CN-122020089-A - Operation violation identification method and system based on self-adaptive model arrangement

CN122020089ACN 122020089 ACN122020089 ACN 122020089ACN-122020089-A

Abstract

The application discloses a self-adaptive model programming-based operation violation identification method and system, relates to the technical field of operation safety supervision, and solves the problems that an operation safety supervision system is insufficient in flexibility and is difficult to adapt to different requirements of different scenes. According to the embodiment of the application, the same set of hardware and model pools can support differentiated supervision strategies through the service scene labels and the model arrangement rule base, new scenes can be quickly adapted through different configurations, the expandability and flexibility of the system are improved, the problem of configuration rigidness is solved, the processing assembly line can be dynamically degraded according to real-time computing resources through the available computing resource information of the computing system, lighter model combination is adopted in high load, the overall response speed of the system is ensured, congestion is avoided, and the resource-sensitive self-adaptive scheduling is realized. By adopting the method provided by the embodiment of the application, the flexibility of the operation safety monitoring system can be improved, and the differentiated monitoring requirements of different operation scenes can be better adapted.

Inventors

WANG HENGCHAO
LIU BIN
WANG SHUTIAN
FAN LIBIN
ZHAO YICHENG

Assignees

北京路遥科技有限公司

Dates

Publication Date: 20260512
Application Date: 20260204

Claims (10)

1. The method for identifying the operation violations based on the adaptive model arrangement is characterized by being applied to computing equipment in an operation violating identification system based on the adaptive model arrangement, and comprises the following steps of: Receiving an identification task request, wherein the request at least comprises monitoring video stream data, monitoring audio stream data and a service scene label, and the monitoring video stream data and the monitoring audio stream data are associated; Based on the business scene label, a corresponding initial model programming diagram is obtained from a predefined model programming rule base, wherein the initial model programming diagram is a directed acyclic graph, the initial model programming diagram comprises a plurality of nodes and directed edges connected with the nodes, the nodes comprise a lightweight single-mode model and/or a multi-mode large model, and the directed edges are used for indicating data flow and trigger logic; acquiring available computing resource information of the system, and performing simplified operation on the initial model layout diagram based on the available computing resource information to obtain an executable model layout diagram; Loading and instantiating all nodes in the executable model programming diagram, and scheduling the monitoring video stream data and the audio stream data to flow through each node in sequence for processing based on the executable model programming diagram; and generating a job violation identification report based on the processing result of the output node of the executable model layout graph.
2. The method of claim 1, wherein the simplifying operation at least comprises one of replacing a node corresponding to the multi-modal large model with a sub-graph formed by a plurality of collaborative lightweight single-modal models, or placing a node in parallel branches of the initial modeling graph, which is more accurate and takes more time, in an inactive state.
3. The method of claim 1, wherein the obtaining the corresponding initial modeling map from a predefined modeling rule base based on the business scenario label comprises: searching a model arrangement diagram corresponding to the business scene label in the model arrangement rule library; Searching a model layout diagram corresponding to a father scene label of the service scene label based on a predefined scene label hierarchical inheritance structure and the service scene label under the condition that the model layout diagram corresponding to the service scene label does not exist; Calculating the semantic vector similarity of the business scene label and each scene label in the model arrangement rule base under the condition that the model arrangement graph corresponding to the father scene label does not exist; selecting a model programming diagram corresponding to the target scene label with the highest semantic vector similarity from the target scene labels with the semantic vector similarity larger than a similarity threshold as the initial model programming diagram; And acquiring the initial model programming diagram based on a search result under the condition that the model programming diagram corresponding to the business scene label exists or the model programming diagram corresponding to the father scene label exists.
4. The method of claim 3, wherein the number of parent scene tags is a plurality, the search result includes a base model layout map corresponding to the parent scene tags one to one, the obtaining the initial model layout map based on the search result includes: Acquiring the basic model programming diagram; Identifying a public functional subgraph and a special functional module of the basic layout diagram; based on the business scene label, acquiring corresponding fusion constraint conditions and optimization targets from the rule base; Based on the fusion constraint conditions, constructing an initial fusion graph by taking the public function subgraph as a basic framework and taking each specific function module as a selective branch; and carrying out structural optimization on the initial fusion map based on the optimization target to generate the initial model programming map.
5. The method of claim 4, wherein the fusion constraints include at least functional integrity constraints, real-time constraints, and security rule constraints, and the optimization objective includes at least minimizing end-to-end delays and maximizing critical violation detection coverage.
6. The method of any of claims 1-5, wherein the simplifying the initial modeling map based on the available computing resource information results in an executable modeling map, comprising: Calculating a resource tension score of the system; Under the condition that the resource tension score is larger than a trigger threshold, selecting a simplified operation with highest benefit cost ratio from a predefined simplified strategy base based on the resource requirements and performance files of all nodes in the initial model layout graph, wherein the benefit cost ratio is the ratio of estimated resource saving benefits to estimated recognition precision loss; And iteratively applying the selected simplified operation until the simplified model layout diagram meets the resource constraint condition or reaches the maximum iteration number, so as to obtain the executable model layout diagram.
7. The method according to claim 2, wherein replacing the node corresponding to the multi-modal large model with a sub-graph made up of a plurality of collaborative lightweight single-modal models comprises: Identifying a target multi-mode large model meeting replacement conditions in the initial model layout diagram; Selecting a plurality of target lightweight single-mode models from a model pool based on a scene-model capability mapping matrix and currently available computing resources of the system, wherein elements in the scene-model capability mapping matrix represent suitability scores of corresponding models in specific scenes, the scores are comprehensively calculated based on historical accuracy, recall rate and delay performance of the models in the specific scenes; And constructing the subgraph based on the plurality of target lightweight single-mode models, wherein the internal nodes of the subgraph are connected in a serial or parallel mode to cooperatively complete the analysis function of the original large model nodes.
8. A job violation identification system based on adaptive modeling, the system comprising: The receiving module is used for receiving an identification task request, wherein the request at least comprises monitoring video stream data, monitoring audio stream data and a service scene label, and the monitoring video stream data and the monitoring audio stream data are associated; The system comprises an acquisition module, a trigger logic module and a trigger logic module, wherein the acquisition module is used for acquiring a corresponding initial model programming diagram from a predefined model programming rule base based on the service scene label, wherein the initial model programming diagram is a directed acyclic graph and comprises a plurality of nodes and directed edges connected with the nodes, the nodes comprise a lightweight single-mode model and/or a multi-mode large model, and the directed edges are used for indicating data flow and the trigger logic; the simplification module is used for acquiring available computing resource information of the system, and carrying out simplification operation on the initial model layout diagram based on the available computing resource information to obtain an executable model layout diagram; The processing module is used for loading and instantiating all nodes in the executable model programming diagram, and scheduling the monitoring video stream data and the audio stream data to sequentially flow through each node for processing based on the executable model programming diagram; and the generation module is used for generating a job violation identification report based on the processing result of the output node of the executable model programming graph.
9. A computing device, comprising: A memory for storing a program; A processor for loading the program to perform the method of any of claims 1-7.
10. A computer readable storage medium, characterized in that the computer readable storage medium comprises a stored program, wherein the program, when run, controls a device in which the computer readable storage medium is located to perform the method of any one of claims 1-7.

Description

Operation violation identification method and system based on self-adaptive model arrangement Technical Field The invention relates to the technical field of operation safety supervision, in particular to an operation violation identification method and system based on self-adaptive model arrangement. Background In the fields of industrial production, engineering construction and the like, operation safety supervision is a key link for guaranteeing personnel life safety and preventing production safety accidents. With the rapid development of artificial intelligence technology, the operation safety monitoring scheme based on AI vision has the advantages of instantaneity, automation, non-contact and the like, and is widely applied to various operation scenes, so that the traditional monitoring mode relying on manual inspection is effectively replaced, and the monitoring efficiency and coverage are greatly improved. At present, the core architecture of the existing artificial intelligence (ARTIFICIAL INTELLIGENCE, AI) visual operation safety supervision system is mainly divided into two types, and the recognition of illegal behaviors is realized in a static and fixed model deployment mode. The first type of architecture adopts a single multi-mode large model, the model integrates multiple perception modes such as vision, infrared and sound, and has higher recognition accuracy, and the design thought is to uniformly analyze all monitoring video streams through a single model, so that full-coverage recognition of various illegal behaviors is realized. The second type of architecture is used for respectively deploying a plurality of light-weight convolutional neural network (Convolutional Neural Network, CNN) small models aiming at specific illegal behaviors such as non-wearing safety helmets, illegal crossing and the like, and improving the recognition speed in a multi-model parallel processing mode so as to meet the basic requirement of real-time supervision. However, the model deployment flow and the identification policy of the existing system are solidified, and cannot be dynamically adjusted according to the difference of actual service scenes and the real-time computing resource load, so that the existing system is difficult to adapt to the differentiated requirements of different scenes. In view of this, there is a need for a method and system for identifying job violations based on adaptive modeling. Disclosure of Invention Aiming at the problems that the existing operation safety supervision system is insufficient in flexibility and is difficult to adapt to the differentiated requirements of different scenes, the invention provides an operation violation identification method and an operation violation identification system based on self-adaptive model arrangement, which can improve the system flexibility and better adapt to the differentiated requirements of different scenes. The specific technical scheme is as follows: in a first aspect, an embodiment of the present application provides a method for identifying a job violation based on adaptive modeling, including: The method comprises the steps of receiving a task identification request, wherein the task identification request at least comprises monitoring video stream data, monitoring audio stream data and a service scene label, the monitoring video stream data and the monitoring audio stream data are associated, acquiring a corresponding initial model arrangement diagram from a predefined model arrangement rule base based on the service scene label, the initial model arrangement diagram is a directed acyclic diagram, the initial model arrangement diagram comprises a plurality of nodes and directed edges connected with the nodes, the nodes comprise a lightweight single-mode model and/or a multi-mode large model, the directed edges are used for indicating data streams and trigger logic, acquiring available computing resource information of the system, simplifying the initial model arrangement diagram based on the available computing resource information to obtain an executable model arrangement diagram, loading and instantiating all the nodes in the executable model arrangement diagram, scheduling the monitoring video stream data and the audio stream data to sequentially flow through each node for processing based on the executable model arrangement diagram, and generating a job violation identification report based on the processing results of the output nodes of the executable model arrangement diagram. Preferably, the simplifying operation at least comprises replacing the node corresponding to the multi-mode large model with a subgraph formed by a plurality of collaborative lightweight single-mode models, or placing the node with higher precision and more time consumption in parallel branches of the initial model layout chart into an inactive state. Preferably, the method comprises the steps of obtaining a corresponding initial model layout diagr