**5. Performance Comparison**

In order to demonstrate the performance of the proposed method, the abnormal pasting process with spatial-temporal data will be identified in this section. The plate thickness data used for performance comparison are obtained from the pasting machine. After data cleaning and clustering analysis, the data of normal pattern and abnormal patterns is extracted and converted to process images. Each type of process patterns includes 200 and 500 process images with 50 × 50 size for training and testing respectively.

The abnormal pasting process images are online recognized using the constructed CNN recognition model. In the real pasting process, because the processing time of a plate is 0.5 s, the plate thickness observations at 50 locations in 25 s can be mapped to a process image with 50 × 50. Thus, the size of the moving identification window is set to 50 × 50. Taking 0.5 s as a stride length, once a new observation is obtained, the sliding window will move one step. As the window slides, the process images formed by the first 64 observations change smoothly and steadily, in which the output results of the CNN model are normal process images, shown in Figure 8. When the window moves to observation 115, the probability output is (0, 0, 0.005, 0.012, 0.003, 0.007, **0.973**, <sup>0</sup>), which indicates that the abnormal process pattern *F*6 happens in the pasting process. The corresponding cause is insufficient conveyor belt tension caused by the electromagnetic fault of the air pressure machine. When the root cause is removed, the plate thickness comes back to the target value nearby. When the identification window moves to the 199th observation, the probability output of CNN model is (0.011, 0.027, 0.02, 0, **0.934**, 0.008, 0, <sup>0</sup>), which means that the abnormal pattern *F*4 is detected. After checking the current process, we find that the plate flatness changes suddenly, which results from the low steel strength of new grids. After replacing this batch of grids with other qualified grids, the process returns to normal. When observation 281 enters the moving window, the probability output is (0.002, 0.008, 0, **0.9825**, 0, 0.003, 0.0045, <sup>0</sup>). Adjusting the levelness of the pasting machine makes the process go back to normal.

From the above application, we can conclude that the proposed method has practicality to the online identification of abnormal process with spatial-temporal data. To further evaluate the performance of proposed CNN recognition model, the UMPCA based recognition method [11] as a benchmark method is considered for comparison. In this UMPCA based method, Bayes Classifier (BC) is utilized to achieve identification, which is denoted as UMPCA-BC. After 100 experiments, the comparison results are obtained and shown in Figure 9.

*Processes* **2020**, *8*, 73

**Figure 8.** Online identification for the pasting process.
