CN-121093160-B - Early warning method for monitoring abnormal state of gas pipe network

CN121093160BCN 121093160 BCN121093160 BCN 121093160BCN-121093160-B

Abstract

The invention provides an early warning method for monitoring abnormal states of a gas pipe network, which relates to the field of gas monitoring and early warning, and adopts a monitoring system comprising a data preprocessing module, a space-time characteristic module, an ST-CGN network module, an edge intelligent optimization module and a result processing module, wherein the ST-CGN network module outputs multi-working condition probability vectors corresponding to sliding window slicing through the input of a three-dimensional characteristic tensor, simultaneously, a space-time convolution layer in the ST-CGN network module models the local dependence and the holding time causality of a space neighborhood relation and a time sliding window at the same time, and the edge intelligent optimization module realizes the edge optimization of the ST-CGN network module through MAML rapid adaptation, knowledge distillation, model pruning and quantization. The method effectively solves the problems of easy data deletion, easy noise interference, insufficient multi-point coupling identification, poor scene generalization capability, hysteresis of monitoring alarm and the like existing in the existing gas pipe network monitoring method.

Inventors

YANG BO
XU JING
ZHU ZHIQUAN
ZHANG TINGHUA
HAN LEI
ZHU ZHIQIN
CAO LONGHAN
LONG XIANHAN
YANG YONG
WAN YUFENG
HUANG JUN
LUO HAOLIN
XIE WEI

Assignees

重庆创元智能仪表系统有限公司

Dates

Publication Date: 20260512
Application Date: 20250612

Claims (8)

1. A warning method for monitoring abnormal states of a gas pipe network is characterized by adopting a monitoring system comprising a data preprocessing module, a space-time characteristic module, an ST-CGN network module, an edge intelligent optimization module and a result processing module, wherein: The data preprocessing module processes the data acquired by the sensor in the pipeline sequentially through sliding window slicing, missing value interpolation, noise filtering, normalization and encoding; the space-time characteristic module extracts space-related characteristics, time dynamics characteristics and causal propagation characteristics from the preprocessing data of each sliding window slice, and splices the space-time characteristics, the time dynamics characteristics and the causal propagation characteristics to form a three-dimensional characteristic tensor; the method comprises the steps of inputting three-dimensional characteristic tensors by an ST-CGN network module, outputting multi-working condition probability vectors corresponding to sliding window slicing, simultaneously modeling a spatial neighborhood relation and local dependence and time causality of a time sliding window by a space-time convolution layer in the ST-CGN network module, realizing edge optimization of the ST-CGN network module by an edge intelligent optimization module through MAML (fast adaptation, knowledge distillation, model pruning and quantization), obtaining probability of corresponding working conditions by a result processing module through the multi-working condition probability vectors of the ST-CGN network module optimized by the edge intelligent optimization module, and triggering alarm judgment according to a preset threshold; The ST-CGN network module performs output prediction according to the input three-dimensional feature tensor X k , wherein the specific flow comprises the steps of firstly inputting the three-dimensional feature tensor X k , generating causal features through a non-linear structural equation model by utilizing causal driving modeling, then performing TCN or GCN selection according to node degree by utilizing self-adaptive modal fusion, then superposing a space-time convolution layer and an LSTM, outputting a prediction result, and finally positioning intervention, calculating an effect quantity and positioning key nodes through inverse facts; The space-time convolution layer in the ST-CGN network module simultaneously models the local dependence of the space neighborhood relation and the time sliding window and maintains the time causality comprises the following steps: Wherein b represents a bias term; Representing an activation function for introducing nonlinearities; Wherein, the Representing a spatial convolution for aggregating neighbor node features: wherein: representing a dynamic causal adjacency matrix, namely, the connection relation between a node i and a node j of the moment t; W s represents the weight matrix of the spatial convolution; the characteristic of the node j at the time t; Representing a temporal convolution, using a causal convolution of dilation to prevent future information leakage: wherein: Weights representing the time convolution, k representing the position within the convolution kernel; Indicating that node i is in The characteristics of the moment; D represents the expansion factor; Output of space-time convolution of each node Splitting into independent time sequences according to nodes, and inputting the independent time sequences into an LSTM network, wherein the core structure of the LSTM network comprises a cell state, a forgetting gate, an input gate and an output gate.
2. The method for early warning of abnormal state of gas pipe network according to claim 1, wherein the sliding window slicing is characterized in that for rapid dynamic change of gas pipe network pressure signals in short time, a sliding window slicing strategy with a length of T w and a step of S is adopted, and each window W k is as follows: wherein t k represents the initial sampling time of the kth window; The risk of initial leakage or blocking pulse omission is reduced by capturing dynamic redundancy and continuity between continuous windows; The missing value interpolation is specifically that for each sliding window, a linear interpolation strategy is adopted for coping with sampling gaps caused by accidental disconnection of a sensor or loss of a data packet: Wherein P i (t) represents the pressure signal value of the ith node at the t-th time; The noise filtering is specifically that for high-frequency random interference in a gas pipe network, an original signal is smoothed by using an exponential weighted moving average: wherein: x (t) represents the original signal value; representing the signal value after the exponentially weighted moving average processing; The normalization and coding is specifically that firstly, the data set of missing value interpolation and noise filtering is normalized: Where x norm represents the normalized value of sample x ci , x ci represents the ci-th sample in the dataset; representing a mean value of the dataset; Representing standard deviation of the dataset; and then, carrying out One-Hot coding on the standardized value, and eliminating the influence of different dimension on model training.
3. The method for monitoring abnormal state of gas pipe network according to claim 2, wherein the spatial correlation characteristic comprises instantaneous pressure difference of two nodes And relative amplitude of : The time dynamics features include sliding window data averages Standard deviation of sliding window data Maximum fluctuation rate of sliding window data Fluctuation direction of sliding window data : Wherein: representing sampling moments in the sliding window for calculating the direction of the fluctuation; Causal propagation characteristics include causal strength Synchronization with time sequence trend : Wherein: representing the pressure variation of the node i in the kth window; Finally, the spatial correlation feature F k , the temporal dynamics feature F s and the causal propagation feature F y are spliced by channels to form a three-dimensional feature tensor X k : 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 method for early warning for monitoring abnormal state of gas pipe network according to claim 3, wherein the generating causal features by causal driving modeling and nonlinear structural equation model comprises the steps of firstly, establishing a nonlinear structural equation model: Wherein Pa (i) represents a parent node set of the node i; representing a nonlinear activation function; features of node j at time (t-t c ); representing pressure-conducting physical parameters from node j to node i; representing a nonlinear mapping function; Representing a random error term; Representing the intensity of dynamic causal effects; Wherein I q 、I k is a weight matrix, I q is used for performing linear transformation on the characteristics of the current moment of the node I, mapping the characteristics to a characteristic space obtained after the linear transformation, generating a query value to capture key information of the current node state, and acting on the characteristics of the node j Physical parameters of pressure conduction Generating a key value through linear transformation, wherein d represents standardization; Softmax () represents the Softmax function; The causal graph is ensured to be acyclic through smooth acyclic loss, the father node set Pa (i) of the node i is prevented from being trapped into causal circulation, and the logic of causal relation in the model is ensured: Wherein tr () represents the trace of the matrix, N represents the number of nodes; representing the Hadamard product, I W represents a weight matrix whose elements are defined by I ij : wherein: representing an indication function, i.e. if node j belongs to the parent node set of node i 1, Otherwise, then Is 0; The L 2 norm, i.e. euclidean norm, is represented.
5. The method for early warning for monitoring abnormal state of gas pipe network according to claim 3, wherein the selecting TCN or GCN according to node degree by utilizing self-adaptive mode fusion comprises defining mode fusion strategy of node i: The TCN is used for low-node degree scenes of linear or tree topology and capturing local time sequence modes: wherein DilatedConv () represents an expansion convolution operation; The characteristic sequence of the node i sampled by the expansion factor D in the time interval [ t-K, t ] is represented, W TCN represents the weight parameter of the TCN layer, K represents the convolution kernel size; The GCN is used for high node degree scenes of complex or ring structures and aggregating neighbor information: wherein: representing normalized adjacency matrix elements containing causal weights The method comprises the steps of representing the connection relation and the weight of a node j and a node i, W GCN representing a weight matrix of a GCN layer, N (i) representing a neighbor node set of the node i, and ReLU () representing an activation function.
6. The method for early warning for monitoring abnormal state of gas pipe network according to claim 3, wherein said steps of locating intervention, calculating effect quantity and locating key node by means of counterfactual feature comprise firstly, applying abnormal intervention to node s to generate counterfactual feature : Wherein: representing an intervention delta that conforms to a normal distribution; then, the predicted value of the stem prognosis is obtained, the predicted difference before and after intervention is compared, and the influence of the node s on the downstream node j is quantified: wherein: Representing the original predicted value; a predicted value representing a negative fact positioning dry prognosis; finally, selecting the node with the largest effect quantity as a key intervention target and realizing the key intervention node Positioning: wherein PD represents a downstream node set; the representation is within the value range of s, and a function which enables the subsequent expression to take the maximum value is searched.
7. The method for early warning for monitoring abnormal state of gas pipe network according to claim 3, wherein the MAML is adapted to enhance the generalization capability of a cross-scene and adapt to new scene: wherein P local represents the local abnormal sample dataset, L cls represents a classification loss function based on the local dataset; Representing model initial parameters; Representing gradient operators with respect to the initial parameters; Representing a learning rate; Representing the updated model parameters; knowledge distillation transfers knowledge to a student model through a teacher model, compresses model scale and improves reasoning efficiency: KL represents the divergence, is used for measuring the difference between output distribution of teacher model and student model after specific transformation; representing an activation function for converting the output of the model into a probability distribution; representing the output of the teacher model for the input data X; representing the output of the student model for input data X; T represents smoothness for controlling the probability distribution; model pruning and quantization are combined with structured pruning and INT8 quantization to generate an efficient edge model: wherein I ys represents an original weight matrix; Representing a weight matrix after pruning operation; representing a pruning operation function; Representing pruning rate.
8. The method for early warning for monitoring abnormal states of a gas pipe network according to claim 3, wherein the input of the result processing module is multiplexed Kuang Gailv vector G= [ G 0 ,G 1 ,…,G 9 ], the result processing module calculates the probability vector of multiple working conditions through a Softmax function to obtain probabilities G i , i=0, 1..9 corresponding to the working conditions, and a primary leakage alarm threshold M 1 and a secondary blocking alarm threshold M 2 are respectively set: If max { g 4 ,g 5 ,g 6 ,g 7 }＞M 1 , triggering a first-stage leakage alarm, if max { g 8 ,g 9 }＞M 2 , triggering a second-stage blocking alarm, and if the two conditions do not occur, judging that the working condition is normal, and not triggering the alarm.

Description

Early warning method for monitoring abnormal state of gas pipe network The invention relates to a split application of patent application number 2025107796621 and the invention name of a gas abnormal condition monitoring alarm system. Technical Field The invention relates to the technical field of gas monitoring and early warning, in particular to an early warning method for monitoring abnormal states of a gas pipe network. Background The abnormal state monitoring is a core link of the operation and maintenance of the gas pipe network, and aims to identify risks in advance and rapidly process faults through real-time and accurate monitoring so as to avoid accidents such as explosion, gas terminals and the like caused by gas leakage. In the prior art, the abnormal state of the gas pipe network is monitored mainly by arranging various sensors (including pressure, flow, gas concentration, temperature and the like) at monitoring points, and by setting a safety threshold and collecting data in real time, the abnormal state of the gas pipe network is monitored and early warning of the abnormal state of the pipe network is completed. However, with the construction and update of urban modernization, the gas pipe network has formed a crisscross network structure, and the single-point threshold alarming mode of various sensors is very easy to ignore the coupling relation and pressure fluctuation transmission among nodes of the official network, so that abnormal information sources cannot be accurately captured, and the problems of false alarm, delay early warning and the like occur. Meanwhile, the single-point threshold alarming method is only used for comparing an instantaneous value with an average value, high-frequency micro-vibration in the leakage process is difficult to accurately capture, and the problems of missing shear, delayed alarming and the like can also occur. In addition, because the urban crisscross gas pipe network structure (namely the gas pipe network is in a linear, tree-shaped, annular and even mixed topological structure form), the calibration of the safety threshold value is difficult (the safety threshold value is usually based on the fitting of manual experience and a large amount of experimental data), the cost is high, the influence of environmental factors is large, the deviation of the finally obtained safety threshold value is large, the accuracy is low, and the early warning precision and the instantaneity are further influenced. Disclosure of Invention Aiming at the problems existing in the prior art, the invention aims to provide an early warning method for monitoring the abnormal state of a gas pipe network, which is used for monitoring and early warning the abnormal state of the gas pipe network, and can effectively solve the problems of easy data deletion, easy noise interference, insufficient multi-point coupling identification, poor scene crossing generalization capability, hysteresis of monitoring and warning and the like existing in the existing gas pipe network monitoring method. The aim of the invention is achieved by the following technical scheme: An early warning method for monitoring abnormal states of a gas pipe network adopts a monitoring system comprising a data preprocessing module, a space-time characteristic module, an ST-CGN network module, an edge intelligent optimization module and a result processing module, wherein: The data preprocessing module processes data acquired by a sensor in a pipeline sequentially through sliding window segmentation, missing value interpolation, noise filtering, normalization and coding, the space-time characteristic module extracts space correlation characteristics, time dynamics characteristics and causal propagation characteristics from the preprocessed data of each sliding window segmentation and splices the space correlation characteristics, the time dynamics characteristics and the causal propagation characteristics to form a three-dimensional characteristic tensor, the ST-CGN network module outputs multi-working condition probability vectors of the corresponding sliding window segmentation through the input of the three-dimensional characteristic tensor, the edge intelligent optimization module rapidly adapts, knowledge distillation, model pruning and quantization through MAML (Model-Agnostic Meta-Learning), and achieves edge optimization of the ST-CGN network module, and the result processing module obtains the probability of the corresponding working condition through the multi-working condition probability vectors of the ST-CGN network module optimized by the edge intelligent optimization module and triggers alarm judgment according to a preset threshold. Based on further optimization of the scheme, the sliding window slicing specifically comprises the steps that aiming at the rapid dynamic change of a gas pipe network pressure signal in a short time, a sliding window slicing strategy with the length of T w and the step of