Axial flow pump cavitation state identification method based on 1D-CNN

Abstract

The invention discloses a 1D-CNN (1D-Convolutional Neural Network)-based axial flow pump cavitation state identification method. The method comprises the following steps: acquiring pressure pulsation signal data after sampling a pressure pulsation signal of an axial flow pump; performing standardized preprocessing on the pressure pulsation signal data to construct an original data set; dividing a training set and a test set; a 1D-CNN neural network model is constructed; the 1D-CNN neural network model is trained, model parameters are optimized, and the trained 1D-CNN neural network model is acquired; and based on the trained 1D-CNN neural network model and the test set, outputting an identification result. The 1D-CNN in the application can directly process the original time domain signal without feature extraction, the recognition accuracy is high, the test set can reach 90.83%, the performance is particularly excellent in the recognition of critical and serious cavitation states, and the method is suitable for a high-precision diagnosis scene.

Description

Technical Field

[0001] The present application belongs to the field of fluid machinery technology, and specifically relates to a cavitation state identification method for an axial flow pump based on 1D-CNN. Background Art

[0002] Axial flow pumps are core power equipment in water conservancy projects, ship propulsion, energy transmission and other fields, and their operating status directly affects the efficiency and reliability of the system. Cavitation, a typical fault phenomenon in the operation of axial flow pumps, occurs when the local pressure is lower than the saturated vapor pressure of the fluid, triggering the generation and collapse of bubbles, resulting in a sudden drop in pump efficiency, increased vibration, and erosion of flow-through components, seriously threatening equipment life and system safety. Traditional cavitation monitoring relies heavily on vibration signal analysis, acoustic emission detection, or pressure pulsation monitoring, but faces significant challenges: empirical methods based on threshold judgment have weak generalization capabilities and are difficult to cope with complex working conditions; time-frequency domain feature extraction relies on manual design, is susceptible to noise interference, and lacks information integrity.

[0003] Patent CN117633520B proposes a recursive image-based method for detecting cavitation initiation in axial-flow turbines. This method obtains a cavitation dataset from a target turbine, including vibration signals and corresponding cavitation state labels at different measuring points on the turbine as cavitation signal data. The vibration signals at different measuring points on the turbine are then constructed into a vibration signal matrix, which is then reconstructed into a phase space. The vibration signal matrix is ​​then mapped into a high-dimensional phase space to create a multivariate phase diagram. The relative distances between phase points within the multivariate phase diagram are then calculated to create a recursive image. An axial-flow turbine cavitation initiation detection model is then constructed and trained. Vibration signals are collected online in real time using sensors, and recursive images are adaptively calculated to detect turbine cavitation. However, this method suffers from data quality issues that are susceptible to interference from multiple factors in actual use, relies on artificial feature construction, and involves a complex process. Consequently, its real-time and anti-interference capabilities are insufficient.

[0004] Traditional visual inspection methods based on image processing or 2D-CNN rely on high-speed cameras, which are costly and difficult to deploy in real time. With the increasing demand for intelligent operations and maintenance in modern industry, there is an urgent need for a cavitation state recognition method that can directly process raw time series signals and achieve both high accuracy and low latency. Summary of the Invention

[0005] The present application provides an axial flow pump cavitation state recognition method based on 1D-CNN to solve the technical problems of existing axial flow pump cavitation state recognition methods relying on high-speed camera equipment, high computational redundancy and insufficient real-time performance.

[0006] To solve the above technical problems, a technical solution adopted in this application is: a cavitation state identification method for an axial flow pump based on 1D-CNN, comprising:

[0007] S1. Based on the pressure sensor, obtain the sampled pressure pulsation signal data;

[0008] S2. Perform standardization preprocessing on the pressure pulsation signal data to construct an original data set;

[0009] S3. Based on the original dataset, divide the dataset into training and test sets.

[0010] S4. Build a 1D-CNN neural network model;

[0011] S5. Based on the training set, train the 1D-CNN neural network model and optimize the model parameters to obtain a trained 1D-CNN neural network model;

[0012] S6. Based on the trained 1D-CNN neural network model and test set, output the recognition results.

[0013] Furthermore, the method for constructing a 1D-CNN neural network model in step S4 includes:

[0014] S41. Construct a feature extraction module based on convolutional layers, batch normalization layers, and ReLU activation functions. The convolutional layers are connected by pooling layers.

[0015] S42. Construct a 1D-CNN neural network model based on the input layer, feature extraction module, fully connected layer, and output layer.

[0016] Furthermore, the method for constructing the 1D-CNN neural network model in step S41 includes:

[0017] Based on formula (1), construct the ReLU activation function;

[0018] f(x)=max(0,x) (1);

[0019] When x>0, output x; when x≤0, output 0.

[0020] Furthermore, the method of training the 1D-CNN neural network model in step S5 includes:

[0021] The model weights are optimized based on the back-propagation algorithm to minimize the loss function, which consists of the cross-entropy loss and the L2 regularization term.

[0022] Further, methods for optimizing model weights include:

[0023] Based on formulas (2)-(3), the first-order moment and second-order moment of the gradient are calculated; where formulas (2)-(3) are:

[0024] m t =β1m t-1 +(1-β1)g t (2);

[0025]

[0026] Based on formulas (4)-(5), the bias correction is obtained; where (4)-(5) are:

[0027]

[0028]

[0029] Among them, g t represents the current time step gradient, β1 and β2 represent the decay rate, m t and v t Represents the first and second moment estimates of the gradient.

[0030] Based on formula (6), update the model weights; where formula (6) is:

[0031]

[0032] Among them, θ t+1 is the updated model weight, η represents the initial learning rate, which is set to 1×10 -3 , ε represents a numerical stability constant.

[0033] Furthermore, the method for obtaining the loss function includes:

[0034] Based on formula (7), the cross entropy loss is obtained; wherein, formula (7) is:

[0035]

[0036] Among them, N represents the number of samples, C represents the number of categories, and y i,c represents the one-hot encoding of the true label, Represents the class probability predicted by the model.

[0037] Based on formula (8), the L2 regularization term is obtained; wherein formula (8) is:

[0038]

[0039] Among them, λ is the regularization coefficient, which is set to 1×10 -3 , Represents all trainable weights of the network.

[0040] The beneficial effects of this application are as follows: the 1D-CNN in this application can directly process raw time-domain signals without feature extraction, achieving a high recognition accuracy rate of 90.83% on the test set. It performs particularly well in identifying critical and severe cavitation states, making it suitable for high-precision diagnostic scenarios. Through the 1D-CNN's ability to abstract the features of raw time-series signals and its lightweight design, it achieves a technological leap from "offline analysis" to "online diagnosis" in cavitation monitoring, providing a reliable solution for the intelligent operation and maintenance of industrial pump fleets. BRIEF DESCRIPTION OF THE DRAWINGS

[0041] Figure 1 1D-CNN-based axial flow pump cavitation state identification method embodiment of the present application is a flow chart;

[0042] Figure 2 This is a structural block diagram of a 1D-CNN neural network model of an embodiment of a cavitation state identification method for an axial flow pump based on 1D-CNN of the present application;

[0043] Figure 3 This is a diagram of the prediction results of the training set of the 1D-CNN neural network model in the first embodiment of the axial flow pump cavitation state identification method based on 1D-CNN of the present application;

[0044] Figure 4 This is a test set prediction result diagram of the 1D-CNN neural network model of the first embodiment of the axial flow pump cavitation state identification method based on 1D-CNN of the present application;

[0045] Figure 5 This is a confusion matrix diagram of the 1D-CNN neural network model training set in the first embodiment of the axial flow pump cavitation state identification method based on 1D-CNN of the present application;

[0046] Figure 6 This is a confusion matrix diagram of the 1D-CNN neural network model test set of an embodiment of the 1D-CNN-based axial flow pump cavitation state recognition method of the present application. DETAILED DESCRIPTION

[0047] In order to make the objectives, technical solutions and advantages of the present invention more clear, the present invention is further described in detail below with reference to specific embodiments.

[0048] In the following description, many specific details are set forth to facilitate a full understanding of the present invention. However, the present invention may also be implemented in other ways different from the description. Therefore, the present invention is not limited to the specific embodiments disclosed in the following specification.

[0049] See Figure 1 , Figure 1This is a flow chart of an embodiment of a method for identifying cavitation states in an axial flow pump based on 1D-CNN. The method includes:

[0050] S1. Based on the pressure sensor, obtain the sampled pressure pulsation signal data.

[0051] Specifically, the pressure pulsation signal of the axial flow pump at different cavitation stages is collected by a pressure sensor, and the pressure pulsation signal data after the pressure pulsation signal of the axial flow pump is sampled is obtained.

[0052] S2. Perform standardized preprocessing on the pressure pulsation signal data to construct the original data set.

[0053] Specifically, the pressure pulsation signals of the axial flow pumps in different cavitation states collected in the above steps were normalized and denoised. Three consecutive cycles of data were extracted as input features for a single sample. Four typical cavitation states—non-cavitation, incipient cavitation, critical cavitation, and severe cavitation—were labeled categories 1, 2, 3, and 4, respectively. 100 samples were selected for each category, for a total of 400 samples in the original dataset.

[0054] S3. Based on the original dataset, divide it into training set and test set.

[0055] Specifically, we used a commonly used 7:3 training set to test set ratio for the original dataset to ensure adequate model training and generalization. Specifically, we selected the first 280 samples from the total sample as the training set for parameter learning of the convolutional neural network model; the remaining 120 samples were used as the test set to evaluate the model's classification performance on unseen samples.

[0056] S4. Build a 1D-CNN neural network model.

[0057] For details, see Figure 2The feature extraction module consists of a one-dimensional convolutional layer of size [2×1], a batch normalization layer, and a ReLU activation function. The 1D-CNN neural network model includes an input layer, a feature extraction module consisting of multiple convolutional and pooling layers, a fully connected layer, and an output layer. In terms of network architecture design, the input layer has dimensions of [1860×1×1], consistent with the number of points contained in each sample in the time domain, fully preserving the temporal characteristics of the original signal. The number of convolutional channels in the first layer is set to 16, and the second layer to 32, enabling the extraction of local features of different scales layer by layer. Max pooling layers of size [2×1] and stride [2×1] are introduced between convolutional layers to achieve spatial downsampling, thereby reducing feature dimensionality and improving computational efficiency. After feature extraction, a fully connected layer is connected for feature fusion and class mapping. Finally, the predicted probability of each cavitation state is output through the Softmax function, and the classification layer provides the final prediction result.

[0058] Among them, ReLU (RectifiedLinearUnit) activation function:

[0059] f(x)=max(0,x) (1);

[0060] When x>0, output x; when x≤0, output 0.

[0061] ReLU activation: First, the input data passes through the convolution layer for linear feature extraction, and then the convolution output is standardized (mean subtraction, standard deviation division, and application of learnable scaling and translation parameters) to improve training stability. The normalized value is input element by element into the ReLU function, thereby introducing nonlinearity and enhancing the model's expressiveness.

[0062] S5. Based on the training set, train the 1D-CNN neural network model and optimize the model parameters to obtain a trained 1D-CNN neural network model.

[0063] Specifically, the 1D-CNN neural network model is trained through the training set. The model continuously optimizes the network weights through the back propagation algorithm to minimize the loss function and improve the classification accuracy. The initial learning rate is set to 1×10 -3 , the L2 regularization coefficient is set to 1×10 -4 , to prevent the model from overfitting. Finally, the generalization ability and recognition effect of the model are verified through the test set, thus ensuring the effectiveness and stability of the classification model under actual working conditions.

[0064] The specific method for optimizing model weights is the Adam (Adaptive Moment Estimation) algorithm. The Adam algorithm is an adaptive learning rate optimization algorithm, and its weight update formula is as follows:

[0065] First calculate the first-order moment (mean) and second-order moment (variance) of the gradient:

[0066] m t =β1m t-1 +(1-β1)g t (2);

[0067]

[0068] Among them, g t represents the current time step gradient, β1 and β2 represent the decay rate (the default values ​​are 0.9 and 0.999 respectively), m t and v t Represents the first and second moment estimates of the gradient.

[0069] Based on formulas (4)-(5), the bias correction (bias for the initial time step) is obtained; where (4)-(5) are:

[0070]

[0071]

[0072] Based on formula (6), update the model weights; where formula (6) is:

[0073]

[0074] Among them, θ t+1 is the updated model weight, η represents the initial learning rate, which is set to 1×10 -3 , ε represents the numerical stability constant (default 10-8).

[0075] The introduction of a piecewise attenuation strategy improves the stability of the model's later training compared to a fixed learning rate. The expression is as follows:

[0076] η t =η×0.01 [t / m] ;

[0077] Among them, η is the initial learning rate set, t represents the current number of iterations, and m is the iteration threshold for adjusting the learning rate. After m training times, the learning rate is adjusted to η. t .

[0078] Since it is a classification task, the loss function consists of cross entropy loss and L2 regularization term.

[0079] Based on formula (7), the cross entropy loss is obtained; wherein, formula (7) is:

[0080]

[0081] Among them, N represents the number of samples, C represents the number of categories, and y i,c represents the one-hot encoding of the true label, Represents the class probability predicted by the model.

[0082] Based on formula (8), the L2 regularization term is obtained; wherein formula (8) is:

[0083]

[0084] Among them, λ is the regularization coefficient, which is set to 1×10 -3 , Represents all trainable weights of the network.

[0085] S6. Based on the trained 1D-CNN neural network model and test set, output the recognition results.

[0086] Specifically, the input data of the test set is passed through layers such as convolution, pooling, and full connectivity, followed by a forward propagation of the predicted probabilities using Softmax. The total loss is calculated by comparing the predicted results with the true labels. Backpropagation is then used to calculate the gradient of the loss function with respect to the weights. Ultimately, the weights are updated using the formula above to gradually reduce the total loss. The Adam algorithm adaptively adjusts the learning rate, combines cross-entropy loss with L2 regularization, and backpropagates the gradient to minimize the total loss function. This process balances the model's fitting ability and generalization performance, ultimately achieving high-precision classification.

[0087] In terms of model training, the present invention adopts the Adam optimization algorithm to update parameters, which has good adaptive learning ability. During the training process, a segmented learning rate decay strategy is adopted, and the initial learning rate is set to 1×10 -3 , the L2 regularization coefficient is set to 1×10 -4 To prevent model overfitting, the training samples are randomly shuffled during each round of training to improve the network's generalization and robustness. Furthermore, a visual monitoring module is enabled during training to observe the changing trends of the model's loss function and recognition accuracy in real time, facilitating analysis and adjustment of model convergence.

[0088] Table 1 shows the classification prediction results of the 1D-CNN model.

[0089]

[0090] See Figure 3-6The 1D-CNN model achieved an overall classification accuracy of 93.21% on the training set and 90.83% on the test set. The 1D-CNN network employed in this study demonstrated high end-to-end recognition accuracy even with direct input of the original signal. It was particularly stable and reliable in distinguishing severe cavitation states, demonstrating the effectiveness and adaptability of 1D-CNN in classifying nonlinear and non-stationary hydraulic signals.

[0091] The above description is merely an embodiment of the present application and does not limit the patent scope of the present application. Any equivalent structure or equivalent process transformation made using the contents of the present application specification and drawings, or directly or indirectly applied in other related technical fields, are also included in the patent protection scope of the present application.

Claims

1. A cavitation state recognition method for an axial flow pump based on 1D-CNN, characterized in that: The following steps are involved: S1. Based on the pressure sensor, obtain the sampled pressure pulsation signal data; S2. performing standardized preprocessing on the pressure pulsation signal data to construct an original data set; S3. Based on the original data set, divide the training set and the test set; S4. Build a 1D-CNN neural network model; S5. Based on the training set, the 1D-CNN neural network model is trained and the model parameters are optimized to obtain a trained 1D-CNN neural network model; S6. Output the recognition result based on the trained 1D-CNN neural network model and the test set.

2. The method according to claim 1, characterized in that The method for constructing a 1D-CNN neural network model in step S4 includes: S41. Construct a feature extraction module based on a convolutional layer, a batch normalization layer, and a ReLU activation function; wherein the convolutional layers are connected by a pooling layer; S42. Construct the 1D-CNN neural network model based on the input layer, the feature extraction module, the fully connected layer and the output layer.

3. The method according to claim 1, characterized in that The method for constructing a 1D-CNN neural network model in step S41 includes: Based on formula (1), the ReLU activation function is constructed; f(x)=max(0,x) (1); When x>0, output x; when x≤0, output 0.

4. The method according to claim 1, wherein The method for training the 1D-CNN neural network model in step S5 includes: The model weights are optimized based on the back-propagation algorithm to minimize the loss function, wherein the loss function consists of a cross-entropy loss and an L2 regularization term.

5. The method according to claim 4, characterized in that Methods for optimizing model weights include: Based on formulas (2)-(3), the first-order moment and the second-order moment of the gradient are calculated; wherein the formulas (2)-(3) are: m t =β1m t-1 +(1-β1)g t (2); Based on formulas (4)-(5), the deviation correction is obtained; wherein, (4)-(5) are: Among them, g t represents the current time step gradient, β1 and β2 represent the decay rate, m t and v t Represents the first and second moment estimates of the gradient. Based on formula (6), update the model weight; wherein, formula (6) is: Among them, θ t+1 is the updated model weight, η represents the initial learning rate, which is set to 1×10 -3 , ε represents a numerical stability constant.

6. The method according to claim 5, characterized in that Methods for obtaining loss functions include: Based on formula (7), the cross entropy loss is obtained; wherein, the formula (7) is: Among them, N represents the number of samples, C represents the number of categories, and y i,c represents the one-hot encoding of the true label, Represents the class probability predicted by the model. Based on formula (8), the L2 regularization term is obtained; wherein, the formula (8) is: Where λ is the regularization coefficient, which is set to 1×10 -3 ; Represents all trainable weights of the network.

Images