Centrifugal pump cavitation diagnosis method and device based on CNN-LAE-LSTM hybrid network
By combining CNN-LAE-LSTM hybrid network with cavitation pressure signals and visual images, the limitations of a single model in centrifugal pump cavitation diagnosis are overcome, achieving high-precision and robust cavitation identification and adapting to real-time diagnosis under complex operating conditions.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- ZHEJIANG SCI-TECH UNIV
- Filing Date
- 2026-01-30
- Publication Date
- 2026-05-15
AI Technical Summary
Existing centrifugal pump cavitation diagnostic models rely on a single cavitation pressure signal, which makes it difficult to simultaneously adapt to the local impact characteristics, temporal evolution characteristics, and multi-source information complementarity of cavitation data, resulting in low diagnostic accuracy, weak noise resistance, and poor dynamic adaptability.
A CNN-LAE-LSTM hybrid network is adopted, which combines cavitation pressure signal and visual image. The signal is processed by Kalman filtering and modal time-frequency decomposition. CNN is used to extract features, LAE is used for dimensionality reduction and noise removal, and LSTM is used to capture temporal correlation, so as to realize the fusion of multi-source information and feature extraction.
It improves the accuracy and robustness of cavitation diagnosis for centrifugal pumps, enhances the accuracy of identifying complex operating conditions, and meets the needs of early warning and early response in industry.
Smart Images
Figure CN121616844B_ABST
Abstract
Description
Technical Field
[0001] This application relates to the field of deep learning cavitation diagnosis technology for centrifugal pumps, specifically to a centrifugal pump cavitation diagnosis method and apparatus using a CNN-LAE-LSTM hybrid network that combines grayscale image network discrimination. Background Technology
[0002] Centrifugal pumps, as core fluid transport equipment in energy, petrochemical, and urban water supply industries, directly determine the continuous operation capability of industrial systems through their operational reliability. Cavitation is a critical fault limiting their lifespan and efficiency. From its initial stage to its severe stage, it can lead to impeller cavitation, efficiency reduction of over 30%, and even safety accidents such as pipeline rupture. Traditional diagnostic methods often rely on time-domain and frequency-domain analysis of single pressure or vibration signals, such as Fourier transform to extract features. However, interference from motor noise and pipeline resonance under dynamic operating conditions can easily lead to feature distortion. The accuracy rate of early cavitation identification is often below 70%, which cannot meet the industrial demand for "early warning and early response." Therefore, precise diagnostic technology is urgently needed.
[0003] To improve diagnostic accuracy, researchers have introduced deep learning models, such as CNNs for local feature extraction and LSTMs for capturing temporal dependencies. However, single models have limitations: CNNs alone struggle to uncover long-term temporal development patterns of cavitation, and LSTMs suffer from significant computational redundancy when directly processing high-dimensional signals. Although some studies have attempted cascaded models, they have not specifically addressed the issues of feature redundancy and missing spatiotemporal correlations. Furthermore, cavitation severity classification often relies on signal thresholds, neglecting the intuitive characteristics of cavitation region morphology—grayscale images clearly show bubble distribution density, but current technologies do not integrate this with sensor signals, leading to blurred classification boundaries and misclassification rates exceeding 20% for moderate and severe cavitation. Existing centrifugal pump cavitation diagnostic model training methods rely on the head drop to distinguish cavitation types within the pump. However, head is a macroscopic indirect quantity that cannot directly reflect the intensity of centrifugal pump cavitation and is prone to significant measurement errors.
[0004] The demands for real-time performance and robustness in industrial cavitation diagnostics are constantly increasing. This requires not only the accurate extraction of spatiotemporal features from sensor signals but also the integration of morphological information from images for fine-grained classification. Linear autoencoders (LAEs), as lightweight dimensionality reduction tools, can reduce computational load while preserving core features, providing an optimized path for the integration of CNNs and LSTMs. However, currently, there is no technology to fuse CNN-LAE-LSTM models with grayscale images for cavitation diagnostics, failing to balance the advantages of signal temporality and image intuitiveness. This results in difficulties in simultaneously achieving diagnostic accuracy and response speed under complex operating conditions, necessitating the construction of a multi-source fusion diagnostic system. Summary of the Invention
[0005] The purpose of this application is to provide a CNN-LAE-LSTM hybrid network centrifugal pump cavitation diagnosis method and device to solve the technical problems in related technologies, such as relying on only a single cavitation pressure signal, the difficulty of a single model to simultaneously adapt to the local impact characteristics, temporal evolution characteristics and multi-source information complementarity requirements of cavitation data, resulting in low diagnostic accuracy, weak noise resistance and poor dynamic adaptability.
[0006] According to a first aspect of the embodiments of this application, a method for cavitation diagnosis of a centrifugal pump using a CNN-LAE-LSTM hybrid network is provided, comprising:
[0007] Collect cavitation data of different degrees under multiple sets of deviated flow and rated flow conditions of centrifugal pumps. The cavitation data includes cavitation pressure signals and cavitation visual images.
[0008] The cavitation pressure signal is subjected to Kalman filtering and modal time-frequency decomposition to obtain the intrinsic mode function components and residual signal data of the denoised cavitation pressure signal; the cavitation visual image is subjected to grayscale image processing to obtain single-channel grayscale image data.
[0009] The intrinsic mode function components and residual signal data are combined with the single-channel grayscale image data to form a training set and a test set;
[0010] A cavitation diagnostic model for centrifugal pumps based on CNN-LAE-LSTM is constructed. The CNN is used to extract local key features of cavitation pressure signals and cavitation visual images; the LAE is used to reduce the dimensionality of features, remove redundancy, and suppress noise; and the LSTM is used to capture the temporal correlation of cavitation pressure signals to adapt to the dynamic characteristics of centrifugal pump operation.
[0011] The centrifugal pump cavitation diagnostic model is trained using the training set and tested using the test set.
[0012] The test data is input into the trained and tested centrifugal pump cavitation diagnostic model to obtain the fault diagnosis results.
[0013] According to a second aspect of the embodiments of this application, a CNN-LAE-LSTM hybrid network centrifugal pump cavitation diagnostic device is provided, comprising:
[0014] The acquisition module is used to acquire cavitation data of different degrees under multiple sets of deviated flow and rated flow conditions of the centrifugal pump. The cavitation data includes cavitation pressure signals and cavitation visual images.
[0015] The data processing module is used to perform Kalman filtering and modal time-frequency decomposition on the cavitation pressure signal to obtain the intrinsic mode function components and residual signal data of the denoised cavitation pressure signal; and to perform grayscale image processing on the cavitation visual image to obtain single-channel grayscale image data.
[0016] The dataset construction module is used to combine the intrinsic mode function components and residual signal data with the single-channel grayscale image data to form a training set and a test set;
[0017] The model building module is used to build a centrifugal pump cavitation diagnosis model based on CNN-LAE-LSTM. The CNN is used to extract local key features of cavitation pressure signals and cavitation visual images; the LAE is used to reduce the dimensionality of features, remove redundancy, and suppress noise; the LSTM is used to capture the temporal correlation of cavitation pressure signals and adapt to the dynamic characteristics of centrifugal pump operation.
[0018] The training and testing module is used to train the centrifugal pump cavitation diagnostic model using the training set and to test the centrifugal pump cavitation diagnostic model using the test set.
[0019] The diagnostic module is used to input the data to be tested into the trained and tested centrifugal pump cavitation diagnostic model to obtain fault diagnosis results.
[0020] According to a third aspect of the embodiments of this application, an electronic device is provided, comprising:
[0021] One or more processors;
[0022] Memory, used to store one or more programs;
[0023] When the one or more programs are executed by the one or more processors, the one or more processors perform the method as described in the first aspect.
[0024] The technical solutions provided by the embodiments of this application may include the following beneficial effects:
[0025] As can be seen from the above embodiments, this application obtains cavitation data, and these parameters can be used to identify the degree of cavitation after cavitation occurs.
[0026] A novel modal time-frequency processing technique and grayscale image method are designed. The novel modal time-frequency processing technique can adaptively decompose cavitation pressure signals, accurately extract cavitation characteristic frequency components, and suppress noise interference; the grayscale image method simplifies the visual information dimension and preserves the key morphological features of the cavitation region. The two work together to achieve multi-source information complementarity, improve feature recognition, provide high-quality input for subsequent diagnostic models, and enhance the accuracy and robustness of cavitation diagnosis.
[0027] A CNN-LAE-LSTM cavitation diagnostic model for centrifugal pumps was designed. CNN automatically extracts key local features from cavitation pressure signals and cavitation visual images; LAE reduces the dimensionality of features and removes redundancy, improving efficiency and suppressing noise; LSTM captures the temporal correlation of signals, adapting to the dynamic characteristics of pump operation. The three technologies work together to enhance robustness and achieve higher accuracy in identifying complex operating conditions such as weak cavitation.
[0028] 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
[0029] 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.
[0030] Figure 1 This is a flowchart illustrating a CNN-LAE-LSTM hybrid network centrifugal pump cavitation diagnostic method according to an exemplary embodiment.
[0031] Figure 2 This is a diagram illustrating a cavitation test bench and monitoring point layout according to an exemplary embodiment; the reference numerals are as follows: 1-pump inlet casing, 2-inducer wheel, 3-impeller, 4-pump casing, 5-mechanical seal, 6-shaft.
[0032] Figure 3 This is a schematic diagram illustrating pressure and vibration signal monitoring points according to an exemplary embodiment.
[0033] Figure 4 This is a schematic diagram of the CNN-LAE-LSTM model algorithm framework according to an exemplary embodiment.
[0034] Figure 5 This is a precision map of a CNN-LAE-LSTM model illustrated according to an exemplary embodiment.
[0035] Figure 6 This is a confusion matrix diagram illustrated according to an exemplary embodiment.
[0036] Figure 7 This is a T-SNE diagram illustrated according to an exemplary embodiment.
[0037] Figure 8 This is a block diagram illustrating a CNN-LAE-LSTM hybrid network centrifugal pump cavitation diagnostic device according to an exemplary embodiment. Detailed Implementation
[0038] Exemplary embodiments will now be described in detail, examples of which are illustrated in the accompanying drawings. When the following description relates to the drawings, unless otherwise indicated, the same numbers in different drawings denote the same or similar elements. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with this application. Rather, they are merely examples of apparatuses and methods consistent with some aspects of this application.
[0039] The terminology used in this application is for the purpose of describing particular embodiments only and is not intended to be limiting of the application. The singular forms “a,” “the,” and “the” used herein are also intended to include the plural forms unless the context clearly indicates otherwise. It should also be understood that the term “and / or” as used herein refers to and includes any and all possible combinations of one or more of the associated listed items.
[0040] It should be understood that although the terms first, second, third, etc., may be used in this application to describe various information, such information should not be limited to these terms. These terms are only used to distinguish information of the same type from one another. For example, without departing from the scope of this application, first information may also be referred to as second information, and similarly, second information may also be referred to as first information. Depending on the context, the word "if" as used herein may be interpreted as "when," "when," or "in response to determination."
[0041] Figure 1 This is a flowchart illustrating a CNN-LAE-LSTM hybrid network centrifugal pump cavitation diagnostic method according to an exemplary embodiment, such as... Figure 1 As shown, the method may include the following steps:
[0042] S1. Collect cavitation data of different degrees under multiple sets of deviated flow and rated flow conditions of the centrifugal pump. The cavitation data includes cavitation pressure signals and cavitation visual images.
[0043] Specifically, such as Figure 2 As shown, the cavitation test bench and monitoring points are arranged. The test platform mainly includes: a water tank 201, an inlet valve 202, an outlet valve 203, an electromagnetic flowmeter 204, an inlet pipe pressure gauge 205, an outlet pipe pressure gauge 206, a centrifugal pump 207, and a pressure sensor array 208 (e.g., ...). Figure 3As shown, the pressure sensor array 208 includes a first sensor 2081, a second sensor 2082, a third sensor 2083, a fourth sensor 2084, a fifth sensor 2085, a sixth sensor 2086, a seventh sensor 2087, an eighth sensor 2088, and a ninth sensor 2089; a sensor power supply 209; a high-speed camera 210; a data acquisition unit 211; and a PC 212, etc., used to collect cavitation pressure signals and cavitation visual images under different pressure conditions. During the test, after the centrifugal pump starts, the sensors (powered by the sensor power supply) collect the pump's status information, while the electromagnetic flowmeter and pressure gauge simultaneously collect flow and pressure data; the high-speed camera records the visual image information of cavitation phenomena inside the pump in real time. The above multi-source signals (including data from sensors, flowmeters, and pressure gauges) are uniformly transmitted to the data acquisition unit, which integrates and preliminarily processes the signals before uploading the data to the PC.
[0044] A pressure sensor array is installed on the outer wall of the inlet casing of the centrifugal pump (e.g.) Figure 3 The cross-sections EE, FF, and GG (1-pump inlet casing, 2-inducer wheel, 3-impeller, 4-pump casing, 5-mechanical seal, 6-pump shaft) are each circumferentially arranged with three pressure sensors spaced 120° apart. Pressure sensors 2081-2083 monitor pressure fluctuations at the leading edge of the inducer wheel, pressure sensors 2084-2086 monitor pressure fluctuations inside the inducer wheel flow channel, and pressure sensors 2087-2089 monitor pressure fluctuations at the tail end of the inducer wheel. The pressure sensors have a sampling frequency of 40000Hz and an accuracy of 2‰, enabling them to obtain high-precision time-domain cavitation pressure signals in different frequency bands within the centrifugal pump's inducer wheel.
[0045] The sample data were obtained from actual experiments and mainly include operating data of centrifugal pumps under normal operating conditions and cavitation conditions. Centrifugal pumps operate normally without cavitation. When cavitation occurs, it is divided into three different cavitation levels according to the severity of cavitation. The normal operating conditions and cavitation conditions of centrifugal pumps constitute four sample categories. Specifically: Sample 0 represents the operating data of a centrifugal pump operating normally without cavitation (corresponding to no cavitation, R0). The three different cavitation levels are distributed as follows: Sample 1 represents the operating data of a centrifugal pump with initial cavitation (corresponding to slight cavitation, R1); Sample 2 represents the operating data of a centrifugal pump with moderate cavitation (corresponding to moderate cavitation, R2); and Sample 3 represents the operating data of a centrifugal pump with severe cavitation (corresponding to severe cavitation, R3).
[0046] Select a centrifugal pump 0.4 Q n 、 0.6 Q n 、 08 Q n、 1.0 Q n 、 1.2 Q n 、 1.4 Q n Flow conditions Q n =4m 3 / h represents the rated flow rate. Under various flow rates, cavitation pressure signals are collected from nine pressure monitoring points using pressure sensors. Cavitation images during centrifugal pump operation are captured by a high-speed camera. These cavitation pressure signals are transmitted to a data acquisition unit and stored as a TDMS format file. This TDMS file is then transferred to a PC terminal, where Origin software reads and processes the cavitation pressure signals. The high-speed camera synchronously records the visual information of the centrifugal pump's cavitation, creating cavitation visual images at a rate of 10,000 high-definition images per second. The cavitation pressure signals are stored on the PC.
[0047] S2. Perform Kalman filtering and modal time-frequency decomposition on the cavitation pressure signal to obtain the intrinsic mode function components and residual signal data of the denoised cavitation pressure signal; perform grayscale image processing on the cavitation visual image to obtain single-channel grayscale image data.
[0048] Specifically, on the PC, Kalman filtering and grayscale image processing are used to further process and analyze the cavitation pressure signal and cavitation visual image. Kalman filtering, as a linear optimal recursive estimation method, effectively filters noisy cavitation pressure signals caused by environmental noise, fluid disturbances, and other interference factors during centrifugal pump operation. It suppresses environmental noise and interference components and can appropriately complete missing signal values, thereby extracting an effective cavitation pressure signal reflecting the centrifugal pump's operating status. In centrifugal pump cavitation diagnosis scenarios, the cavitation pressure signal processed by this Kalman filter can significantly improve the accuracy of cavitation feature recognition, thus enhancing the reliability and accuracy of cavitation diagnosis.
[0049] The specific operation of the Kalman filter is as follows:
[0050] Initialization: Based on the prior characteristics of the cavitation pressure signal, initialize the state vector, state transition matrix, and observation matrix of the Kalman filter;
[0051] The state equation for Kalman filtering is:
[0052]
[0053] In the formula, express k Real value of the system cavitation pressure signal at any given time; : State transition matrix; B control input matrix; : Control input; Process noise;
[0054] The observation equation for Kalman filtering is:
[0055]
[0056] In the formula, :express k Observed values of the system cavitation pressure signal at any given time; H : Represents the observation matrix; Observation noise;
[0057] State prediction: The cavitation pressure signal state value at the next time step is predicted by the state transition matrix, and the predicted state covariance matrix is updated by combining the process noise covariance matrix to complete the one-step prediction.
[0058] Observation Update: Obtain the measured value of the cavitation pressure signal of the centrifugal pump in real time, calculate the residual between the predicted value and the measured value, use the predicted state covariance matrix and the observation noise covariance matrix as input to solve the Kalman gain, use the gain to correct the predicted state vector and covariance matrix, and obtain the optimal estimate of the cavitation pressure signal at the current time.
[0059] The "state prediction-observation update" step is executed iteratively, and the cavitation pressure signal is filtered time-by-time throughout the entire period. Finally, the filtered cavitation pressure signal sequence with environmental noise and disturbances is output.
[0060] The filtered cavitation pressure signal is saved, and modal time-frequency decomposition (MTD) is used to decompose the cavitation pressure signal to reduce signal complexity. The specific operation of the MTD is as follows:
[0061] First, all local extrema of the cavitation pressure signal after Kalman filtering are identified. The upper and lower envelopes are formed by cubic spline interpolation and the average envelope is solved. Then, the first IMF component that meets the conditions is obtained through iterative screening. After deducting this component, the screening process is repeated with the residual signal until the residual cavitation pressure signal is a monotonic or constant sequence. All IMF components and the residual cavitation pressure signal are output to complete the modal time-frequency decomposition of the cavitation pressure signal and realize the separation and extraction of the time-frequency features of the cavitation pressure signal.
[0062] In this embodiment, four types of cavitation pressure signals and cavitation images were selected. Each type of cavitation pressure signal and cavitation image was experimentally measured, resulting in a total of 1000 sets of cavitation pressure signals and cavitation images. Each set of pressure data had a dimension of 307200. Based on the measured cavitation pressure signals and cavitation images, a diagnostic model training set and a test set were constructed. The diagnostic model was then trained, with the training set and test set divided in an 8:2 ratio. Wherein:
[0063] The formula for modal time-frequency decomposition is:
[0064]
[0065] In the formula, This represents the equivalent representation of the original cavitation pressure signal in the time-frequency domain. To take the real part of a complex number, n This represents the number of intrinsic mode functions obtained by modal time-frequency decomposition technique. The first part is the decomposition of the original cavitation pressure signal. i One intrinsic mode function; For the first i Perform Hilbert transform on each eigenmode function.
[0066] Cavitation pressure signal decomposition, using Figure 3 The system uses nine sensors—the first sensor 2081, the second sensor 2082, the third sensor 2083, the fourth sensor 2084, the fifth sensor 2085, the sixth sensor 2086, the seventh sensor 2087, the eighth sensor 2088, and the ninth sensor 2089—to measure and monitor cavitation pressure signals. The data obtained from each measurement by the pressure sensors is then filtered by Kalman filtering to form a cavitation pressure signal matrix. , cti Indicates the emptying type, Indicates the first m Second measurement, n =307200 is the matrix dimension, i.e., the number of sampling points.
[0067]
[0068] Set appropriate K The value is obtained by using modal time-frequency decomposition techniques to analyze the matrix. Decompose, when K When the value is 10, mode aliasing and spurious components caused by under-decomposition and over-decomposition of the signal can be avoided. The matrix is obtained after time-frequency decomposition. .
[0069] =
[0070]
[0071] row vector The matrix after modal time-frequency decomposition, These correspond to sensors 2081 to 2089 respectively; n=307200 is the matrix dimension, i.e., the number of sampling points; p is the measured pressure value; these correspond to sensors 2081 to 2089 respectively. n =307200 is the matrix dimension, i.e., the number of sampling points; p is the measured pressure value; ct_i represents the cavitation type, i=0~3, ct_0 represents no cavitation, ct_1 represents slight cavitation, ct_2 represents moderate cavitation, ct_3 represents severe cavitation; s represents the s-th sample point of ct_i type cavitation, s<1000. This indicates the m-th measurement.
[0072] Furthermore, the final forms of the training and test sets were determined. Cavitation types were determined using image grayscale thresholding, and the 200 sets of cavitation pressure signals measured in the experiment were classified according to the image grayscale thresholds: Voidification Slight cavitation Moderate cavitation, Severe cavitation.
[0073] The cavitation visual image of the centrifugal pump is processed into grayscale image data to obtain single-channel grayscale image data. The core principle of grayscale conversion of cavitation visual images is as follows: The red, green, and blue color components of each pixel in the color cavitation image are assigned weighting coefficients based on the characteristics of human vision. A weighted average algorithm is used to calculate the single-channel grayscale value. This grayscale value replaces the original pixel's three-channel color values, transforming the color cavitation image into a single-channel grayscale image containing only brightness gradients. While removing redundant color information, key features such as the shape, boundary, and brightness contrast of the cavitation region are preserved, providing effective visual feature input for subsequent cavitation pattern diagnosis. To unify the numerical range of grayscale values, a grayscale value normalization mapping formula is further adopted:
[0074]
[0075] In the formula: The pixel grayscale value (integer). The amplitude of a single sampling point in the cavitation visual image. The minimum / maximum value of the cavitation pressure signal. This is a rounding function.
[0076] Linear mapping of grayscale values: Initial cavitation has a small pressure pulse amplitude, corresponding to low grayscale values in the grayscale image; severe cavitation has a large pressure pulse amplitude, corresponding to high grayscale values in the grayscale image. The distribution of light and dark areas in the grayscale image can reflect the difference in the degree of cavitation. By integrating the grayscale area of the grayscale image, the degree of cavitation is quantified; the larger the area, the more severe the cavitation.
[0077] S3. Combine the intrinsic mode function components and residual signal data with the single-channel grayscale image data to form a training set and a test set;
[0078] Through the extraction of key features of cavitation pressure signals and grayscale image processing of cavitation images, four types of cavitation pressure signal matrices are formed. , , , :
[0079]
[0080]
[0081]
[0082]
[0083] The number of non-empty sample points. The number of slightly cavitation sample points. Number of sample points with moderate cavitation Number of heavily cavitation sample points , This represents the average grayscale value. Specifically, data processing is performed on the TDMS file, selecting a centrifugal pump of 0.4. Q n 0.6 Q n 0.8 Q n 1.0 Q n 1.2 Q n 1.4 Q n Four cavitation data sources were used under flow conditions: no cavitation, slight cavitation, moderate cavitation, and severe cavitation.
[0084] The training set and test set were divided in an 8:2 ratio, resulting in a training set containing 800 sample points.
[0085] A test set containing 200 sample points: .
[0086] S4. Construct a centrifugal pump cavitation diagnosis model based on CNN-LAE-LSTM. The CNN is used to extract local key features of cavitation pressure signals and cavitation visual images; the LAE is used to reduce the dimensionality of features, remove redundancy, and suppress noise; the LSTM is used to capture the temporal correlation of cavitation pressure signals and adapt to the dynamic characteristics of centrifugal pump operation.
[0087] Specifically, refer to Figure 4 The centrifugal pump cavitation diagnostic model includes a cavitation data input layer, a training layer, and a testing layer. The training layer includes a two-dimensional double-layer convolutional layer, a pooling layer, a normalization layer, an LAE layer, an LSTM forgetting layer, an LSTM input layer, and an LSTM output layer.
[0088] 1) CNN automatically extracts key local features of cavitation pressure signals and cavitation visual images; the CNN feature extraction module mainly includes the following steps:
[0089] Step 1: CNN Two-Layer Convolution. The cavitation pressure signal of a centrifugal pump is a one-dimensional time-series signal. Its core is feature extraction using the first convolutional layer:
[0090]
[0091] In the formula, Indicates the first layer c One convolutional kernel, For the first c The output characteristic signal of one channel, ReLU is the preferred choice.
[0092] The second convolutional layer extracts deep fusion features.
[0093]
[0094] In the formula, , , indicating the second layer c Two convolutional kernels in the first... c Sub-convolution kernel on 1 input channel, b 2 represents the bias, and the summation operation represents the fusion of the cross-correlation results of all input channels.
[0095] Step 2: Pooling. Max pooling is performed on the cavitation pressure signal feature values as the output, and average pooling gray values are taken from the cavitation visual image as the output.
[0096] Max pooling:
[0097]
[0098] In the formula, S Step size;X The feature vector before pooling; F The pooling kernel size is represented by one dimension, and the length is the other dimension. k ; i The index of the output feature.
[0099] Average pooling: The arithmetic mean of the regions covered by the pooling kernel is taken as the output. The formula is:
[0100]
[0101] In the formula, The number of eigenvalues within the pooling region, i.e., the pooling kernel size. k .
[0102] Step 3: Normalization. By calculating the mean and variance of the features for each batch, the features are normalized to a standard normal distribution. Then, scaling and shifting parameters are used to preserve the expressive power of the features. This is the preferred method for processing time-series features in cavitation diagnosis.
[0103] Calculate the batch mean:
[0104] In the formula, For the current batch, m = H × W (Two-dimensional), or m=L (one-dimensional), x i This refers to a feature value in the batch.
[0105] Calculate batch variance:
[0106] standardization:
[0107] It is a local minimum.
[0108] 2) LAE is used for feature dimensionality reduction and redundancy removal, and noise suppression. The core of the LAE dimensionality reduction and purification module is to perform linear dimensionality reduction and purification on the high-dimensional features output by the CNN, involving both encoding and decoding processes. The encoder's input is the high-dimensional feature vector flattened by CNN pooling, and the output is a low-dimensional encoded vector.
[0109] A single-layer encoder maps the high-dimensional features output by the CNN to a low-dimensional encoded vector through a single linear transformation:
[0110]
[0111] In the formula, Z is the low-dimensional latent coding vector output by the encoder. W e This is the weight matrix for encoding; XThe high-dimensional feature vector input to LAE; b e This is the bias vector.
[0112] A single-layer decoder reconstructs the low-dimensional encoded vector Z into the input dimension X through a linear transformation, which is then used to calculate the reconstruction loss.
[0113]
[0114] In the formula, W d For single-layer decoder weights, b d This is the bias for a single-layer decoder.
[0115] 3) LSTM is used to capture the timing correlation of cavitation pressure signals.
[0116] The LSTM temporal modeling module mainly includes a forget gate, an input gate, and an output gate. These three gates work together to achieve selective memorization and forgetting of temporal features, thus solving the gradient vanishing problem of traditional RNNs.
[0117] Step 1: LSTM forget gate, ignoring invalid cavitation pressure signal noise. Sigmoid output ft This is a vector ranging from 0 to 1, where a value closer to 1 indicates more historical memory is retained, and a value closer to 0 indicates more information is forgotten. In cavitation diagnosis, if... t Initial cavitation characteristics are detected at all times, and the forget gate selectively forgets the historical cavitation pressure signal characteristics of normal operation.
[0118]
[0119] In the formula, This represents the feature vector output at time t. This represents the hidden state at time t-1. W f The weight matrix representing the forget gate; U f The hidden layer weight matrix represents the forget gate; b f The bias vector represents the forget gate; sigmoid is the activation function.
[0120] Step 2: LSTM Input Gate:
[0121]
[0122] In the formula, The sigmoid function controls the proportion of new information input. W i This represents the weight matrix of the input gate. Ui This represents the hidden layer weight matrix of the input gate; b i The bias vector of the input gate
[0123] Step 3: LSTM output gate:
[0124]
[0125] In the formula, The sigmoid function controls the proportion of new information input. W o This represents the weight matrix of the input gate. U o This represents the hidden layer weight matrix of the input gate; b o This represents the bias vector of the input gate.
[0126] Cavitation diagnostic testing was performed using the CNN-LAE-LSTM test set, and the final model accuracy was as follows: Figure 5 and Figure 6 As shown. If the accuracy If the output is positive, output the degree of vacuolation; otherwise, retrain.
[0127] S5. Train the centrifugal pump cavitation diagnostic model using the training set, and test the centrifugal pump cavitation diagnostic model using the test set.
[0128] Specifically, the cavitation pressure signal is saved in Excel format. Data is loaded from the Excel file and input into a CNN. The CNN extracts key local features from the cavitation pressure signal and the cavitation visual image. The high-dimensional local feature sequence output by the CNN is input into a LAE (Latent Image Extractor). The LAE reduces the dimensionality of the features, removes redundancy, and suppresses noise. The low-dimensional temporal latent features output by the LAE are input into an LSTM (Latent Streaming Mechanism). The LSTM captures the temporal correlation of the cavitation pressure signal. The ratio of the training set to the test set is designed to be 8:2. Training is repeated 100 times, with a batch size of 128 samples per round. A reasonable batch size ensures that the model covers a sufficient number of samples in each round of training, while also guaranteeing that each batch contains different types of cavitation states, avoiding unidirectional sample distribution.
[0129] The test set data was not used in model training. Its data distribution is consistent with the training set, containing all vacuolated states, and its preprocessing method is exactly the same as the training set. The data clustering after training is as follows: Figure 7As shown, the data was effectively distinguished. The trained CNN-LAE-LSTM model was switched to "evaluation mode," and the training-specific layers were disabled to avoid parameter changes during inference. The preprocessed test set was then input into the model in batches, forward propagation was performed, the predicted labels were output, and the model's guessed cavitation class was recorded, along with the actual cavitation class of the data. Each result was recorded, and the final output was the training accuracy. ,like Figure 5 . =Number of correctly predicted rounds / Total number of predicted rounds * 100%. Record the confusion matrix and analyze the confusion level of the four cavitation types. If the model accuracy > If the usage requirements are met, proceed to the cavitation data diagnostic stage. If the model accuracy is < If so, the training and test sets are reconstructed, and retraining is performed.
[0130] Cavitation pressure signals possess both "local impulsivity" and "temporal evolution," thus requiring multi-module collaboration: Local characteristics of cavitation pressure signals: The collapse of cavitation bubbles generates high-frequency impact pulses, so CNNs are used to extract these local details; High-dimensional redundancy of cavitation pressure signals: CNN outputs high-dimensional features and high noise levels, so LAEs are used to compress features and remove redundancy, while core cavitation information is preserved through reconstruction constraints; Temporal characteristics of cavitation pressure signals: Cavitation is a dynamic process that evolves from nothing to something, from weak to strong, with features evolving over time, so LSTMs are used to capture this temporal dependency.
[0131] S6. Input the data to be tested into the trained and tested centrifugal pump cavitation diagnosis model to obtain the fault diagnosis results;
[0132] Specifically, if the prediction accuracy is achieved Then, proceed to the prediction stage. Save the test data in Excel format and input it into the prediction layer of the CNN-LAE-LSTM model. Only retain the cavitation pressure signal of the test data, and use the CNN-LAE-LSTM model to predict the cavitation type of the cavitation pressure signal.
[0133] As described in the above embodiments, this application collects cavitation pressure signals of different degrees under the same flow rate of a centrifugal pump, as well as cavitation pressure signals under normal operating conditions. Kalman filtering and modal time-frequency decomposition are performed on the cavitation pressure signals to obtain the intrinsic mode function components and residual signal data of the denoised cavitation pressure signals. The cavitation visual images are processed to obtain single-channel grayscale image data. The intrinsic mode function components and residual signal data of the denoised cavitation pressure signals and the single-channel grayscale image data are combined to form a training set and a test set. A centrifugal pump cavitation diagnosis model based on CNN-LAE-LSTM is constructed. CNN is used to extract local key features of the cavitation visual and pressure signals; LAE is used to reduce the dimensionality of the features, remove redundancy, and suppress noise; LSTM is used to capture the temporal correlation of the signals and adapt to the dynamic characteristics of pump operation. The centrifugal pump cavitation diagnosis model is trained using the training set and the model accuracy is verified using the test set. The data to be predicted is input into the trained centrifugal pump cavitation diagnosis model to obtain the cavitation fault diagnosis result.
[0134] Corresponding to the aforementioned embodiments of the CNN-LAE-LSTM hybrid network centrifugal pump cavitation diagnostic method, this application also provides embodiments of a CNN-LAE-LSTM hybrid network centrifugal pump cavitation diagnostic device.
[0135] Figure 8 This is a block diagram of a CNN-LAE-LSTM hybrid network centrifugal pump cavitation diagnostic device according to an exemplary embodiment. (Refer to...) Figure 8 The device includes:
[0136] Acquisition module 1 is used to acquire cavitation data of different degrees under multiple sets of deviated flow and rated flow conditions of centrifugal pump. The cavitation data includes cavitation pressure signals and cavitation visual images.
[0137] Data processing module 2 is used to perform Kalman filtering and modal time-frequency decomposition on the cavitation pressure signal to obtain the intrinsic mode function components and residual signal data of the denoised cavitation pressure signal; and to perform grayscale image processing on the cavitation visual image to obtain single-channel grayscale image data.
[0138] Dataset construction module 3 is used to combine the intrinsic mode function components and residual signal data with the single-channel grayscale image data to form a training set and a test set;
[0139] Model building module 4 is used to build a centrifugal pump cavitation diagnosis model based on CNN-LAE-LSTM. The CNN is used to extract local key features of cavitation pressure signals and cavitation visual images; the LAE is used to reduce the dimensionality of features, remove redundancy, and suppress noise; the LSTM is used to capture the temporal correlation of cavitation pressure signals and adapt to the dynamic characteristics of centrifugal pump operation.
[0140] Training and testing module 5 is used to train the centrifugal pump cavitation diagnostic model using the training set and to test the centrifugal pump cavitation diagnostic model using the test set.
[0141] Diagnostic module 6 is used to input the data to be tested into the trained and tested centrifugal pump cavitation diagnostic model to obtain fault diagnosis results.
[0142] Regarding the apparatus in the above embodiments, the specific manner in which each module performs its operation has been described in detail in the embodiments related to the method, and will not be elaborated upon here.
[0143] For the device embodiments, since they basically correspond to the method embodiments, the relevant parts can be referred to in the description of the method embodiments. The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate, and the components shown as units may or may not be physical units, that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this application according to actual needs. Those skilled in the art can understand and implement this without creative effort.
[0144] Accordingly, this application also provides an electronic device, comprising: one or more processors; a memory for storing one or more programs; and when the one or more programs are executed by the one or more processors, causing the one or more processors to implement the CNN-LAE-LSTM hybrid network centrifugal pump cavitation diagnosis method as described above.
[0145] Accordingly, this application also provides a computer-readable storage medium storing computer instructions that, when executed by a processor, implement the CNN-LAE-LSTM hybrid network centrifugal pump cavitation diagnosis method described above.
[0146] Other embodiments of this application will readily occur to those skilled in the art upon consideration of the specification and practice of the disclosure herein. This application is intended to cover any variations, uses, or adaptations of this application that follow the general principles of this application and include common knowledge or customary techniques in the art not disclosed herein. The specification and embodiments are to be considered exemplary only.
[0147] It should be understood that this application is not limited to the precise structure described above and shown in the accompanying drawings, and various modifications and changes can be made without departing from its scope.
Claims
1. A CNN-LAE-LSTM hybrid network method for cavitation diagnosis of centrifugal pumps, characterized in that, include: Collect cavitation data of different degrees under multiple sets of deviated flow and rated flow conditions of centrifugal pumps. The cavitation data includes cavitation pressure signals and cavitation visual images. Kalman filtering and modal time-frequency decomposition are performed on the cavitation pressure signal to obtain the intrinsic mode function components and residual signal data of the denoised cavitation pressure signal; The cavitation visual image is subjected to grayscale image processing to obtain single-channel grayscale image data; The intrinsic mode function components and residual signal data are combined with the single-channel grayscale image data to form a training set and a test set; A cavitation diagnostic model for centrifugal pumps based on CNN-LAE-LSTM is constructed. The CNN is used to extract local key features of cavitation pressure signals and cavitation visual images; the LAE is used to reduce the dimensionality of features, remove redundancy, and suppress noise; and the LSTM is used to capture the temporal correlation of cavitation pressure signals to adapt to the dynamic characteristics of centrifugal pump operation. The centrifugal pump cavitation diagnostic model is trained using the training set and tested using the test set. The test data is input into the trained and tested centrifugal pump cavitation diagnostic model to obtain the fault diagnosis results.
2. The method according to claim 1, characterized in that, The cavitation pressure signal includes four cavitation conditions, each of which is determined by... j The signal consists of cavitation pressure signals from several sample points, and each cavitation pressure signal is detected by several pressure sensors.
3. The method according to claim 2, characterized in that, The cavitation pressure signal is detected by an array of pressure sensors. The pressure sensor array is set at multiple cross-sections on the outer wall of the inlet shell of the centrifugal pump. Three pressure sensors are evenly arranged circumferentially in each cross-section, with a spacing of 120°. The nine pressure sensors are used to detect pressure fluctuations at the leading edge, inside and tail of the inducer wheel, respectively.
4. The method according to claim 1, characterized in that, The specific operation of the Kalman filter is as follows: Initialization: Based on the prior characteristics of the cavitation pressure signal, initialize the state vector, state transition matrix, and observation matrix of the Kalman filter; State prediction: The cavitation pressure signal state value at the next time step is predicted by the state transition matrix, and the predicted state covariance matrix is updated by combining the process noise covariance matrix to complete the one-step prediction. Observation Update: Obtain the measured value of the cavitation pressure signal of the centrifugal pump in real time, calculate the residual between the predicted value and the measured value, use the predicted state covariance matrix and the observation noise covariance matrix as input to solve the Kalman gain, use the gain to correct the predicted state vector and covariance matrix, and obtain the optimal estimate of the cavitation pressure signal at the current time. The "state prediction-observation update" step is executed iteratively to filter the cavitation pressure signal at each time step throughout the entire time period, and finally outputs a filtered cavitation pressure signal sequence that suppresses environmental noise and disturbances.
5. The method according to claim 1, characterized in that, The specific operation of the modal time-frequency decomposition is as follows: First, all local extrema of the cavitation pressure signal after Kalman filtering are identified. The upper and lower envelopes are formed by cubic spline interpolation and the average envelope is solved. Then, the first IMF component that meets the conditions is obtained through iterative screening. After deducting this component, the screening process is repeated with the residual signal until the residual cavitation pressure signal is a monotonic or constant sequence. All IMF components and the residual cavitation pressure signal are output to complete the modal time-frequency decomposition of the cavitation pressure signal and realize the separation and extraction of the time-frequency features of the cavitation pressure signal.
6. The method according to claim 1 or 5, characterized in that, The formula for modal time-frequency decomposition is: ； ； In the formula, This represents the equivalent representation of the original cavitation pressure signal in the time-frequency domain. f Indicates frequency, t Indicates time, j It is the imaginary unit. This represents a temporary variable used for integration operations. To take the real part of a complex number, n This represents the number of intrinsic mode functions obtained by modal time-frequency decomposition technique. The first part is the decomposition of the original cavitation pressure signal. i One intrinsic mode function; For the first i Perform Hilbert transform on each eigenmode function.
7. The method according to claim 1, characterized in that, The specific operation for grayscale image processing of the cavitation visual image is as follows: The RGB three-channel values of each pixel in the cavitation visual image are extracted, and the weighted average of the three channels is calculated based on the grayscale algorithm. The average value is then assigned as the grayscale value of the corresponding pixel, thus completing the conversion of the image from color to a single-channel grayscale image.
8. The method according to claim 1, characterized in that, The centrifugal pump cavitation diagnostic model includes a cavitation data input layer, a training layer, and a testing layer. The training layer includes a two-dimensional double-layer convolutional layer, a pooling layer, a normalization layer, an LAE layer, an LSTM forgetting layer, an LSTM input layer, and an LSTM output layer, wherein: The cavitation data input layer is used to receive the intrinsic mode function components and residual signal data after modal time-frequency decomposition of the centrifugal pump, as well as the single-channel grayscale image data after grayscale image processing. The first convolutional layer of the two-dimensional double-layer convolutional layer is used to extract the spectral features of the cavitation pressure signal and visualize the cavitation region in the image; the second convolutional layer is used to extract the combination of cavitation pulses and the implicit features in the frequency domain. The pooling layer includes max pooling and average pooling. Max pooling is used to extract the feature values of the pressure signal and takes the maximum value as the output. Average pooling is used to extract the gray values of the cavitation image and takes the average value as the output. The normalization layer is used to standardize the input data; The LAE layer is used to extract features and reduce the dimensionality of the CNN input data. The LSTM forgetting layer is used to selectively ignore invalid cavitation pressure signal features; The LSTM input layer is used to selectively update the cavitation pressure signal features; The LSTM output layer is used to selectively output effective cavitation pressure signal features.
9. A CNN-LAE-LSTM hybrid network centrifugal pump cavitation diagnostic device, characterized in that, include: The acquisition module is used to acquire cavitation data of different degrees under multiple sets of deviated flow and rated flow conditions of the centrifugal pump. The cavitation data includes cavitation pressure signals and cavitation visual images. The data processing module is used to perform Kalman filtering and modal time-frequency decomposition on the cavitation pressure signal to obtain the intrinsic mode function components and residual signal data of the denoised cavitation pressure signal. The cavitation visual image is subjected to grayscale image processing to obtain single-channel grayscale image data; The dataset construction module is used to combine the intrinsic mode function components and residual signal data with the single-channel grayscale image data to form a training set and a test set; The model building module is used to build a centrifugal pump cavitation diagnosis model based on CNN-LAE-LSTM. The CNN is used to extract local key features of cavitation pressure signals and cavitation visual images; the LAE is used to reduce the dimensionality of features, remove redundancy, and suppress noise; the LSTM is used to capture the temporal correlation of cavitation pressure signals and adapt to the dynamic characteristics of centrifugal pump operation. The training and testing module is used to train the centrifugal pump cavitation diagnostic model using the training set and to test the centrifugal pump cavitation diagnostic model using the test set. The diagnostic module is used to input the data to be tested into the trained and tested centrifugal pump cavitation diagnostic model to obtain fault diagnosis results.
10. An electronic device, characterized in that, include: One or more processors; Memory, used to store one or more programs; When the one or more programs are executed by the one or more processors, the one or more processors implement the method as described in any one of claims 1-8.