An intelligent electrical fire monitoring method and system for buildings

Abstract

The present application discloses a method and system for intelligent electrical fire monitoring in buildings, which relates to the field of fire monitoring and includes: obtaining the distribution box parameters of different floors in a target monitoring area within a preset time period, where the distribution box parameters include leakage current values and temperature values; performing multi-scale decomposition on the distribution box parameters to obtain a plurality of intrinsic mode functions and residual trends; extracting features from each intrinsic mode function and residual trend to obtain an electrical fire feature vector; constructing a fire risk assessment model according to the electrical fire feature vector; and predicting the fire risk level according to the fire risk assessment model. Aiming at the nonlinear relationship between the leakage current and temperature of the distribution box ignored in the prior art, the present application improves the accuracy of building electrical fire monitoring through multi-variable time-delay analysis and ensemble empirical mode decomposition, etc.

Description

Technical Field

[0001] This application relates to the field of fire monitoring, and particularly to a method and system for intelligent electrical fire monitoring in buildings. Background Art

[0002] Traditional methods for electrical fire monitoring in buildings mainly rely on regular manual inspections and simple threshold alarm systems. Such monitoring means are lagging and passive. With the development of Internet of Things and artificial intelligence technologies, a new generation of electrical fire monitoring systems has begun to use intelligent sensing networks to collect electrical parameters in real time, but there are still technical bottlenecks in data processing and risk assessment, especially the limited ability to extract features and perform intelligent recognition on complex electrical signals.

[0003] On the one hand, traditional monitoring systems usually independently monitor and analyze parameters such as leakage current and temperature, ignoring the non-linear coupling relationship between these parameters. In fact, before an electrical device fails, there is often a complex time delay and mutual promotion relationship between abnormal leakage current and temperature. This non-linear relationship is a key indicator for early warning of electrical fires. On the other hand, existing monitoring systems generally use traditional signal processing methods such as Fourier transform or wavelet analysis to perform frequency domain analysis on electrical parameters. These methods have obvious limitations in dealing with non-stationary and non-linear electrical parameters and are difficult to accurately extract fault features. In particular, although the traditional empirical mode decomposition (EMD) technology can be used for non-stationary signal analysis, it is prone to mode mixing, resulting in the electrical fire characteristics being wrongly dispersed into different mode functions, reducing the accuracy of fault identification.

[0004] For example, the related patent CN118781734A discloses a method for intelligent electrical fire monitoring in buildings, including: obtaining the leakage current values and temperature values of distribution boxes on different floors in a target monitoring area at multiple predetermined time points within a predetermined time period; arranging the leakage current values and temperature values of distribution boxes on different floors at multiple predetermined time points in the time dimension and sample dimension to obtain a time series input tensor of electrical fire monitoring parameters; performing feature encoding on the time series input tensor of electrical fire monitoring parameters to obtain a time series feature map of dynamic changes in electrical fire monitoring parameters; performing feature emphasis on the time series feature map of dynamic changes in electrical fire monitoring parameters to obtain an emphasized feature map of electrical fire monitoring; and obtaining an alarm result based on the emphasized feature map of electrical fire monitoring. However, this solution arranges the leakage current values and temperature values in the time dimension and sample dimension, ignoring the non-linear time series correlation between the two. Therefore, the monitoring accuracy of this solution needs to be further improved. Summary of the Invention

[0005] In view of the neglect of the non-linear relationship between the leakage current and temperature in the distribution box in the prior art, the present application provides a building intelligent electrical fire monitoring method and system, which improves the accuracy of building electrical fire monitoring by means of multi-variable time-delay analysis and ensemble empirical mode decomposition, etc.

[0006] One aspect of the present application provides a building intelligent electrical fire monitoring method, including: S1, obtaining the distribution box parameters of different floors in the target monitoring area within a preset time period, where the distribution box parameters include leakage current values and temperature values; S2, performing multi-scale decomposition on the distribution box parameters to obtain a plurality of intrinsic mode functions and residual trends; S3, extracting features from each intrinsic mode function and residual trend to obtain an electrical fire feature vector; S4, constructing a fire risk assessment model according to the electrical fire feature vector; S5, predicting the fire risk level according to the fire risk assessment model.

[0007] Further, the intrinsic mode function represents different frequency oscillation components and corresponding time-varying characteristics in the distribution box parameters; the residual trend represents the long-term change baseline and overall evolution trend of the distribution box parameters.

[0008] Further, S2, performing multi-scale decomposition on the distribution box parameters to obtain a plurality of intrinsic mode functions and residual trends, includes: S21, constructing a multi-variable time series matrix of the leakage current value and temperature value, and calculating the time-delay cross-correlation coefficient of the leakage current value and temperature value; S22, determining the coupling time window between the leakage current value and temperature value according to the peak value of the time-delay cross-correlation coefficient; according to the coupling time window, combining the leakage current value and temperature value into a two-dimensional vector signal; S23, performing ensemble empirical mode decomposition, generating a plurality of noise-assisted signals by adding white noise with different amplitudes to the two-dimensional vector signal multiple times, and each noise-assisted signal is used as a set member; by adding white noise with different amplitudes and constructing set members, the mode mixing problem in traditional empirical mode decomposition is effectively solved, the stability of signal decomposition is improved, so that the extracted intrinsic mode function can more accurately reflect different frequency oscillation components of the distribution box parameters, and the detection sensitivity of high-frequency weak leakage anomalies is improved.

[0009] S24, for each set member, screening the corresponding local extreme points, and constructing upper and lower envelope lines according to the local extreme points; calculating the mean value of the upper and lower envelope lines of each set member, subtracting the corresponding mean value from the set member to obtain a candidate component; S25, calculating the orthogonality index between the candidate component and the corresponding two-dimensional vector signal; obtaining the candidate component with the orthogonality index lower than the preset orthogonality index threshold as the first intrinsic mode function; S26, subtracting the extracted intrinsic mode function from the two-dimensional vector signal obtained in S22, and repeating S23 to S25 for the remaining signal to sequentially extract the next intrinsic mode function; S27, repeating S26 until the remaining signal is a monotonic function, and taking the monotonic function as the residual trend.

[0010] Further, in S21, a multivariate time series matrix of leakage current values and temperature values is constructed, and the time-delay cross-correlation coefficient of the leakage current values and temperature values is calculated, including: obtaining the leakage current data series and temperature data series of different floors within a preset time period; aligning the leakage current data series and temperature data series according to time tags to construct a two-dimensional time series matrix; setting a maximum time-delay range according to the acquisition frequency of the leakage current of the distribution box, where the maximum time-delay range reflects the delay relationship between temperature anomalies and leakage current in the distribution box; within the set maximum time-delay range, moving the temperature data series in sequence with a preset step size to form data pairs under different time-delay conditions; calculating the mutual information coefficient between the data pairs under different time-delay conditions as the time-delay cross-correlation coefficient of the leakage current values and temperature values.

[0011] Further, in S22, the coupling time window between the leakage current value and the temperature value is determined according to the peak value of the time-delay cross-correlation coefficient; according to the coupling time window, the leakage current value and the temperature value are combined into a two-dimensional vector signal, including: constructing a cross-correlation coefficient curve according to the time-delay cross-correlation coefficient; extracting the time-delay value corresponding to the peak value of the cross-correlation coefficient curve. ; taking the time-delay value as the center, obtaining a continuous time-delay interval where the time-delay cross-correlation coefficient is greater than a preset threshold as the coupling time window between the leakage current value and the temperature value; according to the coupling time window, shifting the leakage current data series relative to the temperature data series by time step lengths to obtain the leakage current data series and temperature data series after time alignment; taking the leakage current data series after time alignment as the first-dimensional component and the temperature data series as the second-dimensional component to form a two-dimensional vector signal by combination. By calculating the time-delay cross-correlation coefficient between the leakage current and the temperature value, the non-linear coupling relationship and delay characteristics between the two can be accurately identified, which is more in line with the temperature-leakage physical change law during the failure of electrical components than traditional methods.

[0012] Further, in S3, feature extraction is performed on each intrinsic mode function and the residual trend to obtain an electrical fire feature vector, including: S31, performing Hilbert-Huang transform on each intrinsic mode function to obtain the frequency-domain features of each intrinsic mode function; S32, performing piecewise fitting on the residual trend to obtain trend features; S33, combining the frequency-domain features and trend features to obtain an electrical fire feature vector.

[0013] Further, in S31, perform Hilbert-Huang transform on each intrinsic mode function to obtain the frequency-domain characteristics of each intrinsic mode function, including: extending both ends of each intrinsic mode function to N times the length of the original signal using the mirror symmetry extension method; performing Hilbert transform on the extended intrinsic mode functions to calculate the corresponding analytic signals; extracting the instantaneous amplitude of each intrinsic mode function according to the analytic signals. And the instantaneous phase; calculate the derivative of the instantaneous phase as the instantaneous frequency of each intrinsic mode function, delete the instantaneous frequency corresponding to the extended part, and only retain the instantaneous frequency within the original signal interval. According to the retained instantaneous frequency And the corresponding instantaneous amplitude , construct the Hilbert energy spectrum of each intrinsic mode function. According to the Hilbert energy spectrum , calculate the frequency-domain characteristics of each intrinsic mode function.

[0014] Further, the value range of N is from 1.2 to 1.6.

[0015] Further, in S4, construct a fire risk assessment model according to the electrical fire feature vectors, including: collecting a sample set of electrical fire feature vectors corresponding to different building electrical fire levels; using the electrical fire feature vectors in the sample set as inputs and the real fire levels as outputs to construct a fire risk assessment model based on random forest.

[0016] This application also provides a building intelligent electrical fire monitoring system, including: a data acquisition module that acquires the distribution box parameters of different floors in the target monitoring area within a preset time period, and the distribution box parameters include leakage current values and temperature values; a signal decomposition module that performs multi-scale decomposition on the distribution box parameters to obtain multiple intrinsic mode functions and a residual trend; a feature extraction module that extracts features from each intrinsic mode function and the residual trend to obtain electrical fire feature vectors; a risk prediction module that predicts the fire risk level of the electrical fire feature vectors according to the fire risk assessment model.

[0017] Compared with the prior art, the advantages of this application are as follows:

[0018] Traditional empirical mode decomposition (EMD) is prone to mode mixing, resulting in the electrical fire characteristics being wrongly dispersed into different IMFs, reducing the accuracy of fault identification. Moreover, there is often a temporal correlation between the leakage current and temperature anomalies, but in the existing technologies, the two are often independently decomposed separately, ignoring the mutual correlation. In this application, by calculating the time-delay cross-correlation coefficient between the leakage current and temperature values, the coupling time window between the two is identified, and the two are combined into a two-dimensional vector signal for joint analysis; the ensemble empirical mode decomposition technique is adopted, and ensemble members are constructed by adding multiple groups of white noise with different amplitudes to the signal, effectively overcoming the mode mixing problem; combined with the Hilbert-Huang transform and piecewise fitting technique, the multi-scale characteristics of the leakage current signal are comprehensively extracted, greatly improving the accuracy and foresight of the electrical fire risk assessment. Description of the Drawings

[0019] This application will be further described in the form of exemplary embodiments, which will be described in detail through the drawings. These embodiments are not restrictive. In these embodiments, the same numbers represent the same structures, where:

[0020] Figure 1 is an exemplary flowchart of a building intelligent electrical fire monitoring method shown in some embodiments of this application;

[0021] Figure 2 is an exemplary flowchart of the method for extracting the first intrinsic mode function shown in some embodiments of this application;

[0022] Figure 3 is an exemplary flowchart of constructing an electrical fire feature vector shown in some embodiments of this application. Detailed Embodiments

[0023] The methods and systems provided in the embodiments of this application will be described in detail below with reference to the drawings.

[0024] As Figure 1 shown, obtain the distribution box parameters of different floors in the target monitoring area within a preset time period. The distribution box parameters include leakage current values and temperature values; perform multi-scale decomposition on the distribution box parameters to obtain multiple intrinsic mode functions and residual trends; extract features from each intrinsic mode function and residual trend to obtain an electrical fire feature vector; construct a fire risk assessment model according to the electrical fire feature vector; predict the fire risk level according to the fire risk assessment model.

[0025] S1. First, deploy an intelligent sensor network in the distribution boxes on each floor of the target building to achieve automatic collection of distribution box parameters. The specific implementation method is as follows: Install leakage current transformers and temperature sensors inside the distribution boxes on each floor. The leakage current transformers adopt a magnetic core ring design with a sensitivity of not less than 0.5 mA; the temperature sensors use thermocouple arrays, and multiple collection points are set at key positions inside the distribution boxes to ensure the comprehensiveness of temperature field distribution monitoring. According to the development characteristics of electrical fires, the leakage current sampling frequency is set to 10 Hz - 100 Hz to capture transient leakage changes; the temperature data sampling frequency is set to 0.1 Hz - 1 Hz to meet the monitoring requirements of heat changes. Standardize the collected raw data, including removing power interference, electromagnetic radiation interference, and sensor self-noise, and filling in missing values to ensure data quality. The collected data is stored in the format of "floor number - distribution box number - parameter type - timestamp - parameter value" to form a standardized time series dataset for subsequent time series analysis.

[0026] As Figure 2 shown, S2. Decompose the distribution box parameters at multiple scales to obtain multiple intrinsic mode functions and a residual trend; the intrinsic mode functions represent different frequency oscillation components and corresponding time-varying characteristics in the distribution box parameters; the residual trend represents the long-term change baseline and overall evolution trend of the distribution box parameters.

[0027] S21. Construct a multivariate time series matrix and calculate the time-delay cross-correlation coefficient. Extract the leakage current and temperature data sequences within a specific time period (usually 1 - 7 days) from the database, and keep the sampling time consistency of the two types of data. For the case where the sampling frequencies are inconsistent, align the two types of data in the time dimension through interpolation. Denote the aligned leakage current sequence as and the temperature sequence as to construct a two-dimensional matrix , where represents the two-dimensional time series matrix; t represents the time point. According to the characteristics of electrical faults, set the maximum time-delay range to 24 hours, which covers the time delay from the occurrence of leakage to the obvious change in temperature for most electrical faults. Within the set time-delay range, move the temperature sequence backward in steps of 10 minutes to obtain , where τ represents the time-delay value. For each τ, calculate the mutual information between and : , where is the joint probability density, and are the marginal probability densities, which are calculated by the kernel density estimation method.

[0028] S22. Determine the coupling time window between the leakage current value and the temperature value according to the peak value of the time-delay cross-correlation coefficient; according to the coupling time window, combine the leakage current value and the temperature value into a two-dimensional vector signal. Specifically, organize the different time-delay values τ (usually in the range of 0 - 24 hours) calculated in S21 and their corresponding mutual information coefficients into a corresponding relationship table. Use the time-delay value τ as the abscissa and the mutual information coefficient as the ordinate to plot the cross-correlation coefficient curve. To ensure the smoothness of the curve, use the cubic spline interpolation method to perform continuous interpolation between discrete sampling points to form a smooth cross-correlation coefficient curve. To eliminate possible high-frequency noise interference, apply the Savitzky-Golay filter to the original cross-correlation coefficient curve, set the window width to 7 - 11 sampling points, and set the polynomial order to 3 to reduce the noise impact while retaining the main features of the curve

[0029] Use the multi-scale peak detection algorithm to identify local maxima at different time scales. First, locate the main peak area at a coarser scale, and then accurately locate the peak position at a finer scale. For the detected peak points, verify by calculating the second derivative of the curve to ensure that the second derivative at the peak point is negative and its absolute value is greater than a preset threshold (such as 0.05). From the verified peak points, select the point with the largest mutual information coefficient, and its corresponding time-delay value is . For the case of multiple peaks, if the difference between the highest peak and the second-highest peak is less than 20%, then perform a weighted average (the weights are proportional to the peak values) of the time-delay values of the two peaks as. In practical applications, the time-delay in the electrical fire precursor is usually in the range of 6 - 12 hours

[0030] Set the preset threshold to 70% of the peak mutual information coefficient, that is . This threshold can be adjusted according to the characteristics of the specific building electrical system. It can be set to 60% in high-sensitivity scenarios and 80% in high-specificity scenarios. Starting from , search in both the left and right directions respectively to find the time-delay points where the mutual information coefficient is first lower than the preset threshold, and record them as and . The interval is the coupling time window between the leakage current and the temperature. If the identified window is too wide (>12 hours), then use the adaptive threshold method to increase the threshold to 75% - 85% and recalculate the window boundaries to ensure that the captured coupling relationship has sufficient specificity

[0031] According to the coupling time window , determine the effective data analysis time range and intercept the data sequences of the two parameters within this time range. Shift the leakage current data sequence to the right by a time unit, i.e., This operation aligns the leakage current signal with the temperature signal in time, directly reflecting the physical causal relationship in the data sequence. The offset operation may result in invalid data at the sequence boundary. The "mirror extension" method is used to handle the boundary problem, that is, data points symmetric to the boundary points are added at both ends of the sequence to maintain the continuity and integrity of the sequence.

[0032] For the aligned leakage current sequence and the temperature sequence perform normalization to unify their scales. The normalization adopts the Z-score standardization method: , where X represents the original data point; Z represents the normalized data point; μ is the sequence mean, and σ is the sequence standard deviation. The normalized leakage current sequence is used as the first-dimensional component, and the temperature sequence is used as the second-dimensional component to construct a two-dimensional vector signal , where represents the normalized leakage current sequence; represents the normalized temperature sequence; By calculating the autocorrelation function of each component of the two-dimensional vector, it is ensured that the constructed vector signal retains the main time characteristics of the original data. At the same time, by calculating the cross-correlation function between the two components, the alignment effect is verified. This application takes into account the time-delay characteristic between the abnormal leakage current and the temperature rise during the fault development process of electrical components, which conforms to the physical law of the development of electrical fires. Compared with the traditional method of analyzing each parameter separately, it can better reflect the real fault evolution process.

[0033] S23. Perform ensemble empirical mode decomposition. The white noise generation uses a high-quality pseudo-random number generator to ensure that the noise sequence has a uniform spectral distribution. The noise amplitude setting adopts an adaptive method and is dynamically adjusted based on the signal strength: Let the white noise amplitude be α×σ, where σ is the standard deviation of the original two-dimensional vector signal, and α is the proportionality coefficient. For high signal-to-noise ratio scenarios, the value range of α is from 0.1 to 0.2; for low signal-to-noise ratio scenarios, the value range of α is from 0.2 to 0.3. To obtain a more comprehensive noise coverage, a linearly increasing sequence of α values is designed, increasing from 0.1 to 0.3, with a total of 10 to 20 levels.

[0034] For each noise amplitude level, generate 5 to 10 groups of independent white noise sequences, forming a total of 50 to 200 noise-assisted ensemble members. For the two-dimensional vector signal , where represents the processed leakage current sequence; Add white noise to both components respectively: ; where represents the kth noise-assisted ensemble member; and The noise sequence for the k-th set member, the two are independent but have the same amplitude. A parallel computing architecture is used to process multiple set members to improve the decomposition efficiency. According to the number of server cores, the set members are grouped, and each group is assigned to a computing core for simultaneous processing.

[0035] In this application, white noise with different amplitudes is added to introduce perturbations in different frequency intervals of the signal, prompting components with close frequencies to be separated into different IMFs. After the white noise is introduced, the components with close frequencies that might have been mismerged originally are locally modulated by the noise during the decomposition process, enhancing the difference in their extreme point distributions. By adding random white noise multiple times and averaging, the randomness of the noise is canceled out during the final averaging process, while the different frequency components separated by the noise can be retained in the corresponding IMFs.

[0036] S24, using a multi-scale local extreme value detection algorithm for each set member for processing: First, detect local maxima and minima at the original sampling scale, that is, the points satisfying are the maximum points, and the points satisfying are the minimum points. Anti-interference processing is adopted: If the number of sampling points between two adjacent maximum (or minimum) points is less than 3, then compare the amplitudes of these two maximum (or minimum) points, and retain the one with the larger (or smaller) amplitude. For high-sampling-rate data (>50Hz), sliding window pre-smoothing processing is adopted, and the window width is odd, usually 5 to 11 sampling points, to reduce excessive extreme points caused by noise.

[0037] Use cubic spline interpolation to construct the upper and lower envelopes to ensure smooth transition of the envelopes in the entire time domain: The upper envelope is constructed through all local maximum points. The lower envelope is constructed through all local minimum points. Boundary processing adopts the mirror extension method: Add 3 - 5 mirror extreme points at both ends of the signal to ensure the smoothness and accuracy of the envelopes in the edge region. Spline interpolation adopts the "not-a-knot condition" to ensure smooth transition of the envelopes between extreme points and avoid excessive oscillation.

[0038] Calculate the point-by-point mean of the upper and lower envelopes: , where represents the mean of the upper and lower envelopes of the k-th set member; Subtract the mean from the currently processed signal to obtain the candidate component: ; Set the upper limit of the number of iterations to usually 10 times to prevent over-iteration. If the IMF condition is still not met after 10 iterations, forcefully accept the current candidate component. Verify the candidate component Does it meet the following IMF conditions: Condition 1: The absolute value of the difference between the total number of extreme points and the total number of zero crossings does not exceed 1. Condition 2: Within the entire data range, the relative deviation SD of the envelope mean from zero is lower than the threshold: ; If the above conditions are not met, then is regarded as the new input signal, and the screening process is repeated.

[0039] S25, the orthogonality index (OI) is used to measure the orthogonality between the candidate component and the original signal: ; where is the candidate component obtained in the current iteration, is the k-th set member. The range of the orthogonality index is [0, 1], and the smaller the value, the better the orthogonality. The orthogonality index threshold is dynamically adjusted based on the signal complexity, and the complexity is calculated by the spectral entropy of the signal: For low-complexity signals (spectral entropy < 0.6), the threshold is set to 0.05. For medium-complexity signals (spectral entropy 0.6 - 0.8), the threshold is set to 0.075. For high-complexity signals (spectral entropy > 0.8), the threshold is set to 0.1. Through experimental verification, the leakage current-temperature coupling signal of the distribution box usually belongs to medium-complexity signals.

[0040] For each set member, the candidate component with an orthogonality index lower than the preset threshold is obtained as the first IMF of this member: . When the candidate component of a certain set member still does not meet the orthogonality requirement after MAX_ITER iterations (usually set to 10), the current candidate component is forcibly adopted, and the reliability of this IMF is marked as low in the result.

[0041] The first IMFs obtained for all set members are averaged: , where N represents the total number of set members; represents the first intrinsic mode function extracted from the k-th set member; to reduce the influence of outliers, a modified weighted average method can be used: Calculate the deviation between each and its mean. The IMF with a deviation exceeding 3 times the standard deviation is given a lower weight (0.3), and the rest are given the standard weight (1.0). The weighted average is recalculated according to the weights.

[0042] S26, Subtract the extracted i-th IMF from the original two-dimensional vector signal to obtain the residual signal: ; where, represents the i-th intrinsic mode function, and represents the final IMF result obtained after ensemble averaging; is the original two-dimensional vector signal . Take as the new input signal; denote the remaining signal after the (i - 1)-th decomposition; repeat steps S23 to S25 to extract the next IMF. To improve the calculation efficiency, downsampling can be applied to the remaining signal: as the IMF serial number increases, the main frequency of the signal decreases, and the number of sampling points can be gradually reduced to lower the computational complexity. Anti-aliasing filtering is used for downsampling to ensure that no valid information is lost.

[0043] The recursive process terminates when any of the following conditions is met: Condition 1: The remaining signal becomes a monotonic function, that is, the total number of extreme points ≤ 2. Condition 2: The energy ratio of the remaining signal is lower than the threshold: , where denote the original two-dimensional vector signal, which is composed of the normalized leakage current sequence and temperature sequence. Condition 3: The number of extracted IMFs reaches the preset upper limit MAX_IMF (usually 8 to 10). Condition 4: The spectral correlation coefficient of two consecutive IMFs is higher than 0.85, indicating that the decomposition has tended to be redundant.

[0044] S27, repeat S26 until the remaining signal is a monotonic function, and use the monotonic function as the residual trend. The final remaining signal is used as the residual trend, representing the long-term change baseline of the distribution box parameters. To enhance the smoothness and physical meaning of the trend, a low-pass filter is applied to the residual trend, with the cut-off frequency set to 0.01 Hz to filter out the possible high-frequency components mixed in. For the two-dimensional vector signal, the residual trends of the two components are extracted separately to obtain the long-term evolution trends of the leakage current and temperature parameters of the distribution box.

[0045] For the distribution box monitoring application, different IMFs have clear physical meanings: To usually correspond to high-frequency noise and transient leakage characteristics (frequency range: 1 to 100 Hz), reflecting instantaneous states such as equipment switching and load fluctuations. To often correspond to intermediate-frequency oscillation components (frequency range: 0.1 - 1 Hz), reflecting gradual faults such as electrical component deterioration and poor contact. And the subsequent low-frequency components (frequency < 0.1 Hz) reflect long-term influencing factors such as insulation aging and environmental humidity. The residual trend reflects the equipment load baseline and seasonal change trends.

[0046] As Figure 3 shown, S3, feature extraction is performed on each intrinsic mode function and residual trend to obtain the electrical fire feature vector. S31, before performing the Hilbert transform, it is necessary to solve the distortion problem of the traditional Hilbert transform at the signal boundary. The specific implementation method is as follows: for each intrinsic mode function , at both ends of the time domain, the mirror symmetry extension method is used for extension. At the left end of the signal, mirror reflection about the point t = 0 is adopted to form ; At the right end of the signal, mirror reflection about the point t = T (T is the length of the original signal) is adopted to form .

[0047] The extension coefficient N is usually selected between 1.2 and 1.6, and the selection of the N value follows the following rules: for high-frequency IMFs (the first 1 - 3 IMFs), a smaller N value (1.2 - 1.3) is selected to reduce the computational burden; for medium-frequency IMFs (the 4th - 6th IMFs), a medium N value (1.3 - 1.5) is selected; for low-frequency IMFs (the 7th and subsequent IMFs), a larger N value (1.5 - 1.6) is selected to better handle the low-frequency boundary effect. To avoid discontinuity at the extension seam, at the connection between the original signal and the extended part, the Hanning window function is applied for smooth transition, and the window function length is 5% of the original signal length to ensure the smooth continuity of the entire extended signal.

[0048] Analytical signal calculation: Perform Hilbert transform on the extended to obtain the analytical signal: , where is the Hilbert transform of , and the calculation formula is: ; where j represents the imaginary unit; represents the integration variable.

[0049] Instantaneous amplitude calculation: The instantaneous amplitude is the modulus value of the analytical signal: Apply a moving median filter (window length is 5 - 7 sampling points) to the calculation result to eliminate possible numerical fluctuations.

[0050] Instantaneous phase calculation: The instantaneous phase is the argument of the analytical signal: To avoid phase jumps, the phase unwrapping algorithm is adopted to ensure continuous phase change.

[0051] Instantaneous frequency calculation: The instantaneous frequency is the time derivative of the phase: , and the derivative calculation adopts the five-point central difference formula to improve the numerical calculation accuracy: , where is the sampling interval.

[0052] Boundary data processing: Delete the instantaneous characteristic data corresponding to the extended part, and only retain the instantaneous amplitude and the instantaneous frequency within the original signal interval [0, T]. To avoid inaccuracies at the edge 2 to 3 points caused by differential calculation, the linear extrapolation method is used to reconstruct the values of these points.

[0053] Construct the Hilbert energy spectrum based on the retained instantaneous frequency and instantaneous amplitude , construct the Hilbert energy spectrum: , where δ is the Dirac function, indicating that the energy is concentrated only at the frequency at time t; represents the angular frequency variable; in practical applications, the time-frequency plane is discretized into a grid, and the grid resolution is: time resolution: the original sampling interval , frequency resolution: (maximum frequency - minimum frequency) / 100, usually 0.1 to 1 Hz. When the instantaneous frequency falls into the frequency interval , allocate proportionally to adjacent frequency points to achieve a smooth energy distribution.

[0054] Calculate the frequency-domain characteristic parameters. Marginal spectrum calculation: Integrate the Hilbert energy spectrum in the time dimension to obtain the marginal spectrum: In discrete implementation, use the trapezoidal integration method: .

[0055] Energy concentration calculation: The energy concentration EC is used to quantify the degree of aggregation of the frequency distribution: , the larger the EC value, the more concentrated the energy is at a specific frequency. For a normal distribution box, EC is usually 0.1 to 0.3, and it can reach 0.5 to 0.8 when abnormal. Among them, represents the marginal spectrum; represents the discrete frequency points.

[0056] Main frequency and bandwidth calculation: The main frequency point , is the frequency corresponding to the maximum value of the marginal spectrum. The frequency bandwidth BW is defined as the width of the smallest frequency interval containing 80% of the energy. The calculation steps are: Normalize to such that , sort from largest to smallest to get , find the smallest k such that , and calculate the maximum-minimum difference of these k frequency points as the bandwidth BW.

[0057] Extract the electrical fault characteristic parameters. Divide the frequency range into three frequency bands: low frequency (0 to 10 Hz), medium frequency (10 to 100 Hz), and high frequency (>100 Hz); calculate the proportion of the marginal spectrum energy of each frequency band in the total energy: . For a normal distribution box, the low-frequency energy usually dominates ( ), and when there is an electrical fault, the proportion of medium and high-frequency energy increases significantly. represents the proportion of energy in the low-frequency band (0 to 10 Hz); Represents the energy proportion in the medium frequency band (10 to 100 Hz); Represents the energy proportion in the high and low frequency bands (>100 Hz); Calculate the frequency change rate index, the time derivative of the instantaneous frequency: , calculated using the central difference method; Calculate The mean value of And variance ; Frequency change rate index: , characterizing the severity of frequency fluctuations. Calculate the frequency fluctuation index, calculate the instantaneous frequency The mean value of And standard deviation ; Frequency fluctuation index: ; Usually indicates the existence of abnormal fluctuations. Combine the above features into a frequency domain feature vector: , calculate this feature vector for each IMF, jointly constituting a full frequency domain feature matrix for subsequent fault type identification.

[0058] Specifically, the marginal spectrum : The marginal spectrum reflects the energy distribution of the leakage current signal at each frequency component. In a building electrical system, the leakage current during normal operation usually has relatively stable spectral characteristics, mainly concentrated on the power grid fundamental frequency (50 Hz or 60 Hz) and its harmonics. When pre-faults such as insulation aging and poor contact occur, abnormal energy increases will occur in specific frequency bands. For example, partial discharge before insulation breakdown will generate high-frequency noise; loose connections will generate medium-frequency fluctuations.

[0059] Energy concentration EC: Energy concentration quantifies the degree of aggregation of the frequency distribution. For electrical equipment operating normally, the leakage current spectrum energy is usually relatively concentrated at certain specific frequencies, and the EC value is relatively high. When electrical equipment begins to malfunction, the spectrum will become dispersed, manifested as a decrease in the EC value. For example, insulation aging will cause the spectrum energy to spread to a wider frequency band, generating more random noise.

[0060] Main frequency point And frequency bandwidth BW: The main frequency point is the frequency with the most concentrated energy, and the bandwidth reflects the frequency concentration degree of the signal. In a building electrical system, the main frequency under normal operating conditions is usually fixed and related to the operating characteristics of the equipment. When pre-fault symptoms appear, the main frequency may shift or the bandwidth may increase significantly. For example, motor bearing wear will cause the main frequency to shift; cable insulation aging will cause the frequency band to become wider.

[0061] Energy proportion of each frequency band :Divide the frequency spectrum into three intervals: low frequency (0 to 10 Hz), medium frequency (10 to 100 Hz), and high frequency (> 100 Hz), and calculate the energy proportion of each frequency band. In the building electrical system, different types of faults will show energy anomalies in different frequency bands: Low frequency (0 to 10 Hz): Corresponding to the slow change of the system load, an abnormal value may mean a continuous leakage of large current; Medium frequency (10 to 100 Hz): Includes the power grid fundamental frequency and low-order harmonics, and the anomaly is usually related to conductor connection problems or line faults; High frequency (> 100 Hz): Includes high-order harmonics and noise, and an abnormal value often reflects the pre-discharge before insulation breakdown or the precursor of an arc.

[0062] Frequency Variation Rate Index FVR: Measures the speed and degree of change of the instantaneous frequency over time. When the normal electrical system is operating, the change of the instantaneous frequency is usually slow and corresponds to the load change. When the system has intermittent faults or sudden load changes, the frequency variation rate will increase significantly. For example, when the conductor contact is poor, a frequency mutation will occur; When the equipment starts and stops, it will cause a rapid change in frequency.

[0063] Frequency Fluctuation Index FI: Measures the relative fluctuation degree of the frequency and reflects the frequency stability. A normally operating electrical system has a stable frequency characteristic and a low FI value. Early fault states such as poor contact and insulation aging will lead to frequency instability, manifested as an increase in the FI value. For example, a loose contact point will cause impedance fluctuations, which in turn cause frequency instability.

[0064] S32. Perform piecewise fitting on the residual trend to obtain trend features, including: Perform a sliding window analysis on the residual trend, and determine an appropriate window length according to the change characteristics of the leakage current signal; Apply piecewise linear fitting technology to model the residual trend, and determine the optimal piecewise point position and the slope of each segment by the least squares method; Extract key trend feature parameters, including: The slope and duration of each linear segment, reflecting the change rate of the leakage current; The slope mutation degree between adjacent linear segments, characterizing the degree of change of the leakage characteristics; The time position and amplitude of the key inflection point, indicating a significant change in the leakage mode; Combine the extracted parameters to form a trend feature vector , where k represents the slope, represents the slope change, t represents the inflection point time, v represents the inflection point amplitude, reflecting the long-term evolution characteristics of the distribution system. Through the sliding window analysis and piecewise linear fitting of the residual trend, the long-term evolution law of the leakage current is accurately captured. In particular, the extracted slope mutation degree and key inflection point features can reflect the slow accumulation process of the leakage current under insulation aging and humid environment.

[0065] S33. Combine the frequency domain feature vector of each IMF calculated in S31 with the residual trend feature vector calculated in S32. For n IMFs, form an original feature space of (8×n + 4n - 4) dimensions.

[0066] S4. Collect the building data records of the electrical fires that have occurred, and extract the complete sequence of distribution box parameters in the 72 hours before the fire; obtain the electrical fire accident investigation reports and parameter records from the fire safety department; establish a standardized electrical fire case database, including typical fault data of different types of buildings (commercial, residential, industrial). Simulate typical electrical fault scenarios in the laboratory environment, including insulation aging, line overload, poor contact, etc.; set different environmental conditions (temperature from 20 to 40 °C, relative humidity from 40 to 90%), and record the whole process of fault development; collect at least 50 sets of complete parameter evolution data for each fault type.

[0067] Apply the S1 to S3 processes to the original data to extract standardized electrical fire feature vectors; each sample contains 8×n-dimensional IMF frequency domain features and 4n - 4-dimensional residual trend features, where n is the number of IMFs (usually 6 to 10); perform normalization processing on all feature vectors so that the range of each dimensional feature value is in the interval [0, 1].

[0068] Establish a 5-level risk assessment system: Level 0 (safe): All parameters are within the normal range and there are no abnormalities; Level 1 (low risk): Slight abnormalities occur but are within the acceptable range; Level 2 (medium risk): Obvious abnormalities but not yet reaching the dangerous level; Level 3 (high risk): Seriously deviate from the normal value and require urgent intervention; Level 4 (extremely high risk): The system is in a critical state and a fire may be about to occur.

[0069] Historical fire case annotation: Determine the level according to the time from the occurrence of the fire (level 4 for 0 to 6 hours before the fire, level 3 for 6 to 24 hours before the fire, etc.); Experimental sample annotation: Determine the risk level according to the leakage current value, the severity of the fault, and expert evaluation.

[0070] Adopt a hierarchical random forest architecture, including a basic layer and an integration layer. The basic layer designs three dedicated random forest sub-models: High-frequency feature forest: Focus on the frequency domain features of IMF1 to IMF3; The frequency domain features of IMF1 to IMF3 are sensitive to sudden faults. Medium-frequency feature forest: Focus on the frequency domain features of IMF4 to IMF6; The frequency domain features of IMF4 to IMF6 are sensitive to gradual faults. Trend feature forest: Focus on the residual trend features; The integration layer designs a meta-random forest to integrate the output probabilities of the three sub-models.

[0071] Number of trees: 500 trees for each of the base-layer forests and 300 trees for the integrated-layer forest; Maximum tree depth: 25 - 30 layers, tuned through the validation set; Minimum number of samples per node: 5 samples for leaf nodes and 12 samples for internal nodes; Feature selection criterion: Gini impurity index; Control of tree diversity: Each tree uses the square root of the total number of features as the number of candidate features for splitting. Based on cost-sensitive analysis, an optimal decision threshold is determined for each risk level; A lower threshold (0.6 to 0.7) is adopted for the high-risk level (3 to 4) to improve recall; A higher threshold (0.8 to 0.9) is adopted for the low-risk level (0 - 1) to improve precision.

[0072] S5. Predict the fire risk level according to the fire risk assessment model. Obtain the real-time leakage current value and temperature value of the distribution box in the target monitoring area, process them according to the processes of S2 and S3 to obtain the real-time electrical fire feature vector; Input the real-time electrical fire feature vector into the random forest model constructed in S4 to calculate the fire risk probability value; Map the fire risk probability value to a preset risk level interval, and use the mapping result as the fire risk level. The risk level interval includes four levels: low risk, medium risk, high risk, and extremely high risk; When the fire risk level is high risk or extremely high risk, send a fire risk warning message to the building management system. The warning message includes the risk level, the determination result of the fault type, and the distribution box location information; Store the fire risk level assessment results in a time series manner to form an electrical fire risk evolution trend curve for long-term risk analysis and system optimization.

Claims

1. A building intelligent electrical fire monitoring method, characterized in that: include: S1, obtaining the distribution box parameters of different floors in the target monitoring area within a preset time period, wherein the distribution box parameters include leakage current value and temperature value; S2, multi-scale decomposition of distribution box parameters is performed to obtain multiple intrinsic mode functions and residual trends; S3, extracting features from each intrinsic mode function and residual trend to obtain electrical fire feature vectors; S4, constructing a fire risk assessment model based on the electrical fire feature vector; S5, predicting the fire risk level according to the fire risk assessment model; S2, multi-scale decomposition of distribution box parameters is performed to obtain multiple intrinsic mode functions and residual trends, including: S21, constructing a multivariate time series matrix of leakage current values ​​and temperature values, and calculating the time lag mutual correlation coefficient of the leakage current values ​​and temperature values, including: obtaining leakage current data sequences and temperature data sequences of different floors within a preset time period; aligning the leakage current data sequences and temperature data sequences according to time tags to construct a two-dimensional time series matrix; setting a maximum time lag range according to the leakage current acquisition frequency of the distribution box, wherein the maximum time lag range reflects the delay relationship between the temperature anomaly and the leakage current in the distribution box; within the set maximum time lag range, sequentially moving the temperature data sequence with a preset step length to form data pairs under different time lag conditions; calculating the mutual information coefficient between the data pairs under different time lag conditions as the time lag mutual correlation coefficient of the leakage current value and the temperature value; S22, determining a coupling time window between the leakage current value and the temperature value according to a peak value of a time-delay mutual correlation coefficient; combining the leakage current value and the temperature value into a two-dimensional vector signal according to the coupling time window; S23, performing ensemble empirical mode decomposition, generating multiple noise auxiliary signals by adding white noise of different amplitudes to the two-dimensional vector signal multiple times, each noise auxiliary signal being a member of a set; S24, for each set member, selecting the corresponding local extreme value point, and constructing the upper and lower envelopes according to the local extreme value points; Calculate the mean of the upper and lower envelopes of each set member, subtract the corresponding mean from the set member to obtain the candidate component; S25, calculating the orthogonality index between the candidate component and the corresponding two-dimensional vector signal; Obtaining a candidate component whose orthogonality index is lower than a preset orthogonality index threshold as the first eigenmode function; S26, subtracting the extracted intrinsic mode function from the two-dimensional vector signal obtained in S22, repeating S23 to S25 for the remaining signal, and extracting the next intrinsic mode function in turn; S27, repeat S26 until the remaining signal is a monotonic function, and take the monotonic function as the residual trend.

2. The intelligent building electrical fire monitoring method according to claim 1 is characterized in that: The intrinsic mode function represents the different frequency oscillation components in the distribution box parameters and the corresponding time-varying characteristics; The residual trend represents the long-term variation baseline and overall evolution trend of the distribution box parameters.

3. The intelligent building electrical fire monitoring method according to claim 1, characterized in that: S22, determining a coupling time window between the leakage current value and the temperature value according to the peak value of the time-delay mutual correlation coefficient; combining the leakage current value and the temperature value into a two-dimensional vector signal according to the coupling time window, including: According to the time-lag mutual correlation coefficient, a mutual correlation coefficient curve is constructed; Extract the lag value corresponding to the peak of the cross-correlation coefficient curve ; By time lag As the center, obtain the continuous time lag intervals whose time lag mutual correlation coefficient is greater than the preset threshold value as the coupling time window between the leakage current value and the temperature value; Shift the leakage current data sequence relative to the temperature data sequence according to the coupled time window time steps, and obtain the leakage current data series and temperature data series after time alignment; The time-aligned leakage current data sequence is used as the first-dimensional component, and the temperature data sequence is used as the second-dimensional component, and they are combined to form a two-dimensional vector signal.

4. The intelligent building electrical fire monitoring method according to claim 1, characterized in that: S3, extract the features of each intrinsic mode function and residual trend to obtain the electrical fire feature vector, including: S31, performing Hilbert-Huang transform on each intrinsic mode function to obtain frequency domain characteristics of each intrinsic mode function; S32, performing piecewise fitting on the residual trend to obtain trend characteristics; S33, combining the frequency domain features and the trend features to obtain an electrical fire feature vector.

5. The intelligent building electrical fire monitoring method according to claim 4 is characterized in that: S31, performing Hilbert-Huang transform on each intrinsic mode function to obtain frequency domain characteristics of each intrinsic mode function, including: The mirror symmetric extension method is used to extend the two ends of each intrinsic mode function to N times the length of the original signal; Perform Hilbert transform on each extended intrinsic mode function and calculate the corresponding analytical signal; According to the analytical signal, the instantaneous amplitude of each intrinsic mode function is extracted and instantaneous phase; Calculate the derivative of the instantaneous phase as the instantaneous frequency of each eigenmode function, delete the instantaneous frequency corresponding to the extended part, and only retain the instantaneous frequency within the original signal interval ; According to the instantaneous frequency retained and the corresponding instantaneous amplitude , construct the Hilbert energy spectrum of each eigenmode function ; According to the Hilbert energy spectrum , calculate the frequency domain characteristics of each eigenmode function.

6. The intelligent building electrical fire monitoring method according to claim 5, characterized in that: The value of N ranges from 1.2 to 1.

6.

7. The intelligent building electrical fire monitoring method according to claim 6, characterized in that: S4, based on the electrical fire feature vector, build a fire risk assessment model, including: Collect electrical fire feature vector sample sets corresponding to electrical fire levels in different buildings; The electrical fire feature vector in the sample set is taken as input and the actual fire level is taken as output to construct a fire risk assessment model based on random forest.

8. A building intelligent electrical fire monitoring system, characterized in that: include: The data acquisition module collects the distribution box parameters of different floors in the target monitoring area within a preset time period. The distribution box parameters include leakage current value and temperature value; The signal decomposition module performs multi-scale decomposition on the distribution box parameters to obtain multiple intrinsic mode functions and residual trends; The feature extraction module extracts the features of each intrinsic mode function and residual trend to obtain the electrical fire feature vector; The risk prediction module predicts the fire risk level of the electrical fire feature vector according to the fire risk assessment model; Among them, the distribution box parameters are decomposed at multiple scales, including: S21, constructing a multivariate time series matrix of leakage current values ​​and temperature values, and calculating the time-lag mutual correlation coefficient of the leakage current values ​​and temperature values, including: obtaining the leakage current data sequence and the temperature data sequence of different floors within a preset time period; aligning the leakage current data sequence and the temperature data sequence according to the time label to construct a two-dimensional time series matrix; setting a maximum time lag range according to the leakage current acquisition frequency of the distribution box, and the maximum time lag range reflects the delay relationship between the temperature anomaly and the leakage current in the distribution box; within the set maximum time lag range, sequentially moving the temperature data sequence with a preset step size to form data pairing under different time lag conditions; calculating the mutual information coefficient between the data pairs under different time lag conditions as the time-lag mutual correlation coefficient of the leakage current value and the temperature value; S22, determining a coupling time window between the leakage current value and the temperature value according to a peak value of a time-delay mutual correlation coefficient; combining the leakage current value and the temperature value into a two-dimensional vector signal according to the coupling time window; S23, performing ensemble empirical mode decomposition, generating multiple noise auxiliary signals by adding white noise of different amplitudes to the two-dimensional vector signal multiple times, each noise auxiliary signal being a member of a set; S24, for each set member, selecting the corresponding local extreme value point, and constructing the upper and lower envelopes according to the local extreme value points; Calculate the mean of the upper and lower envelopes of each set member, subtract the corresponding mean from the set member to obtain the candidate component; S25, calculating the orthogonality index between the candidate component and the corresponding two-dimensional vector signal; Obtaining a candidate component whose orthogonality index is lower than a preset orthogonality index threshold as the first eigenmode function; S26, subtracting the extracted intrinsic mode function from the two-dimensional vector signal obtained in S22, repeating S23 to S25 for the remaining signal, and extracting the next intrinsic mode function in turn; S27, repeat S26 until the remaining signal is a monotonic function, and take the monotonic function as the residual trend.

Images