Physically Enriched Neural Networks For Industrial Metal Detection

Restricted (Penn State Only)
- Author:
- Das, Suhrid
- Graduate Program:
- Electrical Engineering
- Degree:
- Master of Science
- Document Type:
- Master Thesis
- Date of Defense:
- October 18, 2024
- Committee Members:
- Vishal Monga, Thesis Advisor/Co-Advisor
Ram Mohan Narayanan, Committee Member
Madhavan Swaminathan, Program Head/Chair
Ankit Tyagi, Special Signatory - Keywords:
- Deep Learning
Machine Learning
Domain-enriched Machine Learning
Metal Detection
Digital Signal Processing
Pattern Recognition
Detection
Classification
Product Signal Adaptation
Artificial Intelligence
Time Series Classification - Abstract:
- Complex Time-series data from industrial sensors have been used since the past few decades. Traditionally, Digital Signal Processing(DSP) and other statistical methods have been leveraged to tackle these problems. Although, these methods have proven to be reliable, they struggle with tasks involving pattern recognition due to huge variations in data attributed to several physical factors which are hard to model mathematically. Here, we consider the crucial problem of metal detection. Products are passed through the detector’s aperture, where specialized sensors measure permittivity and conductivity values. Conventional industrial methods typically employ threshold-based algorithms that identify contamination by detecting strong conductive behavior. This characteristic, combined with other factors such as operational frequency, product orientation, contaminant type, and external noise, significantly complicates the detection process. The application of threshold-based methods in these complex scenarios often leads to a high incidence of false negatives and false positives due to their inability to incorporate domain-specific information. This limitation underscores the need for more sophisticated detection algorithms that can account for the nuanced interplay of various factors affecting these temporal signals in industrial contamination testing. Deep Learning methods are innately impressive in pattern recognition tasks and their advancements have led to increased accuracy in classification-based tasks with the requirement of little domain expertise. However, these algorithms are trained at the cost of huge datasets which in certain industrial scenarios are unachievable due to acquisition costs and time. Our study introduces domain-enriched deep learning algorithms for multivariate time series learnable classification and product signal adaptation, which make the best of both worlds by incorporating state of the art feature extraction abilities of deep learning architectures with domain knowledge to provide increased accuracy with reasonable training. We aim to address two broad tasks - first, the Binary Classification Problem which categorizes products as Clean or Contaminated and second, the Multi-Class Classification Task where we take a step forward and try to identify the type of contaminant present in a Contaminated product. The product signal adaptation modules are able to adaptively extract the contaminant signals as a pre-processing step to the main classification models further scaling up our models to different product types. Our methods perform superior to some of the widely used Traditional Signal Processing and Classical Machine Learning methods.