Manufacturers across industries strive to adopt manufacturing practices that guarantee high quality products. The low-quality products and flawed quality practices result in loss of revenue, reputation and much worse, a severe impact on lives and environment.
For example, Forbes mentioned that the quality and compliance issues in biotechnology industry drained 25-39% of their manufacturing costs in 2021 as a result of 300+ drug and medical device recalls. Ford suspended production and shipment of the F-150 lightning, the electric version of F-150 due to quality issues with the battery resulting in loss of sales and revenue. There are many such examples, with an average manufacturer attributing Cost of Quality (COQ) at 20% of total revenue.
On the other hand, advanced manufacturing methodologies, technologies, and trainings are helping manufacturers improve quality and compliance. However, the adoption of Artificial Intelligence (AI) is helping manufacturers create a predictive quality assurance regime that delivers additional benefits such as:
- Demand-based manufacturing
- Increased capacity & reduced overall cost (material, resources, etc).
- Increased manufacturing throughput.
For example, Deloitte’s Smart Manufacturing group mentioned that “Predictive Quality Analytics will reshape automotive quality, warranty, and recall management and help you getting ahead of the curve.” McKinsey suggested that in biopharma “Predictive quality risk monitoring can improve effective auditing by four times and can also spot signs of trial and site quality risks earlier”.
Quality control is a mechanism that determines if a product is manufactured according to the specification(s). There are four types of quality control mechanisms:
- Process control – focused on standardizing or center lining the asset set parameters, and other process parameters.
- Control charts – focuses on change control, and deviations.
- Acceptance sampling – focused on inspections and sample verification.
- Product quality control – focused on product verification and analysis.
What are the traditional quality control mechanisms?
There are several mechanisms. Statistical Process Control (SPC) technique is one of the widely used statistics-based mechanism that comprises of several techniques (or components) such as center line, upper line, and lower line, as well as the causation-based techniques such as common cause and special cause. In simple terms, the process data is subjected statistical analysis to identify anomalies and underlying causes.
- Create a continuous improvement cycle of qualify, monitor, and improve.
- Provide process owners a clear insight into the process models and their behavior instead of simply prescriptive improvement guidance.
- Reduces the unwanted variation in the process (thus, quality).
What is Predictive quality control?
A mechanism that uses historical process data (manufacturing environment) and predicts the quality of yet to be manufactured product(s). The process data broadly comprises of three kinds: 1) the machine sensor data and the 2) quality measurements data, 3) discrete environmental events. The predictive quality use cases range from detecting machine faults to inspecting product quality. The north-star goal for any data-driven predictive quality initiative is to achieve all quality targets defined in the specification(s). Such goals are typically measured using various Key Performance Indicators including Right First Time (RFT), Yield-Cost-Throughput (YCT), and Overall Equipment Effectiveness (OEE).
AI based systems have several classifications, however, in the context of predictive analytics, they are classified into learning systems, knowledge based intelligent supervisory systems, autonomous systems and self-aware systems. Predictive quality control techniques broadly fall under knowledge based intelligent supervisory systems which are equipped with unsupervised and reinforcement learning machine learning (ML) or deep learning (DL) models.
What are the predictive quality control techniques?
Different quality control techniques are adopted at different stages of manufacturing. Some practices are across industries, and some are specific to certain industries or manufacturing practices / process. Most predictive quality control techniques should be narrow enough to drive a niche and provide an extreme precision and quality conformance, but also broad enough to be made applicable across use cases that are driven of similar processes, material, or equipment (the key being similarities in the behavior patterns of the time-series data along with their causational synergy with the discrete events). Below, I have listed a few concrete examples at a level of detail.
- Visual inspection AI: Inferential AI models are used to analyze the images (or videos) and determine the dents or anomalies. Such anomalies are qualified as problems and their impact using the FMEA (Failure Mode & Effect Analysis) library. Convoluted and Recurrent Neural Network class of algorithms play a significant role in providing the vision inspection capabilities in addition to the clustering algorithms (ranging from K-cluster to more advanced adaptive resonance theory). For unlabeled tagged data, such deep learning models could help determine the anomalies.
- Multi-variate analysis – product validation: For automotive or high-tech electronics, the battery pack assembly is a precision driven process. Any low quality or misaligned connections at the microscopic level could cause significant performance degradations. Quality check of battery joins is an important exercise. The manufacturers have already minimized the error rates to an average of 0.1%. However, applying ReliefF algorithm trained AI model in determining the features using causation technique has helped achieved a significantly higher identification of defects, thus, helping the manufacturer get closer to their zero-defect goal.
- Process optimization: For large discrete product manufacturers, the assembly lines are made of long line of process units, with each unit completing a specific manufacturing function, with some taking as long as several hours to days. When such time intense processes induce faults into the product, there is an immediate impact on loss of material, time, and overall capital. Having a predictive model that determines if the manufacturing process runs within the quality specification will result in improved process, reduced cost, and improved product quality. AutoML models such as Auto-sklearn regression + classifier class could help manufacturers predict with a level of confidence that a product could fail due to a set of specific conditions at a specific process step. Process optimization results in improved product quality.
- Scheduling & sequencing: Manufacturers with wider product mix, different product lines (by capacity, configuration etc), and constrained manufacturing (order, priority, contamination, regulation, resources etc) require an adaptive scheduling mechanism that is responsive to the real-time environmental conditions. Such scheduling offers maximal multi-variate optimization, thus resulting in better quality and compliant products, in addition to achieving an optimal production process. The adaptive scheduling requires Machine learning models that interpret the scenarios based on process, sensor, and event data to provide predictive scheduling or sequencing.
What are the key capabilities required to productize & scale predictive quality?
- Connect: Ability to connect with systems and assets real-time (or near real-time) and retrieve the various types of data: timeseries (sensor data and other metrics – typically drawn from machines or assets) and discrete (event or action data – typically drawn from human activity or automated events via enterprise systems such as CMMS, ERP, inventory system etc).
- Contextualize: develop a run-time schema model that can help build the virtual factory meta-model. The available database types such as graph-database, with a layer of run-time adaptive layer to expand the self-generative model will help capture the entities (assets, systems, and people), relations (processes) and attributes (set points, operating parameters, product parameters etc).
- Cognize: Easily draw a use case specific predictive insights as explained with examples above. It is important to recognize that such models should adhere to following conditions:
- Re-usable: The AutoML mechanisms will automate ML pipeline of data preparation, algorithm selection, feature selection, train model, validate model, deploy model, and monitor model. Such automated model could allow easy integration points into the pipeline to draw continuous improvement of the model. This approach helps in moving the pre-trailed models from use case to use case as long as the context, data behaviors and expected outcomes are within the “ball-park”.
- Explainable or auditable: A process or set of services / tools made available to draw an “xray” into the AI model on the reasoning behind the prediction. An explainable AI model is mandatory for mission critical actions to be performed based on the model’s predictions.
- Edge (or in-line) & cloud (hyperscalars, controllers and SCADA): For achieving certain class of predictive quality control, it is fine to work with historical process and event data and predict a trend, anomaly, or event. Such analytics could be run on cloud or enterprise datacenters. However, certain other classes of predictive quality require real-time or near-real time responses (for example, vision-based product quality analysis followed up machine set-point advisory). For the latter class, its important that analytics are run in/at the factory assets – in other words, run it on controllers or optimized edge devices.
- Predictive Quality methods are critical for manufacturers who strive to identify causative factors and measure Key Performance Indicators (KPIs) that help them achieve high precision quality targets such as Right First Time (RFT), six or nine-sigma on dents, zero-failures etc where re-aligning the process parameters and asset set-points become keys to measure (or predict) quality before the product is produced.
- Productized predictive quality AL models that are reusable across use cases should have certain key parameters – automated ML training pipeline where models can be iteratively trained, grouping of use cases where the data behaviors, patterns and expected outcomes are similar, explainable ai which provides visibility on the reasoning, and a mechanism that has an adaptive ability to draw a virtual factory data fabric with time-series and the discrete event data.
Industry 4.0 | Supply Chain Solutions
6moA very informative article. Thank you Srinivas K.