Cavitation fault diagnosis method for pump unit based on multi-modal physical information neural network

Abstract

The application discloses a pump unit cavitation fault diagnosis method based on a multi-modal physical information neural network, comprising the following steps: calculating vibration logarithmic total energy and spectrum centroid based on vibration signal data, calculating current standard deviation based on current signal data, constructing and training a teacher model composed of three parallel independent sub-networks, constructing and training a student model, wherein the student model comprises a shared vibration encoder, a shared current encoder, a feature fusion layer, a fault classification head and a physical regression head, adopting a weighted sum of a classification loss, a physical consistency loss and a physical evolution trend constraint loss as a total loss function, training the student model through a knowledge distillation mode, and forcing the model to learn a feature manifold conforming to a cavitation physical evolution law. The application embeds a physical mechanism into a neural network through a PINN architecture, and solves the problems of lack of physical interpretability and weak generalization ability of a pure data-driven model.

Description

Technical Field

[0001] This invention relates to the field of fault diagnosis of fluid machinery, and specifically to a method for cavitation fault diagnosis of pump units based on a multimodal physical information neural network. Background Technology

[0002] Centrifugal pumps, as core equipment for industrial fluid transportation, require stable operation. Cavitation is one of the most common fluid dynamic failures in centrifugal pumps. When the local liquid pressure inside the pump drops to the saturated vapor pressure at the delivery temperature, the liquid vaporizes to form bubbles. These bubbles collapse instantaneously when they travel with the liquid flow to the high-pressure area, generating microjet streams and high-intensity shock waves. This continuous localized impact on the impeller surface leads to material erosion, induces severe vibration and noise, and in severe cases, causes a sharp decline in pump performance, head, and efficiency, or even catastrophic accidents such as shaft breakage. Therefore, early identification and severity assessment of cavitation failures are crucial for ensuring industrial safety.

[0003] With the development of artificial intelligence technology, data-driven fault diagnosis methods based on deep learning have gradually replaced traditional signal processing methods and become a current research hotspot. These methods typically collect vibration signals from pump units, automatically extract features through neural networks, and classify faults. Some researchers have proposed a centrifugal pump cavitation diagnosis method based on an improved stacked denoising autoencoder (RIME-SDAE). This method first performs singular value decomposition (SVD) denoising on the triaxial vibration signal, then manually extracts time-domain, frequency-domain, and time-frequency-domain features such as root mean square value, skewness, and sample entropy to construct a feature set, and finally uses the SDAE network for pattern recognition. Although this method achieves high accuracy on specific test sets, it essentially still relies on the distribution patterns of statistical features; the model is highly sensitive to feature selection and lacks a deep understanding of the physical meaning behind the features. Other researchers have proposed a water supply pump cavitation diagnosis method based on an adaptive convolutional neural network (CNN). This method directly uses slices of the original vibration signal as input and utilizes the powerful feature extraction capabilities of CNN to output the probability distribution of the cavitation state end-to-end. This study attempts to adapt the network structure to different working conditions by fine-tuning it, and to optimize it using transfer learning when data is scarce.

[0004] While the aforementioned data-driven methods perform well in experimental settings, they still suffer from significant "black box" defects and a lack of physical consistency in practical engineering applications. The lack of physical interpretability refers to the fact that pure data-driven models, such as CNNs and SDAEs, focus only on the nonlinear mapping between input data and output labels, ignoring the inherent physical operating mechanisms of the pump unit. The model cannot explain why it determines a signal as severely cavitation, which is unreliable in high-reliability industrial scenarios. Physical laws indicate that as cavitation develops from its initial stage to a severe stage, the energy released by bubble collapse typically increases, and the vibration intensity should increase accordingly. However, pure data models are not physically constrained, and when encountering noise interference or operating condition fluctuations not included in the training set, they are prone to outputting predictions that violate physical laws. For example, the model might misclassify a high-energy vibration signal as "normal," or predict that vibration energy decreases when cavitation intensifies. Such "non-monotonic" erroneous predictions are physically unacceptable. Therefore, they are prone to producing predictions that violate common sense physics. Summary of the Invention

[0005] To address the shortcomings of existing technologies, this invention provides a method for cavitation fault diagnosis of pump units based on a multimodal physical information neural network. This method transforms the physical mechanism into a differentiable loss function constraint, forcing the deep neural network to learn feature representations that conform to physical laws. The technical solution of this invention is as follows:

[0006] A method for cavitation fault diagnosis of pump units based on multimodal physical information neural networks includes the following steps:

[0007] S1: Collect vibration and current signal data of centrifugal pumps under different cavitation levels, divide them into time series samples according to a fixed length, label the corresponding cavitation conditions, and construct a training dataset;

[0008] S2: Calculate the two physical proxy indicators, total logarithmic energy of vibration and centroid of the spectrum, based on vibration signal data; calculate the physical proxy indicator, standard deviation of current, based on current signal data.

[0009] S3: Construct and train a teacher model consisting of three parallel independent sub-networks. Each sub-network performs regression prediction on one of three physical agent metrics. Each sub-network includes a vibration encoder, a current encoder, an attention fusion module, a weighted feature fusion module, and a regression prediction head.

[0010] S4: Construct and train a student model, which includes a shared vibration encoder, a shared current encoder, a feature fusion layer, a fault classification head, and a physical regression head; the fault classification head outputs the probability distribution of multiple working conditions, and the physical regression head outputs the predicted values ​​of three physical proxy indicators; load and freeze the pre-trained teacher model, and use the weighted sum of classification loss, physical consistency loss, and physical evolution trend constraint loss as the total loss function to train the student model through knowledge distillation;

[0011] S5: Use the trained student model to diagnose cavitation faults in pump units.

[0012] Furthermore, the total logarithmic energy of vibration is obtained by performing a fast Fourier transform on the vibration signal to obtain the power spectrum, calculating the average power spectrum of all channels, summing the power spectra across the entire frequency band and taking the logarithm; the centroid of the spectrum is obtained by calculating the weighted average frequency of the average power spectrum and normalizing it; the standard deviation of the current is obtained by calculating the mean of the standard deviations of all current channels over the time series.

[0013] Furthermore, the vibration encoder is composed of multiple one-dimensional residual convolutional blocks stacked together, which sequentially map the number of input channels to a high-dimensional feature space, and reduce the sequence length accordingly through convolutional layers; the current encoder adopts a structure symmetrical to the vibration encoder, but uses independent parameters.

[0014] Furthermore, during the training of each sub-network of the teacher model, mean squared error loss combined with attention entropy regularization loss is used for training. The network parameters are optimized through backpropagation algorithm, an early stopping strategy is adopted to prevent overfitting, and the model with the minimum validation loss is saved as the final teacher model.

[0015] Furthermore, the physical regression head includes three independent regression sub-networks, each with the same structure as the regression prediction head of the teacher model, predicting three physical surrogate indices respectively.

[0016] Furthermore, the attention fusion module concatenates the vibration feature vector output by the vibration encoder and the current feature vector output by the current encoder, processes them through two fully connected layers and a nonlinear activation function, and finally outputs the attention weights of the vibration feature vector and the current feature vector through a fully connected layer and another activation function.

[0017] The weighted feature fusion module uses the attention weights to sum the vibration feature vector and the current feature vector, and then performs a linear transformation on the spliced ​​features through a fully connected layer as a residual link to output the final fused features.

[0018] Furthermore, the calculation of the physical evolution trend constraint loss includes the following sub-steps:

[0019] (1) Calculate the mathematical expectation U of the category probability distribution output by the student model, thereby transforming the discrete fault categories into a continuous cavitation severity index;

[0020] (2) Within the training batch, the samples are sorted in ascending order based on the calculated U-values; the spectral centroid is selected as the observed physical state quantity, and the physical evolution gradient between adjacent samples is calculated:

[0021] (3) Construct a first-order evolution equation as a constraint objective;

[0022] (4) Calculate the mean square error between the predicted gradient and the target physical gradient, and use it as the evolution loss function.

[0023] A pump unit cavitation fault diagnosis device based on a multimodal physical information neural network includes one or more processors for implementing a pump unit cavitation fault diagnosis method based on a multimodal physical information neural network.

[0024] An electronic device, comprising:

[0025] One or more processors;

[0026] A storage device for storing one or more programs, which, when executed by the electronic device, enable the electronic device to implement a pump unit cavitation fault diagnosis method based on a multimodal physical information neural network.

[0027] A computer-readable storage medium having a program stored thereon, which, when executed by a processor, implements a method for cavitation fault diagnosis of pump units based on a multimodal physical information neural network.

[0028] The beneficial effects of this invention are as follows:

[0029] 1. Strong physical interpretability: Because this invention adopts a teacher-student knowledge distillation architecture based on PINN, it embeds explicit physical mechanisms into deep neural networks and uses physical consistency loss to force the model feature space to satisfy physical state equations. This achieves the technical effect of avoiding the output of erroneous results that violate physical common sense by traditional black-box models, and significantly improving the credibility of diagnostic results in industrial scenarios.

[0030] 2. Excellent generalization ability: Because this invention introduces a physical evolution trend constraint loss based on the idea of ​​partial differential equations, it constrains the model to follow the continuous physical evolution manifold of cavitation development in the latent space. This enables the model to maintain the monotonicity and continuity of predictions even when faced with intermediate conditions or noise interference not covered in the training data, effectively solving the technical problem of weak generalization ability of pure data-driven models.

[0031] 3. Decoupled design enables precise guidance: Because this invention designs a decoupled teacher model containing three parallel independent sub-networks, it avoids mutual interference between different physical indicators during feature extraction, thereby providing the student model with purer and more accurate physical knowledge supervision, enabling the lightweight student model to have complex physical perception capabilities. Attached Figure Description

[0032] Figure 1 This is a flowchart of a pump unit cavitation fault diagnosis method based on a multimodal physical information neural network, according to an embodiment of the present invention.

[0033] Figure 2 This is a training diagram for the teacher model.

[0034] Figure 3 This is a diagram illustrating the training of the student model.

[0035] Figure 4 A schematic diagram showing the predicted and actual values ​​of the total logarithmic energy of vibration predicted by the teacher model.

[0036] Figure 5 A schematic diagram showing the predicted and actual values ​​of the spectral centroids predicted by the teacher model.

[0037] Figure 6 A diagram illustrating the predicted and actual values ​​of the current standard deviation predicted by the teacher model.

[0038] Figure 7 A schematic diagram showing the predicted and actual values ​​of the total logarithmic energy of vibration predicted by the student model.

[0039] Figure 8 A diagram illustrating the predicted and actual values ​​of the spectral centroids for the student model.

[0040] Figure 9 A diagram illustrating the predicted and actual values ​​of the current standard deviation for the student model.

[0041] Figure 10 The confusion matrix diagram of the predicted and true labels for the student model's identification of cavitation states. Detailed Implementation

[0042] The present invention will be described in detail below with reference to the accompanying drawings and preferred embodiments. The purpose and effects of the present invention will become clearer. It should be understood that the specific embodiments described herein are only for explaining the present invention and are not intended to limit the present invention.

[0043] The cavitation fault diagnosis method for pump units based on multimodal physical information neural networks of this invention transforms the physical mechanism into a differentiable loss function constraint, forcing the deep neural network to learn feature representations that conform to physical laws. The technical approach of this method follows the bidirectional coupling principle of "physics driving data, data verifying physics," and adopts a two-stage knowledge distillation architecture, mainly including the following four core components:

[0044] (1) Construction of physical proxy indicators: Based on the cavitation mechanism of the pump unit, three proxy indicators with clear physical meanings are extracted from vibration and current signals; (2) Independent path teacher model training: A regression model with a multi-path parallel architecture is constructed to accurately learn the mapping relationship of physical indicators; (3) PINN student model construction: A classification network with dual-head output is designed to complete fault classification and physical indicator regression at the same time; (4) Physical constraint distillation training: Physical consistency loss and physical evolution trend loss are used to transfer the physical knowledge of the teacher model to the student model.

[0045] like Figure 1 As shown in the figure, the pump unit cavitation fault diagnosis method based on multimodal physical information neural network provided in this embodiment of the invention specifically includes the following steps:

[0046] Step 1: Collect vibration and current signal data of centrifugal pumps under different cavitation levels, divide the data into time series samples according to a fixed length, label the corresponding cavitation conditions, and construct a training dataset.

[0047] Step 2: Calculate two physical indices based on vibration signal data: the total logarithmic energy of vibration and the centroid of the spectrum. Calculate the standard deviation of the current based on the current signal data. The total logarithmic energy of vibration is obtained by performing a fast Fourier transform on the vibration signal to obtain the power spectrum. After calculating the average power spectrum of Dv channels, the sum of the power spectral densities across the entire frequency band is calculated and the logarithm is taken. The centroid of the spectrum is obtained by calculating the weighted average frequency of the vibration power spectrum and normalizing it. The standard deviation of the current is obtained by calculating the mean standard deviation of the Dc phase current over the time series. The target vector composed of these three physical surrogate indices will serve as the supervision target of the teacher model.

[0048] Based on the physical mechanism of cavitation failure in pump units, this method selects the following three statistical quantities with clear physical meaning and computability: total energy of vibration logarithm, centroid of spectrum, and standard deviation of current as physical proxy indicators.

[0049] 1. Total vibrational logarithmic energy

[0050] This index characterizes the intensity of the shock wave generated by the collapse of cavitation bubbles. The specific calculation method is as follows:

[0051] First, the vibration signal is preprocessed, that is, the original vibration signal is processed. (L is the number of sampling points, D)v The mean-reduced vibration signal is obtained by performing a mean-reduced processing on the vibration channel number. :

[0052]

[0053] Vibration signal after removing the mean Perform a fast Fourier transform, and then calculate the power spectrum P. v (f):

[0054]

[0055]

[0056] Where f represents the vibration frequency.

[0057] For D v The power spectra of each channel are averaged to obtain the average power spectrum. And calculate the total vibrational energy E total .

[0058]

[0059]

[0060] Where, d v Representing 1 to D v One of the vibration channels in the middle, f min To analyze the starting frequency of the frequency band, f max To analyze the cutoff frequency of the frequency band.

[0061] 2. Spectral centroid

[0062] This index characterizes the frequency component migration characteristics caused by cavitation noise. As cavitation deepens, the bubble collapse energy shifts from high frequencies to low frequencies, and the spectral centroid effectively reflects this trend. The calculation process for the spectral centroid is as follows:

[0063] First, the non-DC component is extracted to eliminate the influence of the DC component (f=0).

[0064]

[0065] Among them, F nonzero This represents the non-DC component of the vibration signal.

[0066] The centroid F of the spectrum is obtained by calculating the weighted average frequency of the average power spectrum. centroid :

[0067]

[0068] Finally, normalization is performed to normalize the centroid to the [0,1] interval, resulting in the normalized spectral centroid.

[0069]

[0070] 3. Current standard deviation

[0071] This index characterizes the pulsating effect of changes in the gas-liquid two-phase flow regime on the motor load torque. The calculation process for the current standard deviation is as follows:

[0072] The first step is to perform current detrending processing, removing the trend term from the current signal to highlight its fluctuation characteristics:

[0073]

[0074] Where LPF is a low-pass filter; d c Represents any current path. Indicates the dth c The original current signal of each current channel, Indicates the dth c The current signal after detrending processing of each current channel.

[0075] Then calculate the standard deviation of the three-phase currents and take the average.

[0076]

[0077] The above three physical quantity proxy indicators constitute the target vector. , as the supervision target of the teacher model.

[0078] Step 3: Construct and train a teacher model consisting of three parallel independent sub-networks. Each sub-network performs regression prediction on one of the three physical proxy metrics: predicted total energy of vibration logarithm, spectral centroid, and current standard deviation. Each sub-network includes a vibration encoder, a current encoder, an attention fusion module, a weighted feature fusion module, and a regression prediction head. The model is trained using mean squared error loss combined with attention entropy regularization loss. The network parameters are optimized using the backpropagation algorithm. An early stopping strategy is adopted to prevent overfitting. The model with the minimum validation loss is saved as the final teacher model.

[0079] like Figure 2 As shown, in order to accurately capture the changing patterns of different physical indicators, this invention constructs a multi-path parallel architecture teacher model. This teacher model consists of three independent sub-networks, each of which is specifically responsible for predicting a physical proxy indicator, thus avoiding interference between different physical quantities.

[0080] For the k-th physical proxy metric (k=1,2,3 corresponding to total energy, spectral centroid, and current standard deviation respectively), its subnetwork T(k) includes an input layer, a bimodal encoder, an attention fusion module, and a regression prediction head.

[0081] The vibration input of the input layer is represented as: Current input is represented as The number of current signal channels D here c The value is 3.

[0082] Dual-mode encoders include vibration encoders. and current encoder , used for feature extraction, where:

[0083] The vibration encoder consists of three stacked one-dimensional residual convolutional blocks, which sequentially increase the number of input channels from D... v The sequence length is mapped to 32-dimensional, 64-dimensional, and finally to a 128-dimensional feature space. At the same time, the sequence length is reduced to L, L / 2 and L / 4 respectively by convolutional layers with a stride of 2.

[0084] The current encoder adopts a symmetrical structure to the vibration encoder, but uses independent parameters to sequentially increase the number of input channels from D. c The sequence length is mapped to 32-dimensional, 64-dimensional, and finally to a 128-dimensional feature space. The sequence length reduction method is the same as that of the vibration encoder.

[0085] The output of each encoder is transformed into a 128-dimensional feature vector through a global average pooling operation:

[0086] Vibration eigenvectors

[0087] Current eigenvector .

[0088] Attention Fusion Module A (k) The vibration and current feature vectors are concatenated into a 256-dimensional vector, then processed through a two-layer fully connected network (128-dimensional in the first layer and 64-dimensional in the second layer) and the ReLU nonlinear activation function. Finally, a 2-dimensional fully connected layer and the Softmax function are used to output the attention weights for the two modes. ,satisfy .

[0089] The weighted feature fusion module uses attention weights to perform a weighted summation of the original feature vectors. Simultaneously, the concatenated features are linearly transformed through a 128-dimensional fully connected layer, serving as a residual connection. The final fused features are calculated as follows:

[0090]

[0091] Among them, FC128 [;] represents a 128-dimensional fully connected layer, and [;] represents vector concatenation.

[0092] Regression Prediction Head Each includes a 128-dimensional fully connected layer, a ReLU activation function, and a 32-dimensional fully connected layer. The fused features are then processed by a regression prediction head. After processing, the predicted value of the k-th physical index is finally output. ,Right now:

[0093]

[0094] To prevent the model from over-relying on a single modality, this method introduces an attention entropy regularization term into the loss function of each sub-network used to train the teacher model. For the k-th sub-network, the attention entropy regularization term... Defined as:

[0095]

[0096] Then, the total loss function of the k-th subnetwork is:

[0097]

[0098] Where N is the number of samples and λ is the entropy regularization coefficient.

[0099] The final output of the teacher model is a vector of predicted values ​​for three physical indicators: .

[0100] Step 4: Build and train the student model, reload the original multimodal data and its classification labels, and use the same preprocessing methods. The student model includes a shared encoder, a feature fusion layer, a fault classification head, and a physics regression head. The fault classification head outputs the probability distribution of multiple operating conditions, and the physics regression head outputs the predicted values ​​of three physical surrogate indicators. Load and freeze the pre-trained teacher model, using a weighted sum of classification loss, physical consistency loss, and physical evolution trend loss as the total loss function. Train the student model through knowledge distillation, enabling the student to learn physical laws while maintaining classification accuracy. Save the parameters of the student model with the highest validation accuracy.

[0101] like Figure 3 As shown, the student model adopts a shared encoder architecture, which reduces computational complexity while ensuring physical constraints. The student model accepts the same multimodal input as the teacher model, but extracts features through a dual encoder with shared parameters, and uses a dual-head output structure to simultaneously complete fault classification and physical index regression.

[0102] Shared encoders include shared vibration encoders and shared current encoders, wherein:

[0103] Shared vibration encoder The structure of the vibration encoder is identical to that of the teacher model, but its parameters are shared across the prediction tasks of the three physical surrogate indices. This shared vibration encoder consists of three stacked one-dimensional residual convolutional blocks, sequentially converting channel D... v The vibration input is mapped to 32-dimensional and 64-dimensional features, and the final output is a 128-dimensional feature representation. At the same time, the sequence length is reduced to L, L / 2 and L / 4 respectively.

[0104] Shared current encoder The structure of the shared current encoder is identical to that of the teacher model, but its parameters are shared across the prediction tasks of the three physical surrogate metrics. This shared current encoder consists of three stacked one-dimensional residual convolutional blocks, which sequentially... The channel current input is mapped to 32-dimensional and 64-dimensional, and the final output is a 128-dimensional feature representation. The sequence length reduction method is the same as that of the vibration encoder.

[0105] The encoder output is converted into a 128-dimensional feature vector through global average pooling, as follows:

[0106] Vibration eigenvectors

[0107] Current eigenvector .

[0108] The feature fusion layer concatenates the vibration feature vector and the current feature vector into a 256-dimensional vector. Then, it passes through a 256-dimensional fully connected layer, batch normalization, and the ReLU activation function. Finally, it passes through a 128-dimensional fully connected layer to generate a 128-dimensional fused feature. Specifically, it is expressed as:

[0109]

[0110] Dual-head output structure:

[0111] (1) Fault classification head The fused features are processed through a 128-dimensional fully connected layer and a ReL layer. The activation function is U, and finally, a Q-dimensional fully connected layer and a softmax function are used to output the probability distribution of the Q-class operating conditions. :

[0112]

[0113] (2) Physical regression head It contains three independent regression subnetworks, each with the same regression prediction head as the teacher model, predicting three physical surrogate indicators respectively. , , :

[0114]

[0115] in, Let represent the regression subnetwork of the k-th physical proxy index. Its structure is: 128-dimensional fully connected layer → ReLU activation function → 32-dimensional fully connected layer → output scalar value.

[0116] The student model receives physics knowledge guidance from the teacher model in the following three ways:

[0117] Soft tags provide: During the training phase, the teacher model's predicted physical metrics. This is passed to the student model as a "soft target," causing the student model's physics regression head to output... To get as close as possible to the teacher model's predictions.

[0118] Feature alignment constraints: fused features of the student model During training, it is constrained by physical consistency loss, so that its implicit encoding is consistent with physical laws in feature representation.

[0119] Gradient coupling mechanism: The gradient of the physical regression head is backpropagated to the shared encoder through the fused feature layer, so that the entire feature extraction process is constrained by physical laws, thereby improving the physical interpretability and generalization ability of the classification head.

[0120] This knowledge distillation mechanism ensures that the student model maintains high classification accuracy while its internal feature representation conforms to physical laws, thus achieving a deep integration of physical mechanisms and data-driven approaches.

[0121] This invention innovatively constructs a physical information loss function, transforming domain knowledge into differentiable constraint terms to guide the training process of the student model. The total loss function of the student model consists of three parts: data-driven classification loss, physical consistency loss, and physical evolution trend constraint loss.

[0122] (1) Data-driven classification loss

[0123] The data-driven classification loss uses the standard cross-entropy loss function to measure the performance of the classification task, ensuring that the model can accurately distinguish different degrees of cavitation. For a training batch containing N samples, the classification loss is calculated as follows:

[0124]

[0125] In the formula, y n,j This represents the true label of the nth sample, where j=1,2,3,…,Q represent Q working conditions, and p n,j This represents the probability that the nth sample, predicted by the student model, belongs to the jth class.

[0126] (2) Loss of physical consistency

[0127] Physical consistency loss, also known as knowledge distillation loss, requires that the student model's predicted physical indicators closely approximate the teacher model's predictions. During training, the trained teacher model is treated as a "numerical solver" of the physical system, its output representing the physical state the system should follow given input conditions. Physical consistency loss is defined as:

[0128]

[0129] in, and represents the predicted values ​​of the student model and the teacher model for the k-th physical proxy index of the n-th sample, where k=1,2,3 correspond to the total vibration energy, the centroid of the spectrum, and the standard deviation of the current.

[0130] The physical meaning of this loss lies in forcing the solution space of the student model to satisfy the implicit state equations defined by the teacher model:

[0131]

[0132] This is a data-driven agent-based process for minimizing the residuals of generalized partial differential equations, ensuring that the features extracted by the student model conform to the coupling mechanism of multiple physical quantities.

[0133] (3) Physical evolution trend constraint loss

[0134] This method, based on the evolutionary laws of cavitation development in fluid dynamics, constructs a physical evolution trend constraint loss function based on the idea of ​​partial differential equations. This loss function constrains the training process of the neural network to a manifold that conforms to the physical evolution equations.

[0135] The specific construction steps are as follows:

[0136] (3.1) Definition of the severity of continuous cavitation

[0137] To calculate the evolution gradient of physical quantities, we first calculate the mathematical expectation U of the class probability distribution P output by the student model, thereby transforming discrete fault categories into continuous cavitation severity indicators:

[0138]

[0139] Where j is the category index, P j This represents the corresponding predicted probability.

[0140] U is a continuous variable that can reflect subtle changes in cavitation development.

[0141] (3.2) Sample sorting and gradient calculation

[0142] Within the training batch, samples are sorted in ascending order based on the calculated U-values. The spectral centroid is selected as the observed physical state quantity. Calculate the physical evolution gradient between adjacent samples:

[0143]

[0144] Where i is the sorted sample index. Let be the centroid value of the spectrum predicted for the i-th sample.

[0145] (3.3) Construction of physical evolution equations

[0146] According to the cavitation mechanism, the spectral centroid exhibits a trend of "first rising and then falling" as cavitation develops. Therefore, the following first-order evolution equation is constructed as a constraint objective:

[0147]

[0148] U peak Physical inflection point parameter, representing the physical quantity at this stage. Reaching peak value; Evolution rate coefficient: the sensitivity to changes in physical quantities; This represents the median severity of adjacent samples.

[0149] (3.4) Calculation of evolution constraint loss

[0150] By minimizing the mean squared error between the "predicted gradient" and the "target physical gradient," the feature space learned by the model is forced to strictly follow the above physical equations:

[0151]

[0152] in This represents the number of valid sample pairs used in the calculation. Physical evolution trend loss is used to constrain characteristic manifolds to conform to physical evolution equations.

[0153] The total loss function of the student model is a weighted sum of the three losses mentioned above:

[0154]

[0155] in, and These are weighting coefficients used to balance the importance of each loss term. During training, this total loss function optimizes the student model's parameters using backpropagation, enabling the model to achieve both high classification accuracy and physical consistency.

[0156] Step 5: Conduct model evaluation and testing, namely, evaluate the model's classification performance and physical index prediction accuracy on the test set, and analyze the model performance by plotting the confusion matrix and feature space visualization; finally, use the trained student model to diagnose cavitation faults in the pump unit.

[0157] On the other hand, embodiments of the present invention also provide a pump unit cavitation fault diagnosis device based on a multimodal physical information neural network, including one or more processors, for implementing the pump unit cavitation fault diagnosis method based on a multimodal physical information neural network in the above embodiments.

[0158] The pump unit cavitation fault diagnosis device based on multimodal physical information neural network of this embodiment can be applied to any device with data processing capabilities, such as a computer. The device embodiment can be implemented in software, hardware, or a combination of both. Taking software implementation as an example, as a logical device, it is formed by the processor of any data processing device loading the corresponding computer program instructions from non-volatile memory into memory for execution. From a hardware perspective, in addition to the processor, memory, network interface, and non-volatile memory, the data processing device in this embodiment may also include other hardware depending on its actual functions, which will not be elaborated further.

[0159] This invention also provides a computer-readable storage medium storing a program thereon, which, when executed by a processor, implements the pump unit cavitation fault diagnosis method based on a multimodal physical information neural network described in the above embodiments.

[0160] The computer-readable storage medium can be an internal storage unit of any data processing device as described in any of the foregoing embodiments, such as a hard disk or memory. The computer-readable storage medium can also be an external storage device, such as a plug-in hard disk, smart media card (SMC), SD card, flash card, etc., equipped on the device. Furthermore, the computer-readable storage medium can include both internal storage units of any data processing device and external storage devices. The computer-readable storage medium is used to store the computer program and other programs and data required by the data processing device, and can also be used to temporarily store data that has been output or will be output.

[0161] The effectiveness of the method of the present invention will be further demonstrated through a specific application example below.

[0162] In this embodiment, different cavitation states are simulated by adjusting the inlet pressure of the centrifugal pump. Vibration and current signal data of the centrifugal pump under different cavitation levels are segmented to a fixed length, labeled with operating condition tags, and a complete dataset is formed. The dataset is then divided into training, validation, and test sets in a 7:1.5:1.5 ratio, maintaining a consistent distribution ratio for each category during the division process to ensure the representativeness of the training effect and the reliability of the evaluation results.

[0163] In this embodiment, the teacher model predicts the following values ​​for the three physical indices: total logarithmic energy of vibration, spectral centroid, and current standard deviation: Figure 4 , Figure 5 and Figure 6 As shown in Table 1, the accuracy results of the three physical agent indicators predicted by the teacher model are presented.

[0164] Table 1. Accuracy Results of Physical Agent Volume Indicators

[0165]

[0166] Test results show that the predictions of the three physical proxy indicators prove that the teacher model can accurately fit the physical laws.

[0167] In this embodiment, the predicted and actual values ​​of the total logarithmic energy of vibration predicted by the student model are as follows: Figure 7 As shown, the predicted and actual values ​​of the predicted spectral centroid are as follows: Figure 8 As shown, the predicted and actual values ​​of the predicted current standard deviation are as follows: Figure 9 As shown in the three figures, the predicted values ​​of the three physical surrogate quantities exhibit highly similar waveform trends to the actual values ​​under various cavitation conditions, indicating that the predictions of the three physical surrogate quantities for cavitation conditions have physical significance. The cavitation state results identified by the student model are shown below. Figure 10 As shown in the figure, the diagonal line indicates that the predicted label is consistent with the true label, and the result of the label not being on the diagonal line indicates that the prediction result is biased. The prediction accuracy is 91.3%.

[0168] It will be understood by those skilled in the art that the above descriptions are merely preferred examples of the invention and are not intended to limit the invention. Although the invention has been described in detail with reference to the foregoing examples, those skilled in the art can still modify the technical solutions described in the foregoing examples or make equivalent substitutions for some of the technical features. All modifications and equivalent substitutions made within the spirit and principles of the invention should be included within the scope of protection of the invention.

Claims

1. A method for cavitation fault diagnosis of pump units based on multimodal physical information neural networks, characterized in that, Includes the following steps: S1: Collect vibration and current signal data of centrifugal pumps under different cavitation levels, divide them into time series samples according to a fixed length, label the corresponding cavitation conditions, and construct a training dataset; S2: Calculate the two physical proxy indicators, total logarithmic energy of vibration and centroid of the spectrum, based on vibration signal data; calculate the physical proxy indicator, standard deviation of current, based on current signal data. S3: Construct and train a teacher model consisting of three parallel independent sub-networks. Each sub-network performs regression prediction on one of three physical agent metrics. Each sub-network includes a vibration encoder, a current encoder, an attention fusion module, a weighted feature fusion module, and a regression prediction head. S4: Construct and train a student model, which includes a shared vibration encoder, a shared current encoder, a feature fusion layer, a fault classification head, and a physical regression head; the fault classification head outputs the probability distribution of multiple working conditions, and the physical regression head outputs the predicted values ​​of three physical proxy indicators; load and freeze the pre-trained teacher model, and use the weighted sum of classification loss, physical consistency loss, and physical evolution trend constraint loss as the total loss function to train the student model through knowledge distillation; S5: Use the trained student model to diagnose cavitation faults in pump units; The calculation of the physical evolution trend constraint loss includes the following sub-steps: (1) Calculate the expected value of the class probability distribution of the student model output. U This transforms discrete fault categories into continuous cavitation severity indicators. (2) Within the training batch, based on the calculated U The samples are sorted in ascending order by value; the spectral centroid is selected as the observed physical state quantity, and the physical evolution gradient between adjacent samples is calculated; (3) Construct a first-order evolution equation as a constraint objective; (4) Calculate the mean square error between the predicted gradient and the target physical gradient, and use it as the evolution loss function.

2. The method for cavitation fault diagnosis of pump units based on multimodal physical information neural networks according to claim 1, characterized in that, The total logarithmic energy of vibration is obtained by performing a fast Fourier transform on the vibration signal to obtain the power spectrum, calculating the average power spectrum of all channels, summing the power spectra across the entire frequency band and taking the logarithm; the centroid of the spectrum is obtained by calculating the weighted average frequency of the average power spectrum and normalizing it; the standard deviation of the current is obtained by calculating the mean of the standard deviations of all current channels over the time series.

3. The method for cavitation fault diagnosis of pump units based on multimodal physical information neural networks according to claim 1, characterized in that, The vibration encoder is composed of multiple stacked one-dimensional residual convolutional blocks, which sequentially map the number of input channels to a high-dimensional feature space and reduce the sequence length accordingly through convolutional layers; the current encoder adopts a structure symmetrical to the vibration encoder, but uses independent parameters.

4. The method for cavitation fault diagnosis of pump units based on multimodal physical information neural networks according to claim 1, characterized in that, During training, each sub-network of the teacher model is trained using mean squared error loss combined with attention entropy regularization loss. The network parameters are optimized through backpropagation, and an early stopping strategy is adopted to prevent overfitting. The model with the minimum validation loss is saved as the final teacher model.

5. The method for cavitation fault diagnosis of pump units based on multimodal physical information neural networks according to claim 1, characterized in that, The physical regression head includes three independent regression subnetworks, each with the same structure as the regression prediction head of the teacher model, and each predicts three physical surrogate indices.

6. The method for cavitation fault diagnosis of pump units based on multimodal physical information neural networks according to claim 1, characterized in that, The attention fusion module concatenates the vibration feature vector output by the vibration encoder and the current feature vector output by the current encoder, processes them through two fully connected layers and a nonlinear activation function, and finally outputs the attention weights of the vibration feature vector and the current feature vector through one fully connected layer and another activation function. The weighted feature fusion module uses the attention weights to sum the vibration feature vector and the current feature vector, and then performs a linear transformation on the spliced ​​features through a fully connected layer as a residual link to output the final fused features.

7. A pump unit cavitation fault diagnosis device based on a multimodal physical information neural network, characterized in that, It includes one or more processors for implementing the pump unit cavitation fault diagnosis method based on a multimodal physical information neural network as described in any one of claims 1 to 6.

8. An electronic device, characterized in that, include: One or more processors; A storage device for storing one or more programs, which, when executed by the electronic device, cause the electronic device to implement the pump unit cavitation fault diagnosis method based on a multimodal physical information neural network as described in any one of claims 1 to 6.

9. A computer-readable storage medium, characterized in that, It stores a program that, when executed by a processor, implements the pump unit cavitation fault diagnosis method based on a multimodal physical information neural network as described in any one of claims 1 to 6.

Images