A safety valve checking method and system based on sound signal detection

Abstract

This invention discloses a safety valve verification method and system based on sound signal detection. The method includes: S1, acquiring a raw sound signal set through an array of sound sensors; S2, processing the raw sound signal set obtained in S1 to obtain a sound signal set; S3, extracting multi-dimensional features from the noise-reduced sound signal set, including time-domain features, frequency-domain features, and time-frequency-domain features; S4, fusing the multi-dimensional features and inputting them into a classification model for processing, with the classification model outputting a classification result, including background noise, a slightly opening sound, and a fully activated sound. Thus, the trained classification model can accurately classify based on preprocessed data and corresponding feature samples, effectively distinguishing between slightly opening sounds, fully activated sounds, and background noise, overcoming human error and environmental noise interference, and improving the accuracy and reliability of the activation pressure verification.

Description

Technical Field

[0001] This invention relates to the field of sound recognition technology, and in particular to a safety valve verification method and system based on sound signal detection. Background Technology

[0002] Safety valves are automatic valves widely used in pressure-bearing devices and equipment such as boilers, pressure vessels, and pressure pipelines to prevent accidents caused by excessive pressure. They are an important component of the safety protection system for pressure-bearing special equipment. When the system pressure exceeds a preset safety value, the safety valve activates, releasing the excess medium to the outside of the system and preventing the pressure from rising further. The pressure point at which the safety valve begins to release pressure is called the trip pressure, and the accuracy of the safety valve's trip pressure directly determines the overall safety of the system.

[0003] In practical applications, safety valves require pressure calibration before installation and use. Currently, the determination of the opening pressure is typically done manually by listening to the sound. The calibrator determines the valve's opening pressure by hearing the continuous discharge of the test medium. When the safety valve is in its slightly open state before full opening, there is a faint venting sound. This requires the calibrator's experience to prevent misjudgment, and the test results are related to the calibrator's proficiency. Furthermore, after slightly opening, the pressure needs to be increased until the safety valve reaches full opening. At this point, the valve core will experience a violent bouncing motion accompanied by loud noise. The calibrator needs to concentrate intensely and continuously listen to the faint venting sound and the high-intensity full opening noise for a short period, which can cause hearing loss and psychological stress. Prolonged operation can also lead to fatigue, affecting the accuracy of the manual calibration results. Summary of the Invention

[0004] The purpose of this invention is to overcome the shortcomings of the prior art and provide a safety valve calibration method and system based on sound signal detection. The method can improve the accuracy of safety valve calibration. By deploying sound sensors to collect sound signals in real time during the safety valve start-up calibration process, extracting start-up acoustic features, performing convolutional neural network learning, and automatically determining the start-up pressure state, this calibration method can avoid noise damage to personnel during safety valve start-up pressure calibration, improve the working environment of calibration personnel, and make the calibration results more accurate.

[0005] In order to solve the corresponding technical problems and achieve the corresponding effects, the embodiments of this application are implemented through the following technical solutions:

[0006] In a first aspect, this application discloses a safety valve verification method based on sound signal detection, comprising: S1, synchronously acquiring multiple sound time-domain signals as raw sound data using an array of sound sensors at a set sampling frequency, and using the raw sound data acquired within one gas supply cycle as a set of raw sound signals; S2, processing the raw sound signal set obtained in S1 to obtain a sound signal set; S3, extracting multi-dimensional features from the sound signal set obtained after noise reduction, wherein the multi-dimensional features include time-domain features, frequency-domain features, and time-frequency-domain features; S4, fusing the multi-dimensional features and inputting them into a classification model for processing, wherein the classification model outputs a classification result, wherein the classification result includes background noise, slightly opening sound, and fully opening sound.

[0007] The effect is as follows: In the method of this application embodiment, the trained classification model can accurately classify based on the preprocessed data and the corresponding feature samples, effectively distinguishing between "slight opening sound," "full start-up sound," and "background noise." By accurately identifying the moment corresponding to the "full start-up sound," the pressure value detected by the pressure transmitter at that time can be obtained synchronously and determined as the safety valve's start-up pressure. This algorithm overcomes the problems of human error and environmental noise interference, improving the accuracy and reliability of start-up pressure verification.

[0008] Furthermore, the step of processing the original sound signal set obtained in S1 to obtain a sound signal set includes: processing the original sound signal set obtained in S1, removing missing or abnormal data, and processing the original sound signal set using a preset noise reduction algorithm to process the original sound signal set into a sound signal set.

[0009] Furthermore, the time-domain features include the extracted time-domain maximum, minimum, mean, peak-to-peak value, and standard deviation; the frequency-domain features include the extraction of centroid frequency and average frequency features by performing Fourier transform on each sound time-domain signal in the sound signal set for spectral analysis.

[0010] Furthermore, the extraction steps of the time-frequency domain features include: S31, generating an energy spectrum through short-time Fourier transform; S32, performing logarithmic operations on the energy spectrum to compress the dynamic range and enhance weak signals; S33, generating a logarithmic Mel spectrum based on Mel-scale filter bank processing to simulate the characteristics of human hearing.

[0011] Furthermore, the process of fusing multi-dimensional features and inputting them into the classification model includes: S41, extracting samples from the multi-dimensional features, including: one-dimensional feature samples composed of time-domain features and frequency-domain features, two-dimensional time-series samples obtained by time synchronization based on data collected synchronously from multiple sensors, and RGB image samples obtained by converting the time-frequency analysis results into color images.

[0012] Furthermore, S42 is executed after S41; a neural network model is pre-trained using a preset dataset, and the low-level convolutional layers are frozen during training to extract general edge features, while the high-level convolutional layers are fine-tuned to adapt to the voiceprint abstract features.

[0013] Furthermore, the neural network model is trained into a deep neural network model for noise classification, including a weight locking layer and a parameter update layer.

[0014] Furthermore, it also includes S411A performed before S42; S411A, which fuses the features of the sound signal set processed by the multi-sensor and inputs it into the Transformer model, wherein: the multi-channel signals are merged into a multi-channel image and input into a single Transformer model, and then input into the fully connected layer of the classification model.

[0015] Furthermore, it also includes S411B performed before S42; in S411B, each signal is independently input into a single Transformer model, and multiple Transformer models perform feature fusion before being input into the fully connected layer of the classification model, so as to input the fused features into the fully connected layer of the classification model.

[0016] Secondly, this application discloses a safety valve verification system based on sound signal detection, including a storage device and a processor. The storage device stores a computer program for the aforementioned safety valve verification method based on sound signal detection, and the processor executes the computer program stored in the storage device. Attached Figure Description

[0017] Figure 1 This is a diagram showing the arrangement of the safety valves and various sound sensors in some embodiments of this application;

[0018] Figure 2 This is a schematic diagram of the safety valve verification method based on sound signal detection in some embodiments of this application;

[0019] Figure 3 The energy spectrum diagrams are shown in some embodiments of this application;

[0020] Figure 4 This is a simplified schematic diagram of the training process of the noise classification depth riser network in some embodiments of this application;

[0021] Figure 5 This is a simplified schematic diagram of the multi-channel model and the high-level feature fusion model in some embodiments of this application. Detailed Implementation

[0022] The technical solution of the present invention will be clearly and completely described below with reference to the embodiments. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0023] In existing technologies, some safety valves exhibit a slight opening before tripping, during which a faint leakage sound is emitted from the valve's outlet. Furthermore, the duration of this slight opening is difficult to control accurately, leading to misjudgments during manual inspection that the valve has fully tripped, resulting in significant errors in the calculated tripping pressure. While the calibration process can be quiet, the collected time-domain sound signals can be directly used as the tripping sound signal set. However, in actual calibration workshops, conditions are complex, with significant interference noise, such as worker voices and other equipment noise. The sound sensor picks up both the initial leakage sound and the complex interference noise, resulting in a very low signal-to-noise ratio. Moreover, the sensor then rapidly acquires the loud sound of full tripping, making it difficult for traditional signal amplitude determination to accurately distinguish between the slight opening and full tripping states.

[0024] For this purpose, please refer to Figures 1-5 This invention provides a safety valve calibration method and system based on sound signal detection.

[0025] Before proceeding, let's explain some technical terms:

[0026] ImageNet is a large-scale visual database used in existing technologies for object recognition software research. It has multiple nodes that can be used for model training.

[0027] VGG16 is a deep convolutional neural network model proposed by the Visual Geometry Group (VGG) at Oxford University in 2014.

[0028] In subsequent embodiments, with Figure 1 The following is an example of the arrangement of sound sensors.

[0029] refer to Figure 1Before pressure calibration, the array of acoustic sensors is deployed. Individual acoustic sensors are fixed to the calibration platform (not shown in the figure) using clamping brackets. The first acoustic sensor is positioned 0.3 meters directly in front of the safety valve outlet, and the second acoustic sensor is placed 0.3 meters directly behind the safety valve outlet. Along a horizontal line perpendicular to the first and second acoustic sensors, with the safety valve outlet as the midpoint, the third and fourth acoustic sensors are symmetrically placed on either side, each 0.3 meters from the midpoint. Next, the signal outputs of the four acoustic sensors and the pressure transmitter are connected to the five analog inputs of a multi-channel synchronous data acquisition card, powered by a 12V power supply. Then, the data acquisition card is connected to a host computer. The host computer is used to set the acquisition channels and parameters of the acquisition card. The corresponding acquisition channel of the acquisition card's input terminal must be selected, and the sampling frequency must be greater than twice the highest frequency of the trigger signal to ensure undistorted signal reproduction, thus enabling the acquisition and transmission of sound and pressure signals during pressure calibration.

[0030] The aforementioned trip signal is the high-frequency transient noise generated by the violent impact between the valve core and the valve seat at the moment the safety valve trips.

[0031] After the aforementioned preparations are completed, open the inlet shut-off valve to allow gas to flow through the pipeline to the inlet of the safety valve until the safety valve pressure reaches 90% of the opening pressure. Adjust the shut-off valve to reduce the gas pressure increase rate to 0.01 MPa and maintain a slow pressure increase until the safety valve fully opens. The array sound sensor collects the instantaneous intense noise of the valve core vibration. At this time, the pressure value detected by the pressure transmitter is the opening pressure of the safety valve.

[0032] Next, according to Figure 2 The safety valve verification method based on sound signal detection according to an embodiment of this application specifically includes the following steps:

[0033] S1. Multiple sound time-domain signals are synchronously collected by the aforementioned array sound sensor at a set sampling frequency as raw sound data. The raw sound data collected within one air supply cycle (i.e., from the start of air supply to the valve core returning to its seat) is used as a set of raw sound signals.

[0034] Specifically, during the pressure calibration process, the array of sound sensors arranged as described above synchronously acquires multiple sound time-domain signals according to the set sampling frequency, thereby obtaining multiple initial sound time-domain signals. These acquired sound signals are actually directly obtained raw signals, and therefore are defined as raw sound data, that is, the sound time-domain signals acquired by the four sensors in the aforementioned embodiment. Based on this, all sound time-domain signals acquired within one cycle from the start of air supply to the valve core returning to its seat are taken as a set of raw sound signals, wherein the amplitude of the sound time-domain signals in the raw sound signal set includes both positive and negative directions.

[0035] In some examples, since the opening pressure of different safety valves varies in actual applications, the pressurization time is inconsistent. The gas supply cycle can be set according to the actual situation, and the same sampling cycle can be set for the verification process with similar opening pressure.

[0036] S2. Process the original sound signal set obtained in S1, remove missing or abnormal data, and use a preset noise reduction algorithm to process the original sound signal set into a sound signal set. It can be understood that since the original sound signal collected in real time by the array sound sensor at the pressure calibration site contains a raw mixed signal of the target sound (safety valve working sound) and background noise, and since this raw mixed signal contains the target acoustic signature and environmental noise, noise reduction processing is required.

[0037] In detail, if the original sound signal set is only preprocessed to remove missing, abnormal, or other dirty data, and noise data is not considered, directly using the unprocessed original sound signal set for jump pressure detection may lead to low detection accuracy. Therefore, after obtaining the sound, the preset noise reduction methods used are Singular Value Decomposition (SVD) and wavelet thresholding to denoise the actual sound data. By reducing the noise data in the actual sound data, the signal-to-noise ratio is improved, and the sound signal set is obtained.

[0038] S3. Extract multi-dimensional features from the noise-reduced audio signal set. These multi-dimensional features include time-domain features, frequency-domain features, and time-frequency-domain features. Specifically, the aforementioned time-domain features may include extracting the maximum, minimum, mean, peak-to-peak value, and standard deviation in the time domain. By performing Fourier transform on each audio time-domain signal in the audio signal set and conducting spectral analysis, the centroid frequency and average frequency features are extracted as frequency-domain features.

[0039] The extraction steps for time-frequency domain features include:

[0040] S31. Generate the energy spectrum through short-time Fourier transform (STFT).

[0041] Specifically, by performing a short-time Fourier transform to convert the signal dimension, combining time-domain and frequency-domain information, and presenting it in the form of an energy spectrum, it includes key acoustic characteristic parameters such as frequency, amplitude, and energy. This effectively reflects the relationship between the frequency and amplitude of the original sound signal over time, and also represents the frequency energy distribution of the signal at different times. For example... Figure 3 As shown, the energy spectrum in this example is a linear energy spectrum.

[0042] Furthermore, considering the large size of the energy spectrum obtained after the short-time Fourier transform and the linear frequency scale, which does not match the nonlinear perception characteristics of the human ear, a logarithmic operation is first performed on the linear energy spectrum to compress the dynamic range and improve the discernibility of weak sound components (such as the slight opening sound of a safety valve). Then, a Mel-scale filter bank is applied to process the nonlinear frequency scale that simulates the differences in human ear sensitivity to different frequencies (sensitive to low-frequency differences but insensitive to high-frequency differences). After these two steps, the log-Mel spectrum is finally obtained. Therefore, to obtain sound features of appropriate size, drawing on the auditory characteristics of the human ear, subsequent steps S32-S33 are performed.

[0043] S32. Perform logarithmic operations on the energy spectrum to compress the dynamic range and enhance weak signals.

[0044] S33. Based on Mel-scale filter bank processing, a log-Mel spectrum is generated to simulate the characteristics of human hearing.

[0045] Based on the aforementioned steps, after preprocessing and noise suppression of the original audio signal, feature information of the signal in multiple dimensions such as time domain, frequency domain, and time-frequency domain is extracted to construct a multi-dimensional feature vector, including parameters such as energy distribution, dominant frequency, short-time energy, MFCC coefficient, and envelope curve statistics. Audio signals obtained from multiple stress test experiments are labeled, numbered, and categorized. Based on the extracted multi-dimensional feature vector, an audio signal sample database is constructed.

[0046] Based on this, S4 proceeds by fusing multi-dimensional features and inputting them into the classification model for processing. The classification model outputs the classification results, including background noise, the slightly activated sound, and the fully activated sound. Specifically, S4 includes:

[0047] S41. By extracting time-domain / frequency-domain statistical features from single-channel audio signals in a sound signal set, and concatenating them into a vector, a one-dimensional feature sample is formed by combining time-domain features (maximum value, mean value, peak-to-peak value, etc.) and frequency-domain features (centroid frequency, average frequency, etc.). Based on data synchronously acquired by four sensors, a two-dimensional time-series sample is obtained through time synchronization. The time-frequency analysis results (such as the aforementioned STFT energy spectrum, and then combined with wavelet time-frequency graph) are converted into a color image to obtain RGB image samples. The three channels (R / G / B) can represent the energy intensity or time-frequency coefficients of different frequency components.

[0048] S42. Establish a 16-layer deep convolutional neural network framework and initialize the network parameters. Divide the input audio signal image data with type labels in the audio sample database into training and validation sets in a 7:3 or 8:2 ratio. Use the ImageNet pre-trained VGG16 model, freezing the low-level convolutional layers during training to extract general edge features, and fine-tuning the high-level convolutional layers to adapt to the abstract features of the voiceprint.

[0049] To be more specific, we can combine Figure 4 To understand, the VGG16 model is pre-trained using ImageNet. In this example, based on the functional characteristics of the trained model, it is defined as a deep neural network model for noise classification, which includes a weight locking layer and a parameter update layer.

[0050] During training, the parameters of the lower-level networks are locked, specifically the parameters and weights of the first three convolutional layers. The output layer structure is modified to correspond to the number of sound types. Specifically, the lower-level convolutional layers (convolutional modules 1, 2, and 3) learn edge and contour features from images as generalized features, applicable to both natural and audio signal images. The higher-level convolutional layers, however, have more abstract and concrete features. Therefore, the remaining two convolutional layers and the fully connected module are trained and their parameters adjusted using an audio signal training set, while the validation set is used to assess model performance. The higher-level convolutional layers (convolutional modules 4 and 5) learn more abstract feature representations, namely, abstract voiceprint features, thereby achieving deep transfer learning and establishing a convolutional neural network model based on voiceprint image features. This model is defined as a deep neural network model for noise classification.

[0051] In some preferred embodiments, S411A is also included before S42.

[0052] S411A: The sound signal set processed by multiple sensors is fused into features and input into the Transformer model, wherein: the multiple signals are merged into a multi-channel image and input into a single Transformer model, and then input into the fully connected layer of the classification model.

[0053] refer to Figure 5To understand, a multi-channel model is set up before the classification model. This multi-channel model merges the processed sound signals from four sensors into a multi-channel image, which is then input into a single Transformer model. More specifically, fine-grained classification of voiceprint signals is based on a Transformer neural network model, trained using an attention mechanism. The experimental setup includes multiple sound sensors capable of measuring multi-channel sound signals. A multi-channel model and a high-level feature fusion model are constructed. The multi-channel model works by fusing the signals from multiple sensors at the input layer. The spectrum of each sensor signal occupies one channel, and the entire spectrum is input as a single sample into the Transformer model.

[0054] In some other preferred embodiments, S411B is also included, which is performed before S42.

[0055] S411B. Each signal is independently input into a single Transformer model. Multiple Transformer models are fused before being input into the fully connected layer of the classification model, so that the fused features are input into the fully connected layer of the classification model.

[0056] Continue to refer to Figure 5 To understand this, a high-level feature fusion model is set up before the classification model. This model inputs each signal independently into the Transformer and fuses the features before the fully connected layer. Specifically, the high-level feature fusion model treats each signal from multiple sensors as a separate Transformer input. Only before the fully connected layer and the Softmax classifier, it fuses the features from each signal together, and then inputs them into the aforementioned classification model for classification. This allows for the classification of micro-jump sounds, full-jump sounds, and noise signals, enabling accurate identification of jump pressure through the classified signals.

[0057] S5. When the full start-up sound is detected, the detection value of the pressure transmitter at that moment is obtained simultaneously; this pressure value is determined as the start-up pressure of the safety valve, and the verification result is output.

[0058] Therefore, in the method of this application embodiment, the trained classification model can accurately classify the corresponding feature samples based on the preprocessed data, effectively distinguishing between "slight opening sound," "full start-up sound," and "background noise." By accurately identifying the moment corresponding to the "full start-up sound," the pressure value detected by the pressure transmitter at that time can be obtained synchronously and determined as the safety valve's start-up pressure. This algorithm overcomes the problems of human error and environmental noise interference, improving the accuracy and reliability of start-up pressure verification.

[0059] A safety valve verification system based on sound signal detection according to an embodiment of this application includes a storage device and a processor. The storage device stores a computer program for implementing the safety valve verification method based on sound signal detection described in the foregoing embodiments, and the processor is used to execute the computer program stored in the storage device.

[0060] The above description is merely a preferred embodiment of the present invention. It should be understood that the present invention is not limited to the forms disclosed herein and should not be construed as excluding other embodiments. It can be used in various other combinations, modifications, and environments, and can be altered within the scope of the concept described herein through the above teachings or related technologies or knowledge. Modifications and variations made by those skilled in the art that do not depart from the spirit and scope of the present invention should be within the protection scope of the appended claims.

Claims

1. A safety valve calibration method based on sound signal detection, characterized in that, include: S1. Multiple sound time-domain signals are synchronously collected by the array of sound sensors at a set sampling frequency as raw sound data, and the raw sound data collected within one gas supply cycle is used as a set of raw sound signals. S2. Process the original sound signal set obtained in S1 to obtain a sound signal set; S3. Extract multi-dimensional features from the noise-reduced audio signal set, wherein the multi-dimensional features include time-domain features, frequency-domain features, and time-frequency-domain features; wherein, the extraction steps of the time-frequency-domain features include: S31. Generate the energy spectrum through short-time Fourier transform; S32. Perform logarithmic operations on the energy spectrum to compress the dynamic range and enhance weak signals; S33. Based on Mel scale filter bank processing, generate log-Mel spectrogram to simulate the characteristics of human hearing; S4. The multi-dimensional features are fused and input into a classification model for processing. The classification model outputs a classification result, which includes background noise, a slightly activated sound, and a full jump sound. The process of fusing the multi-dimensional features and inputting them into the classification model includes: S41. Extracting samples from the multi-dimensional features, including: One-dimensional feature samples composed of time-domain and frequency-domain features, two-dimensional time-series samples obtained by time synchronization based on data synchronously collected from multiple sensors, and RGB image samples obtained by converting time-frequency analysis results into color images; S411A is executed after S41; S411A: The sound signal set processed by multiple sensors is fused into features and input into the Transformer model, where: Multiple signals are combined into a multi-channel image and input into a single Transformer model, which is then fed into the fully connected layer of the classification model. S42 is executed after S411A; S42. Train the neural network model using a preset dataset. During the training process, freeze the low-level convolutional layers to extract general edge features and fine-tune the high-level convolutional layers to adapt to the voiceprint abstract features.

2. The safety valve verification method based on sound signal detection according to claim 1, characterized in that, The process of processing the original sound signal set obtained in S1 to obtain the sound signal set includes: The original sound signal set obtained by S1 is processed to remove missing or abnormal data. A preset noise reduction algorithm is used to process the original sound signal set into a sound signal set.

3. The safety valve verification method based on sound signal detection according to claim 1, characterized in that, The time-domain features include the extracted time-domain maximum, minimum, mean, peak-to-peak value, and standard deviation; the frequency-domain features include the extraction of centroid frequency and average frequency features by performing Fourier transform on each time-domain signal in the sound signal set and conducting spectral analysis.

4. The safety valve verification method based on sound signal detection according to claim 1, characterized in that, The neural network model is trained into a deep neural network model for noise classification, including a weight locking layer and a parameter update layer.

5. The safety valve verification method based on sound signal detection according to claim 1, characterized in that, This also includes S411B, which precedes S42; S411B. Each signal is independently input into a single Transformer model. Multiple Transformer models are fused before being input into the fully connected layer of the classification model, so that the fused features are input into the fully connected layer of the classification model.

6. A safety valve calibration system based on sound signal detection, characterized in that, The device includes a storage device and a processor, wherein the storage device stores a computer program for implementing the safety valve verification method based on sound signal detection as described in any one of claims 1-5, and the processor executes the computer program stored in the storage device.

Images