To run an ML/DL model in a high-complexity process, an elaborate model based on hundreds of thousands of sensor data must be developed, and the model must be served with stability through consistent data validation and model improvement. A global semiconductor company has developed an anomaly detection model optimized for the core process using MakinaRocks' Core ML technology accumulated in the manufacturing industry. It also was able to implement a workflow that enables real-time inference and continuous model improvement/deployment through the MLOps solutions(MakinaRocks Link™ & Runway™).

 

Situation

  • Failures in the core process of semiconductor lead frame manufacturing can result in problems such as decreased sales, decreased product reliability, and unexpected interruptions in the manufacturing process.
  • We determined that installing additional devices such as cameras and external sensors for evaluating the physical status of the equipment is restricted in the extended application.
  • Furthermore, it was determined that manual monitoring was impossible when approximately 70 types of data were analyzed and applied in a complex manner when collecting sensor data of core modules during the manufacturing process. As a result, we decided to move forward with ML/DL model-based detection and failure prediction.

Challenge

  • A data-driven insight is provided by an ML/DL-based anomaly detection model. The data collection process of the sensors in each device, however, was not stable.
  • Real-time inference is required based on sensor data collected in seconds. There was, however, no environment for serving the ML/DL model.
  • According to data verification and model improvement, frequent ML/DL model improvement and redeployment are required, and the complexity of operating numerous ML/DL models may increase when considering future expansion.

Solution

 

1. Develop anomaly detection model and provide anomaly score 

  • Data from normal operations for several months is downloaded and used to build an anomaly detection model based on Autoencoder, which has learned the normal operation data distribution.
  • Anomalous events from the past are used as a test set to ensure that the learned model classifies normal situations, anomalous situations, and symptoms before and after the anomaly, and that performance has been improved in response to field conditions.
  • MLOps is used to deploy the anomaly detection model, and the Anomaly Score is provided for a quick assessment of the overall state of the facilities.

  • The Anomaly Score calculates the level of risk by combining all available data. Individual data trends must be used concurrently for cause analysis and data validation.
  • The Reconstruction value of individual sensor data is provided for a data-driven understanding of the cause and consistent data validation.
  • To detect pattern changes in individual data, the model with normal distribution learning compares the inferred value (green) to the actual inflow value (yellow).

 

 

2. Implementation of real-time inference environment utilizing MakinaRocks Runway™

  • A ML/DL model serving environment is required for validating the model's output results based on actual data.
  • Real-time inference required an environment such as a Streaming DB, however, it was not available. In MakinaRocks Runway™, it includes a Kafka module that sequentially feeds the stored DB values into the model as input values to enable real-time inference.

 

 

3. Implementation of ML Model improvement/deployment workflow using MakinaRocks Link™

  • In order to serve the model in real time through MLOps, the developed model must be produced and easily distributed in the industrial field. In addition, pipeline creation and management will is critical for retraining.
  • By utilizing Link™, a pipeline suitable for use in an MLOps is automatically created. 
  • The pipeline that is created by Link™ is automatically registered in Runway™ via 'Export to Runway.' If the condition set by the field conditions is exercised, this pipeline is executed automatically via retraining. In addition, the manager can validate the retrained model and deploy it with a singe click.