| Summary: | 碩士 === 國立虎尾科技大學 === 電機工程系碩士班 === 105 === In the recent years, the 3D printing technology develops rapidly and globally that it is being regarded as an important part of the third industrial revolution.in the recent years. The 3D printing technology simplifies a lot of complex parts of manufactured process. The technology of 3D printing are characterized by efficiency, flexibility, low cost, and customization With the important patents such as fused deposition Modeling (FDM) and stereo lithography (SLA) had expired, more vendors were able to participate in this industry. Because of the modern machinery structure is complexity and long time to use 3D printer, the mechanical failure is happened. Therefore, in order to improve the machine reliability, the fault diagnosis system becomes more important in the future. We expect that the maintenance staff can find the faults quickly and correctly via fault diagnosis system. In this thesis, we use neural algorithms and sensing elements to realize fault diagnosis function in 3D printer.
The proposed 3D printer fault diagnosis system integrates neural network, sensing elements and has the function of graphical interface. In the hardware, we monitor the status of 3D printers by using the MPU6050 accelerometer and Arduino microcontrollers. In the software, we use machine learning of neural network to simulate the nonlinear relationship of each fault of 3D printer. In the thesis, we also use different parameters of neural networks to find out which parameter is suitable for our 3D printer. The network output is then sent to LabVIEW and displayed on the screen to help maintenance staff to find out which part of the machine is faulty.
We use different parameters such as the number of neurons, the number of training times, the learning rate, with different printing graphics for training 3D printer fault diagnosis system to find out the optimal parameter in the system. We use 200 simulation data with different hidden layer and output layer transfer function in the simulation scenario. The recognition rate in different hidden layer and output layer transfer function is 90.5% by tansig and softmax, the recognition rate is 89% by logsog and softmax, the recognition rate is 89% by tansig and tansig, the recognition rate is 83% by logsog and tansig, the recognition rate is 35% by tansig and logsog and the recognition rate is 36% by logsog and logsog. The simulated results show that the best recognition rate is used tansig and softmax of hidden layer and output layer, respectively.
We use 6000 sensing data about 10 minutes with different hidden layer and output layer transfer function in real time working scenario. The recognition rate is 83.5% by tansig and softmax, the recognition rate is 82.5% by logsog and softmax, the recognition rate is 83% by tansig and tansig, the recognition rate is 82.9% by logsog and tansig, the recognition rate is 48.3% by tansig and logsog and the recognition rate is 29.4% by logsog and logsog. The simulated results show that the best recognition rate is used tansig and softmax of hidden layer and output layer, respectively. In conclusions, an intelligent fault diagnosis for 3D printer by neural network is proposed and the optimal parameters of neural network found.
|
|---|
