Gas abnormal working condition monitoring and alarming system

Abstract

The invention provides a gas abnormal working condition monitoring and alarming system, relates to the field of gas monitoring and early warning, and adopts a monitoring system comprising a data preprocessing module, a spatial-temporal characteristic module, an ST-CGN network module, an edge intelligent optimization module and a result processing module. Wherein the data preprocessing module comprises a sliding window fragmentation method, a missing value interpolation method, a noise filtering method, a normalization method and a coding method; the spatial-temporal feature module extracts spatial correlation features, time dynamics features and causal propagation features, and fuses and splices the features into a three-dimensional feature tensor; the ST-CGN network module outputs a multi-working-condition probability vector corresponding to a sliding window fragment through the input of a three-dimensional feature tensor; and the edge intelligent optimization module realizes edge optimization of the ST-CGN network module. The system is used for monitoring and early warning gas abnormal working conditions, solves many problems existing in traditional gas pipe network monitoring, and has the advantages of being rapid in response, high in monitoring precision, high in cross-scene generalization ability and the like.

Description

Technical Field

[0001] The present invention relates to the technical field of gas monitoring and early warning, and particularly to a gas abnormal condition monitoring and alarm system. Background Art

[0002] The monitoring of gas abnormal conditions is a key link to ensure the safe operation of the gas system, and its purpose is to identify and warn potential risks such as gas leakage, pressure abnormality, and equipment failure; at present, the monitoring of gas abnormal conditions mainly adopts the monitoring method of single-point threshold alarm, that is, sensors (pressure, flow, gas concentration sensors, etc.) are deployed at the monitoring points, safety thresholds are set, and data is collected in real time. When the collected data exceeds or is lower than the preset safety threshold, an alarm is triggered. However, with the high-rise and complexity of urban pipe networks, the traditional gas single-point threshold alarm method ignores the physical coupling and pressure fluctuation propagation among nodes in the gas pipe network, resulting in difficulty in identifying the source of abnormal information and thus unable to accurately monitor abnormal information, and problems such as false alarm triggering occur (for example: when a part of the pipeline on the third floor is blocked, the pressure fluctuation will be transiently conducted through the pipeline to the second and fourth floors, generating synchronous or time-delay characteristics, but single-point monitoring cannot capture this multi-point coupling relationship); at the same time, the traditional gas single-point threshold alarm method only compares the instantaneous value or the average value within a period of time with the threshold, and it is difficult to capture the high-frequency micro-vibrations of the leaked gas, resulting in problems such as missed detection or alarm delay, and thus inevitable safety hazards occur (experimental data shows that the standard deviation of pressure in the slight leakage stage increases suddenly by 2.5 to 3 times compared with the normal state, but the average value change in the slight leakage stage is less than 1%. If the single-point threshold alarm method based on the average value threshold is used, it is extremely easy to ignore the slight leakage stage; at the same time, the slow pressure drop caused by pipeline aging, such as a pressure drop of 0.3 to 0.7 kPa per hour, is usually considered as a normal fluctuation under daily use, resulting in alarm delay). In addition, due to the deepening of the urbanization process, the gas pipe networks in different communities or factories are linear, tree-shaped, ring-shaped, or even mixed topologies. The thresholds calibrated based on artificial experience and a large amount of experimental data usually need to be recalibrated for different pipe network structures, with a long debugging cycle, high cost, and being extremely susceptible to external environmental factors, resulting in large deviation and low accuracy of the calibrated thresholds.

[0003] In summary, due to the high-rise, multi-level and complex path of the gas pipe network, the traditional single-point threshold monitoring and alarm method has problems such as easy data loss, susceptibility to noise interference, insufficient multi-point coupling identification, poor cross-scenario generalization ability, and lag in monitoring and alarm. Summary of the Invention

[0004] Aiming at the problems existing in the above-mentioned prior art, the purpose of the present invention is to provide a gas abnormal condition monitoring and alarm system. This system forms an end-to-end closed-loop structure through the bidirectional coupling of data streams and model parameter streams, and is used for the monitoring and early warning of gas abnormal conditions, effectively solving the problems such as easy data loss, susceptibility to noise interference, insufficient multi-point coupling recognition, poor cross-scene generalization ability, and lag in monitoring and alarm existing in the current gas pipeline network monitoring.

[0005] The purpose of the present invention is achieved through the following technical solutions: A gas abnormal condition monitoring and alarm system uses a monitoring system including a data preprocessing module, a spatio-temporal feature module, an ST-CGN network module, an edge intelligence optimization module, and a result processing module; among them: The data preprocessing module processes the data collected by sensors in the pipeline through methods of sliding window sharding, missing value interpolation, noise filtering, normalization, and encoding in sequence; The spatio-temporal feature module extracts spatial correlation features, time dynamics features, and causal propagation features from the preprocessed data of each sliding window shard, and splices them to form a three-dimensional feature tensor; The ST-CGN network module outputs a multi-condition probability vector corresponding to the sliding window shard through the input of the three-dimensional feature tensor; The edge intelligence optimization module realizes the edge optimization of the ST-CGN network module through MAML (Model-Agnostic Meta-Learning), knowledge distillation, model pruning, and quantization; The result processing module obtains the probability of the corresponding condition through the multi-condition probability vector of the ST-CGN network module optimized by the edge intelligence optimization module, and triggers an alarm judgment according to a preset threshold.

[0006] Based on a further optimization of the above solution, the specific sliding window sharding is as follows: aiming at the rapid dynamic changes presented by the gas pipeline network pressure signal in a short time, a sliding window sharding strategy with a length of T w and a step of S is adopted. Each window W k is:

[0007] In the formula: t k represents the starting sampling moment of the k th window; By capturing the dynamic redundancy and continuity between consecutive windows, the risk of missed detection of initial leaks or blocked pulses is reduced; The missing value interpolation is specifically as follows: For each sliding window, to cope with the sampling gaps caused by occasional sensor disconnections or packet losses, a linear interpolation strategy is adopted:

[0008] In the formula: P i (t) represents the pressure signal value of the t th moment and the i th section; The noise filtering is specifically as follows: For the high-frequency random interference in the gas pipeline network, the exponential weighted moving average is used to smooth the original signal:

[0009] In the formula: represents the smoothing coefficient; x(t) represents the original signal value; represents the signal value after the exponential weighted moving average processing; The normalization and encoding are specifically as follows: First, standardize the dataset of missing value interpolation and noise filtering:

[0010] In the formula: x norm represents the x i value of the sample after standardization; x i represents the i th sample in the dataset; represents the mean of the dataset; represents the standard deviation of the dataset; Then, perform One-Hot encoding on the standardized values to eliminate the influence of different dimensions on model training.

[0011] Based on the further optimization of the above scheme, the spatial correlation features include the instantaneous pressure difference between two nodes and the relative amplitude :

[0012] The time dynamics features include the average value of the sliding window data , the standard deviation of the sliding window data , the maximum volatility of the sliding window data and the fluctuation direction of the sliding window data :

[0013] In the formula: Indicates a specific moment within the sliding window for calculating the fluctuation direction (e.g., the last sampling moment within the window, used to compare with the previous moment to determine the pressure fluctuation direction); The causal propagation features include causal strength Synchronization with the temporal trend :

[0014]

[0015] In the formula: Indicates the pressure change of node k in the i -th window; Cov Indicates covariance, Var Indicates variance; Finally, the spatial correlation features F k and the time dynamics features F s are concatenated with the causal propagation features F y through channel concatenation to form a three-dimensional feature tensor X k :

[0016] Where: N Indicates the number of pipeline network nodes; U Indicates the total dimension of the three types of features (e.g., if the spatial correlation features are 2D, the time dynamics features are 5D, and the causal propagation features are 2D, then the total dimension is 9D); R Indicates the number of time steps (i.e., the number of sampling points within the sliding window); Concat() Indicates channel concatenation.

[0017] Based on the further optimization of the above scheme, the specific process for the ST-CGN network module to perform output prediction according to the input three-dimensional feature tensor X k is as follows: First, input the three-dimensional feature tensor X k , generate causal features using causal-driven modeling through a non-linear structural equation model; then, use adaptive modal fusion to select TCN (Temporal Convolutional Network) or GCN (Graph Convolutional Network) according to the node degree; after that, stack the spatio-temporal convolutional layer and LSTM (Long Short-Term Memory Network) to output the prediction result; finally, perform counterfactual localization intervention, calculate the effect size, and locate the key nodes.

[0018] Based on the further optimization of the above solution, the specific process of generating causal features by using causal-driven modeling and through a non-linear structural equation model is as follows: First, establish a non-linear structural equation model:

[0019] In the formula: Pa(i) represents the set of parent nodes of node i ; represents a non-linear activation function (such as the Sigmoid function); represents the parent node j at time ( t - t c )'s feature (such as pressure value); represents node j to node i 's pressure conduction physical parameters (such as pipe resistance, flow rate, etc.); represents a non-linear mapping function for fusing features and physical parameters (for example: MLP, multi-layer perceptron); represents a random error term; represents the dynamic causal effect intensity;

[0020] In the formula: I q , I k both represent weight matrices; among them: I q is used to perform a linear transformation on the feature of node i at the current moment, map it to a specific feature space, and generate a query value to capture the key information of the current node state; act on the feature j of the parent node and the concatenated vector of the pressure conduction physical parameter , and generate a key value through a linear transformation for calculating the relevance with the query vector; d represents standardization; represents a point set; Softmax() represents Softmax function; Ensure that the causal graph is acyclic through a smooth acyclic loss, avoid the set of parent nodes i of node Pa(i) from falling into a causal loop, and ensure the logicality of the causal relationship in the model ( R dag represents the smooth acyclic loss function):

[0021] Wherein: tr() represents the trace of a matrix (i.e., the sum of diagonal elements); N represents the number of nodes; represents the Hadamard product; I W represents the weight matrix, whose elements are defined by I ij as follows:

[0022] Wherein: represents the indicator function, that is, if node j belongs to the parent node set of node i then is 1, otherwise is 0; represents L the L2 norm, i.e., the Euclidean norm.

[0023] Based on the further optimization of the above solution, the specific selection of TCN or GCN according to the node degree by using adaptive modal fusion is as follows: Define the modal fusion strategy of node i as follows:

[0024] Among them, TCN is used for low node degree scenarios with linear or tree-like topologies to capture local temporal patterns:

[0025] Wherein: DilatedConv() represents the dilated convolution operation for capturing local temporal patterns; represents the feature sequence sampled by node i in the time interval t - K, t with a dilation factor D ; W TCN represents the weight parameter of the TCN layer for convolution operation; K represents the convolution kernel size; GCN is used for high node degree scenarios with complex or cyclic structures to aggregate neighbor information:

[0026] Wherein: represents the normalized adjacency matrix element, which includes the causal weight and represents the connection relationship and weight between node j and node i ; W GCN represents the weight matrix of the GCN layer for feature transformation; N(i)Represents the set of neighbor nodes of the node i and is used to aggregate neighbor information; ReLU() Represents the activation function.

[0027] Based on the further optimization of the above scheme, the specific steps of performing counterfactual localization intervention, calculating the effect size, and locating the key nodes are as follows: First, an abnormal intervention is applied to the node s to generate counterfactual features :

[0028] In the formula: Represents the intervention increment, which follows a normal distribution; Then, the predicted value after the intervention is obtained, and the prediction difference before and after the intervention is compared to quantify the influence of the node s on the downstream node j :

[0029] In the formula: Represents the original predicted value; Represents the predicted value after the counterfactual localization intervention; Finally, the node with the largest effect size is selected as the key intervention target to achieve the localization of the key intervention node :

[0030] In the formula: PD Represents the set of downstream nodes; Represents within the value range of s finding the function that maximizes the subsequent expression.

[0031] Based on the further optimization of the above scheme, the MAML fast adaptation is used to improve the cross-scenario generalization ability and quickly adapt to new scenarios:

[0032] In the formula: P local Represents the local abnormal sample dataset; L cls Represents the classification loss function based on the local dataset; Represents the initial parameters of the model; Represents the gradient operator with respect to the initial parameters; Represents the learning rate; Represents the updated model parameters; Knowledge distillation transfers knowledge from the teacher model to the student model, compresses the model size, and improves the inference efficiency:

[0033] In the formula: KL represents divergence, which is used to measure the difference between the output distributions of the teacher model and the student model after a specific transformation; represents an activation function, which is used to convert the output of the model into a probability distribution (for example: Softmax function); represents the output of the teacher model for the input data X ; represents the output of the student model for the input data X ; T represents a parameter used to control the smoothness of the probability distribution; Model pruning and quantization combine structured pruning and INT8 quantization to generate an efficient edge model:

[0034] In the formula: I ys represents the original weight matrix; represents the weight matrix after the pruning operation; represents the pruning operation function; represents the pruning rate, and the pruning rate in the present invention is 0.2.

[0035] Based on the further optimization of the above solution, the input of the result processing module is a multi-condition probability vector G = G 0, G 1,…, G 9], and the result processing module calculates the multi-condition probability vector through the Softmax function to obtain the probability corresponding to each condition g i (); The first-level leakage alarm threshold i=0,1,…,9 1 and the second-level blockage alarm threshold M 2 are set respectively: M If max{ g 4 ,g 5 ,g 6 ,g 7} > M 1, a first-level leakage alarm is triggered; If max{ g 8 ,g 9} > M 2, a second-level blockage alarm is triggered; If neither of the above two situations occurs, it is judged as a normal condition and no alarm is triggered.

[0036] The following are the technical effects of the solution of the present invention: ​The data preprocessing module of this system processes gas data by means of data preprocessing methods such as sliding window slicing, missing value interpolation, noise filtering, normalization, and encoding in sequence, realizing the continuity of short-term dynamic capture, effectively solving problems such as incomplete original time-series data and insufficient signal-to-noise ratio in the gas monitoring process, and ensuring that subsequent feature calculations are based on complete and low-noise time-series data; through the time-series feature module that splices spatial correlation features, time dynamics features, and causal propagation features, a three-dimensional feature tensor is created, thereby accurately depicting the multi-point coupling and dynamic evolution in the gas pipeline network, effectively avoiding problems such as insufficient recognition of multi-point coupling in high-level pipeline networks and difficulty in capturing micro-vibrations in the initial stage of leakage. At the same time, this system, through the ST-CGN network module, not only realizes the combination of physical parameters and data-driven, realizes causal interpretability, but also dynamically selects modes according to node degrees, effectively balances computational efficiency and model expression ability, and improves the real-time response ability of the model; in addition, the ST-CGN network module also obtains key nodes through counterfactual reasoning, forms active operation and maintenance decisions, and improves the robustness of the overall model under complex working conditions. Through the edge intelligence optimization module of MAML fast adaptation, knowledge distillation, model pruning, and quantization, it not only realizes the scenario adaptive connection of the model in a short time, alleviates the problem of scarce abnormal data, but also improves the anti-forgetting ability on non-stationary data streams and improves the overall stability of the system. Brief Description of the Drawings

[0037] Figure 1 It is a structural block diagram of gas abnormal condition monitoring in an embodiment of the present invention. Detailed Embodiments

[0038] The technical solutions in the embodiments of the present invention will be clearly and completely described below. In the following descriptions, specific details such as specific system structures and technologies are proposed for the purpose of illustration rather than limitation, so as to thoroughly understand the embodiments of the present invention.

[0039] Embodiment 1: A gas abnormal condition monitoring and alarm system adopts a monitoring system including a data preprocessing module, a spatio-temporal feature module, an ST-CGN network module, an edge intelligence optimization module, and a result processing module; among them: The data preprocessing module processes the data collected by sensors in the pipeline through methods such as sliding window slicing, missing value interpolation, noise filtering, normalization, and encoding in sequence; Among them, the sliding window slicing is specifically: aiming at the rapid dynamic changes presented by the gas pipeline network pressure signal in a short moment, a sliding window slicing strategy with a length of T w and a step of S is adopted (in this embodiment, T w = 30 seconds, S= 10 seconds, and the window overlap duration is T w - S = 20 seconds), and each window W k is:

[0040] In the formula: t k represents the starting sampling time of the k th window; By capturing the dynamic redundancy and continuity between consecutive windows, the risk of initial leakage or missed detection of blocked pulses is reduced; The missing value interpolation is specifically as follows: For each sliding window, to cope with the sampling gaps caused by occasional disconnection of sensors or packet loss, a linear interpolation strategy is adopted:

[0041] In the formula: P i (t) represents the pressure signal value of the t th section at the i th moment; The noise filtering is specifically as follows: For the high-frequency random interference in the gas pipeline network, the exponential weighted moving average is used to smooth the original signal:

[0042] In the formula: represents the smoothing coefficient (in this embodiment, = 0.3); x(t) represents the original signal value; represents the signal value after exponential weighted moving average processing; The normalization and encoding are specifically as follows: First, standardize the dataset of missing value interpolation and noise filtering:

[0043] In the formula: x norm represents the x i value of the sample after standardization; x i represents the i th sample in the dataset; represents the mean of the dataset; represents the standard deviation of the dataset; Then, perform One-Hot encoding on the standardized values (using the conventional One-Hot encoding in the art) to eliminate the influence of different dimensions on model training.

[0044] The spatio-temporal feature module extracts spatial correlation features, temporal dynamics features, and causal propagation features from the preprocessed data of each sliding window slice, and splices them to form a three-dimensional feature tensor; The spatial correlation features include the instantaneous pressure difference between two nodes and the relative amplitude :

[0045] The temporal dynamics features include the average value of the sliding window data , the standard deviation of the sliding window data , the maximum volatility of the sliding window data and the fluctuation direction of the sliding window data :

[0046] In the formula: represents a specific moment within the sliding window for calculating the fluctuation direction (for example: the last sampling moment within the window, used to compare with the previous moment to determine the pressure fluctuation direction); The causal propagation features include the causal intensity and the synchronization with the temporal trend :

[0047]

[0048] In the formula: represents the pressure change of node k in the i th window; Finally, splice the spatial correlation features F k , the temporal dynamics features F s and the causal propagation features F y through channel splicing to form a three-dimensional feature tensor X k :

[0049] Among them: N represents the number of pipe network nodes; U represents the total dimension of the three types of features (for example: 2 dimensions for spatial correlation features, 5 dimensions for temporal dynamics features, 2 dimensions for causal propagation features, then the total dimension is 9 dimensions); Rrepresents the number of time steps (i.e., the number of sampling points within the sliding window); Concat() represents channel concatenation.

[0050] The ST-CGN network module takes the input of a three-dimensional feature tensor X k , and outputs the multi-condition probability vectors corresponding to the sliding window slices. The specific process is as follows: First, input the three-dimensional feature tensor X k , and use causal-driven modeling to generate causal features through a non-linear structural equation model. Specifically: First, establish a non-linear structural equation model:

[0051] In the formula: Pa(i) represents the set of parent nodes of node i ; represents the non-linear activation function (e.g., Sigmoid function); represents the feature of the parent node j at time ( t - t c ) (such as pressure value); represents the physical parameter of pressure conduction from node j to node i (such as pipeline resistance, flow rate, etc.); represents the non-linear mapping function used to fuse features and physical parameters (e.g., MLP, multi-layer perceptron); represents the random error term; represents the dynamic causal effect intensity;

[0052] In the formula: I q , I k both represent weight matrices; among them: I q is used to perform a linear transformation on the feature of node i at the current moment, map it to a specific feature space, and generate a query value to capture the key information of the current node state; it acts on the feature j of the parent node and the concatenated vector of the pressure conduction physical parameter , and generates a key value through a linear transformation for calculating the correlation with the query vector; d represents standardization; represents a point set; Softmax() represents Softmax function; Ensure that the causal graph is acyclic by means of a smooth acyclic loss, avoiding the parent node set of node i from falling into a causal loop and ensuring the logicality of the causal relationship in the model ( Pa(i) dag R dag denotes the smooth acyclic loss function):

[0053] In the formula: tr() denotes the trace of the matrix (i.e., the sum of the diagonal elements); N denotes the number of nodes; denotes the Hadamard product; I W denotes the weight matrix, the elements of which are defined by I ij :

[0054] In the formula: denotes the indicator function, i.e., if node j belongs to the parent node set of node i , then is 1, otherwise is 0; denotes L the 2-norm, i.e., the Euclidean norm.

[0055] Then, using adaptive modality fusion, select TCN (Temporal Convolutional Network) or GCN (Graph Convolutional Network) according to the node degree, specifically: Define the modality fusion strategy of node i :

[0056] Among them, TCN is used for low-node-degree scenarios with linear or tree-like topologies to capture local temporal patterns: ; In the formula: DilatedConv() denotes the dilated convolution operation for capturing local temporal patterns; denotes the feature sequence sampled by node i in the time interval t - K, t with a dilation factor of D ; W TCN denotes the weight parameter of the TCN layer for convolution operation; K denotes the convolution kernel size; GCN is used for high-node-degree scenarios with complex or cyclic structures to aggregate neighbor information:

[0057] In the formula: represents the element of the normalized adjacency matrix, which contains the causal weight , represents the node j and the node i connection relationship and weight; W GCN represents the weight matrix of the GCN layer, used for feature transformation; N(i) represents the node i set of neighbor nodes, used to aggregate neighbor information; ReLU() represents the activation function.

[0058] After that, stack the spatio-temporal convolutional layer and LSTM (Long Short-Term Memory Network) to output the prediction result.

[0059] Finally, through counterfactual localization intervention, calculating the effect size and localizing the key nodes, specifically: First, apply an abnormal intervention to the node s to generate the counterfactual feature :

[0060] In the formula: represents the intervention increment, which follows a normal distribution; Then, obtain the predicted value after the intervention, and compare the prediction differences before and after the intervention to quantify the influence of the node s on the downstream node j :

[0061] In the formula: represents the original predicted value; represents the predicted value after the counterfactual localization intervention; Finally, select the node with the largest effect size as the key intervention target to achieve the localization of the key intervention node :

[0062] In the formula: PD represents the set of downstream nodes; represents within the range of the value of s find the function that makes the subsequent expression reach the maximum value.

[0063] The edge intelligence optimization module realizes the edge optimization of the ST-CGN network module through MAML (Model-Agnostic Meta-Learning), knowledge distillation, model pruning and quantization; MAML fast adaptation is used to improve cross-scenario generalization ability and quickly adapt to new scenarios:

[0064] Where: P local represents the local abnormal sample dataset; L cls represents the classification loss function based on the local dataset; represents the initial model parameters; represents the gradient operator with respect to the initial parameters; represents the learning rate; represents the updated model parameters; Knowledge distillation transfers knowledge from the teacher model to the student model, compresses the model size, and improves the inference efficiency:

[0065] Where: KL represents the divergence, which is used to measure the difference between the output distributions of the teacher model and the student model after a specific transformation; represents the activation function, which is used to convert the output of the model into a probability distribution (for example: Softmax function); represents the output of the teacher model for the input data X; represents the output of the student model for the input data X; T represents the smoothness used to control the probability distribution; Model pruning and quantization combine structured pruning and INT8 quantization to generate an efficient edge model:

[0066] Where: I ys represents the original weight matrix; represents the weight matrix after the pruning operation; represents the pruning operation function; represents the pruning rate, and the pruning rate in the present invention is 0.2.

[0067] The result processing module obtains the probability of the corresponding working condition through the multi-working-condition probability vector of the ST-CGN network module optimized by the edge intelligence optimization module, and triggers an alarm judgment according to a preset threshold; specifically: the input of the result processing module is the multi-working-condition probability vector G = G 0, G 1,…, G 9], and the result processing module passes through Softmax The function calculates the multi-working-condition probability vector to obtain the probability corresponding to each working conditiong i ( i=0,1,…,9 ); Set the primary leakage alarm threshold M 1 and the secondary blockage alarm threshold M 2 (in this embodiment, M 1 is 0.9, M 2 is 0.85): If max{ g 4 ,g 5 ,g 6 ,g 7} > M 1, then trigger the primary leakage alarm; If max{ g 8 ,g 9} > M 2, then trigger the secondary blockage alarm; If neither of the above two situations occurs, it is judged as normal working condition and no alarm is triggered.

[0068] Embodiment 2: As another preferred implementation of the technical solution of the present invention, on the basis of the above Embodiment 1 solution, the spatio-temporal convolutional layer in the ST-CGN network module simultaneously models the spatial neighborhood relationship and the local dependence of the time sliding window, maintaining time causality, specifically: Spatio-temporal separable convolution:

[0069] In the formula: b represents the bias term; represents the activation function, which is used to introduce non-linearity; Among them, represents the spatial convolution, which is used to aggregate the neighbor node features:

[0070] In the formula: represents the dynamic causal adjacency matrix, that is, at time t node i and neighbor j 's connection relationship; W s represents the weight matrix of the spatial convolution; represents node j at time t 's feature; represents the time convolution, using dilated causal convolution to prevent future information leakage:

[0071] In the formula: represents the weight of the time convolution, kIndicates the position within the convolutional kernel; Indicates a node i At The feature at time ); Indicates the convolutional kernel size; D Indicates the dilation factor; Split the spatio-temporal convolutional output of each node Into independent time series by node and input them into the LSTM network. The core structure of the LSTM network includes a cell state, a forget gate, an input gate, and an output gate. The cell state: stores long-term information and controls the flow of information through the forget gate, input gate, and output gate. The forget gate: determines which historical information to discard. The input gate: determines which parts of the current input are stored in the cell state. The output gate: determines which parts of the cell state are used as the current output.

Claims

1. A gas abnormal condition monitoring and alarming system, characterized in that: Adopt a monitoring system including a data preprocessing module, a spatio-temporal feature module, an ST-CGN network module, an edge intelligence optimization module, and a result processing module; where: The data preprocessing module processes the data collected by the sensors in the pipeline through methods of sliding window sharding, missing value interpolation, noise filtering, normalization, and encoding in sequence; The spatio-temporal feature module extracts spatial correlation features, temporal dynamics features, and causal propagation features from the preprocessed data of each sliding window shard, and splices them to form a three-dimensional feature tensor; The ST-CGN network module outputs a multi-condition probability vector corresponding to the sliding window shard through the input of the three-dimensional feature tensor; The edge intelligence optimization module realizes the edge optimization of the ST-CGN network module through MAML fast adaptation, knowledge distillation, model pruning, and quantization; The result processing module obtains the probability of the corresponding condition through the multi-condition probability vector of the ST-CGN network module optimized by the edge intelligence optimization module, and triggers an alarm judgment according to the preset threshold.

2. The gas abnormal condition monitoring and alarming system according to claim 1, wherein: The specific sliding window slicing is as follows: For the rapid dynamic changes presented by the gas pipeline network pressure signal in a short period of time, a sliding window slicing strategy with a length of T w and a step of S is adopted. Each window W k is: Wherein: t k represents the starting sampling time of the k th window; By capturing the dynamic redundancy and continuity between consecutive windows, the risk of missed detection of initial leakage or blocking pulses is reduced; The missing value interpolation is specifically as follows: for each sliding window, to cope with the sampling gaps caused by occasional disconnection of sensors or packet loss, a linear interpolation strategy is adopted: In the formula: P i (t) represents the t pressure signal value of the i nth section at the The noise filtering is specifically as follows: for the high-frequency random interference in the gas pipeline network, the exponential weighted moving average is used to smooth the original signal: In the formula: represents the smoothing coefficient; x(t) represents the original signal value; represents the signal value after exponential weighted moving average processing; The normalization and encoding are specifically as follows: First, standardize the dataset of missing value interpolation and noise filtering: Wherein: x norm represents a sample x i the value after standardization; x i represents the i th sample in the dataset; represents the mean of the dataset; represents the standard deviation of the dataset; Then, perform One-Hot encoding on the standardized values to eliminate the influence of different dimensions on model training.

3. The gas abnormal condition monitoring and alarming system according to claim 2, wherein: The spatial correlation features include the instantaneous pressure difference between two nodes and the relative amplitude : The time dynamics characteristics include the average value of the sliding window data , the standard deviation of the sliding window data , the maximum volatility of the sliding window data , and the fluctuation direction of the sliding window data : In the formula: represents a specific moment for calculating the fluctuation direction within the sliding window; Causal propagation characteristics include causal strength Synchronization with the temporal trend : In the formula: represents k the pressure change of the node i in the k -th window; Finally, the spatial correlation features F k , the temporal dynamics features F s and the causal propagation features F y are concatenated through channels to form a three-dimensional feature tensor X k : Wherein: N represents the number of pipe network nodes; U represents the total dimension of three types of features; R represents the number of time steps; Concat() represents channel splicing.

4. The gas abnormal condition monitoring and alarming system according to claim 2, characterized in that: The ST-CGN network module performs output prediction based on the input three-dimensional feature tensor X k The specific process for output prediction is as follows: First, input the three-dimensional feature tensor X k , and use causal-driven modeling to generate causal features through a non-linear structural equation model; Then, use adaptive modal fusion to select TCN or GCN according to the node degree; after that, stack spatio-temporal convolutional layers and LSTM to output the prediction result; finally, perform counterfactual localization intervention, calculate the effect size, and locate the key nodes.

5. The gas abnormal condition monitoring and alarming system according to claim 4, characterized in that: The specific method of using causal-driven modeling to generate causal features through a non-linear structural equation model is as follows: First, establish a non-linear structural equation model: In the formula: Pa(i) represents the set of parent nodes i ; represents the non-linear activation function; represents the parent node j at time( t - t c ); represents the physical parameter of pressure conduction from node j to node i ; represents the non-linear mapping function; represents the random error term; represents the strength of the dynamic causal effect; In the formula: I q and I k both represent weight matrices; where: I q is used to linearly transform the features of node i at the current moment, map them to a specific feature space, and generate query values to capture the key information of the current node state; act on the features of the parent node j and the splicing vector of the pressure conduction physical parameter to generate key values through linear transformation; d represents standardization; represents a point set; Softmax() represents Softmax function; Ensure acyclicity of the causal graph by smooth acyclic loss to avoid the set of parent nodes of i from falling into a causal loop and ensure the logic of causal relationships in the model: Pa(i) ​ In the formula: tr() represents the trace of the matrix; N represents the number of nodes; represents the Hadamard product; I W represents the weight matrix, the elements of which are defined by I ij as follows: In the formula: represents the indicator function, that is, if the node j belongs to the set of parent nodes of the node i , then is 1, otherwise is 0; represents L the 2-norm, that is, the Euclidean norm.

6. The gas abnormal condition monitoring and alarming system according to claim 4, characterized in that: The specific method of using adaptive modal fusion to select TCN or GCN according to the node degree is as follows: Define the node i 's modal fusion strategy: Among them, TCN is used for low-node-degree scenarios with linear or tree-like topologies to capture local temporal patterns: Wherein: DilatedConv() represents the dilated convolution operation; represents the node i within the time interval t - K, t , with the dilation factor D sampled feature sequence; W TCN represents the weight parameter of the TCN layer; K represents the convolution kernel size; GCN is used for high-node-degree scenarios with complex or ring structures to aggregate neighbor information: Wherein: represents the element of the normalized adjacency matrix, which contains the causal weight , represents the node j and the node i connection relationship and weight; W GCN represents the weight matrix of the GCN layer; N(i) represents the set of neighbor nodes of the node i ; ReLU() represents the activation function.

7. A gas abnormal condition monitoring and alarming system according to claim 4, characterized in that: The specific method of performing counterfactual localization intervention, calculating the effect size, and locating the key nodes is as follows: First, an abnormal intervention is applied to the node s to generate counterfactual features : Wherein: represents the intervention increment, which conforms to the normal distribution; Then, obtain the predicted values after the intervention, compare the prediction differences before and after the intervention, and quantify the nodes s for downstream nodes j : In the formula: represents the original predicted value; represents the predicted value after counterfactual location intervention; Finally, select the node with the largest effect size as the key intervention target to achieve the key intervention node Location: In the formula: PD represents the set of downstream nodes; represents within the range of values of s find the function that maximizes the subsequent expression.

8. The gas abnormal condition monitoring and alarming system according to claim 4, characterized in that: The input of the result processing module is a multi-condition probability vector G = G 0, G 1,…, G 9], and the result processing module calculates the multi-condition probability vector through Softmax a function to obtain the probability corresponding to each condition g i ( i=0,1,…,9 ); respectively set the primary leakage alarm threshold M 1 and the secondary blockage alarm threshold M 2: If max{ g 4 ,g 5 ,g 6 ,g 7}> M 1, a first-level leakage alarm is triggered; If max{ g 8 ,g 9}> M 2, a secondary blocking alarm is triggered; If neither of the above two situations occurs, it is judged as a normal condition and no alarm is triggered.

Images