CN-122024763-A - Pipeline leakage identification method and system based on acoustic emission feature fusion

Abstract

The invention provides a pipeline leakage identification method and system based on acoustic emission feature fusion. According to the method, one-dimensional time domain features and two-dimensional frequency spectrum features of the 15 ml/s micro-leakage acoustic emission signals are aligned in time, feature fusion is completed along the channel dimension, and acoustic emission self-fusion features with stronger information characterization capability are constructed. Training and identifying different noise backgrounds, different pipeline pressures and different leakage defect working conditions through ViT-transducer models. The result shows that the method can stably distinguish the non-leakage state from the tiny leakage state under different pipeline pressures, different noise environments and different leakage defect conditions. Under the low pressure noise condition, the system identification accuracy is kept at 87%, under the medium noise condition, the average identification accuracy can reach about 92%, and when the pipeline pressure is increased and the noise interference is weakened, the identification accuracy is further increased to about 98%, so that good stability and robustness are achieved.

Inventors

• XU CHANGHANG
• Wang Wenao
• Shang Jiateng
• DU YU
• DU ZHICHENG
• ZHANG YUBIN
• LI NA

Assignees

• 中国石油大学(华东)

Dates

Publication Date

20260512

Application Date

20260204

Claims (2)

1. 1. A pipeline leakage identification method based on acoustic emission feature fusion comprises the following steps: (1) Performing acoustic emission signal micro-leakage tests under different working conditions, and collecting one-dimensional time domain characteristics and two-dimensional frequency spectrum characteristics of leakage signals under the micro-leakage state to form an acoustic emission signal micro-leakage database; (2) Confirming the time domain representation of the one-dimensional time domain feature and the two-dimensional frequency spectrum feature, wherein the one-dimensional time domain feature comprises the information of impact energy, envelope form, duration, waveform rising and attenuation modes of the acoustic emission signal, and the discrete time sequence can be represented as follows: ; ; Wherein, the Representing discrete acoustic emission signal amplitudes; Representing a discrete time index; Representing the total amount of overall sampling; is the natural time scale of the time domain signal, Representing the sampling frequency; The two-dimensional frequency spectrum features comprise frequency components of acoustic emission signals, feature frequency bands and local energy aggregation information, and a frequency spectrum matrix is obtained by short-time Fourier transform: ; Wherein, the Indicating that the signal is at the first Over a time window, frequency bin is Complex spectral values at the time, including amplitude and phase information of the frequency at the time location; a discrete index representing the frequency component; Representing a sequence of magnitudes of the acoustic emission time domain signal; Representing a discrete time index; corresponding to the first time series A short time analysis window; The length of the sample is analyzed each time in the short-time Fourier calculation process; Representing a displacement step between two adjacent time frames; As a window function used in the short-time fourier transform, Is a kernel function used in a short-time fourier transform; Representing the length of a frequency axis obtained by short-time Fourier transform; representing the length of a time axis obtained by short-time Fourier transform; (3) On the premise that the one-dimensional time domain features and the two-dimensional frequency spectrum features have the same time scale, confirming master-slave positions of the one-dimensional time domain features and the two-dimensional frequency spectrum features in fusion, and constructing dominant feature tensors, taking the two-dimensional frequency spectrum features as dominant positions in the fusion process, taking the one-dimensional time domain features as passive adjustment positions, namely ensuring that a matrix structure of the two-dimensional frequency spectrum features is kept unchanged on the basis of facilitating forward propagation, and realizing the fusion by adjusting the matrix structure of the one-dimensional time domain features, wherein the method comprises the following specific implementation steps of: the linear interpolation downsampling method is used for carrying out the length transformation of the frequency spectrum characteristics on the time axis on the premise of not losing complete information of the time scale, and the frequency spectrum matrix is scaled to be convenient for the subsequent input of a transducer: ; Wherein, the Representation mapping to uniform length on a timeline The time index of the latter time-frequency matrix is marked as ; Is a time interpolation kernel for slave To the point of Is mapped by interpolation of (a); by adjusting the dimension structure, the spectrum characteristic is established as the dominant position in the fusion process, and the spectrum characteristic is obtained by the above formula Input 2D-CNN network: ; Wherein, the The extracted spectral feature tensor for 2D-CNN has the shape: ; Outputting the channel number for 2D-CNN; the frequency dimension is reserved for the spectrum characteristic matrix after 2D-CNN processing; the time dimension is unified; (4) At the dominant feature tensor After the structure of the signal is determined, a time domain feature tensor matched with the signal is required to be constructed so as to carry out subsequent fusion, and the method comprises the specific operations of inputting an original one-dimensional time domain signal into a 1D-CNN, extracting key time domain feature information in the signal, and ensuring that a one-dimensional feature matrix is completely consistent with two-dimensional spectrum features in a time dimension based on a time alignment principle: ; ; Wherein, the The time domain feature tensor extracted for the 1D-CNN is shaped as: ; The number of output channels is 1D-CNN; representing the presentation to be Mapping to a time length The shape of the post time domain feature tensor is: ; Mapping kernel for time direction To the point of Is mapped to; (5) In order to enable the one-dimensional feature to be integrated into the structure of the two-dimensional feature, the one-dimensional feature is structurally copied and expanded along the frequency dimension direction of the two-dimensional feature: , ; Wherein, the The time domain characteristic tensor after the expansion of the frequency dimension direction is represented, and the shape is as follows: the method enables the one-dimensional feature to be consistent with the two-dimensional feature in tensor structure on the premise of not adding useless information and not changing the physical meaning of the one-dimensional feature; (6) After the construction of the one-dimensional time domain feature tensor and the two-dimensional frequency spectrum feature tensor is completed, multidimensional feature fusion is carried out, the two-dimensional frequency spectrum feature tensor is taken as a leading frame, and the one-dimensional time domain feature tensor is embedded into a tensor space where the two-dimensional frequency spectrum features are located along a channel dimension on the basis of keeping the physical meaning of the two-dimensional frequency spectrum feature tensor stable and the time dimension unchanged: ; Wherein, the Representing the fused feature tensor, the shape is: ; representing a fusion operation per channel dimension; (7) Fusion features to be constructed The method is input into a Vit-transducer algorithm model to carry out deep feature learning and modeling, and the modeling capability of a self-attention mechanism on the global feature relation is fully utilized to realize automatic mining and expression of key information in the acoustic emission self-fusion feature.
2. 2. A pipeline leak identification system based on acoustic emission feature fusion, comprising: The pipeline tiny leakage simulation module comprises a pressure air source, a pressure regulating valve, a pressure stabilizing manifold, a flowmeter, a control valve, a replaceable leakage orifice plate and a pipeline and valve connecting assembly; the pipeline is used for introducing the pressure air source; The pressure regulating valve, the pressure stabilizing manifold, the control valve, the replaceable leakage orifice plate, the flowmeter and the valve connecting assembly are arranged on the pipeline and used for stabilizing the pressure parameter and the leakage structural form of the pressure air source; The acoustic emission signal acquisition module comprises an acoustic emission sensor, a pre-amplifier and an acquisition processing system, wherein the acoustic emission sensor is arranged at a key position of a pipeline and is used for acquiring a high-frequency acoustic emission signal generated when a pipeline is subjected to micro leakage in real time, the pre-amplifier is used for carrying out pre-amplification and filtering processing on the acquired weak signal, and the acquisition processing system is used for carrying out high-speed sampling and transmitting the acquired weak signal to an upper computer so as to realize stable acquisition and digital storage of the acoustic emission signal in the leakage process and provide an original data basis for subsequent feature extraction and analysis; the long-distance pipeline noise adding module is used for simulating the energy attenuation and environmental noise interference influence experienced by the acoustic emission signal in the long-distance pipeline propagation process, the module is used for carrying out noise superposition processing on the acquired original acoustic emission signal based on an additive Gaussian white noise model, and constructing different noise working conditions by setting signal-to-noise ratio parameters so as to equivalently represent the weak signal characteristics under the long-distance condition, and acoustic emission data samples with multiple noise levels can be generated through the module to provide data support for robustness verification of a subsequent algorithm under the complex working conditions; The self-feature fusion processing and leakage identifying module comprises a personal computer, a one-dimensional time domain signal processing program, a two-dimensional frequency spectrum generating and analyzing program, a self-feature fusion algorithm and a leakage identifying model based on a Vit-transducer, wherein the module respectively performs time domain feature extraction and time-frequency transformation on the acquired acoustic emission signals subjected to noise simulation to construct one-dimensional time domain features and two-dimensional frequency spectrum features, maps the one-dimensional features to feature spaces consistent with the two-dimensional frequency spectrum features in a time domain alignment and frequency expansion mode, fuses the features along the channel direction to form acoustic emission self-fusion feature tensor, and finally inputs the fused features into the Vit-transducer model to perform depth feature learning and classification discrimination so as to realize automatic identification and real-time judgment on different leakage defect states of long-distance pipelines.

Description

Pipeline leakage identification method and system based on acoustic emission feature fusion Technical Field The invention belongs to the field of pipeline leakage acoustic emission detection and identification, and particularly relates to a pipeline leakage identification method and system based on acoustic emission feature fusion. Background Pipes play an indispensable role in industrial production and in people's daily life. In industrial production, pipelines are a key component of mass transport and communication networks, and are widely used for the transport of liquids, gases and particulate materials. Through arranging the pipeline system scientifically and reasonably, high-efficiency and rapid transmission among different positions can be realized, so that the production efficiency is remarkably improved. Especially industrial liquid pipelines often carry the task of transporting high temperature, high pressure, flammable, explosive and highly corrosive materials. However, the pipeline generally forms a huge and complex network system, and involves a plurality of branches, junctions, pump stations, valves and other key components, and in the long-term operation process, the pipeline may be damaged or leaked due to various factors such as aging, mechanical property change, pipe wall corrosion, welding defects or artificial damage. In order to ensure the safe operation of the pipeline, the maintenance environment and personnel safety, an efficient and reliable leakage detection means is required to be adopted for autonomous and real-time monitoring of the pipeline, so that quick identification and early warning are realized at the early stage of leakage, and potential loss is reduced to the greatest extent. The existing leakage detection technology mainly comprises acoustic emission detection, optical fiber detection, ultrasonic detection, infrared detection and the like. The essence is that the leakage identification is finally realized by capturing the fine physical signals generated by the leakage event and analyzing and calculating the signals. The method can ensure the smooth operation of industrial production to a certain extent, and can realize effective judgment on most conventional leakage events. However, existing detection methods still have significant limitations with respect to small leaks that occur in long distance pipelines. The signal strength generated by the tiny leakage is weak, the signal is easy to attenuate rapidly in the propagation process in the pipeline, and meanwhile, the signal is greatly influenced by the interference of environmental noise, so that the characteristics of a single signal are not obvious, and the signals are difficult to reliably capture in a long distance by the traditional technology. The response of the existing method to leakage usually depends on obvious change of signal characteristics, and the early-stage signal of the tiny leakage is weak and the change is not obvious, so that the difficulty of early detection and accurate positioning is further increased. Aiming at the challenge, the invention provides a pipeline tiny leakage identification method and system based on acoustic emission signal self-feature fusion. According to the method, the one-dimensional time domain features and the two-dimensional frequency spectrum features of the acoustic emission signals are fused to form the composite features with larger information quantity and less interference, so that high-precision detection and identification of the tiny leakage are realized. The system comprises a manifold leakage module, an acoustic emission signal acquisition module, a self-feature fusion processing module and a leakage detection and identification module. According to the method, one-dimensional time domain features are stretched and expanded along the frequency direction through time domain alignment, so that the one-dimensional time domain features are consistent with two-dimensional frequency spectrum features, and then fusion is carried out along the channel direction, so that brand new acoustic emission self-fusion features are obtained. The innovative processing mode can fully utilize the multidimensional information of the signals, improve the identification capability of tiny leakage and provide reliable technical support for the safe operation of the pipeline. Disclosure of Invention Aiming at the problem that the existing methane micro-leakage detection means can only acquire a small amount of information with single characteristics, the invention provides a remote positioning technical scheme and a remote positioning system based on infrared and ultrasonic signal characteristic fusion. The global characteristic information containing one-dimensional and two-dimensional signals is obtained by combining the characteristic information obtained by the infrared ultrasonic detection means, so that the accurate positioning of the remote methane micro leakage is realized. In