Feature extraction and pattern recognition of gas pipeline flow noise signals in a strong noisy background

Construction of experimental platform

The experimental platform established in this study mainly includes three parts: the generation system of flow noise signal, the signal acquisition and processing system and the data analysis system. The generation system of flow noise signal is realized by a closed-loop gas circulation system, which consists of air source, pressure regulating valve, flowmeter, pipeline (including straight pipe section, elbow and tee, etc.), microphone receiver and exhaust system. The pressure and flow rate of gas are controlled by adjusting the pressure regulating valve to simulate different gas pipeline working conditions. Secondly, the sensors used in the experiment mainly use PCB model 106B dynamic pressure sensor and AWA model AWA14423 acoustic sensor to capture noise signals. PCB 106B dynamic pressure sensor has high sensitivity and wide frequency response range (2.5 Hz–20 kHz), which is very suitable for detecting low-intensity flowing noise signals. AWA14423 acoustic wave sensor is used to capture environmental noise, which is characterized by good directivity and high signal-to-noise ratio, and can effectively distinguish target signals from background noise. All signal acquisition is preprocessed by special signal processing software, including filtering, amplification and A/D conversion to ensure the accuracy and stability of data. Finally, in order to verify the repeatability and reliability of the experiment, the experiment is repeated for many times at different times and under different environmental conditions, and the collected data was strictly analyzed statistically. The experimental results show that the experimental data have good consistency and repeatability, which proves the stability of the experimental platform and the effectiveness of the sensors used.

Simulation of gas pipeline flow noise

This study primarily focuses on investigating the flow noise of gas inside buried urban gas pipelines. The internal gas flow noise in pipelines is mainly caused by the relative motion between the gas and the solid wall, as well as the changes in the flow conditions that occur when the gas passes through resistance components in the pipeline, leading to the formation of turbulence or vortices. Flow noise simulations and analyses were conducted for elbow sections of different diameters, pressures, and velocities. A total of 75 sets of operating conditions were simulated, as indicated in Table 2.

The pipeline system under the operating condition of DN25, flow velocity of 3 m/s, and outlet pressure of 0.01 MPa was initialized and simulated in a steady state to determine the locations of monitoring points. The steady-state results are shown in Fig. 1.

After conducting steady-state calculations, the sound pressure level distribution map of the pipeline system shows that the flow noise signal is stronger at the tee junction and on the inner surface of the elbow. Based on this observation, receiver signal acquisition points were set at these locations. Additionally, signal acquisition points were also set in the straight pipe section to serve as the receiving points for the gas pipeline flow noise signal before encountering any resistance components. The locations of the receiver points at the tee junction and elbow are shown in Fig. 2.

Signal receiver points were set at the top, bottom, left, and right of each cross-section of the elbow and straight pipe components. In addition, signal receiver points were set at the upper and lower portions as well as the outer surface of the tee junction, resulting in a total of 30 signal receiver points. Figure 3 shows the contour maps of sound pressure level, pressure, and velocity obtained from the flow field simulation analysis.

Figure 3 shows that areas with high sound pressure levels in the sound pressure level cloud map correspond to higher velocities in the velocity cloud map. In the pressure cloud map, the pressure fluctuations are relatively small, except for the three-way junction where the pressure values are higher. Additionally, Fig. 4 shows that the gas pipeline flow noise signal exhibits different spectral characteristics in straight pipe sections and when encountering resistance components such as bends and three-way junctions. These variations follow a regular pattern, with a repeated trend of decreasing sound pressure levels followed by an increase as the frequency increases. At bends and three-way junctions, the flow conditions change, leading to turbulence, eddies, and relative motion between the internal fluid medium and the solid wall. These changes result in a more pronounced variation of the sound pressure level compared to straight pipe sections, and within each frequency range exhibiting regular changes, numerous fluctuations occur. Overall, the sound pressure level decreases. The trend of sound pressure level variation with frequency remains consistent at each signal receiving point in the pipeline system, with no significant differences in numerical values. When compared to straight pipe sections and three-way junctions, bends exhibit the largest fluctuation range in sound pressure level, with a minimum value close to 40 dB and a maximum value exceeding 180 dB.

To validate the accuracy of the simulated flow noise signal characteristics, a comparison was made between the experimental and simulated sound pressure level values at various frequency points for the gas pipeline flow noise signal. This comparison aimed to verify the accuracy of the numerical simulation. Table 3 presents the sound pressure level values at different frequency points for the same measurement point, with a velocity of 3 m/s, obtained from both the experimental and numerical simulation data. Additionally, a collection of strong background noise signals, such as footstep noise, speech, and construction noise, was recorded, totaling 25 sets, for pattern recognition purposes.

By comparing the experimental data and simulated data in Table 3, it shows that the maximum discrepancy between the experimental and simulated results is only 2.68%, indicating a small error. Furthermore, considering that the sensor itself has an error range of ±1.5 dB, this confirms the accuracy of the numerical simulation.

Simulation of surface flow noise signals in soil

The propagation of sound and the vibration of structures are interrelated phenomena. The vibration of structures generates sound waves, while the propagation of sound causes structural vibrations. In this study, simulation was conducted using LMS Virtual.Lab. The specific analysis process is outlined as follows in Fig. 5.

Based on the total length and width of the pipeline system model, as well as the specifications for the burial depth of gas pipelines outlined in the “Urban Gas Design Code 2020”, a pipeline model with a depth of 0.6m was established to extract the gas pipeline flow noise signal propagated through the soil. The three-dimensional solid model of the buried gas pipeline is shown in Fig. 6, and the locations of the signal collection points are indicated in Fig. 7.

The CGNS sound source file of the gas pipeline flow noise signal was input for the operating conditions of DN25, flow velocity of 3 m/s, and outlet pressure of 0.01 MPa. Cloud maps at the same frequency for three different soil porosity conditions (0.4, 0.5, and 0.6) and frequency domain plots of sound pressure levels at the signal collection points on the soil surface are shown in Figs. 8, 9 and 10.

The cloud maps depicting the propagation of gas pipeline flow noise signals in soil under different porosity conditions reveal that there are minimal differences in signal propagation among varying porosities. However, numerical analysis indicates that as the porosity increases, the values of the gas pipeline flow noise signal reaching the soil surface decrease. The frequency domain plots obtained at the signal collection points on the soil surface for different porosity conditions show negligible influence of porosity on the characteristics of the gas pipeline flow noise signal in the soil. To facilitate better comparison, the frequency domain plots of the gas pipeline flow noise signal at the same signal collection point for porosities of 0.4, 0.5, and 0.6 are overlaid on a single coordinate axis, as illustrated in Fig. 11.

From Fig. 11, it can be visually observed that with increasing porosity, the sound pressure level of gas pipeline flow noise on the soil surface slightly decreases, with a numerical reduction of approximately 1–3 dB. The gas pipeline flow noise signal transmitted to the soil surface has the highest sound pressure level value when the porosity is 0.4.

By comparing the surface flow noise signal with the pipeline flow noise signal, and by performing Fourier Transform on the sound pressure level frequency domain plots at the signal collection points on the soil surface and the gas pipeline flow noise sound pressure level frequency domain plots, the corresponding time domain plots for both signals were obtained. The time domain and frequency domain plots of the gas pipeline flow noise signal and the surface noise signal are illustrated in Figs. 12 and 13.

Note: It is assumed that Figs. 12 and 13 are mentioned in the text, but their descriptions are not provided.

Figures 12 and 13 show that the frequency domain plots are not exactly the same. This is because the surface noise signal collected at the signal collection points in the soil is not the same as the noise signal generated by the gas pipeline flow. However, the overall trend indicates that the time domain and frequency domain characteristics of the gas pipeline flow noise signal and the propagated gas pipeline flow noise signal on the soil surface remain unchanged. Both exhibit a decreasing trend in sound pressure level as the frequency increases. Additionally, there are localized fluctuations in sound pressure level within specific frequency ranges, reaching extreme values. The time domain characteristics also exhibit consistent trends.

Signal feature extraction

Signal feature extraction is the process of extracting information from signals and serves as a fundamental and crucial step in various fields such as pattern recognition, intelligent systems, and mechanical fault diagnosis. The wide applicability of feature extraction has led to its extensive use in speech analysis, image recognition, geological surveys, weather forecasting, mechanical fault diagnosis, and almost all branches of science and engineering (Liu, Xiao & Zhu, 2020; He, Wu & Runwei, 2020; Qiao, Khishe & Ravakhah, 2021). Frequency domain-based feature extraction is commonly employed as it provides clear physical interpretations and offers more intuitive feature information compared to time domain waveforms.

The high-frequency noise components in the original signal are eliminated by applying a low-pass filter before any characteristics are extracted from the gas pipeline flow noise signal. In Fig. 14, the red curve represents the result of low-pass filtering applied to the signal.

After applying low-pass filtering to the frequency domain signal, it is segmented using a frame-based approach. Considering that the studied signal is a non-periodic continuous signal, the frame length is set to 400 and the frame shift is set to 300. Figure 15 illustrates the result of signal segmentation using frames.

From the figure, it shows that after segmentation, the signal is divided into 13 frames labeled as S1, S2, …, S13. Each frame contains the combined characteristic information of different frequency bands of the gas pipeline flow noise signal on the soil surface. Therefore, the feature vectors of sound pressure level are extracted for each frame. Upon observing the first frame, it can be noted that the first sample represents the characteristic of significant attenuation in the gas pipeline flow noise signal. Consequently, two sound pressure level feature vectors are selected from the first frame, resulting in a total of 14 sound pressure level feature vectors extracted. Table 4 presents a set of sound pressure level feature vectors, where 1, 2, and 3 correspond to the three signal collection points on the soil surface.

Figure 16 shows that there are minimal differences and no significant gradients in the extracted sound pressure level frames from the three signal points of the pipeline. Directly recognizing the high-dimensional feature vectors calculated at the 14 points would result in excessive computational complexity, and the performance of the recognition algorithm would be compromised due to the strong correlation between each frame signal. Therefore, an optimization of the original feature set is performed using PCA for dimensionality reduction. PCA is a commonly used data dimensionality reduction technique. Its working principle is to transform the original data into a new coordinate system through orthogonal transformation, so that the data variance on the first axis of the new coordinate system is the largest, that is to say, first principal component represents the largest variability of the data, the second principal component is orthogonal to first principal component and represents the second largest variability, and so on. In this way, the main features in the original data set can be represented by less principal components. In this study, PCA method is used to extract the most informative principal components from many characteristics of flow noise signals, which can effectively compress the data while preserving the important signal characteristics as much as possible to reduce the complexity of the subsequent classification model.

Firstly, the data samples are standardized and then covariance is calculated. There is a certain degree of correlation between the feature data of the sound signals at different locations of the pipeline. The covariance matrix aims to measure the linear relationships between the data. Finally, the eigenvalues and eigenvectors of the covariance matrix are computed, and the contribution rates of each eigenvalue are determined. The results are shown in Table 5. Additionally, the cumulative contribution rate of the principal components is depicted in the cumulative contribution rate variation plot shown in Fig. 17.

The more principal components there are, the larger the computed noise in the data, and the corresponding model error will increase. Conversely, if there are too few principal components, some important information may be neglected. Therefore, the method of determining the number of principal components should involve the use of cumulative variance contribution rate. Specifically: (1)${\displaystyle P_{\mathrm{CPV}}=\sum _{i=1}^k{\lambda }_i/\sum _{i=1}^m{\lambda }_i.}$

From the cumulative contribution rate exceeding the first five principal components, it shows that the contribution rate of other key principal components is very low and can be eliminated. Therefore, this study selects the first five principal components.

Neural network pattern recognition

BPNN is a multi-layer feedforward neural network for supervised learning. Its learning process includes two stages: forward propagation and backward propagation. In the forward propagation, the input data is transmitted from the input layer, processed by the hidden layer and then transmitted to the output layer. If the actual output of the output layer does not match the expected output, it enters the backward propagation stage. In the back propagation, the error signal is transmitted back layer by layer according to the gradient descent method, and the weight of each neuron is adjusted according to the error signal to reduce the output error. This process is repeated many times until the error of network output falls within an acceptable range. In the task of pattern recognition of flow noise signals, BPNN can learn the characteristics of distinguishing different types of noise signals (such as gas flow noise and background noise) and classify them effectively. The following processes are used to optimize the neural network algorithm (Cui, Chen & Qu, 2017; Zhang, Dong & Yanf, 2020):

(1) Selection of input node count

The dimensionality of the extracted sound pressure level feature vectors and the classification performance in the actual classification process are considered to determine the number of input nodes in the network.

(2) Selection of output layer node count

To recognize the patterns of buried gas pipeline flow noise signals and strong background noise signals, binary encoding is used to determine the output types. The encoding is set to represent the gas flow noise signal inside the buried gas pipeline, while another encoding represents other strong background noise signals.

(3) Selection of hidden layer node count

The optimal number of hidden layer nodes is determined by analyzing and comparing the training errors under different conditions of hidden layer nodes and considering the variations in different testing errors.

(4) Selection of initial weight values

Three commonly used methods are employed to determine the initial weight values:

(1) Random selection of initial weight values within a closed interval.

(2) Random selection of initial weight values within a small interval near zero.

(3) Selecting different approaches: Initializing the connection weights from the input layer to the hidden layer with small random values, and initializing the connection weights from the hidden layer to the output layer as either −1 or 1.

(5) Selection of learning rate

The learning rate (lr) is generally chosen within the range of 0.001 to 1. Multiple sets of learning rates are compared and validated to select the optimal learning rate for neural network training.

(6) Selection of desired error

The desired error also plays a significant role in the learning and training process of the neural network. Its selection determines the number of hidden layers and the number of nodes. The relationship between the desired error and the accuracy of neural network classification is not inversely proportional. Therefore, in practical selection, it is sufficient to choose a desired error that meets the requirements. Extremely low desired error values can make the network more complex, leading to a decrease in overall accuracy of the neural network, as shown in the Fig. 18.

The BPNN classification model constructed in this study adopts a three-layer topology structure with 5 input nodes, 10 hidden layer nodes, and 2 output nodes. Therefore, the network node topology structure is 5 × 10 × 2. The model is illustrated in Fig. 19.

For the network model aimed at classifying gas flow noise signals, the mean squared error is chosen as the objective function. The network is trained for 1,000 steps with a target error coefficient of 0.001 and a learning rate of 0.001. Based on the extracted gas signal feature database, the usual training-to-testing ratio is 8:2. There are a total of 225 sets of buried gas flow noise feature data, with 200 sets used for model training and the remaining 25 sets for testing. Additionally, there are 25 sets of feature data for strong background noise signals, with 20 sets used for model training and the remaining 5 sets for testing. The normalized results of the sound pressure level feature vector data for buried gas pipeline flow noise and strong background noise signals can be found in Tables 6 and 7, respectively. The trained model is then evaluated by inputting the test data into it to classify gas flow noise signals (Hu & Zhang, 2011). The root mean square error variation curve during the training of the gas signal network is shown in Fig. 20.

In Fig. 20, the mean squared error between the network output and the expected values decreases as the number of iterations increases. When the number of iterations reaches 45, the root mean square error stabilizes, achieving the desired training effect. In order to verify the effectiveness of the method presented in this paper for the classification and recognition of gas pipeline flow noise signals, it is necessary to test the trained network model by inputting two types of sound signal test samples into the trained gas sound signal BP network model for verification. According to the test results shown in the table below, the trained network model has an accuracy rate of up to 97% for classifying and recognizing flow noise signals with a strong noise background.

In order to show the performance of the feature extraction and pattern recognition model used in this study more intuitively, the following data charts and analysis are specially designed to show the performance of different feature extraction methods and their combination with BPNN in classifying gas flow noise signals. Firstly, the following experimental conditions are set: Based on the same data set, the performances of four combined methods, PCA + BPNN, Wavelet Transform (WT) + BPNN, Discrete Fourier Transform (DFT) + BPNN and Fast Fourier Transform (FFT) + BPNN, in feature extraction and pattern recognition are compared. Each method is cross-validated to ensure the reliability of the results. The ability of different feature extraction methods to reduce the data dimension while retaining important signal information is shown in Fig. 21. Figure 21 evaluates the efficiency of various methods by calculating the information loss rate between the original signal and the signal after feature extraction. The data shows that PCA has the best performance in information retention and the lowest loss rate.

Table 8 shows the performance of the above methods in classification accuracy, recall and F1 score after combining with BPNN. The results show that PCA + BPNN combination is superior to other methods in all evaluation indexes, and has the highest classification accuracy and efficiency. The analysis results show that PCA can not only reduce the data dimension more effectively, but also retain the key information better when processing the actual gas flow noise signal. When combined with BPNN, PCA + BPNN has achieved better results in accuracy and recall than other methods. For example, when dealing with complex signals with background noise, PCA + BPNN obviously reduces the misclassification, which proves its effectiveness in identifying flowing noise in strong noise environment. In addition, F1 score, as the harmonic average of accuracy and recall, also shows the advantages of PCA + BPNN in comprehensive evaluation. Through the above data and analysis, it is considered that PCA + BPNN is an efficient and reliable feature extraction and pattern recognition method, which is suitable for identifying gas flow noise signals under strong noise background.

In the experimental results and discussion part, this study makes a more in-depth analysis of the model recognition effect. By comparing the experimental data with the simulation data, it is found that the pattern recognition model based on PCA combined with BPNN achieves 97% recognition accuracy, 96% recall and 96.5% F1 score in the complex noise background. This result is significantly higher than that of traditional pattern recognition methods such as wavelet transform (WT), discrete Fourier transform (DFT) and fast Fourier transform (FFT) combined with BPNN. This shows that the model has good robustness and efficiency in dealing with flowing noise signals with strong noise. Secondly, in order to further optimize the performance of the model, the network structure, learning rate, weight initialization and other parameters of BPNN are carefully adjusted. In particular, by cross-checking the number of different hidden layer nodes, 10 nodes are determined as the best settings. In the aspect of weight initialization, a small range random number initialization method is adopted to reduce the influence of initial weight on model performance. In the choice of learning rate, several groups of experiments are carried out between 0.001 and 0.1, and finally 0.01 is determined as the optimal learning rate, which not only ensures the stability of learning, but also avoids the problem of slow convergence. Finally, the results of this study are compared with the existing research results. For example, Rajasekaran & Kothandaraman (2024) used a similar BPNN model to classify the leakage noise of urban gas pipelines, but its highest accuracy was only 92%, and the performance of the model was better than their research. This may be attributed to the adoption of PCA technology in the feature extraction stage, which effectively reduces the dimension of input features, thus reducing the computational burden of the model and improving the performance of the classifier. Through the above analysis, it is considered that the method proposed in this study can effectively improve the accuracy and efficiency of identifying gas flow noise signals in strong noise environment, and has important practical value for urban gas pipeline safety monitoring.