Centrifugal pump cavitation state recognition method based on multi-source sensor signals

Abstract

The invention discloses a centrifugal pump cavitation state recognition method based on multi-source sensor signals, which comprises the following steps: collecting the multi-source sensor signals of a centrifugal pump at different water temperatures and power frequencies to obtain multi-modal data with cavitation state labels; resampling the multi-modal signal data by using a linear interpolation method, and synchronizing and aligning the signals in a time domain; performing time-frequency analysis on multi-sensor signals by using short-time Fourier transform, performing dimensionality reduction on features through principal component analysis, and extracting one-dimensional features and two-dimensional features; a one-dimensional convolutional neural network and a two-dimensional convolutional neural network are used to carry out deep learning on signal features of different modes, and feature layer fusion is carried out on multi-dimensional features through a feature fusion strategy; classifying the fused feature vectors by using time sequence characteristics in a mode of gating a circulation unit; and adopting a weighted fusion strategy, setting a weight coefficient of each signal source according to the classification accuracy of each signal source in fault diagnosis, and inputting weighted data into a classification identification model support vector machine to obtain a final cavitation state identification result. According to the method, through the synergistic effect of the signals of the multi-source sensor, the recognition precision and robustness of the cavitation state are comprehensively improved, and the method has important economical efficiency and technical popularization value.

Description

Technical Field

[0001] The invention belongs to the field of fluid machinery cavitation fault diagnosis, and in particular relates to a centrifugal pump cavitation state identification method based on multi-source sensor signals. Background Art

[0002] Centrifugal pumps are widely used in industrial production, and their stability and efficiency directly affect the operational safety and economy of the entire system. Cavitation is a common fault in the operation of centrifugal pumps, which often causes serious damage to the pump body and related equipment. According to statistics, about 30% of centrifugal pump failures worldwide are related to cavitation, resulting in economic losses of up to billions of dollars. Therefore, it is particularly important to identify the cavitation state of centrifugal pumps in a timely and accurate manner.

[0003] In practical applications, relying on a single sensor signal to identify the cavitation state of a centrifugal pump is obviously not enough to meet industrial needs. The operating state of a centrifugal pump is complex and changeable, and a single sensor often cannot fully capture its dynamic characteristics. For example, the vibration signal can reflect the abnormality of the mechanical part, but it may not be obvious in the early stage of cavitation; the noise signal provides dynamic information of liquid flow, but is greatly affected by environmental noise interference; the inlet pressure pulsation and outlet pressure pulsation signals can reflect the pressure changes of the fluid in the pump, but it is difficult to accurately judge the degree of cavitation by analyzing them alone; and the torque signal is closely related to the energy loss of the pump. Therefore, it is necessary to integrate the synergistic effects of five types of signals: vibration, noise, inlet pressure pulsation, outlet pressure pulsation, and torque, in order to achieve accurate identification of the cavitation state of a centrifugal pump. By constructing a multi-layer neural network, the deep learning algorithm can automatically extract features from a large number of sensor signals and effectively identify potential cavitation states.

[0004] In summary, the centrifugal pump cavitation state identification method based on multi-source sensor signals proposed in the present invention solves the shortcomings of the traditional centrifugal pump cavitation state identification method by comprehensively utilizing multi-source sensor signals and deep learning algorithms, and significantly improves the accuracy and robustness of cavitation state identification. The implementation of this method will provide effective protection for the safe operation of centrifugal pumps and reduce maintenance costs, and has important industrial application value and promotion prospects. Summary of the invention

[0005] In view of the problem of poor recognition accuracy of the cavitation state of centrifugal pumps in the prior art, the present invention proposes a method for multi-source sensor signals and uses deep learning technology to efficiently process features, which can accurately identify the cavitation state of the fluid in the centrifugal pump.

[0006] The present invention discloses a method for identifying the cavitation state of a centrifugal pump based on multi-source sensor signals, the method comprising the following steps:

[0007] S1, collect multi-source sensor signals of centrifugal pumps at different water temperatures and operating frequencies to obtain multimodal data with cavitation state labels;

[0008] S2, resampling the multimodal signal data using a linear interpolation method to synchronize and align the signals in the time domain;

[0009] S3, use short-time Fourier transform to perform time-frequency analysis on multi-sensor signals, and reduce the dimension of features through principal component analysis to extract one-dimensional features and two-dimensional features;

[0010] S4, use one-dimensional convolutional neural network and two-dimensional convolutional neural network to perform deep learning on signal features of different modalities, and fuse multi-dimensional features at feature layers through feature fusion strategy;

[0011] S5, using the gated recurrent unit method to classify the fused feature vector using the timing characteristics;

[0012] S6. Using a weighted fusion strategy, the weight coefficient of each signal source is set according to the classification accuracy of each signal source in fault diagnosis, and the weighted data is input into the classification recognition model support vector machine to obtain the final cavitation state recognition result.

[0013] Furthermore, the multi-source sensor signals include: a pump casing vibration signal, a 1m distance noise signal, a pump inlet pressure pulsation signal, a pump outlet pressure pulsation signal and a pump shaft torque signal.

[0014] Furthermore, the cavitation status labels include four types: no cavitation, incipient cavitation, strong cavitation, and severe abnormal cavitation.

[0015] Furthermore, the different water temperature conditions are realized by electric heating wires and temperature control cabinets, and the experimental conditions are monitored by temperature sensors. The different power frequency conditions are realized and monitored by electric control cabinets.

[0016] Furthermore, the resampling of the multimodal signal data using the linear interpolation method is specifically as follows: by specifying the polynomials of each segment and their boundary conditions, the curves between the data points are connected to ensure that the curves are smooth and continuous at these points.

[0017] For each spline segment Usually a cubic polynomial, each interval There is a polynomial like this that can be expressed as follows:

[0018] Ensure the smoothness and continuity of the entire curve, and satisfy the conditions of equal function values, equal first-order derivatives, equal second-order derivatives, and natural boundary conditions.

[0019] Cubic spline interpolation is used for resampling of multimodal signal data. First, the coefficients of each interval are calculated by combining the equations of the above conditions. . Use the obtained coefficients to construct cubic spline functions for each interval. Interpolate or extrapolate at a given point and use the constructed spline segment polynomial to calculate at the required position. By combining the polynomials for each interval, a smooth, continuous curve is constructed at the known data nodes to achieve high-precision interpolation on the data set.

[0020] Furthermore, the short-time Fourier transform for time-frequency analysis of multi-sensor signals is specifically: using short-time Fourier transform to obtain the time-frequency distribution of the signal:

[0021] in, Indicates the cavitation signal, represents the time variable, represents a frequency variable, Indicates The window function is centered on , and the Hanning window is applied. The overlap rate of the window is 0.2. This is the obtained time-frequency distribution matrix.

[0022] Furthermore, the dimension reduction of the features by principal component analysis is specifically as follows: Apply covariance analysis to each time frame and get the covariance matrix . Then, the covariance matrix Perform eigenvalue decomposition to obtain eigenvalues and the corresponding eigenvector . Use the difference spectrum method to select the main eigenvalues. Calculate the difference between each eigenvalue and its subsequent eigenvalue. Select the point with the largest difference as the cutoff point to determine the number of principal components. :

[0023] Furthermore, the feature fusion strategy for fusing multi-dimensional features at the feature level is specifically as follows: applying one-dimensional and two-dimensional convolutional neural networks to perform deep learning on the extracted features to obtain highly generalized one-dimensional and two-dimensional feature vectors, which are then concatenated to obtain a fused feature vector.

[0024] Furthermore, the gated recurrent unit is specifically: using the gated recurrent unit to classify the fused feature vector. The core idea of ​​the gated recurrent unit is to use the update gate and the reset gate to decide when to update or forget the old state information. The basic update formula of the gated recurrent unit is as follows:

[0025] Update Gate:

[0026] Reset the gate:

[0027] Candidate hidden states:

[0028] Final hidden state:

[0029] in: Indicates that at time step Input, and denote the hidden states of the current and previous time steps, and Represent the activation vectors of the update gate and reset gate, respectively, controlling how information is passed between units. , , and , , are the model parameters, corresponding to the weight matrix and bias vector for the respective gating and state updates. represents the sigmoid activation function, which is used to limit the gate signal between 0 and 1, and Hyperbolic tangent activation function, used to generate candidate hidden states. To represent vector and vector splicing.

[0030] Furthermore, the weighted fusion strategy is adopted to set the weight coefficient of the signal source according to the classification accuracy of each signal source in fault diagnosis, specifically: the weighted fusion strategy is adopted to set the weight coefficient of the fusion feature vector according to the classification accuracy of each sensor signal in the classification task, and the weighted data is input into the final support vector machine classification recognition model, and the integrated vector It is expressed as:

[0031] in, They are the weights of vibration signal, noise signal, inlet pressure pulsation signal, outlet pressure pulsation signal and torque signal respectively. The weights are calculated by the accuracy of the decision-making of the five single-signal diagnosis models in S5, and the benchmark accuracy is introduced to increase the weight distribution of the advantage model.

[0032] Furthermore, the weighted data is input into the classification recognition model support vector machine, specifically: given the fusion feature vector As input, the SVM model will find an optimal hyperplane: , in is the normal vector of the hyperplane, is the bias term.

[0033] During the training phase, supervised learning techniques are applied to reduce the discrepancy between input data and reconstructed output. The training set consists of labeled integration vectors, 60% of which are used for training, 20% for validation, and the remaining 20% ​​for testing.

[0034] During the model validation process, the validation set is used to evaluate the performance of the model, mainly relying on cross-validation indicators such as accuracy, recall, precision, and F1 score. Based on the feedback of these evaluation indicators, the model parameters are adjusted (such as adjusting the learning rate, kernel function type and its parameters, the number of support vectors, etc.) to enhance the classification performance.

[0035] According to another aspect of the present specification, there is provided a centrifugal pump cavitation state identification system based on multi-source sensor signals, the system comprising: a centrifugal pump multi-source sensor data acquisition module, a signal data resampling module, a multi-dimensional feature extraction module, a multi-dimensional feature fusion module, a single sensor identification and classification module, and a multi-source sensor identification and classification module;

[0036] The centrifugal pump multi-source sensor data acquisition module is used to collect multi-source sensor signals of the centrifugal pump at different water temperatures and operating frequencies to obtain multi-modal data with cavitation state labels;

[0037] The signal data resampling module is used to resample the multimodal signal data using a linear interpolation method to synchronize and align the signals in the time domain;

[0038] The multidimensional feature extraction module is used to perform time-frequency analysis on multi-sensor signals using short-time Fourier transform, and to reduce the dimension of features by principal component analysis to extract one-dimensional features and two-dimensional features;

[0039] The multidimensional feature fusion module is used to use a one-dimensional convolutional neural network and a two-dimensional convolutional neural network to perform deep learning on signal features of different modalities, and to fuse multidimensional features at the feature layer through a feature fusion strategy;

[0040] The single sensor recognition and classification module is used to classify the fused feature vectors using the timing characteristics in a gated recurrent unit manner.

[0041] The multi-source sensor identification and classification module is used to adopt a weighted fusion strategy, set the weight coefficient of the signal source according to the classification accuracy of each signal source in fault diagnosis, and input the weighted data into the classification and identification model support vector machine to obtain the final cavitation state identification result.

[0042] According to the third aspect of the present specification, a centrifugal pump cavitation state identification device based on multi-source sensor signals is provided, comprising a central processing unit, a network adapter and a memory, wherein executable code is stored in the memory, and when the central processing unit executes the executable code, the centrifugal pump cavitation state identification method based on multi-source sensor signals is implemented.

[0043] The beneficial effects of the technical solution provided by the present invention are as follows: the present invention realizes a method and system for identifying the cavitation state of a centrifugal pump based on multi-source sensor signals, solving the problem that a single sensor signal cannot fully capture the cavitation state of a centrifugal pump. Through the collaborative work of multi-source sensors, the recognition accuracy of the cavitation state is effectively improved and the robustness of the system is enhanced. Adaptability tests are carried out for different water temperatures and power frequency conditions to ensure that the method can effectively identify the cavitation state under a variety of environmental conditions, and has broad industrial application prospects.

[0044] In summary, the present invention provides an efficient and reliable centrifugal pump cavitation state identification technology, which can significantly improve the operating safety and economy of the system and provide a new solution for fault diagnosis in the field of fluid machinery. BRIEF DESCRIPTION OF THE DRAWINGS

[0045] Figure 1 A schematic diagram of a centrifugal pump cavitation state identification method based on multi-source sensor signals;

[0046] Figure 2 This is a diagram of the multi-source sensor signal acquisition system for a centrifugal pump;

[0047] Figure 3 This is the flow chart of the cubic spline interpolation algorithm;

[0048] Figure 4 Principle diagram of extracting one-dimensional and two-dimensional features through principal component analysis;

[0049] Figure 5 Schematic diagram of gated recurrent unit;

[0050] Figure 6 It is the principle diagram of weighted fusion strategy and classification recognition model;

[0051] Figure 7 A schematic diagram of a centrifugal pump cavitation state identification device based on multi-source sensor signals provided in an embodiment of the present invention. DETAILED DESCRIPTION

[0052] In order to make the objectives, technical solutions and advantages of the present invention more clear, the embodiments of the present invention are described in further detail below.

[0053] like Figure 1As shown, this embodiment proposes a centrifugal pump cavitation state identification method based on multi-source sensor signals, which is used to realize the state identification of the cavitation degree of the fluid in the centrifugal pump, and solves the problem that it is difficult to accurately identify the cavitation state by monitoring a single sensor signal in engineering practice. See the following description for details:

[0054] S1, such as Figure 2 As shown, the centrifugal pump multi-source sensor signal acquisition system:

[0055] The experimental data collection is undertaken by a data acquisition instrument, a laptop computer and various sensors. The sampling rate is 10240Hz, and the sampling time for a single working condition is 60s. A vibration sensor is installed on the back of the centrifugal pump volute by bolting, the dynamic torque sensor is installed on the connecting shaft between the centrifugal pump and the drive motor, the noise sensor is fixed on a bracket about one meter away from the centrifugal pump, and the pressure pulsation sensor is installed on the pipeline at the inlet and outlet of the centrifugal pump, respectively, and the pressure fluctuation data of the inlet and outlet are recorded at the same time. The temperature sensor is placed in the water tank, which evaluates and controls the experimental conditions by monitoring the water temperature. An ultrasonic flow meter is also installed at the inlet section of the pump.

[0056] Before the experiment begins, first fill the water tank through the water inlet to ensure that the water level in the water tank reaches the predetermined height. Then start the water inlet valve of the centrifugal pump to ensure continuous water supply from the pump. In order to accurately simulate and control different working conditions, a small refrigeration cycle unit and a heating rod are introduced into the experiment to adjust the temperature of the liquid in the water tank. At the same time, the input frequency of the centrifugal pump is adjusted by the frequency converter to change the speed of the pump to cover different working conditions from low temperature to high temperature. There is a 30-minute time interval between each set of experiments so that the system can return to a stable state and prepare for the setting of the next set of experiments.

[0057] In order to determine the cavitation state in the centrifugal pump, the cavitation margin is selected as a parameter to quantify the degree of cavitation. The cavitation margin is an indicator of how effectively the pump can operate without cavitation. The formula is: ,in, Indicates the absolute pressure at the suction port, is the vapor pressure of the liquid, represents the density of the liquid, represents the acceleration due to gravity, Indicates the flow rate of the liquid at the suction port, Represents the friction loss from the liquid surface to the pump inlet.

[0058] S2, such as Figure 3 As shown, the cubic spline interpolation algorithm:

[0059] The cubic spline interpolation algorithm is used to generate a smooth curve passing through a set of known data points. The multimodal signal data is resampled by specifying the polynomials of each segment and its boundary conditions, connecting the curves between the data points to ensure that the curve is smooth and continuous at these points.

[0060] For each spline segment Usually a cubic polynomial, each interval There is a polynomial like this that can be expressed as follows:

[0061] To ensure smoothness and continuity of the entire curve, the following boundary conditions are used:

[0062] Function value equality condition (continuity) ensures that each spline segment is connected at the interval point is continuous:

[0063] The first-order derivative equality condition (smoothness) ensures that the first derivative is continuous at the interval connection point, that is, the curve is smooth:

[0064] The equality condition of the second-order derivative ensures that the second-order derivative is also continuous, that is, the curvature changes smoothly:

[0065] Natural boundary conditions, assuming that the second-order derivative of the spline is zero at both ends, that is, the curve is flat at both ends:

[0066] in, represents the cubic polynomial coefficients for each interval, is the input for interpolation or extrapolation, and It is the first paragraph The interval boundaries of the spline. No. The polynomial of the spline segments. Derivatives of spline polynomials. Second derivative of a spline polynomial.

[0067] Cubic spline interpolation is used for resampling of multimodal signal data. First, the coefficients of each interval are calculated by combining the equations of the above conditions. . Use the obtained coefficients to construct cubic spline functions for each interval. Interpolate or extrapolate at a given point and use the constructed spline segment polynomial to calculate at the required position. By combining the polynomials for each interval, a smooth, continuous curve is constructed at the known data nodes to achieve high-precision interpolation on the data set.

[0068] S3, such as Figure 4 As shown, one-dimensional features and two-dimensional features are extracted:

[0069] Its basic steps are: (1) short-time Fourier transform; (2) principal component analysis; (3) two-dimensional feature extraction; (4) one-dimensional feature extraction.

[0070] Use short-time Fourier transform to obtain the time-frequency distribution of the signal, and then analyze the main changing components of the signal by calculating the covariance matrix and performing eigenvalue decomposition. Select significant eigenvalues ​​through the difference spectrum, and use the eigenvectors corresponding to these eigenvalues ​​to reconstruct the main components of the signal, which are called the two-dimensional features of the signal. Perform fast Fourier transform on the reconstructed main component modulation signal and select the spectrum to obtain the one-dimensional features of the signal.

[0071] (1) Short-time Fourier transform

[0072] The basic formula for short-time Fourier transform is: , in, Indicates the cavitation signal, represents the time variable, represents a frequency variable, Indicates The window function is centered on , and the Hanning window is applied. The overlap rate of the window is 0.2. This is the obtained time-frequency distribution matrix.

[0073] (2) Principal component analysis

[0074] The principal component analysis first performs short-time Fourier transform on the result Apply covariance analysis to each time frame and get the covariance matrix : , in, express In the In vector form, express The mean vector of Indicates the total number of time frames.

[0075] Afterwards, the covariance matrix Perform eigenvalue decomposition to obtain eigenvalues and the corresponding eigenvector : , Among them, the eigenvector describes the main direction of change of the data, while the eigenvalue Indicates the magnitude of the change in that direction.

[0076] The main eigenvalues ​​are selected using the difference spectrum method. The difference between each eigenvalue and its successor is calculated: , in, Represents the difference between adjacent eigenvalues. Select the point with the largest difference as the cutoff point to determine the number of principal components :

[0077] (3) Two-dimensional feature extraction

[0078] According to the selected eigenvalues, rearrange all corresponding eigenvectors in descending order of eigenvalues ​​to construct a new eigenvector matrix . Reconstruct the principal components according to the selected eigenvectors to obtain the two-dimensional feature matrix of the signal , where each column represents a signal component

[0079] in, Indicates the eigenvector matrix corresponding to the maximum The two-dimensional features provide the overall signal structure information, which helps to determine whether the signal deviates from the normal state.

[0080] (4) One-dimensional feature extraction

[0081] Perform fast Fourier transform on the reconstructed main component modulation signal to obtain the spectrum :

[0082] in, Represents imaginary units. The one-dimensional feature provides frequency information and can be used to compare the spectrum characteristics under different operating conditions.

[0083] S4, fusion of one-dimensional features and two-dimensional features:

[0084] Before feature fusion, one-dimensional convolutional neural network and two-dimensional convolutional neural network are used to perform deep learning on signal features of different modalities.

[0085] In a one-dimensional convolutional neural network, the convolution layer mainly processes sequence data. The convolution operation is usually expressed as:

[0086] in, Represents the sequence data output by the one-dimensional convolution operation, is the one-dimensional feature data of the input, is the convolution kernel (filter), and Represents the input sequence and convolution kernel The index of Determines the position of the sliding convolution kernel on the input sequence to produce the corresponding position in the output sequence .

[0087] The convolution layer of a two-dimensional convolutional neural network processes two-dimensional spatial data. The convolution operation can be expressed as:

[0088] in, represents the two-dimensional matrix output of the two-dimensional convolution operation, Represents a two-dimensional matrix of input Apply a 2D convolution kernel (filter) As a result, and The input matrices are The row and column indices in , which traverse all positions of the entire input matrix. In a two-dimensional convolution, the convolution kernel slides with the input matrix at a certain step, multiplies each position, and then adds up all the products to get the output matrix In the corresponding position of The value of .

[0089] After applying one-dimensional and two-dimensional convolutional neural networks to deep learning the extracted features, highly generalized one-dimensional and two-dimensional feature vectors are obtained. They are concatenated to obtain the fused feature vector , respectively represent the feature vectors after the fusion of one-dimensional and two-dimensional features of vibration signal, noise signal, inlet and outlet pressure pulsation signal, and torque signal.

[0090] S5. Feature vector classification

[0091] The gated recurrent unit is used to classify the fused feature vector. The core idea of ​​the gated recurrent unit is to use the update gate and reset gate to decide when to update or forget the old state information. The basic update formula of the gated recurrent unit is as follows:

[0092] Update Gate:

[0093] Reset the gate:

[0094] Candidate hidden states:

[0095] Final hidden state:

[0096] in: Indicates that at time step Input, and denote the hidden states of the current and previous time steps, and Represent the activation vectors of the update gate and reset gate, respectively, controlling how information is passed between units. , , and , , are the model parameters, corresponding to the weight matrix and bias vector for the respective gating and state updates. represents the sigmoid activation function, which is used to limit the gate signal between 0 and 1, and Hyperbolic tangent activation function, used to generate candidate hidden states. To represent vector and vector splicing.

[0097] The update gate determines the previous hidden state How much information will be retained in the current hidden state . Update Gate The calculation uses the sigmoid activation function to ensure The value of is between 0 and 1. This value determines how much previous information is retained and how much new information is added. The reset gate determines how much previous information is forgotten. This is done by Control, also use the sigmoid function. If is close to 0, then the previous hidden state The influence of will be ignored, so that the gated recurrent unit can discard some irrelevant information. Candidate hidden state Based on the current input and the previous hidden state after reset The candidate hidden state is controlled by the reset gate and takes into account the current input and some previous information. The final hidden state is the current candidate hidden state and the previous hidden state The weighted sum of . Update gate Determines the importance of the candidate hidden state and the previous hidden state.

[0098] In the process of implementing the classification and recognition of a set of labeled feature vectors by the gated recurrent unit and evaluating its accuracy, the feature vector processed by S4 is first input into the gated recurrent unit model that has been built and set up.

[0099] During the model training process, supervised learning is performed with the help of known labels, and the network weights are iteratively updated through the back-propagation mechanism and optimization algorithm to minimize the loss function gap between the predicted output and the actual label.

[0100] After training, the accuracy of the single sensor signal classification task is obtained by importing the feature vector of the validation set or test set into the gated recurrent unit model for prediction and then calculating the degree of match between its output classification and the actual label. , respectively represent the feature vectors after the fusion of one-dimensional and two-dimensional features of vibration signal, noise signal, inlet pressure pulsation signal, outlet pressure pulsation signal and torque signal.

[0101] S6. Weighted Fusion

[0102] A weighted fusion strategy is adopted to set the weight coefficient of the fusion feature vector according to the classification accuracy of each sensor signal in the classification task, and the weighted data is input into the final support vector machine classification and recognition model.

[0103] Specifically, the integrated vector It is expressed as:

[0104] in, They are the weights of vibration signal, noise signal, inlet pressure pulsation signal, outlet pressure pulsation signal and torque signal respectively. The weights are calculated by the accuracy of the decision-making of the five single-signal diagnosis models in S5, and the benchmark accuracy is introduced to increase the weight distribution of the advantage model. Its mathematical formula can be expressed as:

[0105] in, is a manually set baseline accuracy. This equation ensures that the weight of each signal source is proportional to the difference in its classification accuracy.

[0106] In order to ensure the relative consistency of the weight coefficients of each signal source, the above proportional relationship is normalized so that the total weight of all signal sources is 1, in the form of:

[0107] After that, the fused feature vector Input into the final support vector machine classification model for classification and recognition. Specifically, given the fusion feature vector As input, the SVM model will find an optimal hyperplane:

[0108] in is the normal vector of the hyperplane, is the bias term. The goal of SVM is to maximize the margin of the hyperplane while satisfying the constraints of all sample points:

[0109] in, For sample The true category label, For sample By solving this optimization problem, SVM finally obtains the optimal hyperplane and can classify and predict new input samples.

[0110] During the training phase, supervised learning techniques are applied to reduce the discrepancy between input data and reconstructed output. The training set consists of labeled integration vectors, 60% of which are used for training, 20% for validation, and the remaining 20% ​​for testing.

[0111] During the model validation process, the validation set is used to evaluate the performance of the model, mainly relying on cross-validation indicators such as accuracy, recall, precision, and F1 score. Based on the feedback of these evaluation indicators, the model parameters are adjusted (such as adjusting the learning rate, kernel function type and its parameters, the number of support vectors, etc.) to enhance the classification performance.

[0112] The tuned support vector machine model is used to identify the integrated vector. If the model output accuracy does not meet the expected standard, repeated training is required to further improve its performance. The specific method includes subdividing the training set into multiple subsets, and simultaneously training multiple support vector machine models for each subset. After the training is completed, the diagnostic result with the highest frequency is selected by majority voting, thereby improving the accuracy and reliability of the diagnosis. In summary, the present invention provides a centrifugal pump cavitation state identification method based on multi-source sensor signals, which successfully overcomes the problem of poor recognition accuracy of traditional centrifugal pump cavitation state identification, and can accurately identify the cavitation state of the fluid in the centrifugal pump, showing superior performance and broad application prospects.

[0113] Corresponding to the method, the present embodiment also provides a centrifugal pump cavitation state identification system based on multi-source sensor signals, the system comprising: a centrifugal pump multi-source sensor data acquisition module, a signal data resampling module, a multi-dimensional feature extraction module, a multi-dimensional feature fusion module, a single sensor identification and classification module, and a multi-source sensor identification and classification module;

[0114] The centrifugal pump multi-source sensor data acquisition module is used to collect multi-source sensor signals of the centrifugal pump at different water temperatures and operating frequencies to obtain multi-modal data with cavitation state labels;

[0115] The signal data resampling module is used to resample the multimodal signal data using a linear interpolation method to synchronize and align the signals in the time domain;

[0116] The multidimensional feature extraction module is used to perform time-frequency analysis on multi-sensor signals using short-time Fourier transform, and to reduce the dimension of features by principal component analysis to extract one-dimensional features and two-dimensional features;

[0117] The multidimensional feature fusion module is used to use a one-dimensional convolutional neural network and a two-dimensional convolutional neural network to perform deep learning on signal features of different modalities, and to fuse multidimensional features at the feature layer through a feature fusion strategy;

[0118] The single sensor recognition and classification module is used to classify the fused feature vectors by using the timing characteristics in a gated recurrent unit manner;

[0119] The multi-source sensor identification and classification module is used to adopt a weighted fusion strategy, set the weight coefficient of the signal source according to the classification accuracy of each signal source in fault diagnosis, and input the weighted data into the classification and identification model support vector machine to obtain the final cavitation state identification result.

[0120] Corresponding to the above-mentioned embodiment of a method for identifying the cavitation state of a centrifugal pump based on multi-source sensor signals, the present invention also provides an embodiment of a device for identifying the cavitation state of a centrifugal pump based on multi-source sensor signals. Figure 7 An embodiment of the present invention provides a centrifugal pump cavitation state identification device based on multi-source sensor signals, including a central processing unit, a network adapter and a memory, wherein the memory stores executable code, and when the central processing unit executes the executable code, the centrifugal pump cavitation state identification method based on multi-source sensor signals is implemented.

[0121] An embodiment of a centrifugal pump cavitation state identification device based on multi-source sensor signals provided by the present invention can be applied to any device with data processing capabilities, and the device with data processing capabilities can be a device or apparatus such as a computer. The device embodiment can be implemented through software, or through hardware or a combination of software and hardware. Taking software implementation as an example, as a device in a logical sense, it is formed by the processor of any device with data processing capabilities in which it is located reading the corresponding computer program instructions in the persistent memory into the memory and running them. From a hardware perspective, if Figure 7As shown in FIG. 1 , a hardware structure diagram of a centrifugal pump cavitation state identification device based on multi-source sensor signals provided by the present invention is provided in any device with data processing capability, except Figure 7 In addition to the central processing unit, network adapter and memory shown, any device with data processing capabilities in which the apparatus in the embodiment is located may also include other hardware, generally based on the actual functions of the device with data processing capabilities, which will not be described in detail.

[0122] The implementation process of the functions and effects of each unit in the above-mentioned device is specifically described in the implementation process of the corresponding steps in the above-mentioned method, and will not be repeated here.

[0123] For the device embodiment, since it basically corresponds to the method embodiment, the relevant parts can refer to the partial description of the method embodiment. The device embodiment described above is only schematic, wherein the units described as separate components may or may not be physically separated, and the components displayed as units may or may not be physical units, that is, they may be located in one place, or they may be distributed on multiple network units. Some or all of the modules may be selected according to actual needs to achieve the purpose of the scheme of the present invention. Ordinary technicians in this field can understand and implement it without paying creative work.

Claims

1. A method for identifying the cavitation state of a centrifugal pump based on multi-source sensor signals, characterized in that: The method includes: S1, collect multi-source sensor signals of centrifugal pumps at different water temperatures and operating frequencies to obtain multimodal data with cavitation state labels; S2, resampling the multimodal signal data using a linear interpolation method to synchronize and align the signals in the time domain; S3, use short-time Fourier transform to perform time-frequency analysis on multi-sensor signals, and reduce the dimension of features through principal component analysis to extract one-dimensional features and two-dimensional features; S4, use one-dimensional convolutional neural network and two-dimensional convolutional neural network to perform deep learning on signal features of different modalities, and fuse multi-dimensional features at feature layers through feature fusion strategy; S5, using the gated recurrent unit method to classify the fused feature vector using the timing characteristics; S6. Using a weighted fusion strategy, the weight coefficient of each signal source is set according to the classification accuracy of each signal source in fault diagnosis, and the weighted data is input into the classification recognition model support vector machine to obtain the final cavitation state recognition result.

2. A centrifugal pump cavitation state identification method based on multi-source sensor signals according to claim 1, characterized in that: The multi-source sensor signals include: a pump casing vibration signal, a 1m distance noise signal, a pump inlet pressure pulsation signal, a pump outlet pressure pulsation signal and a pump shaft torque signal.

3. The method for identifying the cavitation state of a centrifugal pump based on multi-source sensor signals according to claim 1, characterized in that: The different water temperature conditions are realized by electric heating wires and temperature control cabinets, and the experimental conditions are monitored by temperature sensors. The different power frequency conditions are realized and monitored by electric control cabinets.

4. The method for identifying the cavitation state of a centrifugal pump based on multi-source sensor signals according to claim 1, characterized in that: The method of resampling the multimodal signal data using the linear interpolation method is as follows: by specifying the polynomials of each segment and their boundary conditions, the curves between the data points are connected to ensure that the curves are smooth and continuous at these points. For each spline segment Usually a cubic polynomial, each interval There is a polynomial like this that can be expressed as follows: , ensuring the smoothness and continuity of the entire curve, satisfying the equal conditions of function values, first-order derivatives, second-order derivatives and natural boundary conditions Cubic spline interpolation is used for resampling of multimodal signal data. First, the coefficients of each interval are calculated by combining the equations of the above conditions. . Use the obtained coefficients to construct cubic spline functions for each interval. Interpolate or extrapolate at a given point and use the constructed spline segment polynomial to calculate at the required position. By combining the polynomials for each interval, a smooth, continuous curve is constructed at the known data nodes to achieve high-precision interpolation on the data set.

5. The method for identifying the cavitation state of a centrifugal pump based on multi-source sensor signals according to claim 1, characterized in that: The short-time Fourier transform is used to perform time-frequency analysis on multi-sensor signals: the time-frequency distribution of the signal is obtained by using the short-time Fourier transform: ， in, Indicates the cavitation signal, represents the time variable, represents a frequency variable, Indicates The window function is centered on , and the Hanning window is applied. The overlap rate of the window is 0.

2. This is the obtained time-frequency distribution matrix.

6. The method for identifying the cavitation state of a centrifugal pump based on multi-source sensor signals according to claim 1, characterized in that: The dimensionality reduction of features by principal component analysis is specifically as follows: Apply covariance analysis to each time frame and get the covariance matrix . Then, the covariance matrix Perform eigenvalue decomposition to obtain eigenvalues ​​and corresponding eigenvectors . Use the difference spectrum method to select the main eigenvalues. Calculate the difference between each eigenvalue and its subsequent eigenvalue. Select the point with the largest difference as the cutoff point to determine the number of principal components. :

7. The method for identifying the cavitation state of a centrifugal pump based on multi-source sensor signals according to claim 1, characterized in that: The feature fusion strategy for fusing multi-dimensional features at the feature level is specifically as follows: applying one-dimensional and two-dimensional convolutional neural networks to perform deep learning on the extracted features to obtain highly generalized one-dimensional and two-dimensional feature vectors, which are then concatenated to obtain a fused feature vector.

8. The method for identifying the cavitation state of a centrifugal pump based on multi-source sensor signals according to claim 1, characterized in that: The weighted fusion strategy is adopted to set the weight coefficient of the signal source according to the classification accuracy of each signal source in fault diagnosis. Specifically, the weighted fusion strategy is adopted to set the weight coefficient of the fusion feature vector according to the classification accuracy of each sensor signal in the classification task, and the weighted data is input into the final support vector machine classification and recognition model. The integrated vector It is expressed as: ， in, They are the weights of vibration signal, noise signal, inlet pressure pulsation signal, outlet pressure pulsation signal and torque signal respectively. The weights are calculated by the accuracy of the decision-making of the five single-signal diagnosis models in S5, and the benchmark accuracy is introduced to increase the weight distribution of the advantage model.

9. A system for implementing the method according to any one of claims 1 to 9, characterized in that: The system includes: a centrifugal pump multi-source sensor data acquisition module, a signal data resampling module, a multi-dimensional feature extraction module, a multi-dimensional feature fusion module, a single sensor identification and classification module, and a multi-source sensor identification and classification module; The centrifugal pump multi-source sensor data acquisition module is used to collect multi-source sensor signals of the centrifugal pump at different water temperatures and operating frequencies to obtain multi-modal data with cavitation state labels; The signal data resampling module is used to resample the multimodal signal data using a linear interpolation method to synchronize and align the signals in the time domain; The multidimensional feature extraction module is used to perform time-frequency analysis on multi-sensor signals using short-time Fourier transform, and to reduce the dimension of features by principal component analysis to extract one-dimensional features and two-dimensional features; The multidimensional feature fusion module is used to use a one-dimensional convolutional neural network and a two-dimensional convolutional neural network to perform deep learning on signal features of different modalities, and to fuse multidimensional features at the feature layer through a feature fusion strategy; The single sensor recognition and classification module is used to classify the fused feature vectors by using the timing characteristics in a gated recurrent unit manner. The multi-source sensor identification and classification module is used to adopt a weighted fusion strategy, set the weight coefficient of the signal source according to the classification accuracy of each signal source in fault diagnosis, and input the weighted data into the classification and identification model support vector machine to obtain the final cavitation state identification result.

10. A centrifugal pump cavitation state identification device based on multi-source sensor signals, comprising a central processing unit, a network adapter and a memory, wherein the memory stores executable code, and when the central processing unit executes the executable code, the centrifugal pump cavitation state identification method based on multi-source sensor signals is implemented.

Images