Quality Inspection

Data Products

Deep Learning and Computer Vision to increase production quality

Industries:

Industrial - Life Science

Solutions:

Data Products

Technologies:

Azure - PyTorch - Aws - Google Cloud Platform

Scopri

Overview

Visual inspection is an essential step in the production process of many industries, including manufacturing and pharmaceuticals, to examine products for defects or errors. However, this process can be time-consuming and costly, and therefore, only carried out on a sample basis, leading to rework and inefficiencies in downstream production steps.

Inspector training is often complex and not easily scalable.

Artificial intelligence and deep learning algorithms can assist inspectors by “augmenting” their diagnostic capabilities or automating some of the Quality Control processes.

Computer vision can be used to detect a wide range of defects, including surface defects, dimensional errors, and missing or incorrect components. It can also be used to verify products’ compliance with regulations and standards.

In the manufacturing industry, computer vision-based visual inspection can be applied to a variety of products, such as mechanical parts, electronics, and packaging, locating and identifying different types of non-conformities.

Challenges

Human inspection presents several critical issues that can affect the accuracy and efficiency of the quality control process. These include:

Fatigue and boredom: Human inspectors can become tired or bored after performing the same tasks repeatedly, leading to a decrease in attention to detail and an increase in errors.
Limited visibility: Some defects may be difficult to detect by the human eye, especially if they are small or hidden in a complex product.
Subjectivity: Different people may have different interpretations of what constitutes a defect, causing inconsistency in the inspection process.
Limited speed: Human inspection is usually slower than automated inspection, causing bottlenecks in the production process.

Solution

Quantyca has developed several solutions over time based on established CV architectures such as ResNets and Deep Learning frameworks such as fast.ai and Pytorch. The key to the success of these projects lies in the ability to analyze with the client's Domain Experts the fundamental steps of the process and the points where inspection complexity is nested, and to make the best use of available data using fine-tuning techniques starting from already consolidated foundation models. Quantyca supports the client in different phases:

Problem formalisation

Definition of scope, identification of performance metrics, calculation of ROI between inspection cost savings and impact on final quality

Process analysis and breakdown

Each visual inspection process is typically decomposable into sub-problems that can be addressed and optimised individually, simplifying algorithms and increasing overall maintainability

Validation and/or collection of training datasets

Identification of data on which to train the model and possible iterative enrichment and correction to make the process of defining the training ground-truth as robust as possible for the objectives to be achieved

Modelling and training

Literature review and identification of the most suitable foundation models for the problem, on which fine-tuning is to be performed. Implementation and training of the model using standard technologies and frameworks such as Pytorch, Fast.ai and Timm

Performance validation

Rigorous performance analysis of each algorithmic step and estimation of the overall KPI improvement of the AI-enabled process

Deployment in production

Integration of the model with field technologies, both on edge and in hybrid mode with the cloud, exploiting industry standards such as Diconde

Production performance analysis and model drift control

Continuous performance monitoring, real impact assessment and retraining of the model under new operating conditions

Need personalised advice? Contact us to find the best solution!