A method for detecting abnormal connection of valve body

Abstract

A method for detecting abnormal valve connection includes the following steps: step S1. collecting and processing intelligent valve operation data; step S2. establishing a valve body abnormality detection network model; step S3. training the valve body abnormality detection network model; step S4. evaluating the valve body abnormality detection network model; step S5. detecting abnormal valve connection. The method of the invention can detect abnormal valve connection in a timely and comprehensive manner, improve valve operation efficiency, has a wide range of applications, is not limited to the shape and type of the valve, and comprehensively detects abnormal types.

Description

Technical Field

[0001] The present invention relates to the technical field of intelligent valve data analysis, and in particular to a method for detecting abnormal valve body connection. Background Art

[0002] The valves used in pipelines are usually related to the operation of their downstream water supply equipment. Therefore, they are extremely important and require regular inspection and maintenance. The mechanical part of the electric valve consists of a valve body, a valve cover and a valve disc. The valve body is the channel through which water flows; the valve cover is a sealing component connected to the valve body; and the valve disc is the control part for opening and closing the valve. Slippage, blockage, rust, and depression between the valve disc and the valve body will affect the sealing of the valve and even cause the valve to get stuck or damaged. In valve abnormality detection, a scanning device is used to scan the distance information from each point on the valve sealing surface to the reference plane, thereby displaying the surface information of the valve disc, thereby judging whether the surface depression between the valve disc and the valve body is abnormal. However, this method has the following disadvantages:

[0003] 1. A scanner is needed to scan the inside of the valve body, which requires the laser signal emitted by the scanner to penetrate into the inside of the valve. There are certain requirements for the valve material and shape, and there are certain limitations in applicability;

[0004] 2. The surface depression of the valve body and the valve disc is judged by the distance information of the reference point. It is impossible to correctly detect abnormalities such as slip and stick, and it is impossible to comprehensively judge the type of abnormality. Summary of the invention

[0005] The technical problem to be solved by the present invention is to overcome the shortcomings of the above-mentioned background technology and provide a method for detecting valve body connection abnormalities, which can timely and comprehensively detect valve body connection abnormalities, improve valve operation efficiency, have a wide range of applications, are not limited to the shape and type of the valve, and can comprehensively detect the types of abnormalities.

[0006] The technical solution adopted by the present invention to solve the technical problem is a method for detecting abnormal valve body connection, comprising the following steps:

[0007] Step S1. Collection and processing of intelligent valve operation data;

[0008] Step S2. Establishing a valve body abnormality detection network model;

[0009] Step S3. Training a valve body abnormality detection network model;

[0010] Step S4. Evaluate the valve body abnormality detection network model;

[0011] Step S5: Detection of abnormal valve body connection.

[0012] Further, in step S1, the method for collecting and processing the intelligent valve operation data is as follows:

[0013] Step S1-1. Collect valve operation data set Q in an experimental environment: Under different valve operation scenarios, collect sound signal data during valve operation, and label the collected audio by category;

[0014] Step S1-2. Collect the public valve operation data set P on the Internet and label the categories by yourself;

[0015] Step S1-3: uniformly crop and normalize the valve body operation data, and perform short-time Fourier transform to highlight its frequency domain characteristics.

[0016] Further, in step S1-1, sound signal data of the valve operation is collected in different valve operation scenarios, including: when the valve is operating normally, audio is collected using the sound module; when the valve switch is controlled, audio is collected using the sound module; when the valve body is abnormal, audio is collected using the sound module, and the valve body abnormality includes three situations: valve body sliding, valve body sticking, and valve body stuck. The sound module is used to collect audio in the three situations of valve body sliding, valve body sticking, and valve body stuck, respectively.

[0017] Further, in steps S1-1 and S1-2, the category labeling means: when the valve operates normally, the collected audio is labeled as the normal operation audio of the valve; when the valve body slides, the collected audio is labeled as the valve body sliding audio; when the valve body is sticking, the collected audio is labeled as the valve body sticking audio; when the valve body is stuck, the collected audio is labeled as the valve body stuck audio.

[0018] Furthermore, the method of establishing a valve body abnormality detection network model is as follows:

[0019] Step S2-1. Build a feature extraction module, which includes three one-dimensional dilated convolutional layers and one fully connected layer. The first one-dimensional dilated convolutional layer is connected to the second one-dimensional dilated convolutional layer, the second one-dimensional dilated convolutional layer is connected to the third one-dimensional dilated convolutional layer, and the third one-dimensional dilated convolutional layer is connected to the fully connected layer; the kernel size of the first one-dimensional dilated convolutional layer is 5, the dilation coefficient is 1, and the pad parameters after the dilated convolution are filled to the same size as before the dilated convolution; the kernel size of the second one-dimensional dilated convolutional layer is 5, the dilation coefficient is 2, and the dilation coefficient is The pad parameters after the dilated convolution are filled to the same size as before the dilated convolution; the kernel size of the third one-dimensional dilated convolution layer is 5, the dilation coefficient is 4, and the pad parameters after the dilated convolution are filled to the same size as before the dilated convolution; the first one-dimensional dilated convolution layer and the second one-dimensional dilated convolution layer use PReLU as the activation function after the dilated convolution; three layers of one-dimensional dilated convolution layers are used to extract features from the waveform data, and finally a fully connected layer is used to output the features as a fixed 512-dimensional feature. The extracted features can finally characterize the basic characteristics of the sound event;

[0020] Step S2-2. Build two layers of one-dimensional residual dilated convolutional layers. The kernel size of the first layer of one-dimensional residual dilated convolutional layer is 5, the dilation coefficient is 4, and no padding is set; the kernel size of the second layer of one-dimensional residual dilated convolutional layer is 5, the dilation coefficient is 2, and no padding is set; the residual dilated convolution processes the audio of the detected valve in different operating states nonlinearly;

[0021] Step S2-3. Build a fully connected layer, use the fully connected layer to linearly transform the features, and output N-dimensional features; N represents the number of anomalies that need to be detected;

[0022] Step S2-4. Build an activation layer, use the Softmax activation function in the activation layer, and use the Softmax activation function to output the likelihood of the four valve body states; at the same time, adjust the hyperparameters and set the likelihood limit thresholds of the four valve body states. If the likelihood of a certain valve body state is greater than the corresponding likelihood limit threshold, it is determined that the valve body is in this valve body state at this moment, otherwise it is determined that the valve body is not in this valve body state at this moment.

[0023] Further, in step S2-3, N=4, including four valve body states: normal operation of the valve body, valve body sliding, valve body adhesion, and valve body stuck.

[0024] Further, in step S2-4, the principle of using the Softmax activation function to output the likelihood of the four valve body states is:

[0025] The continuous audio is processed using an average weighted method. Specifically, the audio waveforms of the valve in three sections, from 2s to 1s before the point, from 1.5s to 0.5s before the point, and from 1s to the current time point, are sent to the detection module to obtain the likelihood of the valve state in the three time ranges. The likelihood of each valve state at the current moment is obtained by weighted average with weights of 0.1, 0.2, and 0.7, and then compared with the likelihood limit threshold of each valve state. If the obtained likelihood of the valve state is greater than the corresponding likelihood limit threshold, it is determined that the valve body is in that valve state at that moment; otherwise, it is determined that the valve body is not in that valve state at that moment.

[0026] Further, in step S3, the method for training the valve body anomaly detection network model is as follows: divide the training data set and the evaluation data set, select 80% of the public valve body operation data set P and the self-collected valve operation data set Q as the training data set, and the remaining 20% ​​as the evaluation data set; select cross entropy as the loss function, and random gradient descent as the optimization algorithm in the training process, and set the hyperparameters and expected accuracy of the valve body anomaly detection network model; input the training data set into the valve body anomaly detection network model, train the valve body anomaly detection network model, and stop the training when the loss function stops decreasing or the decrease is very small.

[0027] Further, in step S4, the method for evaluating the valve body anomaly detection network model is as follows: inputting the evaluation data set into the valve body anomaly detection network model, evaluating the valve body anomaly detection network model, and obtaining the accuracy of different anomaly detection through absolute value error.

[0028] Further, in step S5, the method for detecting valve body connection abnormality is as follows: the electric valve is in a full-pipe operation state, and the audio data is continuously uploaded by the sound module and input into the detection module for judgment; after the detection module receives the operation audio at the current moment, it inputs the trained valve body abnormality detection network model, and the model finally outputs whether the valve body of the current valve is abnormally connected, and completes the abnormality detection of the valve body; if the valve body connection is abnormal, a warning is issued and the type of abnormality is informed; if the valve body is operating normally, detection is continuously performed to ensure the normal operation of the electric valve.

[0029] Compared with the prior art, the advantages of the present invention are as follows:

[0030] (1) The valve body connection abnormality detection method of the present invention is based on acoustic signal analysis. Data collection only requires placing a recording device next to the valve to detect the valve body connection abnormality. No additional detection tools are required. The method is easy to operate and has a wide range of applications. It is not limited to the shape and type of the valve and is suitable for long-term use on the valve body of a long-distance water pipeline.

[0031] (2) The present invention detects the abnormal type of valve body connection through audio, which is not limited by the limitations of optical characteristics. The abnormal type detection is comprehensive and the valve body connection abnormality can be detected more widely.

[0032] (3) The present invention updates the current operating status of the valve body in real time after detection by the detection module, thereby improving the valve operating efficiency, detecting valve body abnormalities and solving them in a timely manner. BRIEF DESCRIPTION OF THE DRAWINGS

[0033] Figure 1 It is a flow chart of an embodiment of the method of the present invention.

[0034] Figure 2 yes Figure 1 A structural diagram of a valve body abnormality detection network model of the illustrated embodiment. DETAILED DESCRIPTION

[0035] The present invention is further described in detail below with reference to the accompanying drawings and specific embodiments.

[0036] The valve body connection abnormality detection device of this embodiment includes a sound module, a detection module and an alarm module. The sound module is connected to the detection module, and the detection module is connected to the alarm module. The detection module is equipped with a valve body abnormality detection network model, and the detected valve is an electrically controlled valve.

[0037] Reference Figure 1 The valve body connection abnormality detection method of this embodiment includes the following steps:

[0038] Step S1. Collection and processing of intelligent valve operation data;

[0039] Collect valve operation data set Q in the experimental environment: In different valve operation scenarios, collect the sound signal data of valve operation, including: when the valve is operating normally, use the sound module to collect audio; control the valve switch, use the sound module to collect audio; when the valve body is abnormal, use the sound module to collect audio. Valve body abnormalities include valve body sliding, valve body adhesion, and valve body stuck. Use the sound module to collect audio in the three situations of valve body sliding, valve body adhesion, and valve body stuck. The collected audio is labeled with categories and used as training materials. Category labeling means: when the valve is operating normally, the collected audio is labeled as valve normal operation audio; when the valve body slides, the collected audio is labeled as valve body sliding audio; when the valve body sticks, the collected audio is labeled as valve body adhesion audio; when the valve body is stuck, the collected audio is labeled as valve body stuck audio.

[0040] Collect the public valve operation data set P on the Internet and label the categories by yourself.

[0041] The valve operation data are uniformly cropped and normalized, and short-time Fourier transform is performed to highlight its frequency domain characteristics, which facilitates network training and operation.

[0042] Step S2. Establishing a valve body abnormality detection network model;

[0043] Reference Figure 2, step S2-1. Build a feature extraction module, which includes three layers of one-dimensional dilated convolutional layers and one layer of fully connected layer. The first layer of one-dimensional dilated convolutional layer is connected to the second layer of one-dimensional dilated convolutional layer, the second layer of one-dimensional dilated convolutional layer is connected to the third layer of one-dimensional dilated convolutional layer, and the third layer of one-dimensional dilated convolutional layer is connected to the fully connected layer. The kernel size of the first layer of one-dimensional dilated convolutional layer is 5, the dilation coefficient is 1, and the pad parameters after the dilated convolution are filled to the same size as before the dilated convolution; the kernel size of the second layer of one-dimensional dilated convolutional layer is 5, the dilation coefficient is 2, and the pad parameters after the dilated convolution are filled to the same size as before the dilated convolution; the kernel size of the third layer of one-dimensional dilated convolutional layer is 5, the dilation coefficient is 4, and the pad parameters after the dilated convolution are filled to the same size as before the dilated convolution; the first layer of one-dimensional dilated convolutional layer and the second layer of one-dimensional dilated convolutional layer use PReLU as the activation function after dilated convolution. Through dilated convolution, the model can widely extract important features of the valve operation audio on the valve body connection. At the same time, compared with ordinary convolution, the use of dilated convolution can expand its receptive field and enhance the feature generalization ability of the network. Three layers of one-dimensional dilated convolution layers are used to extract features of waveform data. Finally, a fully connected layer is required to output the features as a fixed 512-dimensional feature. The final extracted features can characterize the basic characteristics of sound events and facilitate subsequent abnormal event detection.

[0044] Step S2-2. Build two layers of one-dimensional residual dilated convolutional layers. The kernel size of the first layer of one-dimensional residual dilated convolutional layer is 5, the dilation coefficient is 4, and no padding is set; the kernel size of the second layer of one-dimensional residual dilated convolutional layer is 5, the dilation coefficient is 2, and no padding is set. Here, the two layers of one-dimensional residual dilated convolutional layers use the dilation coefficient opposite to that of the feature extraction module, and no padding is set, so that the feature dimension is further reduced. At the same time, the residual dilated convolution changes the features nonlinearly, and can nonlinearly process the audio of the detected valve under different operating conditions.

[0045] Step S2-3. Build a fully connected layer, pass through two layers of one-dimensional residual expansion convolution layers, the features have undergone sufficient nonlinear changes, and then use the fully connected layer to linearly change the features and output N-dimensional features. N represents the number of anomalies that need to be detected. In this embodiment, N=4, including four valve body states: normal operation of the valve body, valve body sliding, valve body adhesion, and valve body stuck.

[0046] Step S2-4. Build an activation layer, use the Softmax activation function, and use the Softmax activation function to output the likelihood of the four valve states. At the same time, adjust the hyperparameters and set the likelihood limit thresholds of the four valve states. If the likelihood of a certain valve state is greater than the corresponding likelihood limit threshold, it is determined that the valve state is in that state at that moment, otherwise it is determined that the valve state is not in that state at that moment.

[0047] In the actual detection process, due to environmental changes and the requirement for audio continuity, further processing is required after the likelihood is obtained. In order to cope with environmental changes and expand the robustness and versatility of the network, it is necessary to obtain the likelihood of the valve body working normally in the experimental environment as a benchmark reference, denoted as p; when the environment changes, the increase or decrease of the likelihood limit threshold can be determined based on p. At the same time, since the valve operation audio is continuous, this embodiment uses an average weighted method to process continuous audio, specifically as follows: the audio waveforms of the three sections of the valve from 2s to 1s before the point, from 1.5s to 0.5s before, and from 1s before to the current time point are sent to the detection module to obtain the likelihood of the valve body state in the three time ranges, and the likelihood of each valve body state at the current moment is obtained by weighted average with weights of 0.1, 0.2, and 0.7, and then compared with the likelihood limit threshold of each valve body state. If the obtained likelihood of the valve state is greater than the corresponding likelihood limit threshold, it is determined that the valve body is in this valve state at this moment and the valve body is in this classification state. Otherwise, it is determined that the valve body is not in this valve state at this moment and the valve body has not experienced this classification event. If the classification state is an abnormal state, the sound detection module notifies the alarm module of it and warns the user to check the valve.

[0048] Step S3. Training a valve body abnormality detection network model;

[0049] Divide the training data set and the evaluation data set, select 80% of the public valve operation data set P and the self-collected valve operation data set Q as the training data set, and the remaining 20% ​​as the evaluation data set; select cross entropy as the loss function and stochastic gradient descent as the optimization algorithm in the training process, set the hyperparameters and expected accuracy of the valve anomaly detection network model; input the training data set into the valve anomaly detection network model, train the valve anomaly detection network model, and stop training when the loss function stops decreasing or the decrease is very small.

[0050] Step S4. Evaluate the valve body abnormality detection network model;

[0051] The evaluation data set is input into the valve body anomaly detection network model to evaluate the valve body anomaly detection network model, and the accuracy of different anomaly detection is obtained through the absolute value error.

[0052] Step S5. Detecting abnormal valve body connection;

[0053] To detect abnormal valve connection, first put the electric valve in full pipe operation state, use the sound module to continuously upload the audio data and put it into the detection module for judgment. After the detection module receives the current running audio, it inputs the trained valve body abnormality detection network model. The model finally outputs whether the valve body of the current valve is abnormally connected, and completes the abnormality detection of the valve body. If the valve body connection is abnormal, a warning is issued and the abnormality is informed; if the valve body is operating normally, continuous detection is performed to ensure the normal operation of the electric valve.

[0054] The present invention can timely and comprehensively detect abnormal valve body connection, improve valve operation efficiency, has a wide range of applications, is not limited to the shape and type of the valve, and can comprehensively detect abnormal types.

[0055] Those skilled in the art may make various modifications and variations to the present invention. If these modifications and variations are within the scope of the claims of the present invention and their equivalents, then these modifications and variations are also within the protection scope of the present invention.

[0056] The contents not described in detail in the specification are prior art known to those skilled in the art.

Claims

1. A method for detecting abnormal valve body connection, It is characterized in that The following steps are involved: Step S1. Collection and processing of intelligent valve operation data; Step S2. Establishing a valve body abnormality detection network model; Step S3. Training a valve body abnormality detection network model; Step S4. Evaluate the valve body abnormality detection network model; Step S5. Detecting abnormal valve body connection; The method of establishing a valve body abnormality detection network model is as follows: Step S2-1. Build a feature extraction module, which includes three layers of one-dimensional dilated convolutional layers and one layer of fully connected layer. The first layer of one-dimensional dilated convolutional layer is connected to the second layer of one-dimensional dilated convolutional layer, the second layer of one-dimensional dilated convolutional layer is connected to the third layer of one-dimensional dilated convolutional layer, and the third layer of one-dimensional dilated convolutional layer is connected to the fully connected layer; the kernel size of the first layer of one-dimensional dilated convolutional layer is 5, the dilation coefficient is 1, and the pad parameters after the dilated convolution are filled to the same size as before the dilated convolution; the kernel size of the second layer of one-dimensional dilated convolutional layer is 5, the dilation coefficient is 2, and the pad parameters after the dilated convolution are filled to the same size as before the dilated convolution; the kernel size of the third layer of one-dimensional dilated convolutional layer is 5, the dilation coefficient is 4, and the pad parameters after the dilated convolution are filled to the same size as before the dilated convolution; The first one-dimensional dilated convolution layer and the second one-dimensional dilated convolution layer use PReLU as the activation function after dilated convolution; three layers of one-dimensional dilated convolution layers are used to extract features from waveform data, and finally a fully connected layer is used to output the features as a fixed 512-dimensional feature. The extracted features can finally characterize the basic characteristics of the sound event; Step S2-2. Build two layers of one-dimensional residual dilated convolutional layers. The kernel size of the first layer of one-dimensional residual dilated convolutional layer is 5, the dilation coefficient is 4, and no padding is set; the kernel size of the second layer of one-dimensional residual dilated convolutional layer is 5, the dilation coefficient is 2, and no padding is set; the residual dilated convolution processes the audio of the detected valve in different operating states nonlinearly; Step S2-3. Build a fully connected layer, use the fully connected layer to linearly transform the features, and output N-dimensional features; N represents the number of anomalies that need to be detected; Step S2-4. Build an activation layer, use the Softmax activation function in the activation layer, and use the Softmax activation function to output the likelihood of the four valve body states; at the same time, adjust the hyperparameters and set the likelihood limit thresholds of the four valve body states. If the likelihood of a certain valve body state is greater than the corresponding likelihood limit threshold, it is determined that the valve body is in this valve body state at this moment, otherwise it is determined that the valve body is not in this valve body state at this moment.

2. The method for detecting abnormal valve body connection according to claim 1, Features: In step S1, the method for collecting and processing the intelligent valve operation data is as follows: Step S1-1. Collect valve operation data set Q in an experimental environment: Under different valve operation scenarios, collect sound signal data during valve operation, and label the collected audio by category; Step S1-2. Collect the public valve operation data set P on the Internet and label the categories by yourself; Step S1-3: uniformly crop and normalize the valve body operation data, and perform short-time Fourier transform to highlight its frequency domain characteristics.

3. The method for detecting abnormal valve body connection according to claim 2, Features: In step S1-1, sound signal data of valve operation are collected in different valve operation scenarios, including: when the valve is operating normally, audio is collected by using the sound module; when the valve switch is controlled, audio is collected by using the sound module; when the valve body is abnormal, audio is collected by using the sound module, and valve body abnormalities include valve body sliding, valve body sticking, and valve body stuck. The sound module is used to collect audio in the three situations of valve body sliding, valve body sticking, and valve body stuck, respectively.

4. The method for detecting abnormal valve body connection according to claim 3, Features: In steps S1-1 and S1-2, the category labeling means: when the valve operates normally, the collected audio is labeled as the normal operation audio of the valve; when the valve body slides, the collected audio is labeled as the valve body sliding audio; when the valve body is sticking, the collected audio is labeled as the valve body sticking audio; when the valve body is stuck, the collected audio is labeled as the valve body stuck audio.

5. The method for detecting abnormal valve body connection according to claim 1, Features: In step S2-3, N=4, including four valve body states: normal operation of the valve body, valve body sliding, valve body adhesion, and valve body stuck.

6. The method for detecting abnormal valve body connection according to claim 1, Features: In step S2-4, the principle of using the Softmax activation function to output the likelihood of the four valve body states is: The continuous audio is processed using an average weighted method. Specifically, the audio waveforms of the valve in three sections, from 2s to 1s before the point, from 1.5s to 0.5s before the point, and from 1s to the current time point, are sent to the detection module to obtain the likelihood of the valve state in the three time ranges. The likelihood of each valve state at the current moment is obtained by weighted average with weights of 0.1, 0.2, and 0.7, and then compared with the likelihood limit threshold of each valve state. If the obtained likelihood of the valve state is greater than the corresponding likelihood limit threshold, it is determined that the valve body is in that valve state at that moment; otherwise, it is determined that the valve body is not in that valve state at that moment.

7. The method for detecting abnormal valve body connection according to claim 2, Features: In step S3, the method for training the valve body anomaly detection network model is as follows: divide the training data set and the evaluation data set, select 80% of the public valve body operation data set P and the self-collected valve operation data set Q as the training data set, and the remaining 20% ​​as the evaluation data set; select cross entropy as the loss function, and random gradient descent as the optimization algorithm in the training process, and set the hyperparameters and expected accuracy of the valve body anomaly detection network model; input the training data set into the valve body anomaly detection network model, train the valve body anomaly detection network model, and stop the training when the loss function stops decreasing or the decrease is very small.

8. The method for detecting abnormal valve body connection according to claim 7, Features: In step S4, the method for evaluating the valve body anomaly detection network model is as follows: inputting the evaluation data set into the valve body anomaly detection network model, evaluating the valve body anomaly detection network model, and obtaining the accuracy of different anomaly detection through absolute value error.

9. The method for detecting abnormal valve body connection according to claim 1, Features: In step S5, the method for detecting abnormal valve body connection is as follows: the electric valve is in a full-pipe operation state, and the audio data is continuously uploaded by the sound module and input into the detection module for judgment; after the detection module receives the operation audio at the current moment, it inputs the trained valve body abnormality detection network model, and the model finally outputs whether the valve body of the current valve is abnormally connected, completing the abnormality detection of the valve body; if the valve body connection is abnormal, a warning is issued and the type of abnormality is informed; if the valve body is operating normally, detection is performed continuously to ensure the normal operation of the electric valve.

Images