Automatic Defect Recognition (ADR) is a term that is gaining traction within the NDT world. However, neither ADR or the functions such a system should deliver are clearly defined.
At JetSoft we prefer to split ADR into two stages, using the terminology Automatic Indication Detection (AID) and Automatic Part Sentencing (APS).
The problem with ADR
The scope of ADR isn’t clear. It seems, depending upon your source, ADR can range from a system which highlights abnormalities to an inspector, to one which can perform full sentencing (go/no go), eliminating the need for an inspector at all. In practice, most end users anticipate a solution resembling the latter, but this may not be true of the products promoted by equipment vendors.
AID and APS
At JetSoft we don’t use the term ADR, due to the lack of clarity on what this means, instead we use the terms:
- Automatic Indication Detection (AID)
- Automatic Part Sentencing (APS)
Defect detection is a two step process. The first step is the discovery of an indication or discontinuity, the second step is evaluating that indication against a specification to determine whether it is a defect. ADR may (or may not – depending upon who you talk to) be considered to be the automation of this two-step process. However, in JetSoft’s experience each stage can be automated separately, indeed it would be difficult not to.
JetSoft refers to the automation of the first step as Automatic Indication Detection (AID) with the results highlighted to an inspector or used as an independent assessment and compared with the manual outcome.
The second step, qualifying the indications against defect specifications, JetSoft calls Automatic Part Sentencing (APS). Because once a system is confidently automatically qualifying defects it must logically be able to decide if the part passes or failed.
Is ADR overhyped?
News about ADR (go/no go) using deep learning and machine vision is rife and in my view racing towards the peak of inflated expectations on the Hype Cycle.
The Hype Cycle, developed by the technology firm Gartner, details the phases a new technology goes through to reach maturity and widespread adoption.
The scale of the peak and trough of this graph is exacerbated by the lack of clear definition of ADR. It is not difficult to imagine a situation where customers and suppliers differ on the expected functionality, resulting in a negative view of ADR and its possibilities.
However, ADR will happen. There is a large commercial appetite to reduce the cost of inspection and new advances in computation technologies prove that it is possible technically. But it may take a while to get there and require more investment (time and money) by users than anticipated. The market should be prepared not to expect a single turnkey algorithm that performs ADR on any component for any defect, but, should rather, anticipate solutions that require dedicated configuration, development and training based on what the users is looking for and the component/material type.
We hear a lot about the benefits involved with ADR, but perhaps less so of the realism of implementing and delivering them.
What to expect
JetSoft is amongst several companies utilising these new technologies, with working systems, using deep learning and machine learning techniques in place in production environments.
It is JetSoft’s experience that performing accurate automatic detection of defects takes dedicated time, development, and extensive testing. Furthermore, for an AID system to be effective, additional information beyond inspection data is required. Including component details and upstream process data, both of which are critical.
It is important to note that in every case the preceding step to achieving an APS must be a fully working, and empirically proven AID system.
As always, we would love to hear your thoughts on this, and if you have any questions, or would like to know more about what we do and the systems we have developed please get in contact.