Industry 4.0: from point data to business impact

Stage 1: Data

Due to IoT Sensors in the assembly line, our data sets are unstructured, organized, and semi-structured. We created a Big Data Analytics platform driven by Hadoop and Spark. In near real time, IoT sensor data and other operational process data from across factories are assimilated. Each distinct source requires a unique ingestion approach based on the variety, velocity, validity, and volume of data. The solution used "schema on read," which means that the data is not checked prior to loading but rather when a query is submitted; this results in a very quick initial load because no data is read and there is no danger for subsequent phases. As a result, we built an automated quality engineering method for reconciling, ensuring the quality, completeness, and absence of crucial data points. It is integrated with an escalation-based notification system, which ensures that the appropriate team is notified immediately. Without this mechanism, it would have been a case of garbage-in, garbage-out.

Across all assignments, we evaluated > 1TB of data from a diverse sample of IoT machines and assembled robots with 76+ million operational process data and 600+ business drivers.

Stage 2: How to travel from Data to Information

After defining the problem statement for predicted failures in the manufacturing area and acquiring data, we began data pre-processing. This is crucial for ensuring that acquired data is clean, complete, and curate. To begin, without context for data, we cannot analyze anything in the past or harness it for future decision-making. We labeled raw data prior to beginning pre-processing with AI algorithms. Data labeling is a time-consuming and laborious procedure that entails data annotation, tagging, classification, moderation, and processing. There are numerous manual techniques, however we choose to label using machines assisted by machine learning/artificial intelligence technologies. We concentrated on a strategy that is somewhat unknown but quite effective. Self-supervised learning is a methodology for autonomously labeling data by matching and exploiting connections between various input indicators. Self-supervised learning is ideally suited for changing settings because industrialized models may be constantly learned in production. At this point, our data has been turned into information.

Figure: Self-supervised learning

Stage 3: How to leap forward from Information to Knowledge

To achieve the Knowledge state of the data, pre-processing of the data was conducted. This included data normalization, identifying duplicates/quality, imputation of missing values, and identifying extreme data points that were outliers. We used distance-based algorithms and genetic algorithms for pattern matching in order to build a golden copy of part ID identification, cleaning, and standardization based on placed orders and varied BOMs. Missing value imputation was performed using machine learning algorithms to predict missing values. All machine learning algorithms contribute to the principle of continuous quality assurance at each stage of the journey. The entire process of standardization and quality assurance was automated based on the DevOps principles. Each change was logged, analyzed, and communicated to stakeholders.

The historical data for machine failures was skewed, with the minority class (percentage of failures) ranging between 3% and 5%. It would be misleading to feed such data into predictive maintenance systems. Collecting further failure data was also not a possibility. We used advanced machine learning techniques to turn an imbalanced data set into a balanced training dataset, but left the test/validation datasets alone to avoid bias during the validation process.

Each dataset has a few extreme observations, and ours was no exception. We had to identify extreme observations that should be ignored when creating models. To avoid outliers entering into the models, we applied AI approaches such as Cook's distance and other similar strategies. We are now pushing onward to the level of Knowledge.

Additionally, we encountered a cold start issue everywhere new units were planted, and we lacked sufficient data to train our models. The use of machine learning-based similarity algorithms enabled us to distinguish between new and old objects based on a variety of criteria. Thus, nearest-matched historical product/machine data was utilized to predict, and it was refreshed with each use of the model output via a feedback loop procedure that gave back deviations in prediction to the model as a self-learning feature.

The challenge of obtaining representative test data

Although test data is typically included in the initial data set, we were unable to obtain production data due to security, privacy, and local laws. As a result, we used Unsupervised Deep Learning algorithms to generate production-grade data based on production systems without exposing real production data. However, the generated data retained the same pattern as the live production data. Now, data validation tests are automated using continuous integration. Anytime a new data-related feature becomes available, it integrates with the master flow via automated pipelines. Following integration, we used automated continuous delivery (CD) pipelines to push the feature to all candidate platforms with the necessary approval and exception processes. It enabled real-time training and validation of machine learning solutions within the corresponding platform using the necessary dataset. Section 6 provides a deep insight into synthetic data.

Stage 4: Time to fly towards Wisdom

When we established a relationship between the objective function and various extra variables, we arrived at the Knowledge stage of data. To avoid misleading results, pairwise correlation, scatter plots, and VIF (Variance Inflation Factor) were used to identify correlations between independent features. Classification set of machine learning models aided in the development of machine intelligence by establishing a trend behind failures over the last 24 months and enabling fully trained solutions to take on the future part of the Wisdom phase. When machine learning techniques are used, significant statistics such as true negatives, true positives, false positives, and false negatives are generated, as well as a confusion matrix is established to help grasp critical metrics. We built functions to compare these indicators to predefined thresholds and to notify the appropriate stakeholder at the appropriate moment. After we were satisfied with the trained model's performance on the training data set, we compared it to the validation data set. Generally, if there is a considerable difference in metrics between the training and validation datasets, this is referred to as an overfitting problem, which reduces prediction accuracy when using new data sets. We built a function to check for overfitting and automated it to generate alerts in the event of a deviation from the preset overfitting threshold.
Now, we face a more difficult task: maintaining a high level of quality while progressing toward Wisdom. Here, we forecasted failures for the following 24 hours based on qualified features.
Regular data points were generated by IoT devices, and the solution projected the chance of failure per machine and each plant. Once the model emits results, we created an automatic layered notification procedure into three categories. We observed models deteriorating after a quarter, which was detected by the business team before the technical team. It was time to restart the model. Using statistical techniques, we found drift in the model's features and created continuous training pipelines for automatic training after recalibration is confirmed and accepted.

Step 5: The Holy Grail of business impact

Finally, we tied outcomes to the business purpose of improving the bottom line and published considerable improvements in decreasing plant/machine shutdowns and optimizing both fixed and variable costs.

We observed the importance of instilling continuous quality at each step. In contrast to transaction assignments, we must constantly monitor the accuracy of current models, assess their impact, and determine whether or not to refresh our models with new data. Continuous quality cannot be achieved without the use of intelligent automation powered by advanced machine learning/artificial intelligence algorithms that can learn and optimize themselves with each execution of the solution. Rather than relying on bundled products' drag-and-drop functionality, it is strongly encouraged to write code-based scripts based on open source stacks, since each script can be registered, automated, adjusted, and customized to meet our own needs. Due to their scalability, self-learning capabilities, and statistical approaches, AI/ML techniques are recommended for incorporating continuous quality into the entire solution as well as individual steps.

As mentioned previously, the first hurdle is correctly labeling the data. Our historical records lacked labels denoting failure and non-failure. We used AI/NLP algorithms to analyze error logs and textual data for key terms and flagged records as failure or non-failure. Now, the next task is to have fewer failure records, i.e. the minority class was less than 5%. As a result, it has an unbalanced data collection and datasets that are skewed toward one class. It is unsuitable for AI/ML techniques to produce accurate and dependable results, and a few machines lacked the necessary data to train AI/ML models. We addressed the imbalanced data set issue by oversampling machine learning approaches via SMOTE algorithms and converting the data set into an imbalanced data set only during the training phase. Additionally, to address the absence of data difficulty in a few cases, we generated synthetic data using Unsupervised Deep Learning algorithms (Gans).

Figure: Quality Engineering during preprocessing and feature engineering

The team utilized a variety of classification algorithms on training data sets, including logistic regression and tree-based models. Techniques for decision trees and ensembles — Random Forest and XGBoost. After adjusting and validating all algorithms many times, we built a bespoke ensemble methodology with voting for each method based on data segment. Following training, models are validated and deployed in production.

Automotive acceptance criterion testing and A/B testing were the solution's highlights, as we not only created a single challenger solution but also implemented four challenger models with the purpose of answering the following:

1. When is it necessary to replace an obsolete model with a new one?
2. How often should we proactively retrain our models?
3. When is it appropriate to revert to a previous version of the model?