Distribution box fire early warning method and system based on multi-source information fusion

Abstract

The invention discloses a distribution box fire early warning method and system based on multi-source information fusion, and particularly relates to the technical field of distribution box fire early warning. Temperature, smoke, acoustic vibration and current harmonic signals are synchronously acquired, time sequence alignment and amplitude normalization are performed, and a dynamic baseline is generated in real time; extracting multi-scale statistics and morphological characteristics according to an adaptive window, and calibrating an operation mode through unsupervised clustering; then retrieving a matched baseline signature in a historical feature library, and calculating a multi-modal standardization deviation of a current window; mapping the deviation index into a graph node, combining a covariance edge weight, a phase synchronization index and a smoke discrete index, obtaining a coupling risk coefficient through a fusion model, and outputting a comprehensive risk index; and finally, carrying out multi-scale rate and acceleration analysis on the comprehensive risk index, triggering three-level early warning of attention, warning and danger by adopting a dynamic threshold, and correcting the threshold by utilizing operation and maintenance feedback self-learning, thereby effectively solving the problems of early warning response lag and unknown early warning level.

Description

Technical Field

[0001] This invention relates to the field of fire early warning technology for distribution boxes, and more specifically, to a method and system for fire early warning of distribution boxes based on multi-source information fusion. Background Technology

[0002] As a core node in low-voltage power distribution systems, distribution boxes are constantly exposed to diverse environments such as factories, high-rise buildings, old communities, and temporary construction sites. They house heat-generating components like circuit breakers, contactors, busbars, and secondary control circuits, and are simultaneously subjected to complex operating conditions involving alternating dust, oil mist, high humidity, high temperatures, and impact vibration. In recent years, the industry has widely deployed multi-modal sensors, including those for temperature, smoke, current, and acoustic vibration, both inside and outside distribution boxes. These sensors are then uploaded to the cloud via IoT edge gateways, and combined with rule bases or simple machine learning models to generate fire warnings. However, most existing solutions rely on fixed threshold triggers or single-modal anomaly detection: an alarm is triggered directly when the temperature exceeds a preset upper limit, the smoke concentration exceeds a certain percentage, or the harmonic distortion exceeds an empirical value. Further approaches, while introducing multi-source fusion, often focus on the union logic of simultaneous threshold exceedances, lacking dynamic correlation metrics and trend analysis. With the cyclical fluctuations of power load, baseline drift due to component aging, and random superposition of on-site interference, such static, single-scale criteria often cause false alarms during high-load transients or peak environmental dust. They also fail to be reported during slow failure processes such as hidden insulation degradation and gradual increase in contact resistance because they do not reach the threshold. This causes maintenance personnel to become fatigued with alarm information and makes it difficult to pinpoint the true source of the fire in a timely manner. Summary of the Invention

[0003] To overcome the aforementioned deficiencies in existing technologies, this invention provides a method and system for early warning of fires in distribution boxes based on multi-source information fusion. This method synchronously normalizes multi-source signals such as temperature, smoke, acoustic vibration, and current harmonics to generate dynamic baselines. It then uses unsupervised clustering to identify operating modes, calculates multimodal deviations based on historical signatures, constructs a correlation graph, and introduces phase synchronization and smoke dispersion indices to obtain a comprehensive risk index. Combined with multi-scale rate and acceleration dynamic thresholds, it achieves self-learning hierarchical early warning, thereby solving the problems mentioned in the background technology.

[0004] To achieve the above objectives, the present invention provides the following technical solution: a fire early warning method for distribution boxes based on multi-source information fusion, comprising:

[0005] Step S1: Collect temperature, smoke, acoustic vibration, and current harmonic signals, perform timing synchronization and amplitude normalization, correct sampling time difference and amplitude difference, and generate dynamic baseline;

[0006] Step S2: Based on the dynamic baseline, the normalized signal is windowed for a fixed duration, multi-scale statistical and morphological features are extracted, and unsupervised clustering is used to divide the operating mode in order to construct a feature vector set;

[0007] Step S3: Retrieve a baseline signature that matches the operating mode from the historical feature database, calculate the multimodal deviation of the current window features from the baseline signature, and output the standardized deviation index for subsequent coupling calculation;

[0008] Step S4: Map the deviation index to nodes, set edge weights according to covariance, introduce the phase synchronization index and smoke dispersion index, obtain the coupling risk coefficient through the fusion model and update the edge weights to generate a comprehensive risk index.

[0009] Step S5: Perform multi-scale rate and acceleration analysis on the comprehensive risk indicators, determine the risk level based on the dynamic threshold, trigger attention, warning or danger warning, and record the operation and maintenance feedback adaptive correction threshold.

[0010] Preferably, step S1 includes:

[0011] Step 101: Compensate for the delay of each sensing node by Kalman prediction and combine it with the dominant frequency phase-locked micro-correction to achieve microsecond-level timing synchronization of multi-source signals;

[0012] Step 102: Normalize the synchronized signal using quantile mapping and generate a dynamic baseline by convolving a local Gaussian kernel and a global exponential decay kernel to suppress slow environmental drift.

[0013] Step 103: Use synchronous confidence to drive Bayesian backtracking and smooth reconstruction of missing timestamps, and output a continuous corrected time series;

[0014] Step 104: Perform a dual-threshold assessment of stability and instantaneous drift rate on the dynamic baseline. Only when the thresholds are met is the baseline confirmed to be valid and the synchronization confidence and baseline stability output.

[0015] Preferably, step S2 includes:

[0016] Step 201: Adjust the window duration and overlap rate in real time based on the synchronization confidence. When the synchronization confidence decreases, widen the window and reduce the overlap ratio.

[0017] Step 202: Extract three types of statistical features—amplitude, spectrum, and morphology—within each window, and encode them relative to the dynamic baseline to form a baseline-relative feature vector;

[0018] Step 203: Use the density-adjusting kernel clustering algorithm to perform unsupervised clustering on the baseline relativized feature vectors, generate running mode labels and write them into the window index table.

[0019] Preferably, step S2 includes:

[0020] The baseline relativized feature vector is input into the density self-adjusting kernel clustering algorithm to adaptively determine the cluster center. After noise correction using the neighborhood consistency coefficient multidimensional reassignment, a stable cluster is generated. The cluster fingerprint is constructed using the skewness, kurtosis and modulus of the cluster center features and hashed to form a globally unique cluster label. The cluster label, along with the window timestamp and cluster stability, is written into the runtime mode metadata database for quick retrieval of the baseline signature in step S3.

[0021] Preferably, step S3 includes:

[0022] Step 301: Retrieve candidate baseline signatures in the historical feature database according to the operating mode label, and calculate the confidence level based on stability index, dynamic baseline consistency and feature similarity, and select the target baseline signature;

[0023] Step 302: Compare the current window features with the target baseline signature, calculate the four types of standardized deviation indices (amplitude, spectrum, phase, and dynamic baseline) and write them into the deviation buffer.

[0024] Preferably, step S4 includes:

[0025] Step 401: Map the standardized deviation index to temperature deviation nodes, smoke deviation nodes, acoustic deviation nodes and current deviation nodes, and assign the edge weights between nodes according to the historical covariance and Mahalanobis distance and write them into the edge weight matrix.

[0026] Step 402: Enhance the edge weights of temperature deviation nodes and current deviation nodes using the phase synchronization index, implement a suppression and enhancement strategy for the relevant edge weights of smoke deviation nodes using the smoke dispersion index, and update the edge weights by outputting the coupling risk coefficient through the fusion model to generate a comprehensive risk index.

[0027] Preferably, the method for obtaining the phase synchronization index is as follows:

[0028] Within the calibration time series Tc(t), perform Hilbert transforms on the temperature-normalized signal Ŝtemperature(t) and the current harmonic-normalized signal Ŝelectric(t) respectively to obtain the instantaneous phase sequences φtemperature(n) and φelectric(n). Calculate the sample-by-sample phase difference φΔ(n) = φtemperature(n) - φelectric(n), and obtain the phase synchronization index PSI according to the following formula:

[0029]

[0030] Where j represents the sequential number of the sampling point, and the total number of samples N corresponds to the number of sampling points in the current time window; the amplitude is obtained after summing the complex exponents and normalized, and the phase synchronization index PSI takes a value range of 0–1. The closer the value is to 1, the higher the phase lock between the temperature and current harmonic dominant frequencies, revealing the abnormal strength of electrothermal coupling.

[0031] Preferably, the method for obtaining the smoke dispersion index is as follows:

[0032] Using the normalized smoke signal Ŝsmoke(t) in the corrected time series Tc(t) as input, the peak sequence p_i is extracted by the adaptive thresholding method, where i = 1…M;

[0033] Calculate the probability of each peak amplitude relative to the total peak amplitude. The smoke dispersion index (SDI) is calculated using the following formula:

[0034]

[0035] Among them, the peak probability q_i is dimensionless, the number of peaks M is the number of peaks within the window, i represents the sequential number of peaks, and the denominator lnM normalizes the information entropy; the smoke dispersion index SDI ranges from 0 to 1. The higher the value, the more dispersed the peak amplitude distribution of the smoke signal, indicating that the smoke concentration fluctuates greatly and is unstable, and is more likely to correspond to early combustion or a surge in environmental dust.

[0036] Preferably, step S5 includes:

[0037] Step 501: Obtain the short-term rate by short-window differencing and the long-term rate by long-window linear regression slope for the comprehensive risk index, and obtain the change acceleration curve by low-pass filtering differencing;

[0038] Step 502: Based on the dynamic threshold, the comprehensive risk index, short-term rate, long-term rate and change acceleration are mapped to three levels of early warning: attention, warning and danger. The dynamic threshold and weighting factor are adaptively corrected using operation and maintenance feedback.

[0039] To achieve the above objectives, the present invention provides the following technical solution: a fire early warning system for distribution boxes based on multi-source information fusion, comprising:

[0040] Synchronization and normalization module: Collects temperature, smoke, acoustic vibration, and current harmonic signals, performs timing synchronization and amplitude normalization, corrects sampling time difference and amplitude difference, and generates dynamic baseline;

[0041] Feature clustering module: Based on the dynamic baseline, the normalized signal is windowed at a fixed duration, multi-scale statistical and morphological features are extracted, and unsupervised clustering is used to divide the operation mode in order to construct a feature vector set;

[0042] Baseline matching module: retrieves baseline signatures that match the operating mode from the historical feature database, calculates the multimodal deviation of the current window features from the baseline signature, and outputs the standardized deviation index for subsequent coupling calculation;

[0043] Fusion reasoning module: Maps deviation indicators to nodes, sets edge weights according to covariance, introduces phase synchronization index and smoke dispersion index, obtains coupling risk coefficient through fusion model and updates edge weights, and generates comprehensive risk index;

[0044] Trend warning module: Performs multi-scale rate and acceleration analysis on comprehensive risk indicators, determines the risk level based on dynamic thresholds, triggers attention, warning or danger warnings, and records operation and maintenance feedback to adaptively correct the thresholds.

[0045] The technical effects and advantages of this invention are as follows:

[0046] (1) This invention performs time synchronization and amplitude normalization processing on temperature signals, smoke signals, acoustic vibration signals and current harmonic signals collected by multi-source sensors. Combined with dynamic baseline construction, time window division and unsupervised clustering to identify the operation mode, it realizes accurate alignment and feature extraction of multimodal data. It effectively solves the problem that traditional static judgment methods cannot reflect the signal drift caused by load fluctuations, environmental changes and equipment aging. By retrieving baseline signatures in the historical feature library and calculating real-time deviation measurement, it realizes a fine characterization of the deviation of the actual operation state. It effectively solves the contradiction of false alarms and false alarms caused by fixed thresholds or simple difference analysis.

[0047] (2) This invention maps each modal deviation metric to graph nodes, calculates the phase synchronization index and smoke dispersion index respectively, and inputs these two indices into a predefined fusion model to calculate the coupling risk coefficient. This achieves deep fusion and weighted inference of multi-source features, effectively solving the problems of low accuracy of single-modal early warning and conflict of multi-dimensional features. By performing trend acceleration analysis on the generated comprehensive risk index and using adaptive thresholds for hierarchical early warning, it achieves early identification and multi-level prompts of potential fire risks, effectively solving the problems of delayed early warning response and unclear early warning levels, thereby significantly improving the reliability and practicality of the system. Attached Figure Description

[0048] Figure 1 This is a flowchart of the fire early warning method for distribution boxes according to the present invention. Detailed Implementation

[0049] Exemplary embodiments of the present disclosure will now be described in more detail with reference to the accompanying drawings. While exemplary embodiments of the present disclosure are shown in the drawings, it should be understood that the present disclosure may be implemented in various forms and should not be limited to the embodiments set forth herein. Rather, these embodiments are provided so that this disclosure will be thorough and complete, and will fully convey the scope of the disclosure to those skilled in the art.

[0050] At the same time, it should be understood that, for ease of description, the dimensions of the various parts shown in the accompanying drawings are not drawn according to actual scale.

[0051] The following description of at least one exemplary embodiment is merely illustrative and is in no way intended to limit the scope of this application and its application or use.

[0052] Techniques, methods, and equipment known to those skilled in the art may not be discussed in detail, but where appropriate, such techniques, methods, and equipment should be considered part of the specification.

[0053] See Figure 1 The present invention provides a flowchart of a fire early warning method for electrical distribution boxes. Figure 1 The fire early warning method for distribution boxes based on multi-source information fusion shown includes:

[0054] Step S1: Collect temperature, smoke, acoustic vibration, and current harmonic signals in the distribution box environment, perform timing synchronization and amplitude normalization, correct sampling time difference and amplitude difference, and generate dynamic baseline;

[0055] Step S1 specifically includes:

[0056] Step 101: Kalman Prediction Alignment

[0057] Using a global time anchor point as a reference, a time drift state vector containing the delays of four sensing nodes (temperature, smoke, acoustic vibration, and current harmonics) is constructed. An improved Kalman filter is used to recursively derive this vector, dynamically compensating for network jitter and node clock errors, and achieving coarse timing synchronization at the millisecond level.

[0058] Step 102: Phase-locked self-calibration

[0059] Based on Kalman coarse alignment, the instantaneous phase of the dominant frequency of the temperature signal and the current harmonic is extracted, and the phase residual between the two is calculated. If the phase residual exceeds the adaptive threshold, a fine-tuning offset is injected into the corresponding channel to correct the remaining microsecond-level time error and complete the synchronization of the two rounds of iteration.

[0060] Step 103: Self-perceptive backtracking smoothing

[0061] A synchronization confidence index is constructed based on the cumulative value of the synchronization residual. If the confidence is lower than the set threshold, Bayesian backtracking smoothing is triggered to reconstruct missing or misaligned timestamps to ensure that the corrected time series is continuous without breaks.

[0062] Step 104: Distribution Mapping Normalization

[0063] On a unified time grid, the quantile vectors of each signal in the short time window are calculated. A distribution transformation function based on boxline mapping is used to map the original amplitude to a dimensionless sequence with zero mean and unit variance, thereby resisting abnormal spikes and unifying the amplitude scale.

[0064] Step 105: Multi-scale trend filtering

[0065] The normalized signal is convolved with a local Gaussian kernel and a global exponential decay kernel to obtain the local trend and the global trend. Then, a weighting factor is generated based on the synchronization confidence, and the two trends are fused proportionally to output a dynamic baseline that is updated in real time as the environment evolves.

[0066] Step 106: Baseline Robustness Assessment and Refresh

[0067] The stability and instantaneous drift rate of the dynamic baseline within the observation window are calculated. The baseline is considered valid only if both indicators are below the preset threshold. If the condition is not met, the window is extended and the calculation is repeated. Finally, a stable dynamic baseline, synchronization confidence, and baseline stability are output, providing a reference for subsequent window division and cluster radius selection.

[0068] In summary, step S1 forms a closed loop from top to bottom: prediction-phase-locking-smoothing-normalization-filtering-evaluation. Kalman recursion and phase-locking work together to achieve microsecond-level synchronization, and self-sensing smoothing ensures stable time continuity. Distribution mapping and dual-scale filtering unify amplitude and dynamically track the environment. Robust evaluation eliminates potential drift in advance. Through deep coupling of the time and amplitude domains, the data coordinate system output by this step is both highly accurate and adaptable to operating conditions, providing a robust, reusable, and conflict-free parameter system for subsequent operation mode clustering and risk coupling.

[0069] Step S2: Based on the dynamic baseline, the normalized signal is windowed for a fixed duration, multi-scale statistical and morphological features are extracted, and unsupervised clustering is used to divide the operating mode in order to construct a feature vector set;

[0070] Background: The dynamic baseline has been constructed in step S1, and the timing and amplitude scales of multi-source signals have been unified. However, in actual operation, the distribution box exhibits a multi-peak, discrete state distribution, and a single global baseline cannot subdivide the various operating conditions caused by load changes, environmental disturbances, and device aging. Step S2 aims to divide the continuous time series into time windows with consistent granularity that can reflect detailed differences. For each window, a set of feature vectors covering amplitude, spectrum, and phase morphology are extracted, and unsupervised clustering is used to automatically classify them into several operating modes. The operating mode serves as the reference object for subsequent deviation measurement and as the semantic coordinate for dynamic risk inference. The entire process needs to be executed in real time within the limited computing power at the edge. Therefore, an adaptive windowing strategy and a lightweight density-driven clustering framework are designed. Through linkage with the dynamic baseline and synchronous confidence, the high timeliness and robustness of window construction, feature extraction, and mode classification are ensured.

[0071] Step S2 specifically includes:

[0072] Step 201: Read the normalized signal Ŝi(t) and dynamic baseline Bdyn(t) on the calibration time series Tc(t), calculate the window duration Δtwin and overlap rate ηlap in real time based on the synchronization confidence, complete the adaptive windowing, and generate a unique timestamp tk for each window; extract three sets of statistical features—amplitude, spectrum, and morphology—within each window to form a feature vector f(k), and then subtract the four baseline values ​​Bdyn(k) at the center point of the window to obtain the baseline relativized feature vector f(k). At the same time, write the window label, timestamp tk, and feature vector f(k) into the feature buffer; the feature vector f(k) maintains the integrity of multi-scale information and eliminates slow-changing environmental interference.

[0073] Furthermore, step S2 includes an adaptive windowing strategy for window duration and overlap rate, specifically:

[0074] The adaptive windowing strategy is driven by both synchronization confidence and environmental stability. During periods of high synchronization confidence, a shorter window duration is used to capture the details of rapidly changing signals. As synchronization confidence decreases, the window duration is lengthened to suppress misjudgments caused by jitter, while simultaneously reducing window overlap to minimize redundant calculations and smooth transient noise. To ensure that the segmentation boundary aligns with the actual signal change points, the dynamic baseline oscillation coefficient is evaluated before each window segmentation. If the oscillation coefficient exceeds the historical stability threshold, a minimum window duration is backtracked and re-evaluated, effectively avoiding truncating the drastic transition period at the window edge. The output of the window segmentation includes the window start and end timestamps, the dynamic baseline value at the window center point, and a synchronization confidence snapshot, all written into the window index table for subsequent feature extraction and clustering scheduling.

[0075] Mechanism Explanation: The window duration and overlap rate are simultaneously correlated with synchronization confidence and oscillation coefficient, ensuring that window partitioning considers both time alignment quality and signal stability. This solves the problem of fixed windows either being too wide to bury abrupt changes or too narrow to introduce noise in time-varying scenarios. Synchronization confidence is derived from the timing synchronization framework of the first step, thus eliminating name drift; the oscillation coefficient is directly calculated from the dynamic baseline, ensuring responsiveness. This strategy dynamically adjusts the window boundaries, ensuring that the feature extraction stage always operates on a timescale that is most sensitive to and stable with the actual field conditions.

[0076] Furthermore, step S2 includes a hierarchical extraction method for multi-scale statistical features, specifically:

[0077] The normalized signal within each window is fed into a hierarchical feature extraction pipeline. The first layer calculates metrics in the time domain such as amplitude peak-to-valley difference, sustained temperature rise slope, smoke signal rise duration, average acoustic pulse interval, and average current harmonic envelope to characterize slow trends and macroscopic dynamics. The second layer performs short-time frequency analysis in the spectral domain, extracting fundamental frequency energy, the proportion of third harmonic energy, the proportion of total higher harmonic energy, and the acoustic high-frequency energy bandwidth to capture high-frequency characteristics generated by sudden electric arcs or mechanical collisions. The third layer uses Hilbert transform in the morphological domain to obtain instantaneous phase and envelope, calculates phase difference metrics for temperature and current signals, calculates envelope peak volatility for acoustic vibration signals, and calculates envelope duration for smoke signals, thereby characterizing the nonlinear coupling features of the signal morphology. The three layers of features are concatenated into a feature vector and associated with a window index table.

[0078] Mechanism Explanation: Hierarchical feature extraction covers time, frequency, and morphology simultaneously, resulting in a dense yet dispersed distribution of windows in space. This means samples of the same operating mode cluster tightly, while samples of different modes are far apart, laying the foundation for high-quality partitioning by subsequent clustering algorithms without manual annotation. Since short-time frequency analysis and Hilbert transform are both conventional signal processing techniques, they do not produce proprietary indicators that are difficult to interpret, thus avoiding the "unclear function" issue mentioned in Article 26, Paragraph 3 of patent examination.

[0079] Furthermore, step S2 includes a dynamic baseline relativization coding mechanism, specifically:

[0080] To eliminate the drift effect caused by slow environmental changes, the feature vector of each window is subtracted from the dynamic baseline vector corresponding to the same timestamp before being written to the clustering cache, introducing the concept of relative deviation. The dynamic baseline comes from the multi-scale trend filtering in step S1, with each of the four signals having a one-dimensional reference value. After relativization, amplitude-type indicators become deviations from the baseline, spectral-type indicators become energy gains relative to the baseline spectrum, and morphological indicators represent deviations from the baseline phase and envelope. After relativization, the sensitivity of the feature vector to slowly changing factors such as solar radiation, temperature, and long-term load changes is significantly reduced, while maintaining high sensitivity to early signs of sudden faults. After encoding, the feature vector is written to the feature cache and awaits retrieval by the clustering scheduler.

[0081] Mechanism Explanation: Dynamic baseline relativization allows clustering algorithms to see in space the deviations that truly describe anomalies, rather than the natural drift of absolute values, thereby eliminating the overlap of multi-peak data at the macro scale and significantly improving clustering resolution.

[0082] Step 202: Read data in batches from the baseline relativized feature vector f(k) and corresponding timestamp tk output in Step 201. Use density self-adjusting kernel clustering algorithm to adaptively determine cluster centers by dual threshold screening of local density scaling and relative distance. Recursively assign the remaining samples according to the kernel function distance. Thus, the high-dimensional feature space is divided into several dense operating mode clusters without any prior labels. Calculate the neighborhood consistency coefficient κk for the initial clustering results. When the consistency coefficient is lower than the threshold, immediately reassign suspicious windows to the nearest center to remove isolated noise and converge the cluster shape. After the clustering is stable, calculate the mean of the central features for each cluster to generate a cluster fingerprint vector. Construct a three-dimensional fingerprint by statistical skewness, kurtosis, and modulus and perform hash mapping to obtain a globally unique cluster label Lc. At the same time, write the label Lk, timestamp tk, and stability of the cluster to which each window belongs into the mode label table, which is used as a unique key value for retrieving and matching the baseline signature, ensuring that subsequent multimodal deviation measurement can accurately correspond to the window operating mode.

[0083] Furthermore, step S2 includes unsupervised clustering and pattern cluster labeling, specifically:

[0084] In the clustering stage, a density self-adjustment mechanism is used to determine the cluster centers. First, the local density estimate of each window in space is calculated and sorted according to density. Then, the center is determined based on a dual threshold of distance separation and local density. The dual threshold is not a static constant, but is adaptively updated according to the number of windows in the current batch and the overall density mean, so as to ensure that a reasonable number of centers can still be selected when the amount of data decreases or increases sharply. After the centers are determined, a kernel density recursive strategy is used to map the remaining windows to the nearest center to form the initial cluster.

[0085] Mechanism Explanation: Density-adjustable kernel clustering avoids the problems of excessive cluster cutting or merging caused by traditional fixed density thresholds. By statistically analyzing the current sample density distribution in real time, it automatically adjusts the number of cluster centers as the pattern complexity increases or decreases, solving the pain point of operational mode mismatch when multi-peak and single-peak conditions alternate in the field, while ensuring that the computational complexity is controllable within a linear range.

[0086] Furthermore, step S2 includes neighborhood consistency verification and reassignment strategy, specifically:

[0087] After the initial clustering is completed, a neighborhood consistency coefficient is calculated for each window, which is the proportion of the window's neighbors in the same cluster to its nearest neighbors. If the neighborhood consistency coefficient is lower than the set consistency threshold, the window is marked as a suspicious sample and an attempt is made to reassign it to a neighboring cluster center. During the reassignment process, not only Euclidean distance is considered, but also the difference in the synchronization confidence of the dynamic baseline and the oscillation coefficient when the window is generated, to ensure that samples within the same cluster remain compact and consistent in the three dimensions of time, space and dynamic baseline. The neighborhood consistency check effectively filters noisy windows to prevent them from pulling the cluster center off track or acting as isolated noise that affects subsequent bias measurements.

[0088] Mechanism Explanation: Neighborhood consistency verification performs multi-dimensional reassignment of suspicious samples, taking into account spatial distance, temporal consistency, and dynamic baseline coupling characteristics. This enables posterior correction of clustering results, which is more comprehensive than traditional reassignment based solely on spatial distance. It integrates data cleaning and clustering optimization into the same process, achieving lightweight, integrated processing.

[0089] Furthermore, step S2 includes cluster fingerprint generation and semantic tag persistence, specifically:

[0090] After determining the cluster structure, the mean value of the feature vectors within each cluster is calculated to obtain the cluster center feature. Three statistical indicators—skewness, kurtosis, and feature modulus length—are then calculated for the cluster center feature to form a cluster fingerprint. The cluster fingerprint is hashed to obtain semantic labels, which are then written to the runtime mode metadata database. The metadata database records metadata such as the cluster fingerprint, window start and end time periods, the number of cluster members, and the distribution of synchronization confidence. Subsequently, when a new window arrives, historical clusters with high fingerprint similarity are prioritized in the metadata database, enabling rapid incremental updates without rerunning the full clustering. Mechanism explanation: Cluster fingerprints not only improve the runtime mode persistence efficiency but also allow the historical feature library to directly index the corresponding baseline signature through the label during step three, saving multiple rounds of matching calculations and reducing latency. The hash mapping method uses common segmented low-collision hashing, avoiding private encoding that is uninterpretable to the outside world.

[0091] In summary, step S2 transforms multimodal data from a uniform scale into feature vectors depicting operational details through adaptive windowing and feature extraction, and assigns stable and traceable pattern labels to each time window using unsupervised clustering. The adaptive windowing strategy addresses the difficulty of applying uniform window granularity under conditions of fluctuating synchronous confidence and environmental oscillations; hierarchical feature extraction and dynamic baseline relativistic encoding ensure that clustering focuses on differences directly related to potential anomalies rather than slow environmental changes; density-adjusting kernel clustering, neighborhood consistency verification, and cluster fingerprint persistence provide multiple guarantees for real-time performance, purity, and traceability in pattern partitioning. The output pattern labels and window indices will be directly used in step S3 to accurately match historical baseline signatures and calculate the bias metric.

[0092] Step S3: Retrieve a baseline signature that matches the operating mode from the historical feature database, calculate the multimodal deviation of the current window features from the baseline signature, and output the standardized deviation index for subsequent coupling calculation;

[0093] Background: In step S2, time window segmentation and unsupervised clustering based on dynamic baselines have been completed, and operating mode labels corresponding to each time window have been output. Although this process allows signals to be reasonably grouped, mode labels alone are insufficient to support subsequent comprehensive risk inference. This is because different operating modes will exhibit completely different degrees of signal deviation under historical operating conditions. Without a baseline signature based on historical features as a reference, it will be impossible to accurately determine the difference between the current window signal and the normal operating condition. The goal of step S3 is to establish this reference system: by searching the historical feature database, the baseline signature corresponding to the current operating mode is found, and then the difference between the current window features and the baseline signature is quantified to obtain a representative deviation measure. The deviation measure can establish a unified measurement standard among different modes and operating modes, thereby providing reliable input data for step S4 to calculate the phase synchronization index, smoke dispersion index, and comprehensive risk index.

[0094] Step S3 specifically includes:

[0095] Step 301: Receive the window label Lk and timestamp tk output in step 202. Locate candidate records in the first-level index of the historical feature database using label Lk as the primary key. Based on cluster fingerprint similarity, refine the selection in the second-level index and extract several baseline signatures with the smallest distance to the current cluster fingerprint. Compare the consistency between the dynamic baseline snapshot of the candidate baseline signature and the current dynamic baseline snapshot, and calculate the comprehensive confidence score by combining three factors: signature stability index, snapshot difference, and distance between the current window feature and the cluster center feature. When the comprehensive confidence score is higher than a preset threshold, determine the candidate baseline signature with the highest confidence score as the standard reference for the current operating mode and write it into the matching result table. If the current window is determined to be in a stable state after subsequent quality verification, its features are used to incrementally update the baseline signature, realizing adaptive maintenance of the historical feature database and ensuring that subsequent deviation calculations are always based on the latest normal operating conditions.

[0096] Furthermore, step S3 includes historical feature database retrieval and baseline signature matching, specifically:

[0097] The historical feature database uses the operating mode label as the primary index key, with each operating mode corresponding to one or more baseline signature records. Each baseline signature record consists of a pattern cluster central feature vector, a dynamic baseline snapshot, stability indicators, and timestamp information, all derived from statistical analysis of a large amount of data under normal operating conditions. The feature database supports incremental updates; when a new data window is determined to be in a stable state and passes quality verification, its features can be used to update the baseline signature of that pattern to maintain its up-to-date representativeness. The entire indexing process remains consistent with the pattern labels and window indexes output in step S2, thus ensuring complete integration between retrieval and pattern parsing.

[0098] Furthermore, the selection and confidence assessment of the baseline signature in step S3 are specifically as follows:

[0099] After locating candidate baseline signature records, a confidence score is calculated for each record. The confidence score is based on a comprehensive evaluation of three factors: first, the stability index of the candidate baseline signature; second, the degree of difference between the candidate baseline signature and the current dynamic baseline snapshot; and third, the similarity between the current window features and the candidate baseline center feature vector. The stability index reflects whether the baseline signature has been in a non-anomaly state for a long period of time, and is recorded as the stability score Ps by the proportion of baseline signatures that have been in a non-anomaly state continuously within the same label historical window. The dynamic baseline difference reflects the consistency between the current operating conditions and the historical background, and is expressed as the Euclidean distance between the current dynamic baseline snapshot and the candidate snapshot. The dynamic baseline difference score is denoted as Pd after normalization. Feature similarity directly reflects the distance between patterns, and is denoted as Pf after inverse normalization of the Mahalanobis distance from the current window feature vector to the candidate baseline center feature vector. The comprehensive confidence score is calculated by combining these three factors. The three scores are linearly weighted according to weights w1, w2, and w3 to obtain the comprehensive confidence score C = w1·Ps + w2·Pd + w3·Pf, where the weights satisfy w1 + w2 + w3 = 1. If the confidence score of a candidate baseline signature exceeds the set threshold, it is used as the standard reference for the current operating mode. If multiple records meet the conditions, the one with the highest confidence score is selected.

[0100] Step 302: Calculation and Standardization of Deviation Measurement

[0101] Read the baseline signature and window feature vector f̃(k) determined in step 301, and calculate four types of deviations: amplitude, spectrum, phase, and dynamic baseline. The amplitude deviation is the absolute difference between the current window amplitude index and the baseline amplitude index, normalized by the amplitude range. The spectrum deviation is represented by the inverse value of the overlap between the current energy distribution and the baseline energy distribution. The phase deviation is obtained by comparing the instantaneous phase difference with the envelope volatility. The dynamic baseline deviation is the difference between the current window baseline snapshot and the baseline signature snapshot, normalized. Based on the discriminative power of each operating mode in historical anomaly detection, the weights are adaptively set, and the four types of deviations are weighted and summarized to generate the original comprehensive deviation value. If the comprehensive deviation exceeds the historical fluctuation limit, the anomaly enhancement function is activated to amplify the deviation contribution in a nonlinear manner. After the enhancement is completed, unified standardization is performed in combination with the historical fluctuation range of each operating mode, and four single-mode standardized deviation indices and one comprehensive standardized deviation index are output. These, along with the window timestamp, are written into the deviation buffer for direct use when constructing the deviation graph node in step 401.

[0102] Furthermore, step S3 includes multimodal joint calculation of the deviation metric, specifically:

[0103] After obtaining the baseline signature corresponding to the current operating mode, the deviation of the current window features relative to the baseline signature is calculated. The deviation metric covers multimodal features, including amplitude features, spectral features, phase features, and dynamic baseline differences. The deviation of amplitude features is obtained by normalizing the absolute difference between the current window amplitude index and the baseline signature amplitude index. The deviation of spectral features is represented by the degree of difference between the current window spectral energy distribution and the baseline spectral distribution, calculated using energy ratio or distribution overlap index. The deviation of morphological features is obtained by comparing phase difference and envelope volatility. The deviations of all operating modes are weighted and aggregated, and the weights of different modes are dynamically adjusted by their discriminative power in historical anomaly detection. The weighted deviation values ​​are standardized to a dimensionless unified scale, allowing direct comparison of deviation results under different operating modes and different signal modes.

[0104] Furthermore, step S3 includes anomaly enhancement and standardization of the deviation metric, specifically:

[0105] To improve the responsiveness of deviation metrics in abnormal scenarios, an anomaly enhancement mechanism is introduced. When the deviation value exceeds the normal fluctuation range, its contribution is amplified through a nonlinear function, enabling the deviation metric to show a significant improvement in the early anomaly stage. This enhancement mechanism combines the historical deviation distribution and the upper and lower bounds of the current dynamic baseline to ensure that the amplification effect is not triggered by instantaneous noise. The enhanced deviation value is mapped to a dimensionless scoring interval through standardization transformation, which can be directly used in subsequent steps. The standardization process considers the historical fluctuation intervals of different operating modes, thereby enabling the deviation metric to be comparable across modes. The output of this process is a standardized deviation index for each mode and a comprehensive deviation index, both of which are passed to step S4 for the construction of deviation map nodes.

[0106] Mechanism Explanation: The anomaly enhancement mechanism amplifies the difference between anomalous signals and baseline conditions, improving early warning capabilities, while the standardization mechanism ensures that deviation indices can be uniformly measured across modalities and modes, eliminating comparison barriers caused by different signal characteristics and mode backgrounds. This design enhances sensitivity and versatility, allowing subsequent comprehensive risk analysis to directly utilize standardized deviation indices without requiring additional adjustments for different modes.

[0107] Step S3 Summary: Step S3 provides a quantitative reference for the difference between the current operating condition and the normal operating condition through historical feature database retrieval, baseline signature matching, and multimodal deviation measurement. The hierarchical indexing and dynamic update mechanism of the historical feature database ensure the accuracy and timeliness of baseline signature selection, while multidimensional confidence assessment further improves the reliability of reference matching. The deviation measurement adopts multimodal joint calculation and introduces anomaly enhancement and standardization processing, making the deviation index sensitive to early anomalies and allowing direct comparison across modes and conditions. These results are passed to step S4 as the basic input for deviation map node construction and comprehensive risk index calculation.

[0108] Step S4: Map the deviation index to nodes, set edge weights according to covariance, introduce the phase synchronization index and smoke dispersion index, obtain the coupling risk coefficient through the fusion model and update the edge weights to generate a comprehensive risk index.

[0109] Background: The first three steps have already output three types of key inputs: first, standardized deviation indices organized by time windows; second, the basic data (phase sequence, smoke peak sequence) required for the phase synchronization index and smoke dispersion index; and third, operating mode labels and their window timestamps used to measure the strength of correlation between different modes. The core task of step S4 is to aggregate these heterogeneous inputs into a multimodal deviation graph on a unified time axis, and introduce the phase synchronization index and smoke dispersion index as node features within the graph structure. After calculating the coupling risk coefficient and edge-weighted inference, a comprehensive risk index that can be directly measured across all modes is output. To balance real-time performance and interpretability, this step adopts a two-layer processing chain: the first layer maps the deviation indices and operating modes to graph nodes and establishes initial edge weights; the second layer, driven by the phase synchronization index and smoke dispersion index, calculates the coupling risk coefficient through a fusion model and dynamically adjusts the edge weights, thereby enabling the comprehensive risk index to maintain high sensitivity under complex operating conditions.

[0110] The core objective of step S4 is to convert the multimodal deviation index into a unified graph structure representation, and to dynamically update the edge weights through a fusion model driven by the phase synchronization index and the smoke dispersion index, ultimately generating a comprehensive risk index under the graph reasoning framework, providing high-resolution input for trend acceleration analysis.

[0111] Furthermore, the method for obtaining the phase synchronization index is as follows:

[0112] First, within the calibration time series Tc(t), perform Hilbert transforms on the temperature-normalized signal Ŝtemperature(t) and the current harmonic-normalized signal Ŝelectric(t) respectively to obtain the instantaneous phase sequences φtemperature(n) and φelectric(n). Calculate the sample-by-sample phase difference φΔ(n) = φtemperature(n) - φelectric(n), and then obtain the phase synchronization index PSI according to the following formula:

[0113]

[0114] Where j represents the sequential number of the sampling point, and the total number of samples N corresponds to the number of sampling points in the current time window; the amplitude is obtained after summing the complex exponents and normalized, and the phase synchronization index PSI takes a value range of 0–1. The closer the value is to 1, the higher the phase lock between the temperature and current harmonic dominant frequencies, revealing the abnormal strength of electrothermal coupling.

[0115] Furthermore, the method for obtaining the smoke dispersion index is as follows:

[0116] Using the normalized smoke signal Ŝsmoke(t) in the corrected time series Tc(t) as input, the peak sequence p_i is extracted by the adaptive thresholding method, where i = 1…M;

[0117] Calculate the probability of each peak amplitude relative to the total peak amplitude. The smoke dispersion index (SDI) is calculated using the following formula:

[0118]

[0119] Among them, the peak probability q_i is dimensionless, the number of peaks M is the number of peaks within the window, i represents the sequential number of peaks, and the denominator lnM normalizes the information entropy; the smoke dispersion index SDI ranges from 0 to 1. The higher the value, the more dispersed the peak amplitude distribution of the smoke signal, indicating that the smoke concentration fluctuates greatly and is unstable, and is more likely to correspond to early combustion or a surge in environmental dust.

[0120] Step S4 specifically includes:

[0121] Step 401: Deviation Graph Node Construction and Initial Edge Weight Setting

[0122] Read the four single-modal standardized deviation indices output in step 302—temperature deviation, smoke deviation, acoustic deviation, and current deviation—and construct a deviation vector based on the window timestamp tk. Establish a two-dimensional index surface using the window timestamp as the row index and the standardized deviation index as the column index, mapping each deviation index to a temperature deviation node, smoke deviation node, acoustic deviation node, and current deviation node. Assign the node weights directly to the corresponding standardized deviation scores (ranging from zero to one hundred), and write the operating mode label Lk into the node metadata. Call the deviation covariance matrix of the operating mode with the same label from the historical feature library, calculate the Mahalanobis distance between the current deviation vector and the historical covariance space, and back-normalize the distance results to generate the initial edge weights between the four nodes in the same window. Higher edge weight values ​​indicate stronger coordinating deviations between the two modes. Write the node weights, edge weight matrix, and neighborhood average edge weight into the deviation graph data structure, providing a complete topological foundation for step 402 to dynamically update the edge weights and calculate the coupling risk coefficient based on the phase synchronization index and smoke dispersion index.

[0123] Step 402: After reading the node weights and initial edge weight matrix output in Step 401, align the phase synchronization index time series with the timestamp tk, convert the phase synchronization index into a synchronization weight increment using piecewise linear mapping, and apply it only to the edge between the temperature deviation node and the current deviation node; when the phase synchronization index is in the background interval, maintain the original edge weight; when it enters the sensitive interval, increase the edge weight by the increment ratio; when it exceeds the high-risk threshold, directly increase the edge weight to the maximum value and reset the increment benchmark at the end of the window; read the smoke dispersion index time series, and implement a suppression-re-enhancement strategy for all incident edges of the smoke deviation node: maintain the original value in the normal interval, reduce the edge weight to offset environmental noise when it enters the moderate dispersion interval, and when the smoke dispersion index exceeds the high-risk threshold, re-enlarge the edge weight and enable local subgraph isolation to prevent false diffusion; the updated edge weight matrix, together with the node weights, is input into the fusion model to calculate the coupling risk coefficient, thereby generating a new comprehensive risk index and writing it into the risk buffer for multi-scale rate and acceleration analysis in Step 501.

[0124] Furthermore, step S4 includes a method for calculating the modal correlation of the initial edge weights, specifically as follows:

[0125] After creating nodes, initial edge weights need to be assigned to pairs of nodes within the same window to represent the correlation strength between different modal deviations. The covariance matrix of each modal deviation index under the same operating mode in the historical feature database is read, and the Mahalanobis distance of the current window deviation vector in this covariance space is calculated. A smaller Mahalanobis distance indicates a closer approximation to the historical normal state, corresponding to a lower edge weight; a larger distance indicates that both modes deviate from the normal state, corresponding to a higher edge weight. Edge weights are normalized to zero to a range, providing an initial reference for calculating the coupling risk coefficient of the next technical point. After the edge weight update is written into the graph structure, the average edge weight of the neighborhood is recorded for each node to form a local density index for subsequent identification of high-risk clusters.

[0126] Scenario Description: In a scenario where a distribution box experiences a simultaneous increase in acoustic vibration and current harmonics due to loose cables, the Mahalanobis distance between the acoustic deviation node and the current deviation node will be much greater than the distance between them and the temperature deviation node, thus significantly increasing their edge weights. The increase in the average edge weight of the neighborhood indicates that these two modes in the fusion model have a high degree of resonance, which requires extra attention.

[0127] Furthermore, step S4 includes a dynamic update method for edge weights driven by the phase synchronization index and the smoke dispersion index, specifically as follows:

[0128] Phase synchronization index driven synchronization weight adjustment strategy

[0129] The phase synchronization index time series is read and mapped to the connection edge between the temperature deviation node and the current deviation node. Specifically, a synchronization weight increment that varies with the phase synchronization index is added to the initial edge weight. The increment calculation adopts a piecewise linear mapping: when the phase synchronization index is below the upper threshold of the background confidence interval, the edge weight is not increased; when the phase synchronization index enters the sensitive interval, the edge weight increases linearly proportionally; when the phase synchronization index exceeds the high-risk threshold, the edge weight jumps directly to the maximum value. This strategy allows the electrothermal coupling characteristics to be dynamically enhanced in the graph structure according to the risk level, while avoiding the problem of continuous edge weight expansion due to long-term high exponents.

[0130] Scenario description: The phase synchronization index will slowly rise in the early stage of local overload, and the edge weight increment will increase accordingly, indicating that the phase lock of the model temperature and current harmonics is strengthening; when the overload intensifies and causes the phase synchronization index to jump suddenly, the edge weight jumps to the maximum value, and the graph inference immediately identifies that the coupling risk of this mode has entered the danger zone, triggering a significant increase in the subsequent comprehensive risk index.

[0131] Discrete weight suppression mechanism driven by smoke discrete index

[0132] In contrast to the enhancement logic of the phase synchronization index, the smoke dispersion index primarily affects all incident edges of smoke deviation nodes. When the smoke dispersion index is within the normal range, the edge weights remain unchanged. When the smoke dispersion index rises to the moderate dispersion range, the edge weights between smoke nodes and non-smoke nodes are reduced to weaken the local disturbances caused by the smoke dispersion itself. When the smoke dispersion index exceeds the high-risk threshold, the edge weights between smoke nodes and other nodes are increased again, but at the same time, local subgraph isolation is activated to prevent abnormal smoke signal propagation from misleading the overall graph reasoning. This suppression-re-enhancement strategy ensures that smoke signal anomalies can be captured without causing misjudgments of the overall risk index during the local dispersion stage.

[0133] Scenario Description: During the rainy season when dust is high, the smoke dispersion index may rise due to environmental factors but not exceed the high-risk threshold. The suppression mechanism promptly reduces the smoke-related edge weights, preventing environmental dust from being mistaken for fire smoke. Once an electric arc actually causes continuous smoke generation, the scenario description exceeds the high-risk threshold, the edge weights are increased again and locally isolated, and the model can quickly lock the node clusters that resonate with the smoke anomaly and other modes, achieving accurate localization.

[0134] Step S4 Summary: Step S4 integrates the multimodal deviation index, phase synchronization index, and smoke dispersion index using a graph structure method, enabling a hierarchical representation of various risk information in the topological space. Deviation nodes and covariance edge weights jointly characterize the basic deviation relationships, while the phase synchronization index and smoke dispersion index play a dual role in amplifying key couplings and filtering environmental noise when dynamically adjusting the edge weights. The resulting comprehensive risk index retains the global perspective of cross-modal information fusion while possessing the ability to analyze local anomalies in key couplings, providing high-resolution, low-noise input for the trend acceleration analysis in Step S5. Through strict terminology locking, dimensionless indexing, and dynamic edge weight adjustment, this scheme achieves deep coupling of multi-source information in the spatial and temporal domains, avoiding the problems of insufficient sensitivity or excessive false alarms in complex environments caused by traditional fixed threshold methods.

[0135] Step S5: Perform multi-scale rate and acceleration analysis on the comprehensive risk indicators, determine the risk level based on the dynamic threshold, trigger attention, warning or danger warning, and record the operation and maintenance feedback adaptive correction threshold.

[0136] Background: The comprehensive risk index has already been output as a dimensionless sequence that increases over time through the graph reasoning framework in step S4, and forms a continuous time series in the deviation buffer. Compared with the local early warning signal of a single mode, the comprehensive risk index integrates the deviation information and coupling relationship of four signals: temperature, smoke, acoustic vibration and current harmonics. It is the most direct global quantity reflecting the overall fire risk of the distribution box. However, the comprehensive risk index itself is still a measure of the intensity of transient risk and has not yet reflected the two important dimensions of "risk change speed" and "risk change acceleration" that describe the trend of risk evolution. Without quantitative analysis of the growth rate and acceleration of the indicators, it is impossible to accurately determine whether the fire risk is a short-term fluctuation or a continuous increase, making it difficult to issue graded warnings at the most appropriate time window. The goal of step S5 is to fill this gap: through the trend acceleration analysis chain, the comprehensive risk indicators are mapped into two derived curves of change rate and acceleration, identifying the continuous upward characteristics of risk from a multi-scale perspective; at the same time, combined with the dynamic threshold adjustment strategy, clear warning levels are assigned to different upward levels and alarm logic is triggered; in the algorithm design, this step adopts three complementary measures: hierarchical trend filtering, dynamic threshold adaptation, and graded alarm mapping, to ensure that it can both suppress instantaneous noise and capture early signs of slow risk increase.

[0137] The core objective of step S5 is to construct multi-scale change rate and acceleration curves using the time series of comprehensive risk indicators as input, and to determine the early warning level and output alarm signals in real time through a dynamic threshold and hierarchical mapping mechanism.

[0138] Step S5 specifically includes:

[0139] Step 501: Read the continuous comprehensive risk index time series in the risk buffer. Slide sample the same series using two scales: short window and long window. For the short window, perform first-order differencing on the index to obtain the short-term rate. For the long window, calculate the slope using local linear regression to form the long-term rate. After unified normalization, the two are concatenated into a velocity vector. The weighting factors are dynamically adjusted according to the local variance of the short-window and long-window rates. The smaller weight represents high noise and is suppressed, while the larger weight represents a stable trend and is strengthened. The weighted velocity vector is low-pass filtered and then second-order differencing is calculated to obtain the changing acceleration curve. The curve is synchronized with the window timestamp and written into the rate buffer. At the same time, the attention threshold, warning threshold, and danger threshold in the dynamic threshold table are updated according to the median acceleration of the past hour and stored in the threshold buffer to provide a real-time threshold benchmark for step 502.

[0140] Step 502: Read the comprehensive risk index, velocity vector, acceleration curve, and dynamic threshold table sequentially. First, determine the range of the comprehensive risk index value. Then, combine the relative relationship between the short-term rate and changing acceleration curve and the thresholds at each level to map the current warning level into four levels: normal, attention, warning, and danger. When the level triggers attention or above, the alarm actuator simultaneously activates the local audio-visual equipment and pushes the alarm information to the operation and maintenance platform, recording the alarm time, level, and risk context to the alarm log. After the operation and maintenance personnel finish processing, write the results to the feedback interface. Based on the feedback, automatically adjust the learning rate of the dynamic threshold table and weight factors. If there is a false alarm, increase the threshold or decrease the weight. If it is effective, maintain or tighten the threshold, forming a threshold self-learning closed loop driven by practical results. Write the updated threshold table back to the threshold cache for use in the next cycle.

[0141] Furthermore, step S5 includes multi-scale trend acceleration analysis, specifically:

[0142] First, slide two time windows across the time series of the comprehensive risk indicator:

[0143] One group consists of short windows used to capture rapid mutations;

[0144] The other set consists of long windows, used to track slow ascents;

[0145] The first difference of the comprehensive risk index sequence within the short window is performed to obtain the short-term rate of change sequence; the local linear regression slope estimation is performed on the sequence within the long window to obtain the long-term rate of change sequence.

[0146] The two types of velocities are normalized using a uniform dimensionless scale and then concatenated into a velocity vector, which is then mapped back to the same time axis.

[0147] A weighting factor is introduced, which is dynamically adjusted according to the local variance of the short window rate and the long window rate: when the short window rate fluctuates sharply, its weight is reduced to avoid short-term noise dominating; when the long window rate rises slightly over a long period of time, its weight is appropriately increased to ensure that the gradual upward trend receives sufficient attention.

[0148] After normalization and weight adjustment, the velocity vector is written into the velocity buffer to provide input for calculating acceleration.

[0149] Mechanism Explanation: The dual-window strategy is adopted because rapid heating or arc discharge can often cause a sharp increase in the comprehensive risk index within ten seconds, while chronic aging of the conductor or long-term increase in ambient temperature takes several minutes to manifest. By measuring the rate in the two time windows separately and adaptively weighting it according to the noise intensity, the problem of missed detection or false alarm caused by focusing on only one time scale is avoided. In addition, the slope of local linear regression is better at smoothing noise in long windows than simple difference, making the long-term rate sensitive to slow trends and unaffected by local fluctuations.

[0150] Furthermore, step S5 includes generating the changing acceleration curve and the threshold triggering parameter, specifically:

[0151] After the velocity vector is written to the rate buffer, the second-order difference is calculated to obtain the changing acceleration curve. To avoid amplification of small noises in the velocity vector that could cause oscillations in the acceleration curve, the velocity vector is first low-pass filtered before the difference is calculated. After the acceleration curve is generated, a dynamic threshold table is synchronized according to the window timestamp: the threshold table stores three levels of acceleration thresholds, corresponding to the three levels of attention, warning, and danger; the threshold update mechanism is dynamically adjusted based on the median absolute value of acceleration over the past hour: if the overall measured acceleration over the past hour increases, the threshold is moved up; otherwise, it is moved down, thus ensuring that the threshold can drift synchronously with the environment and equipment status; the updated threshold is written to the threshold buffer for the hierarchical alarm mapping call in step 502.

[0152] Mechanism Explanation: The acceleration curve reveals the change in the rate of risk growth, while dynamic threshold adjustment ensures that alarm sensitivity does not become unbalanced due to long-term background changes. For example, during the high temperatures of summer, the comprehensive risk index may remain high, but a synchronous upward shift of the threshold can prevent frequent false alarms; when cold air arrives and the background risk decreases, a downward shift of the threshold can promptly detect a new round of abnormal growth.

[0153] Furthermore, step S5 includes hierarchical early warning mapping and alarm execution, specifically as follows:

[0154] The hierarchical early warning mapping is as follows: First, the comprehensive risk index, velocity vector, acceleration curve, and dynamic threshold table are read. For each time point, the following judgments are made sequentially: If the comprehensive risk index is below the lower threshold and the acceleration curve has not triggered the attention level threshold, the normal state is maintained; if the comprehensive risk index is between the lower and middle thresholds and any velocity component exceeds the short-window rate threshold, an attention level early warning is output; if the comprehensive risk index is between the middle and upper thresholds and the acceleration curve exceeds the warning threshold, a warning level early warning is output; when the comprehensive risk index exceeds the upper threshold or the acceleration curve crosses the danger level threshold, a danger level early warning is immediately output. The mapping rules clearly distinguish the weights of comprehensive level quantities and dynamic quantities to prevent simple numerical spikes or instantaneous noise from triggering incorrect alarm levels. In this invention, the early warning signal can be transmitted via wireless networking.

[0155] Mechanism Explanation: In a scenario where current harmonics gradually increase and temperature rises slowly, the comprehensive risk index may continue to hover around the threshold. However, as long as the acceleration curve does not exceed the warning threshold, it will not enter the excessive warning level. Conversely, when arc discharge suddenly occurs and the acceleration curve rises rapidly, even if the current comprehensive risk index has not exceeded the upper threshold, an early alarm will be triggered due to dangerous acceleration, thus achieving an advanced response.

[0156] Alarm Execution: Once the warning level is determined by the hierarchical mapping, the alarm executor is invoked to issue multi-channel alarms: local audible and visual alarms, push notifications from the operations and maintenance platform, and SMS and email notifications, etc. After each alarm, the alarm executor records the alarm time, alarm level, and risk context, including comprehensive risk indicators, velocity vectors, and acceleration curve values. After the operations and maintenance personnel process the alarm, they write the processing results back to the processing feedback interface, which then performs reinforcement learning-based updates to the threshold adaptively: if the operations and maintenance personnel confirm that the dangerous level alarm is valid, the current threshold is maintained; if the operations and maintenance personnel report a false alarm, the learning rate of the upper and lower thresholds or weighting factors is adjusted to reduce similar false alarms.

[0157] Step S5 Summary: A two-tiered processing chain—multi-scale trend acceleration analysis and hierarchical early warning mapping—transforms the comprehensive risk indicator from static numerical values ​​to dynamic evolution curves. Based on dynamic thresholds, rate of change, and acceleration, it outputs real-time alerts at three levels: Attention, Warning, and Danger. Multi-scale rate and acceleration calculations ensure that both rapidly changing and slowly escalating risks can be captured; adaptive dynamic thresholds and weights guarantee the alerts remain sensitive even under long-term environmental drift; hierarchical mapping and closed-loop feedback make the alarm levels closer to real-world risks, and threshold configurations are continuously optimized through maintenance feedback. The entire process is strictly integrated with the outputs of the previous four steps, maintaining consistency in all terminology and parameter names, forming a complete closed-loop technology chain from signal acquisition to hierarchical alarms. This fully demonstrates the invention's creativity and engineering feasibility in distribution box fire early warning scenarios.

[0158] In conclusion, the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A fire early warning method for distribution boxes based on multi-source information fusion, characterized in that, include: Step S1: Collect temperature, smoke, acoustic vibration, and current harmonic signals, perform timing synchronization and amplitude normalization, correct sampling time difference and amplitude difference, and generate dynamic baseline; Step S2: Based on the dynamic baseline, the normalized signal is windowed according to a fixed duration, multi-scale statistical and morphological features are extracted, and unsupervised clustering is used to divide the operation mode to construct a feature vector set; Step S3: The baseline signature matching the operation mode is retrieved from the historical feature library, the multimodal deviation of the current window features from the baseline signature is calculated, and the standardized deviation index is output for subsequent coupling calculation; Step S4: The deviation index is mapped to nodes, edge weights are set according to covariance, phase synchronization index and smoke dispersion index are introduced, the coupling risk coefficient is obtained through the fusion model and the edge weights are updated to generate a comprehensive risk index; Step S5: Multi-scale rate and acceleration analysis is performed on the comprehensive risk index, the risk level is determined according to the dynamic threshold, attention, warning or danger warning is triggered, and the operation and maintenance feedback adaptive correction threshold is recorded.

2. The method for early warning of fires in distribution boxes based on multi-source information fusion according to claim 1, characterized in that, Step S1 includes: Step 101: Compensating for the delay of each sensing node through Kalman prediction and combining it with the dominant frequency phase-locked micro-correction to achieve microsecond-level timing synchronization of multi-source signals; Step 102: Normalizing the synchronized signal using quantile mapping and generating a dynamic baseline by convolving a local Gaussian kernel and a global exponential decay kernel to suppress slow environmental drift; Step 103: Using synchronization confidence to drive Bayesian backtracking smooth reconstruction of missing timestamps and outputting a continuous corrected time series; Step 104: Performing dual-threshold evaluation of stability and instantaneous drift rate on the dynamic baseline, confirming the baseline as valid only when the threshold is met and outputting synchronization confidence and baseline stability.

3. The method for early warning of fires in distribution boxes based on multi-source information fusion according to claim 1, characterized in that, Step S2 includes: Step 201: Adjusting the window duration and overlap rate in real time based on the synchronization confidence; widening the window and reducing the overlap ratio when the synchronization confidence decreases; Step 202: Extracting three types of statistical features—amplitude, spectrum, and morphology—within each window, and forming a baseline relativized feature vector relative to the dynamic baseline; Step 203: Performing unsupervised clustering on the baseline relativized feature vector using a density-adjusting kernel clustering algorithm, generating running mode labels, and writing them into the window index table.

4. The method for early warning of fires in distribution boxes based on multi-source information fusion according to claim 3, characterized in that, Step S2 includes: inputting the baseline relativized feature vector into the density self-adjusting kernel clustering algorithm to adaptively determine the cluster center; using the neighborhood consistency coefficient multidimensional reassignment to correct noise and generate stable clusters; constructing cluster fingerprints using the skewness, kurtosis and modulus of the cluster center features and hashing them into globally unique cluster labels; and writing the cluster labels, along with the window timestamp and cluster stability, into the running mode metadata database for quick retrieval of baseline signatures in step S3.

5. The method for early warning of fires in distribution boxes based on multi-source information fusion according to claim 1, characterized in that, Step S3 includes: Step 301: Retrieve candidate baseline signatures in the historical feature library according to the operating mode label, and calculate the confidence level based on stability index, dynamic baseline consistency and feature similarity, and select the target baseline signature; Step 302: Compare the current window features with the target baseline signature, calculate the four types of standardized deviation indexes of amplitude, spectrum, phase and dynamic baseline and write them into the deviation buffer.

6. The method for early warning of fires in distribution boxes based on multi-source information fusion according to claim 1, characterized in that, Step S4 includes: Step 401: Mapping the standardized deviation index to temperature deviation nodes, smoke deviation nodes, acoustic deviation nodes and current deviation nodes, assigning the edge weights between nodes according to the historical covariance and Mahalanobis distance and writing them into the edge weight matrix; Step 402: Using the phase synchronization index to enhance the edge weights of the temperature deviation nodes and current deviation nodes, using the smoke dispersion index to implement a suppression and re-enhancement strategy for the relevant edge weights of the smoke deviation nodes, and updating the edge weights by outputting the coupling risk coefficient through the fusion model to generate a comprehensive risk index.

7. The method for early warning of fires in distribution boxes based on multi-source information fusion according to claim 6, characterized in that, The method for obtaining the phase synchronization index is as follows: Within the correction time series Tc(t), perform Hilbert transform on the temperature-normalized signal Ŝtemperature(t) and the current harmonic-normalized signal Ŝelectricity(t) respectively to obtain the instantaneous phase sequences φtemperature(n) and φelectricity(n). Calculate the sample-by-sample phase difference φΔ(n), and then obtain the phase synchronization index PSI according to the following formula: Where j represents the sequential number of the sampling point, and the total number of samples N corresponds to the number of sampling points in the current time window; the amplitude is obtained after summing the complex exponents and normalized, and the phase synchronization index PSI takes a value range of 0–1. The closer the value is to 1, the higher the phase lock between the temperature and current harmonic dominant frequencies, revealing the abnormal strength of electrothermal coupling.

8. The method for early warning of fires in distribution boxes based on multi-source information fusion according to claim 6, characterized in that, The method for obtaining the smoke dispersion index is as follows: taking the normalized smoke signal Ŝsmoke(t) in the corrected time series Tc(t) as input, the peak sequence p_i is extracted using an adaptive threshold method, where i = 1…M; the probability of each peak amplitude relative to the total peak amplitude is calculated. The smoke dispersion index (SDI) is calculated using the following formula: Among them, the peak probability q_i is dimensionless, the number of peaks M is the number of peaks within the window, i represents the sequential number of peaks, and the denominator lnM normalizes the information entropy; the smoke dispersion index SDI ranges from 0 to 1. The higher the value, the more dispersed the peak amplitude distribution of the smoke signal, indicating that the smoke concentration fluctuates greatly and is unstable, and is more likely to correspond to early combustion or a surge in environmental dust.

9. The method for early warning of fires in distribution boxes based on multi-source information fusion according to claim 1, characterized in that, Step S5 includes: Step 501: Obtain the short-term rate by short-window differential and the long-term rate by long-window linear regression slope for the comprehensive risk index, and obtain the change acceleration curve by low-pass filtering differential; Step 502: Map the comprehensive risk index, short-term rate, long-term rate and change acceleration to three levels of early warning: attention, warning and danger, based on the dynamic threshold, and adaptively correct the dynamic threshold and weight factor using operation and maintenance feedback.

10. A fire early warning system for distribution boxes based on multi-source information fusion, characterized in that, include: Synchronization and normalization module: Collects temperature, smoke, acoustic vibration, and current harmonic signals, performs timing synchronization and amplitude normalization, corrects sampling time difference and amplitude difference, and generates dynamic baseline; Feature clustering module: Based on the dynamic baseline, the normalized signal is windowed at a fixed duration, multi-scale statistical and morphological features are extracted, and unsupervised clustering is used to divide the operation mode in order to construct a feature vector set; Baseline matching module: retrieves baseline signatures that match the operating mode from the historical feature database, calculates the multimodal deviation of the current window features from the baseline signature, and outputs the standardized deviation index for subsequent coupling calculation; Fusion inference module: maps the deviation index to nodes, sets edge weights according to covariance, introduces phase synchronization index and smoke dispersion index, obtains the coupling risk coefficient through the fusion model and updates the edge weights, and generates a comprehensive risk index. Trend warning module: Performs multi-scale rate and acceleration analysis on comprehensive risk indicators, determines the risk level based on dynamic thresholds, triggers attention, warning or danger warnings, and records operation and maintenance feedback to adaptively correct the thresholds.

Images