| Abstract [eng] |
The aim of this Master's final project is to investigate the diagnostic system of mechatronic devices. To achieve this goal, the following tasks are set: to examine the diagnostic methods of mechatronic machines; to analyze the methods and tools for processing images and temporal signals that help to classify the ports of industrial controllers; to create a prototype of the diagnostic algorithm for industrial controllers integrated in mechatronic machines; to investigate the efficiency of image processing methods in identifying the states of the ports of an industrial controller; to investigate the diagnostic algorithm of the mechatronic device by qualitative and quantitative parameters. The software used in the project is MATLAB R2020b. The first chapter describes the diagnostic methods of mechatronic machines, based on models, signals, and data. Ports of PLCs from different manufacturers are presented and problems related to their state reading using image recognition are described. The second part presents the research methodology and system. The analysed object is a palletizing device for bread boxes controlled by a Siemens controller. The accuracy and speed of reading controller ports using YOLOv4 and YOLOv4 Tiny neural networks were examined; the performance and accuracy of networks under different environmental and network training conditions were evaluated. The impact of video capturing, environmental lighting, and the size of the neural network input on detection accuracy and performance were examined. The third part analyses the LSTM neural network for data forecasting and tests the fault detection algorithm, successfully detecting system sensor faults. It was determined that the trained LSTM network is intended for detecting sensor faults and is not designed to detect actuator faults or system air pressure drop errors. The fourth chapter provides recommendations and suggestions for the use of the system. Conclusions and research results: 1. It has been determined that diagnostic methods are divided into analytical and hardware-based. Analytical diagnostic methods are divided into methods based on the system model, system signals, and system data. The hardware diagnostic method is based on equipment duplication, where faults are detected by comparing the signals issued by duplicated devices, and any detected difference is considered a system failure. 2. By analysing the ports of industrial controllers, it was found that light indications mounted in the controller are used to display system states. It was determined that having a pre-compiled system event log and using neural networks, system failures can be detected, and the internal logic of the system can be restored and depicted with a Petri net. 3. A prototype of the diagnostic algorithm for industrial controllers integrated into mechatronic machines, capable of detecting system failures, was developed. The created system is suitable only for detecting sensor faults, it cannot detect actuator faults. 4. It was experimentally determined that when the video resolution is 1920 × 1080 px, the accuracy reaches 93.1% in the case of diffused lighting, while in the case of uneven LED lighting – only 70.2%. The research results confirmed that the speed of image processing with the YOLOv4 Tiny neural network reaches 49 frames per second using the GPU "RTX 4070", and 22 frames per second using the CPU "7600x". The optimal input size for performance using the YOLOv4 neural network is 352 × 352 px, while using YOLOv4 Tiny – 806×806 px. 5. During the research, it was found that the fault is detected within 300 ms from the start of the sensor signal disappearance. It was also determined that the system works correctly when there are 10 controller inputs and 6 outputs in it. |
