Centrifugal pump fault diagnosis method and device based on acoustic radiation signal processing

Abstract

This application relates to the field of centrifugal pump fault diagnosis technology, specifically to a method and apparatus for centrifugal pump fault diagnosis based on acoustic radiation signal processing. The method involves acquiring the acoustic radiation signal of a centrifugal pump to obtain a test acoustic radiation signal; performing wavelet packet decomposition on the test acoustic radiation signal; processing the sub-frequency band information in different ways to obtain a reconstructed signal; extracting various features from the reconstructed signal; combining the features to obtain a dataset; inputting the dataset into a pre-trained centrifugal pump fault diagnosis model; and outputting the centrifugal pump acoustic radiation signal fault diagnosis result. This application utilizes a combination of wavelet packet decomposition and singular value decomposition for acoustic radiation signal feature extraction, effectively improving the accuracy and efficiency of centrifugal pump acoustic radiation signal fault identification. It can accurately identify centrifugal pump fault types even under strong background noise interference, which is significant for risk warning of centrifugal pump faults.

Description

Technical Field

[0001] This application relates to the field of centrifugal pump fault diagnosis technology, specifically to a centrifugal pump fault diagnosis method and apparatus based on acoustic radiation signal processing. Background Technology

[0002] Centrifugal pumps are a common type of mechanical equipment, playing a vital economic and social role in chemical production and daily life. Intelligent diagnostic methods, through machine learning-based intelligent analysis and mining, extract fault information from equipment status monitoring data, enabling early fault diagnosis and achieving scientific and economical maintenance. Therefore, conducting centrifugal pump fault diagnosis is crucial for ensuring the safe and economical operation of equipment.

[0003] Currently, commonly used fault feature extraction methods are generally based on vibration signals. However, acquiring vibration signals often requires installing various sensors on the equipment itself and wiring them. On the one hand, some industrial sites lack the necessary space; on the other hand, the complex working conditions in industrial sites require testing personnel to set up sensors in relatively harsh environments. In contrast, acoustic radiation signals can be acquired using corresponding sound signal sensors, eliminating the need to install sensors on the equipment itself. This method has high practical value, is easy to operate, and has a promising future.

[0004] However, the acoustic radiation signal of centrifugal pumps is easily interfered with by environmental noise. During the acquisition process, noise can contaminate the original acoustic radiation signal, making it unsuitable as direct input for fault diagnosis. Noise reduction of the acquired acoustic radiation signal is a crucial step in centrifugal pump fault diagnosis. Existing techniques generally employ improved variational mode decomposition (VMD) methods to decompose the signal, using envelope entropy difference coefficients and artificial bee colony algorithms to optimize the penalty factor and the number of decomposition levels. By utilizing the characteristic frequency ranges and absolute energy changes of IMF1 and IMF2, the development and changes in cavitation within the centrifugal pump can be reflected relatively accurately. Alternatively, a 2D-HMM centrifugal pump fault diagnosis method based on wavelet packets can be used. Using wavelet packets as a signal decomposition tool, this method can better reflect the signal characteristics of different frequency bands compared to Fourier transform.

[0005] While the improved variational mode decomposition method described above can transfer the signal component acquisition process to a variational framework and effectively avoid problems such as mode aliasing, over-envelope, under-envelope, and boundary effects, this method requires manual selection of the decomposition level K and the penalty factor a. The selection of the penalty factor a and the decomposition level K affects the decomposition effect of the method, and the parameter setting method cannot obtain the optimal parameter combination. Similarly, traditional wavelet packet decomposition also has the same problem. The decomposition level and threshold processing method need to be manually selected. Theoretically, different signals require different selection strategies, but manual selection is inefficient, difficult to adapt to different signal conditions, and often results in different effects for different signals, thus losing the meaning of noise reduction.

[0006] Therefore, how to efficiently and accurately reduce the noise of the acoustic radiation signal of centrifugal pumps, construct an effective dataset to train a diagnostic model with good diagnostic performance, and correctly diagnose the fault type of centrifugal pumps are urgent problems to be solved. Summary of the Invention

[0007] In view of the shortcomings of the prior art described above, the purpose of this application is to provide a centrifugal pump fault diagnosis method and device based on acoustic radiation signal processing. The method is based on wavelet packet-singular value decomposition and uses a residual neural network to train the decomposed and reconstructed signal. This is to solve the problem that the prior art cannot efficiently and accurately process the acoustic radiation signal of centrifugal pump and diagnose its fault category.

[0008] To achieve the above and other related objectives, this application provides a centrifugal pump fault diagnosis method based on acoustic radiation signal processing, comprising:

[0009] Acquire the acoustic radiation signal of the centrifugal pump, and preprocess the acoustic radiation signal to obtain the acoustic radiation signal to be measured;

[0010] The acoustic radiation signal to be measured is subjected to wavelet packet decomposition, and the energy proportion of each sub-band after wavelet packet decomposition is calculated. ;

[0011] According to the sub-band energy ratio Different processing methods are applied to the sub-frequency band information to obtain the reconstructed signal;

[0012] Extract various features from the reconstructed signal, and combine the features to obtain a dataset;

[0013] Input the dataset into the pre-trained centrifugal pump fault diagnosis model, and output the fault diagnosis results of the centrifugal pump acoustic radiation signal.

[0014] In an optional embodiment of this application, the step of acquiring the acoustic radiation signal of the centrifugal pump and preprocessing the acoustic radiation signal includes:

[0015] The acoustic radiation signal of the centrifugal pump is collected in real time by an acoustic radiation signal sensor, which is aligned with the pump body of the centrifugal pump.

[0016] Remove outliers and trend terms from the acoustic radiation signal of the centrifugal pump.

[0017] In an optional embodiment of this application, the wavelet packet decomposition of the acoustic radiation signal to be measured includes:

[0018] Based on the characteristics of the acoustic radiation signal to be measured, the basis functions of wavelet packet decomposition are selected;

[0019] The number of wavelet packet decomposition layers is determined based on the spectral characteristics of the acoustic radiation signal to be measured.

[0020] The acoustic radiation signal to be measured is decomposed into wavelet packets based on the basis functions and the number of decomposition layers.

[0021] In an optional embodiment of this application, the step of basing the energy ratio of the sub-bands... Different processing methods are applied to the sub-band information to obtain the reconstructed signal, including:

[0022] when >50%, the sub-band information is completely preserved without processing;

[0023] when When the percentage is less than 1%, the sub-band information is set to zero.

[0024] When 1%≤ When the value is ≤50%, singular value decomposition is performed on the sub-band information.

[0025] In an optional embodiment of this application, the singular value decomposition process on the sub-band information includes:

[0026] The sub-frequency band information is reconstructed into a sub-frequency band signal to obtain a first sub-signal;

[0027] The first sub-signal is constructed into a Hankel matrix using a delayed embedding method to obtain the first matrix;

[0028] Based on the eigenvalues ​​of the first matrix, obtain the singular values ​​of the first matrix;

[0029] Perform difference spectrum analysis on the singular values ​​to determine the truncation order p of the singular values, retain the first p singular values, and set the subsequent singular values ​​to zero.

[0030] The first matrix is ​​reconstructed using the processed singular values ​​to obtain the second matrix;

[0031] The row vectors of the second matrix are concatenated end to end and the data at the concatenation points is removed to obtain the second sub-signal;

[0032] Accumulate all the second sub-signals to obtain the reconstructed signal.

[0033] In an optional embodiment of this application, the step of extracting various features of the reconstructed signal and combining the features to obtain a dataset includes:

[0034] The waveform, frequency spectrum, power spectrum, and envelope spectrum of the reconstructed signal are extracted.

[0035] The four types of maps obtained are arranged in a 2×2 combination to obtain the dataset.

[0036] In an optional embodiment of this application, the training process of the centrifugal pump fault diagnosis model includes:

[0037] Improved residual neural networks;

[0038] The improved residual neural network is trained to obtain the centrifugal pump fault diagnosis model.

[0039] In an optional embodiment of this application, the improved residual neural network includes:

[0040] Adjust some parameters of the fc1000 layer, softmax layer, and classification output layer in the residual neural network;

[0041] The large-kernel convolution in the residual neural network is decomposed and replaced by multiple layers of small convolutional kernels;

[0042] Improve the method for calculating loss;

[0043] Increase the number of iterations of the residual neural network.

[0044] In an optional embodiment of this application, training the improved residual neural network includes:

[0045] Obtain a training set, which is the processed acoustic radiation signal of a centrifugal pump;

[0046] Set the number of iterations in each round to a first preset value, and set the total number of loops to a second preset value;

[0047] The training set is fed into the improved residual neural network for iterative training. When the number of iterations reaches a first preset value, this round of training is completed.

[0048] Once the previous round of training is completed, the training set is randomly shuffled and used in the next round of training to continue training.

[0049] When the training process loops to the second preset value, the training ends to obtain the centrifugal pump fault diagnosis model.

[0050] To achieve the above and other related objectives, this application also provides a centrifugal pump fault diagnosis device based on acoustic radiation signal processing, comprising:

[0051] Signal acquisition unit: used to acquire the acoustic radiation signal of the centrifugal pump, and preprocess the acoustic radiation signal to obtain the acoustic radiation signal to be measured;

[0052] Wavelet packet decomposition unit: Used to perform wavelet packet decomposition on the acoustic radiation signal to be measured, and calculate the energy proportion of each sub-band after wavelet packet decomposition. ;

[0053] Reconstructed signal unit: used to reconstruct the signal based on the energy proportion of the sub-band. Different processing methods are applied to the sub-frequency band information to obtain the reconstructed signal;

[0054] Feature extraction unit: used to extract various features of the reconstructed signal and combine the features to obtain a dataset;

[0055] Fault diagnosis unit: Used to input the dataset into the pre-trained centrifugal pump fault diagnosis model and output the fault diagnosis results of the centrifugal pump acoustic radiation signal.

[0056] This application discloses a centrifugal pump fault diagnosis method based on acoustic radiation signal processing. The method involves acquiring the acoustic radiation signal of the centrifugal pump, preprocessing the acoustic radiation signal to obtain a target acoustic radiation signal, performing wavelet packet decomposition on the target acoustic radiation signal, and calculating the energy proportion of each sub-band after wavelet packet decomposition. According to the energy proportion of the sub-band The process involves processing sub-frequency band information in different ways to obtain a reconstructed signal; extracting various features from the reconstructed signal and combining these features to obtain a dataset; inputting the dataset into a pre-trained centrifugal pump fault diagnosis model to output the centrifugal pump acoustic radiation signal fault diagnosis result. This application utilizes a combination of wavelet packet decomposition and singular value decomposition for acoustic radiation signal feature extraction, which better highlights fault information, effectively improves the accuracy of subsequent fault identification, and the algorithm's speed is higher than traditional feature extraction methods, resulting in higher diagnostic efficiency. Furthermore, this application uses a residual neural network to train the model. Introducing residual blocks into the residual neural network significantly increases the number of layers and avoids network degradation phenomena seen in traditional neural networks, solving the problem of excessively deep networks affecting training. In addition, this application achieves non-contact measurement through acoustic radiation signals for centrifugal pump fault diagnosis, avoiding exposure of measurement personnel to hazardous gas environments.

[0057] It should be understood that the above general description and the following detailed description are exemplary and explanatory only, and do not limit this application. Attached Figure Description

[0058] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with this application and, together with the description, serve to explain the principles of this application. It is obvious that the drawings described below are merely some embodiments of this application, and those skilled in the art can obtain other drawings based on these drawings without any inventive effort. In the drawings:

[0059] Figure 1 This is a schematic diagram illustrating the ResNet50 neural network model structure, as shown in an exemplary embodiment of this application.

[0060] Figure 2 This is a schematic flowchart illustrating a centrifugal pump fault diagnosis method based on acoustic radiation signal processing, as shown in an exemplary embodiment of this application.

[0061] Figure 3 This is a schematic diagram illustrating the singular value decomposition process in an exemplary embodiment of this application;

[0062] Figure 4 This is a flowchart illustrating an exemplary embodiment of the improved residual neural network method of this application;

[0063] Figure 5 This is a schematic diagram illustrating the training process of the improved residual neural network as an exemplary embodiment of this application;

[0064] Figure 6 This is a schematic diagram of the framework of a centrifugal pump fault diagnosis device based on acoustic radiation signal processing, as illustrated in an exemplary embodiment of this application.

[0065] Figure 7 This is a schematic diagram illustrating the test results of a centrifugal pump fault diagnosis model, as shown in an exemplary embodiment of this application. Detailed Implementation

[0066] To enable those skilled in the art to better understand the present application, the technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present application, and not all embodiments. Based on the embodiments of the present application, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of the present application. In addition, the illustrations provided in the following embodiments are only schematically illustrating the basic concept of the present application. Therefore, the drawings only show components related to the present application and are not drawn according to the actual number, shape and size of components in the implementation. In the actual implementation, the type, quantity and proportion of each component can be arbitrarily changed, and the layout of the components may also be more complex.

[0067] It should be noted that the terms "comprising" and "having" and any variations thereof in the specification, claims and accompanying drawings of this application are intended to cover non-exclusive inclusion. For example, a process, method, system, product or device that includes a series of steps or units is not necessarily limited to those steps or units that are explicitly listed, but may include other steps or units that are not explicitly listed or that are inherent to such process, method, product or device.

[0068] The following explains the technical terms used in this application:

[0069] Wavelet packet decomposition, also known as optimal subband tree structure, is a further optimization of wavelet transform. Its main algorithm is as follows: based on wavelet transform, at each level of signal decomposition, in addition to further decomposing the low-frequency subband, it also further decomposes the high-frequency subband. Finally, by minimizing a cost function, the optimal signal decomposition path is calculated, and this path is used to decompose the original signal.

[0070] Singular Value Decomposition (SVD): Traditional eigenvalue decomposition is only suitable for extracting features from square matrices. However, in practical applications, most data correspond to matrices that are not square matrices. The matrix may be a sparse matrix with many zeros, which requires a lot of storage and wastes space. In this case, it is necessary to extract the main features. SVD represents any complex matrix by multiplying three smaller and simpler submatrices. These three smaller matrices describe the important properties of the larger matrix.

[0071] Residual Neural Networks (ResNet): The main contribution was the discovery of the "degeneration phenomenon" and the application of "shortcut connections" to address it, which greatly eliminated the training difficulties of excessively deep neural networks. The "depth" of neural networks broke through 100 layers for the first time, and the largest neural networks even exceeded 1000 layers.

[0072] Please refer to Figure 1 , Figure 1This is a schematic diagram illustrating the ResNet50 neural network model structure, as shown in an exemplary embodiment of this application. Figure 1 As shown, in an exemplary embodiment, a ResNet50 neural network model is used to construct the neural network model, that is, a 50-layer residual neural network is used to train the dataset. Figure 1 It can be divided into three parts: left, middle and right. The left part is the overall structure of ResNet50, the middle part is the specific structure of each stage of ResNet50, and the right part is the specific structure of the Bottleneck layer.

[0073] First, it's important to explain the left side. ResNet is divided into 5 stages. Stage 0 has a relatively simple structure and can be considered as a preprocessing of the INPUT. The last 4 stages all consist of Bottleneck layers and have similar structures. Stage 1 contains 3 Bottleneck layers, and the remaining 3 stages contain 4, 6, and 3 Bottleneck layers respectively.

[0074] Secondly, the middle part needs to be explained. In Stage 0, (3,224,224) refers to the number of channels, height, and width of the input INPUT, i.e., (C,H,W). We assume the height and width of the input are equal, so we use (C,W,W). The first layer of Stage 0 includes three sequential operations: convolution (CONV), where 7×7 refers to the kernel size, 64 refers to the number of kernels (i.e., the number of channels output by this convolutional layer), and / 2 indicates a stride of 2; a BN layer; and a ReLU activation function. The second layer in Stage 0 is MAXPOOL, i.e., a max pooling layer, with a kernel size of 3×3 and a stride of 2. (64,56,56) represents the number of channels, height, and width of the output of this stage, where 64 equals the number of kernels in the first convolutional layer of this stage, and 56 equals 224 / 2 / 2 (a stride of 2 halves the input size). In general, in Stage 0, an input of shape (3,224,224) passes through a convolutional layer, a batch normalization (BN) layer, a ReLU activation function, and a max-pooling layer to produce an output of shape (64,56,56). In Stage 1, the input shape is (64,56,56), and the output shape is (256,56,56). BTNK1 and BTNK2 are two structures of the Bottleneck layer. Stages 2, 3, and 4 have similar structures to Stage 1 and will not be described further here.

[0075] Finally, it's important to explain the right-hand side. BTNK2 has two variable parameters, C and W, which are c and W in the input shape (C,W,W). Let the input with shape (C,W,W) be... Let the three convolutional blocks to the left of BTNK2 (and the associated BN and ReLU) be functions The two are added together and then passed through a ReLU activation function to obtain the output of BTNK2. The shape of this output is still (C, W, W), which is the same as the input of BTNK2 mentioned above. With output With the same number of channels, BTNK1 has four variable parameters: C, W, C1, and S. Compared to BTNK2, BTNK1 has one more convolutional layer on the right side, which can be expressed as a function. BTNK1 corresponds to the input. With output The different number of channels is precisely what this added convolutional layer will handle. Become This serves to match the differences between input and output dimensions. and (if the number of channels is the same), then summation can be performed. .

[0076] Please see Figure 2 , Figure 2 This is a schematic flowchart illustrating a centrifugal pump fault diagnosis method based on acoustic radiation signal processing, as shown in an exemplary embodiment of this application. Figure 2 As shown, in an exemplary embodiment, the centrifugal pump fault diagnosis method based on acoustic radiation signal processing includes at least steps S210 to S250, which are described in detail below:

[0077] In step S210, the acoustic radiation signal of the centrifugal pump is acquired, and the acoustic radiation signal is preprocessed to obtain the acoustic radiation signal to be measured.

[0078] It should be noted that acquiring the acoustic radiation signal of the centrifugal pump and preprocessing the signal includes: real-time acquisition of the acoustic radiation signal of the centrifugal pump using an acoustic radiation signal sensor aligned with the pump body; and removal of outliers and trend terms from the acoustic radiation signal. This application constructs a method and system for centrifugal pump fault diagnosis based on acoustic radiation signals, achieving non-contact measurement and avoiding exposure of measurement personnel to hazardous gas environments.

[0079] In step S220, wavelet packet decomposition is performed on the acoustic radiation signal to be measured, and the energy proportion of each sub-band after wavelet packet decomposition is calculated. .

[0080] First, it should be noted that wavelet packet decomposition of the acoustic radiation signal to be measured includes:

[0081] First, based on the characteristics of the acoustic radiation signal to be measured, the basis functions for wavelet packet decomposition are selected. Among the selectable wavelet packet basis function types, the "dBN" wavelet basis function is selected according to the performance of the signal to be processed, with N set to 4, i.e., the "dB4" basis function is used to decompose the signal.

[0082] Secondly, based on the spectral characteristics of the acoustic radiation signal to be measured, the number of wavelet packet decomposition layers is determined. The number of wavelet packet decomposition layers mainly affects the signal decomposition effect. If the number of wavelet packet decomposition layers is too small, more noise radiation signal may be included, resulting in an insignificant noise reduction effect. If the number of wavelet packet decomposition layers is too large, the overall calculation time will increase exponentially, and the signal will be split too finely, which will also affect the final noise reduction effect. The number of wavelet packet decomposition layers is selected according to the following formula:

[0083] ,

[0084] Where N is the number of wavelet packet decomposition levels, Fs is the signal sampling frequency, fp is the peak frequency of the signal spectrum, and fsp is the second-highest frequency of the signal spectrum. For the acquired acoustic radiation signal, a 5-level wavelet packet decomposition is used.

[0085] Finally, wavelet packet decomposition is performed on the acoustic radiation signal to be measured based on the basis function and the number of decomposition layers.

[0086] It should also be noted that the energy proportion of each sub-band after wavelet packet decomposition is calculated. Specifically, this includes: analyzing the time-frequency information of the signal contained in the wavelet sub-bands, and performing targeted processing on the energy proportion of the sub-bands, which can effectively improve the algorithm's running speed. The energy proportion is calculated according to the following formula:

[0087] ,

[0088] in, Ej represents the energy percentage, where Ej is the energy of the j-th sub-band in the last layer, j=1,2,3..., and Esum is the sum of the energies of all sub-bands in the last layer.

[0089] In step S230, based on the sub-band energy ratio Different processing methods are applied to the sub-band information to obtain the reconstructed signal.

[0090] It should be noted that the statement based on the sub-band energy ratio... Different processing methods are applied to the sub-band information to obtain the reconstructed signal, specifically including: when >50%, the sub-frequency band information is completely preserved without processing; when When <1%, the sub-frequency band information is set to zero; when 1% ≤ When the value is ≤50%, singular value decomposition is performed on the sub-band information.

[0091] In step S240, various features of the reconstructed signal are extracted and combined to obtain a dataset.

[0092] It should be noted that after signal reconstruction is completed, various features are extracted from the reconstructed signal, and these features are combined to form a dataset, which is then used for training the subsequent model. The extraction of various features from the reconstructed signal specifically includes: First, extracting the waveform, frequency spectrum, power spectrum, and envelope spectrum of the reconstructed signal. The waveform reflects the time-domain information of the signal, the frequency spectrum reflects the frequency-domain information, the power spectrum reflects the change in signal power with frequency, and the envelope spectrum reflects the hidden information after demodulation. Second, the four types of spectra are arranged in a 2×2 combination to create images, and these images are archived to create a dataset.

[0093] In step S250, the dataset is input into the pre-trained centrifugal pump fault diagnosis model, and the centrifugal pump acoustic radiation signal fault diagnosis result is output.

[0094] It should be noted that the training process of the centrifugal pump fault diagnosis model includes: improving the residual neural network so that the neural network can map one-to-one to the fault types of the centrifugal pump, thereby improving the classification accuracy of the neural network; and training the improved residual neural network to obtain the centrifugal pump fault diagnosis model.

[0095] Please see Figure 3 , Figure 3 This is a schematic diagram illustrating the singular value decomposition process in an exemplary embodiment of this application. Figure 3 As shown, in an exemplary embodiment, the singular value decomposition process includes at least steps S310 to S370, which are described in detail below:

[0096] In step S310, the sub-frequency band information is reconstructed into a sub-frequency band signal to obtain a first sub-signal.

[0097] It should be noted that, in a specific embodiment, 1%≤ The coefficients of the sub-bands that can reflect the weights from ≤50% of the sub-band information are reconstructed into the sub-band signal, i.e., the first sub-signal X. j (N), where j is the j-th sub-band and N is the number of sub-signal data points. This represents the energy percentage of the sub-band.

[0098] In step S320, the first sub-signal is constructed into a Hankel matrix using the delayed embedding method to obtain the first matrix.

[0099] It should be noted that the first sub-signal X j (n) is constructed as a Hankel matrix using the delayed embedding method. The expression for constructing the Hankel matrix is:

[0100] ,

[0101] Where A is the first sub-signal X j The Hankel matrix of (n), i.e., the first matrix.

[0102] In step S330, the singular values ​​of the first matrix are obtained based on the eigenvalues ​​of the first matrix.

[0103] It should be noted that singular value decomposition is performed on the first matrix A to obtain its singular values. The formula for singular value decomposition is:

[0104] ,r,

[0105] in, For A H Eigenvalues ​​of A > > >... >0, = =...=0.

[0106] In step S340, differential spectral analysis is performed on the singular values ​​to determine the truncation order p of the singular values, the first p singular values ​​are retained, and the subsequent singular values ​​are set to zero.

[0107] It should be noted that, according to the singular value decomposition theory, the singular values ​​of the first matrix A... The singular value difference spectrum (SVD) can reflect the concentration of signal and noise energy. The first p larger singular values ​​mainly reflect the useful signal, while the smaller singular values ​​mainly reflect noise. Setting these singular values ​​to zero removes noise from the signal. The formula for the SVD is: The singular value difference spectrum is plotted, the position of the maximum abrupt change point p is found, the first p singular values ​​are retained, and the subsequent singular values ​​are set to zero. This step can improve the computation speed of subsequent networks and remove most of the noise in the signal.

[0108] In step S350, the first matrix is ​​reconstructed using the processed singular values ​​to obtain the second matrix.

[0109] It should be noted that the reconstruction of the first matrix using the processed singular values ​​is performed using the following formula:

[0110] ,

[0111] in, For the reconstructed second matrix, U is derived from A. T The m×m matrix V is composed of the eigenvectors of A, and V is composed of A. T The eigenvectors are composed of an n×n matrix, where U and V are both unitary matrices.

[0112] In step S360, the row vectors of the second matrix are connected end to end and the connection point data is removed to obtain the second sub-signal.

[0113] It should be noted that, in one specific embodiment, the second matrix... The row vectors are connected end to end and the data at the connection points are removed to form the reconstructed sub-band signal X. j (N), which is the second sub-signal.

[0114] In step S370, all the second sub-signals are accumulated to obtain the reconstructed signal.

[0115] It should be noted that, in one specific embodiment, the processed sub-band signals X are... j Adding the two signals (N) together, we obtain the final reconstructed signal Y(N). The reconstruction formula is as follows:

[0116] .

[0117] Please see Figure 4 , Figure 4 This is a schematic flowchart illustrating an improved residual neural network method according to an exemplary embodiment of this application. Figure 4 As shown, in an exemplary embodiment, the improved residual neural network method includes at least steps S410 to S440, which are described in detail below:

[0118] In step S410, some parameters of the fc1000 layer, softmax layer and classification output layer in the residual neural network are adjusted.

[0119] It should be noted that the FC1000 layer, the softmax layer, and the classification output layer belong to... Figure 1In the OUTPUT module, some parameters of the fc1000 layer, softmax layer, and classification output layer in the residual neural network are adjusted so that the residual neural network can map one-to-one to the fault types of centrifugal pumps.

[0120] In step S420, the large kernel convolution in the residual neural network is decomposed and replaced by multiple layers of small convolution kernels.

[0121] It should be noted that decomposing the large convolutional kernels in the residual neural network and replacing them with multiple layers of small convolutional kernels can not only reduce the network parameters, but also increase the network depth and improve network capacity and complexity.

[0122] In step S430, the method for calculating the loss degree is improved.

[0123] It should be noted that the purpose of improving the loss function is to enable the residual neural network to converge quickly and accelerate its computation speed. The default loss algorithm for residual neural networks is relatively slow, but this improvement speeds up the network's computation time. The improved loss function is calculated using the symmetric JS div loss. The formula for calculating the loss value of the nth sample in JS div loss is as follows:

[0124] ,

[0125] in, It is the output of the neural network, and has been normalized and logarithmized; These are the actual labels (probability is the default).

[0126] In step S440, the number of iterations of the residual neural network is increased.

[0127] It should be noted that the more iterations a residual neural network undergoes, the higher the classification accuracy will be.

[0128] Please see Figure 5 , Figure 5 This is a schematic diagram illustrating the training process of the improved residual neural network, as shown in an exemplary embodiment of this application. Figure 5 As shown, in an exemplary embodiment, training the improved residual neural network includes at least steps S510 to S550, which are described in detail below:

[0129] In step S510, a training set is obtained, which is the processed acoustic radiation signal of the centrifugal pump.

[0130] First, it should be noted that in one specific embodiment, the raw acoustic radiation signal of the centrifugal pump was acquired using a centrifugal pump test bench. The test bench mainly consists of a centrifugal pump, a pressurized water station, a control console, and a signal acquisition unit. The centrifugal pump uses an HTE40-315 signal, and the motor speed is 2950 r / min. The centrifugal pump was subjected to shaft misalignment, removal of base bolts, and increased water pressure at the water station to simulate misalignment, bolt loosening, and cavitation faults. Acoustic radiation signal data were collected under normal, misaligned, bolt loosening, and cavitation conditions, with 200 sets of data selected for each operating state. Each set of sample data has 2560 sampling points.

[0131] It should also be noted that the raw signals collected above need to be processed. First, for the signal, the dB4 wavelet basis function is selected to perform wavelet packet decomposition. The number of decomposition layers is selected as five layers. After decomposition, 32 sub-bands are generated. The energy proportion of each sub-band is calculated in turn, and singular value decomposition is performed based on the energy proportion information. Finally, the processed sub-band signals are reconstructed to obtain the reconstructed acoustic radiation signal. Among them, 90% of the data in each group is selected and labeled as the training set for model training, and 10% of the data is used as the test set for model training.

[0132] In step S520, the number of iterations in each round is set to a first preset value, and the total number of cycles is set to a second preset value.

[0133] It should be noted that, in one specific embodiment, the number of iterations in each round is set to nine, and the total number of loops is set to thirty rounds.

[0134] In step S530, the training set is fed into the improved residual neural network for iterative training. When the number of iterations reaches a first preset value, this round of training is completed.

[0135] It should be noted that, in one specific embodiment, during each training round, after nine iterations, the trained neural network is validated in real time using a test set. Real-time validation allows technicians to easily observe the results. Generally, the accuracy rate increases with each validation, but if it drops excessively, technicians can promptly stop training and make corresponding adjustments to the network, thereby reducing the time cost during the training process.

[0136] In step S540, after the previous round of training is completed, the training set is randomly shuffled and put into the next round of training to continue training.

[0137] In step S550, when the training process loops to the second preset value, the training ends to obtain the centrifugal pump fault diagnosis model.

[0138] It should be noted that, in a specific embodiment, the training process can be terminated after thirty rounds, resulting in a centrifugal pump fault diagnosis model. The model's effectiveness is then validated using a test set, and the feasibility of the residual neural network is determined based on the results. The confusion matrix of the test results is shown below. Figure 7 As shown, calculations show that the accuracy of this centrifugal pump fault diagnosis model is 98%, indicating a good effect on centrifugal pump fault diagnosis.

[0139] Please see Figure 6 , Figure 6 This is a schematic diagram illustrating the framework of a centrifugal pump fault diagnosis device based on acoustic radiation signal processing, as shown in an exemplary embodiment of this application. Figure 6 As shown, in an exemplary embodiment, the centrifugal pump fault diagnosis device 600 based on acoustic radiation signal processing includes at least:

[0140] Signal acquisition unit 610: used to acquire the acoustic radiation signal of the centrifugal pump, and preprocess the acoustic radiation signal to obtain the acoustic radiation signal to be measured.

[0141] Wavelet packet decomposition unit 620: Used to perform wavelet packet decomposition on the acoustic radiation signal to be measured, and calculate the energy proportion of each sub-band after wavelet packet decomposition. .

[0142] Reconstruction signal unit 630: used to reconstruct the signal based on the energy ratio of the sub-band. Different processing methods are applied to the sub-band information to obtain the reconstructed signal.

[0143] Feature extraction unit 640: used to extract various features of the reconstructed signal and combine the features to obtain a dataset.

[0144] Fault diagnosis unit 650: Used to input the dataset into the pre-trained centrifugal pump fault diagnosis model and output the fault diagnosis results of the centrifugal pump acoustic radiation signal.

[0145] It should be noted that the centrifugal pump fault diagnosis device 600 based on acoustic radiation signal processing in this embodiment is a device corresponding to the centrifugal pump fault diagnosis method based on acoustic radiation signal processing described above. The functional modules in the centrifugal pump fault diagnosis device 600 may correspond to the corresponding steps in the centrifugal pump fault diagnosis method based on acoustic radiation signal processing. The centrifugal pump fault diagnosis device 600 based on acoustic radiation signal processing in this embodiment can be implemented in conjunction with the centrifugal pump fault diagnosis method based on acoustic radiation signal processing. Accordingly, the relevant technical details mentioned in the centrifugal pump fault diagnosis device 600 based on acoustic radiation signal processing in this embodiment can also be applied to the centrifugal pump fault diagnosis method based on acoustic radiation signal processing described above.

[0146] It should be noted that the functional modules of the centrifugal pump fault diagnosis device 600 based on acoustic radiation signal processing described above can be fully or partially integrated into a single physical entity, or they can be physically separated. Furthermore, these modules can be implemented entirely in software via processing element calls; they can be fully implemented in hardware; or some modules can be implemented in software via processing element calls, while others are implemented in hardware. Additionally, these modules can be fully or partially integrated together, or implemented independently. The processing element mentioned here can be an integrated circuit with signal processing capabilities. During implementation, some or all steps of the above method, or the functional modules mentioned above, can be completed through integrated logic circuits in the hardware of the processor element or through software instructions.

[0147] In summary, the centrifugal pump fault diagnosis method and apparatus disclosed in this application, based on acoustic radiation signal processing, utilizes a combination of wavelet packet decomposition and singular value decomposition for signal denoising, better highlighting fault information. This effectively improves the accuracy of subsequent fault identification and offers higher speed and diagnostic efficiency compared to traditional signal denoising algorithms. Furthermore, this application employs a residual neural network to train the model. Introducing residual blocks into the residual neural network significantly increases the number of layers while avoiding network degradation, thus solving the problem of excessively deep networks affecting training. Moreover, this application achieves non-contact measurement through acoustic radiation signals for centrifugal pump fault diagnosis, preventing personnel from being exposed to hazardous gas environments. Overall, the centrifugal pump fault diagnosis method provided in this application can accurately identify centrifugal pump fault types even under strong background noise interference, which is significant for risk warning of centrifugal pump faults.

[0148] The above embodiments are merely illustrative of the principles and effects of this application and are not intended to limit this application. Any person skilled in the art can modify or alter the above embodiments without departing from the spirit and scope of this application. Therefore, all equivalent modifications or alterations made by those skilled in the art without departing from the spirit and technical concept disclosed in this application should still be covered by the claims of this application.

Claims

1. A method for diagnosing centrifugal pump faults based on acoustic radiation signal processing, characterized in that, The method includes: Acquire the acoustic radiation signal of the centrifugal pump, and preprocess the acoustic radiation signal to obtain the acoustic radiation signal to be measured; The acoustic radiation signal to be measured is subjected to wavelet packet decomposition, and the energy proportion of each sub-band after wavelet packet decomposition is calculated. ; According to the sub-band energy ratio Different processing methods are applied to the sub-frequency band information to obtain the reconstructed signal; wherein, the process is based on the energy proportion of the sub-frequency band. Different processing methods are applied to the sub-band information to obtain the reconstructed signal, including: when >50%, the sub-frequency band information is completely preserved without processing; when When <1%, the sub-frequency band information is set to zero; when 1% ≤ When the value is ≤50%, singular value decomposition is performed on the sub-band information; Extract various features from the reconstructed signal, and combine the features to obtain a dataset; Input the dataset into the pre-trained centrifugal pump fault diagnosis model, and output the fault diagnosis results of the centrifugal pump acoustic radiation signal.

2. The centrifugal pump fault diagnosis method based on acoustic radiation signal processing according to claim 1, characterized in that, The acquisition of the acoustic radiation signal from the centrifugal pump and the preprocessing of the acoustic radiation signal include: The acoustic radiation signal of the centrifugal pump is collected in real time by an acoustic radiation signal sensor, which is aligned with the pump body of the centrifugal pump. Remove outliers and trend terms from the acoustic radiation signal of the centrifugal pump.

3. The centrifugal pump fault diagnosis method based on acoustic radiation signal processing according to claim 1, characterized in that, The wavelet packet decomposition of the acoustic radiation signal to be measured includes: Based on the characteristics of the acoustic radiation signal to be measured, the basis functions of wavelet packet decomposition are selected; The number of wavelet packet decomposition layers is determined based on the spectral characteristics of the acoustic radiation signal to be measured. The acoustic radiation signal to be measured is decomposed into wavelet packets based on the basis functions and the number of decomposition layers.

4. The centrifugal pump fault diagnosis method based on acoustic radiation signal processing according to claim 1, characterized in that, The singular value decomposition process for the sub-frequency band information includes: The sub-frequency band information is reconstructed into a sub-frequency band signal to obtain a first sub-signal; The first sub-signal is constructed into a Hankel matrix using a delayed embedding method to obtain the first matrix; Based on the eigenvalues ​​of the first matrix, obtain the singular values ​​of the first matrix; Perform difference spectrum analysis on the singular values ​​to determine the truncation order p of the singular values, retain the first p singular values, and set the subsequent singular values ​​to zero. The first matrix is ​​reconstructed using the processed singular values ​​to obtain the second matrix; The row vectors of the second matrix are concatenated end to end and the data at the concatenation points is removed to obtain the second sub-signal; Accumulate all the second sub-signals to obtain the reconstructed signal.

5. The centrifugal pump fault diagnosis method based on acoustic radiation signal processing according to claim 1, characterized in that, The step of extracting various features from the reconstructed signal and combining the features to obtain a dataset includes: The waveform, frequency spectrum, power spectrum, and envelope spectrum of the reconstructed signal are extracted. The four types of maps obtained are arranged in a 2×2 combination to obtain the dataset.

6. The centrifugal pump fault diagnosis method based on acoustic radiation signal processing according to claim 1, characterized in that, The training process of the centrifugal pump fault diagnosis model includes: Improved residual neural networks; The improved residual neural network is trained to obtain the centrifugal pump fault diagnosis model.

7. The centrifugal pump fault diagnosis method based on acoustic radiation signal processing according to claim 6, characterized in that, The improved residual neural network includes: Adjust some parameters of the fc1000 layer, softmax layer, and classification output layer in the residual neural network; The large-kernel convolution in the residual neural network is decomposed and replaced by multiple layers of small convolutional kernels; Improve the method for calculating loss; Increase the number of iterations of the residual neural network.

8. The centrifugal pump fault diagnosis method based on acoustic radiation signal processing according to claim 6, characterized in that, The training of the improved residual neural network includes: Obtain a training set, which is the processed acoustic radiation signal of a centrifugal pump; Set the number of iterations in each round to a first preset value, and set the total number of loops to a second preset value; The training set is fed into the improved residual neural network for iterative training. When the number of iterations reaches a first preset value, this round of training is completed. Once the previous round of training is completed, the training set is randomly shuffled and used in the next round of training to continue training. When the training process loops to the second preset value, the training ends to obtain the centrifugal pump fault diagnosis model.

9. A centrifugal pump fault diagnosis device based on acoustic radiation signal processing, characterized in that, The apparatus for diagnosing centrifugal pump faults based on acoustic radiation signal processing as described in any one of claims 1 to 8 includes: Signal acquisition unit: used to acquire the acoustic radiation signal of the centrifugal pump, and preprocess the acoustic radiation signal to obtain the acoustic radiation signal to be measured; Wavelet packet decomposition unit: Used to perform wavelet packet decomposition on the acoustic radiation signal to be measured, and calculate the energy proportion of each sub-band after wavelet packet decomposition. ; Reconstructed signal unit: used to reconstruct the signal based on the energy proportion of the sub-band. Different processing methods are applied to the sub-frequency band information to obtain the reconstructed signal; Feature extraction unit: used to extract various features of the reconstructed signal and combine the features to obtain a dataset; Fault diagnosis unit: Used to input the dataset into the pre-trained centrifugal pump fault diagnosis model and output the fault diagnosis results of the centrifugal pump acoustic radiation signal.

Images