Pneumatic control valve fault diagnosis method and device based on microwave vibration measurement

Abstract

The invention discloses a pneumatic control valve fault diagnosis method and device based on microwave vibration measurement, and relates to the technical field of industrial process control. The method comprises the steps that a non-contact control valve rod vibration signal is collected; constructing a multi-scale time-frequency domain dual-channel feature fusion network structure, and extracting time sequence dynamic features of a valve rod vibration signal in a time domain branch through combination of multi-scale one-dimensional convolution and a bidirectional gating circulation unit; the frequency domain branch converts the one-dimensional signal into a two-dimensional time-frequency image, and uses a multi-scale two-dimensional convolutional network to extract spectrum texture features; introducing a channel attention mechanism to adaptively learn feature importance weight, and adopting a cross attention mechanism to deeply fuse time-frequency domain features; and based on a preset control valve fault category, performing fault classification on the fused control valve time-frequency features to obtain a control valve classification result. According to the method, complementary information of different modes is fully excavated, fault features of the control valve are fully extracted, and therefore accurate diagnosis of the faults of the control valve is achieved.

Description

Technical Field

[0001] This application relates to the field of industrial process control technology, and in particular to a method and device for diagnosing faults in pneumatic control valves based on microwave vibration measurement. Background Technology

[0002] Control valves, as key actuators in industrial process control systems, are widely used in process industries such as petrochemicals, power generation, metallurgy, and pharmaceuticals. Their operating status directly affects the safety, stability, and economy of the production process. Operating long-term in harsh environments such as high temperature, high pressure, and strong corrosion, they are prone to various failure modes, including valve seat corrosion, valve plug jamming, and actuator failure. Therefore, research on control valve fault diagnosis technology to achieve accurate fault identification is of great significance for ensuring industrial production safety and improving equipment reliability. In recent years, control valve fault diagnosis technology has developed rapidly, and existing research methods can be mainly divided into two categories: model-based diagnostic methods and data-driven diagnostic methods.

[0003] In existing technologies, control valve fault diagnosis mainly relies on the analysis of signals such as pressure, flow rate, and valve body vibration. To address the challenges of processing multivariate time series data and poor generalization of unseen faults in multi-valve system fault diagnosis, Jeong et al. constructed a one-dimensional convolutional neural network (1D CNN) model based on flow and pressure data collected by sensors. By extracting temporal features through convolutional blocks and fully connected layers, they achieved fault valve location identification and severity quantification. The model demonstrated robust performance for both trained and unseen fault scenarios, providing an efficient solution for valve fault diagnosis.

[0004] However, flow and pressure signals have a delayed response and are easily affected by factors such as operating condition fluctuations and pipeline characteristics. They essentially reflect the response results after valve action and lack sensitivity to early fault characteristics caused by changes in the internal state of the actuator, such as wear of valve internals, spring fatigue, and mechanical loosening. This makes it difficult to meet the need for accurate diagnosis of the health status of key internal components of control valves. Collecting only valve body vibration signals mainly reflects the overall vibration characteristics of the valve body. Valve body vibration is easily affected by external interferences such as overall structural stiffness, pipeline vibration, and fluid pulsation, which dilutes or masks fault feature information. It fails to directly obtain valve stem vibration features containing richer fault information. Moreover, existing algorithms have a single feature extraction method and cannot fully explore the complementary information of signals in different domains, resulting in insufficient extraction of fault features of control valves. Summary of the Invention

[0005] Therefore, it is necessary to provide a method and device for diagnosing pneumatic control valve faults based on microwave vibration measurement to address the aforementioned technical problems.

[0006] The present invention adopts the following technical solution: This invention provides a method for diagnosing faults in pneumatic control valves based on microwave vibration measurement, comprising: Acquire the valve stem vibration signal of the pneumatic control valve; Vibration features at different scales of the control valve stem vibration signal are extracted in parallel using convolutional kernels of different scales and then concatenated according to the channel dimension to obtain multi-scale valve stem vibration fusion features. The multi-scale valve stem vibration fusion features are then abstracted layer by layer through two-layer convolutional neural networks and pooling operations to extract deep fault features. The channel attention of the deep fault features at different channel dimensions is calculated, and the temporal dependency relationship of the channel attention-weighted deep fault features is modeled using a bidirectional GRU network to obtain the temporal domain features of the control valve. The vibration signal of the control valve stem is subjected to STFT image transformation to obtain a two-dimensional control valve time-frequency image. The two-dimensional control valve time-frequency image is abstracted layer by layer through a four-layer two-dimensional convolutional neural network and pooling operation to extract deep spectral texture features. The channel attention of the deep spectral texture features in different channel dimensions is calculated, and the spectral texture features after channel attention weighting are spatially aggregated using global average pooling and fully connected layers to obtain the frequency domain features of the control valve. The time-domain features and frequency-domain features of the control valve are fused to obtain the time-frequency features of the control valve; Based on the preset control valve fault categories, the time-frequency characteristics of the control valve are classified into faults to obtain the control valve classification results.

[0007] Preferably, the control valve classification results are obtained through a multi-scale time-frequency domain dual-channel feature fusion model; The multi-scale time-frequency domain dual-channel feature fusion model includes: a time-domain branch composed of a multi-scale convolutional neural network, a channel attention mechanism, and a bidirectional gated recurrent unit cascaded together; a frequency-domain branch composed of four layers of two-dimensional convolutional neural networks with different kernel sizes cascaded together with a channel attention mechanism; a feature fusion module based on cross-attention; and a fault classification module; wherein the time-domain branch and the frequency-domain branch are connected in parallel.

[0008] Preferably, acquiring the valve stem vibration signal of the pneumatic control valve specifically includes: The vibration signal of the control valve stem is collected by a microwave displacement sensor using microwave vibration measurement technology, with a collection frequency of 1000Hz.

[0009] Preferably, calculating the channel attention weights of deep fault features in different channel dimensions specifically includes: By using global average pooling to compress the spatial information of each channel of the control valve fault characteristics, a control valve fault characteristic channel descriptor is generated. The nonlinear dependencies between the fault feature channel descriptors of each control valve are learned through a two-layer fully connected network, and channel attention weights are generated. The importance of the control valve fault features in each channel is obtained by multiplying the channel attention weights by the control valve fault features channel by channel.

[0010] Preferably, when performing STFT image transformation on the control valve stem vibration signal, a Hanning window is used as the window function, with a window length of 128 points and a step size of 64 points; the STFT image transformation is used to convert the one-dimensional control valve stem vibration signal into a two-dimensional control valve time-frequency image of size 65×11.

[0011] Preferably, the time-domain features and frequency-domain features of the control valve are fused to obtain the time-frequency features of the control valve, specifically including: Using time-domain features as query vectors and frequency-domain features as key-value vectors, we calculate the attention weights of the time-domain features. The frequency domain features are used as the query vector, and the time domain features are used as the key and value vectors. The attention weight of the frequency domain features is calculated. Based on the attention weights of the time-domain features and the frequency-domain features, the time-domain features and frequency-domain features of the control valve are weighted and fused to obtain the time-frequency features of the control valve.

[0012] Preferably, the preset control valve fault categories include: normal state, sensor fault, pressure reducing valve fault, actuator leakage fault, valve stem friction fault, and valve jamming fault.

[0013] The present invention also provides a pneumatic control valve fault diagnosis device based on microwave vibration measurement, characterized in that it includes: The data acquisition module is used to acquire the valve stem vibration signal of the pneumatic control valve; The temporal feature extraction module is used to extract vibration features of different scales of the control valve stem vibration signal in parallel using convolutional kernels of different scales and concatenate them according to the channel dimension to obtain multi-scale valve stem vibration fusion features. The multi-scale valve stem vibration fusion features are abstracted layer by layer through two layers of convolutional neural network and pooling operation to extract deep fault features. A channel attention mechanism is introduced to calculate the channel attention weights of fault features in different channel dimensions, and a bidirectional GRU network is used to model the temporal dependency of the channel attention-weighted control valve fault features to obtain the temporal features of the control valve. The frequency domain feature extraction module is used to perform STFT image transformation on the vibration signal of the control valve stem to obtain a two-dimensional control valve time-frequency image. The two-dimensional control valve time-frequency image is abstracted layer by layer through a four-layer two-dimensional convolutional neural network and pooling operation to extract deep spectral texture features. A channel attention mechanism is introduced to calculate the channel attention weights of the spectral features in different channel dimensions, and global average pooling and fully connected layers are used to spatially aggregate the channel attention-weighted spectral texture features to obtain the frequency domain features of the control valve. The time-frequency domain feature fusion module is used to fuse the time-domain features and frequency-domain features of the control valve to obtain the time-frequency features of the control valve; The fault classification module is used to classify the time-frequency characteristics of the control valve based on the preset control valve fault categories, and obtain the control valve classification result.

[0014] The present invention provides a computer-readable storage medium storing a computer program, which, when executed by a processor, implements the above-mentioned method for diagnosing pneumatic control valve faults based on microwave vibration measurement.

[0015] The present invention provides a computer device, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the program, it implements the above-mentioned method for diagnosing pneumatic control valve faults based on microwave vibration measurement.

[0016] The above-mentioned at least one technical solution adopted in this invention can achieve the following beneficial effects: In the pneumatic control valve fault diagnosis method based on microwave vibration measurement provided by this invention, multi-scale vibration features are extracted in parallel using convolutional kernels of different sizes and then stitched together in channel dimensions to construct multi-scale fusion features. This effectively solves the limitations of traditional methods, such as single feature extraction and lack of multi-scale perception, and can simultaneously capture local transient fluctuations and global trend changes, thereby comprehensively mining fault characterization information. Fault features are further refined through two layers of deep convolution, enhancing the hierarchy and discriminative power of the features. A channel attention mechanism is introduced to calculate the weights of each channel, and a bidirectional GRU network is used to extract the temporal dependencies of fault features, obtaining the time-domain features of the control valve. This achieves adaptive screening and enhancement of control valve fault features, suppressing noise and redundant information. The valve stem vibration signal is subjected to STFT transformation to generate a two-dimensional time-frequency image, and spectral texture features are extracted from it, extending the fault information of the control valve from the time domain to the frequency domain, making up for the shortcomings of single-domain analysis. The frequency domain branch also introduces a channel attention mechanism to screen key spectral features as the frequency domain features of the control valve, ensuring the effectiveness and representativeness of the frequency domain information. By fusing the time-domain and frequency-domain features of the control valve, a time-frequency fusion feature is formed, which realizes the deep complementarity of time-frequency dual-modal information, giving full play to the synergistic effect of time-domain dynamic characteristics and frequency-domain texture features in control valve fault diagnosis, thereby effectively improving the accuracy of fault diagnosis.

[0017] Furthermore, the method of this invention overcomes the problems of existing methods that rely on pressure and flow signals with delayed response and are easily interfered with, or that only collect valve body vibration signals, resulting in the inability to directly obtain valve stem vibration characteristics containing richer fault information, as well as the problems of traditional contact sensors being difficult to install, easily introducing additional loads, and interfering with the dynamic characteristics of the valve stem. By non-contactly collecting valve stem vibration signals based on microwave vibration measurement technology, valve stem vibration data reflecting the state of key components such as valve core, spring, and seals are directly obtained, ensuring the authenticity and quality of the signal from the source, and providing a foundation for solving the problems of "limited signal acquisition and indirect fault information". Attached Figure Description

[0018] The accompanying drawings, which are included to provide a further understanding of this application and form part of this application, illustrate exemplary embodiments and are used to explain this application, but do not constitute an undue limitation of this application. In the drawings:

[0019] Figure 1 A flowchart illustrating a fault diagnosis method for pneumatic control valves based on microwave vibration measurement provided by the present invention; Figure 2 A multi-scale time-frequency domain dual-channel feature fusion model architecture diagram for a pneumatic control valve fault diagnosis method based on microwave vibration measurement provided by the present invention; Figure 3 A multi-scale time-domain branch architecture diagram of a pneumatic control valve fault diagnosis method based on microwave vibration measurement provided by the present invention; Figure 4 The GRU network structure diagram of a pneumatic control valve fault diagnosis method based on microwave vibration measurement provided by the present invention; Figure 5 A schematic diagram of the channel attention mechanism in a pneumatic control valve fault diagnosis method based on microwave vibration measurement provided by the present invention; Figure 6 A multi-scale frequency domain branch architecture diagram of a pneumatic control valve fault diagnosis method based on microwave vibration measurement provided by the present invention; Figure 7 A schematic diagram illustrating the cross-attention mechanism of a pneumatic control valve fault diagnosis method based on microwave vibration measurement provided by this invention. Figure 8 The structural diagram of microwave vibration measurement for a pneumatic control valve fault diagnosis method based on microwave vibration measurement provided by the present invention; Figure 9 A schematic diagram of a control valve fault simulation test bench for a pneumatic control valve fault diagnosis method based on microwave vibration measurement provided by the present invention; Figure 10 The confusion matrix of six fault mode test sets for a fault diagnosis method for pneumatic control valves based on microwave vibration measurement provided by the present invention; Figure 11 The classification feature visualization results of a pneumatic control valve fault diagnosis method based on microwave vibration measurement provided by the present invention; Figure 12 The confusion matrix of 11 fault mode test sets of the DAMADICS platform for a fault diagnosis method of pneumatic control valve based on microwave vibration measurement provided by the present invention. Figure 13 A multi-level T-SNE feature visualization diagram of a pneumatic control valve fault diagnosis method based on microwave vibration measurement provided by the present invention; Figure 14 A schematic diagram of a pneumatic control valve fault diagnosis device based on microwave vibration measurement provided by the present invention; Figure 15 A diagram of a computer device for implementing a fault diagnosis method for pneumatic control valves based on microwave vibration measurement, provided by the present invention. Detailed Implementation

[0020] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of this application will be clearly and completely described below in conjunction with specific embodiments and corresponding drawings. Obviously, the described embodiments are only a part of the embodiments of this application, and not all of them. All other embodiments obtained by those skilled in the art based on the embodiments in the specification without creative effort are within the scope of protection of this application.

[0021] The technical solutions provided by the various embodiments of this application are described in detail below with reference to the accompanying drawings.

[0022] Figure 1 This is a schematic diagram of a fault diagnosis method for pneumatic control valves based on microwave vibration measurement according to the present invention, which specifically includes the following steps: S101: Acquire the valve stem vibration signal of the pneumatic control valve.

[0023] Optionally, the vibration signal of the control valve stem is acquired using a microwave displacement sensor based on microwave vibration measurement technology, with an acquisition frequency of 1000Hz.

[0024] Specifically, to verify the classification performance of the proposed multi-scale time-frequency domain dual-channel feature fusion method in control valve fault diagnosis, the model was first validated on a constructed control valve fault simulation test bench, and then further validated on the DAMADICS platform. Validation using datasets from two different sources fully demonstrates the effectiveness and generalization ability of the proposed method.

[0025] To simulate control valve failure data under real-world operating conditions, this embodiment constructed a dedicated control valve failure simulation test bench, see [link to relevant documentation]. Figure 9The test bench mainly consists of an ATS pneumatic control valve, an air compressor pump, an air supply pipeline system, a back pressure regulating valve, a data acquisition system, and a host computer monitoring system. Among them, the ATS pneumatic control valve adopts a pneumatic diaphragm actuator with a nominal diameter of DN100 and a stroke of 30mm, which can realize 0-100% opening adjustment.

[0026] A microwave displacement sensor based on microwave vibration measurement technology is used to collect valve stem vibration signals, such as... Figure 8 As shown, the sampling frequency is 1000Hz. Compared with contact accelerometers, microwave displacement sensors have the following significant advantages: First, the non-contact measurement method avoids mechanical wear between the sensor and the valve stem, and does not impose additional load on the normal operation of the valve; second, it has a long measurement distance and is suitable for harsh environments.

[0027] The collected dataset was divided into training, validation, and test sets at a ratio of 70%, 15%, and 15%, respectively, with no overlap between the three sets to avoid data leakage. All input data underwent z-score normalization to eliminate dimensional differences and improve model convergence stability. Data was continuously collected for 100 seconds for each operating state, with a data length of 100,000. A sliding window method was used to segment the original valve stem vibration signal, ensuring a 50% overlap between adjacent samples to maintain temporal continuity. Each operating state had 284 samples, totaling 1704 samples across the six operating states. During training, the batch size was set to 32, all convolutional neural network layers used the ReLU activation function, the Adam optimizer was selected, and an adaptive learning rate was used to effectively accelerate gradient descent convergence. The initial learning rate was set to 0.001, and the loss function was classification cross-entropy to adapt to the needs of multi-class classification tasks.

[0028] S102: The vibration features of the control valve stem vibration signal at different scales are extracted in parallel using convolutional kernels of different scales and spliced ​​according to the channel dimension to obtain multi-scale valve stem vibration fusion features; the multi-scale valve stem vibration fusion features are abstracted layer by layer through two-layer convolutional neural networks and pooling operations to extract deep fault features; the channel attention of deep fault features at different channel dimensions is calculated, and the temporal dependency relationship of the deep fault features after channel attention weighting is modeled using a bidirectional GRU network to obtain the temporal domain features of the control valve.

[0029] Optionally, the channel attention of deep fault features in different channel dimensions is calculated, specifically including: compressing the spatial information of each channel of the control valve fault features through global average pooling to generate control valve fault feature channel descriptors; learning the nonlinear dependency relationship between each control valve fault feature channel descriptor through a two-layer fully connected network to generate channel attention weights; and multiplying the channel attention weights with the control valve fault features channel by channel to obtain the importance of the control valve fault features in each channel.

[0030] Specifically, see Figure 3 The time-domain branch, as the core module of the time-frequency domain dual-channel parallel structure, is mainly responsible for extracting temporal features from the original valve stem vibration signal of the control valve. The valve stem vibration signal can directly reflect the working state of the valve, and different types of faults often cause the valve stem vibration signal to exhibit specific fluctuation patterns and abnormal characteristics. Based on this, the time-domain branch adopts a cascaded structure of multi-scale convolutional neural network, channel attention mechanism, and bidirectional gated recurrent unit to achieve in-depth mining of fault features.

[0031] Considering the significant differences in the characteristic scales of different types of control valve faults in valve stem vibration signals, the temporal branch first designed a parallel multi-scale convolutional structure. For the input one-dimensional valve stem vibration signal, the model simultaneously uses three types of one-dimensional convolutional kernels of sizes 3, 5, and 7 for parallel feature extraction. Smaller kernels primarily capture local transient features of the signal, while larger kernels can perceive trends over a longer time range. Each convolutional branch outputs 32 feature channels, which, after processing with the ReLU activation function, are fused into a 96-dimensional multi-scale feature representation through channel dimension concatenation. Subsequently, the fused features undergo further feature extraction through two layers of deep convolutional networks, with batch normalization and pooling operations performed after each layer. This design allows the model to fully exploit fault feature information at different levels. To enhance the model's selective attention to important features, the temporal branch introduces a channel attention mechanism after the deep convolutional network. This mechanism adaptively learns the importance weights of each feature channel, thereby strengthening feature channels highly relevant to fault diagnosis while suppressing interference from noise and redundant information.

[0032] After spatial feature extraction, the temporal branch employs a two-layer bidirectional GRU network to fully exploit temporal dependencies in the feature sequence by simultaneously processing sequence information in both forward and backward directions. Subsequently, a fully connected layer maps the temporal features into a fixed-dimensional feature vector. This feature vector not only preserves the temporal dynamic characteristics of the original valve stem vibration signal but also integrates multi-scale spatial feature information, providing a foundation for subsequent fusion with frequency domain features.

[0033] Among them, the Bidirectional Gated Recurrent Unit (BiGRU) is an improved recurrent neural network structure. By combining a gated recurrent unit (GRU) with a bidirectional processing mechanism, it can effectively capture long-term dependencies and bidirectional contextual information in sequence data. The GRU is a simplified variant of LSTM, and its network is as follows: Figure 4As shown, information flow is controlled by introducing reset and update gate mechanisms. Compared to LSTM, GRU has fewer parameters and a simpler structure, yet still maintains good performance. The core computation process of GRU can be represented as:

[0034] ; in, To reset the door, To update the door, In the candidate hidden state, Update to hide state. This represents the sigmoid activation function, and * represents element-wise multiplication. These are the weight matrices for the reset gate, update gate, and candidate states, respectively. For the corresponding bias term, For a moment Input, For a moment The hidden state.

[0035] BiGRU consists of two independent GRU layers: a forward GRU that processes information from the beginning to the end of the sequence, and a backward GRU that processes information from the end to the beginning of the sequence. For time... The output of BiGRU is composed of the hidden states in two directions:

[0036] ; in, and These represent the hidden states for the forward and backward directions, respectively, and [;] represents the vector concatenation operation.

[0037] Channel attention is an adaptive feature recalibration technique. In deep convolutional neural networks, different convolutional kernels extract different types of features, and these features contribute differently to the final task. Channel attention enables the network to learn autonomously and emphasize important feature channels. The principle diagram is shown below. Figure 5 As shown.

[0038] A Squeeze-and-Excitation (SE) structure is used to implement the channel attention mechanism. For the input feature map... ,in These represent the height, width, and number of channels of the feature map, respectively. The computation process of the channel attention mechanism includes three stages: squeezing, excitation, and multiplication.

[0039] During the compression stage, global average pooling is used to compress the spatial information of each channel into a scalar, resulting in a channel descriptor: ; in Indicates the first The passage is in the location The eigenvalue at that location.

[0040] During the activation phase, the non-linear dependencies between channels are learned through a two-layer fully connected network: ; in, For channel weight vectors, and For learnable parameters, To reduce the proportion, Represents the ReLU activation function. The activation function is sigmoid. The final attention weights are obtained by fusing the results of the two pooling methods:

[0041] ; The output of channel attention is the channel-wise product of the original feature map and the attention weights: .

[0042] S103: Perform STFT image transformation on the vibration signal of the control valve stem to obtain a two-dimensional control valve time-frequency image; abstract the two-dimensional control valve time-frequency image layer by layer through a four-layer two-dimensional convolutional neural network and pooling operation to extract deep spectral texture features; calculate the channel attention of the deep spectral texture features in different channel dimensions, and use global average pooling and fully connected layers to spatially aggregate the spectral texture features after channel attention weighting to obtain the frequency domain features of the control valve; Optionally, when performing STFT image transformation on the control valve stem vibration signal, a Hanning window is used as the window function, with a window length of 128 points and a step size of 64 points; the STFT image transformation is used to convert the one-dimensional control valve stem vibration signal into a two-dimensional control valve time-frequency image of size 65×11.

[0043] Specifically, the frequency domain branch, as another core component of the time-frequency dual-channel parallel structure, extracts the fault characteristics of the control valve from a frequency domain perspective through time-frequency transformation, compensating for the limitations of the time domain branch. Unlike the time domain branch, which directly processes the original signal, the frequency domain branch converts the one-dimensional valve stem vibration signal into a two-dimensional time-frequency image through short-time Fourier transform (STFT), thereby capturing the frequency distribution of the signal. The structure of the multi-scale frequency domain branch is as follows: Figure 6 As shown.

[0044] STFT obtains the joint time-frequency distribution of the signal by sliding a fixed-length window function across the time-domain signal and performing a Fourier transform on the signal within each window. Specifically, for an input signal of length 700, the frequency domain branch uses a 128-point Hanning window with a stride of 64 to generate a 65×11 time-frequency image. A four-layer two-dimensional convolutional neural network with different kernel sizes is used to extract features from the time-frequency image. Batch normalization and pooling operations are configured after each convolutional layer, and finally, a fully connected layer outputs the same dimension as the time-domain branch.

[0045] Converting one-dimensional time-series signals into two-dimensional time-frequency images offers significant advantages. Time-frequency images can visually display the distribution characteristics of signal energy on the time-frequency plane. Different fault types often exhibit unique texture patterns, such as the periodic energy concentration in a specific frequency band due to valve stem jamming, or the abnormal enhancement of high-frequency components caused by valve blockage. This conversion transforms the originally implicit time-domain dynamic features into visualized spectral texture features, providing richer feature representations for fault diagnosis.

[0046] The Short-Time Fourier Transform (STFT) is an important time-frequency analysis method for analyzing non-stationary signals. Unlike the traditional Fourier Transform, which can only obtain the global spectrum of a signal, the STFT decomposes the signal into a series of locally stationary segments by sliding a window function on the time domain and performs Fourier transforms on each segment, thereby obtaining the spectral information of the signal as it changes over time.

[0047] For discrete signals Its STFT is defined as: ; in, For time frame indexing, For frequency bin index, For frame shift, For FFT points, Let be the window function. This study uses the Hanning window as the window function, and its expression is:

[0048] ; in, The length of the window.

[0049] S104: The time-domain features and frequency-domain features of the control valve are fused to obtain the time-frequency features of the control valve.

[0050] Optionally, the time-domain features and frequency-domain features of the control valve are fused to obtain the time-frequency features of the control valve. Specifically, this includes: using the time-domain features as a query vector and the frequency-domain features as a key-value vector, calculating the attention weight of the time-domain features; using the frequency-domain features as a query vector and the time-domain features as a key-value vector, calculating the attention weight of the frequency-domain features; and weighting and fusing the time-domain features and frequency-domain features of the control valve according to the attention weights of the time-domain features and the frequency-domain features to obtain the time-frequency features of the control valve.

[0051] Specifically, cross-attention is an important feature interaction technique in deep learning. Its core idea lies in establishing correlation mappings between different feature representations to achieve effective information transfer and fusion. In the field of multimodal learning, cross-attention can capture the inherent correlations between data from different modalities, thereby improving the model's representation learning ability. The principle is as follows: Figure 7 As shown.

[0052] Let the feature matrix extracted from the original signal using method one be... The feature matrix extracted by method two is ,in and These represent the dimensions of the time-domain and frequency-domain features, respectively. The calculation process of the cross-attention mechanism is as follows:

[0053] Method 1 cross-attention to Method 2: through linear transformation Mapped to a query vector. Mapped to key vectors and value vectors: ; in, , and The weight matrix is ​​a learnable matrix. For the embedded dimension.

[0054] Calculate attention weights: ; Output: The representation of F1 enhanced by F2 is: ; Method 2's cross-attention to Method 1: Similarly, F2 enhanced by F1 is represented as follows: .

[0055] S105: Based on the preset control valve fault categories, classify the time-frequency characteristics of the control valve for faults to obtain the control valve classification result.

[0056] Optionally, the preset control valve fault categories include: normal state, sensor fault, pressure reducing valve fault, actuator leakage fault, valve stem friction fault, and valve jamming fault.

[0057] The test bench was designed to simulate six typical control valve operating states: (1) Normal operating state, where the control valve operates normally according to the set signal; (2) Sensor failure, which is simulated by adjusting the preload of the valve positioner feedback linkage mechanism to make the valve position feedback signal received by the positioner deviate from the actual valve opening; (3) Pressure reducing valve failure, which is simulated by adjusting the air supply pressure to make it deviate from the normal range; (4) Actuator leakage failure, which is achieved by setting a controllable leakage point in the pneumatic pipeline; (5) Increased valve stem friction, which is simulated by using an adjustable mechanical device to increase the valve stem movement resistance; (6) Valve jamming failure, which is simulated by using a mechanical limit device to simulate the valve jamming phenomenon at a specific position. Each failure state can be controlled and reproduced by corresponding methods to ensure the reliability of experimental data.

[0058] Optionally, the control valve classification results are obtained through a multi-scale time-frequency domain dual-channel feature fusion model; The multi-scale time-frequency domain dual-channel feature fusion model includes: a time-domain branch composed of a multi-scale convolutional neural network, a channel attention mechanism, and a bidirectional gated recurrent unit cascaded together; a frequency-domain branch composed of four layers of two-dimensional convolutional neural networks with different kernel sizes cascaded together with a channel attention mechanism; a feature fusion module based on cross-attention; and a fault classification module; wherein the time-domain branch and the frequency-domain branch are connected in parallel.

[0059] Specifically, this embodiment provides a multi-scale time-frequency domain dual-channel feature fusion network model for control valve fault diagnosis, the overall framework of which is as follows: Figure 2 As shown, the network consists of four modules: a time-domain branch, a frequency-domain branch, a cross-attention-based feature fusion module, and a fault classification module. First, the raw signal data is preprocessed. The time-domain branch captures the temporal dynamic features of the valve stem vibration signal, and the frequency-domain branch extracts spectral texture features. Then, the cross-attention-based feature fusion module achieves the fusion of time and frequency features, fully exploring the correlation between different modes. Finally, a fully connected layer and a Softmax function complete the fault classification.

[0060] Confusion matrices are important tools for evaluating the performance of classification models. By comparing the predicted results with the true labels, they can intuitively demonstrate the model's ability to identify different categories and the patterns of misclassification. To verify the effectiveness of the proposed time-frequency domain dual-channel feature fusion method in control valve fault diagnosis, this embodiment constructs a 6×6 confusion matrix, as shown below. Figure 10 As shown.

[0061] The matrix includes classification results for the normal state (D0) and five typical fault modes: sensor fault (D1), pressure reducing valve fault (D2), air leakage fault (D3), friction fault (D4), and jamming fault (D5). The model's average classification accuracy is 96.25%. The confusion matrix shows that labels D4 and D5 were accurately identified, while labels D0, D1, D2, and D3 exhibited varying degrees of confusion. This result demonstrates that the proposed model can effectively extract fault features of control valves and exhibits good classification ability.

[0062] Analysis of the fault mechanisms reveals a clear overlap between pressure-reducing valve malfunctions and sensor malfunctions, primarily due to the similarity in their impact paths on the control system. Firstly, both types of faults affect the positioning accuracy of the control valve, leading to similar deviations in the valve stem vibration signal. Secondly, the unstable gas supply pressure caused by a pressure-reducing valve malfunction indirectly affects the vibration feedback signal, producing similar behavior to a sensor malfunction. Furthermore, under certain operating conditions, both faults may produce similar low-frequency oscillation characteristics, further complicating differentiation.

[0063] To comprehensively evaluate the model's performance in control valve fault diagnosis, this study uses precision, recall, and F1-score as the main evaluation metrics to measure the model's classification effectiveness. Their expressions are as follows: ; ; ; Where TP represents a true positive, FP represents a false positive, and FN represents a false negative.

[0064] Table 1 shows the model's classification performance evaluation metrics for six control valve operating states. Overall, the model's average precision, recall, and F1 score all exceed 96%, with each metric for each category exceeding 90%, validating the model's excellent diagnostic performance in control valve fault diagnosis tasks.

[0065] Table 1 Evaluation metrics for classification performance on the test set To further verify the feature learning ability and separability of the feature space of the proposed model, this study employs the t-Distributed Stochastic Neighbor Embedding (t-SNE) algorithm to perform dimensionality reduction and visualization analysis on the high-dimensional features extracted by the model. t-SNE is a nonlinear dimensionality reduction technique that can map data points in a low-dimensional space to a low-dimensional space for visualization while preserving the local structure of data points in the high-dimensional space. Figure 11The distribution of the test set samples on a two-dimensional plane after t-SNE dimensionality reduction is shown, where the scatter points of different colors represent six different operating states of the control valve.

[0066] from Figure 11 The feature distribution reveals a clear clustering structure across the six operating states, indicating that the model has successfully learned discriminative feature representations for each category. Friction faults (D4) and jamming faults (D5) form independent and compact clusters, consistent with their accurate classification in the confusion matrix. Normal state (D0) and leakage faults (D3) also exhibit good clustering characteristics. Sensor faults (D1) and pressure reducing valve faults (D2) show slight overlap, revealing misclassification between them in the feature space. t-SNE visualization results demonstrate that through multi-scale feature extraction in the time domain and time-frequency texture analysis in the frequency domain, combined with feature fusion using a cross-attention mechanism, the model can map the original high-dimensional valve stem vibration signal to a feature space with good discriminative power, achieving high-precision control valve fault diagnosis.

[0067] To highlight the superiority of the proposed time-frequency domain dual-channel feature fusion method, this study selected several models for comparative experiments. The comparative methods included: Convolutional Neural Networks (CNNs) as a classic spatial feature extraction model; Long Short-Term Memory Networks (LSTMs) as a typical temporal modeling method; a CNN-LSTM combined model to verify the effectiveness of simple feature fusion; and the performance of the time-domain and frequency-domain branches in the proposed method when used individually. All comparative experiments were conducted under the same data preprocessing and training parameters. Table 2 shows the performance comparison results of each method on the test set.

[0068] Table 2 Performance Comparison of Different Methods Experimental results show that the traditional single-model CNN and LSTM achieve accuracies of 75.00% and 81.25%, respectively, indicating relatively limited performance. The CNN-LSTM combined model improves the accuracy to 86.67%, but still remains below 90%. The time-domain and frequency-domain branches, used individually, achieve accuracies of 89.58% and 82.92%, respectively, validating the effectiveness of feature fusion. The proposed model achieves an accuracy of 96.25%, representing improvements of 6.67% and 13.33% compared to using the time-domain and frequency-domain branches individually, respectively, and improvements of 21.25%, 15%, and 9.58% compared to CNN, LSTM, and CNN-LSTM, respectively. The comparison of experimental results in the table above shows that the proposed model exhibits the best evaluation metrics, achieving accurate diagnosis of control valve fault types by fully mining and fusing time- and frequency-domain information.

[0069] DAMADICS (Development and Application of Methods for Actuator Diagnosis in Industrial Control Systems) is used for control valve performance evaluation and fault detection research. This platform simulates the dynamic behavior of industrial control valves under various operating conditions through precise mathematical modeling and physical simulation. It can generate valve stem vibration signal data containing different fault modes, providing reliable support for the development and verification of fault diagnosis algorithms.

[0070] Using control valve stem vibration simulation data generated based on the DAMADICS platform, 11 different fault modes were selected from Table 3. These fault modes include common failure modes in actual industrial scenarios such as control valve failure, positioner failure, and motor failure, fully reflecting the diversity and complexity of control valve faults. Each fault causes the valve stem vibration signal to exhibit unique time-domain fluctuations and frequency-domain distribution characteristics, providing a reliable experimental verification basis for the proposed method.

[0071] Table 3 Dataset Fault Description To verify the effectiveness and generalization ability of the proposed model in fault diagnosis, this study conducted tests and validations on the DAMADICS platform dataset. Figure 12 The confusion matrix of the proposed model on the test set is shown, with an average classification accuracy of 99.24%. The confusion matrix reveals that 8 out of 11 fault modes were accurately identified, with only a few misclassifications observed for labels C0, C1, and C8. The clear dark bands along the diagonal demonstrate that the model performs excellently not only in laboratory environments but also possesses reliable fault diagnosis capabilities in industrial scenarios.

[0072] Table 4 details the model's classification performance metrics for 11 fault modes. The precision, recall, and F1-score for all fault modes all exceed 94%, with average values ​​for each metric exceeding 99%. Compared to the experimental results from a self-built testbed, the performance metrics on the DAMADICS dataset are generally higher. This is mainly because the self-built testbed inevitably encounters interference factors such as environmental noise, mechanical vibration, and sensor measurement errors in the actual physical environment, while the simulation platform provides relatively clean signal conditions. The experimental results demonstrate the model's good classification performance and stability in fault diagnosis tasks.

[0073] Table 4 Evaluation Indicators for Classification Performance on the Test Set To further verify the model's feature learning performance on the DAMADICS dataset, see [link / reference]. Figure 13 This paper presents the distribution of 11 fault modes in a two-dimensional feature space after t-SNE dimensionality reduction. From the feature maps of the original signal data, it can be observed that the fault samples are relatively dispersed in the feature space, with blurred boundaries between categories. After processing by the time-domain branch, multi-scale convolution and bidirectional GRU networks effectively extract the temporal dynamic features of the valve stem vibration signal, causing some fault categories to begin to show a clustering trend, although some degree of inter-class aliasing still exists. The frequency-domain branch captures the frequency-domain information of the signal through STFT transform and convolution processing. Some fault modes exhibit feature distributions complementary to the time-domain features. Compared to the clustering patterns shown by the time-domain features, some fault categories that are difficult to distinguish in the time domain show a more obvious separation trend in the frequency-domain space, verifying that the model can extract complementary fault feature information from both the time and frequency domains.

[0074] The fused feature maps demonstrate that the cross-attention machine effectively utilizes the complementary features of the time-frequency and frequency branches, significantly improving inter-class separation and reducing overlapping regions, showcasing the advantages of cross-attention feature fusion. The final feature maps further optimize the clustering results, with more compact class distributions and clearer boundaries, and only a few classes overlapping, proving the model's effectiveness in improving fault diagnosis performance.

[0075] This embodiment addresses the problems of limited feature extraction and insufficient information mining in control valve fault diagnosis. It proposes a fault diagnosis method based on the fusion of microwave vibration measurement and time-frequency domain features. The main work and conclusions are as follows: A fault diagnosis model based on multi-scale time-frequency domain dual-channel feature fusion was constructed. A parallel network architecture with two branches in the time and frequency domains was designed. The time-domain branch uses multi-scale one-dimensional convolutional kernels to extract local features from different receptive fields, combined with a BiGRU network to capture temporal dependencies. The frequency-domain branch converts vibration signals into time-frequency images through STFT and extracts spectral texture features using a multi-scale two-dimensional convolutional network. A channel attention mechanism was introduced to adaptively adjust feature weights, and a cross-attention mechanism was used to achieve deep fusion of time-frequency domain features, fully exploiting multi-modal complementary information. A pneumatic control valve fault simulation test bench was built, simulating five typical faults, including valve stem jamming and actuator leakage, by adjusting key parameters such as air source pressure and valve stem friction. Microwave vibration measurement technology was innovatively applied to the field of control valve fault diagnosis. A microwave displacement sensor was used to achieve non-contact acquisition of valve stem vibration signals, effectively avoiding interference from contact measurements on valve stem motion characteristics, ensuring the authenticity of fault signals and measurement accuracy, and providing reliable data support for model training and validation.

[0076] The effectiveness of the proposed method was verified on a control valve fault simulation test bench. Experimental results show that the method achieves an average diagnostic accuracy of 96.25%, with average precision, recall, and F1 score all exceeding 96%, representing improvements of 21.25%, 15%, and 9.58% compared to CNN, LSTM, and CNN-LSTM methods, respectively. Ablation experiments confirm the crucial role of the time-frequency feature fusion mechanism in improving model performance. t-SNE visualization results show that the fused features have clearer class boundaries and better clustering effects, demonstrating that the time-frequency domain dual-channel feature fusion strategy can effectively improve the accuracy of control valve fault diagnosis.

[0077] The model's generalization ability was further validated on the DAMADICS platform. For 11 fault modes, including control valves, actuators, and valve positioners, the proposed method achieved an average classification accuracy of 99.24%, with accurate identification of 8 fault modes. The average precision, recall, and F1-score all exceeded 99%. Experimental results demonstrate that the model exhibits excellent diagnostic performance, good generalization ability, and engineering application value.

[0078] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this invention.

[0079] The above describes a method for diagnosing pneumatic control valve faults based on microwave vibration measurement, provided by one or more embodiments of the present invention. Based on the same idea, the present invention also provides a corresponding device for diagnosing pneumatic control valve faults based on microwave vibration measurement, such as... Figure 14 As shown.

[0080] Figure 14 A schematic diagram of a pneumatic control valve fault diagnosis device based on microwave vibration measurement provided by the present invention is shown. The device includes: The data acquisition module is used to acquire the valve stem vibration signal of the pneumatic control valve; The temporal feature extraction module 1401 is used to extract vibration features of different scales of the control valve stem vibration signal in parallel using convolutional kernels of different scales and to concatenate them according to the channel dimension to obtain multi-scale valve stem vibration fusion features; the multi-scale valve stem vibration fusion features are abstracted layer by layer through two layers of convolutional neural network and pooling operation to extract deep fault features; the channel attention of deep fault features in different channel dimensions is calculated, and the temporal dependency relationship of the deep fault features after channel attention weighting is modeled using a bidirectional GRU network to obtain the temporal features of the control valve; The frequency domain feature extraction module 1402 is used to perform STFT image transformation on the vibration signal of the control valve stem to obtain a two-dimensional control valve time-frequency image; the two-dimensional control valve time-frequency image is abstracted layer by layer through a four-layer two-dimensional convolutional neural network and pooling operation to extract deep spectral texture features; the channel attention of the deep spectral texture features in different channel dimensions is calculated, and the spectral texture features after channel attention weighting are spatially aggregated using global average pooling and fully connected layers to obtain the frequency domain features of the control valve; The time-frequency domain feature fusion module 1403 is used to fuse the time-domain features and frequency-domain features of the control valve to obtain the time-frequency features of the control valve. The fault classification module 1404 is used to classify the time-frequency characteristics of the control valve based on the preset control valve fault categories to obtain the control valve classification result.

[0081] Specific limitations regarding the microwave vibration-based pneumatic control valve fault diagnosis device can be found in the above-described limitations of the microwave vibration-based pneumatic control valve fault diagnosis method, and will not be repeated here. Each module in the aforementioned microwave vibration-based pneumatic control valve fault diagnosis device can be implemented entirely or partially through software, hardware, or a combination thereof. These modules can be embedded in or independent of the processor in a computer device, or stored in the computer device's memory as software, so that the processor can call and execute the corresponding operations of each module.

[0082] The present invention also provides a computer-readable storage medium storing a computer program that can be used to execute the above-described... Figure 1 A fault diagnosis method for pneumatic control valves based on microwave vibration measurement is provided.

[0083] The present invention also provides Figure 15 The schematic diagram of the computer device shown is as follows: Figure 15 As shown, at the hardware level, this computer device includes a processor, internal bus, network interface, memory, and non-volatile memory, and may also include other hardware required for business operations. The processor reads the corresponding computer program from the non-volatile memory into memory and then executes it to achieve the above. Figure 1 A fault diagnosis method for pneumatic control valves based on microwave vibration measurement is provided.

[0084] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium, and when executed, it can include the processes of the embodiments of the methods described above. Any references to memory, storage, databases, or other media used in the embodiments provided by this invention can include at least one of non-volatile and volatile memory. Non-volatile memory can include read-only memory (ROM), magnetic tape, floppy disk, flash memory, or optical storage, etc. Volatile memory can include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM can be in various forms, such as static random access memory (SRAM) or dynamic random access memory (DRAM), etc.

Claims

1. A method for diagnosing a fault of a pneumatic control valve based on microwave vibration measurement, characterized by, The method comprises the following steps: acquiring a valve rod vibration signal of a pneumatic control valve; extracting vibration features of different scales of the valve rod vibration signal of the control valve in parallel using different scale convolution kernels and splicing according to channel dimensions to obtain multi-scale valve rod vibration fusion features; and abstracting the multi-scale valve rod vibration fusion features layer by layer through a two-layer convolutional neural network and a pooling operation to extract deep fault features; calculating channel attention of the deep fault features in different channel dimensions, and modeling the deep fault features weighted by the channel attention using a bidirectional GRU network to obtain a time domain feature of the control valve; performing STFT image transformation on the valve rod vibration signal of the control valve to obtain a two-dimensional control valve time-frequency image; and abstracting the two-dimensional control valve time-frequency image layer by layer through a four-layer two-dimensional convolutional neural network and a pooling operation to extract deep spectral texture features; calculating channel attention of the deep spectral texture features in different channel dimensions, and performing spatial aggregation on the spectral texture features weighted by the channel attention using a global average pooling and a fully connected layer to obtain a frequency domain feature of the control valve; fusing the time domain feature of the control valve and the frequency domain feature of the control valve to obtain a time-frequency feature of the control valve; based on a preset control valve fault category, classifying the time-frequency feature of the control valve to obtain a control valve classification result.

2. The method according to claim 1, wherein the method is characterized by, The control valve classification result is obtained by a multi-scale time-frequency domain dual-channel feature fusion model; The multi-scale time-frequency domain dual-channel feature fusion model comprises a time domain branch formed by cascading a multi-scale convolutional neural network, a channel attention mechanism and a bidirectional gated recurrent unit, a frequency domain branch formed by cascading a four-layer two-dimensional convolutional neural network with different convolution kernel sizes and a channel attention mechanism, a feature fusion module based on cross attention and a fault classification module; wherein the time domain branch and the frequency domain branch are connected in parallel.

3. The method according to claim 1, wherein the method is characterized by, The method for acquiring the valve rod vibration signal of the pneumatic control valve specifically comprises: collecting the valve rod vibration signal of the control valve by a microwave displacement sensor of a microwave vibration measurement technology, and the collection frequency is 1000 Hz.

4. The method of claim 1, wherein the method comprises: The method for calculating the channel attention weight of the deep fault features in different channel dimensions specifically comprises: generating a control valve fault feature channel descriptor by compressing the spatial information of each channel of the control valve fault features through a global average pooling; generating a channel attention weight by learning the nonlinear dependency relationship between each control valve fault feature channel descriptor through a two-layer fully connected network; multiplying the channel attention weight with the control valve fault features channel by channel to obtain the importance of the control valve fault features in each channel.

5. The method of claim 1, wherein the method comprises: In the step of performing STFT image transformation on the valve rod vibration signal of the control valve, a Hanning window is used as a window function, the window length is 128 points, and the step length is 64 points; the STFT image transformation is used to convert the one-dimensional valve rod vibration signal of the control valve into a two-dimensional control valve time-frequency image with a size of 65*11.

6. The method of claim 1, wherein the method comprises: The method for fusing the time domain feature of the control valve and the frequency domain feature of the control valve to obtain the time-frequency feature of the control valve specifically comprises: calculating the attention weight of the time domain feature by taking the time domain feature as a query vector, and taking the frequency domain feature as a key and a value vector; The frequency domain feature is taken as a query vector, the time domain feature is taken as a key and a value vector, and the attention weight of the frequency domain feature is calculated; According to the attention weight of the time domain feature and the attention weight of the frequency domain feature, the control valve time domain feature and the control valve frequency domain feature are weighted and fused to obtain a control valve time-frequency feature.

7. The method of claim 1, wherein the method comprises: The preset control valve fault category includes a normal state, a sensor fault, a pressure reducing valve fault, an actuator air leakage fault, a valve stem friction fault, and a valve sticking fault.

8. A microwave-based vibration measurement device for diagnosing faults in a pneumatic control valve, characterized by The method comprises the following steps: The data acquisition module is configured to acquire a valve stem vibration signal of a pneumatic control valve. The time domain feature extraction module is configured to extract vibration features of different scales of the control valve valve stem vibration signal in parallel by using different scale convolution kernels and splice the vibration features according to a channel dimension to obtain a multi-scale valve stem vibration fusion feature; and to perform layer-by-layer abstraction on the multi-scale valve stem vibration fusion feature by using a two-layer convolutional neural network and a pooling operation to extract deep fault features. The channel attention of the deep fault features in different channel dimensions is calculated, and the deep fault features weighted by the channel attention are subjected to time sequence dependency modeling by using a bidirectional GRU network to obtain a control valve time domain feature. The frequency domain feature extraction module is configured to perform an STFT image transformation on the control valve valve stem vibration signal to obtain a two-dimensional control valve time-frequency image; and to perform layer-by-layer abstraction on the two-dimensional control valve time-frequency image by using a four-layer two-dimensional convolutional neural network and a pooling operation to extract deep spectral texture features. The channel attention of the deep spectral texture features in different channel dimensions is calculated, and the spectral texture features weighted by the channel attention are subjected to spatial aggregation by using a global average pooling and a fully connected layer to obtain a control valve frequency domain feature. The time-frequency domain feature fusion module is configured to fuse the control valve time domain feature and the control valve frequency domain feature to obtain a control valve time-frequency feature. The fault classification module is configured to perform fault classification on the control valve time-frequency feature based on a preset control valve fault category to obtain a control valve classification result.

Images