Fault arc-caused fire early warning method based on multi-sensor fusion

Abstract

The invention relates to a fault arc-caused fire early warning method based on multi-sensor fusion, and belongs to the field of fault arc fire. The method comprises the following steps: establishing an arc electrical fire cause fault simulation test platform, selecting a typical single load and a combined load to perform a test before normal operation of a line and the occurrence of an arc cause fire, and establishing a fault arc waveform database and an arc fire database; the power-on detection method for the fault arc detection device is based on dictionary learning, current characteristics are represented through a sparse matrix, and fault classification is carried out in combination with an SVM. The fluctuation range of a plurality of characteristic quantities of the line in a normal state and when a fault arc occurs is analyzed, a threshold value is determined according to the periodic increase ratio of the characteristic quantities in different states, a final combined load detection threshold value is formed by complementation of a plurality of threshold values, and fault arc detection is realized; and based on the arc fire database, training a double-layer LSTM model under Bayesian optimization, and identifying multi-sensor signals in a time window to realize fire early warning.

Description

Technical Field

[0001] The present invention belongs to the field of fault arc fires, and particularly relates to a method for early warning of fault arc-caused fires based on multi-sensor fusion. Background Art

[0002] Since the invention of electric energy, it has penetrated into our daily life and become an indispensable and important part. It has greatly facilitated people's production and life, but at the same time, it is also accompanied by risks. Especially electrical safety hazards. Once a fire is triggered, the losses caused are often huge and immeasurable. Electrical fires are often accompanied by high temperatures, toxic smoke, and rapidly spreading fires, and sometimes even result in loss of life.

[0003] In many fire cases, fault arcs are important factors causing fires. The circuit may generate fault arcs due to reasons such as aging, poor contact, damaged insulation, and high-load operation. Traditional overload or short-circuit protection cannot detect fault arcs in the circuit. When a fault arc occurs, it is easy to cause local overheating, directly or indirectly leading to the occurrence of a fire.

[0004] Compared with a single-sensor monitoring system, a multi-sensor system can improve the accuracy of fault diagnosis and the timeliness of response by comprehensively analyzing data from different sources. Multi-sensor early warning of arc-caused fires has obvious advantages in the early warning and protection of arc-caused fires through earlier and more comprehensive evaluations. Summary of the Invention

[0005] The purpose of the present invention is to provide a method for early warning of fault arc-caused fires based on multi-sensor fusion, which can prevent the power-on misoperation of the fault arc detection device and achieve fault arc detection and fire early warning.

[0006] To achieve the above purpose, the technical solution of the present invention is: A method for early warning of fault arc-caused fires based on multi-sensor fusion, including:

[0007] By building a simulation test platform for fault arc-caused electrical fire causes, selecting typical single loads and combined loads to conduct tests during normal circuit operation and before the occurrence of arc-caused fires, and establishing a fault arc waveform database and an arc fire database;

[0008] Proposing a power-on detection method for a fault arc detection device based on dictionary learning, characterizing current features through a sparse matrix, and combining SVM for fault classification;

[0009] Analyzing the fluctuation ranges of multiple characteristic quantities of the circuit in the normal state and when a fault arc occurs, determining thresholds based on the periodic increase ratios of different state characteristic quantities, and using multiple thresholds to complement each other to form a final combined load detection threshold to achieve fault arc detection;

[0010] Based on the arc fire database, a double-layer LSTM model under Bayesian optimization is trained to identify multi-sensor signals within a time window to achieve fire warning.

[0011] In an embodiment of the present invention, the arc electrical fire causation fault simulation test platform includes a fault arc simulation generating device, the main circuit of the test platform, a data acquisition module, a control and status detection module, and a host computer.

[0012] In an embodiment of the present invention, the selected loads of the arc electrical fire causation fault simulation test platform include 3A, 6A, 13A, 20A pure resistive loads, and seven typical non-linear electrical equipment, including fluorescent lamps, halogen lamps, dimming lamps, switching power supplies, vacuum cleaners, air compressors, and electric hand tools; the main circuit of the test platform consists of multiple electrical circuits. The fault arc simulation generating device is switched to the set load fault point through the host computer and the control and status detection module, and at the same time, load sockets are reserved to select the load combination according to needs; the data acquisition module extracts the normal operation and fault arc current characteristic indexes of each electrical circuit, and the characteristic indexes include waveform indexes, kurtosis indexes, frequency characteristic ratios, and frequency centroids to construct a fault arc waveform database; multiple sensors in the data acquisition module are used to extract the data of the key positions of the line during normal operation and before the occurrence of arc-caused fires to construct an arc fire database.

[0013] In an embodiment of the present invention, the power-on detection method of the fault arc detection device can prevent the misoperation of the fault arc detection device during power-on.

[0014] In an embodiment of the present invention, the power-on detection method of the fault arc detection device constructs a sparse matrix to characterize the current waveform characteristics through dictionary learning based on the fault arc waveform database and combines SVM to achieve fault arc classification.

[0015] In an embodiment of the present invention, the power-on detection method of the fault arc detection device is specifically implemented as follows:

[0016] Integrate the fault arc waveform database into a current matrix X, the size of the matrix is m*n, where m represents the signal length of each current sample, and n represents the number of samples, including fault arc currents and normal current samples under different loads;

[0017] X≈DA

[0018] In the formula, D represents the dictionary matrix extracted from the original signal, and A represents the sparse coefficient matrix;

[0019] The MAP method is used for dictionary learning. The MAP method is based on the principle of maximum a posteriori estimation and updates the dictionary by minimizing the reconstruction error:

[0020] D(k+1) = D (k) + βR T RD (k)

[0021] R = X - D (k) A (k)

[0022] A (k) = D (k)-1 X

[0023] Wherein, D (k) represents the dictionary matrix of the current iteration, β represents the learning rate, R is the residual matrix, and A (k) represents the sparse matrix of the current iteration;

[0024] The final dictionary matrix D and sparse coefficient matrix A are obtained by optimizing the following objective function:

[0025]

[0026] Wherein, is the reconstruction error, used to measure the difference between the original signal and the reconstructed signal, λ is the sparsity regularization parameter, used to control the strength of sparsity, and |A||1 is the L1 norm, used to promote the sparsity of the sparse coefficient matrix;

[0027] The sparse coefficient vector A obtained by dictionary learning i is used as the input of the SVM, and the core information of the signal is extracted through sparse representation; The SVM classifies by finding a hyperplane that maximizes the margin between different classes, and the hyperplane constraint condition is satisfied:

[0028] w T A i + b = 0

[0029] Wherein, w represents the normal vector of the hyperplane, and b represents the bias term of the hyperplane;

[0030] The SVM is trained by optimizing the following objective function:

[0031]

[0032] Wherein, C represents the penalty parameter, which controls the penalty degree of the SVM for misclassification, and δ i represents the slack variable, allowing some samples to be misclassified on the decision boundary;

[0033] The classification decision function is:

[0034] f(x) = w T B i + b

[0035] Wherein, Bi Let \(i\) denote the time-window current signal collected in real time, and \(b\) denote the bias term of the hyperplane. If \(f(x)>0\), it is considered a positive-class sample and the current line is considered normal. At this time, the fault arc detection algorithm is entered. Otherwise, if \(f(x)<0\), it is considered a negative-class sample, and the current signal within the time window is resampled until a normal line signal appears.

[0036] In an embodiment of the present invention, the fault arc detection is specifically implemented as follows:

[0037] Perform fast Fourier decomposition on the current in the line when a fault arc occurs to obtain the harmonic contents of each order, and select four time-frequency domain characteristic indexes, namely the waveform index \(C\) f , the kurtosis index \(K\) v , the frequency characteristic ratio \(SR\) f , and the frequency centroid \(w\);

[0038] The threshold of a single characteristic parameter can be used as a judgment benchmark to identify fault arcs. However, the thresholds of each characteristic quantity are different under different loads. Relying solely on the threshold of a certain load as the threshold of the entire fault arc identification system will lead to misjudgment and missed judgment. Therefore, starting from the cycle increase ratio of the characteristic quantity, when using the cycle increase ratio of the characteristic quantity in different states as the judgment basis, a unified threshold is obtained to achieve fault arc detection for a relatively large number of load combinations. The expression of the characteristic quantity increase ratio \(\mu\) is:

[0039]

[0040] In the formula, \(x\) is the characteristic quantity calculated in the current cycle, \(x_0\) is the characteristic quantity calculated in the reference cycle, and \(x\) max is the larger value between \(x\) and \(x_0\);

[0041] Initially, current is collected with a two-cycle window, the characteristic quantity is extracted, and the cycle increase ratios of the waveform index, kurtosis index, frequency characteristic ratio, and frequency centroid are calculated respectively. To ensure that the reference value is the characteristic quantity extracted when the line is operating normally, the normal characteristic quantity accumulator threshold \(A\) th is equal to 3, that is, when three of the four cycle increase ratios of the characteristic quantities are less than the corresponding thresholds, the characteristic quantity extracted from the current in the latter cycle is determined as the reference value;

[0042] Subsequently, single-cycle current collection is started, and four characteristic quantities are extracted; after the characteristic quantity of the newly collected current is extracted, the cycle increase ratio with the characteristic quantity reference value is calculated. To prevent the cycle increase ratio of a certain characteristic quantity during normal operation from fluctuating beyond the threshold under different load combinations, the fault characteristic quantity accumulator threshold \(B\) thGreater than or equal to 3, that is, when there are three feature quantity period increase ratios greater than the corresponding fluctuation threshold values, it is considered that a faulty arc may occur. At this time, the faulty arc accumulator C is incremented by 1, and the normal current reference value is not updated; when the faulty feature quantity accumulator threshold B th is less than 3, if C is greater than 0, it is decremented by 1, and the previously collected feature quantity is updated to the normal feature quantity;

[0043] Finally, in order to prevent accidental mutation points caused by the influence of line interference signals, take C th equal to 3. When the faulty arc accumulator accumulates to its threshold C th it is determined that a faulty arc has occurred in the line.

[0044] In an embodiment of the present invention, the transformation expression of the fast Fourier transform is as follows:

[0045]

[0046] In the formula, n is the number of discrete points of the discrete signal sequence, where k is the frequency index, i is the current index, and X k is the k-th frequency component in the frequency domain, and x i is the n-th sampling point in the time domain;

[0047] The waveform index C f The expression is:

[0048]

[0049] In the formula, N is the number of sampling points in a single period, and i k is the k-th sampling point;

[0050] The kurtosis index K v The expression is:

[0051]

[0052] The frequency characteristic ratio SR f The expression is:

[0053]

[0054] In the formula, I j is the amplitude of the j-th harmonic;

[0055] The expression of the frequency centroid w is:

[0056]

[0057] In an embodiment of the present invention, the multi-sensor includes a temperature sensor, a carbon monoxide sensor, a smoke sensor, a humidity sensor, a flame sensor, a current sensor, a high-frequency noise sensor, and a sound sensor disposed at key positions of the line.

[0058] In an embodiment of the present invention, based on an arc fault fire database, a double-layer LSTM model under Bayesian optimization is trained to identify multi-sensor signals within a time window to achieve fire warning. The specific implementation method is as follows:

[0059] Integrate the arc fault fire database into a sample matrix m*n, where m is the number of samples and n is the specific number of sensors; perform normalization processing on the sample matrix to eliminate the dimension difference, and divide it into a training set and a test set for model training and evaluation; select 80% of the data as the training set and 20% of the data as the test set, and classify them, with normal being 0 and abnormal being 1;

[0060] Initialize the number of units, dropout rate, learning rate, and batch size of the double-layer LSTM model, construct a double-layer LSTM model, with a dropout layer following each layer of the LSTM network. The output result of the first layer of the LSTM network is processed by the first dropout layer and used as the input of the second layer of the LSTM network, and finally the result is output through a fully connected layer; the first layer of the LSTM extracts short-term features by gradually processing the time series of the original data; captures the data dependence patterns within a shorter time range, and these short-term features are passed to the second layer of the LSTM, which is responsible for identifying global dependencies within a longer time range and extracting more complex potential signals related to fire prediction;

[0061] Use a loss function to measure the gap between the model output and the true label. When the model prediction value is close to the true label, the loss function is small, otherwise it is large; during the training process of the model, the hyperparameters are adjusted by Bayesian optimization to minimize the loss function;

[0062] Loss function B expression

[0063]

[0064] where N is the total number of samples, y i is the classification label of the actual sample, is the probability that the sample is 1. In the classification task, is a probability value from 0 to 1, that is, the error between the training set and the test set, which is used to measure the training effect of the model;

[0065] Bayesian optimization uses a surrogate model to approximately calculate the minimization of the loss function of the double-layer LSTM model; this surrogate model predicts the value of the objective function at unevaluated points based on historical evaluation results and estimates the uncertainty of the prediction. By continuously modifying and iterating the number of units, dropout rate, learning rate, and batch size of the double-layer LSTM model, the minimum loss function is obtained while ensuring the iteration converges, and the corresponding hyperparameters are retained;

[0066] Calculate the accuracy, precision, recall, and F1 score under the final double-layer LSTM model, and set the lower limits of the minimum values of the accuracy and F1 score. If the minimum values of the accuracy and F1 score calculated by the final double-layer LSTM model do not meet the requirements, readjust the ratio of the training set and the test set. The adjustment range of the training set is 70 - 85%, and the adjustment range of the test set is exactly the opposite, which is 15 - 30%.

[0067] The multi-sensor uses the sampled data within twenty seconds as a time window, inputs it into the trained double-layer LSTM model optimized by Bayesian, outputs the hidden state vector of the second layer of LSTM, and then uses the second layer of LSTM model as the input of the fully connected layer to obtain the probability of a fire occurring.

[0068] The expression of the fire occurrence probability y is as follows:

[0069] y = f(Wh + b)

[0070] In the formula, f is the activation function, W is the weight matrix, h is the hidden state vector of the second layer of LSTM, b is the bias vector, and the weight matrix W and bias vector b of the fully connected layer are obtained through model training iteration. When it is set that y is greater than a certain threshold, it is considered that there is a fire risk and a warning is issued, and this threshold is defined by the on-site environment.

[0071] Compared with the prior art, the present invention has the following beneficial effects:

[0072] 1. A method for detecting power-on line faults of a faulty arc detection device based on dictionary query combined with SVM classification is proposed. Based on the multi-load faulty arc data set, this method uses dictionary query to obtain the dictionary matrix and sparse matrix to characterize the current characteristics, and combines SVM for sparse matrix feature classification to detect the power-on line of the faulty arc detection device, preventing the misoperation of the faulty arc detection device when powered on.

[0073] 2. A faulty arc detection algorithm with strong adaptability and high accuracy and multi-feature threshold complementarity is proposed. Through the faulty waveform characteristics under single load and combined load, this algorithm compares the fluctuation ranges of each characteristic quantity of the line in the normal state and when a faulty arc occurs, determines the threshold based on the periodic increase ratio of different state characteristic quantities, uses multiple thresholds to complement each other to form the final combined load detection threshold, and proposes the flow of the faulty arc detection algorithm.

[0074] 3. A multi-sensor fire warning method based on a double-layer LSTM model optimized by Bayesian is proposed. This method uses multi-sensors to collect data before the normal operation of the line and the occurrence of arc-caused fires to construct an arc fire database. The double-layer LSTM model under Bayesian optimization is trained through the arc fire database to identify multi-sensor signals within the time window and achieve fire warning. Description of the Drawings

[0075] Figure 1 This is the flowchart of the multi - feature fusion fault arc detection algorithm of the present invention.

[0076] Figure 2 This is the flowchart for training the fire warning of the double - layer LSTM model optimized by Bayesian of the present invention. Specific implementation manners

[0077] The technical solution of the present invention will be specifically described below with reference to the accompanying drawings.

[0078] The present invention provides a method for fire warning caused by fault arcs based on multi - sensor fusion, including:

[0079] By building a simulation test platform for the causes of arcing electrical fires, selecting typical single - load and combined - load for tests during normal line operation and before the occurrence of arcing - caused fires, a fault arc waveform database and an arc fire database are established;

[0080] A power - on detection method for a fault arc detection device based on dictionary learning, characterizing current features through a sparse matrix, and combining SVM for fault classification is proposed;

[0081] Analyze the fluctuation ranges of multiple characteristic quantities of the line in the normal state and when a fault arc occurs, determine thresholds based on the periodic increase ratios of different state characteristic quantities, and use multiple thresholds to complement each other to form the final combined - load detection threshold to achieve fault arc detection;

[0082] Based on the arc fire database, train a double - layer LSTM model under Bayesian optimization to identify multi - sensor signals within a time window and achieve fire warning.

[0083] The following is the specific implementation process of the present invention.

[0084] The current waveforms generated by fault arcs are complex and random, and it is difficult to conduct a comprehensive and detailed study on the diversity of load types in reality. The present invention builds a simulation test platform for the causes of arcing electrical fires, selects typical single - load and combined - load for tests during normal line operation and before the occurrence of arcing - caused fires, and establishes a fault arc waveform and an arc fire database. The test platform consists of a fault arc simulation generation device, the main line of the test platform, a data acquisition module, a control and status detection module, and an upper computer.

[0085] The test platform selects several common residential loads, including 3A, 6A, 13A, and 20A pure resistive loads, as well as seven typical non-linear electrical equipment specified in the GB / T 31143 standard, including fluorescent lamps, halogen lamps, dimmable lamps, switching power supplies, vacuum cleaners, air compressors, and electric hand tools. The main circuit of the test platform consists of multiple electrical circuits. The fault arc simulation generator can be switched to the set load fault point through the upper computer, and load sockets are reserved for self-selection of load combinations according to requirements.

[0086] Extract the characteristic indexes of the normal operation and fault arc current of each electrical circuit. The characteristic indexes include waveform index, kurtosis index, frequency characteristic ratio, and frequency centroid, and construct a fault arc database. Use multi-sensors to extract data at key positions of the line during normal operation and before the occurrence of arc-caused fires, and construct an arc fire database.

[0087] (1) Power-on fault detection method for fault arc detection device

[0088] At present, many fault arc detection devices achieve fault arc detection by comparing the periodic characteristic differences before and after the fault. When there is a fault arc in the line before the fault arc detection device is powered on, the fault arc detection device cannot achieve normal detection. The present invention sets up a power-on fault detection function to prevent the misoperation of the fault arc detection device when it is powered on. Based on the fault arc waveform database, through dictionary learning, a sparse matrix is constructed to represent the current waveform characteristics, and SVM is combined to realize fault arc classification.

[0089] Integrate the arc waveform database into a current matrix X, and the size of the matrix is m*n. Where m represents the signal length of each current sample, and n represents the number of samples, including fault arc currents and normal current samples under different loads.

[0090] X≈DA

[0091] In the formula, D represents the dictionary matrix extracted from the original signal, and A represents the sparse coefficient matrix.

[0092] Use MAP for dictionary learning. The MAP method is based on the principle of maximum a posteriori estimation and updates the dictionary by minimizing the reconstruction error.

[0093] D (k+1) =D (k) +βR T RD (k)

[0094] R=X - D (k) A (k)

[0095] A (k) =D (k)-1 X

[0096] In the formula, D (k) represents the dictionary matrix of the current iteration, β represents the learning rate, R is the residual matrix, and A (k) represents the sparse matrix of the current iteration.

[0097] The final dictionary matrix D and sparse coefficient matrix A are obtained by optimizing the following objective function.

[0098]

[0099] In the formula, is the reconstruction error, which is used to measure the difference between the original signal and the reconstructed signal. λ is the sparsity regularization parameter, which is used to control the strength of sparsity. |A||1 is the L1 norm, which is used to promote the sparsity of the sparse coefficient matrix.

[0100] Taking the sparse coefficient vector A i obtained by dictionary learning as the input of the SVM, the core information of the signal can be more effectively extracted through sparse representation, reducing noise and redundancy. The support vector machine classifies by finding a hyperplane that maximizes the margin between different classes, and the hyperplane constraint conditions are satisfied:

[0101] w T A i +b = 0

[0102] In the formula, w represents the normal vector of the hyperplane, and b represents the bias term of the hyperplane.

[0103] The SVM is trained by optimizing the following objective function:

[0104]

[0105] In the formula, C represents the penalty parameter, which controls the penalty degree of the classifier for misclassification. δ i represents the slack variable, allowing some samples to be misclassified on the decision boundary.

[0106] The classification decision function is:

[0107] f(x) = w T B i +b

[0108] In the formula, B i represents the current signal of the time window collected in real time, and b represents the bias term of the hyperplane. If f(x)>0, it is considered a positive class sample, and it is considered that the current line is normal. At this time, the fault arc detection algorithm is entered. On the contrary, if f(x)<0, it is considered a negative class sample, and the current signal in the time window is resampled until the normal signal of the line appears.

[0109] (2) Implementation of the Fault Arc Detection Algorithm (see Figure 1 )

[0110] Whether applied to linear loads or nonlinear loads, the circuit exhibits high stability during normal operation, and its characteristic parameters usually only fluctuate within a small range, which is in sharp contrast to the circuit in the fault state.

[0111] When a fault arc occurs, the high-frequency components of the current in the line increase significantly, and the harmonic contents of each order are decomposed by implementing the fast Fourier transform. The thresholds of each characteristic quantity are different under different loads. Relying solely on the threshold of a certain load as the threshold of the entire fault arc identification system will lead to misjudgment and missed judgment. Four time-frequency domain characteristic indexes are selected, namely the waveform index C f , the kurtosis index K v , the frequency characteristic ratio SR f , and the frequency centroid w.

[0112] The transformation expression of the fast Fourier transform

[0113]

[0114] In the formula, n is the number of discrete points of the discrete signal sequence.

[0115] The waveform index C f Expression

[0116]

[0117] In the formula, N is the number of single-cycle sampling points, and i k is the kth sampling point.

[0118] The kurtosis index K v Expression

[0119]

[0120] The frequency characteristic ratio SR f Expression

[0121]

[0122] In the formula, I j is the amplitude of the jth harmonic.

[0123] The expression of the frequency centroid w

[0124]

[0125] For waveform indicators, due to the distortion of the current waveform caused by the presence of a faulty arc, the faulty current waveforms of various loads are distorted compared to the normal operating current waveforms, and the waveform indicators all increase to varying degrees. For the kurtosis indicator, due to the presence of a faulty arc, the sharpness of the data distribution and the zero rest time are increased, resulting in a skewed distribution of the current waveform, and the kurtosis indicators of various loads all increase to varying degrees. Especially for the switched-mode power supply load, when the conduction angle conducts, the amplitude surges caused by the re-ignition of the arc lead to extremely obvious changes in the waveform indicator and the kurtosis indicator. For the frequency characteristic ratio and the frequency centroid frequency domain characteristic quantity, the electronic switched-mode power supply and the dimming lamp carry odd harmonics during normal operation at 90°, and their frequency characteristic ratios are correspondingly large. When a faulty arc occurs, the fundamental wave amplitude decreases and the harmonic content increases, and the frequency characteristic ratios and the frequency centroids of various loads all increase to varying degrees.

[0126] The four characteristic parameters extracted in the experiment showed obvious differences between the normal and faulty operating states under various load conditions. This indicates that for a specific load type, the threshold of a single characteristic parameter can be used as a judgment criterion to identify faulty arcs. However, the thresholds of each characteristic quantity under different loads are different. Simply relying on the threshold of a certain load as the threshold of the entire faulty arc identification system will lead to misjudgments and missed judgments. Whether it is the time-domain characteristic quantity or the frequency-domain characteristic quantity, there is an overlap between the faulty-state characteristic quantities and the normal-state characteristic quantities of different loads, and the value of a certain characteristic quantity cannot be directly selected as the basis for judging faulty arcs. When a faulty arc occurs in the circuit, due to the randomness of the faulty arc, the waveform periodicity is severely damaged. In order to find the threshold of the characteristic quantity suitable for more loads, we can start from the cycle increase ratio of the characteristic quantity.

[0127] Expression of the characteristic quantity increase ratio μ

[0128]

[0129] In the formula, x is the characteristic quantity calculated in the current cycle, x0 is the characteristic quantity calculated in the reference cycle, and x max is the larger value between x and x0.

[0130] When the load is operating normally, the fluctuations of the same characteristic quantity are relatively stable. The cycle increase ratios of the normal current characteristic quantities of the four selected characteristic quantities under single-load and combined-load parallel operation are generally lower than those of the faulty arc current characteristic quantities. Therefore, when using the cycle increase ratio of the characteristic quantity in different states as the judgment basis, a unified threshold can be obtained to achieve the detection of faulty arcs under a relatively large number of load combinations.

[0131] Initially, current is collected with a window of two cycles, feature quantities are extracted, and the waveform index, kurtosis index, frequency feature ratio, and the period increase ratio of the feature quantity of the frequency centroid are calculated respectively. To ensure that the reference value is the feature quantity extracted during normal line operation, the normal feature quantity accumulator threshold A th is equal to 3, that is, when three of the four feature quantity period increase ratios are less than the corresponding thresholds, it is determined that the feature quantity extracted from the current in the latter cycle is used as the reference value.

[0132] Subsequently, single-cycle current collection is started, and four types of feature quantities are extracted. After extracting the feature quantity of the newly collected current, the period increase ratio of the feature quantity is calculated with the feature quantity reference value. To prevent the period increase ratio of a certain feature quantity during normal operation from fluctuating beyond the threshold under different load combinations, the fault feature quantity accumulator threshold B th is greater than or equal to 3, that is, when three of the feature quantity period increase ratios are greater than the corresponding fluctuation thresholds, it is considered that a fault arc may occur. At this time, the fault arc accumulator C is incremented by 1, and the normal current reference value is not updated. When the fault feature quantity accumulator threshold B th is less than 3, if C is greater than 0, it is decremented by 1, and the feature quantity collected last time is updated as the normal feature quantity.

[0133] Finally, to prevent accidental mutation points caused by the influence of line interference signals, take C th equal to 3. When the fault arc accumulator accumulates to its threshold C th , it is determined that a fault arc has occurred in the line.

[0134] (3) Implementation of multi-sensor fire warning (see Figure 2 )

[0135] Before a fire occurs, it contains a lot of information. The precursors of a fire are often accompanied by a decrease in environmental humidity, an increase in temperature, and the generation of fault arcs. When there is smoldering, smoke is generated, and carbon monoxide is generated under insufficient combustion conditions. There may be slight sounds of burning objects, such as objects cracking due to heat.

[0136] In addition to using a current sensor to extract the characteristics of the fault arc waveform for fault diagnosis, multiple temperature sensors, carbon monoxide sensors, smoke sensors, humidity sensors, flame sensors, current sensors, high-frequency noise sensors, and sound sensors are also used to detect at key positions on the line. Based on the fire database, the present invention trains a Bayesian-optimized double-layer LSTM model for identifying multi-sensor signals within a time window to achieve fire warning.

[0137] Integrate the arc fault database into a sample matrix \(m\times n\), where \(m\) is the number of samples and \(n\) is the number of specific sensors. Normalize the sample matrix to eliminate the dimension difference, and divide it into a training set and a test set for model training and evaluation. Select 80% of the data as the training set and 20% of the data as the test set, and classify them. Normal is 0 and abnormal is 1.

[0138] The search space defines the value range of the hyperparameters to be adjusted in the LSTM temperature fault detection model in Bayesian optimization. Bayesian optimization will search for the optimal parameter combination within this range. Initialize the number of units, dropout rate, learning rate, and batch size of the double-layer LSTM model. The number of units affects the feature extraction ability, complexity, training time, and generalization performance of the model. The dropout rate represents the proportion of neurons that will be randomly set to zero during training to prevent overfitting. The learning rate determines the parameter update amplitude of the model in each iteration. Too high may cause the model to skip the optimal solution, and too low will result in a slow training process. The batch size determines the number of samples used for each model update. A smaller batch size can make the model adjust more finely but will increase the training time, while a larger batch size may accelerate the training.

[0139] Build an LSTM model, receive the initial values of the number of units and dropout rate set in the previous step, and build a two-layer LSTM network model. Each layer of the LSTM network is followed by a dropout layer. The output result of the first layer of the LSTM network is processed by the first dropout layer and used as the input of the second layer of the LSTM network. Finally, the result is output through a fully connected layer. The first layer of the LSTM extracts short-term features by gradually processing the time series of the original data and captures the data dependence patterns within a shorter time range. These short-term features are passed to the second layer of the LSTM, which is responsible for identifying global dependencies within a longer time range and extracting more complex potential signals related to fire prediction.

[0140] Use the loss function to measure the gap between the model output and the true label. When the model prediction value is close to the true label, the loss function is small, and vice versa. The model adjusts the hyperparameters through Bayesian optimization during training to minimize the loss function.

[0141] Expression of loss function B

[0142]

[0143] where \(N\) is the total number of samples, \(y\) i is the classification label of the actual sample. is the probability that the sample is 1. In the classification task, is a probability value from 0 to 1, that is, the error between the training set and the test set, used to measure the training effect of the model.

[0144] Bayesian optimization uses a surrogate model to approximately calculate the loss function minimized by the LSTM. This surrogate model predicts the value of the objective function at unevaluated points based on historical evaluation results and estimates the uncertainty of the prediction. It is necessary to reasonably define the number of iterations of the Bayesian optimization model. By continuously modifying and iterating the number of units, dropout rate, learning rate, and batch size of the double-layer LSTM model, while ensuring iterative convergence, the loss function is minimized, and the corresponding hyperparameters are retained.

[0145] Under the final LSTM model, accuracy, precision, recall, and F1-score are calculated, and the minimum lower limits of accuracy and F1-score are set. If the minimum values of accuracy and F1-score calculated by the final LSTM model do not meet the requirements, the ratio of the training set and the test set is readjusted. The adjustment range of the training set is 70 - 85%, and the adjustment range of the test set is exactly the opposite, which is 15 - 30%.

[0146] The multi-sensor takes the sampling data within twenty seconds as a time window and inputs it into the trained double-layer LSTM model optimized by Bayesian optimization. The hidden state vector of the second layer of LSTM is output, and then the second layer of LSTM model is used as the input of the fully connected layer to obtain the probability of a fire occurring.

[0147] Expression of the fire occurrence probability y

[0148] y = f(Wh + b)

[0149] In the formula, f is the activation function, W is the weight matrix, h is the hidden state vector of the second layer of LSTM, and b is the bias vector. The weight matrix W and the bias vector b of the fully connected layer are obtained through model training iteration. When y is greater than a certain threshold, it is considered that there is a fire risk and a warning is issued. This threshold needs to be defined according to the on-site environment.

[0150] The above are the preferred embodiments of the present invention. All changes made according to the technical solution of the present invention, when the functional effects produced do not exceed the scope of the technical solution of the present invention, shall fall within the protection scope of the present invention.

Claims

1. A fault arc-caused fire warning method based on multi-sensor fusion, characterized in that, Including: By building a simulation test platform for the causes of arcing electrical fires, selecting typical single loads and combined loads for testing during normal line operation and before the occurrence of arcing-caused fires, a fault arc waveform database and an arc fire database are established. A power-on detection method for a fault arc detection device is proposed, which is based on dictionary learning, characterizes current features through a sparse matrix, and combines SVM for fault classification. Analyze the fluctuation ranges of multiple characteristic quantities of the line in the normal state and when a fault arc occurs, determine the threshold based on the periodic increase ratio of different state characteristic quantities, and use multiple thresholds to complement each other to form the final combined load detection threshold to achieve fault arc detection. Based on the arc fire database, train a double-layer LSTM model optimized by Bayesian to identify multi-sensor signals within a time window to achieve fire warning.

2. The method for early warning of fault arc-caused fire based on multi-sensor fusion according to claim 1, wherein, The simulation test platform for the causes of arcing electrical fires includes a fault arc simulation generation device, the main line of the test platform, a data acquisition module, a control and status detection module, and a host computer.

3. The method for early warning of fire caused by faulty electric arcs based on multi-sensor fusion according to claim 2, wherein, The loads selected for the simulation test platform for the causes of arcing electrical fires include 3A, 6A, 13A, and 20A pure resistive loads, as well as seven typical non-linear electrical appliances, including fluorescent lamps, halogen lamps, dimmable lights, switching power supplies, vacuum cleaners, air compressors, and electric hand tools; the main line of the test platform consists of multiple electrical circuits. The fault arc simulation generation device is switched to the set load fault point through the host computer and the control and status detection module, and at the same time, load sockets are reserved to freely select load combinations according to requirements; the data acquisition module extracts the current characteristic indexes of normal operation and the occurrence of fault arcs in each electrical circuit. The characteristic indexes include waveform indexes, kurtosis indexes, frequency characteristic ratios, and frequency centroids, and a fault arc waveform database is constructed. Use multiple sensors in the data acquisition module to extract data at key positions of the line during normal operation and before the occurrence of arcing-caused fires to construct an arc fire database.

4. The method for early warning of fire caused by faulty electric arc based on multi-sensor fusion according to claim 1, characterized in that, The power-on detection method for the fault arc detection device can prevent the misoperation of the fault arc detection device when powered on.

5. The method for early warning of fire caused by faulty electric arcs based on multi-sensor fusion according to claim 1 or 4, characterized in that The power-on detection method for the fault arc detection device is based on the fault arc waveform database, constructs a sparse matrix to characterize the current waveform features through dictionary learning, and combines SVM to achieve fault arc classification.

6. The method for early warning of fire caused by faulty electric arc based on multi-sensor fusion according to claim 1 or 4, characterized in that, The power-on detection method for the fault arc detection device is specifically implemented as follows: Integrate the fault arc waveform database into a current matrix X, the size of the matrix is m*n, where m represents the signal length of each current sample, and n represents the number of samples, including fault arc currents and normal current samples under different loads. X≈DA In the formula, D represents the dictionary matrix extracted from the original signal, and A represents the sparse coefficient matrix. The MAP method is used for dictionary learning. The MAP method is based on the principle of maximum a posteriori estimation and updates the dictionary by minimizing the reconstruction error. D (k+1) = D (k) + βR T RD (k) R = X - D (k) A (k) A (k) = D (k)-1 X where D (k) represents the dictionary matrix of the current iteration, β represents the learning rate, R is the residual matrix, and A (k) represents the sparse matrix of the current iteration; The final dictionary matrix D and sparse coefficient matrix A are obtained by optimizing the following objective function: In the formula, is the reconstruction error, which is used to measure the difference between the original signal and the reconstructed signal. λ is the sparsity regularization parameter, which is used to control the strength of sparsity. ||A||1 is the L1 norm, which is used to promote the sparsity of the sparse coefficient matrix; The sparse coefficient vector A obtained by dictionary learning i is used as the input of the SVM, and the core information of the signal is extracted through sparse representation; the SVM classifies by finding a hyperplane that maximizes the margin between different classes, and the hyperplane constraint conditions are satisfied as follows: w T A i +b = 0 In the formula, w represents the normal vector of the hyperplane, and b represents the bias term of the hyperplane. The SVM is trained by optimizing the following objective function: where C represents the penalty parameter, which controls the degree of penalty for misclassification by the SVM, and δ i represents the slack variable, allowing some samples to be misclassified on the decision boundary; The classification decision function is: f(x) = w T B i + b where B i represents the time-window current signal collected in real time, and b represents the bias term of the hyperplane; if f(x)>0, it is considered a positive-class sample and the current line is considered normal, and at this time, the fault arc detection algorithm is entered; otherwise, if f(x)<0, it is considered a negative-class sample, and the current signal within the time window is resampled until the normal signal of the line appears.

7. The method for early warning of fault arc-caused fire based on multi-sensor fusion according to claim 1, wherein The specific implementation of fault arc detection is as follows: When a faulty arc occurs, the current in the circuit is subjected to fast Fourier decomposition to obtain the harmonic contents of each order, and four time-frequency domain characteristic indexes are selected, namely the waveform index C f , the kurtosis index K v , the frequency characteristic ratio SR f , and the frequency centroid w; The threshold of a single characteristic parameter can be used as a judgment benchmark to identify faulty arcs; however, the thresholds of each characteristic quantity under different loads are different. Simply relying on the threshold of a certain load as the threshold of the entire faulty arc identification system will lead to misjudgment and missed judgment. Therefore, starting from the periodic increase ratio of characteristic quantities, when using the periodic increase ratio of characteristic quantities in different states as the judgment basis, a unified threshold is obtained to achieve the detection of faulty arcs under a relatively large number of load combinations. The expression of the characteristic quantity increase ratio μ is: where x is the characteristic quantity calculated in the current period, x0 is the characteristic quantity calculated in the reference period, and x max is the larger value between x and x0; Initially, current is collected with a two - period window, feature quantities are extracted, and the waveform index, kurtosis index, frequency feature ratio, and the feature quantity period increase ratio of the frequency centroid are calculated respectively; to ensure that the reference value is the feature quantity extracted during normal line operation, the normal feature quantity accumulator threshold A th is equal to 3, that is, when three of the four feature quantity period increase ratios are less than the corresponding thresholds, the feature quantity extracted from the current in the subsequent period is determined as the reference value; Subsequently, start the current acquisition for a single cycle and extract four characteristic quantities; After extracting the feature quantity of the newly collected current, perform a feature quantity periodic increase ratio calculation with the feature quantity reference value; in order to prevent the fluctuation of a certain feature quantity periodic increase ratio from exceeding the threshold under different load combinations, take the fault feature quantity accumulator threshold B th Greater than or equal to 3, that is, when there are three feature quantity periodic increase ratios greater than the corresponding fluctuation thresholds, it is considered that a fault arc may occur. At this time, the fault arc accumulator C is incremented by 1, and the normal current reference value is not updated; when the fault feature quantity accumulator threshold B th Less than 3, if C is greater than 0, it is decremented by 1, and the feature quantity collected last time is updated to the normal feature quantity; Finally, to prevent accidental mutation points caused by the influence of line interference signals, take C th equal to 3. When the fault arc accumulator accumulates to its threshold C th , it is determined that a fault arc has occurred in the line.

8. The method for early warning of fire caused by faulty arcs based on multi-sensor fusion according to claim 7, characterized in that The transformation expression of the fast Fourier transform is as follows: where n is the number of discrete points of the discrete signal sequence, where k is the frequency index, i is the current index, and X k is the k-th frequency component in the frequency domain, and x i is the n-th sampling point in the time domain; Waveform Index C f The expression is: where N is the number of sampling points in a single period, and i k is the k-th sampling point; Kurtosis index K v The expression is: Frequency characteristic ratio SR f The expression is as follows: where I j is the amplitude of the j-th harmonic; The expression of the frequency centroid w is:

9. The method for early warning of fire caused by faulty arc based on multi-sensor fusion according to claim 1, characterized in that, The multi-sensors include a temperature sensor, a carbon monoxide sensor, a smoke sensor, a humidity sensor, a flame sensor, a current sensor, a high-frequency noise sensor, and a sound sensor arranged at key positions of the line.

10. The method for early warning of fire caused by faulty electric arcs based on multi-sensor fusion according to claim 1 or 9, characterized in that, Based on the arc fire database, train a double-layer LSTM model under Bayesian optimization to identify multi-sensor signals within a time window and achieve fire early warning. The specific implementation method is: Integrate the arc fire database into a sample matrix m*n, where m is the number of samples and n is the specific number of sensors; normalize the sample matrix to eliminate the dimension difference, and divide it into a training set and a test set for model training and evaluation; select 80% of the data as the training set and 20% of the data as the test set, and classify them. Normal is 0 and abnormal is 1; Initialize the number of units, dropout rate, learning rate, and batch size of the double-layer LSTM model, construct a double-layer LSTM model, and follow an exit layer after each layer of the LSTM network. The output result of the first layer of the LSTM network is processed by the first exit layer and used as the input of the second layer of the LSTM network. Finally, the result is output through a fully connected layer; the first layer of the LSTM extracts short-term features by gradually processing the time series of the original data; captures the data dependence patterns within a relatively short time range, and these short-term features are passed to the second layer of the LSTM. The second layer of the LSTM is responsible for identifying global dependencies within a longer time range and extracting more complex potential signals related to fire prediction; Use a loss function to measure the gap between the model output and the true label. When the model prediction value is close to the true label, the loss function is small, otherwise it is large; During the training process of the model, the hyperparameters are adjusted by Bayesian optimization to minimize the loss function; The expression of the loss function B where N is the total number of samples, and y i is the classification label of the actual sample, is the probability that the sample is 1. In the classification task, is a probability value from 0 to 1, that is, the error between the training set and the test set, which is used to measure the training effect of the model; Bayesian optimization uses a surrogate model to approximately calculate the minimization of the loss function of the double-layer LSTM model; this surrogate model predicts the value of the objective function at unevaluated points based on historical evaluation results and estimates the uncertainty of the prediction. By continuously modifying and iterating the number of units, dropout rate, learning rate, and batch size of the double-layer LSTM model, the loss function is minimized while ensuring the iteration converges, and the corresponding hyperparameters are retained; Calculate the accuracy, precision, recall rate, and F1 score under the final double-layer LSTM model, and set the minimum lower limits of the accuracy and F1 score; If the minimum values of the accuracy and F1 score calculated by the final double-layer LSTM model do not meet the requirements, readjust the ratio of the training set and the test set; the adjustment range of the training set is 70-85%, and the adjustment range of the test set is exactly the opposite, which is 15-30%; The multi-sensor uses the sampling data within twenty seconds as a time window and inputs it into the trained double-layer LSTM model optimized by Bayesian to output the hidden state vector of the second-layer LSTM, and then uses the second-layer LSTM model as the input of the fully connected layer to obtain the probability of a fire occurring; The expression of the fire occurrence probability y is: y = f(Wh + b) In the formula, f is the activation function, W is the weight matrix, h is the hidden state vector of the second-layer LSTM, and b is the bias vector. The weight matrix W and the bias vector b of the fully connected layer are obtained through model training iteration; when it is set that y is greater than a certain threshold, it is considered that there is a fire risk and a warning is issued, and this threshold is defined by the on-site environment.

Images