2nd ICAI 2022

International Conference on Automotive Industry 2022

Mladá Boleslav, Czech Republic

of the transport set is indicated in the form of the percentage of empty platforms during the implemented operations. In addition, the application indicates the percentage of time during which the transport set carried out transport work in excess of the norm for the operation. In the red square there are safety risks defined on the basis of the sequence of the transport set appearance in the place not matching the defined sequences of routes.

Figure 7: Final web application screen shot

Source: Own elaboration

4. Conclusion The presented solution has been implemented in AUTOPART battery plant located in Poland, the supplier of starter batteries for vehicles. In our opinion, the solution can be widely used in industry. Not only in automotive sector, but everywhere where there are events recording devices available in production area and there are repetitive routines of internal transport. The use of transfer learning methods allows to significantly shorten the process implementation of deep learning solutions. Classical machine learning methods require much more expertise than the deep methods cited in the paper. Applications based on deep learning using an appropriate degree of generalization allow to achieve high computational efficiency and versatility of solutions that can be used in other applications after minor modifications to the size of input vectors or network output layers. The project presented in the report is a good example confirming the effectiveness and ease of using pre-trained networks in industrial/commercial applications. The project achieved an SSE of 140 which is a satisfactory result. In addition, the time required to perform the analysis for a single day was reduced by nearly 93% which allows for continuous analysis rather than taking a snapshot of a day as before. Additionally, the observer’s influence on the results was eliminated. No potential for improvement was seen with the classical methods. On the other hand, with continuous observation using YOLO, nearly 30% potential to use the transport means for other purposes became apparent. Errors in data collection are minimized to zero. What remains to be improved is the frame processing speed (7 frames per second using commercial PC), which is not sufficient. As a countermeasure the number of frames per second was reduced, which allows to process data at a satisfactory rate. The possibility of implementing additional conditions should be noted. In the presented

115