The relentless progress of artificial intelligence technologies is constantly creating new valuable development opportunities for industrial operations. New technical capabilities with the promise of replacing repetitive human inspections, ensuring product quality and streamlining operative processes are being created at a pace never seen before.
However, even though artificial intelligence capabilities keep becoming more general and applicable to diverse problems, implementing an effective end-to-end solution in a concrete industrial setting still requires – and will continue to require – a specific set of human skills. Several aspects related to technology selection, performance, integration and process optimization have to be carefully considered in order to create a solution that reliably and accurately fulfills the key operational needs.
Focus on the right AI technologies
Although the generic capabilities of the leading large language models capture most of the public attention, AI model architectures specific to the given industrial needs are often needed for effective solutions.
In the domain of industrial quality assurance, solutions are often based on computer vision technologies. The state of art in computer vision technologies is progressing extremely fast and producing increasingly capable visual analysis models for various applications. It is easy to find online sources with very compelling demonstrations for certain typical use cases. However, when trying to create a truly production-level solution and reach the last few percentage points of application accuracy and reliability in an industrial setting, generic out-of-the box solutions mostly fall short. In practice, it requires a lot of technical knowledge and experience to pick up the optimal techniques for the given problem. Furthermore, in many cases a combination of technologies is required to achieve the right level of performance, as discussed in https://topdatascience.com/advantages-of-combining-different-models/.
Utilization of pre-trained foundation models for computer vision is a great productivity boost for solution development. At the same time, reaching the accuracy and reliability required by a production-level industrial application typically requires specific training (fine-tuning) of the models.
Consider the full solution
While the focus is often on AI technologies, the final performance of an industrial computer vision -based solution will depend on a number of other system components.
First of all, camera capabilities are essential for capturing the right type of information. The key questions for camera selection are related to the properties of the target objects and the physical environment:
- What are the dimensions of the features to be detected and analyzed?
- Is color information relevant for the task at hand? If so, what colors are the most relevant?
- Is the target moving? How fast?
Lighting is often an undervalued component of a computer vision solution but can be crucial in reaching the full potential of both the camera equipment and the selected CV technologies. In fact, all the camera-related questions above also impact the lighting requirements. For example, the movement of the target object as well as its reflective qualities impact the light source specifications. As a result of varying local conditions, in most practical industry applications configuring a customised lighting solution will considerably increase the performance of the system.
Industrial solutions often require real-time operation in order to be responsive to the production processes. In most cases it is important to consider the performance requirements for running the AI models. Sometimes fully cloud-based deployments work well, but in many cases the solutions need to run on the edge. To enable that successfully, attention needs to be paid to the performance optimization of both image processing operations and inference model pipeline.
Integrate and optimize the process
For an industrial AI solution to become truly effective its deployment needs to reach beyond the limits of computer screens. Operators working in industrial production settings have other priorities than working in front of a computer workstation. Therefore, for example quality assurance and security solutions are often integrated directly with factory floor alarm systems. In many quality assurance cases it is even more effective to integrate the solution output directly to production line automation, for example to discard faulty products or stop the production line in case of more severe issues.
While tactical AI solutions often bring quick benefits in saving human labor and removing errors in the process, the most significant benefits result from enabling significant process changes. A well-engineered industrial quality assurance solution can create meaningful business impacts by removing unnecessary process phases, enabling faster introduction of new products or even enabling new levels of customer service. Reaching these requires deep understanding of the underlying business process drivers. In some cases the new quantitative process data produced by a computer vision -based quality assurance solution can be combined with other numeric data and used for holistic production process modeling and optimization on a larger scale.
Build for maintainability
Modern industrial processes need to be dynamic to support fast changes in technology and the business environment. New products and production technologies are constantly being introduced and product quality requirements in different markets may change. If process-critical AI-based solutions are not properly designed to be maintained and developed together with the business, a significant part of the overall solution cost can be associated with keeping up with the changes.
In effective solutions the update process for business-led changes is considered and supported from the start. Also the hardware specifications need to take future changes into account.
There are several typical dynamics in the domain of industrial computer vision quality assurance, requiring flexibility from the solutions.
- Target object variation: the visual properties of the computer vision targets will continue to change beyond the solution development phase, due to various external factors such as seasonal variation or product optimization. It is crucial for the solution to either adapt directly or have a routine means to manage small changes so that it can be maintained on the factory floor without extensive investment of engineering effort or long system downtimes.
- New product introductions will continue to happen for any up-to-date production process. Although many AI-based computer vision models generalize remarkably well to new data, updates with possible re-tuning are often required for fully new products. The process for this should also be preplanned and managed for efficiency.
- Changing quality criteria: regulatory or market pressures can lead to changes in product inspection criteria. Although these often relate to much slower changes than for example product variation, it is again important to be able to adapt efficiently without rebuilding the solution.
- Production line duplication: product success in the marketplace may lead to a need for rapid production capability scaling. In such situations the computational inference capability, whether in the cloud or on the edge, needs to be scalable as quickly. Also, the camera and lighting setups need to be well understood to enable fast duplication.
Conclusions
AI and computer vision are creating significant opportunities for optimizing industrial processes and unlocking entirely new value in production. However, realization of the gains enabled by new technology requires a specific mix of application capabilities, from keeping abreast of the latest AI developments to ensuring the solution deployments match the business targets. Working with an experienced and competent development partner to get the most out of the AI transformation is definitely the way to go.
Author:
Otto Pulkkinen, Head of Industry Solutions
