Method for monitoring leakage of liquefaction pump set based on intelligent analysis of sound wave characteristics

Through intelligent analysis of acoustic features and deep learning models, the real-time and sensitivity problems of leakage detection during the transmission of liquefied gas pump sets are solved, and all-round leakage monitoring and control are achieved, which improves the safety of the petrochemical industry.

CN118936776APending Publication Date: 2024-11-12TIANJIN WITNSOFT TECH
View PDF 0 Cites 1 Cited by

Patent Information

Application Number
CN202411181508.6
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2024-08-27
Publication Date
2024-11-12

AI Technical Summary

Technical Problem

In the petrochemical industry, leakage incidents occur frequently during the transmission of liquefied gas pump groups, resulting in property losses and safety hazards. It is difficult for existing monitoring methods to achieve real-time, all-round and high-sensitivity leakage detection.

Method used

Using a method based on intelligent analysis of sound wave characteristics, the leakage sound wave signal is collected through the acoustic wave reception sensor, preprocessing and feature extraction, and automatic identification is achieved using deep learning models (such as convolutional deep belief network) to realize real-time monitoring of leakage of liquefied pump group.

Benefits of technology

It realizes all-weather, non-contact, long-distance and safe leakage monitoring, with high sensitivity and fast response capabilities, and provides precise control and support for leakage of liquefied pump sets.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN118936776A_ABST
    Figure CN118936776A_ABST
Patent Text Reader

Abstract

The invention relates to a method for monitoring leakage of a liquefaction pump set based on intelligent analysis of sound wave characteristics, which comprises the following steps of: acquiring a leakage sound wave signal: arranging a sound wave receiving sensor near a relevant position where leakage of the liquefaction pump set is easy to occur so as to acquire the leakage sound wave signal; preprocessing the leakage sound wave signal: preprocessing the original leakage sound wave signal received by the sound wave sensor, wherein the preprocessing comprises de-noising and filtering processing; extracting leakage sound wave signal features, including extracting a Mel-frequency cepstrum coefficient; preparing pump set leakage sound wave feature training data; aiming at sound wave signal characteristics of pump set leakage, a deep learning mathematical model is established and trained, and automatic identification of liquefied pump set leakage based on sound wave signals is realized; and after model training is completed, the recognition precision of the model is evaluated and optimized repeatedly until a satisfactory result is achieved, and then the model is solidified.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention is oriented to the field of petrochemical industry, relates to the transmission working process of a liquefied pump group, and provides a method for monitoring leakage of a liquefied pump group based on intelligent analysis of acoustic wave characteristics. Background Art

[0002] In recent years, with the rapid development of the petrochemical industry, various safety incidents have occurred frequently. Especially in the process of liquefied gas pump group transmission, leakage often occurs due to pipeline corrosion or loose sealing or excessive pressure and flow; leakage can lead to serious safety hazards in various aspects such as property loss, environmental pollution and personnel safety, which must be taken seriously and require more practical real-time monitoring methods to be given for comprehensive and reasonable monitoring, control and treatment.

[0003] At present, the detection methods for monitoring liquefied pump group leakage are gradually diversified. On the basis of traditional chemical monitoring methods, intelligent detection methods based on emerging technologies such as vision have also gradually developed in recent years. Here, an intelligent method for monitoring liquefied pump group leakage based on acoustic imaging technology is proposed. As a new technical attempt and supplement, it can better monitor and warn the transmission process of the wild flower pump group from all directions and angles, and timely discover safety hazards and safety problems. Summary of the invention

[0004] The purpose of the present invention is to provide a method for monitoring leakage of a liquefied gas pump set based on intelligent analysis of acoustic wave characteristics during the liquefied gas transmission process in the petrochemical industry. The technical solution of the present invention is as follows:

[0005] A method for monitoring leakage of a liquefied pump unit based on intelligent analysis of acoustic wave characteristics comprises the following steps:

[0006] (1) Collecting acoustic wave signals of leakage: Arrange acoustic wave receiving sensors near the relevant locations where leakage is likely to occur in the liquefied pump group to collect acoustic wave signals of leakage.

[0007] (2) Leakage acoustic wave signal preprocessing: The original leakage acoustic wave signal received by the acoustic wave sensor is preprocessed, and the preprocessing includes denoising and filtering.

[0008] (3) Extracting the characteristics of the leakage acoustic wave signal, including extracting the Mel frequency cepstrum coefficients

[0009] (4) Prepare the pump leakage acoustic wave feature training data

[0010] (5) Based on the acoustic wave signal characteristics of pump group leakage, a deep learning mathematical model is established and trained to realize the automatic identification of liquefied pump group leakage based on acoustic wave signals.

[0011] (6) After the model training is completed, the recognition accuracy of the model is evaluated and optimized repeatedly until satisfactory results are achieved and the model is solidified.

[0012] Furthermore, the method of step (3) is as follows:

[0013] 1) Use a first-order high-pass filter for pre-emphasis processing;

[0014] 2) Divide the leakage acoustic wave signal after pre-emphasis processing into frames, and multiply each frame by a window function for windowing.

[0015] 3) Performing fast Fourier transform on each frame signal after frame division and windowing to obtain the spectrum of each frame, and then obtaining the spectral line energy of the leakage sound wave signal.

[0016] 4) The obtained spectral line energy is passed through the Mel triangular filter bank, and the start, middle and cutoff frequencies of each triangular filter in the Mel domain are converted to the linear frequency domain; filtering is performed to obtain the energy of the filter bank, and then a logarithmic operation is performed to obtain the Fbank feature.

[0017] 5) Perform IDFT on the basis of Fbank features to obtain Mel frequency cepstral coefficient MFCC features.

[0018] Furthermore, a Hamming window function may be used for windowing.

[0019] Furthermore, the method of step (5) is as follows:

[0020] Convolutional deep belief network (CDBN) is constructed by simulating the structure of deep belief network (DBN), multiple CRBMs are stacked, and probability maximum pooling is added on the top of each CRBM, that is, the maximum value of the probability is selected in each region, and the regional division has no overlap. The multi-layer network model is trained using the greedy layer-by-layer training method, and the output of each pooling layer is used as the input of the visual layer of the next layer of CRBM network. After training a layer of network, its weights and biases are maintained, and it is activated as the input of the next layer of intelligent recognition network model;

[0021] After constructing the multi-layer network model, the training data set is imported to train the network model. During the training process, unsupervised learning and supervised learning will be performed alternately. The method is as follows: the data input layer is combined with the first hidden layer to form the first layer of the CRBM network, and unlabeled data is imported for unsupervised learning to obtain the initial weights and biases. Then, labeled data is imported on this basis, and supervised learning is performed through the softmax classifier to obtain the adjusted weights and biases. Then, the pooling layer in the first layer of the CRBM network is used to combine with the next hidden layer to form the second layer of the CRBM network, and the corresponding weights and biases are obtained through unsupervised learning. Then, supervised learning is performed through the classifier, and the weights and biases of the two layers are adjusted as the input of the next layer.

[0022] The beneficial effects of the present invention are as follows: The monitoring algorithm based on intelligent analysis of acoustic wave characteristics provided by the present invention has the advantages of strong real-time performance, high sensitivity, strong adaptability, and fast response time. It can realize all-weather, non-contact, long-distance, and safe monitoring, and provides strong technical support for intelligent supervision and precise control of liquefied pump group leakage. BRIEF DESCRIPTION OF THE DRAWINGS

[0023] Figure 1 Schematic diagram of the implementation principle flow chart. DETAILED DESCRIPTION

[0024] The present invention and its application scenarios are described below.

[0025] 1. Collect the acoustic signal of the leak

[0026] Acoustic wave receiving sensors are arranged near the relevant locations where the liquefied pump group is prone to leakage to collect the acoustic wave signals of the leakage, and these acoustic wave receiving sensors are connected to the intelligent analysis and monitoring server.

[0027] 2. Leakage acoustic wave signal preprocessing

[0028] A series of preprocessing steps are performed on the original leakage acoustic wave signals received by the acoustic wave sensor to convert them into a digital form suitable for computer analysis. In this process, the acoustic wave signal is first sampled by a sampler, and its continuous analog signal is discretized into a digital signal at a certain frequency; then a quantization operation is performed to convert the voltage value of each sampling point into a finite number of numerical levels, so that the acoustic wave information can be digitally stored and transmitted.

[0029] 3. Noise elimination and filtering of leakage acoustic wave signals

[0030] In order to ensure the effectiveness and accuracy of subsequent analysis, the pre-processed digital signal needs to be denoised and filtered. Denoising is to remove background noise and other non-target sound wave interference; filtering focuses on separating effective signals within a specific frequency band. These pre-processing methods can effectively improve the quality of the signal and reduce the impact of noise on subsequent recognition algorithms.

[0031] 4. Extract leakage acoustic wave signal characteristics

[0032] Then comes the feature extraction stage, extracting parameters that reflect the essential characteristics of the sound wave from the preprocessed digital signal. Currently, Mel Frequency Cepstral Coefficient (MFCC) is a new and widely used feature representation method, which contains important information such as the main frequency, energy distribution and harmonic structure of the sound wave.

[0033] (1) Pre-emphasis:

[0034] Pre-emphasis is to increase the energy of the high-frequency part of the signal. Since high-frequency signals decay quickly during transmission, but the high-frequency part contains many features that are beneficial to sound wave recognition, the energy of the high-frequency part needs to be increased in the feature extraction part.

[0035] The pre-emphasis filter is a first-order high-pass filter. Given a time domain input signal x[n], the signal after pre-emphasis is:

[0036] y(n)=x(n)-α·x(n-1) (1)

[0037] Among them, 0.9≤α≤1.0.

[0038] (2) Frame division and windowing:

[0039] Since the sound wave signal has non-stationary characteristics and short-term stationary characteristics, it is necessary to divide the sound wave signal into frames. One frame has N sampling points. If the value of N is 256 or 512, the time covered is about 20 to 30ms. In order to avoid excessive changes between two adjacent frames and smooth transitions, there will be an overlapping area between two adjacent frames. This overlapping area contains M sampling points. Usually the value of M is about 1 / 2 or 1 / 3 of N.

[0040] In order to alleviate spectrum leakage, each frame needs to be multiplied by a window function for windowing. Commonly used window functions include rectangular window and Hamming window. The process of windowing actually cuts off the signal in the time domain. The window function is multiplied with the signal in the time domain, which is equivalent to convolution of the frequency domain representation.

[0041] This embodiment adopts the Hamming window function, which is defined as:

[0042]

[0043] Where L is the window length.

[0044] (3) Frequency domain transformation:

[0045] Since it is usually difficult to see the characteristics of a signal from its transformation in the time domain, it is usually converted into energy distribution in the frequency domain for observation. Different energy distributions can represent the characteristics of different sound waves. The sound wave frames after the previous step are transformed from the time domain to the frequency domain, and the spectral characteristics are obtained through conversion. The specific steps are as follows:

[0046] 1) FFT

[0047] Since it is usually difficult to see the characteristics of a signal from its transformation in the time domain, it is usually converted into energy distribution in the frequency domain for observation. Different energy distributions can represent the characteristics of different sound waves. Therefore, after multiplying the Hamming window, each frame must be subjected to a fast Fourier transform to obtain the energy distribution in the spectrum. The spectrum of each frame is obtained by performing a fast Fourier transform on each frame of the signal after framing and windowing.

[0048]

[0049] Where x(n) is the input sound wave signal, and N represents the number of Fourier transform points.

[0050] 2) Spectral line energy

[0051] The spectrum line energy of the sound wave signal is obtained by taking the modulus square of the spectrum after FFT.

[0052] (4) Mel filter bank and logarithmic operation

[0053] The obtained energy spectrum is passed through a Mel triangular filter bank to define a filter bank with M filters (the number of filters is close to the number of critical bands). The filter used is a triangular filter with a center frequency of f(m). The formula of the Mel filter bank is:

[0054]

[0055] The starting, middle and cutoff frequencies of each triangular filter in the Mel domain are converted to the linear frequency domain; the energy spectrum is filtered to obtain the energy of the filter bank, and then a logarithmic operation is performed to obtain the Fbank feature.

[0056] (5) Cepstrum transform

[0057] Based on the Fbank feature, the IDFT is performed to obtain the Mel-frequency cepstral coefficient (MFCC) feature; the cepstral transform is to inversely transform the logarithm of the spectrum to extract the MFCC feature. The specific steps are as follows: first calculate the logarithmic spectrum, then perform discrete cosine transform (DCT) on the logarithmic spectrum to obtain the Mel-frequency cepstral coefficient (MFCC) feature.

[0058] 5. Prepare the pump group leakage acoustic wave feature training data set

[0059] The pump group leakage acoustic wave signals are collected multiple times and relevant features are extracted to form a large number of training data sets for use.

[0060] 6. Develop a deep learning model algorithm for automatic identification of pump group leakage

[0061] By adopting automatic identification technology based on deep learning, appropriate mathematical algorithms and parameters are used to model the typical characteristics of the acoustic wave signals of pump group leakage, and a deep learning mathematical model is developed to realize the automatic identification function of liquefied pump group leakage based on acoustic wave signals.

[0062] First, a convolutional deep belief network (CDBN) is constructed by simulating the structure of a deep belief network (DBN). That is, multiple CRBMs (convolutional restricted Boltzmann machines) are stacked, and probability maximum pooling is added on top of each CRBM, that is, the maximum probability is selected for each region, and the region division does not overlap. A greedy layer-by-layer training method similar to DBN is used to train the network algorithm model for intelligent classification and recognition. The output of each pooling layer is used as the input of the visual layer of the next CRBM network. After training a layer of network, its weights and biases are maintained, and it is activated as the input of the next layer of intelligent recognition network model.

[0063] After constructing the multi-layer network model, import the pre-prepared training data set to train the network model. During the training process, unsupervised learning (unlabeled data) and supervised learning (labeled data) will be performed alternately. First, the data input layer is combined with the first hidden layer to form the first layer of the CRBM network. Unlabeled data is imported for unsupervised learning to obtain the initial weights and biases. Then, labeled data is imported on this basis, and supervised learning is performed through the softmax classifier to obtain the adjusted weights and biases. Then, the pooling layer in the first layer of the CRBM network is used to combine with the next hidden layer to form the second layer of the CRBM network. The corresponding weights and biases are obtained through unsupervised learning. Then, supervised learning is performed through the classifier, and the weights and biases of the two layers are adjusted as the input of the next layer. This learning process can be analogized layer by layer, and the entire intelligent recognition network model can be trained in the same way.

[0064] After the model training is completed, the recognition accuracy of the model is evaluated through reasonable means and optimized repeatedly until satisfactory results are achieved and the model is solidified.

[0065] 7. Field Application

[0066] The optimized and solidified system is deployed at the user's application site. It can be used as a single monitoring method or combined with other monitoring methods to achieve real-time monitoring of the leakage of the liquefied pump group.

Claims

1. A method for monitoring leakage of a liquefied pump unit based on intelligent analysis of acoustic wave characteristics, comprising the following steps: (1) Collecting acoustic wave signals of leakage: Arrange acoustic wave receiving sensors near the relevant locations where leakage is likely to occur in the liquefied pump group to collect acoustic wave signals of leakage. (2) Leakage acoustic wave signal preprocessing: Preprocess the original leakage acoustic wave signal received by the acoustic wave sensor. Preprocessing includes denoising and filtering. (3) Extracting the characteristics of the leakage acoustic wave signal, including extracting the Mel frequency cepstrum coefficients (4) Prepare the pump leakage acoustic wave feature training data (5) Based on the acoustic wave signal characteristics of pump group leakage, a deep learning mathematical model is established and trained to realize the automatic identification of liquefied pump group leakage based on acoustic wave signals. (6) After the model training is completed, the recognition accuracy of the model is evaluated and optimized repeatedly until satisfactory results are achieved and the model is solidified.

2. The method for monitoring leakage of a liquefied pump unit based on intelligent analysis of acoustic wave characteristics according to claim 1 is characterized in that: The method of step (3) is as follows: 1) Use a first-order high-pass filter for pre-emphasis processing; 2) Divide the leakage acoustic wave signal after pre-emphasis processing into frames, and multiply each frame by a window function for windowing. 3) Performing fast Fourier transform on each frame signal after frame division and windowing to obtain the spectrum of each frame, and then obtaining the spectral line energy of the leakage sound wave signal. 4) The obtained spectral line energy is passed through a Mel triangular filter bank to convert the start, middle and cutoff frequencies of each triangular filter in the Mel domain into the linear frequency domain; After filtering and obtaining the energy of the filter bank, a logarithmic operation is performed to obtain the Fbank feature. 5) Perform IDFT on the basis of Fbank features to obtain Mel frequency cepstral coefficient MFCC features.

3. The method for monitoring leakage of a liquefied pump unit based on intelligent analysis of acoustic wave characteristics according to claim 2 is characterized in that: The Hamming window function is used for windowing.

4. The method for monitoring leakage of a liquefied pump unit based on intelligent analysis of acoustic wave characteristics according to claim 1 is characterized in that: The method of step (5) is as follows: Convolutional deep belief network (CDBN) is constructed by simulating the structure of deep belief network (DBN), multiple CRBMs are stacked, and probability maximum pooling is added on the top of each CRBM, that is, the maximum value of the probability is selected in each region, and the regional division has no overlap. The multi-layer network model is trained using the greedy layer-by-layer training method, and the output of each pooling layer is used as the input of the visual layer of the next layer of CRBM network. After training a layer of network, its weights and biases are maintained, and it is activated as the input of the next layer of intelligent recognition network model; After constructing the multi-layer network model, the training data set is imported to train the network model. During the training process, unsupervised learning and supervised learning are performed alternately. The method is as follows: the data input layer is combined with the first hidden layer to form the first layer of the CRBM network, unlabeled data is imported for unsupervised learning, and the initial weights and biases are obtained. Then, labeled data is imported on this basis, and supervised learning is performed through the softmax classifier to obtain the adjusted weights and biases. The pooling layer in the first CRBM network is then used to form the second CRBM network in combination with the next hidden layer. The corresponding weights and biases are obtained through unsupervised learning. Then, supervised learning is performed through the classifier, and the weights and biases of the two layers are adjusted as the input of the next layer.

Citation Information

Cited By

  • Oil and gas equipment leakage acoustic characteristic intelligent identification early warning method and system

    CN121230973A