Oil and gas leak detection system based on multi-sensor fusion and pattern recognition

Abstract

This application relates to the field of industrial safety monitoring technology and discloses an oil and gas leak detection system based on multi-sensor fusion and pattern recognition. The system includes a multimodal data acquisition unit, a data preprocessing and feature extraction unit, a leak scenario identification unit, a leak source location unit, and an alarm and linkage control unit. The method comprises the following steps: first, multi-source sensor signals are acquired and processed into a feature sequence; then, the sequence is input into a long-short-term memory network model to identify the leak scenario; then, based on the scenario classification results, the optimal algorithm is adaptively selected to calculate the leak source location; finally, the scenario and location information are integrated to trigger a precise alarm and perform linkage control. The present invention evaluates sensor reliability to enhance data quality, utilizes a long-short-term memory network to identify the leak scenario and adaptively selects a location algorithm, and finally optimizes the emergency response through reinforcement learning, achieving a leap from precise detection to intelligent closed-loop control.

Description

Technical Field

[0001] The present invention relates to the technical field of industrial safety monitoring, and in particular to an oil and gas leakage detection system based on multi-sensor fusion and pattern recognition. Background Art

[0002] In the oil and gas industry, oil and gas leak detection systems play a vital role in ensuring safety during production, storage, and transportation. Currently, mainstream technical solutions typically monitor potential leaks by deploying point-contact sensors, open optical path gas detectors, or infrared thermal imaging devices in key areas.

[0003] However, existing detection technologies still have inherent limitations in practical applications. For one thing, the raw signals acquired by these systems are highly susceptible to interference from complex on-site environmental noise, climate change, and sensor performance drift. Traditional data processing methods struggle to effectively eliminate anomalous data and mitigate the effects of faults, thus impacting the stability and reliability of detection results.

[0004] On the other hand, a deeper deficiency lies in the limited ability of most existing systems to analyze leak events, primarily confined to determining whether a leak has occurred. They generally lack the ability to deeply identify the underlying leak patterns, making it impossible to effectively distinguish between leaks with vastly different physical characteristics and severity, such as a sudden, high-pressure jet leak versus a persistent, slow seepage leak.

[0005] Therefore, the present invention proposes an oil and gas leakage detection system based on multi-sensor fusion and pattern recognition to solve the shortcomings of the existing technology. Summary of the Invention

[0006] In response to the shortcomings of the existing technology, the present invention provides an oil and gas leak detection system based on multi-sensor fusion and pattern recognition, which solves the problems of low positioning accuracy, single alarm information and rigid emergency response strategy caused by the inability of oil and gas leak detection technology to effectively distinguish leakage scenarios.

[0007] To achieve the above objectives, the present invention is implemented through the following technical solutions: an oil and gas leak detection system based on multi-sensor fusion and pattern recognition, the system comprising:

[0008] Multimodal data acquisition unit, used to obtain original signals from multiple sensors of different principles in the monitoring area and form a time series of original signals;

[0009] a data preprocessing and feature extraction unit, configured to process the original signal time series to construct a comprehensive feature vector at each time step;

[0010] a leakage scenario recognition unit, configured to input a feature sequence consisting of the comprehensive feature vector into a preset long short-term memory network model, so as to generate a leakage scenario classification result representing the leakage pattern through the long short-term memory network model;

[0011] a leakage source locating unit, configured to adaptively select a locating algorithm according to the leakage scene classification result and calculate the location information of the leakage source;

[0012] An alarm and linkage control unit is used to combine the leakage scene classification result with the location information of the leakage source to trigger an alarm signal and execute a linkage control instruction.

[0013] Preferably, the multimodal data acquisition unit is composed of a contact sensor, an open optical path infrared detector and an ultrasonic sensor array deployed in the monitoring area;

[0014] The contact sensor is used to obtain the oil and gas concentration point measurement values ​​at key nodes in the monitoring area;

[0015] The open optical path infrared detector works according to the Beer-Lambert law. By analyzing the light intensity attenuation of infrared light between the transmitting end and the receiving end, combined with the preset gas absorption coefficient and optical path length, it calculates the integrated measurement value of oil and gas concentration covering a large space.

[0016] The ultrasonic sensor array is used to acquire a broadband acoustic signal generated at a leakage point due to leakage of high-pressure gas.

[0017] Preferably, when processing the original signal time series, the data preprocessing and feature extraction unit uses a Kalman filter to perform recursive processing in order to obtain an optimal estimate of the true state of the system from the original signal with noise. The recursive processing is completed through a cyclic iteration of a prediction phase and an update phase:

[0018] Prediction stage, based on the state estimate of the previous moment and the state transfer matrix , calculate the state prediction value at the current moment , and update the predicted covariance matrix ;

[0019] Update phase, combined with the actual observation value at the current moment , through the Kalman gain Correct the state prediction value to obtain the final state estimate at the current moment , the calculation formula of the state estimate is:

[0020] ;

[0021] Where, Representatives in The final state estimate obtained after combining the observation values ​​at all times; Representatives in The state prediction value is based only on historical data at all times; Representatives in Kalman gain at time t; Representatives in The actual observation value vector from the sensor at each moment; is the observation matrix;

[0022] The data preprocessing and feature extraction unit further obtains the state estimation value As input, the comprehensive feature vector is constructed at each time step, and the comprehensive feature vector contains the state estimate and the time change rate of the state estimate.

[0023] Preferably, before constructing the comprehensive feature vector, the data preprocessing and feature extraction unit further includes:

[0024] Predicting a theoretical state value of a sensor by monitoring the readings of adjacent sensors of the sensor in the area;

[0025] Calculate the deviation between the state estimation value of the certain sensor and the theoretical state value ;

[0026] The deviation Mapped to a standardized credibility indicator , an exponential decay model is used to map the increase in deviation to a nonlinear decrease in credibility. The calculation formula of the credibility index is:

[0027] ;

[0028] Where, Preset parameters for adjusting the sensitivity of the deviation to the credibility index; is the natural exponential function;

[0029] The data preprocessing and feature extraction unit uses the credibility index The state estimation value from the certain sensor is weighted, and the weighted state estimation value is used to construct the comprehensive feature vector.

[0030] Preferably, the leakage scenario recognition unit processes the feature sequence formed by the comprehensive feature vector, captures the dependency of data in the time dimension, and adopts the long short-term memory network model for learning;

[0031] The long short-term memory network model regulates information flow through internal forget gates, input gates, and output gates, and updates its cell state to achieve selective memory of time series information;

[0032] The cell state at each moment The updating process is the fusion of the cell state at the previous moment The cell state update formula obtained after the input information at the current moment is:

[0033] ;

[0034] Where, is the cell state at the current moment; For the last moment The cell state; is the output of the forget gate; is the output of the input gate; It is the element-wise product operation; is the hyperbolic tangent activation function; is the weight matrix; is the hidden state at the previous moment; is the comprehensive feature vector at the current moment; is the bias vector;

[0035] The leakage scenario recognition unit ultimately generates the leakage scenario classification result based on the output of the long short-term memory network model, and the leakage scenario classification result includes a high-pressure jet leakage scenario and a slow leakage scenario.

[0036] Preferably, the leakage source locating unit adaptively selects a locating algorithm and calculates the leakage source location information in the following manner:

[0037] When the received leakage scenario classification result is a high-pressure jet leakage scenario, in order to use the acoustic signal generated in this scenario for positioning, the leakage source positioning unit calls an arrival time difference positioning algorithm based on an ultrasonic sensor array. The algorithm determines the leakage source position by solving the intersection of a set of hyperbolic equations. Each hyperbola is determined by the time difference between the acoustic signal arriving at any pair of sensors. The hyperbola equation is:

[0038] ;

[0039] Where, , is the coordinate of the leakage source to be solved; , and , are the coordinates of the pair of sensors respectively; is the speed of sound; is a measurement of the time difference between the arrival of the acoustic signal at the pair of sensors;

[0040] When the received leakage scenario classification result is a slow leakage scenario, in order to use the gas concentration field distribution in this scenario for positioning, the leakage source positioning unit calls the gas reverse diffusion model positioning algorithm. This algorithm uses iterative optimization to find a set of virtual leakage source parameters that can best fit the current concentration readings of each sensor. This process is achieved by solving the following optimization problem:

[0041] ;

[0042] In the formula, the goal of optimization is to find the optimal leakage source strength Coordinates of the leak source , ; is the total number of sensors; For the The actual concentration measurement value of each sensor; The gas diffusion model is based on the leakage source intensity , leak source coordinates , , wind speed vector and atmospheric stability level The predicted The concentration value at each sensor location.

[0043] Preferably, the alarm and linkage control unit generates and executes the linkage control instruction after combining the leakage scene classification result and the location information of the leakage source. The linkage control instruction is used to send a control signal to the emergency response equipment deployed in the monitoring area to drive the emergency response equipment to perform operations such as cutting off the leakage source, reducing the oil and gas concentration in the leakage area, or diluting the leaked oil and gas.

[0044] Preferably, the linkage control instruction can be adaptively optimized according to the real-time effect of emergency disposal, and the alarm and linkage control unit further includes:

[0045] The entire emergency response process is modeled as a reinforcement learning problem, in which the state is a combination of the leakage scenario classification results and the real-time sensor readings, the action is the emergency response combination corresponding to the linkage control instruction, and the reward is determined by the change in the monitored leakage hazard level after executing the linkage control instruction;

[0046] Through the learning algorithm, the goal is to maximize the expected value of the long-term cumulative reward and continuously optimize the decision-making strategy. The strategy specifies which action to choose in any state. The goal of the learning algorithm is to solve the optimal strategy defined by the following formula :

[0047] ;

[0048] Where, is the strategy to be optimized; is the expected function; Refers to the strategy of seeking to maximize the objective function Operation; is the time step index; is a discount factor used to balance the importance of current rewards and future rewards; For the Rewards obtained in the following steps: For the The status of the step; For the Steps according to strategy The action of choice;

[0049] The alarm and linkage control unit is configured to determine the optimal strategy according to the solution. , dynamically adjust and generate the linkage control instructions that are finally executed.

[0050] Preferably, the alarm signal triggered by the alarm and linkage control unit is composed of an audible and visual alarm signal acting on the scene, a remote alarm information sent to a preset mobile terminal, and a voice alarm information broadcast at the monitoring center; the content of the remote alarm information includes the location information of the leakage source calculated by the leakage source positioning unit, and the leakage scene classification result generated by the leakage scene recognition unit.

[0051] The present invention also provides an oil and gas leakage detection method based on multi-sensor fusion and pattern recognition, the method comprising the following steps:

[0052] S1. Collect the original signals output by various sensors with different principles in the monitoring area and form a time series of the original signals;

[0053] S2. Processing the original signal time series to construct a comprehensive feature vector at each time step, thereby forming a feature sequence composed of the comprehensive feature vectors;

[0054] S3, inputting the feature sequence into a preset long short-term memory network model to generate a leakage scene classification result representing the leakage pattern through the long short-term memory network model;

[0055] S4. Adaptively selecting a positioning algorithm based on the leakage scene classification result, and calculating the location information of the leakage source using the positioning algorithm;

[0056] S5. Combining the leakage scene classification result with the location information of the leakage source, triggering an alarm signal and executing a linkage control instruction.

[0057] The present invention provides an oil and gas leak detection system based on multi-sensor fusion and pattern recognition. It has the following beneficial effects:

[0058] 1. The present invention uses a long short-term memory network model to conduct in-depth analysis of multidimensional feature sequences and realize the classification of leakage patterns. The system then adaptively activates the optimal positioning algorithm based on the classification results. It can effectively process and accurately locate leakage scenarios with different physical characteristics, such as high-pressure injection and slow leakage. Compared with the existing technology that uses a single, fixed detection logic and positioning method, the present invention solves the defects of poor adaptability and difficulty in ensuring positioning accuracy when facing complex and changeable actual leakage conditions.

[0059] 2. This invention proposes a data preprocessing solution that combines real-time sensor reliability assessment. By dynamically calculating the reliability index of each sensor and weighting the data, it effectively suppresses the interference of abnormal data on the overall analysis, improving the quality of input features and the system's fault tolerance. Compared with the simple data fusion or filtering processing methods used in existing technologies, it solves the problem of reduced reliability of the entire detection system and the susceptibility to misjudgment or missed detection caused by the failure or performance drift of a single sensor.

[0060] 3. This invention constructs the emergency response process as a reinforcement learning model, enabling the linkage control unit to continuously optimize its control strategy based on real-time feedback from leak development, aiming to maximize long-term cumulative rewards. This transition from passive execution to intelligent decision-making, empowering emergency response with adaptive optimization capabilities. This overcomes the shortcomings of existing linkage control systems, which suffer from rigid logic and single response measures. It also addresses the inability to dynamically adjust based on response effectiveness in complex emergency situations, resulting in low emergency response efficiency.

[0061] 4. By integrating multimodal perception, scene recognition, and adaptive positioning, this invention establishes a complete perception-cognition-decision-making information processing chain. The system's final output alarm signal not only indicates the occurrence of a leak but also includes key qualitative and quantitative information about the leak source and pattern. This enriches the connotation and practical value of alarm information. Compared to traditional technologies that only provide a leak presence / absence judgment or a single physical quantity reading, this invention addresses the fundamental limitation of alarm information, which is limited to a single dimension and cannot provide sufficient situational awareness and decision-making basis for emergency response. BRIEF DESCRIPTION OF THE DRAWINGS

[0062] Figure 1 This is a system architecture diagram of the present invention;

[0063] Figure 2Flow chart of the method of the present invention. DETAILED DESCRIPTION

[0064] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the drawings in the present specification. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0065] See also Figure 1 The embodiment of the present invention provides an oil and gas leakage detection system based on multi-sensor fusion and pattern recognition, the system comprising:

[0066] Multimodal data acquisition unit, used to obtain original signals from multiple sensors of different principles in the monitoring area and form a time series of original signals;

[0067] In this embodiment, the oil and gas leak detection system based on multi-sensor fusion and pattern recognition begins with a multimodal data acquisition unit. This unit acquires a variety of raw physical signals in real time from the monitored physical space using a variety of sensors based on different operating principles. It then integrates and synchronizes these signals to form a time-stamped raw signal time series, providing comprehensive, unprocessed initial data for the system's subsequent analysis and decision-making units.

[0068] Specifically, the hardware structure of the multimodal data acquisition unit is a composite sensor network composed of multiple sensor types that complement and collaborate with each other. It aims to cover the physical manifestations that may be produced by oil and gas leakage incidents of different types and sizes through information redundancy and complementarity at the physical level.

[0069] In the specific implementation, the unit is composed of contact sensors, open optical path infrared detectors and ultrasonic sensor arrays deployed in the monitoring area to form a three-dimensional, multi-dimensional perception system.

[0070] Contact sensors primarily measure oil and gas concentrations at specific, critical points within the monitoring area. Typically, these sensors are deployed in high-risk locations identified through historical experience or risk assessments, such as around valves on storage tanks, at flange connections on pipelines, and near pump seals. Deployed in close proximity to these potential leak sources, they can sensitively detect early, small-scale oil and gas leaks.

[0071] In one possible implementation, contact sensors can be catalytic combustion sensors, electrochemical sensors, or metal oxide semiconductor sensors to adapt to different monitoring environments and target gases. Catalytic combustion sensors are sensitive to a broad spectrum of combustible gases; electrochemical sensors are highly selective for specific toxic gases (such as hydrogen sulfide); and metal oxide semiconductor sensors are used in cost-sensitive applications requiring fast response. Together, these sensors form a comprehensive monitoring network for critical locations.

[0072] Open optical path infrared detectors are designed to provide non-contact, regional monitoring of large, open, or semi-open spaces. These detectors consist of an infrared transmitter and receiver, typically deployed at the boundaries or diagonals of the monitored area, ensuring their optical path covers the largest possible volume.

[0073] The detector operates according to the Beer-Lambert Law. When an infrared beam of a specific wavelength emitted by the transmitter passes through an air medium containing the target oil and gas, the beam energy is absorbed by the oil and gas molecules and attenuated. By accurately measuring the change in infrared light intensity at the receiver relative to the initial intensity at the transmitter, and combining this with known physical parameters, the integral or average concentration of the oil and gas along the entire optical path can be calculated. This concentration calculation is based on the following relationship:

[0074] ;

[0075] Where, is the calculated average concentration of oil and gas along the optical path; It is the light intensity actually measured by the infrared receiving end; The initial light intensity of the infrared emission end is usually pre-calibrated in a clean environment without target gas. is the gas absorption coefficient of the target oil and gas component to the infrared light of the specific wavelength, which is a known physical constant; is the optical path length between the infrared transmitter and the receiver.

[0076] The measurement method effectively makes up for the limitation of contact sensors that can only perform point measurements. It can detect oil and gas clouds that have spread or diffused in a larger space, which is of great significance for assessing the scope of leakage impact.

[0077] Ultrasonic sensor arrays are designed specifically to detect specific leak scenarios, namely high-pressure gas leaks. When gas stored in a pressure vessel or high-pressure pipeline is ejected into the atmosphere at high speed through a crack or small hole, the intense turbulence and shear effects generate a high-energy, broadband acoustic signal at the leak point. This signal typically falls within the ultrasonic frequency range, above the human hearing threshold.

[0078] The array consists of multiple ultrasonic sensors distributed in a specific spatial geometry. These sensors are configured to operate synchronously to capture the broadband acoustic signal generated at the leak point due to a high-pressure gas leak. Each sensor not only records the waveform and intensity of the acoustic signal but, more importantly, records the arrival time of the acoustic signal with extremely high temporal precision. This highly accurate arrival time information is the core basis for subsequent sound source localization calculations.

[0079] Since the ultrasonic signal is generated by the physical behavior of leakage itself, its propagation is not affected by the wind field, and the background noise in the environment is low. Therefore, this technology has an extremely high signal-to-noise ratio and reliability for the detection of high-pressure jet leaks.

[0080] This multimodal data acquisition unit aggregates data from all of the aforementioned sensors (contact sensor concentration readings, integrated concentration values ​​calculated by an open-light infrared detector, and acoustic signal waveforms collected by an ultrasonic sensor array). Using a synchronized clock mechanism, each piece of collected data is precisely timestamped, forming a unified, multi-dimensional raw signal time series that is then transmitted to the system's subsequent data preprocessing and feature extraction units for further processing.

[0081] Data preprocessing and feature extraction unit, used to process the original signal time series to construct a comprehensive feature vector at each time step;

[0082] In this embodiment, the data preprocessing and feature extraction unit receives the original signal time series formed by the aforementioned multimodal data acquisition unit, and converts these original signals, which are usually accompanied by measurement noise and random disturbances, into a structured feature representation with higher information density, namely, a comprehensive feature vector, for subsequent analysis by the leakage scenario recognition unit.

[0083] Specifically, when processing the original signal time series, this embodiment uses a Kalman filter to recursively process the time series data to extract the optimal estimate that reflects the system's true physical state from the noisy signal. The Kalman filter is a highly efficient autoregressive filter that can optimally estimate the state of a dynamic system in the presence of uncertainty. This recursive processing is accomplished through a cyclic iteration of the prediction phase and the update phase.

[0084] In the prediction phase, the unit is based on the previous moment ( ) The optimal state estimate obtained And a preset state transition matrix that describes the inherent dynamic behavior of the system , to calculate the current moment ( ) state prediction value At the same time, the uncertainty associated with the state estimation, that is, the error covariance matrix, also increases from the previous moment Update to the current prediction covariance matrix .

[0085] In the update phase, the unit takes the actual observation value at the current moment (i.e. the original signal obtained from the multimodal data acquisition unit) is introduced into the calculation. By calculating the Kalman gain , the gain is used as a weight coefficient to balance the uncertainty of the state prediction value and the uncertainty of the actual observation value. The state prediction value is corrected using the Kalman gain to obtain the final state estimate at the current moment This state estimate, which combines the predicted information with the observed information, is a more accurate description of the current state of the system. The calculation and update process of the state estimate can be expressed by the following state update equation:

[0086] ;

[0087] Where, Representatives in The final state estimate obtained after combining the observation values ​​at all times; Representatives in The state prediction value is based only on historical data at all times; Representatives in Kalman gain at time t; Representatives in The actual observation value vector from the sensor at each moment; is the measurement matrix, which establishes the mapping relationship between the state space and the observation space.

[0088] In one possible implementation, in order to further improve the robustness of the system in the face of failure or performance drift of a single or partial sensor, the data preprocessing and feature extraction unit also performs a set of operations to evaluate and utilize the real-time reliability of each sensor before constructing the comprehensive feature vector.

[0089] This operation predicts the theoretical state value of a sensor based on the readings of neighboring sensors within the monitoring area. This prediction is based on the physical assumption that the physical process of gas diffusion strongly correlates the readings of spatially adjacent sensors. This theoretical state value is generated using spatial interpolation algorithms, such as the inverse distance weighted method or kriging, using the estimated state values ​​of surrounding sensors.

[0090] Then, the deviation between the state estimate of a sensor (obtained by Kalman filtering) and the theoretical state value predicted in the previous step is calculated. The magnitude of this deviation directly reflects the degree of consistency between the sensor's reading and the local physical field formed by its neighboring sensors. A sensor with a normal reading will have a small deviation; a faulty sensor's reading will deviate significantly from the surrounding trend, resulting in a larger deviation.

[0091] To convert this deviation value into a standardized, easy-to-use indicator, the unit converts the deviation Mapped to a standardized credibility indicator In this embodiment, an exponential decay model is used, which can map the linear increase of deviation to the nonlinear decrease of credibility, thereby more effectively suppressing abnormal deviations. The calculation formula of the credibility index, that is, the credibility index model, is defined as follows:

[0092] ;

[0093] Where, For the The sensors in The credibility index of the moment, its value range is between 0 and 1; is the natural exponential function; is a pre-set positive real number parameter used to adjust the sensitivity of the deviation to the credibility index. The larger the value, the more drastically the credibility decreases with the increase of deviation; The calculated The sensors in Deviation of time.

[0094] Data preprocessing and feature extraction unit to calculate the credibility index The state estimates from a particular sensor are weighted, using the weights as the weights. These weighted state estimates are then used to construct the comprehensive feature vector. This way, data from highly reliable sensors dominates the subsequent analysis, while the influence of data from unreliable sensors is significantly reduced.

[0095] After obtaining the filtered and reliability-weighted state estimates of each sensor, the unit constructs a comprehensive feature vector at each time step. This comprehensive feature vector forms the basis for input into downstream models for pattern recognition. To comprehensively characterize a leak event, the comprehensive feature vector includes not only the current state estimate itself, which reflects the static intensity and spatial distribution of the leak, but also the time rate of change of the state estimate. This rate of change is calculated by calculating the difference between the current state estimate and the previous state estimate, directly reflecting the dynamic trend of the leak. Combining static and dynamic information forms a comprehensive feature vector that comprehensively characterizes the leak event, thereby forming a feature sequence composed of comprehensive feature vectors.

[0096] a leakage scenario recognition unit, configured to input a feature sequence consisting of a comprehensive feature vector into a preset long short-term memory network model, so as to generate a leakage scenario classification result representing the leakage pattern through the long short-term memory network model;

[0097] In this embodiment, the leakage scenario recognition unit receives and processes the feature sequence composed of comprehensive feature vectors generated by the data preprocessing and feature extraction unit. This unit performs deep learning analysis on this feature sequence to ultimately generate a leakage scenario classification result that characterizes the underlying leakage pattern.

[0098] Specifically, to effectively process the feature sequence composed of comprehensive feature vectors and deeply capture the complex dependencies of data across time, this embodiment utilizes a pre-trained long-short-term memory (LSTM) network model for learning and recognition. This model was chosen because the evolution of oil and gas leaks is inherently a dynamic time process. Different leak scenarios (e.g., instantaneous high-pressure jets versus long-term, slow leaks) manifest in sensor data through specific patterns in the evolution of features over time. As a variant of recurrent neural networks, the unique internal structure of LSTM networks enables them to effectively learn and memorize long-term dependencies in time series, thereby accurately distinguishing leak patterns.

[0099] In one possible implementation, the LSTM model uses a unique internal gating mechanism to regulate the flow of information within the network and continuously update a cell state throughout the entire sequence processing process, thereby achieving selective memory of time series information. This gating mechanism is primarily implemented by three core components: the forget gate, the input gate, and the output gate. The forget gate determines which old information from the previous cell state is forgotten; the input gate determines which new input information is stored in the current cell state; and the output gate controls what information is output to the next layer or the next time step based on the current updated cell state.

[0100] The cell state at each moment the model processes the sequence data The updating process does not simply overwrite the old information with the new information, but organically integrates the cell state at the previous moment. This fusion mechanism ensures that key historical information is retained while new trends are promptly incorporated. The cell state update process can be accurately described by the following cell state update formula:

[0101] ;

[0102] Where, is the cell state at the current moment; For the last moment The cell state; is the output of the forget gate; is the output of the input gate; It is the element-wise product operation; is the hyperbolic tangent activation function; is the weight matrix; is the hidden state at the previous moment; is the comprehensive feature vector at the current moment; is the bias vector.

[0103] After inputting the complete feature sequence into the long-short-term memory network model for processing, the leak scenario recognition unit ultimately generates a leak scenario classification result based on the output state of the long-short-term memory network model through a classification layer (e.g., a fully connected layer followed by a softmax activation function). This classification result is a clear label that indicates to which preset leak mode the currently monitored event belongs. In this embodiment, the leak scenario classification results specifically include two scenarios with significantly different physical characteristics and emergency response requirements: a high-pressure jet leak scenario and a slow seepage scenario. This classification result is passed to the subsequent leak source localization unit as a direct basis for its selection of a localization algorithm.

[0104] A leakage source positioning unit is used to adaptively select a positioning algorithm based on the leakage scene classification result and calculate the location information of the leakage source;

[0105] In this embodiment, the leak source localization unit receives the leak scenario classification results output by the aforementioned leak scenario identification unit. This unit's core function is to adaptively select and execute a matching localization algorithm based on the physical characteristics of the leak as characterized by the classification results, ultimately calculating and outputting the leak source's location information. This adaptive localization strategy aims to utilize the optimal physical model and data source for different leak patterns, thereby improving localization accuracy and reliability.

[0106] Specifically, the leak source location unit is pre-configured with at least two location algorithms, each corresponding to a specific leak scenario. The unit's operating logic is a conditional branching process based on the classification results.

[0107] When the received leak scenario classification result is a high-pressure jet leak, the system determines that the current leak event is accompanied by a strong release of acoustic energy. To accurately locate the source using the characteristic acoustic signals generated in this scenario, the leak source localization unit uses the Time-Difference-of-Arrival (TDOA) localization algorithm based on the ultrasonic sensor array.

[0108] The physical basis of this algorithm is that after the sound wave is emitted from the leakage point, it propagates in all directions at the speed of sound. When it reaches any two sensors at different positions in space (such as sensor and sensors ). For a fixed leak source, this time difference is constant. The set of all possible locations that produce this constant time difference forms a hyperbola on a two-dimensional plane with the two sensors as the foci. By using multiple pairs of sensors in the ultrasonic sensor array, multiple such hyperbolas can be determined, and the unique intersection of these hyperbolas is the location of the leak source.

[0109] The algorithm locates the leak source by solving the intersection of a set of hyperbolic equations. Each hyperbola is determined by the time difference between the acoustic signal reaching any pair of sensors. The hyperbola equation is:

[0110] ;

[0111] Where, , is the coordinate of the leakage source to be solved; , and , A pair of sensors and sensors coordinates of The propagation speed of the acoustic signal in the current environment medium. This value can be corrected in real time according to environmental parameters such as temperature and humidity. The acoustic signal actually measured and recorded by the ultrasonic sensor array reaches the sensor and sensors The time difference between the measurements.

[0112] When the received leak scenario classification result is a slow leak, the system determines that the acoustic characteristics of the current leak are not obvious, but are mainly manifested in the changes in the gas concentration field distribution in space. To use the gas concentration field distribution in this scenario for positioning, the leak source positioning unit uses the gas reverse diffusion model positioning algorithm.

[0113] The algorithm essentially involves a model inversion problem based on actual measurement data. Through an iterative optimization process, it seeks a set of virtual leak source parameters that best fits the current concentration readings of each sensor. This process is accomplished by solving an optimization problem: finding a set of leak source parameters (location and intensity) that minimizes the overall error between the predicted concentration values ​​at each sensor location, as forward-calculated by the gas diffusion model, and the actual concentration values ​​measured by the sensors.

[0114] In one possible implementation, this process is achieved by solving the following optimization problem:

[0115] ;

[0116] In the formula, the goal of optimization is to find the optimal leakage source strength Coordinates of the leak source , ; is the total number of sensors; For the The actual concentration measurement value of each sensor; The gas diffusion model is based on the leakage source intensity , leak source coordinates , , wind speed vector and atmospheric stability level The predicted The concentration value at each sensor location is continuously adjusted by using algorithms such as gradient descent or particle swarm optimization. , , until the sum of the square errors converges to a minimum value. , This is the best estimate of the location identified as the leak source.

[0117] An alarm and linkage control unit, which is used to combine the leakage scene classification results with the location information of the leakage source, trigger the alarm signal and execute the linkage control instructions;

[0118] In this embodiment, the alarm and linkage control unit serves as the final output and execution terminal of the entire system. Its main function is to receive and integrate the high-level semantic information analyzed by the previous-level unit, namely the leakage scene classification results and the location information of the leakage source, and based on this, generate and execute two types of key outputs: alarm signals and linkage control instructions.

[0119] In a possible implementation, after the unit receives the classification result and location information confirming the leakage, it will start the alarm and linkage control process in parallel.

[0120] In terms of triggering alarm signals, this unit aims to ensure that leakage information can be quickly and comprehensively transmitted to all relevant parties through a multi-modal and multi-channel alarm matrix. Specifically, the alarm signal triggered by the alarm and linkage control unit is composed of the following parts:

[0121] One is the sound and light alarm signal acting on the leakage site. This unit will send a start signal to the sound and light alarm deployed in the monitoring area, driving it to emit high-decibel sirens and high-intensity warning flashes to achieve the most direct and rapid warning to on-site workers.

[0122] The second is a remote alarm message sent to a pre-set mobile terminal. This unit uses an integrated wireless communication module to send alarm information to the personal mobile devices of pre-set managers and emergency response team members via text message or app push. Notably, this remote alarm message is highly informative. It goes beyond a simple warning and explicitly includes the leak source location information calculated by the leak source location unit and the leak scenario classification results generated by the leak scenario identification unit.

[0123] The third is the voice alarm information broadcast at the monitoring center. This unit can drive the voice broadcast system in the central monitoring room, broadcasting detailed information including the leak location and leak scene in a clear voice loop, ensuring that on-duty personnel can immediately obtain complete incident details.

[0124] Regarding the execution of linkage control commands, upon triggering an alarm, the unit generates and executes linkage control commands based on the received leak information. These commands essentially send preset control signals to various emergency response devices deployed within the monitoring area, prompting them to automatically execute a series of preliminary emergency response operations aimed at preventing the disaster from escalating.

[0125] Generally speaking, these operations may include sending a closing command to the emergency shut-off valve of the pipeline upstream of the leakage point to cut off the source of the leakage; starting the forced ventilation system near the leakage area to actively reduce the oil and gas concentration in the local space; or starting the fire sprinkler and water curtain system to dilute and cool the diffused oil and gas cloud.

[0126] In a more advanced implementation, in order to enable the linkage control instructions to break away from the limitations of fixed plans and perform adaptive optimization based on the real-time effects during the emergency response process, the decision logic of the alarm and linkage control unit is constructed as a reinforcement learning problem.

[0127] This implementation modeled the entire emergency response process as a reinforcement learning model. In this model, the system's "state" is composed of the leakage scene classification results and the real-time readings of all sensors at the current moment. This state vector provides comprehensive environmental perception for system decision-making. The system's "action" is the linkage Control commands correspond to various possible combinations of emergency response plans. For example, a specific action might be "Close valve A, simultaneously activate fan B and sprinkler C." The system's "reward" is a quantitative indicator determined by the change in the leak risk level detected by sensors after executing a specific action. If the oil and gas concentration within the station decreases significantly after executing the action, the system receives a positive numerical reward; conversely, if the risk level increases, a negative numerical reward or penalty is awarded.

[0128] This unit uses a pre-defined learning algorithm to iteratively optimize a decision-making strategy with the ultimate goal of maximizing the expected value of long-term cumulative rewards. The strategy specifies which action the system should choose in any given state. The core goal of the learning algorithm is to solve the optimal strategy defined by the following formula:

[0129] ;

[0130] Where, is the strategy to be optimized, which defines the mapping relationship from state to action; is the expected function, which means finding the average value of the internal value under all possibilities; Refers to the strategy of seeking to maximize the objective function Operation; is the time step index; A discount factor used to balance the importance of current rewards and future rewards. It is a preset constant between 0 and 1, used to balance the importance of current immediate rewards and future long-term rewards in decision-making; For the Rewards obtained in the following steps: For the The status of the step; For the Steps according to strategy Select the action.

[0131] The alarm and linkage control unit uses the optimal strategy obtained through offline training and online learning , dynamically adjusting and generating the linkage control instructions that ultimately need to be executed. This means that the unit has the ability to learn from experience and can generate more reasonable and effective emergency control decisions based on real-time feedback than fixed plans.

[0132] See also Figure 2 The present invention also provides an oil and gas leakage detection method based on multi-sensor fusion and pattern recognition, the method comprising the following steps:

[0133] S1. Collect the original signals output by various sensors with different principles in the monitoring area and form a time series of the original signals;

[0134] Step S1 specifically involves deploying a composite sensing network composed of sensors with different operating principles at the industrial sites or along the pipelines to be monitored. This network can include contact sensors for measuring oil and gas concentrations at specific points, open-optical infrared detectors for integrating concentrations over a wide range, and ultrasonic sensor arrays for capturing the high-frequency acoustic signals generated by high-pressure gas injection. The system continuously acquires measurement signals from these sensors and, through a unified clock system, accurately timestamps each piece of data, thereby integrating data from different physical dimensions into a synchronized, multi-channel time series of raw signals.

[0135] S2. Processing the original signal time series to construct a comprehensive feature vector at each time step, thereby forming a feature sequence composed of the comprehensive feature vectors;

[0136] Step S2 aims to deeply process and purify the original signal time series obtained in the previous step in order to extract the core information that can effectively characterize the state of the system. Specifically, the system first uses a filtering algorithm, such as Kalman filtering, to recursively process the original data full of noise and random disturbances, with the aim of obtaining a smoother estimate that better reflects the actual physical state. Furthermore, in order to improve the robustness of the system, the system will also evaluate the real-time reliability of each sensor, for example, by comparing the consistency of the reading of a certain sensor with the reading of its neighboring sensors, and weighting the information of the sensor according to the assessed credibility. Finally, at each time step, the unit combines the filtered and reliability-weighted state estimate, as well as the rate of change of the estimate over time, into a comprehensive feature vector. Over time, these vectors form a feature sequence that can fully describe the dynamic evolution of the leakage event.

[0137] S3. Inputting the feature sequence into a preset long short-term memory network model to generate a leakage scene classification result representing the leakage pattern through the long short-term memory network model;

[0138] Step S3 performs pattern recognition on the feature sequence generated in the previous step to determine the underlying type of leak. The system inputs this feature sequence into a pre-trained long-short-term memory network model. As a deep learning model particularly adept at processing time series data, this network can effectively capture the long-term dependencies and specific patterns of feature vectors evolving over time. By analyzing these dynamic patterns, the model can distinguish leak events with different physical causes, such as high-pressure instantaneous injection due to equipment rupture or continuous slow leakage due to seal aging, and ultimately output a clear leak scenario classification result.

[0139] S4. Adaptively select a positioning algorithm based on the leakage scene classification result, and use the positioning algorithm to calculate the location information of the leakage source;

[0140] Step S4 is a decision-based adaptive positioning process. After receiving the leakage scenario classification result generated in the previous step, the unit will not adopt a single fixed positioning algorithm, but will intelligently select the most appropriate positioning technology based on the classification result. For example, when the classification result is "high-pressure jet leakage scenario", the system will determine that the acoustic signal at this moment is the best positioning basis, so it will call the arrival time difference positioning algorithm based on the ultrasonic sensor array, and infer the sound source position by solving the time difference between the sound wave signal arriving at different sensors. On the contrary, when the classification result is "slow leakage scenario", the acoustic signal is weak, and the system will call the positioning algorithm based on the gas diffusion model instead. By analyzing the reading distribution of the current gas concentration sensors, the system will inversely calculate the location of the leakage source that is most likely to cause this concentration field.

[0141] S5. Combine the leakage scene classification result with the location information of the leakage source, trigger an alarm signal and execute a linkage control instruction;

[0142] Step S5 closes the loop of the entire detection and response process. This unit combines the leak type (classification results from step S3) with the leak location (calculated information from step S4) to generate and execute a precise response action.

[0143] On the one hand, the system triggers a multi-level alarm system, including activating on-site sound and light alarms to alert on-site personnel, sending remote alarm information containing the precise location and type of leak to the manager's mobile terminal, and making a voice broadcast in the central control room;

[0144] On the other hand, the system will automatically generate and execute a series of linkage control instructions based on the location and severity of the leak, such as sending a shutdown signal to the pipeline valve upstream of the leak point to cut off the gas source, or starting emergency response equipment such as ventilation and spraying near the leak point to actively control and mitigate the risks brought by the leak.

[0145] While embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions, and variations may be made to these embodiments without departing from the principles and spirit of the invention, and that the scope of the invention is defined by the appended claims and their equivalents.

Claims

1. The oil and gas leakage detection system based on multi-sensor fusion and pattern recognition is characterized by: The system comprises: Multimodal data acquisition unit, used to obtain original signals from multiple sensors with different principles in the monitoring area and form a time series of original signals; a data preprocessing and feature extraction unit, configured to process the original signal time series to construct a comprehensive feature vector at each time step; When processing the original signal time series, the data preprocessing and feature extraction unit uses a Kalman filter to perform recursive processing to obtain the optimal estimate of the real state of the system from the noisy original signal. The recursive processing is completed through a cyclic iteration of the prediction phase and the update phase: Prediction stage, based on the state estimate of the previous moment and the state transfer matrix , calculate the state prediction value at the current moment , and update the predicted covariance matrix ; Update phase, combined with the actual observation value at the current moment , through the Kalman gain Correct the state prediction value to obtain the final state estimate at the current moment , the calculation formula of the state estimate is: ； Where, Representatives in The final state estimate obtained after combining the observation values ​​at all times; Representatives in The state prediction value is based only on historical data at all times; Representatives in Kalman gain at time t; Representatives in The actual observation value vector from the sensor at each moment; is the observation matrix; The data preprocessing and feature extraction unit further obtains the state estimation value As input, constructing the comprehensive feature vector at each time step, the comprehensive feature vector including the state estimate and the time rate of change of the state estimate; a leakage scenario recognition unit, configured to input a feature sequence consisting of the comprehensive feature vector into a preset long short-term memory network model, so as to generate a leakage scenario classification result representing the leakage pattern through the long short-term memory network model; a leakage source locating unit, configured to adaptively select a locating algorithm according to the leakage scene classification result and calculate the location information of the leakage source; An alarm and linkage control unit is used to combine the leakage scene classification result with the location information of the leakage source to trigger an alarm signal and execute a linkage control instruction.

2. The oil and gas leakage detection system based on multi-sensor fusion and pattern recognition according to claim 1 is characterized in that: The multimodal data acquisition unit is composed of a contact sensor, an open optical path infrared detector and an ultrasonic sensor array deployed in the monitoring area; The contact sensor is used to obtain the oil and gas concentration point measurement values ​​at key nodes in the monitoring area; The open optical path infrared detector works according to the Beer-Lambert law. By analyzing the light intensity attenuation of infrared light between the transmitting end and the receiving end, combined with the preset gas absorption coefficient and optical path length, it calculates the integrated measurement value of oil and gas concentration covering a large space. The ultrasonic sensor array is used to acquire a broadband acoustic signal generated at a leakage point due to leakage of high-pressure gas.

3. The oil and gas leakage detection system based on multi-sensor fusion and pattern recognition according to claim 1 is characterized in that: Before constructing the comprehensive feature vector, the data preprocessing and feature extraction unit further includes: Predicting a theoretical state value of a sensor by monitoring the readings of adjacent sensors of the sensor in the area; Calculate the deviation between the state estimation value of the certain sensor and the theoretical state value ; The deviation Mapped to a standardized credibility indicator , an exponential decay model is used to map the increase in deviation to a nonlinear decrease in credibility. The calculation formula of the credibility index is: ； Where, Preset parameters for adjusting the sensitivity of the deviation to the credibility index; is the natural exponential function; The data preprocessing and feature extraction unit uses the credibility index The state estimation value from the certain sensor is weighted, and the weighted state estimation value is used to construct the comprehensive feature vector.

4. The oil and gas leakage detection system based on multi-sensor fusion and pattern recognition according to claim 1 is characterized in that: The leakage scene recognition unit processes the feature sequence composed of the comprehensive feature vector, captures the dependency of data in the time dimension, and adopts the long short-term memory network model for learning; The long short-term memory network model regulates information flow through internal forget gates, input gates, and output gates, and updates its cell state to achieve selective memory of time series information; The cell state at each moment The updating process is the fusion of the cell state at the previous moment The cell state update formula obtained after the input information at the current moment is: ； Where, is the cell state at the current moment; For the last moment The cell state; is the output of the forget gate; is the output of the input gate; It is the element-wise product operation; is the hyperbolic tangent activation function; is the weight matrix; is the hidden state at the previous moment; is the comprehensive feature vector at the current moment; is the bias vector; The leakage scenario recognition unit ultimately generates the leakage scenario classification result based on the output of the long short-term memory network model, and the leakage scenario classification result includes a high-pressure jet leakage scenario and a slow leakage scenario.

5. The oil and gas leakage detection system based on multi-sensor fusion and pattern recognition according to claim 4 is characterized in that: The leakage source positioning unit adaptively selects the positioning algorithm and calculates the leakage source position information in the following manner: When the received leakage scenario classification result is a high-pressure jet leakage scenario, in order to use the acoustic signal generated in this scenario for positioning, the leakage source positioning unit calls an arrival time difference positioning algorithm based on an ultrasonic sensor array. The algorithm determines the leakage source position by solving the intersection of a set of hyperbolic equations. Each hyperbola is determined by the time difference between the acoustic signal arriving at any pair of sensors. The hyperbola equation is: ； Where, , is the coordinate of the leakage source to be solved; , and , are the coordinates of the pair of sensors respectively; is the speed of sound; is a measurement of the time difference between the arrival of the acoustic signal at the pair of sensors; When the received leakage scenario classification result is a slow leakage scenario, in order to use the gas concentration field distribution in this scenario for positioning, the leakage source positioning unit calls the gas reverse diffusion model positioning algorithm. This algorithm uses iterative optimization to find a set of virtual leakage source parameters that can best fit the current concentration readings of each sensor. This process is achieved by solving the following optimization problem: ； In the formula, the goal of optimization is to find the optimal leakage source strength Coordinates of the leak source , ; is the total number of sensors; For the The actual concentration measurement value of each sensor; The gas diffusion model is based on the leakage source intensity , leak source coordinates , , wind speed vector and atmospheric stability level The predicted The concentration value at each sensor location.

6. The oil and gas leakage detection system based on multi-sensor fusion and pattern recognition according to claim 1 is characterized in that: After combining the leakage scene classification result and the location information of the leakage source, the alarm and linkage control unit generates and executes the linkage control instruction. The linkage control instruction is used to send a control signal to the emergency response equipment deployed in the monitoring area to drive the emergency response equipment to perform operations such as cutting off the leakage source, reducing the oil and gas concentration in the leakage area, or diluting the leaked oil and gas.

7. The oil and gas leakage detection system based on multi-sensor fusion and pattern recognition according to claim 6 is characterized in that: The linkage control instructions can be adaptively optimized according to the real-time effect of emergency response, and the alarm and linkage control unit also includes: The entire emergency response process is modeled as a reinforcement learning problem, in which the state is a combination of the leakage scenario classification results and the real-time sensor readings, the action is the emergency response combination corresponding to the linkage control instruction, and the reward is determined by the change in the monitored leakage hazard level after executing the linkage control instruction; Through the learning algorithm, the goal is to maximize the expected value of the long-term cumulative reward and continuously optimize the decision-making strategy. The strategy specifies which action to choose in any state. The goal of the learning algorithm is to solve the optimal strategy defined by the following formula : ； Where, is the strategy to be optimized; is the expected function; Refers to the strategy of seeking to maximize the objective function Operation; is the time step index; is a discount factor used to balance the importance of current rewards and future rewards; For the Rewards obtained in the following steps: For the The status of the step; For the Steps according to strategy The action of choice; The alarm and linkage control unit is configured to determine the optimal strategy according to the solution. , dynamically adjust and generate the linkage control instructions that are finally executed.

8. The oil and gas leakage detection system based on multi-sensor fusion and pattern recognition according to claim 1 is characterized in that: The alarm signal triggered by the alarm and linkage control unit is composed of an audible and visual alarm signal acting on the scene, a remote alarm information sent to a preset mobile terminal, and a voice alarm information broadcast at the monitoring center; the content of the remote alarm information includes the location information of the leakage source calculated by the leakage source positioning unit, and the leakage scene classification result generated by the leakage scene recognition unit.

9. An oil and gas leakage detection method based on multi-sensor fusion and pattern recognition, applied to the system according to any one of claims 1 to 8, characterized in that: The method comprises the following steps: S1. Collect the original signals output by various sensors with different principles in the monitoring area and form a time series of the original signals; S2. Processing the original signal time series to construct a comprehensive feature vector at each time step, thereby forming a feature sequence composed of the comprehensive feature vectors; S3, inputting the feature sequence into a preset long short-term memory network model to generate a leakage scene classification result representing the leakage pattern through the long short-term memory network model; S4. Adaptively selecting a positioning algorithm based on the leakage scene classification result, and calculating the location information of the leakage source using the positioning algorithm; S5. Combining the leakage scene classification result with the location information of the leakage source, triggering an alarm signal and executing a linkage control instruction.

Images