CN-122020577-A - Scene perception information fusion decision method for industrial field multi-source heterogeneous data

CN122020577ACN 122020577 ACN122020577 ACN 122020577ACN-122020577-A

Abstract

The invention discloses a scene perception information fusion decision method of industrial field multi-source heterogeneous data, which comprises an edge layer, a cloud layer and an interaction layer to form a three-level cooperative framework, and belongs to the technical field of industrial field data processing. The method comprises the specific steps of S1, collecting industrial field multi-source heterogeneous original data by an edge layer and performing self-adaptive preprocessing to obtain cleaned multi-source characteristic data, S2, processing the multi-source characteristic data by the edge layer through a light scene perception model to output a preliminary scene judgment result, S3, constructing and updating a causal relation graph by a cloud layer based on industrial business knowledge and edge layer increment data, processing edge layer uploading data through a causal reasoning fusion model to output a decision result and issuing the decision result through an interaction layer, and S4, optimizing a causal reasoning fusion model and light scene perception model parameters by the cloud layer based on a decision execution result fed back by the edge layer, and issuing optimization parameters to the edge layer through the interaction layer to realize dynamic iteration.

Inventors

HU FENGJUN
FENG YANYONG
GU HANJIE
SUN YULIANG
DAI GUOYONG
FANG JUNWEI

Assignees

杭州和利时自动化有限公司
浙江树人学院

Dates

Publication Date: 20260512
Application Date: 20260415

Claims (10)

1. The scene perception information fusion decision method of the industrial field multi-source heterogeneous data is characterized by comprising an edge layer, a cloud layer and an interaction layer to form a three-level collaborative architecture, and the decision method comprises the following specific steps of: s1, collecting industrial field multi-source heterogeneous original data by an edge layer, and performing self-adaptive pretreatment to obtain cleaned multi-source characteristic data; s2, the edge layer processes the multi-source feature data through the light scene perception model, outputs a preliminary scene judgment result, the emergency scene directly triggers a local decision, and the non-emergency scene uploads the multi-source feature data and the preliminary scene judgment result to the cloud layer through the interaction layer; S3, the cloud layer builds and updates a causal relation graph based on industrial business knowledge and edge layer incremental data, processes edge layer uploading data through a causal reasoning fusion model, outputs a decision result and transmits the decision result through an interaction layer; and S4, the cloud layer optimizes the causal reasoning fusion model and the lightweight scene perception model parameters based on the decision execution result fed back by the edge layer, and transmits the optimized parameters to the edge layer through the interaction layer to realize dynamic iteration.
2. The scene perception information fusion decision method of industrial field multi-source heterogeneous data according to claim 1, wherein the adaptive preprocessing in the step S1 adopts a meta-learning driven processing mode, specifically comprises three sub-steps of data type adaptive recognition, dynamic noise filtering and scene missing data complement, and is suitable for the dynamically changing environment of the industrial field without manual intervention.
3. The scene-aware information fusion decision method of industrial field multi-source heterogeneous data according to claim 2, wherein said data type adaptive identification comprises the following sub-steps: S11, extracting three core element characteristics of time sequence, discreteness and space of data; S12, training an identification model by adopting a model-independent element learning algorithm, and quickly adapting to new data types by utilizing a preset part of samples; s13, inputting meta-characteristics of the data to be processed into a trained recognition model, and outputting one or more data types of a time sequence continuous type, a time sequence discrete type, an image type and a text type.
4. The scene-aware information fusion decision method of industrial field multi-source heterogeneous data according to claim 2, wherein said dynamic noise filtering comprises the following sub-steps: SQ11, dynamically adjusting a wavelet threshold according to time sequence sensor data and combining with the running state of equipment, and filtering high-frequency components after wavelet decomposition to keep effective signals; SQ12, aiming at video data, utilizing the background fixed characteristic of the industrial scene, and adopting a Gaussian mixture background modeling and foreground segmentation algorithm to distinguish a moving target from environmental noise; SQ13, aiming at the text log, adopting an industrial term dictionary matching and noise word filtering method, and reserving a core text containing an operation instruction and alarm information.
5. The scene perception information fusion decision method for the industrial field multi-source heterogeneous data is characterized in that the light scene perception model is constructed through a knowledge distillation technology and comprises two stages of cloud multi-mode converter teacher model training and edge layer light chemical raw model distillation, after the teacher model training is finished, the teacher model is compressed in a mode of reducing the number of layers of an encoder, reducing the number of attention heads and reducing the feature vector quantization to obtain a student model, and the higher precision of the model is reserved while the light weight of the model is ensured.
6. The scene perception information fusion decision method of the industrial field multi-source heterogeneous data according to claim 5, wherein in the step S2, scene perception logic of a lightweight scene perception model comprises the steps of calculating contribution degrees of different modal features to scene recognition through a lightweight self-attention mechanism in such a way that mutual information of the modal features and scene classification labels is obtained, the higher the mutual information is, the higher the weight is, the weighted multi-modal features are fused and input into a student model, and a preliminary scene judgment result comprising scene types and confidence degrees is output.
7. The scene perception information fusion decision method of industrial field multi-source heterogeneous data according to claim 1, wherein in the step S3, the construction and updating of the causal relationship graph comprises the following sub-steps: S31, defining data characteristics, equipment states and scene events as causal graph nodes by combining industrial business knowledge, combing causal relations among the nodes and constructing an initial causal relation graph through directed edge connection; S32, based on incremental data uploaded by the edge layer, causal discovery is carried out through condition independence test and causal direction judgment, new causal relation nodes and edges are supplemented, and causal strength weights are updated to ensure causal graph timeliness.
8. The scene perception information fusion decision method of the industrial field multi-source heterogeneous data is characterized in that the processing process of an inference fusion model comprises the steps of filtering features which are not in causal relation with a decision target by adopting a back door adjustment method based on a causal relation graph, calculating causal effect values of all filtered feature nodes on the decision target nodes and taking the causal effect values as fusion weights, weighting and fusing to obtain fusion feature vectors, constructing a Bayesian network inference model based on the fusion feature vectors and the causal relation graph, and outputting three-dimensional decision information comprising decision results, visual interpretation and business text interpretation.
9. The scene-aware information fusion decision method of industrial field multi-source heterogeneous data according to claim 1, wherein the dynamic iteration in step S4 comprises the following sub-steps: S41, labeling decision execution results fed back by the boundary layer, and constructing reward and punishment signals based on the labeling results; S42, optimizing conditional probability parameters of the Bayesian network by adopting a gradient descent method in combination with punishment signals and fusion feature vectors; s43, retraining a cloud teacher model based on the feedback data, and carrying out secondary distillation on the edge layer student model to update parameters; And S44, transmitting the optimized model parameters to an edge layer and a cloud layer in an incremental transmission mode to finish updating.
10. The scene perception information fusion decision method of the industrial field multi-source heterogeneous data is characterized in that in the step S4, the core functions of an interaction layer comprise data encryption transmission, increment synchronization and breakpoint continuous transmission, the encryption transmission adopts a preset encryption algorithm and an industrial Ethernet protocol, a secret key is updated according to a preset period, an edge layer only uploads the preprocessed characteristic data during increment synchronization, an abnormal scene additionally uploads key original data, a cloud layer only issues changed model parameters, the breakpoint continuous transmission is achieved through local caching of the edge end and cloud integrity verification, and the network recovery is followed by automatic continuous transmission of missing data.

Description

Scene perception information fusion decision method for industrial field multi-source heterogeneous data Technical Field The invention belongs to the technical field of industrial field data processing, and particularly relates to a scene perception information fusion decision method of industrial field multi-source heterogeneous data. Background Massive multi-source heterogeneous data can be generated in the industrial field operation process, and the data comprise time sequence sensor data, video monitoring data, PLC control signals, text operation logs and the like, and are core bases for realizing industrial scene perception and intelligent decision. At present, the processing and fusion decision of multi-source heterogeneous data in the industrial field mainly adopts a mode of combining a traditional data fusion method with single architecture deployment, so as to realize the monitoring and decision of industrial scenes. Through deep analysis of the prior art, the following core defects are found, and the decision requirement of high precision, real-time performance and high reliability of an industrial field is difficult to meet: (1) The multi-source data fusion relies on statistical correlation, and misjudgment is easy to occur, wherein the prior art only fuses based on statistical correlation among data, and causal correlation and false correlation cannot be distinguished. For example, there may be a statistical correlation of industrial field environmental humidity to equipment failure, but no direct causality, and existing methods are prone to misjudging it as a valid correlation, resulting in unnecessary downtime loss. (2) The scene perception and decision framework is single, the real-time performance and the accuracy cannot be balanced, the full-edge deployment framework is limited by the computing power of edge equipment, the model accuracy is low, the full-cloud deployment framework needs to transmit full amount of original data, the network delay is high, and the real-time response requirement of an industrial field emergency scene cannot be met. (3) The decision result lacks of interpretability and the landing difficulty is high, the existing AI fusion decision model is mostly a black box model, the output result has no definite business interpretation, and the industrial operators are difficult to understand and trust, so that the actual landing application of the technology on the industrial site is limited. (4) The data preprocessing has poor pertinence and weak adaptability, the existing preprocessing method needs to manually define data types and filtering parameters, cannot adapt to the dynamically changing environment of the industrial field, has poor preprocessing effect and influences the accuracy of the subsequent fusion decision. (5) The AI model training process is incomplete, iterative optimization is not complete, the key contents such as training data, training steps, loss function design and the like of the AI model are not clear in the prior art, and the model performance is difficult to continuously improve due to the lack of a model iterative optimization mechanism based on actual decision feedback of an industrial field. Therefore, there is a need to develop a scene-aware information fusion decision method for multi-source heterogeneous data in industrial fields to solve the problems in the prior art. Disclosure of Invention In order to make up the defects of the prior art, the invention aims to provide a scene perception information fusion decision method for multi-source heterogeneous data in an industrial field, which can realize efficient processing and intelligent decision of the multi-source heterogeneous data in the industrial field, has a simple structure and is convenient to use, so as to solve the problems in the background art. In order to achieve the above purpose, the present invention provides the following technical solutions: The scene perception information fusion decision method of the industrial field multi-source heterogeneous data comprises an edge layer, a cloud layer and an interaction layer, a three-level collaborative architecture is formed, and the decision method specifically comprises the following steps: s1, collecting industrial field multi-source heterogeneous original data by an edge layer, and performing self-adaptive pretreatment to obtain cleaned multi-source characteristic data; s2, the edge layer processes the multi-source feature data through the light scene perception model, outputs a preliminary scene judgment result, the emergency scene directly triggers a local decision, and the non-emergency scene uploads the multi-source feature data and the preliminary scene judgment result to the cloud layer through the interaction layer; S3, the cloud layer builds and updates a causal relation graph based on industrial business knowledge and edge layer incremental data, processes edge layer uploading data through a causal reasoning fusion mo