CN-122020466-A - Multi-mode model-based gas use abnormal behavior identification method and system

CN122020466ACN 122020466 ACN122020466 ACN 122020466ACN-122020466-A

Abstract

The invention discloses a gas use abnormal behavior identification method and system based on a multi-mode model, and belongs to the technical field of intelligent gas safety and artificial intelligence. The invention builds an explicit gas use structure causal model, and defines a causal relation diagram among hidden variables such as basic gas consumption, high-power equipment start-stop, leakage event, environmental temperature influence and the like. And acquiring multi-mode time sequence data, training a decoupling representation learning model combining a variation self-encoder and causal constraint, and decomposing observation data into independent characteristic components of hidden variables in a causal model of a corresponding structure. When an abnormality occurs, the change of each hidden variable component in the normal state and the abnormal state is compared, the source variable causing the abnormality is positioned, and the influence path of the father node is traced back along the causal graph of the structural causal model, so that the interpretable root cause diagnosis from the abnormality detection is realized. The invention improves the accuracy of abnormality identification, provides operational diagnosis conclusions such as equipment faults or abnormal behaviors and the like, and improves the intelligent level of gas safety operation and maintenance.

Inventors

CHEN XINGMING
HUANG ZHONGFENG
LIU DONGSHUANG
SU JUN

Assignees

重庆燃气集团股份有限公司沙坪坝分公司
重庆沙坪坝交通实业有限公司

Dates

Publication Date: 20260512
Application Date: 20260128

Claims (9)

1. A method for identifying abnormal gas usage based on a multimodal model, the method comprising: constructing a structural causal model of a gas use scene, wherein the structural causal model comprises a plurality of hidden variable nodes and a causal direction and a functional relation defined between the hidden variable nodes; Collecting multi-mode time sequence observation data of gas use of a target user or region and preprocessing the data; The decoupling characterization learning model maps the multi-mode time sequence observation data code to a decoupled hidden space so that each dimension or dimension group of the hidden space corresponds to a specific hidden variable in the structural causal model and is independently controllable; reasoning the real-time or historical multi-mode observation data by using the trained decoupling representation learning model to obtain a corresponding decoupling hidden variable representation; based on the decoupling hidden variable characterization, calculating the deviation degree of the healthy baseline of each hidden variable component, identifying hidden variables with the deviation degree exceeding a preset threshold value through comparative analysis, and marking the hidden variables as suspected root dependent variables; and according to a causal graph defined by the structural causal model, starting from the suspected root dependent variable, performing backward causal tracing, analyzing the state of an upstream parent node variable, and generating a diagnosis report containing root cause positioning and causal path interpretation.
2. The method for identifying abnormal gas usage based on multi-modal model as claimed in claim 1, The hidden variables at least comprise basic gas consumption, high-power equipment starting, leakage events and environmental temperature influences; The multi-mode time sequence observation data at least comprises gas flow time sequence data, household pressure time sequence data, environment temperature data and equipment switch event logs.
3. The method for identifying abnormal gas usage based on multi-modal model as claimed in claim 1, The decoupling characterization learning model adopts a framework based on a variation self-encoder, and introduces a causal constraint loss function.
4. The method for identifying abnormal gas usage based on a multimodal model as claimed in claim 3, wherein the causal constraint comprises: Independent constraint, namely, enabling each dimension to be independent by minimizing mutual information among hidden variable dimensions or maximizing total correlation; Causal intervention constraint, namely simulating human intervention on a certain hidden variable in training data, wherein only the representation of a corresponding dimension is required to be changed after model coding, and other dimensions are kept stable; And (3) prior distribution constraint, namely matching prior distribution conforming to physical meaning of each hidden variable dimension, and carrying out constraint through KL divergence.
5. The method for identifying abnormal gas usage based on multi-modal model as claimed in claim 4, The causal intervention constraint is realized by constructing a sample pair containing known and single factor changes in a training data set, and decoupling the sample pair before and after the same factor changes into a characteristic learning model during training, wherein the difference of coding vectors of the model is mainly concentrated in hidden variable dimensions corresponding to the factor.
6. The method for identifying abnormal gas usage based on multi-modal model as claimed in claim 1, The method comprises the steps of carrying out normalization and alignment processing consistent with a training stage on observation data of a fixed time window acquired in real time or in history to form a standardized input tensor, then inputting the tensor into an encoder network of a deployed variational self-encoder, and carrying out forward propagation reasoning by the encoder to output a structured hidden space vector which is divided into specific dimension groups according to a preset design and respectively and uniquely corresponds to hidden variables defined in SCM.
7. The method for identifying abnormal gas usage based on multi-modal model as claimed in claim 1, The method for establishing the health base line comprises the steps of obtaining characterization of each hidden variable by reasoning through historical data reserved in a training phase in a model training phase, and calculating statistical distribution parameters of each dimension of the hidden variable, wherein the deviation degree calculation adopts a Markov distance or a z-score method based on a score.
8. The method for identifying abnormal gas usage based on multi-modal model as claimed in claim 1, The specific method for the backward causal tracing comprises the steps of searching all direct father nodes in the structural causal model aiming at marked suspected root dependent variables, analyzing whether the characterization of the father node variables is in an abnormal state or whether the historical trend of the father node variables shows degradation or not by utilizing a decoupling characterization learning model, if the father node is abnormal, adding the father node variables as a collaborative root cause, and continuing tracing upwards until the states of the exogenous variables or all the nodes are normal.
9. A multi-modal model based abnormal gas usage behavior recognition system, the system comprising: The system comprises a structural causal model definition module, a causal model generation module and a causal model generation module, wherein the structural causal model definition module is used for constructing a structural causal model of a fuel gas use scene, and comprises a plurality of hidden variable nodes and causal direction and function relations between the hidden variable nodes; The multi-mode data acquisition and preprocessing module is used for acquiring multi-mode time sequence observation data of gas use of a target user or region; The decoupling characterization learning model is used for mapping the multi-modal time sequence observation data code to a decoupling hidden space so that each dimension or dimension group of the hidden space corresponds to a specific hidden variable in the structural causal model and is independently controllable; Jie Ouyin variable representation acquisition module, which is used to utilize the trained decoupling representation learning model to infer real-time or historical multi-modal observation data to obtain the corresponding decoupling hidden variable representation; the root cause analysis module is used for calculating the deviation degree of the healthy base line of each hidden variable component based on the decoupling hidden variable characterization, identifying hidden variables with the deviation degree exceeding a preset threshold value through comparative analysis, and marking the hidden variables as suspected root cause variables; And the diagnostic report generation module is used for carrying out backward causal tracing from the suspected root dependent variable according to a causal graph defined by the structural causal model, analyzing the state of an upstream parent node variable and generating a diagnostic report containing root cause positioning and causal path interpretation.

Description

Multi-mode model-based gas use abnormal behavior identification method and system Technical Field The invention relates to the technical field of gas safety monitoring and artificial intelligence, in particular to a gas use abnormal behavior identification method and system based on a multi-mode model. Background With the popularization of intelligent gas meters of the Internet of things, abnormal gas use detection based on data driving becomes a research hotspot. The prior art mainly focuses on two types, namely a threshold method based on rules or statistics and a mode identification method based on machine learning and deep learning. However, the existing methods suffer from two core drawbacks, first, that is, correlation rather than causal, that existing models (e.g., deep learning models) are good at finding complex correlation patterns from data, but cannot distinguish causal relationships between variables. When an anomaly is detected, the system can only report "an anomaly pattern occurred", but cannot answer "why this anomaly was caused". For example, the abnormal rise in flow may be due to "on hanging furnace" (normal behavior) or "medium pressure pipe leakage" (dangerous failure) which is difficult to distinguish by the existing method. Secondly, the black box decision lacks the interpretability, namely, the detection model based on the deep neural network is like a black box, the decision process is difficult to understand, the operation and maintenance personnel cannot trust any unexplained alarm, and accurate emergency treatment is difficult to carry out according to the alarm. Therefore, a technology capable of not only detecting anomalies with high precision but also explaining the physical or behavioral sources behind the anomalies is urgently needed in the field of gas safety, so as to realize the crossing from 'perception anomalies' to 'understanding anomalies', thereby guiding precise intervention. Disclosure of Invention The embodiment of the invention aims to provide a method for identifying abnormal gas use behavior based on a multi-mode model, which aims to solve the problem that the prior gas abnormality detection technology can only identify abnormal phenomena and is difficult to locate root causes because causal relations among variables cannot be distinguished and model decision processes cannot be interpreted. The embodiment of the invention is realized in such a way that a gas use abnormal behavior identification method based on a multi-mode model comprises the following steps: constructing a structural causal model of a gas use scene, wherein the structural causal model comprises a plurality of hidden variable nodes and a causal direction and a functional relation defined between the hidden variable nodes; Collecting multi-mode time sequence observation data of gas use of a target user or region and preprocessing the data; The decoupling characterization learning model maps the multi-mode time sequence observation data code to a decoupled hidden space so that each dimension or dimension group of the hidden space corresponds to a specific hidden variable in the structural causal model and is independently controllable; reasoning the real-time or historical multi-mode observation data by using the trained decoupling representation learning model to obtain a corresponding decoupling hidden variable representation; based on the decoupling hidden variable characterization, calculating the deviation degree of the healthy baseline of each hidden variable component, identifying hidden variables with the deviation degree exceeding a preset threshold value through comparative analysis, and marking the hidden variables as suspected root dependent variables; and according to a causal graph defined by the structural causal model, starting from the suspected root dependent variable, performing backward causal tracing, analyzing the state of an upstream parent node variable, and generating a diagnosis report containing root cause positioning and causal path interpretation. Another object of the embodiments of the present invention is to provide a system for identifying abnormal gas usage based on a multi-modal model. The system comprises: The system comprises a structural causal model definition module, a causal model generation module and a causal model generation module, wherein the structural causal model definition module is used for constructing a structural causal model of a fuel gas use scene, and comprises a plurality of hidden variable nodes and causal direction and function relations between the hidden variable nodes; The multi-mode data acquisition and preprocessing module is used for acquiring multi-mode time sequence observation data of gas use of a target user or region; The decoupling characterization learning model is used for mapping the multi-modal time sequence observation data code to a decoupling hidden space so that each dimension or dimension group of the hidden space