High-consequence area oil and gas pipeline vibration monitoring and early warning method and system based on time sequence multi-axis vibration sensing equipment
By deploying time-series multi-axis vibration sensing devices on oil and gas pipelines, performing P/S wave separation and deep learning model training, the problems of response lag and positioning accuracy in monitoring oil and gas pipelines in high-consequence areas were solved, achieving high-precision and low-cost construction behavior identification and alarm.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- INNER MONGOLIA WESTERN NATURAL GAS CO LTD
- Filing Date
- 2025-12-26
- Publication Date
- 2026-04-14
AI Technical Summary
Existing technologies for monitoring oil and gas pipelines in high-consequence areas suffer from problems such as delayed response, high false alarm rate, and poor positioning accuracy. In particular, in complex soil environments, vibration signals attenuate rapidly and propagation paths are uncertain, making it impossible to effectively identify minor but dangerous construction activities.
By employing a time-series multi-axis vibration sensing device, triaxial vibration signals from multiple nodes are simultaneously acquired, P/S wave physical separation and feature extraction are performed, and a multi-task learning model of a deep time-series network is constructed to achieve construction behavior recognition and alarm.
It achieves high-precision positioning within a radius of 200-300 meters, reduces the false alarm rate by an order of magnitude, meets the requirements of second-level early warning and minute-level response, and can operate independently in cloudless and fiber-optic environments. It is suitable for high-humidity, high-salt, and dusty outdoor environments, reducing costs and improving deployment convenience.
Smart Images
Figure CN121854772A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of pipeline monitoring technology, specifically to a method and system for monitoring and early warning of vibration in high-consequence oil and gas pipelines based on time-series multi-axis vibration sensing equipment. Background Technology
[0002] As oil and gas pipeline networks continue to expand, the proportion of pipelines crossing high-consequence areas has increased significantly, and third-party construction damage has become one of the main causes of pipeline safety accidents. Traditional monitoring methods mainly rely on manual inspections or single vibration sensors, which suffer from problems such as response lag, high false alarm rates, and inability to accurately locate faults. For example, in areas with heavy traffic or active construction, environmental noise can easily trigger false alarms, while minor but dangerous digging activities may go undetected.
[0003] In recent years, although monitoring systems using distributed optical fiber (DAS) or single vibration nodes have been developed, they generally suffer from drawbacks such as high cost, complex deployment, weak signal recognition capabilities, and a lack of effective separation and utilization of P-waves and S-waves. Especially in complex soil environments, the vibration signal attenuates rapidly and the propagation path is uncertain, making it difficult to meet engineering requirements in terms of positioning accuracy.
[0004] Therefore, there is a need to provide a method and system for monitoring and early warning of vibration in high-consequence oil and gas pipelines based on time-series multi-axis vibration sensing devices to solve the above problems. Summary of the Invention
[0005] This invention provides a method and system for monitoring and early warning of vibration in high-consequence oil and gas pipelines based on a time-series multi-axis vibration sensing device, in order to solve existing problems.
[0006] The first aspect of this invention provides a method for monitoring and early warning of vibration in high-consequence oil and gas pipelines based on a time-series multi-axis vibration sensing device. This method employs the following technical solution, including: Simultaneously collect triaxial vibration signals corresponding to multiple different nodes along the high-consequence zone of the oil and gas pipeline, and obtain the construction behavior category corresponding to the triaxial vibration signal at each moment; The triaxial vibration signal is subjected to P / S wave physical separation processing to obtain P wave signal and S wave signal; the time domain features and frequency domain features corresponding to P wave signal and S wave signal are organized into multi-dimensional vectors corresponding to P wave signal and S wave signal according to axis and wave mode, and the multi-dimensional vectors are reduced to obtain lightweight feature vectors corresponding to P wave signal and S wave signal. A dataset is constructed based on the lightweight feature vectors corresponding to P-wave and S-wave signals, the triaxial vibration signal corresponding to each time moment, the construction behavior category label, the event hazard level label, and the P-wave signal arrival time label corresponding to each time moment. A multi-task learning model based on a deep temporal network is constructed, and the multi-task learning model is trained on a dataset to obtain a trained target multi-task learning model. The multi-task learning model includes an encoder and a multi-task output head. The encoder is used to encode the input feature vector into shared features. The multi-task output head includes a classification head for outputting the construction behavior category based on the shared features, a risk head for outputting the event hazard level based on the shared latent layer representation, and a regression head for outputting the arrival time of the P-wave signal based on the shared latent layer representation. The lightweight feature vectors corresponding to the triaxial vibration signal, P-wave signal and S-wave signal at the current moment are input into the target multi-task learning model to obtain the construction behavior category and confidence level, event hazard level and P-wave signal arrival time at the current moment. When the confidence level is greater than or equal to the preset alarm threshold, the alarm location is obtained based on the time difference of arrival of the P-wave signal of different nodes at the current time; the alarm level is determined according to the event hazard level, and the alarm location, the type of construction behavior at the alarm location, and the alarm level are output.
[0007] A further technical solution of the present invention is that the step of performing P / S wave physical separation processing on the triaxial vibration signal to obtain the P-wave signal and the S-wave signal is as follows: High-pass filtering or polynomial fitting methods are used to correct the triaxial vibration signal; The corrected triaxial vibration signal is denoised by wavelet thresholding to obtain the denoised triaxial vibration signal. The P / S wave physical separation of the denoised triaxial vibration signal is performed to obtain the P wave signal and the S wave signal.
[0008] A further technical solution of the present invention is that the step of reducing the dimensionality of a multidimensional vector to obtain the lightweight feature vectors corresponding to the P-wave signal and the S-wave signal is as follows: performing principal component analysis or linear discriminant analysis on the multidimensional vector to obtain the lightweight feature vectors corresponding to the P-wave signal and the S-wave signal.
[0009] A further technical solution of the present invention is that the encoder includes: Multi-scale 1D-CNN layers are used to extract multi-scale local temporal features of the input features; The channel attention layer is used to dynamically weight multi-scale local temporal features to obtain the first target feature; A bidirectional LSTM layer is used to add contextual features to the first target feature to form the second target feature; And a temporal self-attention layer, used to dynamically weight the second target features corresponding to the time step to obtain shared features.
[0010] A further technical solution of the present invention is that the loss function of the multi-task learning model is:
[0011] In the formula, This represents the loss function of a multi-task learning model. The loss function representing the classification head; The loss function representing the risk head; The loss function representing the regression head; Indicates the learning parameters of the classification head; The learning parameters represent the risk head; This represents the learning parameters of the regression head.
[0012] A further technical solution of the present invention is that the loss function of the classification head adopts the Focal Loss loss function; the loss function of the risk head adopts the ordinal regression loss function; and the loss function of the regression head adopts the Huber Loss loss function.
[0013] A further technical solution of the present invention is to solve the alarm location based on the time difference of arrival of P-wave signals at different nodes at the current time and using a least squares optimization algorithm.
[0014] A further technical solution of the present invention is to use the exponentially weighted moving average (EWMA) algorithm to obtain the alarm threshold at each time point.
[0015] A further technical solution of the present invention includes incremental fine-tuning of the weights of the target multi-task learning model, wherein the specific incremental fine-tuning steps are as follows: A low-rank adapter is introduced in at least one preset layer of the target multi-task learning model to incrementally update the original weight matrix; When the input feature vector is input into this preset layer, its output features are obtained by the combined action of the original weight matrix and the low-rank adapter. Based on newly collected data or manual verification results, while freezing the original weight matrix, only the parameters in the low-rank adapter are trained, thereby achieving lightweight incremental fine-tuning of the model.
[0016] A second aspect of the present invention provides a vibration monitoring and early warning system for high-consequence oil and gas pipelines based on a time-series multi-axis vibration sensing device, comprising: The signal acquisition module is used to synchronously acquire triaxial vibration signals corresponding to multiple different nodes along the high-consequence zone of the oil and gas pipeline, and obtain the construction behavior category corresponding to the triaxial vibration signal at each moment. The signal processing module is used to perform P / S wave physical separation processing on the triaxial vibration signal to obtain P wave signal and S wave signal; organize the time domain features and frequency domain features corresponding to the P wave signal and S wave signal into multi-dimensional vectors corresponding to the P wave signal and S wave signal according to the axis and wave mode; and reduce the dimensionality of the multi-dimensional vectors to obtain lightweight feature vectors corresponding to the P wave signal and S wave signal. The dataset construction module is used to construct a dataset based on the lightweight feature vectors corresponding to the triaxial vibration signal, P-wave signal and S-wave signal at each time point, as well as the construction behavior category label, event hazard level label and P-wave signal arrival time label at each time point. The neural network module is used to construct a multi-task learning model based on a deep temporal network, and to train the multi-task learning model on a dataset to obtain a trained target multi-task learning model. The multi-task learning model includes an encoder and a multi-task output head. The encoder is used to encode the input feature vector into shared features. The multi-task output head includes a classification head for outputting the construction behavior category based on the shared features, a risk head for outputting the event hazard level based on the shared latent layer representation, and a regression head for outputting the arrival time of the P-wave signal based on the shared latent layer representation. The alarm module is used to input the lightweight feature vectors corresponding to the triaxial vibration signal, P-wave signal, and S-wave signal at the current moment into the target multi-task learning model to obtain the construction behavior category and confidence level, event hazard level, and P-wave signal arrival time at the current moment. When the confidence level is greater than or equal to the preset alarm threshold, the alarm location is obtained based on the time difference of P-wave signal arrival at different nodes at the current moment; the alarm level is determined according to the event hazard level, and the alarm location, construction behavior category at the alarm location, and alarm level are output.
[0017] The beneficial effects of this invention are: 1. Intelligent recognition capability of construction behavior based on deep temporal network: This invention sets up a multi-task learning model of LSTM-CNN with channel-time dual attention mechanism, which takes the time sequence of multi-axis vibration signal features as input and identifies construction behaviors such as excavation, drilling, road rolling and knocking in an end-to-end manner.
[0018] 2. This invention achieves a horizontal positioning error of ≤30 meters within a monitoring radius of 200–300 meters through triaxial sensing, P / S wave physical separation, time synchronization, and TDOA least squares optimization, meeting the engineering accuracy requirements for spatial positioning in high-consequence areas. It constructs a three-dimensional criterion of "waveform, time series, and energy," combining wavelet denoising, kurtosis / RMS features, and deep learning discrimination to effectively distinguish between transient construction impacts and stable noises such as vehicles and wind and rain. The measured false alarm rate is ≤8%, reducing it by more than an order of magnitude compared to traditional methods. Furthermore, over 90% of the computation is completed at the edge, uploading only ≤32-dimensional lightweight feature vectors, resulting in an end-to-end alarm response time of ≤3 seconds, meeting the safety control requirements of "second-level early warning and minute-level response."
[0019] 3. The entire system (sensor + edge computing + server + platform) can operate independently in cloud-free and fiber-optic-free environments, complying with data security and on-site autonomy standards. It supports 72-hour caching and breakpoint resumption, ensuring continuous monitoring during network interruptions. The vibration sensing device integrates solar power and a 4G communication module, has an IP67 protection rating, and is suitable for high-humidity, high-salt, and dusty outdoor environments, significantly improving long-term deployment reliability and maintenance convenience. Based on a sliding time window, the alarm threshold is dynamically adjusted, automatically adapting to different conditions and interference source characteristics to achieve differentiated intelligent sensing tailored to each location, effectively solving regional false alarms or missed alarms. The vibration sensing device, combined with an edge computing module, can be deployed arbitrarily near the pipeline centerline, independent of network and power. Compared to distributed fiber optic systems, it eliminates the need for pipeline excavation, making it safer for pipeline networks, especially more convenient and flexible for upgrading existing networks. Maintenance only requires replacing faulty equipment, eliminating the need for complete replacement and secondary excavation, keeping costs under control. Attached Figure Description
[0020] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.
[0021] Figure 1 This is a flowchart illustrating a method for monitoring and early warning of vibration in high-consequence oil and gas pipelines based on a time-series multi-axis vibration sensing device, according to the present invention. Figure 2 This is a schematic diagram of the neural network structure in an embodiment of the present invention. Detailed Implementation
[0022] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.
[0023] An embodiment of the present invention provides a method and system for monitoring and early warning of vibration in high-consequence oil and gas pipelines based on a time-series multi-axis vibration sensor, as follows: Figure 1 and Figure 2 As shown, it includes: S1. Acquire triaxial vibration signals and corresponding construction behavior categories; Specifically, triaxial vibration signals are simultaneously collected from multiple different nodes along the high-consequence zone of the oil and gas pipeline, and the construction behavior category corresponding to the triaxial vibration signal at each moment is obtained.
[0024] For example, in one specific embodiment, time-series multi-axis vibration sensing devices are deployed at multiple different nodes in the high-consequence zone of the oil and gas pipeline along its length. Each time-series multi-axis vibration sensing device integrates a high-precision triaxial MEMS accelerometer. Microsecond-level time alignment is achieved through GPS or a wired synchronization mechanism to synchronously collect the triaxial vibration signal corresponding to each node. Based on the triaxial vibration signal at each moment, the corresponding construction behavior category label can be obtained. Specifically, in this embodiment, a monitoring node is set up every 100-200 meters along the oil and gas pipeline in the high-consequence zone. Each monitoring node is equipped with 4 sets of TR devices (12 in total), distributed in a convex quadrilateral shape to surround the potential construction area. The coordinates of each node are accurately measured by RTK and entered into the system.
[0025] S2. Obtain the lightweight feature vectors of the P-wave and S-wave signals corresponding to the triaxial vibration signals; Specifically, the triaxial vibration signal is subjected to P / S wave physical separation processing to obtain P-wave signal and S-wave signal; the time domain features and frequency domain features corresponding to P-wave signal and S-wave signal are organized into multi-dimensional vectors corresponding to P-wave signal and S-wave signal according to axis and wave type, and the multi-dimensional vectors are reduced to obtain lightweight feature vectors corresponding to P-wave signal and S-wave signal.
[0026] For example, in one specific embodiment, the steps of performing P / S wave physical separation processing on the triaxial vibration signal to obtain P-wave signal and S-wave signal are as follows: the triaxial vibration signal is corrected by using high-pass filtering or polynomial fitting method; wavelet threshold denoising is performed on the corrected triaxial vibration signal to obtain denoised triaxial vibration signal; and P / S wave physical separation is performed on the denoised triaxial vibration signal to obtain P-wave signal and S-wave signal.
[0027] In this embodiment, the triaxial vibration signal is a triaxial broadband vibration signal with X, Y, and Z axes. In this embodiment, the Z-axis signal dominates the P-wave (longitudinal wave), and the X or Y-axis signal dominates the S-wave (transverse wave). Combining the physical law that the P-wave arrives before the S-wave, the effective signal windows of the P-band and S-band are extracted to obtain the P-wave signal and the S-wave signal.
[0028] For example, in one specific embodiment, the time-domain features corresponding to the P-wave signal and the S-wave signal are: RMS, peak-to-peak value, kurtosis, and energy entropy; the frequency-domain features corresponding to the P-wave signal and the S-wave signal are: dominant frequency, frequency band energy ratio, spectral centroid, and spectral flatness; the time-domain features and frequency-domain features corresponding to the P-wave signal and the S-wave signal are organized into multi-dimensional vectors according to the axis and wave type, and the dimensionality is reduced to ≤32 dimensions through PCA (principal component analysis) or LDA (linear discriminant analysis) to form lightweight feature vectors corresponding to the P-wave signal and the S-wave signal.
[0029] S3. Construct the dataset; Specifically, a dataset is constructed based on the lightweight feature vectors corresponding to the triaxial vibration signal, P-wave signal and S-wave signal at each time point, as well as the construction behavior category label, event hazard level label and P-wave signal arrival time label at each time point. For example, in one specific embodiment, the triaxial vibration signal vector of step S1 and the lightweight feature vectors corresponding to the P-wave signal and S-wave signal of step S2 are combined to form an input feature vector. That is, each time step of the input feature vector includes: triaxial vibration signal, lightweight feature corresponding to P-wave signal, and lightweight feature corresponding to S-wave signal.
[0030] For example, in one specific embodiment, the event hazard level label is set to low, medium and high, that is, the construction behavior category label, the event hazard level label and the P-wave signal arrival time label corresponding to each moment.
[0031] At this point, the dataset has been constructed based on the input feature vector.
[0032] S4. Construct a multi-task learning model and train it; Specifically, a multi-task learning model based on a deep temporal network is constructed, and the multi-task learning model is trained on a dataset to obtain a trained target multi-task learning model. The multi-task learning model includes an encoder and a multi-task output head. The encoder is used to encode the input feature vector into shared features. The multi-task output head includes a classification head for outputting the construction behavior category based on the shared features, a risk head for outputting the event hazard level based on the shared latent layer representation, and a regression head for outputting the arrival time of the P-wave signal based on the shared latent layer representation.
[0033] For example, in one specific embodiment, the feature encoder includes: a multi-scale 1D-CNN layer, a channel attention layer, a bidirectional LSTM layer, and a temporal self-attention layer. The multi-scale 1D-CNN layer is used to extract multi-scale local temporal features of the input features; the channel attention layer is used to dynamically weight the multi-scale local temporal features to obtain a first target feature; the bidirectional LSTM layer is used to add contextual perceptual features to the first target feature to form a second target feature; and the temporal self-attention layer is used to dynamically weight the second target feature corresponding to the time step to obtain shared features. Specifically, in this embodiment, the expression for the multi-scale 1D-CNN layer is:
[0034] through After convolution fusion: Input dimensions: B ×5×400, Output Dimensions: B ×64×400, used to extract multi-scale local temporal patterns; In the formula, This represents the feature data after initial processing by a convolutional neural network (CNN). ( ) indicates splicing features; Indicates the input tensor; This represents a 1D convolution with a kernel size of k; Represents the real number field; This indicates the batch size.
[0035] In this embodiment, the channel attention layer dynamically weights the 64 feature channels output by the multi-scale 1D-CNN layer, highlighting the channels most sensitive to the current construction event (e.g., in the "excavation" event, kurtosis and high-frequency energy channels are enhanced) and suppressing irrelevant channels. Its expression is as follows:
[0036] In the formula, This represents a feature table obtained by dynamically weighting multi-scale local temporal features. This represents the feature data after initial processing by a convolutional neural network (CNN). This represents the Sigmoid function; , This represents a fully connected neural network. Represents the ReLU function; Indicates global average pooling; In this embodiment, a bidirectional LSTM layer is used to capture the action evolution pattern of construction behavior (such as "three consecutive excavations" having a temporal pattern). That is, the bidirectional LSTM layer utilizes both past and future contexts. Therefore, the expression for the bidirectional LSTM layer is:
[0037] In the formula, The first part represents the weighted feature sequence, and BiLSTM represents a bidirectional long short-term memory network. For LSTM output.
[0038] In this embodiment, the temporal self-attention layer is used to weight the target features corresponding to the 400 time steps output by the bidirectional LSTM layer, focusing on key impact moments (such as the instant of the first excavation) and suppressing steady noise segments (such as vehicles passing at a constant speed). The expression for the temporal self-attention layer is:
[0039]
[0040] In the formula, This represents the importance weight at millisecond t; This represents a 256-dimensional fixed vector representing the key patterns learned during training. Let represent the "contextual understanding" result of LSTM in millisecond t, which is a 256-dimensional numeric vector; This means 400 The result of an importance-weighted average; This indicates that LSTM is used to sum and traverse variables. The "contextual understanding" results.
[0041] For example, in one specific embodiment, the shared feature layer of the decoder includes a classification head for outputting the construction behavior category, a risk head for outputting the event hazard level, and a regression head for outputting the arrival time of the P-wave signal, wherein the loss function of the multi-task learning model is:
[0042] In the formula, This represents the loss function of a multi-task learning model. The loss function representing the classification head; The loss function representing the risk head; The loss function representing the regression head; Indicates the learning parameters of the classification head; The learning parameters represent the risk head; This represents the learning parameters of the regression head; by introducing learning parameters, the multi-task learning model automatically balances the optimization difficulty of multiple tasks, avoiding one task dominating the training process. As a regularization term, it ensures the learning parameters It will not approach 0 indefinitely (to avoid numerical instability), which forces the model to control the magnitude of uncertainty while reducing loss.
[0043] The loss function for the classification head is the Focal Loss function. ,Right now:
[0044] It should be noted that in this embodiment, the Focal Loss function is used instead of the traditional cross-entropy loss function (CE) to solve the problem of imbalanced samples of construction behavior categories (such as fewer "excavation" samples and more "vehicle" samples). The Softmax / Sigmoid probability output by the model represents the confidence level. For true labels (one-hot encoded) The modulating factor is when When the value is close to 1 (easily separable samples), the term approaches 0, and the loss is "suppressed"; when... For samples close to 0 (difficult to distinguish), this term is close to 1, and the loss is preserved.
[0045] Traditional multi-class classification (such as Softmax) treats "low," "medium," and "high" as unordered categories, ignoring their inherent ordinal relationship (low < medium < high). Ordinal regression, on the other hand, introduces K-1 learnable thresholds, transforming category prediction into an interval judgment of a potential continuous score S. For three levels (K=3), two learnable thresholds are defined: By comparison The relationship with the threshold determines the level: In this way, the "distance" between levels is automatically learned by the threshold interval, naturally preserving the ordinal relationship, i.e., the loss function of the risk head. The expression is:
[0046] in, The risk level is used to determine the alarm level after an alarm is triggered.
[0047] The regression head uses the Huber Loss loss function. ,Right now:
[0048] in, These are the actual arrival times of the P-wave signal, labeled in milliseconds (ms), provided by manual annotation or a high-precision cross-correlation algorithm. It is the arrival time of the P-wave signal predicted by the multi-task learning model, in milliseconds (ms). It is the inflection point threshold, which controls the boundary between quadratic loss and linear loss. In this system, it is set... The time interval is set to 10 ms to accommodate the physical characteristics of vibration signal propagation and the range of annotation errors. It should be noted that in this system, the arrival time of the P-wave signal is a continuous value regression problem. Traditionally, using MSE (Root Mean Square Error) would make the system overly sensitive to outliers. Huber Loss, however, is a robust loss function that combines the advantages of MSE and MAE: when the prediction error is small (i.e., ... Using quadratic function (MSE) gradient smoothing facilitates fast convergence; when the prediction error is large ( The linear function (MAE) is used to prevent the model from being "biased" by outliers. The arrival time of the P-wave in the regression head result is used as input to the TDOA module to calculate the alarm location.
[0049] S5 outputs the alarm location, the type of construction activity at the alarm location, and the alarm level; Specifically, the lightweight feature vectors corresponding to the triaxial vibration signal, P-wave signal, and S-wave signal at the current moment are input into the target multi-task learning model to obtain the construction behavior category and confidence level, event hazard level, and P-wave signal arrival time at the current moment. When the confidence level is greater than or equal to the preset alarm threshold, the alarm location is obtained based on the time difference of P-wave signal arrival at different nodes at the current moment. The alarm level is determined according to the event hazard level, and the alarm location, construction behavior category at the alarm location, and alarm level are output.
[0050] For example, in one specific embodiment, the step of obtaining the alarm location based on the time difference of arrival of the P-wave signal of different nodes at the current time is as follows: based on the time difference of arrival of the P-wave signal of different nodes at the current time, the alarm location is solved by using the least squares optimization algorithm.
[0051] For example, in one specific embodiment, the confidence level obtained from the classification head... Only confidence level is used to decide whether to generate an alert. Only then does it enter the "alarm generation" process, alarm threshold. It is maintained by an online adaptive method, specifically: the alarm threshold at each time step is obtained using the Exponentially Weighted Moving Average (EWMA) algorithm, where the adaptive expression is:
[0052] in, For the current moment Background noise mean estimation, For the current moment Background noise standard deviation estimation It is the current moment. Dynamic alarm threshold, These are the feature values (RMS, kurtosis, energy entropy) collected at the current moment. This is the smoothing coefficient in the EWMA algorithm, which controls the rate at which historical data is forgotten; a typical value is 0.98. It is the safety margin coefficient, which determines the degree of leniency in the threshold. It is the current moment. The alarm threshold. For example, at a node next to a highway, Automatically rise, thus Adjust the speed upwards to avoid false alarms from highway traffic signals. At points with lower background noise levels, such as farmland, [further adjustments are needed]. Lower Sensitive, ensuring that even minute signals are not missed. The results of manual review are updated using a Bayesian mechanism. Accumulate historical data to achieve "increasing accuracy with use".
[0053] For example, in one specific embodiment, the method further includes incremental fine-tuning of the weights of the target multi-task learning model. The specific incremental fine-tuning steps are as follows: a low-rank adapter is introduced into at least one preset layer of the target multi-task learning model to incrementally update the original weight matrix; when the input feature vector is input to the preset layer, its output feature is obtained by the combined action of the original weight matrix and the low-rank adapter; based on newly collected data or manual verification results, while freezing the original weight matrix, only the parameters in the low-rank adapter are trained, thereby achieving lightweight incremental fine-tuning of the model. In this embodiment, a LoRA low-rank adapter is inserted into the bidirectional LSTM layer and the temporal self-attention layer, where the rank... ≤8, new parameter count ≤1%, incremental training completed in 10 minutes at the edge, no need to revert to the original data; for any original weight matrix W, its forward propagation is modified as follows:
[0054] in, To output the feature vector, Given the input feature vector, and the update quantity It is decomposed into low-rank components:
[0055] In the formula, These are the weights of the original model; 1 is the low-rank adapter matrix; 2 is the low-rank adapter matrix; , Original weights The dimension is r, where r is the low-rank number.
[0056] For events that do not trigger alarms, the system automatically records them to the normal event log. The log includes a timestamp, device ID, lightweight feature vector, confidence level, predicted risk level, and judgment result, used to construct an environmental noise baseline and negative sample set. All events (including alarm events, normal events, and false alarms or missed alarms confirmed by manual review) are structured and stored in a historical event database. This database supports retrieval by device, time, and event type, and provides high-quality labeled data for subsequent model optimization and threshold adjustment. Based on this, the system supports a sample augmentation mechanism: data augmentation is performed on samples in the historical event database to alleviate the scarcity of construction event samples and class imbalance, improving the robustness of the model in new scenarios. When the system detects abnormal signal quality of a TR device or a significant change in the environment, it triggers an input recalibration process: using stable noise segments in the normal event log, it recalculates the sensor normalization parameters (such as the mean and standard deviation of the Z-score), and optionally performs model fine-tuning at the edge via the LoRA mechanism to ensure the consistency and effectiveness of feature inputs.
[0057] This embodiment also includes: after outputting alarm information such as alarm location, construction activity type at the alarm location, and alarm level, the alarm information is marked on a GIS map through a visualization platform and pushed to the web interface of maintenance personnel. The system supports linkage with on-site video monitoring to realize automatic alarm screen pop-ups.
[0058] A vibration monitoring and early warning system for high-consequence oil and gas pipelines based on a time-series multi-axis vibration sensor includes: a signal acquisition module, a signal processing module, a dataset construction module, a neural network module, and an alarm module. The signal acquisition module synchronously acquires triaxial vibration signals corresponding to multiple different nodes along the high-consequence zone of the oil and gas pipeline. The signal processing module performs P / S wave physical separation processing on the triaxial vibration signals to obtain P-wave and S-wave signals. The time-domain and frequency-domain features corresponding to the P-wave and S-wave signals are organized into multi-dimensional vectors corresponding to the P-wave and S-wave signals according to axis and wave type. The multi-dimensional vectors are then dimensionality-reduced to obtain lightweight vectors corresponding to the P-wave and S-wave signals. The system employs a feature vector generation module. The dataset construction module generates lightweight feature vectors based on the triaxial vibration signal, P-wave signal, and S-wave signal at each time step. The neural network module constructs a multi-task learning model based on a deep temporal network and trains it on the dataset to obtain a trained target multi-task learning model. This model consists of two parts: a feature encoder and a multi-task output head. The feature encoder receives a sequence of lightweight feature vectors from the edge processing unit and gradually abstracts and fuses spatiotemporal information through a series of neural network layers, ultimately generating a fixed-dimensional shared latent layer representation for downstream tasks. The multi-task output head, based on the shared representation, executes three parallel tasks: a classification head for outputting construction behavior categories, a risk level head for outputting event hazard levels, and a regression head for outputting the precise arrival time of the P-wave.
[0059] The alarm module is used to input the lightweight feature vectors corresponding to the triaxial vibration signal, P-wave signal, and S-wave signal at the current moment into the target multi-task learning model to obtain the construction behavior category and confidence level, event hazard level, and P-wave signal arrival time at the current moment. When the confidence level is greater than or equal to the preset alarm threshold, the alarm location is obtained based on the time difference of P-wave signal arrival at different nodes at the current moment; the alarm level is determined according to the event hazard level, and the alarm location, construction behavior category at the alarm location, and alarm level are output.
[0060] This invention also provides a vibration monitoring and early warning device for high-consequence oil and gas pipelines based on a time-series multi-axis vibration sensing device. The device includes a multi-axis vibration sensing array comprising multiple device substrates. Each substrate integrates a triaxial MEMS accelerometer to ensure that the multiple triaxial MEMS accelerometers are deployed along the high-consequence area of the oil and gas pipeline. The device substrates are installed using a directional deep-buried method to ensure that the Z-axis of the triaxial MEMS accelerometers integrated on the substrates is strictly vertically downward and the X / Y axes are horizontally oriented, effectively decoupling the propagation characteristics of P-waves and S-waves, providing a physical basis for subsequent waveform separation and high-precision TDOA positioning. A low-power embedded processor or AI acceleration chip is installed in the device substrate to perform baseline correction, wavelet denoising, P / S wave physical separation, multi-dimensional feature extraction, and LDA supervised dimensionality reduction, generating lightweight feature vectors. This unit supports at least 72 hours of raw feature data. Based on local caching, it possesses the ability to resume interrupted downloads and update edge models online, ensuring data integrity and system robustness in complex field environments. It also includes a visual alarm platform, IoT data access server, application logic server, and relational database server deployed in the on-site computer room. These are used to receive multi-node feature data, run a multi-task learning model with LSTM-CNN multi-task deep learning and TDOA least squares localization algorithm using a channel-temporal dual attention mechanism, and simultaneously output the construction behavior category, event hazard level, and sound source location. The visual alarm platform integrates a GIS geographic information system to dynamically display pipeline routes, equipment topology, alarm locations, construction behavior categories, and confidence levels, and pushes hierarchical alarm information through multiple channels including SMS, mobile APP, and Web. The platform supports linkage with video surveillance systems, automatically retrieving the corresponding area's footage when an alarm is triggered, realizing a closed-loop response mechanism of "sound source localization—video verification—manual confirmation."
[0061] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.
Claims
1. A method for vibration monitoring and early warning of oil and gas pipelines in high-consequence areas based on a time-series multi-axis vibration sensing device, characterized in that, include: Simultaneously collect triaxial vibration signals corresponding to multiple different nodes along the high-consequence zone of the oil and gas pipeline, and obtain the construction behavior category label corresponding to the triaxial vibration signal at each moment; The triaxial vibration signal is subjected to P / S wave physical separation processing to obtain P wave signal and S wave signal; the time domain features and frequency domain features corresponding to P wave signal and S wave signal are organized into multi-dimensional vectors corresponding to P wave signal and S wave signal according to axis and wave mode, and the multi-dimensional vectors are reduced to obtain lightweight feature vectors corresponding to P wave signal and S wave signal. A dataset is constructed based on the lightweight feature vectors corresponding to P-wave and S-wave signals, the triaxial vibration signal corresponding to each time moment, the construction behavior category label, the event hazard level label, and the P-wave signal arrival time label corresponding to each time moment. A multi-task learning model based on a deep temporal network is constructed, and the multi-task learning model is trained on a dataset to obtain a trained target multi-task learning model. The multi-task learning model includes an encoder and a multi-task output head. The encoder is used to encode the input feature vector into shared features. The multi-task output head includes a classification head for outputting the construction behavior category based on the shared features, a risk head for outputting the event hazard level based on the shared latent layer representation, and a regression head for outputting the arrival time of the P-wave signal based on the shared latent layer representation. The lightweight feature vectors corresponding to the triaxial vibration signal, P-wave signal and S-wave signal at the current moment are input into the target multi-task learning model to obtain the construction behavior category and confidence level, event hazard level and P-wave signal arrival time at the current moment. When the confidence level is greater than or equal to the preset alarm threshold, the alarm location is obtained based on the time difference of arrival of the P-wave signal of different nodes at the current time; the alarm level is determined according to the event hazard level, and the alarm location, the type of construction behavior at the alarm location, and the alarm level are output.
2. The method for monitoring and early warning of vibration in high-consequence oil and gas pipelines based on a time-series multi-axis vibration sensing device according to claim 1, characterized in that, The steps for performing P / S wave physical separation processing on triaxial vibration signals to obtain P-wave and S-wave signals are as follows: High-pass filtering or polynomial fitting methods are used to correct the triaxial vibration signal; The corrected triaxial vibration signal is denoised by wavelet thresholding to obtain the denoised triaxial vibration signal. The P / S wave physical separation of the denoised triaxial vibration signal is performed to obtain the P wave signal and the S wave signal.
3. The method for monitoring and early warning of vibration in high-consequence oil and gas pipelines based on a time-series multi-axis vibration sensing device according to claim 1, characterized in that, The steps to reduce the dimensionality of a multidimensional vector to obtain the lightweight feature vectors corresponding to the P-wave and S-wave signals are as follows: perform principal component analysis or linear discriminant analysis on the multidimensional vector to obtain the lightweight feature vectors corresponding to the P-wave and S-wave signals.
4. The method for monitoring and early warning of vibration in high-consequence oil and gas pipelines based on a time-series multi-axis vibration sensing device according to claim 1, characterized in that the encoder... include: Multi-scale 1D-CNN layers are used to extract multi-scale local temporal features of the input features; The channel attention layer is used to dynamically weight multi-scale local temporal features to obtain the first target feature; A bidirectional LSTM layer is used to add contextual features to the first target feature to form the second target feature; And a temporal self-attention layer, used to dynamically weight the second target features corresponding to the time step to obtain shared features.
5. The method for monitoring and early warning of vibration in high-consequence oil and gas pipelines based on a time-series multi-axis vibration sensing device according to claim 1, characterized in that, The loss function for the multi-task learning model is: In the formula, This represents the loss function of a multi-task learning model. The loss function representing the classification head; The loss function representing the risk head; The loss function representing the regression head; Indicates the learning parameters of the classification head; The learning parameters represent the risk head; This represents the learning parameters of the regression head.
6. A method for monitoring and early warning of vibration in high-consequence oil and gas pipelines based on a time-series multi-axis vibration sensing device, as described in claim 5, is characterized in that... The loss function for the classification head is Focal Loss; the loss function for the risk head is ordinal regression; and the loss function for the regression head is Huber Loss.
7. The method for monitoring and early warning of vibration in high-consequence oil and gas pipelines based on a time-series multi-axis vibration sensing device according to claim 1, characterized in that, Based on the time difference of P-wave signal arrival at different nodes at the current moment, the alarm location is solved using the least squares optimization algorithm.
8. The method for monitoring and early warning of vibration in high-consequence oil and gas pipelines based on a time-series multi-axis vibration sensing device according to claim 1, characterized in that, The alarm threshold at each time point is obtained using the Exponentially Weighted Moving Average (EWMA) algorithm.
9. A method for monitoring and early warning of vibration in high-consequence oil and gas pipelines based on a time-series multi-axis vibration sensing device, as described in claim 1, is characterized in that... It also includes incremental fine-tuning of the weights of the target multi-task learning model. The specific incremental fine-tuning steps are as follows: A low-rank adapter is introduced in at least one preset layer of the target multi-task learning model to incrementally update the original weight matrix; When the input feature vector is input into this preset layer, its output features are obtained by the combined action of the original weight matrix and the low-rank adapter. Based on newly collected data or manual verification results, while freezing the original weight matrix, only the parameters in the low-rank adapter are trained, thereby achieving lightweight incremental fine-tuning of the model.
10. A vibration monitoring and early warning system for oil and gas pipelines in high-consequence areas based on a time-series multi-axis vibration sensor, characterized in that, include: The signal acquisition module is used to synchronously acquire triaxial vibration signals corresponding to multiple different nodes along the high-consequence zone of the oil and gas pipeline, and obtain the construction behavior category corresponding to the triaxial vibration signal at each moment. The signal processing module is used to perform P / S wave physical separation processing on the triaxial vibration signal to obtain P wave signal and S wave signal; organize the time domain features and frequency domain features corresponding to the P wave signal and S wave signal into multi-dimensional vectors corresponding to the P wave signal and S wave signal according to the axis and wave mode; and reduce the dimensionality of the multi-dimensional vectors to obtain lightweight feature vectors corresponding to the P wave signal and S wave signal. The dataset construction module is used to construct a dataset based on the lightweight feature vectors corresponding to the triaxial vibration signal, P-wave signal and S-wave signal at each time point, as well as the construction behavior category label, event hazard level label and P-wave signal arrival time label at each time point. The neural network module is used to construct a multi-task learning model based on a deep temporal network, and to train the multi-task learning model on a dataset to obtain a trained target multi-task learning model. The multi-task learning model includes an encoder and a multi-task output head. The encoder is used to encode the input feature vector into shared features. The multi-task output head includes a classification head for outputting the construction behavior category based on the shared features, a risk head for outputting the event hazard level based on the shared latent layer representation, and a regression head for outputting the arrival time of the P-wave signal based on the shared latent layer representation. The alarm module is used to input the lightweight feature vectors corresponding to the triaxial vibration signal, P-wave signal, and S-wave signal at the current moment into the target multi-task learning model to obtain the construction behavior category and confidence level, event hazard level, and P-wave signal arrival time at the current moment. When the confidence level is greater than or equal to the preset alarm threshold, the alarm location is obtained based on the time difference of P-wave signal arrival at different nodes at the current moment; the alarm level is determined according to the event hazard level, and the alarm location, construction behavior category at the alarm location, and alarm level are output.