Converter valve fault monitoring method based on non-contact microphone array

Abstract

The invention relates to the technical field of converter valve fault diagnosis, and provides a converter valve fault monitoring method based on a non-contact microphone array. The method comprises the following steps: collecting microphone array signals in a valve hall in real time; array signal noise reduction processing is carried out to improve the signal-to-noise ratio of signals; voice activity detection is carried out, abnormal sound signals are found out, and further fault judgment is carried out; and voiceprint feature extraction: extracting Mel spectrum features of the audio as model input, and constructing an improved deep residual shrinkage network model to accurately identify the fault type of the converter valve. The non-contact microphone array is adopted to collect sound signals in the converter valve hall, and the array signal noise reduction technology is matched, so that the problem that the sound receiving effect of a single microphone is poor is solved; meanwhile, the principle that Mel spectrum texture features of different sounds are different is utilized, voice recognition is converted into image classification, and the recognition accuracy of converter valve voiceprint faults is improved.

Description

Technical Field

[0001] The invention relates to the technical field of converter valve fault diagnosis, and in particular to a converter valve fault monitoring method based on a non-contact microphone array. Background Art

[0002] The converter valve is a key device in the UHV DC converter station, and it undertakes the functions of rectification and inversion in the entire UHV DC transmission project. The main electrical equipment contained in the converter valve includes thyristors, gate units, saturated reactors, damping capacitors and damping resistors. Failure of any component will cause a very serious production accident. Because the mechanical structure will change before the component is damaged, there will be abnormal sounds that are different from the sounds during normal operation. By detecting whether there is abnormal sound in the commutation component, it is possible to warn of the occurrence of faults in advance and issue an alarm, providing maintenance personnel with the opportunity to repair in time, avoiding potential greater accidents and dangers, and providing effective help.

[0003] Since the equipment inside the valve hall is all high-voltage, except for special microphones, general microphones can only be installed in the patrol corridor. Due to the long distance and high ambient noise, the signal-to-noise ratio of the sound of the converter valve collected by the microphone is not high, making it difficult to locate and analyze the fault of the converter valve; therefore, it is necessary to reduce the noise of the ambient sound to ensure effective monitoring of the fault phenomenon of the converter valve. At the same time, due to the complex noise of the collected sound signal, the accuracy of fault identification is severely challenged. Summary of the invention

[0004] The purpose of the present invention is to solve the above problems existing in the prior art and to provide a converter valve fault monitoring method based on a non-contact microphone array.

[0005] A method for monitoring a converter valve fault based on a non-contact microphone array comprises the following steps:

[0006] Step S1: microphone array signal acquisition: a microphone array is installed in the valve hall corridor, and the ambient sound in the valve hall corridor is collected by using the microphone array, and a corresponding sound array signal is generated. The ambient sound includes the sound generated during the operation of the converter valve and other noises;

[0007] Step S2: noise reduction of sound array signals: grouping the collected sound array signals and performing noise reduction processing on each group of sound array signals, filtering out ambient sounds and interference sounds other than those of the detection object, so as to obtain noise-reduced sound array preprocessing data, wherein the detection object is a running converter valve;

[0008] Step S3: Performing voice activity detection on the sound array preprocessing data, detecting the starting point of the abnormal sound signal in the sound array preprocessing data by using a voice activity detection algorithm, filtering out stable and worthless voice signals, so as to obtain the sound signal data of the current converter valve;

[0009] Step S4: Extracting voiceprint features from the sound signal data of the current converter valve: extracting the mel time-frequency spectrum of the sound signal data of the current converter valve as the current input data;

[0010] Step S5: Fault type identification: obtain the normal sound and abnormal sound during the operation of the converter valve to extract the benchmark mel spectrum features, use the benchmark mel spectrum features as training samples to train the deep residual shrinkage network model, and use the trained deep residual shrinkage network model to identify the current input data to determine whether the sound signal data of the current converter valve is abnormal sound signal data, and after determining that the sound signal data of the current converter valve is abnormal sound signal data, output a prompt message.

[0011] In the above-mentioned converter valve fault monitoring method based on non-contact microphone array, a non-contact microphone array is used to collect the sound signal in the converter valve hall, and the array signal noise reduction technology is used to solve the problem of poor sound collection effect of a single microphone. The principle of different Mel spectrum texture features of different sounds is used to convert speech recognition into an image classification problem. By constructing an improved deep residual shrinkage network model for fault identification, the recognition accuracy of converter valve voiceprint faults is improved. BRIEF DESCRIPTION OF THE DRAWINGS

[0012] Figure 1 The present invention is a flow chart of a converter valve fault monitoring method based on a non-contact microphone array.

[0013] Figure 2 It is a topological diagram of a microphone array in a converter valve fault monitoring method based on a non-contact microphone array of the present invention.

[0014] Figure 3 This is a network structure diagram of the identification model in the converter valve fault monitoring method based on the non-contact microphone array of the present invention. DETAILED DESCRIPTION

[0015] In order to enable those skilled in the art to better understand the technical solutions in the present application, the technical solutions in the embodiments of the present application will be clearly and completely described below in conjunction with the drawings in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, not all of the embodiments. Based on the embodiments in the present application, all other embodiments obtained by ordinary technicians in this field without creative work should fall within the scope of protection of this application.

[0016] like Figure 1-3 As shown, the present application discloses a converter valve fault monitoring method based on a non-contact microphone array, comprising the following steps:

[0017] Step S1: microphone array signal acquisition: a microphone array is installed in the corridor of the valve hall, and the ambient sound in the corridor of the valve hall is collected in real time by using the microphone array, and a corresponding sound array signal is generated. The ambient sound includes the sound and other noises generated during the operation of the converter valve. Among them, according to the different spatial dimensions, the microphone array topology structure can be divided into three categories: one-dimensional linear array, two-dimensional planar array and three-dimensional space array. In step S1, the microphone array is arranged in the corridor of the converter valve hall using a two-dimensional planar circular array, wherein the array z-axis faces the converter valve.

[0018] Step S2: noise reduction of sound array signal: noise reduction processing is performed on the collected sound array signal transmitted in the form of data frames, and the ambient sound and interference sound of non-detection objects are filtered out to obtain noise-reduced sound array preprocessing data, wherein the detection object is a running converter valve.

[0019] Step S3: Perform voice activity detection on the sound array preprocessing data, detect the starting point of the abnormal sound signal in the sound array preprocessing data by using the voice activity detection algorithm, filter out the stable and worthless voice signals, so as to obtain the sound signal data of the current converter valve.

[0020] Step S4: Extracting voiceprint features from the sound signal data of the current converter valve: extracting the mel time-frequency spectrum of the sound signal data of the current converter valve as the current input data; wherein, the librosa library or the torchaudio library is used to extract the mel time-frequency spectrum of the sound signal data of the current converter valve, and the obtained time-frequency spectrum corresponding to each frame is scaled to a figure with consistent length and width for use by the deep residual shrinkage network model.

[0021] Step S5: Fault identification: obtain the normal sound and abnormal sound during the operation of the converter valve to extract the benchmark mel spectrum features, use the benchmark mel spectrum features as training samples to train the deep residual shrinkage network model, use the trained deep residual shrinkage network model to identify the current input data to determine whether the sound signal data of the current converter valve is abnormal sound signal data, and output prompt information after determining that the sound signal data of the current converter valve is abnormal sound signal data. Among them, the network of the deep residual shrinkage network model is composed of an input layer, a convolution layer, an improved residual shrinkage unit, a batch normalization layer, an adaptive parameter corrected linear unit, a global average pooling layer and a fully connected layer. The structure: the input layer is the current input data input, which is a 224*224 mel spectrum here; the convolution layer is composed of several convolution kernels, and the convolution kernel is used to extract features from the input feature map; the batch normalization (Batch Normalization, referred to as BN) layer is a commonly used technology in deep learning. By standardizing the input of each layer, the input data distribution of each layer is relatively stable, so that a larger learning rate can be used for training, which speeds up the convergence speed and reduces the internal covariate offset (Internal Covariate Shift), which makes the model more adaptable to different input data, improves the generalization ability, helps to alleviate the problems of gradient disappearance and gradient explosion, and makes the training process more stable; adaptive parameter correction linear unit, activation function, add nonlinear ability to the network to improve the network expression ability; residual shrinkage unit: in practical applications, data is often affected by noise, interference or outliers, which may reduce the performance of deep learning models. Traditional neural network structures may find it difficult to effectively deal with these problems because they usually assume that the data is clean and have no special mechanism to deal with noise and outliers. The residual shrinkage unit is proposed to solve the above problems. It can automatically identify and suppress noise and outliers in the input data by introducing attention mechanism and soft thresholding operation, thereby improving the robustness and generalization ability of the model; the pooling layer follows the convolution layer to prevent overfitting; the fully connected layer collects and classifies the features extracted by the previous convolution layer and pooling layer, and classifies the output values ​​with the Sofimax function.

[0022] Furthermore, the method further includes the following steps: after determining that the sound signal data of the current converter valve is abnormal sound signal data, searching for a corresponding fault type from a pre-stored sample library, and displaying the found fault type.

[0023] Furthermore, the method further includes the following steps: after determining that the sound signal data of the current converter valve is abnormal sound signal data, searching for a corresponding fault type from a pre-stored sample library, and when the corresponding fault type is not found, prompting manual judgment and input of the fault type;

[0024] After the current input data is matched with the manually input fault type, it is stored in the sample library.

[0025] In this implementation, considering that most noises in the actual environment are additive noises, and the research on additive noises is the most common at present, and additive noises can be roughly divided into periodic noises, impulse noises, broadband noises and co-channel speech interferences, in step S2, a beam forming algorithm is used to perform noise reduction processing on the array signal, and the beam forming algorithm combines and processes the multi-channel microphone signals, and uses the sound array signal to turn the microphone array to the target direction, so as to enhance the sound signal in the target direction and suppress the interference signal in the non-target direction, instead of physically moving the microphone array. The specific steps are as follows:

[0026] Step S201: construct a far-field acoustic model. Assume that there are M microphones and the received data of the i-th microphone is x i (t), i = 1, 2…, M, and the far-field sound source is s(t), then the output signal can be expressed as:

[0027] x i (t) = s(t-δ i )

[0028] Among them, δ i is the time delay of the ith microphone relative to the center of the array. The microphone is converted to x after digital-to-analog conversion. i (n); the output signal after beamforming can be expressed as:

[0029]

[0030] where a i，k is the k-th order weight of the filter of the ith generator, y i (n) is the output of the i-th channel after passing through the filter;

[0031] Step S202: Find the optimal filter coefficient and convert the output signal of S201 into the frequency domain, which is expressed as:

[0032] Y(e jω )=F(θ,ω)S(e jω )

[0033]

[0034] Where F(θ, ω) is the transmission response function of the beamforming algorithm at angle θ, θ is the arrival direction angle from the sound source to the array, and a is the optimal filter coefficient. The minimum mean square error criterion is used to approximate F(θ, ω) to the desired beam pattern B(θ):

[0035]

[0036] in, is the angle of deviation from the desired direction, the above formula can be rewritten as:

[0037]

[0038]

[0039] z(θ,ω)=[z1(θ,ω),...,z M (θ, ω)]

[0040] Among them, H represents the conjugate transpose operation. From the above formula, it can be seen that the transmission response function is a linear combination of the filter coefficients. Therefore, the integral optimal solution of the least mean square can be converted into the optimal solution of numerical calculation:

[0041]

[0042] Among them, N and L are the dimensions of spatial domain and frequency division respectively, and the optimal filter coefficient finally solved is:

[0043]

[0044] in,() -1 Represents the inverse matrix.

[0045] In the above steps S201 and S202, the input y i (n) is multi-channel data, and the output signal y(n) is the combined signal. a is actually an array of filter coefficients, and the coefficients in different directions are different. In this application, the z-axis of the array points to the converter valve, which is actually a beam in a fixed direction. By selecting the beam coefficient facing the converter valve, signals in other directions can be suppressed, thereby filtering out sound signals in directions other than the converter valve, that is, noise interference signals.

[0046] In this embodiment, the voice activity detection is only used to detect whether an abnormal sound source different from the ambient background sound appears. Therefore, it is not necessary to accurately identify the starting point and the ending point of the signal. It is only necessary to find the signal different from the normal signal. Therefore, in step S3, the signal frequency band energy detection method is used to detect whether an abnormal sound signal appears in the sound array preprocessing data. The specific steps are as follows:

[0047] Step S301: converting a frame of sound array pre-processed data subjected to noise reduction in step S2 into the frequency domain;

[0048] Step S302: Calculate the energy of the entire frequency band and the energy of each frequency point in the frequency domain;

[0049] Step S303: Compare the calculated two frequency band energies with the pre-stored frequency band energy of the background noise. If the energy exceeds the set threshold, it is determined to be a foreign sound. The expression is as follows:

[0050]

[0051]

[0052] Where X(f) is the spectrum of a frame of time domain sampling signal, N(f) is the background noise spectrum, K is the frequency point number of full-band data, Ki is the frequency point number of each sub-band data, T fall is the full-band detection result, T fi is the detection result of each frequency point. In order to avoid missing abnormal sound, if any of the two exceeds the threshold, it is determined that the signal has abnormal sound; if it is determined to be abnormal sound, the next step is performed; if it is determined to be normal, it returns to step S1 to process the next frame signal.

[0053] In this way, the data noise reduction technology of the array signal is used to remove the non-converter valve ambient sound interference in the valve hall and suppress the background noise of the microphone itself, thereby increasing the sound collection effect for the converter valve, filtering the steady signal through the voice activity detection algorithm, and reducing computing resources; by converting the sound signal into a Mel-time spectrum image to replace the sound signal as the input of the model, the model has strong interpretability due to the different textures of different sounds on the Mel spectrum; an improved deep residual shrinkage network is constructed, and by tuning the model, accurate recognition of the ambient sound and fault sound of the converter valve can be achieved. Therefore, the present application can realize early warning and fault diagnosis of converter valve faults, thereby improving the stable operation and reliability of the converter valve.

Claims

1. A method for monitoring a converter valve fault based on a non-contact microphone array, characterized in that: The following steps are included: Step S1: microphone array signal acquisition: a microphone array is installed in the valve hall corridor, and the microphone array is used to collect the ambient sound in the valve hall corridor in real time, and generate a corresponding sound array signal. The ambient sound includes the sound generated during the operation of the converter valve and other noises; Step S2: performing noise reduction processing on the sound array signal collected and transmitted in the form of data frames, filtering out ambient sound and interference sound other than the detection object, so as to obtain noise-reduced sound array preprocessing data, wherein the detection object is the running converter valve; Step S3: Performing voice activity detection on the sound array preprocessing data, detecting the starting point of the abnormal sound signal in the sound array preprocessing data by using a voice activity detection algorithm, filtering out stable and worthless voice signals, so as to obtain the sound signal data of the current converter valve; Step S4: Extracting voiceprint features from the sound signal data of the current converter valve: extracting the mel time-frequency spectrum of the sound signal data of the current converter valve as the current input data; Step S5: Fault identification: obtain the normal sound and abnormal sound during the operation of the converter valve to extract the benchmark mel spectrum features, use the benchmark mel spectrum features as training samples to train the deep residual shrinkage network model, use the trained deep residual shrinkage network model to identify the current input data to determine whether the sound signal data of the current converter valve is abnormal sound signal data, and output a prompt message after determining that the sound signal data of the current converter valve is abnormal sound signal data.

2. The converter valve fault monitoring method based on non-contact microphone array according to claim 1 is characterized in that: The following steps are also included: After determining that the sound signal data of the current converter valve is abnormal sound signal data, the corresponding fault type is searched from the pre-stored sample library, and the found fault type is displayed.

3. The converter valve fault monitoring method based on non-contact microphone array according to claim 1 or 2, characterized in that: The following steps are also included: After determining that the sound signal data of the current converter valve is abnormal sound signal data, searching for the corresponding fault type from the pre-stored sample library, and when the corresponding fault type is not found, prompting manual judgment and input of the fault type; After the current input data is matched with the manually input fault type, it is stored in the sample library.

4. The converter valve fault monitoring method based on non-contact microphone array according to claim 3 is characterized in that: In step 1, the microphone array is arranged in a corridor in the converter valve hall using a two-dimensional planar circular array, wherein the array z-axis faces the converter valve.

5. The converter valve fault monitoring method based on non-contact microphone array according to claim 3 is characterized in that: In step S2, a beamforming algorithm is used to perform noise reduction processing on the array signal. The beamforming algorithm combines the multi-channel microphone signals and uses the sound array signal to turn the microphone array to the target direction, thereby enhancing the sound signal in the target direction and suppressing the interference signal in the non-target direction. The specific steps are as follows: Step S201: construct a far-field acoustic model. Assume that there are M microphones and the received data of the i-th microphone is x i (t), i = 1, 2..., M, the far-field sound source is s(t), then the output signal can be expressed as: x i (t)=s(t-δ i ) Among them, δ i is the time delay of the ith microphone relative to the center of the array; the microphone is converted to x after digital-to-analog conversion i (n); the output signal after beamforming can be expressed as: where a i,k is the k-th order weight of the filter of the i-th generator, y i (n) is the output of the i-th channel after passing through the filter; Step S202: Find the optimal filter coefficient and convert the output signal of S201 into the frequency domain, which is expressed as: Y(e jω )=F(θ,ω)S(e jω ) Where F(θ,ω) is the transmission response function of the beamforming algorithm at angle θ, θ is the arrival direction angle from the sound source to the array, and a is the optimal filter coefficient. The minimum mean square error criterion is used to approximate F(θ,ω) to the desired beam pattern B(θ): in, is the angle of deviation from the desired direction, the above formula can be rewritten as: z(θ,ω)=[z1(θ,ω),...,z M (θ,ω)] Among them, H represents the conjugate transpose operation. From the above formula, it can be seen that the transmission response function is a linear combination of the filter coefficients. Therefore, the integral optimal solution of the least mean square can be converted into the optimal solution of numerical calculation: Among them, N and L are the dimensions of spatial domain and frequency division respectively, and the optimal filter coefficient finally solved is: in,() -1 Represents the inverse matrix.

6. The converter valve fault monitoring method based on non-contact microphone array according to claim 3 is characterized in that: In step 3, a signal frequency band energy detection method is used to detect whether an abnormal sound signal appears in the sound array preprocessing data. The specific steps are as follows: Step S301: converting a frame of sound array pre-processed data subjected to noise reduction in step S2 into the frequency domain; Step S302: Calculate the energy of the entire frequency band and the energy of each frequency point in the frequency domain; Step S303: Compare the calculated two frequency band energies with the pre-stored frequency band energy of the background noise. If the energy exceeds the set threshold, it is determined to be a foreign sound. The expression is as follows: Where X(f) is the spectrum of a frame of time domain sampling signal, N(f) is the background noise spectrum, K is the frequency point number of full-band data, Ki is the frequency point number of each sub-band data, T fall is the full-band detection result, T fi is the detection result of each frequency point. In order to avoid missing abnormal sound, if any of the two exceeds the threshold, it is determined that the signal has abnormal sound; if it is determined to be abnormal sound, the next step is performed; if it is determined to be normal, it returns to step S1 to process the next frame signal.

7. The converter valve fault monitoring method based on non-contact microphone array according to claim 3 is characterized in that: In step S4, the librosa library or the torchaudio library is used to extract the mel time-frequency spectrum of the sound signal data of the current converter valve, and the obtained time-frequency spectrum corresponding to each frame is scaled to a graph with consistent length and width for use by the deep residual shrinkage network model.

8. The converter valve fault monitoring method based on non-contact microphone array according to claim 3 is characterized in that: The network of the deep residual shrinkage network model consists of an input layer, a convolutional layer, an improved residual shrinkage unit, a batch normalization layer, an adaptive parameter corrected linear unit, a global average pooling layer and a fully connected layer.

Images