CN-121982625-A - Intelligent joint defense system and protection method based on AI vision

CN121982625ACN 121982625 ACN121982625 ACN 121982625ACN-121982625-A

Abstract

The invention discloses an intelligent joint defense system and a protection method based on AI vision, which realize comprehensive breakthrough of safety protection performance by integrating multisource image input, a high-speed portrait detection model based on YOLOv optimization, a special PLC communication protocol and a protection mechanism combining software and hardware, reduce false alarm rate to below 0.1%, basically eliminate missing report, ensure that total delay of AI reasoning and system linkage is less than 1 second, ensure that the speed of each frame of the model is lower than 20ms, and are remarkably superior to that of the traditional infrared raster and the conventional AI vision scheme, and simultaneously have multi-mode output, environment self-adaptive start-stop and flexible regional demarcation capability, realize high-reliability and high-real-time global safety joint defense with low cost on the premise of not needing large-scale reconstruction of the traditional workshop PLC system, effectively ensure production continuity and personnel safety, and have remarkable economic and social benefits.

Inventors

Rui Xingwen
GONG BIN
DING JIE
YU JUN
Xia Sijuan
LIU JIAN
GAO FEI

Assignees

南京长安汽车有限公司

Dates

Publication Date: 20260505
Application Date: 20251217

Claims (10)

1. Intelligent joint defense system based on AI vision, its characterized in that includes: The image acquisition unit is used for acquiring image data from various input sources, wherein the input sources comprise at least one of a local camera, a network camera and a video file; The core processing unit is connected with the image acquisition unit, internally provided with an optimized portrait detection model and used for carrying out real-time reasoning on input image data and judging whether personnel exist in a target detection area or not; The PLC is electrically connected with the regional safety control module of the workshop equipment; The communication module is used for establishing data connection between the core processing unit and the PLC and transmitting signals through a communication protocol; When the core processing unit continuously detects personnel for multiple frames, the communication module sends a trigger signal to the PLC, and the PLC responds to the trigger signal and outputs a control signal to the regional safety control module to force the workshop equipment to stop; When the core processing unit continuously detects no personnel for multiple frames, the personnel manually resets the detection signal through a safety door reset button outside the body, after resetting, the detection signal is delayed for a preset time through a delay module, a reset signal is sent to the PLC, and the PLC responds to the reset signal to release the control signal.
2. The intelligent joint defense system based on AI vision as set forth in claim 1, wherein the optimized portrait detection model is a YOLOv architecture-based model with an inference speed of less than or equal to 20 ms/frame, and a total delay from picture capture to recognition result output is less than 1 second.
3. The intelligent joint defense system based on AI vision as in claim 1, wherein the core processing unit further comprises an interaction configuration module that provides a graphical user interface based on PyQT framework development for configuring at least one of a detection region ROI, a sensitivity threshold, a PLC configuration setting, an enterprise WeChat push recipient, and an enabled input source.
4. The intelligent joint defense system based on AI vision as in claim 3, wherein the core processing unit further comprises a multi-output module for synchronously performing at least one of the following operations when a person is detected: pushing early warning information to a designated enterprise WeChat pushing receiver, driving a voice broadcasting device to broadcast on site, marking and displaying personnel positions and detection time on a display terminal, and automatically storing a detection result picture containing the personnel positions to a designated path.
5. The intelligent joint defense system based on AI vision as in claim 1 or 4, wherein the core processing unit performs real-time reasoning on the image data, comprising: preprocessing input image data, and converting the input image data into a format for processing a portrait detection model; carrying out layered extraction on image features through a backbone network based on CSPDARKNET-53 structures; feature fusion is carried out through a neck network comprising a feature FPN pyramid network and a PAN path aggregation network; predicting center coordinates, width and height, confidence and class probability of the portrait by using a detection head of an Anchor-Free frame; filtering the predicted original result to remove redundancy and error frames, and obtaining a final portrait detection result.
6. The intelligent joint defense system based on AI vision as set forth in claim 1, wherein the communication protocol is an S7 communication protocol for Siemens S7 series PLCs, and the trigger signal and the reset signal are implemented by a Boolean variable value written in a data block DB block of the PLCs, wherein setting a variable to TRUE represents triggering shutdown and setting to FALSE represents resetting and canceling shutdown instructions.
7. The intelligent joint defense system based on AI vision as set forth in claim 1, wherein said delay module is integrated within said PLC for turning off a delay timer TOF.
8. An intelligent joint defense method based on AI vision, characterized in that it is applied to the system according to any one of claims 1-7, said method comprising: the initialization step is that a core processing unit is started, system parameters are configured through an interactive configuration module, a portrait detection model is loaded and preheated, and communication connection is established with a PLC; acquiring environment sensitivity data, transmitting the environment sensitivity data to a core processing unit, and automatically starting and stopping a personnel detection function based on a comparison result of the sensitivity data and a preset threshold value; If the sensitivity data is more than or equal to a preset threshold value and reaches a starting condition, acquiring real-time image data of a set current input source through an image acquisition unit, and detecting real-time personnel by utilizing the portrait detection model; triggering and stopping, namely if a person is detected by continuous multiframes, generating a triggering signal by a core processing unit, and sending the triggering signal to a PLC (programmable logic controller) through a communication protocol so as to stop the PLC control workshop equipment; And a resetting step, in which if no personnel are detected in the continuous multiframes, the personnel are reset outside the line, and after the resetting time reaches the resetting delay preset time, a resetting signal is sent to the PLC so that the PLC controls workshop equipment to release the shutdown state.
9. The intelligent joint defense method based on AI vision as defined in claim 8, wherein the system parameters configured in the initializing step include a detection region ROI defined by a polygon, a sensitivity data preset threshold, a continuous detection frame number required for triggering, and a preset time for PLC delay reset.
10. The intelligent joint defense method based on AI vision according to claim 8, further comprising, during the triggering and stopping steps, synchronously executing at least one of enterprise WeChat pushing, voice broadcasting, result displaying and picture saving multiple output operations through multiple output modules.

Description

Intelligent joint defense system and protection method based on AI vision Technical Field The invention belongs to the technical field of industrial safety protection and artificial intelligence intersection, and particularly relates to an intelligent joint defense system and a protection method based on AI vision. Background In an automobile manufacturing workshop, regional safety protection is a key link for guaranteeing production safety and preventing personnel and equipment from accidental collision. At present, the technical application in the field can be mainly divided into two major categories, namely a traditional mechanical/electronic protection technology and a conventional AI visual detection technology. Traditional mechanical/electronic protection technologies, such as an infrared grating, a safety protection door, a scram button and the like, are widely used. Such techniques typically consist of a detection unit (e.g., infrared transmitters and receivers), a signal transmission line, and an execution unit (e.g., field device energy control relay). The working principle of the system depends on the triggering of physical signals, for example, when infrared light beams are blocked or the state of a protective door switch is changed, the system stops the triggering equipment, so that the regional isolation and the safety protection are realized. With the development of artificial intelligence technology, conventional AI visual inspection technology has begun to find exploratory application in certain industrial scenarios. For example, publication CN112348728a, a method for detecting personnel intrusion in an industrial scenario, discloses a face detection scheme based on deep learning. The core architecture of the technology comprises an image acquisition module (a local camera which usually adopts a single path), an algorithm processing module (a portrait recognition model based on a convolutional neural network CNN) and a result output module (outputting a simple text prompt or driving an alarm lamp). The working process comprises the steps of collecting an image of a monitoring area through a camera, judging whether personnel exist in the image or not by utilizing an algorithm model, and generating an alarm signal if the personnel are identified. However, such techniques fail to provide deep fusion and stable interaction with plant control systems (e.g., PLCs). Although the prior art provides safety protection capabilities to a certain extent, they all suffer from significant limitations and disadvantages: The traditional mechanical/electronic protection technology firstly has limited coverage range, equipment such as an infrared grating and the like needs to be arranged continuously along the edge of a protection area, and the irregularly-shaped area is difficult to cover effectively. Secondly, the risk of false alarm and false alarm is high, namely, non-target objects such as a carrying robot, a moving metal workpiece and the like which are operated in a workshop are easy to accidentally block infrared light beams, so that equipment is triggered and stopped by mistake, and if personnel pass through a detection area in a bending, quick traversing and other modes, the false alarm of a system is possibly caused due to incomplete shielding of the light beams, so that potential safety hazards are left. Finally, the flexibility and the suitability are poor, when the layout of workshops is adjusted according to production requirements, the workshops need to be disassembled and wired again, the process is time-consuming and labor-consuming, and the production efficiency is affected. The conventional AI visual detection technology has the advantages that firstly, the input and output modes are single, only a single local camera is usually supported as an input source, multiple inputs such as the existing network camera, pictures and videos of a factory cannot be compatible, the output end is limited to a simple local alarm prompt, diversified and remote early warning means such as enterprise WeChat message pushing, on-site voice broadcasting and detection result picture archiving are lacked, and the comprehensiveness and coverage range of safety early warning are limited. Secondly, the real-time performance is poor, and the high-requirement industrial scene cannot be met. The model reasoning speed of the prior art is generally slow, for example, the reasoning speed recorded in CN112348728A is about 60 ms/frame, the delay from picture grabbing to final recognition result often exceeds 2 seconds, and the emergency safety response requirement of an automobile manufacturing workshop on equipment rapid shutdown (usually within 1 second) cannot be met. Thirdly, the device is not linked with the PLC control system, the conventional visual detection technology is not linked with the on-site PLC device control system, the technical requirement that the device is triggered to stop after an abnormal result is