reduce time‐consuming manual segmentation.
2.4.3 Tool State Diagnosis
Generally speaking, tool wear is concomitant to vibration and it gradually increases due to long‐term usage. In this case, the vibration is acquired from the cutting tool used in the side milling at a sampling rate of 2,000 Hz and the data of the cutting tool records from new to worn status. The entire data set is available in [19] via IEEE DataPort.
Features are extracted from time‐domain signals into the number of 24 WPT nodes based on the 4‐level WPT manner. Although differences in time‐domain signals between new and worn statuses are small as shown in the upper portion of Figure 2.29, the values of the 13, 14, and 15 WPTs of worn tool signals are clearly different from those of a new tool as portrayed in the middle portion of Figure 2.29. Here, the bandwidths of the 13, 14, and 15 WPTs are about 750–812.5, 812.5–875, and 875–937.5 Hz, respectively. The detailed amplitudes of each frequency band are illustrated in the lower portion of Figure 2.29.
Figure 2.29 Comparison of time‐domain signals (upper portion), WPT features (middle portion), and frequency‐domain spectrums (lower portion) between new and worn statuses.
Figure 2.30 illustrates four energy distributions of the 32 WPT features extracted from the X‐axis and Y‐axis vibrations under four cutting depths (from 4 to 7 mm). Note that, main differences of amplitudes exist among high‐frequency bands (especially from 13 to 15) between new and worn statuses. These features provide the AEN model with useful data source to extract information for the tool state diagnosis.
Figure 2.30 WPT distribution results for different cutting depths in the X and Y axis (node number counted from 0): (a) the new tool in X‐axis; (b) the worn tool in X‐axis; (c) the new tool in Y‐axis (d) the worn tool in Y‐axis.
Hence, 32 WPT‐based features of the X‐axis and Y‐axis serve as the inputs to the encoder in an AEN model. As shown in Figure 2.31, four SFs (fAE1, fAE2, fAE3, and fAE4) extracted from the fourth layer in encoder can be used as a compressed representation of the original feature set to reduce the number of feature dimensions; the left side (sample nos. 1–133) and right side (sample nos. 134–266) represent the new and worn cutting tools, respectively. Finally, the four SFs show their capability in classifying the new or the worn tool.
Figure 2.31 Comparison of four SFs extracted by using an AEN for samples of new and worn tools.
The accuracy of four feature sets are compared using a random forest (RF) model and evaluated in a cross‐validation scenario. As shown in Table 2.4, the average accuracies of tool state diagnosis when using 32 WPT‐based features and 4 fAE features are 89.5 and 69.1%, respectively. Furthermore, when the cutting depth is added as one of the inputs, the average accuracies are improved to 90.9 and 81.7%, respectively. This indicates that the accuracy by applying WPT‐based SFs is better than that by utilizing AE‐based SFs regardless of whether the cutting depth is added to the feature set or not.
Table 2.4 Tool diagnosis example: results of using an RF model.
| Model inputs | Average (%) | Best (%) | Worst (%) |
|---|---|---|---|
| 32 WPT SFs (X/Y axis with level = 4) | 89.5 | 95.0 | 81.2 |
| 32 WPT SFs (X/Y axis with level = 4) + Cutting depth | 90.9 | 96.2 | 85.0 |
| fAE1~fAE4 | 69.1 | 77.5 | 60.0 |
| fAE1~fAE4 + Cutting depth | 81.7 | 90.0 | 76.2 |
2.4.4 Tool Diagnosis using Loading Data
Retaining experienced machine operators is difficult because of poor manufacturing environment. Some forming machines are now equipped with pressure sensors to indicate operators with machine status for compensating their inexperience. To detect failures in a forming process, a pressure sensor (load cell) is attached to a forging die to detect variation in the forging signals of a bolt‐forging machine. Forging failures and loads are generally strongly correlated, but the load distribution may vary with numerous failure modes.
Further, an issue regarding big data exists in identifying failures after long‐term data collection. The cycle time for forming a bolt is only 0.3 s; thus, the signal length is approximately 300 points under the 1 kHz sampling rate of the sensor. The amount of data collected daily is almost 10 MB for one forming machine with data collection performed for 22 hours/day when using eight pressure sensors of four stages (i.e. 8 channels × 1000 data samples/second × 3600 seconds/hour × 22 hours/day). Thus, how to automatically diagnose failure modes from loading data in a forging process becomes a challenge.
Observing Figure 2.32, a forging load (pressure)‐stroke curve demonstrates two intervals: T1 and T2, which can be defined to indicate the characteristics of fastener forming. T1 is an interval representing the time from the die contacting the workpiece to forming after the material has exceeded its yield strength and shown plastic deformation; whereas T2 is the time taken for the cavity to be completely filled after T1.
Figure 2.32 A forging load (pressure)‐stroke curve.
The forging energy during T2 is mainly related to the geometric variation of the die, as illustrated in Figure 2.32. A feature engineering method is used to extract features, including those in time and frequency domains, from the load‐stroke signal in interval T2. For example, avg, std, kurt, RMS, skew, and max are extracted