Failure prediction through a model-driven machine learning method

For the purpose of quality control in some production lines each product undergoes an end of line test. Early detection of products that will fail during or at the end of the test, can save time and material. Besides, it prevents hazards and damages to the test equipment and the product which can be caused by operating defective products. Utilizing data analysis for prediction of such anomalies can be challenging due to the non-stationary and multi-phased nature of the test/measurement process. This paper presents a model-driven machine learning method (using welding machine test data), where instead of fully data driven anomaly detection analysis, the analysis starts with approximate modeling of temperature evolution at the location of temperature sensors which is governed by physical process of heat transfer, originated from any unspecified heat source. This approach allows decision making about the test result just at the beginning of the test. In addition, since this method reduces the dimensionality of the input space, the machine learning model can be trained with much less number of samples, which is an enabler for many cases where access to training data is limited.