Converter valve working condition diagnosis method, device and equipment based on voiceprint feature extraction

Abstract

The invention discloses a converter valve working condition diagnosis method, device and equipment based on voiceprint feature extraction, and relates to the technical field of high-voltage direct-current converter valve working condition diagnosis. The method comprises the following steps: acquiring original audio data of an operation state of the high-voltage converter valve, and carrying out framing, windowing and normalization preprocessing on a collected sound signal; separating various source signals in the mixed sound signals by using a blind source separation algorithm to obtain abnormal sound signals of the high-voltage converter valve; respectively extracting MFCC features and wavelet packet energy spectrums of the sound signals of the high-voltage converter valve by adopting a joint feature extraction method, and dynamically adjusting parameters through an improved zebra optimization algorithm ZAO; and inputting the extracted combined optimization features into a convolutional neural network, and judging the fault type of the abnormal operation state of the high-voltage converter valve. According to the method, the reliability of voiceprint feature extraction of the high-voltage converter valve and the accuracy of operation state evaluation are improved, the robustness is high, and possible potential safety hazards are avoided.

Description

Technical Field

[0001] This invention belongs to the field of high voltage DC converter valve operating condition diagnosis technology, and in particular relates to a converter valve operating condition diagnosis method, device and equipment based on acoustic feature extraction. Background Technology

[0002] High-voltage converter valves are core equipment in high-voltage direct current (HVDC) transmission systems, primarily responsible for converting electrical energy between AC and DC. Their operating status directly affects the safety and stability of the power system. With the increasing complexity of power systems and the expanding application scenarios of high-voltage converter valves, higher demands are being placed on monitoring their operating status and providing early warning of faults.

[0003] Traditional methods for monitoring the operating status of high-pressure converter valves primarily rely on real-time monitoring of electrical parameters (such as voltage, current, and temperature). However, these methods often struggle to detect potential faults promptly and accurately, especially for early, subtle anomalies, failing to provide effective warnings. In recent years, with the development of signal processing technology, acoustic signature extraction technology has gradually become an effective means of monitoring the status of high-pressure equipment. Acoustic signals can reflect physical phenomena such as mechanical vibration, structural deformation, and abnormal discharge during equipment operation. By analyzing the acoustic characteristics of converter valves under different operating conditions, their working status can be assessed non-invasively, offering advantages of speed and sensitivity. However, research on acoustic signature extraction technology for high-pressure converter valves is relatively limited, and a highly effective technical system has not yet been established.

[0004] Based on this, the present invention proposes a method and device for diagnosing the operating conditions of high-voltage DC converter valves based on acoustic signal feature extraction. It extracts acoustic features that can characterize the operating status of high-voltage converter valves and performs fault classification and diagnosis through convolutional neural networks, providing a non-contact, reliable, and efficient means of condition monitoring and fault early warning to make up for the shortcomings of existing technologies. Summary of the Invention

[0005] This invention provides a method, apparatus, and equipment for diagnosing the operating conditions of converter valves based on acoustic signature extraction. The purpose is to overcome the shortcomings of existing high-pressure converter valve operating condition diagnosis technologies, such as "poor robustness of feature extraction, low accuracy of fault classification, and strong subjectivity of parameter adjustment." It provides a non-contact, highly reliable, and highly accurate operating condition diagnosis solution, improves the reliability of acoustic signature extraction of high-pressure converter valves and the accuracy of operating status assessment, and provides technical support for the safe and stable operation of converter valves.

[0006] To solve the above-mentioned technical problems, the present invention is achieved through the following technical solution:

[0007] The converter valve operating condition diagnosis method based on acoustic signature feature extraction of the present invention adopts the following technical solution, including the following steps:

[0008] S1: Acquire the raw audio data of the high-pressure converter valve's operating status, and perform preprocessing such as frame segmentation, windowing, and normalization on the acquired signal;

[0009] S2: Using a blind source separation algorithm, various source signals in the mixed sound signal are separated to obtain the abnormal sound signal of the high-pressure converter valve;

[0010] S3: The MFCC features and wavelet packet energy spectrum of the high-pressure converter valve sound signal are extracted by a joint feature extraction method, and the parameters are dynamically optimized by the improved Zebra Optimization Algorithm (ZAO).

[0011] S4: Input the extracted combined features into a pre-trained convolutional neural network to determine the type of fault in the abnormal operating state of the high-pressure converter valve.

[0012] Optionally, in step S1, the preprocessing of the high-pressure converter valve acoustic signal is performed by framing, windowing, and normalizing the original acoustic signal.

[0013] Optionally, in step S2, the pre-judged mixed signal is subjected to blind source separation. The Fast Independent Component Analysis (FastICA) method is used, which is based on the principle of maximizing negative entropy. The objective function is iteratively optimized to achieve rapid extraction of each independent source signal.

[0014] In this process, after the blind source is separated, the kurtosis, a feature index with obvious distinguishability, is used to identify the abnormal signals after separation and extract the abnormal fault signals of the high-pressure converter valve.

[0015] Optionally, in step S3, the wavelet packet method is used to decompose and extract the signal energy spectrum feature parameters. The specific method is as follows:

[0016] A1: First, decompose the abnormal sound signal of the high-pressure converter valve;

[0017] A2: Next, calculate the energy in different frequency ranges;

[0018] A3: Finally, construct the energy spectrum characteristic parameters.

[0019] Optionally, in step S3, the MFCC characteristic parameters of the abnormal acoustic signal of the high-pressure converter valve after blind source separation are obtained by performing preprocessing, framing, windowing, FFT, power spectrum calculation, Mel filtering, logarithmic calculation, DCT transformation and other steps to obtain Mel frequency cepstral coefficients that can represent the spectral characteristics of the acoustic signal.

[0020] Optionally, step S3 further includes: weighting and merging the two feature parameters to form a mixed feature matrix of the high-pressure converter valve acoustic signal, and dynamically adjusting parameters such as the number of wavelet packet decomposition layers, MFCC pre-emphasis coefficients, number of Mel filters, and weighting of MFCC and wavelet packet energy spectra according to different operating conditions using an improved zebra optimization algorithm.

[0021] Optionally, step S4 specifically includes:

[0022] The combined and optimized sound characteristics of the high-pressure converter valve are used as network input. There are four types of faults: heavy overload, loose clamps, DC bias, and short circuit impact.

[0023] Pre-trained Convolutional Neural Networks (CNNs) include convolutional layers, pooling layers, ECA channel attention mechanism, Flatten layers, and fully connected layers.

[0024] The Softmax function for classification prediction: gives the predicted probability of the current sample belonging to each type of fault and outputs the judgment result.

[0025] A high-voltage DC converter valve operating condition diagnostic device based on acoustic signal feature extraction, comprising:

[0026] The data acquisition and preprocessing module acquires the sound signal of the high-pressure converter valve during operation and performs preprocessing such as framing, windowing and normalization on the acquired sound signal.

[0027] The abnormal signal extraction module classifies various source signals using a blind source separation algorithm, and further filters and identifies them using kurtosis index to extract the fault sound signal of the high-pressure converter valve.

[0028] The feature extraction module extracts the combined MFCC and wavelet packet energy spectrum features of the high-pressure converter valve fault sound signal, and dynamically adjusts the parameters through an improved ZOA algorithm.

[0029] The fault classification module uses a pre-trained convolutional neural network to receive combined feature parameters as input and classify fault types.

[0030] An electronic device includes a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the program, it implements the steps of the high-voltage DC converter valve condition diagnosis method based on acoustic signal feature extraction.

[0031] A non-transitory computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the high-voltage DC converter valve condition diagnosis method based on acoustic signal feature extraction.

[0032] The present invention has the following advantages over the prior art:

[0033] (1) This invention utilizes an improved ZOA algorithm to dynamically optimize parameters, significantly enhancing the sensitivity of features to abnormal equipment states. This algorithm effectively avoids the subjectivity and limitations of manual parameter tuning by iteratively searching for the optimal parameter combination. Based on a balance mechanism between global search and local development, ZOA can accurately locate the optimal solution adaptable to different working conditions in a complex parameter space, thereby enhancing the robustness of feature extraction and reducing the risk of overfitting;

[0034] (2) This invention constructs a fusion model of convolutional neural network and ECA channel attention mechanism to achieve accurate classification of four types of faults. ECA adaptively enhances the channel weights corresponding to fault features (such as high-frequency discharge bands and low-frequency vibration components) through a local cross-channel interaction mechanism, suppresses environmental noise interference, and improves the signal-to-noise ratio. This lightweight module enhances feature expression with extremely low computational overhead, and combined with the feature extraction capabilities of CNN, provides scientific and reliable diagnostic guidance for converter valve faults and improves the accuracy of condition assessment;

[0035] (3) This invention helps reduce the manual cost of detecting the operating status of high-pressure converter valves and provides a non-contact method for monitoring the status of high-pressure converter valves. It is convenient to use and has high applicability.

[0036] Of course, any product implementing this invention does not necessarily need to achieve all of the advantages described above at the same time. Attached Figure Description

[0037] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0038] Figure 1 This is a schematic diagram of the high-pressure converter valve acoustic signature extraction method according to an embodiment of the present invention;

[0039] Figure 2 This is a schematic diagram of the wavelet packet energy spectrum feature parameter extraction process;

[0040] Figure 3 This is a schematic diagram of the MFCC feature parameter extraction process;

[0041] Figure 4 This is a flowchart illustrating the Zebra optimization algorithm.

[0042] Figure 5 This is a structural diagram of the fault classifier model in an embodiment of the present invention. Detailed Implementation

[0043] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0044] like Figure 1-5 As shown, this invention provides an embodiment of a method for extracting acoustic signature features of a high-pressure converter valve during operation.

[0045] Figure 1 This is a schematic flowchart of the method for extracting acoustic signature features of a high-pressure converter valve according to an embodiment of the present invention, as follows: Figure 1 As shown, the process includes the following steps:

[0046] Step S1: Obtain the raw audio data of the high-pressure converter valve's operating status, and perform preprocessing such as frame segmentation, windowing, and normalization.

[0047] Specifically, step S1 above includes:

[0048] Step S11: A non-contact, wall-mounted microphone is used near the high-pressure converter valve to acquire the raw audio data of the valve's operating status. The sampling rate is 48kHz, and the data is represented as follows:

[0049] S = [s1, s2, ..., s N ]

[0050] Among them, s N Let N be the signal amplitude at the i-th sampling point; N is the number of sampling points.

[0051] Step S12: Frame segmentation is performed on the original audio signal. Frame segmentation involves shifting and weighting the non-stationary audio signal through a window of a specific length, thereby dividing it into short-term stationary signal segments. An overlapping segmentation method is used, meaning that there is a certain overlap between the previous and subsequent frames. The frame shift relationship is as follows:

[0052]

[0053] Where M is the number of frames; L is the frame length; N is the length of the audio signal, which is a natural number in the range of [0, L-1]; and a is the overlap rate.

[0054] Step S13 involves windowing each signal segment (the windowing operation is performed on data frames, and the window length is equal to the frame length) to mitigate the impact of spectral leakage. This invention selects a Hamming window with good time and frequency aggregation characteristics; the window function is:

[0055]

[0056] Where L is the length of the Hamming window, determined by the frame length and sampling rate; in the above formula, 0.54, 0.46, and π are inherent parameters of the Hamming window, designed to form the window shown in the figure below. The purpose of windowing is to prevent spectral leakage during FFT.

[0057] Step S14: The acquired sound signal is processed into a data frame unit of L (in this embodiment, L = 2000) consecutive data points, and normalized to x. n , represented as:

[0058]

[0059] Among them, s 2000*m+1 ,s 2000*m+2 ,...,s 2000*(m+1) These are the signal amplitudes at the 2000*m+1, 2000*m+2, and 2000*(m+1) sampling points, respectively, where m represents the m-th data frame into which the sound signal is divided, and max{} is the maximum value function used to take the maximum value of the elements in the sequence.

[0060] Step S2 involves performing blind source separation on the pre-judged abnormal signal to separate various source signals from the mixed acoustic signal. Blind source separation, as a signal processing method, can separate various source signals from the aliased signal based on the independence of each sound source.

[0061] Specifically, such as Figure 2 As shown, step S2 above includes:

[0062] Step S21 involves using the FastICA method to perform blind source separation on the pre-judged mixed signal, enabling rapid extraction of each independent source signal. It is worth noting that the separation operation can more effectively extract characteristic parameters of abnormal signals, minimizing the impact and interference of background noise from the high-pressure converter valve operation.

[0063] Step S22 involves further identifying and filtering the target acoustic signal, finding abnormal acoustic signals from each separated signal, and utilizing kurtosis, a feature index with significant discriminative power, as follows:

[0064]

[0065] Kurtosis is calculated using one frame of data, where Xn represents the nth value in the frame. n is a natural number ranging from [0, L]. Kurtosis reflects the non-Gaussian nature of a signal. Normal signals typically approximate a Gaussian distribution, while fault signals often contain spikes or outliers, exhibiting significant non-Gaussian characteristics. In a specific embodiment, a kurtosis threshold K = 3 is set to filter out abnormal signals.

[0066] Step S3: Obtain the MFCC characteristic parameters of the abnormal acoustic signal of the high-pressure converter valve after blind source separation, and use wavelet packet decomposition to extract the signal energy spectrum characteristic parameters.

[0067] Specifically, the process of extracting signal energy spectrum feature parameters using wavelet packet decomposition is as follows:

[0068] Step S311: Taking a three-layer example, perform three-layer wavelet packet decomposition on the signal; extract the signal features of the eight frequency component intervals from low to high in the third layer;

[0069] Step S312: Sum the energies of different intervals to obtain the energies of 8 frequency intervals;

[0070] Step S313: Construct a new energy spectrum feature vector and divide the energy of each interval by the total energy to obtain the energy percentage;

[0071] Combining the above steps and Figure 2 The specific process involves dynamically adjusting the number of decomposition layers according to different operating conditions during implementation to obtain characteristic parameters that characterize the acoustic signal properties of the high-pressure converter valve. The specific wavelet basis selection and the number of layers need to be determined based on the characteristics of the specific fault type. For example, short-circuit impulse is characterized by impulse features, which manifest as drastic amplitude changes in a short period of time in the time domain and contain broadband components in the frequency domain, but the energy is concentrated in a specific time period. The coif3 wavelet basis or db6 wavelet basis is suitable.

[0072] The selection of wavelet bases for different fault signals needs to be dynamically adjusted based on the model's performance and is not fixed.

[0073] Specifically, in combination Figure 3 As shown, the process for extracting the MFCC characteristic parameters of the acoustic signal from the high-pressure converter valve is as follows:

[0074] Step S321: Perform pre-emphasis processing on the signal.

[0075] For example, pre-emphasis is generally achieved using a first-order high-pass filter, and the expression for the pre-emphasized signal b(n) is:

[0076] b(n)=a(n)-αa(n-1),1<n<L;

[0077] Where a(n) is the sound signal at the current moment, a(n-1) is the sound signal at the previous moment, and α is the pre-emphasis coefficient, usually taken as 0.94 to 0.97; this formula means to compensate for high frequency attenuation, equalize the signal spectrum, improve the signal-to-noise ratio, and reduce distortion in signal transmission.

[0078] Step S322: Perform endpoint detection on the pre-emphasized signal, determine the start point and end frame of the input signal, and monitor the range of the valid input signal.

[0079] For example, suppose the high-pressure converter valve signal at a short moment is a(n) with length L, and its energy P is calculated using the following formula:

[0080]

[0081] Its zero-crossing rate Q is:

[0082]

[0083] Where sgn[·] is the sign function; n is the index of the 0th to Lth frame of data, i.e., a natural number in the range [0, L-1]; a(n) represents the nth number in a frame. In signal processing, the zero-crossing rate (ZCR) is an important statistic describing signal characteristics. It measures the number of times a signal crosses the zero line from the positive direction to the negative direction, or from the negative direction to the positive direction, per unit time (or per unit number of samples). It essentially reflects the "activity" of the signal, and is especially sensitive to high-frequency components or abrupt changes in signals.

[0084] Step S323: Perform a frame segmentation operation on the signal, dividing a complete high-voltage converter valve acoustic signal into several sub-frame signals, and then processing each sub-frame signal.

[0085] Step S324: Windowing is applied to each frame of signal.

[0086] For example, let the frame signal be a(n), the window function be h(n), and the windowed signal be Ah(n), then we get:

[0087] Ah(n)=a(n)h(n),0≤n≤L-1;

[0088] In this invention, a Hamming window is used for windowing to mitigate the impact of spectral leakage. Step S325 involves performing a Fast Fourier Transform on each frame of the signal.

[0089] Q(i,k)=FFT[ah m (n)]

[0090] Where i and k are general descriptions of complex numbers, and Q obviously represents the result of the Fourier transform;

[0091] Furthermore, calculate its spectral line energy, and then square the modulus:

[0092] E(i,k)=|Q(i,k)| 2

[0093] Step S326: Perform Mel frequency filtering on the signal.

[0094] For example, the frequency of each frame is obtained, and the energy of the filter is essentially the spectral energy E(i,k) multiplied by the corresponding value I in the frequency domain of the filter within the frequency range. s (k) and add them together:

[0095]

[0096] Where s is the number of filters, and S is dynamically adjusted according to the actual operating conditions.

[0097] Step S327: First, take the logarithm of the calculated energy, then use DCT to transform the frequency spectrum factor to the time domain. The resulting cepstral coefficients are the MFCC coefficients.

[0098]

[0099] In the above formula, cos is the cosine transform, which involves performing a discrete cosine transform on the logarithmic energy. D refers to the order of the MFCC coefficients, typically between 12 and 16. For example, after processing the high-pressure converter valve signal using the wavelet packet method, each frame of energy spectrum extracted represents a 32-dimensional feature vector, while after MFCC extraction, each frame of cepstral features represents a 40-dimensional feature vector. The wavelet packet energy spectrum and MFCC feature parameters of the high-pressure converter valve are arranged in ascending order, and the first 10 feature parameters are taken as the main feature parameters of the signal. The two features are then fused using a weighted method, as shown below:

[0100] F=δMFCC+βWPD

[0101] Where F represents the fusion feature, δ and β are the hyperparameters controlling the MFCC feature and the wavelet packet energy spectrum feature WPD, respectively, and δ+β=1.

[0102] Furthermore, the improved Zebra Optimization Algorithm (ZOA) is used to find the optimal hyperparameters based on actual working conditions. The ZOA is a metaheuristic optimization algorithm inspired by zebra herd behavior. It solves complex optimization problems by simulating the foraging, defense, and migration behaviors of zebra groups. The process is as follows: Figure 4 As shown.

[0103] In this embodiment, the improved ZOA algorithm optimizes the hyperparameters of the feature extraction part, and the steps are as follows:

[0104] 1. Initialization: Determine the optimization objective as the optimal combination of feature parameters, set parameters such as the initial population size and the maximum number of iterations, and set the current best individual as "Pioneer Zebra (PZ)".

[0105] 2. Fitness calculation: Calculate the fitness value of each individual in the population based on the feature extraction effect (fault identification accuracy), and select the better value to update the individual position.

[0106] 3. Dynamic policy update: A dynamic probability P is generated based on the number of iterations. If the escape policy is triggered, the position is updated through the Lévy flight distribution function (enhancing global search); if the attack policy is triggered, the "attract-disturb" policy is used to converge toward the optimal solution (strengthening local exploitation), while a chord function term is introduced to avoid local optima.

[0107] 4. Iteration Termination: Update the optimal value of the population in real time. When the maximum number of iterations is reached, output the optimized hyperparameters.

[0108] Step S4 involves inputting the extracted combined features into a pre-trained convolutional neural network to determine the type of fault in the abnormal operating state of the high-pressure converter valve. There are four types of fault sounds: heavy overload, loose clamps, DC bias, and short-circuit impact.

[0109] Step S41: Pre-train a convolutional neural network for further feature extraction and processing to identify the fault types of abnormal signals in the high-pressure converter valve. The root mean square loss function is used, and the network parameters are updated using stochastic gradient descent. The training process is iterated until the training loss function fully converges, and the model parameters are saved. The loss function is represented using cross-entropy loss.

[0110] Step S42: Input the combined feature matrix from step S3 into the pre-trained convolutional neural network for classification and prediction to determine the fault category of the abnormal signal of the high-pressure converter valve.

[0111] Specifically, Figure 5 This is a diagram of the convolutional neural network structure, which includes two convolutional pooling layers, ECA channel attention, one flattened layer, and two dense layers. The convolutional layers use filters with shared weights to extract features from the input matrix, employing one-dimensional convolution with a 3*10 kernel. The pooling layers select and filter the extracted features, downsampling to reduce feature dimensionality while retaining key information, using max pooling with a 3×3 window size and a stride of 2. In practical processing of high-pressure converter valve acoustic signal features, the 3×3 window size and stride of 2 effectively retain valuable features for fault diagnosis while reducing data volume, preventing model overfitting, and improving the model's generalization ability.

[0112] The specific network structure is as follows:

[0113] 1. Input layer

[0114] Input feature dimension: (N×10), where N is the short-time frame length after framing (such as the single-frame data processing unit length in the example), and 10 is the core feature parameter after filtering (wavelet packet energy spectrum and MFCC fusion feature).

[0115] 2. First-layer convolutional pooling module

[0116] Convolutional layer 1:

[0117] Kernel size: 3×10 (time dimension × feature dimension), where "3" adapts to feature changes within a short time frame, and "10" covers all 10 input features to achieve global correlation capture between features;

[0118] Number of convolution kernels: 16 (for basic fault feature extraction, balancing accuracy and computational cost);

[0119] Step size: 1 (time dimension), no padding (Valid mode), preserves intra-frame detail features;

[0120] Activation function: ReLU (enhances nonlinear feature representation).

[0121] Pooling layer 1:

[0122] Type: Max pooling;

[0123] Window size: 3×1 (time dimension × feature dimension), stride: 2×1;

[0124] Function: Reduces the dimensionality of the output features of convolutional layer 1, retains key features (such as fault-related energy mutations and spectral fluctuations), and reduces the parameter size.

[0125] 3. Second-layer convolutional pooling module

[0126] Convolutional layer 2:

[0127] Kernel size: 3×1 (time dimension × feature dimension, based on the feature channels after dimensionality reduction in the first layer);

[0128] Number of convolution kernels: 32 (extracting high-order correlated features, such as complex fault modes with multiple features coupled);

[0129] Step size: 1 (time dimension), no padding;

[0130] Activation function: ReLU.

[0131] Pooling layer 2:

[0132] Type: Max pooling;

[0133] Window size: 3×1, step size: 2×1;

[0134] Function: Further reduce dimensionality, focus on core fault characteristics, and reduce the risk of overfitting.

[0135] ECA channel attention mechanism:

[0136] Function: Through local cross-channel interaction, it adaptively enhances the weight of fault-sensitive feature channels (such as high-frequency discharge feature channels and low-frequency vibration feature channels), suppresses noise interference, improves the signal-to-noise ratio, and strengthens the expression of key fault features.

[0137] 4. Feature flattening and fully connected layers

[0138] Flatten layer:

[0139] Function: Flatten the multidimensional feature matrix (e.g., (M×1×32)) after the second pooling layer into a one-dimensional vector (length M×32, where M is the time dimension after dimensionality reduction), to prepare for the input of the fully connected layer.

[0140] Fully connected layer 1:

[0141] Number of neurons: 128 (feature fusion and nonlinear transformation to meet fault classification requirements);

[0142] Activation function: ReLU.

[0143] Fully connected layer 2:

[0144] Number of neurons: 4 (corresponding to four types of faults: heavy overload, loose clamps, DC bias, and short circuit impact);

[0145] Output: The probability distribution of the four types of faults is converted by the Softmax function.

[0146] 5. Output layer

[0147] Output results: Predicted probabilities of four types of faults, with the fault type corresponding to the highest probability selected as the diagnostic result.

[0148] ECA Channel Attention: Enables the model to learn and evaluate the importance of features in each channel, dynamically adjust the channel output weights, thereby encouraging the network to pay more attention to features that are beneficial to the current task, and further improving the network's feature extraction and identification / diagnosis capabilities.

[0149] Flatten layer: Expands multi-dimensional features into one-dimensional vectors, which is convenient for fully connected layers to process.

[0150] Dense layer (fully connected layer): The first layer is used for feature fusion, and the second layer outputs the classification result.

[0151] In step S43, the second fully connected layer uses the Softmax function (normalized exponential function) to convert the neuron output into a probability distribution of four fault types.

[0152] The preferred embodiments of the present invention disclosed above are merely illustrative of the invention. These preferred embodiments do not exhaustively describe all details, nor do they limit the invention to the specific implementations described. Clearly, many modifications and variations can be made based on the content of this specification. This specification selects and specifically describes these embodiments to better explain the principles and practical applications of the invention, thereby enabling those skilled in the art to better understand and utilize the invention. The invention is limited only by the claims and their full scope and equivalents.

Claims

1. A method for diagnosing the operating conditions of a converter valve based on acoustic signature feature extraction, characterized in that, Includes the following steps: S1. Obtain the raw audio data of the high-pressure converter valve's operating status, and perform data frame division, windowing, and normalization preprocessing on the acquired sound signal; S2. Perform blind source separation on the preprocessed signal to separate various source signals in the mixed sound signal and obtain the abnormal sound signal of the high-pressure converter valve. S3. The MFCC features and wavelet packet energy spectrum of the high-pressure converter valve sound signal are extracted by a joint feature extraction method, and the parameters are dynamically adjusted by the improved Zebra Optimization Algorithm (ZAO). S4. Input the combined optimized features of the extracted wavelet packet energy spectrum and MFCC into the pre-trained convolutional neural network for identification to determine the type of fault in the high-pressure converter valve's operating status.

2. The method for diagnosing the operating conditions of a converter valve based on acoustic signature extraction according to claim 1, characterized in that, Step S1 specifically includes the following steps: Acquire sound signals generated during the operation of high-pressure converter valves using a non-contact method; The acquired acoustic signals are preprocessed by framing, windowing, and normalization.

3. The method for diagnosing the operating conditions of a converter valve based on acoustic signature extraction according to claim 1, characterized in that, The S2 step specifically includes: Fast Independent Component Analysis (FastICA) is used to separate various source signals from aliased signals; The kurtosis feature index is used to further identify and filter the target acoustic signal.

4. The method for diagnosing the operating conditions of a converter valve based on acoustic signature extraction according to claim 1, characterized in that, The S-step specifically includes the following steps: The energy spectrum feature parameters of the high-pressure converter valve acoustic signal are extracted by wavelet packet decomposition. Specifically, the high-pressure converter valve acoustic signal is decomposed, the energy in different frequency ranges is calculated, and finally the energy spectrum feature parameters are constructed. Extracting MFCC features from the acoustic signal of the high-pressure converter valve; The improved ZAO optimization algorithm dynamically adjusts parameters based on actual operating conditions.

5. The method for diagnosing the operating conditions of a converter valve based on acoustic signature extraction according to claim 4, characterized in that, Extracting the MFCC features of the acoustic signal from the high-pressure converter valve includes the following steps: By performing preprocessing, framing, windowing, FFT, power spectrum calculation, Mel filtering, logarithmic calculation, and DCT transformation on the acoustic signal of the high-pressure converter valve, a set of coefficients that can represent the spectral characteristics of the acoustic signal are obtained.

6. The method for diagnosing the operating conditions of a converter valve based on acoustic signature extraction according to claim 4, characterized in that, The improved ZAO optimization algorithm dynamically adjusts parameters based on actual operating conditions, including the following steps: The improved Zebra optimization algorithm dynamically adjusts parameters such as the number of wavelet packet decomposition layers, MFCC pre-emphasis coefficients, number of Mel filters, and weights for MFCC and wavelet packet energy spectrum allocation according to different working conditions.

7. The method for diagnosing the operating conditions of a converter valve based on acoustic signature extraction according to claim 4, characterized in that, The S4 step specifically includes: The network input uses the optimized sound characteristics of the high-pressure converter valve. There are four types of faults: heavy overload, loose clamps, DC bias, and short circuit impact. Convolutional Neural Networks (CNNs) consist of convolutional layers, pooling layers, channel attention mechanism (ECA), flattened layers, and fully connected layers. The Softmax function provides the predicted probability of the current sample belonging to each type of fault and outputs the judgment result.

8. A high-voltage DC converter valve condition diagnosis device based on acoustic signal feature extraction, used to implement the converter valve condition diagnosis method based on acoustic signature feature extraction as described in any one of claims 1-7, characterized in that, include: The data acquisition and preprocessing module acquires the sound signal of the high-pressure converter valve during operation and performs preprocessing such as framing, windowing and normalization on the acquired sound signal. The abnormal signal extraction module classifies various source signals using a blind source separation algorithm, and further filters and identifies them using kurtosis index to extract the fault sound signal of the high-pressure converter valve. The feature extraction module extracts the combined MFCC and wavelet packet energy spectrum features of the high-pressure converter valve fault sound signal, and dynamically adjusts the parameters through an improved ZOA algorithm. The fault classification module uses a pre-trained convolutional neural network to receive combined feature parameters as input and classify fault types.

9. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the program, it implements the steps of the high-voltage DC converter valve condition diagnosis method based on acoustic signal feature extraction as described in any one of claims 1 to 7.

Images