A two-stage method for identifying disturbances along a gas pipeline

Abstract

The present application belongs to the field of deep learning and security, and particularly relates to a two-stage interference identification method along an oil and gas pipeline, comprising: collecting original signals of distributed optical fibers of an optical fiber pipeline; segmenting the signals after noise reduction to obtain a large number of signal samples; inputting all the signal samples into a pre-classification model for pre-classification, taking the signal samples with pre-classification results as interference events as a data set of an event identification main algorithm; extracting feature vectors of all signals in the data set; identifying the intrusion signals of the distributed optical fiber pipeline through a cascade forest model, and outputting the results as event types. The present application increases a pre-selection link to effectively reduce the possibility that a non-interference state is misidentified as an interference event, and the main algorithm part focuses on enhancing the ability to distinguish ordinary interference and third-party destruction events threatening the safety of the pipeline, both of which make the system more accurately identify which events are threats requiring alarm and which actions do not require alarm.

Description

Technical Field

[0001] The present invention belongs to the field of deep learning and security, and specifically relates to a two-stage method for identifying interference along oil and gas pipelines. Background Art

[0002] In recent years, distributed optical fiber vibration sensing (DOVS) technology has attracted widespread attention in fields such as intelligent security due to its high sensitivity, immunity to electromagnetic interference, and low cost. It has been applied to perimeter security, oil and gas pipeline safety warnings, and structural health monitoring, particularly in protecting long-distance pipelines.

[0003] However, optical fiber sensing is susceptible to environmental influences such as wind and rain, pedestrian activity, and animal activity, so these harmless events can cause unexpected false alarms in the system. Furthermore, the complexity and similarity of vibration signals can lead to errors in vibration type identification. Time delay has also become a problem plaguing this sensing technology, with intrusion identification times exceeding seven seconds, making it difficult to respond to emergencies. Therefore, developing reliable, real-time pattern recognition methods to identify these harmless vibration events, reduce false alarm rates, and improve recognition accuracy is of great practical significance and guiding significance for improving the practical application capabilities of distributed optical fiber sensing and better ensuring the safe operation of natural gas pipelines.

[0004] Vibration signal type recognition relies on a well-performing classifier. This classifier is responsible for inputting the features of various intrusion signals into a classification model for training. The classification model is then used to quickly identify the vibration signal type. This requires the model to possess strong sample learning capabilities to achieve high-precision classification. Currently, the distributed fiber field primarily uses one-dimensional time series data as raw samples for signal feature extraction. Classifiers primarily include machine learning-based and deep learning-based model classification algorithms. Machine learning-based models include support vector machines (SVMs), relevance vector machines (RVMs), linear discriminant analysis (LDA), Gaussian mixture models (GMMs), and random forests (RFs). While machine learning methods such as support vector machines can achieve good classification results with small sample sizes, they have significant limitations for multi-classification tasks. Furthermore, some methods, such as GMMs, rely heavily on the selection of initial values ​​for classification performance. Deep learning-based algorithms and models, such as convolutional neural networks (CNNs), long short-term memory networks (LSTMs), and echo state networks (ESNs), have also been applied. Because deep learning performs feature extraction in an automated manner, it can learn more useful features by building models with multiple hidden layers and extensive training data, thereby improving classification accuracy. A high-precision method based on the combination of empirical mode decomposition (EMD) and radial basis function (RBF) neural networks has been proposed. This method uses the energy ratio of the intrinsic mode functions (IMFs) obtained from intrusion signals through EMD decomposition as the basis for classification. However, the EMD method can cause modal mixing problems in discontinuous signals, seriously affecting accuracy.

[0005] In summary, the existing technology has the following problems:

[0006] 1) Low real-time performance. Most existing research requires the collection of large amounts of historical data for training, which results in significant delays in actual application and makes it difficult to ensure the real-time capabilities required for field operations.

[0007] 2) Lack of regional considerations. The environment surrounding long-distance, cross-regional oil and gas pipelines is highly complex, with diverse features such as mountains, farmland, rivers, and roads. Sensor signals vary significantly depending on the geological characteristics, but most current research has not addressed this issue.

[0008] 3) Lack of applicability: Most research remains in the laboratory and has not been deployed in actual long-distance transportation pipeline networks, so it cannot meet actual needs. Summary of the Invention

[0009] To overcome the shortcomings of existing technologies, the present invention provides a two-stage method for identifying interference along oil and gas pipelines. This method first uses a wavelet threshold method to denoise the raw intrusion signal. It then pre-classifies the denoised signal using a dual sliding window algorithm. Finally, a cascade forest algorithm learns the basic features of the images corresponding to different intrusion signals to identify different types of intrusion events. This method is used to accurately identify which actions along the pipeline require alarms and which do not during long-distance pipeline safety monitoring.

[0010] The purpose of the present invention is achieved through the following technical solutions:

[0011] A two-stage method for identifying interference along oil and gas pipelines includes:

[0012] Step 1: collecting the original signal of the distributed optical fiber of the optical fiber pipeline in time series and performing noise reduction processing;

[0013] Step 2: Segment the denoised signal to obtain a large number of signal samples. If the window length is w l , step length is s l , then the signal of length l can be divided into n samples;

[0014]

[0015] Step 3: Input all signal samples into the pre-classification model for pre-classification. The pre-classification types are divided into interference events and normal events. The signal samples with the pre-classification result of interference events are used as the data set of the main event recognition algorithm.

[0016] Step 4, extracting feature vectors from the signals filtered out in step 3;

[0017] Step 5: The main event recognition algorithm adopts the cascade forest model to identify the distributed optical fiber pipeline intrusion signal, and the output result is the event type.

[0018] Furthermore, step 1 uses variational mode decomposition to reduce noise on the signal. Through variational mode decomposition, a given signal can be decomposed into K modes. The original signal is decomposed into several intrinsic mode functions through variational mode decomposition, and then the high-frequency modes are reconstructed to obtain a clean signal.

[0019] Furthermore, the variational mode decomposition used in step 1 is:

[0020]

[0021] Where f represents the original signal to be processed, t represents time, * is the convolution operator, j is the complex number symbol, represents the differential, δ(t) is the Dirac function, μ k represents the kth modal component (IMF) obtained by decomposition, ω k Represents the center frequency corresponding to each mode. The spectrum of each mode is modulated to the corresponding baseband through Hilbert transform and expressed as

[0022] Furthermore, in step 3, a double sliding window algorithm is used as a pre-classification model.

[0023] Furthermore, in step 3, the interference event classification method is: use the fluctuation amplitude to determine whether the disturbance event has occurred, let X represent the size of the first-layer sliding window, recorded as the small window, let Y represent the size of the second-layer sliding window, recorded as the large window, if more than N points in the small window have amplitudes exceeding the set threshold, then the small window is marked as an abnormal window; if more than M abnormal windows are found in the large window, the signal segment is marked as a disturbance event.

[0024] Furthermore, a dynamic threshold setting method is used in the dual sliding window algorithm. The threshold is adjusted between different defense sectors according to the statistical properties of the signal. The threshold setting includes two steps:

[0025] First, the average value of the waveform under normal events is set as the initial threshold of each defense sector;

[0026] Then, the threshold is adjusted by combining the double sliding window algorithm.

[0027] Furthermore, the time domain features of the waveform extracted in step 4 include multiple ones of maximum value, minimum value, peak-to-peak value, average value, absolute average value, root mean square, variance, standard deviation, energy, peak factor, skewness factor, gap factor, shape factor, pulse factor, and margin factor.

[0028] Furthermore, the cascade forest model in step 5 uses a cascade structure to sequentially refine the model's predictions. At each level of the cascade, a set of weak classifiers are trained on a subset of the data. Each classifier generates an estimate of the class distribution by calculating the percentage of different classes at the leaf node where the relevant instance is located, and then averaging the values ​​across all trees in the same forest. Each level of the cascade receives feature information processed by the previous level and outputs its processing results to the next level. The weak classifiers are combined to form a strong classifier, which is then used to classify the remaining data.

[0029] Furthermore, the event types in step 5 include multiple types of highway disturbance, railway disturbance, mountain and forest disturbance, and excavation damage.

[0030] Compared with the prior art, the present invention has the following beneficial effects:

[0031] 1) Reduced false alarm rate: The addition of a pre-selection step effectively reduces the possibility of non-interference states being mistakenly identified as interference events. The main algorithm focuses on strengthening the ability to distinguish between ordinary interference and third-party sabotage events that threaten pipeline safety. Both enable the system to more accurately identify which events are threats that require alarms and which actions do not require alarms;

[0032] 2) Improve real-time performance: Perform a pre-selection step before the main algorithm to reduce the number of samples that the main algorithm needs to classify, which can significantly improve computational efficiency. BRIEF DESCRIPTION OF THE DRAWINGS

[0033] Figure 1 Flowchart of the present invention;

[0034] Figure 2 Schematic diagram of the pre-classification model structure of the present invention;

[0035] Figure 3 This is a schematic diagram of the main algorithm structure of event recognition of the present invention. DETAILED DESCRIPTION

[0036] The present invention will be further described below with reference to the accompanying drawings.

[0037] Oil and gas pipelines traverse complex geological environments and infrastructure. For example, pipelines may traverse diverse environments, including densely populated towns, rural areas, farmland, rivers, roads, and railways. Vibration levels vary significantly along these geographically diverse routes. For example, the waveform intensity defining a damaging event differs significantly between areas along disruptive highways and railways and those in relatively secluded farmland and mountainous areas. Accurately identifying which actions along the pipeline require alerts and which do not during long-distance pipeline safety monitoring has become a key task for natural gas pipeline optical cable early warning and leak detection systems.

[0038] The present invention aims to provide an effective two-stage strategy to identify third-party damage along the entire route, while identifying different environments such as mountains, forests, roads, and railways. During the long-distance pipeline safety monitoring process, it can accurately identify which events are threats that require alarms and which actions do not require alarms. The first stage aims to identify and delete normal operating conditions without interference from the input data set to ensure that only interference events can be used as input to the main algorithm in the second stage. This is a pre-selection step to reduce the number of samples that the main algorithm needs to classify, which can significantly improve computational efficiency; in addition, it can filter out non-interference fragments to help improve the accuracy of interference identification and reduce the false alarm rate. The second stage uses a machine learning algorithm to learn the basic features of images corresponding to different intrusion signals to realize the identification of different types of intrusion events.

[0039] See also Figure 1-Figure 3A two-stage method for identifying interference along an oil and gas pipeline comprises the following steps:

[0040] Step 1: collect the original signal of the distributed optical fiber of the optical fiber pipeline according to the time series and perform noise reduction processing.

[0041] Among them, the present invention first performs noise reduction processing on all signals, and variational mode decomposition (VMD) is used here. Variational mode decomposition is a variational model that adaptively determines the relevant frequency bands and simultaneously estimates the corresponding modes, which can solve the noise problem existing in the input signal. The model has strong interpretability and is supported by powerful mathematical theory. The most important advantage of this method is that it avoids modal aliasing. Through this method, a given signal can be decomposed into K modes. The present invention decomposes the original signal into 8 intrinsic mode functions (IMFs) through VMD, and then reconstructs the high-frequency modes to obtain a clean signal. This method can improve the noise reduction performance without causing signal distortion.

[0042] The essence of variational mode decomposition is to solve a constrained variational problem:

[0043]

[0044] Where μ k represents the kth IMF obtained by decomposition, ω k Represents the center frequency corresponding to each mode, and the spectrum of each mode modulated to the corresponding baseband is expressed as

[0045] Step 2: Segment the noise-reduced signal to obtain a large number of signal samples with a time span of 30 seconds.

[0046] Among them, if the window length is w l , step length is s l , then the signal of length l can be divided into n samples;

[0047]

[0048] Step 3: Input all signal samples into the pre-classification model for pre-classification, and use the signal samples with the pre-classification results as interference events as the data set of the main algorithm.

[0049] Step 4: Extract feature vectors from the signals selected in step 3.

[0050] Step 5: Use the cascade forest model to identify distributed fiber optic pipeline intrusion signals. The output result is the event type, which is divided into four categories: highway disturbance, railway disturbance, mountain and forest disturbance, and excavation damage.

[0051] In the natural gas pipeline optical cable early warning leak detection system, pattern recognition is one of the core technologies, and the extraction of the detection signal feature vector is one of the most critical links in the pattern recognition module. The feature vector of the optical fiber disturbance signal must meet the following three conditions:

[0052] Features must be unique. Different types of disturbance signals should have their own unique characteristic attributes and overlap with the features of other types of disturbances.

[0053] Features must be stable and will not change with the passage of time, the number of samples, the external environment, and other factors. That is, the inherent characteristics of the signal always exist and remain stable.

[0054] Features must be universal, meaning that the same features can distinguish different types of events. In particular, in distributed fiber-optic early warning systems, feature data can be quantified, facilitating quantitative processing.

[0055] The time-domain characteristics of a waveform are how the distributed optical fiber waveform changes over time. Some time-domain characteristics, such as the waveform's maximum and minimum values, can be perceived visually. Other characteristics require computation, such as the waveform's average value, variance, and short-term energy over a period of time. Waveform vibrations caused by different factors often exhibit differences in their associated time-domain characteristics. As shown in Table 1, the waveform time-domain characteristics extracted by this invention specifically include the following 16: maximum value, minimum value, peak-to-peak value, average value, absolute average value, root mean square (RMS), variance, standard deviation, energy, crest factor, skewness factor, gap factor, shape factor, pulse factor, and margin factor.

[0056] Table 1 Time domain features that can be extracted from distributed fiber waveforms

[0057]

[0058]

[0059] In practical applications, approximately 95% of real-time signals indicate that the device or system is operating normally without any disturbance events, while only 5% indicate that a disturbance event has occurred. If all signals were fed into a machine learning classification model, resources would be occupied, resulting in low timeliness. To improve accuracy and computational efficiency, this invention uses a double sliding window (TDSW) algorithm as the first stage of pre-classification, distinguishing between intrusion events and normal operation before identifying specific categories.

[0060] The waveform in normal working state is relatively smooth, but the waveform of a disturbance event fluctuates greatly. Therefore, the amplitude of the fluctuation can be used to determine whether a disturbance event has occurred. Let X represent the size of the first sliding window, recorded as the small window, and let Y (as long as the sample produced by the present invention) represent the size of the second sliding window, recorded as the large window. If more than N points in the small window have amplitudes exceeding the set threshold, the small window is marked as an abnormal window; if more than M abnormal windows are found in the large window, the signal segment is marked as a disturbance event.

[0061] The amplitude threshold in this algorithm uses a dynamic thresholding approach, whereby the threshold is adjusted across different defense sectors based on the statistical properties of the signal. This allows for more adaptive and robust screening. Threshold setting requires two steps: first, the average value of the waveform under normal operating conditions is set as the initial threshold for each defense sector; then, the threshold is slightly adjusted using a dual sliding window.

[0062] For the second-stage main algorithm for event recognition, the present invention uses a cascade forest-based model to perform signal recognition and classification. The cascade forest model can be seen as a deep collection of random forest algorithms, which uses a cascade structure to sequentially refine the model's predictions. At each level of the cascade, a collection of weak classifiers is trained on a subset of the data. Each classifier generates an estimate of the class distribution by calculating the percentage of different types at the leaf nodes where the relevant instances are located, and then averaging all the trees in the same forest. Each level of the cascade receives the feature information processed by its previous level and outputs its processing results to the next level. The weak classifiers are combined to form a strong classifier, which is then used to classify the remaining data.

[0063] Using cascade forests requires only limited training time and has good pattern matching performance. These advantages are fully consistent with the original intention of designing a fast, efficient, and accurate strategy. The key idea behind the cascade forest algorithm is to use a cascade structure to increase model complexity, which is why it can handle data with a large number of features or variables and capture their nonlinear relationships. In addition, it also provides a simple mechanism by which the number of cascade levels can be adaptively determined so that the model complexity can be automatically set. At the same time, it is also highly robust and can handle noisy data or outlying data points.

[0064] In summary, the two-stage interference identification strategy proposed in this invention not only achieves the basic function of identifying various types of interference events, but also effectively reduces the possibility of non-interference states being mistakenly identified as interference events, and strengthens the ability to distinguish interference events that do not threaten pipeline safety from third-party sabotage events that may cause damage. This invention proposes a dual sliding window algorithm, which aims to perform a preselection step before the main algorithm to reduce the number of samples that the main algorithm needs to classify, significantly improving computational efficiency. Furthermore, it can filter out non-interference segments to help improve the accuracy of interference identification and reduce the false alarm rate.

[0065] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the above embodiments, or replace some or all of the technical features therein with equivalents. However, these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A two-stage method for identifying interference along oil and gas pipelines, characterized in that: include: Step 1: collecting the original signal of the distributed optical fiber of the optical fiber pipeline in time series and performing noise reduction processing; Step 2: Segment the denoised signal to obtain a large number of signal samples. If the window length is w1 and the step length is s1, a signal of length l can be segmented into n samples: ； Step 3: Use the double sliding window algorithm as the pre-classification model, input all signal samples into the pre-classification model for pre-classification, and classify them into interference events and normal events. The signal samples with the pre-classification result of interference events are used as the data set of the main event recognition algorithm; The classification method for interference events is as follows: the fluctuation amplitude is used to determine whether a disturbance event has occurred. Let X represent the size of the first sliding window, recorded as the small window, and let Y represent the size of the second sliding window, recorded as the large window. If more than N points in the small window have amplitudes exceeding the set threshold, the small window is marked as an abnormal window; if more than M abnormal windows are found in the large window, the signal segment is marked as a disturbance event. The dual sliding window algorithm uses a dynamic threshold setting method. The threshold is adjusted between different defense sectors based on the statistical properties of the signal. The threshold setting includes two steps: First, the average value of the waveform under normal events is set as the initial threshold of each defense sector; Then, the threshold is adjusted by combining the double sliding window algorithm; Step 4, extracting feature vectors from the signals filtered out in step 3; Step 5: The main event recognition algorithm adopts the cascade forest model to identify the distributed optical fiber pipeline intrusion signal, and the output result is the event type.

2. A two-stage method for identifying interference along an oil and gas pipeline according to claim 1, characterized in that: The step 1 uses variational mode decomposition to reduce the noise of the signal. Through variational mode decomposition, the given signal can be decomposed into K modes. The original signal is decomposed into several intrinsic mode functions through variational mode decomposition, and then the high-frequency modes are reconstructed to obtain a clean signal.

3. A two-stage method for identifying interference along an oil and gas pipeline according to claim 1, characterized in that: The core idea of ​​the variational mode decomposition used in step 1 is to construct and solve the variational problem: ； In the formula, f represents the original signal to be processed, t represents the time, is the convolution operator, j is a complex number symbol, represents the differential, δ(t) is the Dirac function, μ k represents the kth IMF modal component obtained by decomposition, ω k Represents the center frequency corresponding to each mode. The spectrum of each mode is modulated to the corresponding baseband through Hilbert transform and expressed as .

4. A two-stage method for identifying interference along an oil and gas pipeline according to claim 1, characterized in that: The time domain features of the waveform extracted in step 4 include multiple ones of maximum value, minimum value, peak-to-peak value, average value, absolute average value, root mean square, variance, standard deviation, energy, peak factor, skewness factor, gap factor, shape factor, pulse factor, and margin factor.

5. A two-stage method for identifying interference along an oil and gas pipeline according to claim 1, characterized in that: In step 5, the cascade forest model uses a cascade structure to sequentially refine the model's predictions. At each level of the cascade, a set of weak classifiers are trained on a subset of the data. Each classifier generates an estimate of the class distribution by calculating the percentage of different classes at the leaf node where the relevant instance is located, and then averaging the values ​​across all trees in the same forest. Each level of the cascade receives feature information processed by the previous level and outputs its processing results to the next level. The weak classifiers are combined to form a strong classifier, which is then used to classify the remaining data.

6. A two-stage method for identifying interference along an oil and gas pipeline according to claim 1, characterized in that: The event types in step 5 include multiple types of highway disturbance, railway disturbance, mountain and forest disturbance, and excavation damage.

Images