A Mold Damage Monitoring Algorithm for Power Metallurgy Molding Machines Using Bidirectional Long Short-Term Memory on an Internet of Things Platform

Abstract

In this paper, an analysis and monitoring algorithm is proposed for mold health evaluation using vibration data. Two inertial measurement units (IMUs) and an embedded system are first used to acquire vibration data from a powder metallurgy molding machine. These data are collected on an Internet of Things (IoT) platform using the Message Queueing Telemetry Transport (MQTT) protocol. For data analysis, the vibration signal on the Z axis is segmented to label the contact section of the upper and middle molds, and the corresponding vibration data of the stamping friction on the X, Y, and Z axes are extracted. Using only historical vibration data from normal stamping, a Bidirectional Long Short-Term Memory (Bi-LSTM) model with an attention mechanism is trained to predict normal stamping vibrations several minutes in advance. By comparing the predicted stamping vibrations with the observed data at the current time, the mean square errors (MSEs) are calculated to evaluate the health status of the mold. Several ablation experiments were conducted to assess the performance of the trained model. The average MSE values for normal samples and abnormal samples were smaller than 0.5 and larger than 1.0, respectively. The experimental results confirm that the trained prediction model and evaluation indicators can effectively notify operators in advance. An early warning system using vibration data for mold damage was successfully implemented, enhancing predictive maintenance.

Keywords: Internet of Things (IoT); deep learning; inertial measurement unit (IMU); intelligence system; mean square error; vibration data.

Figure 1

The damaged mold. (a) A horizontal crevice (Chinese words) can be seen in the red circle; (b) A close-up image of the crevice.

Figure 2

A diagram of the proposed method.

Figure 3

The vibration data acquisition devices and data collection platform. (a) Data acquisition hardware; (b) Data bus; (c) Platform based on the MQTT protocol.

Figure 4

The mounted vibration data acquisition devices; red circles depict the positions of the upper and middle molds. (a) Front view of the molding machine; (b) Close-up shot of the upper mold.

Figure 5

The stamping cycle of the upper mold.

Figure 6

The Z axis vibration signal in a stamping cycle. Two peaks are detected in Step 3 as drawn in the two red circles.

Figure 7

The extracted vibration signals on the X, Y, and Z axes for feature extraction. (a) The signals of the IMU on the upper mold; (b) The signals of the IMU on the middle mold.

Figure 8

The network architecture of Bi-LSTM with an attention mechanism when parameters of ‘feature dimension’ and ‘time step’ are 6000 and 20, respectively. Here, symbol ‘+’ represents an addition operation.

Figure 9

The training process of the deep model.

Figure 10

The prediction process of the deep model.

Figure 11

A comparison between the predicted and observed data with an MSE value = 0.1985 in the training dataset (inside test). (a) The data on the X axis; (b) The data on the Y axis; (c) The data on the Z axis.

Figure 12

A comparison between the predicted and the observed data under the healthy condition with an MSE value = 0.2551 (outside test). (a) The data on the X axis; (b) The data on the Y axis; (c) The data on the Z axis.

Figure 13

A comparison between the predicted and observed data in the damaged condition with an MSE value = 1.1076. (a) The data on the X axis; (b) The data on the Y axis; (c) The data on the Z axis.

Figure 14

The MSE values between the predicted vectors and observed vectors. (a) The MSE values in the health status over two hours; (b) The MSE values in the damage status over two hours; (c) The MSE values of (a) after filtering with a 1 × 3 window size; (d) The MSE values of (b) after filtering with a 1 × 3 window size.