cal ×(V digit -V offset )

[0044] Where C smoke is the standardized smoke concentration value, V digitis the digital signal value converted by ADC, unit: bit;

[0045] Further, the converted concentration sequence is smoothed by using a sliding average filtering algorithm (window length N=5) to eliminate short-period fluctuations caused by electromagnetic noise and sampling jitter, and a stable smoke concentration change curve is obtained;

[0046] Further, the concentration threshold C thr and the alarm sensitivity coefficient S sens are set according to the sensor technical parameters;

[0047] Through ADC analog-to-digital conversion, signal calibration, filtering and normalization processing, the collected analog voltage signal is converted into digital, standardized and stable smoke concentration feature data, realizing seamless fusion with subsequent multi-channel sensor data sets;

[0048] For example, an optical smoke sensor with an output range of 0-5V is selected in the deployment scene of a certain substation, and an ADS1115 chip is selected for the ADC module, with a resolution of 16 bits and a sampling frequency of 2kHz. Through calibration experiments, K cal =0.85% / mV, V offset =0.12V, C min =0% / m, C max =5% / m. The sensor detects a voltage output of 2.52V at a certain time, and the digital value V digit =32700bit is obtained through ADC conversion. The concentration C smoke =0.85×(2.52-0.12) =2.04% / m is calculated by substituting the formula. The short-period jitter amplitude is reduced from 0.15% / m to 0.03% / m by performing 5-point sliding mean filtering on the continuous sampling data. The standardized concentration feature input C norm =(2.04-0) / (5-0)=0.408 is obtained, which is used as the smoke concentration channel value of the time period environmental data set, providing accurate input for multi-dimensional scene semantic modeling. Compared with the traditional unfiltered method in the experimental environment, the signal-to-noise ratio is improved by about 12dB, significantly improving the discrimination accuracy of subsequent false alarm discrimination;

[0049] S1.2: The real-time temperature data collected by the temperature sensor is processed by sliding window mean filtering to eliminate abnormal peaks caused by transient electromagnetic interference and obtain a stable temperature change feature sequence;

[0050] The real-time temperature signal collected by the temperature sensor deployed on site in the substation is processed by a sliding window mean filtering method (window length N winBased on the product of the sampling frequency F s and the electromagnetic interference duration T imp , the peak signal caused by the transient electromagnetic pulse interference is smoothed and suppressed;

[0051] Further, by calculating the weighted mean T avg (k) in the sliding window, the temperature value at the center time of the window is processed to minimize deviation, and a more stable change trend sequence relative to the background temperature is obtained;

[0052] Further, an outlier rejection algorithm is used to eliminate isolated points in the smoothed sequence that exceed three times the standard deviation threshold, and linear interpolation is used to complete the data of the rejected points to achieve data continuity and consistency;

[0053] Further, the corrected temperature sequence is input into a first-order difference module to extract the temperature change rate feature, and significant change events are screened out through threshold comparison, realizing sensitive capture of transient temperature rise;

[0054] Through the combination of sliding mean filtering, outlier rejection and change rate extraction processing, the original temperature collection signal is converted into a temperature feature sequence that can reflect the real temperature change trend after noise suppression, realizing the expected technical effect of providing reliable input features for subsequent multi-channel data fusion;

[0055] For example, in a certain substation application scenario, a digital temperature sensor is selected, with an output accuracy of ±0.1℃, a sampling frequency F s set to 10Hz, and an electromagnetic interference duration T imp of about 0.5s, then the window length N win = F s × T imp = 5. The center weight coefficients w i are [1, 2, 3, 2, 1] in turn, and the original sampling data at a certain time is [25.1, 25.2, 27.8, 25.3, 25.2]℃, after weighted mean calculation T avg= (1 x 25.1 + 2 x 25.2 + 3 x 27.8 + 2 x 25.3 + 1 x 25.2) / (1 + 2 + 3 + 2 + 1) = 26.271℃, the influence of the peak 27.8℃ is significantly weakened. After removing the outlier point 29.5℃ and linear interpolation with the values [26.1, 26.3] on both sides, the temperature is 26.2℃. After first-order difference extraction, a rapid temperature rise event of ΔT(k) = 0.35℃ is detected at time k, which is higher than the threshold of 0.3℃, and is marked as a significant change. Compared with the original signal, the processed temperature change curve has a peak interference amplitude reduction of about 80%, and the detection sensitivity of the temperature rise event is improved by about 15%, effectively supporting the input accuracy of the false alarm discrimination algorithm;

[0056] S1.3: Obtain the environmental humidity information using the capacitive humidity sensor, and complete the missing or abnormal data points by linear interpolation algorithm to generate a continuous and complete humidity feature time series;

[0057] S1.4: Collect the electromagnetic field strength data inside the substation using a high-frequency electromagnetic field probe, and extract the electromagnetic spectrum features by Fast Fourier Transform (FFT) algorithm to identify strong electromagnetic interference sources that may affect the fire alarm system;

[0058] S1.5: Perform threshold judgment on the ground resistance value output by the ground line state monitoring module, and record the event and generate line ground state marker data when line ground anomaly is detected, which is used for context association in subsequent false alarm analysis;

[0059] S1.6: Integrate the above multi-channel sensor data, and attach time stamp, meteorological information obtained through weather data interface, load type identification provided by SCADA system, and operation and maintenance activity log recorded by operation and maintenance management system to form a structured and scenario-based environmental data set, which provides input for multi-period semantic reconstruction.

[0060] The step S2: normalize and denoise the collected sensor data to eliminate the influence of different dimensions and random noise on subsequent analysis. Specifically, it includes:

[0061] S2.1: Normalize the collected multi-channel sensor data such as smoke concentration, temperature change, humidity, electromagnetic field strength, and line ground state signal. Use the maximum and minimum normalization method to map each sensor data to the [0, 1] interval to eliminate the feature weight imbalance problem caused by dimension difference and obtain standardized sensor data with uniform dimension;

[0062] The maximum and minimum normalization method (parameters: minimum value Xmin and maximum value Xmax of each channel historical statistics) is used to process the collected multi-channel sensor raw data such as smoke concentration, temperature change, humidity, electromagnetic field strength, and line ground state signal. The normalized data Xnorm is calculated as follows:min Xmax max ), to achieve the unified mapping function across the types of sensor dimensions;

[0063] Further, by inputting the data X i (t) of each channel, the normalization conversion is performed, and the calculation formula is:

[0064]

[0065] wherein X norm (t) is the normalized dimensionless characteristic value, X min,i , X max,i are the minimum and maximum values of the channel i obtained by calibration or historical statistics, respectively, to ensure that the data is mapped to the [0, 1] interval;

[0066] Further, for the possible extreme outliers, the quantile clipping algorithm (parameters: lower quantile Q low and upper quantile Q high , such as 2% and 98% quantile) is used to perform boundary truncation on the original data exceeding the range before normalization, to reduce the distortion influence of abnormal measurement points on the normalization ratio calculation;

[0067] Further, after the clipped original data is input into the normalization formula, the proportional correction is performed by using the channel-level precision factor a i to achieve the comparability enhancement under different sensor precision levels, and the correction formula is:

[0068] X′ norm (t) = a i · X norm (t)

[0069] wherein a i can be dynamically set according to the factory precision level or historical error distribution of the sensor;

[0070] Further, by using the batch processing mechanism, the vectorized normalization operation is performed on the multi-channel data matrix collected in the same batch, and the normalization parameter log file (records X min , X max , Q low , Q high , a i ) is generated, to be called when the normalization condition needs to be traced back in the subsequent steps such as wavelet denoising and feature dimension reduction;

[0071] By combining the maximum-minimum normalization with the quantile clipping and precision factor correction processing methods, the original multi-source sensor data is uniformly converted into a standardized feature matrix with consistent numerical range and dimension-unified, realizing the expected technical effect of cross-channel feature weight balance and subsequent time-frequency domain noise reduction, and improving the comparability of clustering modeling;

[0072] For example, in a certain substation environment, the original data range of smoke concentration is obtained by historical statistics X min = 0.02% / m, X max = 4.95% / m, the temperature change data range is 10.5℃ to 48.3℃, the humidity data range is 35% RH to 92% RH, the electromagnetic field intensity range is 0.05V / m to 4.5V / m, and the line grounding resistance value range is 0.1Ω to 5.3Ω. After 2% and 98% quantile point truncation processing, the abnormal peak of smoke concentration is clipped from 5.8% / m to 4.95% / m. In the normalization calculation, the smoke concentration X = 2.15% / m at a certain sampling time, the corresponding normalized value is X norm = (2.15-0.02) / (4.95-0.02) = 0.434. Combined with the precision factor α = 1.02, the final normalized result is X' norm = 0.443. The numerical values of the five-channel feature matrix generated by batch processing are limited to [0, 1], and the standard deviation difference ratio of different channels before normalization is 2.87, which is reduced to 0.05 after normalization, proving that the dimension consistency of multi-channel features has been significantly improved, meeting the input requirements of the subsequent S2.2 wavelet denoising algorithm. In this scene, the feature input after this normalization preprocessing improves the noise suppression performance in the time-frequency domain, and the overall recognition accuracy is improved by about 6.5% in the subsequent false alarm discrimination experiment;

[0073] S2.2: Based on the wavelet transform algorithm, the normalized sensor data is processed in the time-frequency domain to suppress the high-frequency random disturbance caused by electromagnetic interference and line noise, and the wavelet denoised time domain signal sequence is obtained;

[0074] S2.3: The wavelet denoised signal is subjected to sliding window mean filtering processing, and the signal is subjected to local smoothing operation based on the set time window length, so as to further eliminate residual noise and retain signal trend characteristics, and the smoothed filtered sensor signal output is obtained;

[0075] S2.4: The principal component analysis (PCA) method is used to process the feature dimension reduction of multi-channel sensor data, the covariance matrix between features is calculated, and the principal component vector is extracted, so as to remove redundant information and retain low-dimensional feature representation with the best discriminant ability, and the dimension-reduced sensor feature matrix is obtained;

[0076] S2.5: Variance normalization is performed on the dimension-reduced sensor feature matrix, and the variance values of each principal component feature are used for weighted adjustment to improve the separability of the features in subsequent clustering and fuzzy reasoning, so as to obtain a final input feature vector set used for scene semantic modeling.

[0077] The step S3: based on the spatio-temporal graph neural network, the normalized multi-dimensional sensor data are analyzed for clustering, and multiple time period environment sub-regions with similar scene semantics are divided, and the typical feature patterns of fire signals and interference noises in each sub-region are learned. As shown in Fig. 2 , specifically comprising:

[0078] S3.1: Based on the normalized multi-dimensional sensor data, the sensor data including smoke concentration, temperature change, humidity, electromagnetic field intensity, line grounding state signal, a high-dimensional feature space of the substation operating environment is constructed to form a node feature vector for graph structure modeling;

[0079] S3.2: The graph adjacency relationship modeling is performed on the node feature vector in the high-dimensional feature space, and a dynamic spatio-temporal adjacency matrix is constructed based on the time stamp and spatial correlation to represent the transfer relationship of the environment state between different time points;

[0080] Based on the node feature vector in the high-dimensional feature space, a time correlation calculation method (parameters: time interval Δt and maximum correlation time window T max ) is used to realize the quantification of the time adjacency relationship between nodes at different sampling times, and a time correlation weight matrix W t is obtained;

[0081] Further, by a spatial correlation calculation method (parameters: physical position encoding L i , sensor topology matrix S base ), the calculation of the spatial correlation weight between the sensor nodes located in adjacent physical positions is realized, and a spatial correlation weight matrix W s is generated;

[0082] Further, a weighted fusion strategy (parameters: time-space weight coefficients α, β, satisfying α+β=1) is used to perform linear weighted fusion on the time correlation weight matrix W t and the spatial correlation weight matrix W s , and calculate a dynamic spatio-temporal weight matrix W ts , the formula is:

[0083] W ts (ij)=α·W t (ij)+β·W s (ij)

[0084] Wherein, W ts(ij) represents the integrated correlation weight of node i and node j under dynamic spatio-temporal relationship.

[0085] Further, by threshold pruning method (parameters: adjacency threshold θ adj ), the weight elements of W ts below the threshold are assigned to zero, and the high correlation node pairs are retained, generating a sparse dynamic spatio-temporal adjacency matrix A ts to reduce the interference of irrelevant edge pairs on subsequent graph neural network calculation;

[0086] Further, by using the adjacency matrix normalization method (parameters: symmetric normalization method), the operation is performed on A ts to ensure the normalization of node degree to alleviate the influence of uneven degree distribution, and obtain the dynamic spatio-temporal adjacency matrix

[0087] Through the construction of the dynamic spatio-temporal adjacency matrix, the environmental state transition relationship between different time points can be accurately represented, and redundant connections can be eliminated while retaining key spatio-temporal dependencies, achieving high effectiveness and low redundancy of graph neural network input data;

[0088] For example, in the actual application of a substation, the collected multi-channel sensor node data includes: smoke concentration node N1, temperature node N2, humidity node N3, electromagnetic field intensity node N4, and ground resistance node N5, with a sampling frequency of 1 Hz and a maximum time correlation window T max set to 300 seconds. The Pearson correlation coefficient formula is used to calculate the time correlation weight, and when calculating the time correlation weight, the node pairs with Pearson coefficient exceeding 0.85 within Δt≤60s are assigned with the corresponding weight, and the rest is set to zero. In the calculation of spatial weight, the physically adjacent nodes are assigned with 1.0, the nodes in the same region but not adjacent are assigned with 0.5, and the nodes in different regions are assigned with 0.1. The fusion weight coefficient is set to α=0.6 and β=0.4, and the matrix W ts is obtained after weighted calculation. The threshold θ adj =0.4 is used for pruning, and the sparse adjacency matrix is obtained, with the connection rate reduced from 100% to 42%, effectively eliminating weakly correlated edges. Finally, after symmetric normalization processing, the In the subsequent graph convolution calculation of S3.3, the node feature extraction accuracy is improved by about 5.8% compared with the non-sparse processing, and the calculation complexity is reduced by about 27%, verifying the effectiveness of the method;

[0089] S3.3: Input the dynamic spatio-temporal adjacency matrix and node feature vector into the spatio-temporal graph neural network model, extract the spatial correlation between nodes through graph convolution operation, and capture the time evolution trend of environmental state through time gating mechanism, to generate low-dimensional embedding features with spatio-temporal semantics;

[0090] Dynamic spatio-temporal adjacency matrix output based on the previous step Corresponding node feature vector matrix Using a graph convolutional neural network (GCN) convolution kernel calculation method (parameters: graph convolution order K, convolution weight matrix W g ), the spatial correlation between nodes is realized. The feature aggregation process is based on the transferability of the normalized adjacency matrix, and the single-layer graph convolution operation is realized by the following formula:

[0091]

[0092] Where H (l) is the node feature matrix of the lth layer, and σ is a nonlinear activation function (such as ReLU);

[0093] Further, by stacking multiple layers of graph convolution and controlling the convolution order K, the spatial relationship feature fusion across the local adjacency range is realized, and the batch normalization operation is applied after each convolution to alleviate the gradient disappearance and improve the stability of the feature distribution, and the intermediate node representation matrix H s with global spatial correlation information is obtained.

[0094] Further, based on the time-gated recurrent unit (GRU) method (parameters: hidden state dimension d h , time step T w ), the feature sequence arranged in time order in H s is input to the GRU unit, and the fusion ratio of the historical state and the current input is controlled by the update gate z t and the reset gate r t to capture the time evolution pattern of the environment state. The update process of the GRU unit is defined by the following formula:

[0095] z t = σ(W z x t + U z h t-1 )

[0096] r t = σ(W r x t + U r h t-1 )

[0097]

[0098] Where x t is the input feature vector at time t, and h tis the hidden state at time t, and is the Hadamard product;

[0099] Further, by applying an attention mechanism (parameters: attention weight matrix W a , normalization method softmax) on the GRU output sequence, the attention weight of each time step feature to the overall spatio-temporal embedding is calculated, the importance of the key time segment is enhanced, and the time-weighted spatio-temporal embedding feature matrix H st is obtained.

[0100] Further, H st is input into a fully connected feature compression layer (parameters: target dimension d embed , activation function tanh), and the high-dimensional spatio-temporal feature is mapped to a low-dimensional embedding representation to reduce data redundancy and retain discriminative spatio-temporal semantics.

[0101] By combining the dynamic spatio-temporal adjacency matrix driven graph convolution with the time gating mechanism, the node feature vector of the previous step is converted into a low-dimensional spatio-temporal embedding feature containing spatial topology dependence and time evolution information, realizing the unified representation of the spatio-temporal pattern of the substation multi-channel environment data.

[0102] For example, in a certain 500kV substation, the graph convolution order K = 2 is set, the convolution weight matrix initial value is initialized using the Xavier uniform distribution, the batch size is 64 time node sequences. The GRU hidden state dimension d h = 64, the time step T w = 300s, the attention mechanism uses softmax normalization and sets the weight matrix dimension to After time weighting, it is mapped to a d embed = 16-dimensional embedding space. After convolution and time gating processing, the Fisher discriminant rate of the spatio-temporal embedding vector between classes is improved from 0.72 to 0.91, and the silhouette coefficient in subsequent clustering analysis is improved by 12.4% compared with that without spatio-temporal feature fusion, effectively verifying the accuracy and robustness of the method in extracting complex environmental change patterns.

[0103] S3.4: Perform clustering analysis on the low-dimensional embedding features output by the spatio-temporal graph neural network, and use the improved K-means++ algorithm combined with the silhouette coefficient to optimize the number of clusters to divide the multi-period environment sub-area with similar scene semantics.

[0104] Based on the low-dimensional spatio-temporal embedding feature matrix H output by S3.3, the improved K-means++ clustering initialization method (parameters: maximum number of iterations Iter max , initial center number Kinit , random seed Seed km ), to achieve fast and balanced selection of high-density cluster centers, and reduce the sensitivity of initial centers to clustering convergence accuracy;

[0105] Further, when performing K-means clustering, the Silhouette Coefficient evaluation mechanism is introduced to evaluate the quality of the clustering results for different K values. The Silhouette Coefficient calculation formula is:

[0106]

[0107] Where a(i) is the average distance between sample i and the remaining samples in the cluster, b(i) is the average distance between the sample and all samples in the nearest neighbor cluster, S(i) takes a value in the range [-1, 1], and the closer to 1 indicates that the cluster division is more reasonable.

[0108] Further, by traversing K∈[K min , K max ], the average Silhouette Coefficient is calculated for each K as the optimal clustering number, ensuring that the divided time period environment sub-area achieves the optimal balance in tightness and separation;

[0109] Further, the clustering results under K opt are taken as the output of multiple time period environment sub-area division, generating a scene semantic label vector for each sample, providing accurate time period-scene partition basis for subsequent S3.5 feature pattern learning;

[0110] By combining the improved K-means++ initialization and the Silhouette Coefficient adaptive optimization, the low-dimensional spatiotemporal embedding feature vector is effectively mapped to a multiple time period environment sub-area with high semantic consistency, achieving high-robust scene clustering effect of substation multi-dimensional environmental data;

[0111] For example, in a certain 500kV substation, the low-dimensional embedding feature dimension is set to d embed = 16, the sample size is N = 86400 (corresponding to 24 hours of 1Hz sampling frequency). Set K min = 2, K max = 12, the maximum number of iterations Iter max = 300, the random seed Seed km = 42. In the improved K-means++ initialization, the first center is randomly selected, and the remaining centers are selected according to D min (e i ) 2The probability sampling is performed on the distribution to generate an initial center set of K=8. The average silhouette coefficient under different K is calculated during the clustering process, and the result shows that K=5 is the optimal value , which is better than 0.796 of K=4 and 0.823 of K=6. Therefore, K opt =5. Under the optimal cluster number, the scene semantic labels of the sub-intervals accurately distinguish five types of scenes, i.e., stable operation state of sunny high temperature and low humidity, heavy rain high humidity and high electromagnetic interference state, night low temperature high load state, maintenance high electromagnetic pulse state, and normal operation background state. When used for subsequent fire signal pattern learning, the similarity of the signals in the same class is improved by 14.6% and the difference between classes is improved by 18.3% compared with random clustering, effectively improving the accuracy of the scene matching link in false alarm discrimination.

[0112] S3.5: Based on the divided multiple time period environment sub-intervals, the fire signals and interference noises in each sub-interval are learned for feature patterns. Principal component analysis and discrete Fourier transform are used to extract typical frequency domain and time domain features to construct abnormal signal feature templates and noise feature templates under each sub-interval.

[0113] The step S4: a fuzzy confidence interval model is constructed to express the alarm signal confidence of each feature detection channel in a fuzzy set, and a fuzzy confidence weighted table of abnormal signals and a typical noise set under each time period are established in combination with the historical scene distribution. As shown in Fig. 3 , specifically comprising:

[0114] S4.1: The fire alarm feature signals collected by each sensor channel are normalized to eliminate the dimension inconsistency caused by the difference between sensor types, and a standardized feature vector set is obtained as the basic input data for fuzzy confidence modeling;

[0115] The fire alarm feature signals collected by the multi-channel sensor are normalized by the Z-score standardization method (parameters: mean μ i , standard deviation σ i ), to realize zero-mean unit-variance transformation of each channel signal, so as to eliminate the dimension difference and balance the contribution weight of the feature to the model;

[0116] Further, the normalization calculation is performed on the original feature value x i (k) of the sensor by the following formula:

[0117]

[0118] where x′ i (k) is the normalized feature value of the i-th channel and the k-th sampling point, μ i is the mean value in the historical window of the channel, and σ i is the corresponding standard deviation value, which ensures that the feature data distribution is balanced.

[0119] Further, for the feature channels that present long-tail distribution or non-Gaussian distribution, Box-Cox power transformation (parameter: λ value is selected according to maximum likelihood estimation) is adopted to realize data approximate normal processing and generate a feature vector with optimized distribution form, thereby improving the stability of subsequent fuzzy set modeling;

[0120] Further, linear interval scaling method (parameter: interval lower limit a = 0, upper limit b = 1) is applied to map the normalized feature data to a unified closed interval [0, 1], thereby realizing interval normalization of feature values;

[0121] Further, combined with sliding window statistical smoothing processing (parameter: window length W f , step size Δ w ), short-term mean adjustment is performed on the mapped feature vector to suppress transient deviation caused by random jitter, thereby generating a smooth normalized feature matrix;

[0122] Through the multi-stage normalization and standardization chain processing mode, the alarm feature signals output by different types of sensors are converted into a standardized feature vector set with unified dimension, stable distribution and direct applicability to fuzzy confidence model construction, thereby realizing comparability and synergy of data in each channel in fuzzy set calculation;

[0123] For example, in a certain 500 kV substation, five types of sensor features, i.e. smoke concentration (mg / m 3 ), temperature (℃), humidity (%RH), electromagnetic field intensity (V / m) and grounding resistance (Ω) are collected. For the temperature channel raw data, the window mean value is μ T = 35.2℃, the standard deviation is σ T = 4.6℃, and a temperature sample T = 42.5℃ is converted into T' = (42.5-35.2) / 4.6 ≈ 1.587 using Z-score normalization. For the electromagnetic field intensity channel, non-Gaussian distribution is detected, and the λ value is determined to be 0.21 through maximum likelihood estimation. After Box-Cox transformation, the data skewness is reduced from 1.74 to 0.05. Then, interval scaling method (a = 0, b = 1) is applied. Assuming that the minimum value after transformation is -2.3 and the maximum value is 3.7, then the value of a certain point of the channel -0.5 is mapped into x" = (-0.5+2.3) / (3.7+2.3) ≈ 0.3. Finally, the smoothing window length W f = 15 seconds and the step size Δ w = 5 seconds, and the smoothed normalized feature matrix is output for S4.2 fuzzy C-means clustering. In the validation set, the dimension difference elimination rate of the normalized channels is more than 99%, and the convergence speed of the membership calculation of the fuzzy set modeling is improved by about 17%;

[0124] S4.2: Based on the historical false alarm and real alarm event data, the fuzzy C-means clustering algorithm is used to fuzzy partition the standardized feature vector set, identify the fuzzy noise patterns in each feature channel that are highly related to false alarms, and construct a preliminary fuzzy noise set;

[0125] Standardized feature vector set output based on S4.1 The fuzzy C-means (FCM) clustering algorithm (parameters: number of clusters C num , fuzzy weighting index m, convergence threshold ∈, maximum number of iterations Iter max ) is used to partition the samples in each feature channel in the fuzzy membership space to identify fuzzy noise patterns highly related to false alarms.

[0126] The initial random distribution generation method (parameters: random seed Seed fcm ) is used to initialize the membership matrix Ensure that the membership sum of each sample to each cluster is 1.

[0127] S4.3: Statistical analysis is performed on the historical alarm data under each period scene label, a confidence membership function for each feature channel is established based on fuzzy set theory, and a fuzzy confidence interval model related to the period is generated to express the uncertainty of the alarm signal in different scenes.

[0128] Preliminary fuzzy noise set F noise output based on S4.2 and the corresponding standardized feature vector set X std , the period grouping statistical analysis method (parameters: scene label set L scene , time resolution Δt) is used to cluster the historical alarm data by period and scene to obtain the feature channel sample subset under each period scene label;

[0129] Further, for the sample subset of each feature detection channel i, the interval frequency statistical method (parameters: interval division number B bin , kernel density smoothing bandwidth h kde ) is used to calculate the confidence probability density distribution p i (c) of the channel in the current period, where c represents the alarm confidence value of the channel.

[0130] Further, based on fuzzy set theory, the confidence membership function ildeu i (c) of each channel is constructed, and the triangular, trapezoidal or Gaussian membership function model is preferred, and the function parameters are optimized by the least squares fitting method to make the shape of ildeu i (c) optimally approximate the historical distribution p i (c).

[0131] Further, the Gaussian membership function is described in the following formula:

[0132]

[0133] Where μ i is the mean value of confidence of channel i in this period, σ i is the standard deviation, both of which are calculated from the historical sample subset;

[0134] Further, for the samples identified as noise patterns by F noise , the corresponding c value is given a penalty weight w pen <1> in the fitting process to reduce the impact of noise patterns on the shape of membership function and increase the weight ratio of real alarm patterns in fuzzy expression;

[0135] Further, the fitted confidence membership function of each channel is stored as a period-related fuzzy confidence interval model The model is continuously defined in the interval [c min , c max ] and can output the membership degree corresponding to any confidence value;

[0136] Through fuzzy set modeling and noise penalty fitting, the alarm confidence of each feature channel under different time periods and scenes is converted into a calculable membership function, realizing the structured expression of alarm signal uncertainty;

[0137] For example, in a 500kV substation, set the time resolution Δt = 1 hour, and divide the historical alarm data of the whole year into 5 scenes according to the time period and scene label, a total of 120 time period-scene units. For the smoke concentration channel, set B bin = 50, calculate the confidence probability density in the current period (scene C2), the mean value μ sm = 0.72, and the standard deviation σ sm = 0.15. Select the Gaussian membership function, and get the function through historical sample fitting:

[0138]

[0139] Combined with the high-frequency noise cluster samples identified by F noise , the c value is given a penalty weight w pen = 0.5, and the fitting is performed again. The maximum value of the membership function remains at 0.99 and shows obvious attenuation in the c<0.4 region, and the peak sensitivity of the noise pattern is reduced by 26%. The output The fuzzy confidence weighted table is calculated based on the fuzzy confidence interval model output and the historical frequency distribution of false alarms and real alarms in each period, so that the average membership degree of real alarm samples is improved by about 18.7% relative to the unpenalized process, effectively improving the real alarm discrimination ability of the model;

[0140] S4.4: Based on the membership degree value output by the fuzzy confidence interval model, the fuzzy confidence weighted table of the abnormal signal in each period is calculated and constructed by combining the historical frequency distribution of false alarms and real alarms in each period, which is used to quantify the probability weight of false alarm occurrence in different scenes;

[0141] S4.5: The fuzzy confidence weighted table is fused with the fuzzy noise set to model the period-feature joint fuzzy criterion space as the input basis for the false alarm discrimination decision in the subsequent multi-period fusion reasoning module.

[0142] The step S5: When the alarm signal is triggered, the system calls the historical data of the same scene according to the current scene label, makes a weighted judgment on the alarm signal under the multi-period fuzzy confidence model, and fuses the detection data before and after the period to make Bayesian inference on the abnormal change trend. Specifically, it includes:

[0143] S5.1: Based on the scene label of the current alarm signal, the historical fuzzy confidence weighted table and the typical noise set in the corresponding period are called as the reference input for the abnormal judgment in the current period to construct a scene-adaptive discrimination basis;

[0144] S5.2: The confidence of the alarm signal in the current period is mapped to a fuzzy set in each feature detection channel to generate a fuzzy confidence interval, and compared with the historical noise set to identify whether the current signal falls into the fuzzy region of the known interference mode;

[0145] The confidence vector C cur of each feature detection channel of the alarm signal in the current period is F , and the fuzzy set mapping method is used (parameters: fuzzy confidence membership function of the corresponding period Each channel confidence scalar is mapped to a membership value distribution on a continuous interval to characterize the uncertainty range of the alarm feature;

[0146] Further, the lower limit L i and the upper limit U i of the confidence of each channel i are calculated, and the solution of the membership degree corresponding to the confidence value reaching the preset threshold α is taken as the interval boundary to obtain the fuzzy confidence interval CI i = [L i , U i ], which realizes the quantitative expression of the fuzzy uncertainty of each channel alarm signal;

[0147] Further, the obtained fuzzy confidence interval set CI cur = [CI1,..., CI F ] is compared with each noise cluster center interval NI noise in the historical noise set F j The interval intersection and union comparison is performed using the interval overlap rate formula:

[0148]

[0149] where |·| represents the interval length, NI i,j is the confidence interval of the jth noise pattern in the noise set in channel i, realizing the quantitative evaluation of the similarity between the current confidence interval and the historical noise pattern;

[0150] Further, based on the overlap rate vector IOR i = [IOR i,1 ,..., IOR i,p ] of all channels under each noise pattern, the average overlap rate is calculated to reflect the degree to which the overall characteristics of the alarm signal are covered by the known interference pattern;

[0151] Through the interval intersection and union comparison algorithm, the fuzzy confidence interval is compared with the noise pattern for multi-channel and multi-noise cluster overlap analysis, and the overlap degree index vector IOR cur is output, providing quantitative input for subsequent weighted scoring and trend Bayesian fusion, realizing the early quantitative evaluation of the risk of the current alarm signal falling into the known interference pattern;

[0152] For example, in a certain 500kV substation, the confidence values of the current period alarm signal in five channels are [0.73, 0.65, 0.41, 0.58, 0.47] respectively, and the Gaussian membership function of the corresponding period C2 is used and the cutoff threshold α = 0.6 is set, the temperature channel interval [0.69, 0.77] and the electromagnetic field channel interval [0.55, 0.61] are obtained by solving the equation The historical noise set F noise contains three noise clusters, whose intervals in the temperature channel are [0.65, 0.72], [0.50, 0.60], and [0.80, 0.88] respectively. The temperature channel current interval overlap rate IOR T,1 = (0.72-0.69) / (0.77-0.65)≈0.25 with the first noise interval, and the overlap rates with the second and third clusters are 0 and 0 respectively. After calculating the overlap rates of the five channels with all noise clusters, the average overlap rate This indicates that the overall alarm signal is unlikely to fall into a noise pattern. During the verification phase, the processing link effectively distinguished over 92% of real alarm samples with an overlap rate of less than 0.3 in noise intervals, reducing the probability of false alarms caused by high-frequency interference.

[0153] S5.3: Based on the overlap between the fuzzy confidence interval and the historical noise set, a weighted score calculation is performed on the current alarm signal. The fuzzy membership function is used to nonlinearly weight the confidence of each feature channel to generate a comprehensive anomaly score within the time period.

[0154] The confidence fuzzy interval CI of each characteristic channel of the current alarm signal output by S5.2. cur = [CI1, CI2, ..., CI] F [and historical noise set F] noise The overlap index IOR cur A weighted scoring method is used (parameter: weight vector W = [w1, w2, ..., w...). F Fuzzy membership function The system uses a type and a nonlinear weighted index γ to comprehensively quantify the abnormal significance of the current alarm signal in each channel.

[0155] The effective confidence score S of channel i is calculated using the following formula. i :

[0156]

[0157] in, Let [L] be the average overlap rate between channel i and all historical noisy and blurred intervals. i U i [ ] represents the boundary of the fuzzy confidence interval for this channel. γ is the membership function of the current time period scene. The integral is used to accumulate the membership quality within the interval, and γ is used to control the weighted reinforcement degree of the high membership interval.

[0158] Furthermore, the scores of all channels are weighted and combined according to a preset weight vector W to calculate the comprehensive anomaly score S for the current time period. total :

[0159]

[0160] Wherein, weight w i The setting is based on the relative distinguishability of the feature channels in historical real-time alarm identification, and is preferably obtained through mutual information or feature importance evaluation;

[0161] Furthermore, the calculated S total Perform normalization mapping (parameter: minimum and maximum values ​​S) min ,Smax Derived from historical score distribution, a normalized comprehensive anomaly score S is formed. norm :

[0162]

[0163] Normalization ensures the comparability of scores across different time periods and avoids the influence of score ranges within specific time periods;

[0164] Furthermore, regarding S norm Introducing an abnormal enhancement factor E adj =1+(σ c / μ c ), where σ c With μ c These are the standard deviation and mean of the current confidence interval width across all channels, respectively. This ratio reflects the consistency between channels, thereby adjusting the weight of outlier scores: increasing scores in high-consistency scenarios and suppressing scores in low-consistency scenarios, resulting in the corrected comprehensive score S. final =S norm ×E adj ;

[0165] S is generated by fusing fuzzy membership integral, overlap rate penalty, and channel weight. final The fuzzy confidence interval mapping results of stage S5.2 are transformed into quantifiable comprehensive anomaly scores within a time period, thereby enhancing the true anomaly features and suppressing interference features in false alarm sensitive scenarios.

[0166] For example, in a 500kV substation, the confidence intervals for the five alarm signal channels—smoke, temperature, humidity, electromagnetic field, and grounding—are [0.70, 0.78], [0.69, 0.77], [0.38, 0.44], [0.55, 0.61], and [0.46, 0.50], respectively, with corresponding average overlap rates of 0.12, 0.18, 0.65, 0.22, and 0.58. Taking the temperature channel as an example, a Gaussian membership function is used. Calculate interval integrals Taking γ = 1.2, we obtain S T ≈(1-0.18)×(0.118) 1.2 ≈0.079. S obtained through five-channel calculation. i The vector is [0.081, 0.079, 0.021, 0.044, 0.025], and the weight vector W is [0.28, 0.27, 0.15, 0.18, 0.12]. The weighted combination yields S. total ≈0.065. Historical rating distribution range [S] min S max = [0.015, 0.093], normalized to get S norm≈0.714. Calculate the mean channel width μ. c ≈0.06, standard deviation σ c ≈0.015, abnormal enhancement factor E adj If the score is approximately 1.25, then the overall score S is adjusted. final ≈0.892. This value is close to the preset upper limit for real alarm discrimination, indicating that the alarm signal has a high degree of anomaly significance in the current time period. When it is passed to the S5.4 stage for Bayesian fusion correction with the trend of adjacent time periods, it can improve the overall real alarm recognition accuracy by more than 15%.

[0167] S5.4: Retrieve detection data under the same or similar scene labels in adjacent time periods (previous and next time periods), extract the change trend of its anomaly score and fuzzy confidence, and perform trend fusion correction on the current anomaly score based on the Bayesian inference model to enhance the ability to distinguish the continuity of anomaly signals.

[0168] S5.5: Combine the weighted score of the current time period with the Bayesian inference results of the previous and next time periods to generate the final multi-time period fusion confidence value, and classify the alarm signal as false alarm or real alarm based on the set discrimination threshold, and output the discrimination result to the false alarm decision module.

[0169] Step S6: Based on the inference results, perform a false alarm discrimination decision. If the signal exhibits high-confidence anomalies across multiple time periods and scenarios and is outside the noise fuzzy set, it is marked as a high-confidence real alarm. If the confidence interval and the noise set have a significant intersection, it is marked as a high-probability false alarm and pushed to the manual review queue. Specifically, this includes:

[0170] S6.1: Analyze the fuzzy confidence intervals output by the multi-time period fusion inference stage, extract the upper and lower confidence limits of the current alarm signal in each time period, and form a structured confidence dataset as the input basis for false alarm judgment decision;

[0171] S6.2: Based on the fuzzy set operation rules, calculate the area of ​​intersection between the confidence interval of the current alarm signal and the typical noise set to quantitatively evaluate the similarity between the alarm signal and the historical false alarm pattern and obtain the noise matching index.

[0172] The structured confidence dataset CI based on the output of the multi-time fusion inference stage. struct = [CI1, CI2, ..., CI F ] and typical noise set F noise The noise mode range NI i,j The fuzzy interval intersection operation method is adopted (parameter: interval boundary L). i U i Noise interval boundary This allows for the assessment of the coverage of current alarm signals and historical noise patterns.

[0173] Further, the intersection interval length |CI i ∩NI i,j | of each feature channel i is calculated by interval operation formula

[0174]

[0175] where |CI i ∩NI i,j | is the intersection length of the current confidence interval of the i-th channel and the confidence interval of the j-th noise pattern, and the result is 0 if there is no overlap;

[0176] Further, the corresponding union interval length |CI i ∪NI i,j | is calculated, using the formula:

[0177]

[0178] Further, the intersection length and the union length are substituted into the intersection overlap rate calculation formula:

[0179]

[0180] The proportion value reflects the uncertainty proportion of the current alarm signal overlapping with the historical noise pattern j on the channel i;

[0181] Further, the intersection overlap rate vector IOR i = [IOR i,1 ,..., IOR i,p ] of each channel i under all noise patterns j = 1,..., p is subjected to statistical aggregation operation, and the arithmetic mean is used to obtain the noise matching degree NM i of the channel:

[0182]

[0183] Further, the matching degree vector NM vec = [NM1, NM2,..., NM F ] of all channels is formed, and the weighted average (parameter: channel weight vector W nm derived from historical recognition accuracy) is used to obtain the overall noise matching degree indicator NM total :

[0184]

[0185] By the above fuzzy set interval operation and intersection rate weighted aggregation processing, the multi-channel fuzzy confidence interval is associated with the historical noise pattern to be mapped into a quantifiable noise matching degree index, realizing the accurate evaluation of the similarity between the current alarm signal and the false alarm pattern;

[0186] For example, in the current alarm event of a certain 220 kV substation, the confidence intervals of the five channels (smoke, temperature, humidity, electromagnetic field, grounding) are [0.68, 0.76], [0.70, 0.78], [0.39, 0.45], [0.54, 0.60], [0.44, 0.50] respectively, and the historical noise set contains three types of interference patterns, whose intervals on the temperature channel are [0.69, 0.74], [0.50, 0.60], [0.75, 0.82] respectively. For the temperature channel and the first pattern: the intersection length |CI T ∩NI T,1 |=min(0.78, 0.74)-max(0.70, 0.69)=0.04, the union length 0.78-0.69=0.09, and the intersection rate ≈0.444. The intersection rates with the second and third patterns are 0 and ≈0.333 respectively, and the average value is NM T ≈0.259. After calculating the matching degree for all channels, it is assumed that NM vec =[0.21, 0.259, 0.61, 0.18, 0.52], under the weight W nm =[0.28, 0.27, 0.15, 0.18, 0.12], the overall noise matching degree index NM total ≈0.31. This value is below the system threshold T nm =0.45, and it is determined that the risk of falling into the noise pattern is low, providing a reliable quantitative basis for the false alarm tendency judgment of S6.3;

[0187] S6.3: Compare the noise matching degree index with the preset false alarm judgment threshold. If the intersection area exceeds the threshold, it is determined that the current signal has a high probability of false alarm characteristics, and a false alarm tendency marker is output. Otherwise, it is marked as a potential real alarm signal;

[0188] S6.4: Based on the Bayesian inference result, the confidence degrees of multi-period abnormal change trends are weighted and fused to calculate a comprehensive alarm confidence index, so as to enhance the recognition ability of the persistent fire characteristics, and output a fusion confidence vector;

[0189] S6.5: According to the joint judgment result of the fusion confidence vector and the false alarm tendency marker, the final false alarm discrimination decision is executed: if the signal shows high confidence anomaly in multiple periods and the noise intersection is small, it is marked as a high-confidence real alarm and the fire alarm response process is triggered; otherwise, it is marked as a high-probability false alarm signal and pushed to the artificial review queue.

[0190] The step S7: periodically annotating the false alarm and real alarm cases with scene and updating the fuzzy set, using the incremental learning mechanism to optimize the fuzzy confidence model parameters, to realize the long-term adaptive optimization to the complex environment changes of the substation. Specifically, it includes:

[0191] S7.1: Based on the historical false alarm and real alarm event records, the environmental label information of each alarm triggering period is structured and extracted, including timestamp, weather state, load type, operation identification, to construct a scene semantic label set strongly associated with the alarm event;

[0192] S7.2: The extracted scene semantic label set and corresponding alarm feature data are modeled by a spatio-temporal graph neural network to generate multi-period scene clusters with semantic consistency, to identify the typical scene category to which the current alarm event belongs;

[0193] S7.3: The false alarm and real alarm cases belonging to the same scene cluster are annotated with fuzzy confidence intervals, and the alarm confidence of each detection channel is expressed in intervals based on fuzzy set theory, to form a fuzzy confidence sample set under the scene cluster;

[0194] For false alarm and real alarm cases in the same scene cluster, the input data is the scene cluster label output by S7.2, the alarm feature vector of each detection channel, and its confidence record in the historical event library;

[0195] The fuzzy set annotation method (parameters: channel set F, historical event set E scene ) is used to realize the aggregation calculation of the alarm confidence samples of each detection channel, and the minimum value L f and the maximum value U f of the confidence of the channel under the scene cluster are obtained, to form the preliminary interval boundary;

[0196] Further, through the interval smoothing algorithm (parameter: smoothing coefficient α derived from the historical event stability index), the preliminary interval boundary is contracted or expanded, to suppress outliers and preserve the core distribution, to obtain the modified interval [L f ′, U f ′], to improve the robustness of the interval expression;

[0197] Further, based on the fuzzy set theory, the confidence membership function μ f (c) of each channel is defined, in the form of triangular or trapezoidal membership function, combined with the interval boundary [L f ′, U f ′] to set the core area and the transition area, to realize the mapping of the alarm confidence to the fuzzy domain:

[0198]

[0199] where c is the alarm confidence sample value, C mid is the core region center point, and μ f (c) represents the membership degree of the confidence belonging to the "effective fire" fuzzy set;

[0200] Further, the membership degree distribution of all historical event samples under the scene cluster and each channel is calculated to generate a fuzzy confidence sample matrix M scene , where the matrix element M scene (e, f) = μ f (c e,f ) represents the membership degree of event e on channel f;

[0201] Through the fuzzy confidence sample matrix and the interval boundary set, the result of the previous step is converted into a scene cluster level fuzzy confidence sample set, which realizes the basic data preparation for subsequent incremental learning and parameter optimization;

[0202] For example, in the scene cluster SC5 of a certain 220 kV substation, the historical false alarm and real alarm events of the smoke channel in the confidence degree range are [0.42, 0.88], which is corrected to [0.45, 0.85] after processing by the smoothing coefficient α = 0.1, and the core region center point C mid = 0.65. For a sample with a confidence degree c = 0.72, the membership degree μ smoke (0.72) = (0.85-0.72) / (0.85-0.65) = 0.65. Similarly, the temperature channel [0.50, 0.80] is corrected to the interval [0.52, 0.78], C mid = 0.65, and the confidence degree c = 0.60 corresponds to the membership degree μ temp (0.60) = (0.60-0.52) / (0.65-0.52) ≈ 0.615. The samples of each channel are sequentially calculated and mapped to the [0, 1] fuzzy domain to form the M scene matrix. For example, the vector of event E12 in the matrix is [0.65, 0.615, 0.42, 0.70, 0.38]. This set is used as a training sample for dynamic optimization of model parameters in the incremental learning of S7.4, which can significantly improve the adaptability and anti-interference of the confidence model under different scenes;

[0203] S7.4: Based on the incremental learning mechanism, the model parameters of the fuzzy confidence sample set are updated, the online support vector regression algorithm is used to dynamically optimize the confidence degree distribution parameters in the fuzzy confidence weighted table, so as to improve the adaptability of the model to the emerging or evolving scene characteristics;

[0204] S7.5: synchronize the updated fuzzy confidence model parameters to the multi-time period fusion inference module, and compare the false alarm discrimination results before and after the model update to generate a model optimization effect evaluation report as the decision basis for subsequent model iteration.

[0205] The step S8: dynamically adjust the threshold parameters of the discrimination model based on the historical false alarm events and the environmental change period, to improve the adaptability of the model to the feature shift caused by environmental drift and sensor aging. Specifically, it includes:

[0206] S8.1: cluster analysis is performed on the time stamp, environmental feature parameters and false alarm type label of the historical false alarm events to identify the periodic distribution pattern of the false alarm events. The repetitive periodic features of the false alarm events are extracted based on the clustering algorithm to obtain the periodic feature vector of the false alarm events as the input basis for model drift trend;

[0207] S8.2: model the long-term change trend of the multi-dimensional environmental parameters of the substation, extract the trend term and periodic term of the environmental variables such as temperature, humidity and electromagnetic field intensity based on time series analysis method, and generate the environmental drift trend sequence as the external driving factor for adjusting the discrimination model parameters;

[0208] S8.3: model the output drift of the multi-channel sensor based on the sensor aging model, calculate the sensor output drift coefficient matrix by using the deviation between the historical calibration data and the current output value of the sensor, to quantify the influence degree of sensor aging on the alarm features;

[0209] S8.4: input the periodic feature vector of the false alarm events, the environmental drift trend sequence and the sensor output drift coefficient matrix into the multivariate regression model, calculate the dynamic adjustment amount of the discrimination model threshold parameters based on the deviation between the model output and the actual false alarm when the false alarm occurs, and form the parameter update increment vector;

[0210] S8.5: based on the parameter update increment vector, the alarm threshold, the fuzzy confidence interval boundary and the abnormal recognition sensitivity parameters in the current discrimination model are updated online, the historical update amount and the current increment are fused by using the sliding window weighting method to realize the progressive adjustment of the model parameters, and ensure that the discrimination model still has high robustness and discrimination accuracy under the background of environmental drift and sensor aging.

[0211] The technical scheme of the present application has been described in conjunction with the preferred embodiments shown in the drawings, but those skilled in the art can easily understand that the protection scope of the present application is obviously not limited to these specific embodiments. Those skilled in the art can make equivalent changes or replacements to the related technical features without departing from the principles of the present application, and the technical schemes after these changes or replacements will fall within the protection scope of the present application.

[0212] The above merely provides the preferred embodiments of the present application, but not for limiting the present application; for those skilled in the art, the present application can have various modifications and changes. Any modification, equivalent replacement, improvement, etc. within the spirit and principle of the present application shall be included in the protection scope of the present application.

Claims

1. A method for effectively assessing false alarms in a fire alarm system within a substation from multiple dimensions, characterized in that, Includes the following steps: S1: Collect multi-channel sensor data within the substation and combine it with timestamps, weather information, load type, and operation and maintenance activity logs to form a scenario-based environmental dataset; S2: Normalize and denoise the collected sensor data to obtain normalized multidimensional sensor data. S3: Based on the spatiotemporal graph neural network, cluster analysis is performed on the normalized multidimensional sensor data to divide the environmental sub-regions with similar scene semantics into multiple time periods, and the typical characteristic patterns of fire signals and interference noise in each sub-region are learned. S4: Construct a fuzzy confidence interval model, express the confidence of alarm signals of each feature detection channel in a fuzzy set, and establish a fuzzy confidence weighted table of abnormal signals and a typical noise set for each time period in combination with historical scene distribution; S5: When an alarm signal is triggered, historical data of the same scene is called according to the current scene label. The alarm signal is weighted and judged under the multi-time period fuzzy confidence model, and the detection data of the previous and subsequent time periods are fused to perform Bayesian inference on the abnormal change trend. S6: Execute false alarm discrimination decision based on the reasoning results. If the signal shows high confidence anomaly in multiple time periods and scenarios and is out of the noise fuzzy set, it is marked as a high confidence real alarm. If there is a significant intersection between the confidence interval and the noise set, it is marked as a high probability false alarm and pushed to the manual review queue.

2. The method for effectively judging false alarms in a fire alarm system within a substation from multiple dimensions, as described in claim 1, is characterized in that... Following step S6, the following is also included: S7: Periodically perform scene labeling and fuzzy set updates for false alarms and real alarm cases, and optimize the parameters of the fuzzy confidence model using an incremental learning mechanism; S8: Based on historical false alarm events and environmental change cycles, dynamically adjust the threshold parameters of the discrimination model.

3. The method for effectively judging false alarms in a fire alarm system within a substation from multiple dimensions, as described in claim 1, is characterized in that... Step S1 specifically includes: Based on a multimodal sensor network deployed at the substation, the analog voltage signal output by the smoke concentration sensor is acquired, and then converted into a digital signal sequence through an ADC analog-to-digital converter module to obtain standardized smoke concentration characteristic data. A sliding window mean filter is applied to the real-time temperature data collected by the temperature sensor to obtain a stable temperature change characteristic sequence. Ambient humidity information is acquired using a capacitive humidity sensor, and missing or abnormal data points are filled in using a linear interpolation algorithm to generate a humidity feature time series. A high-frequency electromagnetic field probe was used to collect electromagnetic field intensity data inside the substation, and electromagnetic spectrum features were extracted using a fast Fourier transform algorithm. The grounding resistance value output by the grounding line status monitoring module is judged by a threshold. When an abnormal grounding is detected, the event is recorded and grounding status marker data is generated. The standardized smoke concentration feature data, the stable temperature change feature sequence, the humidity feature time series, the electromagnetic spectrum feature, and the line grounding status marker data are integrated, and timestamps, meteorological information, load type identifiers, and operation and maintenance activity logs are added to form a structured scenario-based environmental dataset.

4. The method for effectively judging false alarms in a fire alarm system within a substation from multiple dimensions, as described in claim 3, is characterized in that... In step S1, temperature data acquisition and processing includes applying weighted center-based sliding window mean filtering, outlier removal, and linear interpolation repair to the raw signal from the temperature sensor, and introducing a first-order difference module to extract temperature change rate features to enhance the detection sensitivity of instantaneous temperature rise events.

5. The method for effectively judging false alarms in a fire alarm system within a substation from multiple dimensions, as described in claim 1, is characterized in that... Step S2 specifically includes: The collected multi-channel sensor data is normalized, and the maximum-minimum normalization method is used to map each sensor data to a unified interval to obtain standardized sensor data under a unified dimension. Based on the wavelet transform algorithm, the standardized sensor data under the unified dimension is denoised in the time-frequency domain. The wavelet coefficients of each signal are extracted and a threshold is set to shrink the coefficients to obtain the time-domain signal sequence after wavelet denoising. The wavelet-denoised time-domain signal is subjected to sliding window mean filtering, and the signal is locally smoothed based on a set time window length to obtain the smoothed and filtered sensor signal output. Principal component analysis is used to perform feature dimensionality reduction on multi-channel sensor data. The covariance matrix between each feature is calculated and the principal component vector is extracted to obtain the dimensionality-reduced sensor feature matrix. The variance normalization process is performed on the dimensionality-reduced sensor feature matrix, and a weighted adjustment is performed based on the variance values ​​of each principal component feature to obtain the input feature vector set for scene semantic modeling.

6. The method for effectively judging false alarms in a fire alarm system within a substation from multiple dimensions, as described in claim 1, is characterized in that... Step S3 specifically includes: Based on normalized multidimensional sensor data, a high-dimensional feature space of the substation operating environment is constructed to form node feature vectors for graph structure modeling. Graph adjacency modeling is performed on the node feature vectors in the high-dimensional feature space, and a dynamic spatiotemporal adjacency matrix is ​​constructed based on timestamps and spatial correlations; The dynamic spatiotemporal adjacency matrix and the node feature vector are input into the spatiotemporal graph neural network model. The spatial correlation between nodes is extracted through graph convolution operation, and the temporal evolution trend of the environmental state is captured through the time gating mechanism to generate low-dimensional embedded features. Cluster analysis is performed on the low-dimensional embedding features, and the improved K-means++ algorithm is used in combination with the silhouette coefficient to optimize the number of clusters, thereby dividing the multi-time-period environmental sub-regions with similar scene semantics. Based on the multi-time period environmental sub-regions, feature pattern learning is performed on the fire signals and interference noise in each sub-region. Principal component analysis and discrete Fourier transform are used to extract typical frequency domain and time domain features, and abnormal signal feature templates and noise feature templates are constructed for each sub-region.

7. The method for effectively judging false alarms in a fire alarm system within a substation from multiple dimensions, as described in claim 6, is characterized in that... In step S3, spatiotemporal feature extraction combines a dynamic spatiotemporal adjacency matrix and a node feature input graph convolutional network with a time-gated recurrent unit, integrates convolutional layer output and attention mechanism, and performs joint modeling of the spatial topology and temporal evolution of multi-channel environmental data to obtain a low-dimensional embedding vector for scene clustering.

8. The method for effectively judging false alarms in a fire alarm system within a substation from multiple dimensions, as described in claim 1, is characterized in that... Step S4 specifically includes: The fire alarm characteristic signals collected by each sensor channel are processed by feature normalization to obtain a standardized feature vector set; Based on historical false alarm and real alarm event data, the standardized feature vector set is fuzzily partitioned using the fuzzy C-means clustering algorithm to construct a preliminary fuzzy noise set; Statistical analysis is performed on historical alarm data under scene labels for each time period. Based on fuzzy set theory, confidence membership functions for each feature channel are established to generate a time-related fuzzy confidence interval model. Based on the membership values ​​output by the fuzzy confidence interval model, and combined with the historical frequency distribution of false alarms and real alarms in each time period, a fuzzy confidence weighted table of abnormal signals in each time period is calculated and constructed. The fuzzy confidence weighted table and the fuzzy noise set are fused and modeled to generate a time-feature joint fuzzy criterion space.

9. The method for effectively judging false alarms in a fire alarm system within a substation from multiple dimensions, as described in claim 8, is characterized in that... In step S4, the establishment of the fuzzy confidence interval model includes applying Z-score standardization, Box-Cox power transformation and interval scaling to the alarm feature signal, combined with sliding window smoothing, inputting the fuzzy C-means clustering algorithm, identifying false alarm association clusters to form a fuzzy noise set, and then fitting the confidence membership function based on historical samples of each time period, and using a penalty weighting method for noise samples to improve the ability to express the confidence of real alarms.

10. The method for effectively judging false alarms in a fire alarm system within a substation from multiple dimensions, as described in claim 1, is characterized in that... Step S5 specifically includes: Based on the scene label of the current alarm signal, retrieve the historical fuzzy confidence weighted table and typical noise set within the corresponding time period to construct the scene adaptive discrimination basis; The confidence scores of the alarm signals in the current time period are mapped using fuzzy sets in each feature detection channel to generate fuzzy confidence intervals. These intervals are then compared with the retrieved historical noise sets to identify whether the current signal falls into the fuzzy region of a known interference pattern. Based on the overlap between the fuzzy confidence interval and the historical noise set, a weighted score calculation is performed on the current alarm signal. The fuzzy membership function is used to nonlinearly weight the confidence of each feature channel to generate a comprehensive anomaly score within the time period. Retrieve detection data under the same or similar scene labels in adjacent time periods, extract the change trend of its abnormal scores and fuzzy confidence, and perform trend fusion correction on the current abnormal scores based on the Bayesian inference model; By combining the weighted score of the current time period with the Bayesian inference results of the previous and next time periods, a final multi-time period fusion confidence value is generated. Based on the set discrimination threshold, the alarm signal is classified as a false alarm or a real alarm, and the discrimination result is output to the false alarm decision module.

Images