A high-voltage circuit breaker mechanical fault online diagnosis method based on voiceprint recognition

Abstract

The application discloses a kind of high-voltage circuit breaker mechanical fault online diagnosis methods based on voiceprint identification, including sound sensor, state monitoring front end, cloud server and customer terminal.The sound sensor is installed switchgear shell, sound sensor is connected with state monitoring front end, and state monitoring front end sends to cloud server after signal processing according to the set communication time;Cloud server stores monitoring data, and simultaneously based on sound data history record, carries out data preprocessing, algorithm analysis, and identifies fault;Meanwhile, cloud server sends real-time monitoring data to customer terminal, and identification result, and user inquires historical data and high-voltage circuit breaker equipment fault identification result by customer terminal.The application judges whether high-voltage circuit breaker occurs mechanical fault by the mode of circuit breaker action sound signal monitoring and algorithm analysis identification, to ensure the safe and stable operation of circuit breaker and provide reference for its equipment maintenance plan.

Description

TECHNICAL FIELD

[0001] The application belongs to the field of artificial intelligence technology, online monitoring and fault diagnosis of power equipment, and particularly relates to a high-voltage circuit breaker mechanical fault online diagnosis method based on voiceprint recognition. BACKGROUND

[0002] The high-voltage circuit breaker plays a very important protection and control role in the power system by breaking, closing and carrying the normal or abnormal current of the operating line, and its operating state has a significant impact on the stability and reliability of the power system. In order to meet the requirements of arc control, the high-voltage circuit breaker must maintain good mechanical properties.

[0003] The main defects of the high-voltage circuit breaker generally include body faults, operating mechanism faults and secondary circuit faults. At present, operating mechanism faults and secondary circuit faults are still the main fault types of high-voltage circuit breakers; and according to the statistical data of the occurrence of circuit breaker faults, 63.2% of the circuit breaker faults in China are caused by operating mechanisms. Among the faults of the circuit breaker body, the spring mechanism faults are mostly manifested as spring fatigue, jamming and the like, thereby causing problems such as the decrease of the action reliability of the circuit breaker and the insufficient opening and closing speed. During the transportation and assembly of the circuit breaker, the vibration may also cause the loosening of the screws and other components, thereby causing the gap variation fault, and the excessively large or small gap may cause the problems of the failure of the action to reach the position or the insufficient kinetic energy, thereby causing the failure of the opening and closing.

[0004] In order to improve such problems, it has become a necessary solution to obtain various physical signals of the circuit breaker through monitoring means, and then analyze and diagnose the faults, which is the development trend in the field of circuit breaker fault detection.

[0005] Therefore, an effective, easy-to-implement, high-accuracy and robust circuit breaker fault diagnosis online monitoring method has great value for maintaining the normal operation of the power system, and changes the method of scheduled maintenance (SM) according to the production plan and experience, which is to check, disassemble and replace parts at regular intervals to prevent equipment damage, secondary damage and production loss, into a more economic and reasonable condition-based maintenance (CBM) maintenance strategy based on the actual operating state of the system, which is to perform periodic (or continuous) state monitoring on the main (or necessary) parts, and then implement the maintenance activities when there is evidence that the system needs maintenance.

[0006] The operation state of the power equipment can be effectively identified by the online monitoring method, and the occurrence and development of the deterioration fault can be predicted and prevented in time. The research results show that the waveform stability of the current is good and simple and intuitive, but the current signal cannot be used to comprehensively diagnose various faults of the circuit breaker, especially cannot be used to diagnose the faults occurring outside the electromagnet part, and these faults are the main faults of the high-voltage circuit breaker; although the vibration signal can detect the body fault and the operating mechanism fault, because the vibration signal needs to be in contact with the circuit breaker body, and the acceleration sensor for detecting the vibration signal can be displaced during the movement of the circuit breaker, the signal transmitted each time is inconsistent, and the accuracy of the diagnosis is affected. SUMMARY

[0007] To solve the above problems, the application discloses a high-voltage circuit breaker mechanical fault online diagnosis system and method based on voiceprint recognition, which can accurately monitor the sound signal of the circuit breaker equipment in real time through a non-contact mode, identify the fault condition through a background algorithm, and send the obvious abnormal state to the customer terminal in time, so as to not only make up for the shortcomings of the existing detection means and fill the gap of the high-voltage circuit breaker online monitoring technology, but also effectively monitor the operation state of the high-voltage circuit breaker in real time, discover potential fault defects in time, and further ensure the safe and stable operation of the distribution network in the jurisdiction.

[0008] To achieve the above purpose, the technical scheme of the application is as follows:

[0009] A high-voltage circuit breaker mechanical fault online diagnosis method based on voiceprint recognition, comprising a sound sensor, a state monitoring front end, a cloud server and a customer terminal. The sound sensor is installed on the switch cabinet shell and connected with the state monitoring front end, the state monitoring front end performs Mel filtering processing on the signal, records audio data when the circuit breaker is about to act each time, and sends the audio data to the cloud server; the cloud server stores the monitoring data, and performs data preprocessing, algorithm analysis and abnormal fault identification; the cloud server sends the real-time monitoring data and the abnormal fault identification results of the circuit breaker body and the operating mechanism to the customer terminal; the user queries the historical data and the fault identification results through the customer terminal, so as to obtain the fault condition of the circuit breaker and arrange maintenance accordingly.

[0010] Specifically, the following steps are included:

[0011] (1) First, the sound sensor picks up the equipment action sound and transmits it to the state monitoring front end for filtering preprocessing when the circuit breaker acts.

[0012] (2) A deep neural network containing a high-order frequency domain feature extraction module, a time sequence signal extraction module and a softmax classification module is designed, which can be used for classifying fault samples after training to achieve the effect of fault diagnosis.

[0013] (3) Simulate a variety of fault scenes and record sound samples, train the neural network model for different actions and different fault conditions, and use the trained model as the final fault diagnosis algorithm.

[0014] Further, in step (1), the signal is pre-emphasized and framed, windowed, and passed through a mel filter bank: the formula for pre-emphasizing the sampled digital signal is:

[0015] y(n) = x(n) - a x(n-1) (1)

[0016] where a is 0.95, n ∈ the set of integers with maximum value being the time of the sound signal × signal sampling rate sr, and minimum value being 1.

[0017] The window function formula when windowing is:

[0018]

[0019] where a is 0.46, n ∈ the set of integers with maximum value being the frame length (25 ms) × signal sampling rate sr, and minimum value being 1, and N is the frame length (25 ms) × signal sampling rate sr.

[0020] The scale of the mel filter bank is selected as 32 equal values in [0, fm_max], and converted to mel frequency and projected onto the frequency interval [0, N / 2] using the following formula:

[0021]

[0022]

[0023] where N is the frequency range selected for Fourier transform, and fs is the sampling frequency

[0024] The formula for generating the mel filter matrix is as follows:

[0025] H m (n) = 0, n < k(m-1)

[0026]

[0027]

[0028] H m (n) = 0, n ≥ k(m+1)

[0029] where the set of k values is the mel frequency vector generated by the formulas (3) and (4) in the previous step.

[0030] Further, the network structure of the frequency domain feature extraction module and the time sequence information feature extraction module in step (2) is as shown in Figure 4 , wherein the frequency domain feature extraction includes two modules ResBlock and expansion-Resblock, which are respectively used as the main structural components of the feature extraction process.

[0031]

[0032]

[0033]

[0034] wherein formula (6) is the formula of the conv layer, is the frequency domain feature of the upper layer for the range, and k is a learnable parameter corresponding to the feature; formula (7) is the formula of the BN layer, X is the feature corresponding to all batches of each channel, and γ and β are learnable parameters. Formula (8) is the formula of the relu layer, wherein x is all feature values after the perception of the conv layer and the normalization of the BN layer.

[0035] Time sequence information extraction module:

[0036] z t = σ (W z (x t ) + U z (h t-1 )) (update gate) # (9)

[0037] r t = σ (W r (x t ) + U t-1 (r t )) (reset gate)

[0038]

[0039]

[0040] wherein z t is the update gate value in the GRU module, r t is the reset gate value, is the hidden unit value, and h t is the value of the input next time step.

[0041] The last classification module uses FCN+softmax function to fit the occurrence probability of each fault, wherein the formula of the softmax is as follows:

[0042]

[0043] wherein w is a learnable parameter.

[0044] Further, in step (3), the loss function when training the neural network is a cross-entropy loss function, and the formula is as follows:

[0045]

[0046] wherein y i is the label value of the training data, and a i is the fitting probability value of the softmax layer output.

[0047] The beneficial effects of the present application are:

[0048] The present application accurately judges whether the circuit breaker body and operating mechanism have obvious abnormal faults by the sound signals during the operation of the circuit breaker, such as the sound of energy storage motor energy storage, so as to achieve the purpose of timely knowing the running state of the high-voltage circuit breaker and timely troubleshooting of degradation faults. The present application provides a high-voltage circuit breaker mechanical fault online diagnosis system and method based on voiceprint recognition, which not only makes up for the shortcomings of existing detection means and fills the gap of high-voltage circuit breaker online monitoring technology, but also further ensures the safe and stable operation of the power distribution network in the jurisdiction. BRIEF DESCRIPTION OF DRAWINGS

[0049] Figure 1 is the working principle diagram of the present application;

[0050] Figure 2 is the data processing flow chart of the state monitoring front end;

[0051] Figure 3 is the working flow chart of the recognition algorithm;

[0052] Figure 4 is the working flow chart of the time-frequency spectrum high-order feature extraction module in the recognition algorithm;

[0053] Figure 5 is the working flow chart of the time sequence feature extraction module in the recognition algorithm;

[0054] Figure 6 is the working flow chart of the background system. DETAILED DESCRIPTION

[0055] The present application will be further illustrated below in combination with the drawings and specific embodiments, and it should be understood that the following specific embodiments are only used to illustrate the present application and not to limit the scope of the present application.

[0056] For example, Figure 1As shown, this invention provides an online diagnostic method for mechanical faults in high-voltage circuit breakers based on voiceprint recognition. The method includes a sound sensor, a condition monitoring front-end, a cloud server, and a client terminal. Its working principle is as follows: the sound sensor monitors the sound signal during the opening and closing of the circuit breaker and transmits the monitoring signal to the condition monitoring front-end. The condition monitoring front-end performs filtering and preprocessing to form monitoring data. The condition monitoring front-end sends the monitoring data to the cloud server according to a set communication cycle. The cloud server stores the monitoring data, performs data preprocessing, and performs algorithm analysis to identify abnormal fault phases. The cloud server sends real-time monitoring data and abnormal fault identification results to the client terminal. Users can query historical data and abnormal phase identification results through the client terminal to obtain information about the mechanical faults in the high-voltage circuit breaker.

[0057] The specific steps are as follows:

[0058] 1. Audio signal preprocessing

[0059] like Figure 2 As shown, we process the recorded audio data directly at the state detection front end, reducing data dimensionality and lowering the cost of signal transmission. The specific process is as follows:

[0060] The first step is to pre-emphasize the speech signal. Since high-frequency signals often have less energy than low-frequency signals, we increase the proportion of high-frequency signal energy relative to low-frequency signal energy to effectively and fully utilize the high-frequency features during signal processing. For the digital signal obtained from sound sampling, the following operations are performed:

[0061] y(n)=x(n)-α·x(n-1)#(1)

[0062] The second step is frame segmentation. We need to perform a Fourier transform on the speech data to convert the information from the time domain to the frequency domain. However, if we perform an FFT on the entire speech segment, we will lose temporal information. Therefore, we assume that the frequency information remains unchanged within a very short time period t. Performing a Fourier transform on a frame of length t will give us a spectrum representing the time and frequency domain information of this sound signal. To ensure a smooth transition between frames, we use a frame length of 25ms and a frame shift of 10ms. Multiplying this by the signal sampling frequency fs gives us the number of sampling points and the number of overlapping sampling points for each frame.

[0063] The third step is windowing. For the signal obtained after framing, in order to eliminate spectral leakage, we set the value outside the window to 0. Here we use a Hamming window, the formula of which is as follows:

[0064]

[0065] Where n∈[1,N] and is an integer, and N is the number of sampling points corresponding to the above frame length. The W matrix is ​​then multiplied by the original signal matrix.

[0066] The fourth step is Fourier transform, DFT is performed on each frame of signal, and FFT is used for calculation to obtain a spectrogram representing the time domain and frequency domain information of the sound signal.

[0067] 2. Mel filtering is performed to extract Fbank features

[0068] For high-dimensional sound time-frequency spectrum signals, a filter bank simulating human ear perception is used to filter the original signal. After using this filter, the proportion of low-frequency signals in the signal is higher than that of high-frequency signals. The relationship between frequency value and mel scale is as follows:

[0069]

[0070]

[0071] We take 0 as the minimum value of the mel frequency and fs / 2 as the maximum value of the mel frequency, and bring them into formula (3) to obtain the maximum and minimum values of the mel scale Mmin and Mmax. Between (Mmin, Mmax), 34 mel scales are selected equally, and they are converted into frequency values by the inverse function of the above formula, and we obtain a 1*34 vector Fm. Scale transformation is performed on the Fm matrix, and the formula is as follows:

[0072]

[0073] The mel frequency values between the scales [0, N / 2] are obtained.

[0074] K = [k(1), k(2), …, k(33), k(34)], and the following formula is used to calculate the probability that N frequency components belong to 30 frequency bands, and the data of 32 filter banks are composed according to the probability. The formula is as follows:

[0075] H m (n) = 0, n < k(m-1)

[0076]

[0077]

[0078] H m (n) = 0, n ≥ k(m+1)

[0079] After obtaining the filter matrix Hm, it is multiplied by the time-frequency spectrum matrix to obtain the Fbank features based on the mel scale. The feature dimension is [m, n], where m is the number of mel filters and n is the number of frames contained in an audio data after framing.

[0080] 3. Training of neural network for identifying Fbank fault features

[0081] First, simulate the fault of high-voltage circuit breaker and record the fault sample, as the sample set for training neural network, we take five kinds of faults as examples here, such as closing spring fatigue, oil buffer oil leakage, energy storage motor voltage is too low, operating mechanism jamming, base instability, etc., among which the closing spring fatigue, operating mechanism jamming use the sound when closing for fault diagnosis; the closing spring fatigue, oil buffer oil leakage, operating mechanism jamming use the sound when opening for fault diagnosis, and the energy storage motor voltage is too low uses the sound when energy storage for fault diagnosis. For the five kinds of faults, the invention trains three neural network models with the same structure and different parameters to perform the final fault diagnosis.

[0082] For the sample set, the invention labels each fault sample with a fault category, and converts the label into one-hot encoding. Taking the fault sample when opening as an example, the sound signal when opening is used to diagnose the closing spring fatigue, oil buffer oil leakage and operating mechanism jambing, so the one-hot encoding converts the labels 1, 2 and 3 of the three fault samples into [1, 0, 0], [0, 1, 0] and [0, 0, 1], which will be used for subsequent comparison with the probability value output by the softmax layer during neural network training and calculation of loss value.

[0083] The structure of the neural network is designed, and for the time and frequency domain characteristics of the time-frequency spectrogram, convolutional layers and recurrent neural network layers are used for feature extraction, and fully connected neural networks and softmax layers are used for fault sample classification, and the overall structure is as shown in Figure 1 First, the Fbank feature is down-sampled and high-order feature extraction is performed using the convolutional layer (formula 6), and then the BN layer (formula 7) is input to perform batch normalization on all data after each layer of feature extraction, so as to improve the convergence speed of the network. Then, a nonlinear activation function (formula 8) is used to give the network better robustness, and a residual structure is used for connection, using Figure 2 modules and repeating multiple times.

[0084]

[0085]

[0086]

[0087] z t = σ(W z (x t )+ U z (h t-1 ))(update gate) # (9)

[0088] r t = σ(W r (xt )+U r (h t-1 ))(reset gate)

[0089]

[0090]

[0091] After obtaining the frame-level high-order feature representation with lower dimension and higher representation, the GRU gated recurrent neural network (formula 9) is used to extract the time domain information, and the network structure is generally as shown in Figure 3 , wherein the values of z t and r t are controlled by the activation function to be between 0 and 1. The closer the value of z t is to 1, the more important the state quantity passed from the previous time step is, and the closer the value of z t is to 0, the more important the hidden state value of this time step is. The closer the value of r t is to 1, the more important the value of the previous time step is to the hidden state value of this time step, and vice versa.

[0092] After obtaining the time-frequency domain high-order feature representation, the feature is input into the fully connected neural network (FCN) and the softmax classifier (formula 10) for the final classification of the fault sample. The function of the softmax classifier is to scale the input data to a value between 0 and 1, and use this value to fit the probability of an event occurring.

[0093]

[0094]

[0095] After setting the random parameters of the network, the labeled five-class fault sample data is fed into the neural network for fault sample classification, and the gradient descent learning is performed based on the cross-entropy loss function (formula 11). After convergence, the network parameters with the best recognition effect on the validation set are obtained as the final recognition network.

[0096] Taking the sound fault diagnosis at the time of switching as an example, for the audio data recorded in the front end, Fbank feature extraction is first performed, and then the trained neural network is input for feature extraction. Finally, a (5*1) size vector is obtained through the softmax layer, which is composed of five values in the range of 0-1, representing the probabilities of four faults and the normal device, respectively. The decision result with the maximum probability value is taken as the final fault diagnosis result.

[0097] 4. Use the trained model to perform fault diagnosis

[0098] When the audio data from the front end is obtained, the parameter k is obtained, if k=0, it is the closing action of the circuit breaker and the sound signal is transmitted, if k=1, it is the opening action of the circuit breaker and the sound signal is transmitted, if k=2, it is the action of the energy storage motor of the circuit breaker and the sound signal is transmitted. After logical judgment, different pre-trained models are used for fault diagnosis according to different sounds, and a vector with the same dimension as the fault category number is obtained, and the maximum value is the fault category corresponding to the position of the one-hot encoding.

[0099] The overall background cloud workflow is as shown in Figure 6

[0100] The application accurately judges whether the high-voltage circuit breaker has obvious mechanical faults by monitoring the sound signal of the high-voltage circuit breaker in action through non-contact monitoring and using artificial intelligence technology, so as to achieve the purpose of knowing the health status of the circuit breaker in time and checking the possible faults in advance, and provides a high-voltage circuit breaker fault online diagnosis method based on voiceprint recognition, which makes up for the deficiency of the existing detection means, fills the gap of the high-voltage circuit breaker online monitoring technology, and further ensures the safe and stable operation of the power distribution network in the jurisdiction.

[0101] It should be noted that the above content only illustrates the technical idea of the present application, and cannot limit the protection scope of the present application. For ordinary skilled persons in the technical field, some improvements and refinements can be made without departing from the principles of the present application, and these improvements and refinements fall within the protection scope of the claims of the present application.​

Claims

1. A method for online diagnosis of mechanical faults in high-voltage circuit breakers based on voiceprint recognition, characterized in that: It includes a sound sensor, a status monitoring front-end, a cloud server, and a client terminal. The sound sensor is installed in the switch cabinet housing and connected to the status monitoring front-end. The status monitoring front-end performs Mel filter processing on its signal and records audio data every time the circuit breaker is about to operate and sends it to the cloud server. The cloud server stores the monitoring data and performs data preprocessing and algorithm analysis to identify abnormal faults. The cloud server sends real-time monitoring data, as well as the results of abnormal fault identification of the circuit breaker body and operating mechanism, to the client terminal. Users can query historical data and fault identification results through the client terminal to obtain the fault status of the circuit breaker and arrange maintenance accordingly. Specifically, the following steps are included: (1) First, when the circuit breaker operates, the sound sensor picks up the sound of the equipment operating and transmits it to the status monitoring front end for pre-processing operations such as pre-emphasis, framing, windowing, and Mel filtering. For signal pre-emphasis, framing, windowing, and passing through a Mel filter bank: the formula for pre-emphasis of the sampled digital signal is: （1）； Where α is 0.95, n∈ is the set of integers whose maximum value is the time of the sound signal × the signal sampling rate sr and whose minimum value is 1; The formula for the window function when adding a window is: （2）； Where α is 0.46, n∈ is the set of integers with the maximum value being frame length × signal sampling rate sr and the minimum value being 1, and N is the frame length × signal sampling rate sr; The Mel filter bank is scaled with 32 equal divisions in the range [0, fm_max], and converted to Mel frequencies using the following formula, which is then projected onto the frequency range [0, N / 2]: （3）； （4）； in N is the frequency range selected for the Fourier transform, and fs is the sampling frequency. ; The formula for generating the Mel filter matrix is ​​as follows: ； （5）； ； ； The set of k values ​​is the Mel frequency vector generated by formulas (3) and (4) in the previous step; (2) Design a deep neural network that includes a high-order frequency domain feature extraction module, a time-series signal extraction module, and a softmax classification module. After training, it can be used to classify fault samples to achieve the effect of fault diagnosis. The frequency domain feature extraction module includes two modules, ResBlock and expansion-Resblock, which serve as the main structural components of the feature extraction process. （6）； （7）； （8）； Formula (6) is the formula for the conv layer. is the frequency domain feature of the upper layer for the range, and k is the learnable parameter corresponding to the feature; Formula (7) is the BN layer formula, X is the feature corresponding to all batches of each channel, and γ and β are learnable parameters; Formula (8) is the ReLU layer formula, where x is all feature values ​​after perception by the conv layer and normalization by the bn layer; The timing information extraction module in step (2) is shown in formula (9): （9）； ； ； ； in For the update gate in the GRU module, To reset the gate value, To hide the unit value, Enter the value for the next time step; The final classification module uses the FCN+softmax function to fit the probability of occurrence for each fault, where the softmax formula is as follows: ； Where w is a learnable parameter; (3) Simulate various fault scenes and record sound samples. Train a neural network model for sound samples of different actions and fault conditions, and use the trained model as the final fault diagnosis algorithm. The loss function used when training a neural network is the cross-entropy loss function, as shown in the following formula: ； in For the label values ​​of the training data, This represents the fitted probability value output by the softmax layer.

Images