A deep learning-based cable early fire heat generation sound signal recognition method

Abstract

The application discloses a kind of early heating sound signal identification methods of cable fire based on deep learning, its steps include: through microphone device acquisition cable temperature sound signal;Combining normalized least mean square NLMS algorithm, echo cancellation and filtering processing are carried out to sound signal;Mel frequency cepstrum coefficient MFCC of processed sound signal is extracted, sample library is formed, and is input into neural network to carry out training identification, and sound signal training model is constructed;Call training model, real-time detect whether cable has over-temperature sign, prevent cable fire from happening.The application adopts neural network training in deep learning and identifies the acoustic emission signal generated in the early heating of underground cable, to realize the early detection of underground cable fire, and provide a new method for the early detection of fire in urban power grid underground cable.

Description

Technical Field

[0001] This invention relates to the field of cable fire detection, and in particular to a method for identifying early-stage heating acoustic signals in cable fires based on deep learning. Background Technology

[0002] With the rapid development of the national economy and urbanization, the demand for power infrastructure is also increasing. Power cables, with their advantages of saving ground space, high safety, and reliability, have been widely used in power systems. Currently, the most widely used cable insulation material is cross-linked polyethylene (XLPE), a high-molecular-weight rubber-plastic material. Excessive operating temperature can lead to thermal aging of the material, resulting in a decrease in its insulation performance. Therefore, the maximum operating temperature of XLPE cables is required to be below 90℃. However, transmission lines themselves have impedance, converting a certain amount of electrical energy into heat energy during current transmission, causing the lines to heat up. As the service life of cables increases and user loads increase year by year, the heating of lines becomes increasingly serious, with conductor temperatures reaching as high as 250℃ during short-circuit faults. Investigations into the causes of fires in power systems have found that most are due to power cables operating under high-voltage power supply for extended periods, leading to cable overheating and ultimately fires. Therefore, sufficient attention should be paid to the detection and early warning of underground cable fires in urban areas.

[0003] In existing technologies, the detection and early warning of urban underground cable fires mainly fall into two categories. One category uses traditional detectors, such as smoke detectors for detecting fire smoke and heat detectors for detecting fire temperature. However, due to the flammability of the cable material itself, when these detectors alarm, the fire has often already spread along the cables, producing large amounts of high-temperature and toxic smoke. Furthermore, the detectors are limited by their own precision and quality, resulting in high instability in detection effectiveness and accuracy. The other category relies on computer-based video fire identification methods. While these methods offer high accuracy and fast response, they are susceptible to lighting conditions. Problems arise when the detected object is not completely covered by the camera, leading to blind spots that cannot be predicted. Therefore, researching and developing more accurate and economical online cable temperature measurement technology to monitor cable operating temperature in real time and accurately identify and respond to overheating hazards is of great significance for ensuring the safety and reliability of the power grid. Summary of the Invention

[0004] The problem to be solved by this invention is to provide a method for early heating sound signal identification of cable fires based on deep learning, which provides a new method for early detection of fires in underground cables of urban power grids.

[0005] The present invention adopts the following technical solution:

[0006] A deep learning-based method for identifying early-stage heating acoustic signals in cable fires includes the following steps:

[0007] S1, collects the acoustic signal of the cable heating up through a microphone device;

[0008] S2, combined with the Normalized Least Mean Square (NLMS) algorithm, performs echo cancellation and filtering on the acoustic signal;

[0009] S3. Extract the Mel-frequency cepstral coefficients (MFCCs) of the processed acoustic signal, form a sample library, input it into the neural network for training and recognition, construct an acoustic signal recognition model, and identify early acoustic signals of cable fires through multiple rounds of training.

[0010] S4 invokes the trained model to detect signs of overheating in cables in real time, preventing cable fires.

[0011] Specifically, in step S2, the Normalized Least Mean Square (NLMS) algorithm normalizes the tap weight adjustment using the squared Euclidean norm of the tap input vector. The specific function expression is as follows:

[0012]

[0013] Where n represents the number of iterations, For input signal, Let μ(n) be the column vector of filter tap weights obtained in the nth iteration, μ(n) be the convergence factor, μ(n) be the step size used in the nth iteration, and e(n) be the input. The error between the expected response and the actual output of the time filter.

[0014] Furthermore, in step S3, the extraction of the Mel-frequency cepstral coefficients (MFCCs) of the acoustic signal includes: preprocessing the original audio, framing, windowing, Fourier transform, Mel filter bank, logarithmic operation, and dynamic feature extraction steps. The weighted dynamic difference method is used as the feature parameter of the acoustic signal, which includes the static and dynamic features of the acoustic signal.

[0015] The weighted dynamic difference coefficient is calculated as follows:

[0016] The Mel-frequency cepstral coefficients of the acoustic signal are combined with its first-order and second-order differences with certain weights to form a new characteristic parameter M. new The formula is:

[0017] M new =M + a·ΔM + b·Δ 2 M

[0018] In this model, the MFCC coefficient is 1, the first-order difference coefficient is a, and the second-order difference coefficient is b. These coefficients are combined to obtain the 13-dimensional characteristic parameters of the acoustic signal. M represents the static characteristics of the audio signal, ΔM represents the dynamic characteristics of the audio signal, and Δ... 2 M is its balance factor, and the formula is:

[0019]

[0020]

[0021] Where k is a constant, and n is 1, 2, 3, ...

[0022] Preferably, in the above formula for calculating the weighted dynamic difference coefficients, the first-order difference coefficient a is 1 / 3, the second-order difference coefficient b is 1 / 6, and k is 2.

[0023] Furthermore, the neural network mentioned in step S3 is a convolutional neural network (CNN), which includes: 1 input layer, 2 convolutional layers, 2 pooling layers, 1 fully connected layer, and 1 output layer; wherein, the input layer inputs a 13-dimensional acoustic signal feature vector, the convolutional and pooling layers extract acoustic signal features, and the fully connected layer classifies acoustic signals.

[0024] Specifically, in the Convolutional Neural Network (CNN), for the input 13-dimensional acoustic signal feature vector x, the convolutional layer uses a 3×3 convolutional kernel for convolution, and the activation function is ReLU, as shown in the following formula:

[0025] f(x) = max(0, x)

[0026] After two convolutional layers and one fully connected layer, the input data dimension is compressed from 13×99×1 to a one-dimensional array, and the output dimension is 2. Then, the softmax activation function is used to convert the output result into a probability distribution. The formula of the softmax activation function is as follows:

[0027]

[0028] Where z(i) is the i-th element of vector z, and vector z has a total of j elements.

[0029] Furthermore, in step S1, the acoustic signal of cable heating is collected through a simulated cable heating platform. The simulated cable heating platform includes: a heating rod, a short-circuit protector, a microphone and bracket, a cable support grid, and an exhaust fan. The heating rod simulates the cable core that gradually heats up after a cable fault occurs in the circuit. The microphone bracket is used to fix the microphone. The cable support grid is used to arrange the cable. The exhaust fan is used to remove the fumes emitted during cable heating in a timely manner.

[0030] Furthermore, the sample library mentioned in step S3 includes acoustic signals of cables overheating induced by heating rods of different power and acoustic signals of cables of different radii overheating induced by heating rods, respectively simulating the acoustic signals generated when cables of different radii in different circuit loads and different cable lines experience overheating.

[0031] Compared with the prior art, the present invention, employing the above technical solution, has the following technical effects:

[0032] 1. This invention provides the acquisition of acoustic signals of cables overheating induced by heating rods of different power and cables of different radii induced by heating rods. It simulates the acoustic signals generated when cables in different circuit loads and different cable lines overheat in reality. It provides a relatively complete coverage of various early fire acoustic signals generated in actual cable operation, and the accuracy of identifying early cable overheating is higher than most existing fire detection methods.

[0033] 2. This invention applies the Normalized Least Mean Square (NLMS) algorithm to the field of intelligent cable fire detection technology. By using the NLMS algorithm to process the collected acoustic signals, the accuracy of acoustic signal processing is improved.

[0034] 3. This invention uses a neural network in deep learning to train and identify acoustic emission signals generated by underground cables due to heat generation in the early stages of a fire, so as to realize the early detection of underground cable fires and provide a new method for the early detection of fires in underground cables of urban power grids. Attached Figure Description

[0035] Figure 1 This is a flowchart of the method for identifying early-stage heating sound signals in cable fires according to the present invention;

[0036] Figure 2 This is a flowchart of the MFCC coefficient extraction process for acoustic signals according to the present invention;

[0037] Figure 3 This is a schematic diagram of the convolutional neural network structure of the present invention. Detailed Implementation

[0038] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of the application will be further described in detail below with reference to the accompanying drawings. The described embodiments are only a part of the embodiments involved in this invention. All non-innovative embodiments based on this invention by other researchers in the art are within the protection scope of this invention.

[0039] like Figure 1 As shown, this invention provides a method for identifying early-stage heating sound signals in cable fires based on deep learning, characterized by the following steps:

[0040] S1, collects the acoustic signal of the cable heating up through a microphone device;

[0041] S2, combined with the Normalized Least Mean Square (NLMS) algorithm, performs echo cancellation and filtering on the acoustic signal;

[0042] S3. Extract the Mel-frequency cepstral coefficients (MFCCs) of the processed acoustic signal, form a sample library, input it into the neural network for training and recognition, construct an acoustic signal recognition model, and identify early acoustic signals of cable fires through multiple rounds of training.

[0043] S4 invokes the trained model to detect signs of overheating in cables in real time, preventing cable fires.

[0044] In a preferred embodiment of the present invention, step S2 involves using the Normalized Least Mean Square (NLMS) algorithm to perform echo cancellation and filtering on the acoustic signal. The Normalized Least Mean Square (NLMS) algorithm is an improved version of the Least Mean Square (LMS) algorithm, which overcomes the problems of slow convergence speed and data dependency in the LMS algorithm.

[0045] When processing audio signals, adaptive filters are commonly used for noise reduction. These filters automatically adjust their parameters based on previously obtained parameters to adapt to the statistical characteristics of unknown, random variations in signal and noise, thus achieving optimal filtering. The Least Mean Square (LMS) adaptive algorithm minimizes the mean square error between the desired response and the filter output signal. It estimates the gradient vector based on the input signal during iteration and updates the weight coefficients to achieve optimality. NLMS, or Normalized LMS, is an improved version of LMS. In LMS, the adjustment of the tap weights is proportional to the tap input vector. When the tap input vector is large, LMS encounters the problem of gradient noise amplification.

[0046] To overcome this problem, this method employs the NLMS algorithm, which normalizes the tap weight adjustments using the squared Euclidean norm of the tapped input vector. NLMS can be viewed as an LMS algorithm with a time-varying step size parameter; regardless of whether the data is irrelevant or relevant, NLMS exhibits a faster convergence speed than standard LMS. The functional expressions for LMS and NLMS are as follows.

[0047] LMS:

[0048] NLMS:

[0049] Where n represents the number of iterations, For input signal, Let μ(n) be the column vector of filter tap weights obtained in the nth iteration, μ(n) be the convergence factor, μ(n) be the step size used in the nth iteration, and e(n) be the input. The error between the expected response and the actual output of the time filter.

[0050] Then, the Mel-frequency cepstral coefficients (MFCCs) of the acoustic signal are extracted, such as... Figure 2 As shown, the method includes preprocessing, framing, windowing, Fourier transform, Mel filter bank, logarithmic operation, and dynamic feature extraction steps. This method uses weighted dynamic difference coefficients as feature parameters of the acoustic signal.

[0051] The preprocessing section denoises the original audio and enhances the effective sound signal. The denoising method involves importing the audio into Adobe Audition, denoising the audio noise, and enhancing the sound effect of the target sound signal, so as to obtain more accurate feature values ​​when extracting audio feature vectors.

[0052] Next, the audio signal is processed by framing, windowing, and Fourier transform to obtain more accurate and detailed audio spectral characteristics.

[0053] This method employs a fixed-frame-length framing approach, dividing the entire audio signal into frames of fixed length, and then processing each frame. A Hamming window is used to window the data in each frame; the formula is as follows:

[0054]

[0055] Where t is time and T is the sampling period;

[0056] Windowing can reduce the impact of spectral leakage in the subsequent Fourier transform, while the Mel filter bank takes into account that the Mel frequency scale is more consistent with the auditory characteristics of the human ear. The relationship between the Mel frequency and the actual frequency is as follows:

[0057]

[0058] It should be noted that the auditory characteristics of the human ear are consistent with the increase of Mel frequency, showing a linear distribution with the actual frequency below 1000Hz, and a logarithmic increase above 1000Hz.

[0059] Next, a weighted dynamic difference method is used as the characteristic parameter of the acoustic signal, which includes the static and dynamic characteristics of the acoustic signal.

[0060] Then, the weighted dynamic difference coefficients are calculated, specifically as follows:

[0061] The Mel-frequency cepstral coefficients of the acoustic signal are combined with its first-order and second-order differences with certain weights to form a new characteristic parameter M.new The formula is:

[0062] M new =M + a·ΔM + b·Δ 2 M

[0063] In this model, the MFCC coefficient is 1, the first-order difference coefficient is a, and the second-order difference coefficient is b. These coefficients are combined to obtain the 13-dimensional characteristic parameters of the acoustic signal. M represents the static characteristics of the audio signal, ΔM represents the dynamic characteristics of the audio signal, and Δ... 2 M is its balance factor, and the formula is:

[0064]

[0065]

[0066] Where k is a constant, and n = 1, 2, 3...

[0067] Preferably, the MFCC coefficient is 1. In the above formula for calculating the weighted dynamic difference coefficient, the first-order difference coefficient a is 1 / 3, the second-order difference coefficient b is 1 / 6, and k is 2.

[0068] In another preferred embodiment of the present invention, the neural network in step S3 is a convolutional neural network (CNN), comprising: an input layer, a convolutional layer, a pooling layer, a fully connected layer, and an output layer;

[0069] Preferred, such as Figure 3 As shown, considering the characteristics of cable acoustic signals, this embodiment designs the Convolutional Neural Network (CNN) as one input layer, two convolutional layers, two pooling layers, one fully connected layer, and one output layer. The input layer takes in a 13-dimensional acoustic signal feature vector. Convolutional layer 1 is followed by pooling layer 1, then convolutional layer 2 and pooling layer 2. The two convolutional layers and two pooling layers extract acoustic signal features, and finally, the fully connected layer classifies the acoustic signal.

[0070] In the convolutional neural network (CNN), the input 13-dimensional acoustic signal feature vector x is processed. The convolutional layers use 3×3 convolutional kernels, and the ReLU activation function is used, as shown in the following formula:

[0071] f(x) = max(0, x)

[0072] After two convolutional layers and one fully connected layer, the input data dimension is compressed into a one-dimensional array. Preferably, the input data dimension is 13×99×1 and the output dimension is 2. Then, the softmax activation function is used to convert the output result into a probability distribution for subsequent classification operations.

[0073]

[0074] Where z(i) is the i-th element of vector z, and vector z has a total of j elements.

[0075] Finally, the convolutional neural network divides the sample library into a training set and a test set. The training set is used for multiple training sessions, and the test set is input into the trained model for simulation testing. After multiple rounds of training to improve accuracy, the test results show that the model can achieve an accuracy of 98.93% in recognizing cable overheating sound signals.

[0076] It should be noted that the cable heating sound signal in step S1 is collected through a simulated cable heating platform. The simulated cable heating platform includes: a heating rod, a short circuit protector, a microphone and bracket, a cable support grid, and an exhaust fan. The heating rod simulates the cable core that gradually heats up after a cable fault occurs in the circuit. The microphone bracket is used to fix the microphone. The cable support grid is used to arrange the cable. The exhaust fan is used to remove the flue gas emitted during the cable heating in a timely manner.

[0077] It is particularly important to note that the sample library mentioned in step S3 includes acoustic signals of cables overheating induced by heating rods of different power and acoustic signals of cables of different radii overheating induced by heating rods, respectively simulating the acoustic signals generated when cables of different radii in different circuit loads and different cable lines experience overheating.

[0078] This invention aims to train and identify acoustic emission signals generated by underground cables in the early stage of temperature rise using a neural network in deep learning. It combines the Normalized Least Mean Square (NLMS) algorithm to perform echo cancellation and filtering on the acoustic signals, extracts the Mel-frequency cepstral coefficients (MFCCs) of the processed acoustic signals, and inputs them into a sample library for training and identification in the neural network. Finally, it calls the trained model to detect whether there are signs of overheating in the cable in real time, so as to realize the early detection of underground cable fires and provide a new method for the early detection of fires in underground cables of urban power grids.

[0079] The above description is only a preferred embodiment of the present invention. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the principle of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. A method for identifying early-stage heating acoustic signals in cable fires based on deep learning, characterized in that, Includes the following steps: S1, collects the acoustic signal of the cable heating up through a microphone device; S2, combined with the Normalized Least Mean Square (NLMS) algorithm, performs echo cancellation and filtering on the acoustic signal; S3. Extract the Mel-frequency cepstral coefficients (MFCCs) of the processed acoustic signal to form a sample library, input it into the neural network for training and recognition, construct an acoustic signal recognition model, and identify early acoustic signals of cable fires through multiple rounds of training; the sample library includes acoustic signals of cables overheating induced by heating rods of different power and acoustic signals of cables of different radii overheating induced by heating rods, respectively simulating the acoustic signals generated when cables of different radii in different circuit loads and different cable lines experience overheating. S4 calls the trained model to detect in real time whether there are signs of overheating in the cable, thus preventing cable fires. In step S3, weighted dynamic difference coefficients are used as characteristic parameters of the acoustic signal, which include the static and dynamic characteristics of the acoustic signal. The weighted dynamic difference coefficient is calculated as follows: The Mel-frequency cepstral coefficients of the acoustic signal are combined with its first-order and second-order differences with certain weights to form new characteristic parameters. The formula is: ； Wherein, the first-order difference coefficients are The second-order difference coefficients are By combining and processing these parameters, a 13-dimensional characteristic parameters for the acoustic signal are obtained, where M represents the static characteristics of the audio signal. Indicates the dynamic characteristics of audio. As its balancing factor, the formula is: ； ； in, is a constant, and n represents the number of iterations.

2. The method for identifying early-stage heating acoustic signals in cable fires based on deep learning according to claim 1, characterized in that, In step S2, the Normalized Least Mean Square (NLMS) algorithm normalizes the tap weight adjustment using the squared Euclidean norm of the tap input vector. Specifically, the function expression is as follows: ； in, For input signal, This is the column vector of filter tap weights obtained in the nth iteration. Let be the convergence factor, and represent the step size used in the nth iteration. For input The error between the expected response and the actual output of the time filter.

3. The method for identifying early-stage heating sound signals in cable fires based on deep learning according to claim 2, characterized in that, In step S3, the extraction of the Mel cepstral coefficients (MFCC) of the acoustic signal includes the following steps: preprocessing the original audio, framing, windowing, Fourier transform, Mel filter bank, logarithmic operation, and dynamic feature extraction.

4. The method for identifying early-stage heating sound signals in cable fires based on deep learning according to claim 1, characterized in that, In the formula for calculating the weighted dynamic difference coefficients, the first-order difference coefficients... 1 / 3, second-order difference coefficient It is 1 / 6. Take 2.

5. The method for identifying early-stage heating acoustic signals in cable fires based on deep learning according to claim 1, characterized in that, In step S3, the neural network is a convolutional neural network (CNN), which includes: 1 input layer, 2 convolutional layers, 2 pooling layers, 1 fully connected layer, and 1 output layer; wherein, the input layer inputs a 13-dimensional acoustic signal feature vector, the convolutional and pooling layers extract acoustic signal features, and the fully connected layer classifies acoustic signals.

6. The method for identifying early-stage heating sound signals in cable fires based on deep learning according to claim 5, characterized in that, The Convolutional Neural Network (CNN) divides the sample library into a training set and a test set. The training set is used for multiple training sessions to identify samples, and the test set is input into the trained model for simulation testing. Multiple rounds of training are conducted to improve the accuracy.

7. The method for identifying early-stage heating acoustic signals in cable fires based on deep learning according to claim 1, characterized in that, In step S1, the acoustic signal of cable heating is collected through a simulated cable heating platform. The simulated cable heating platform includes: a heating rod, a short circuit protector, a microphone and bracket, a cable support grid, and an exhaust fan. The heating rod simulates the cable core that gradually heats up after a cable fault occurs in the circuit. The microphone bracket is used to fix the microphone. The cable support grid is used to arrange the cable. The exhaust fan is used to remove the flue gas emitted during cable heating in a timely manner.

Images