Industrial processes and machines depend on predictability and accuracy. Unexpected patterns, or anomalies, in sensor data can indicate problems such as component failures or sensor deterioration. AI-based anomaly detection that helps engineers identify these potential problems early can help optimize maintenance schedules and improve process efficiency. AI is expected to play a significant role in manufacturing, as 86% of manufacturing industry executives recognize that smart factories will drive competitiveness within the next five years.


Modern manufacturing plants have become increasingly complex, making traditional anomaly detection methods insufficient. Engineers and technicians have had to manually inspect data or rely on automated alerts when sensor values exceed defined thresholds. Engineers cannot analyze thousands of sensors simultaneously, so they are likely to miss anomalies that occur in complex, hidden patterns across many sensors.

These challenges are driving engineers in today’s manufacturing industries to use AI to improve the scale and accuracy of anomaly detection. AI algorithms are trained on vast amounts of data from thousands of sensors, enabling them to accurately identify complex anomalies that are undetectable by the human eye. Manufacturing organizations can combine the scale of AI with the domain knowledge of engineers to build comprehensive anomaly detection solutions.

■ AI-based anomaly detection solution design

Designing an AI-based anomaly detection solution is a comprehensive process, from planning and data collection to deployment and integration. Engineers need to have a deep understanding of both the algorithm development and operational environments to develop solutions that can effectively identify potential problems.

■ Planning and data collection

Problem definition is the beginning of the design process for an AI-based anomaly detection system. This includes assessing available sensor data, components or processes, and the types of anomalies that can occur. For organizations new to AI, it is important to start with a limited-scope proof-of-concept project. Successful results from a proof-of-concept project can provide clear value to the organization before moving on to larger initiatives.

High-quality data is critical to AI systems. Engineers must first define what constitutes an anomaly and the conditions under which data is classified as an anomaly. Data collection involves continuous monitoring of equipment and processes using sensors and manual verification to ensure data accuracy.

■ Data exploration and preprocessing

Most data for industrial anomaly detection comes from sensors that collect time series data, such as temperature, pressure, vibration, voltage, and other measurements collected over a long period of time. It may also include related items such as environmental data, maintenance records, and operating parameters. The first step in designing an anomaly detection algorithm involves cleaning and preprocessing the data to make it suitable for analysis. This includes changing and reorganizing the data format, extracting the elements related to the problem, handling missing values, and removing outliers.

The next step is to choose an anomaly detection technique. This requires an evaluation of the characteristics of the data, the properties of the anomaly, and the available computational resources.

■ Model selection and training

Experimenting with different training approaches for AI models is crucial to finding the best model for a given dataset. Broadly speaking, AI techniques can be divided into supervised and unsupervised learning approaches, depending on the type of data available.

○ Map learning
Supervised learning is used for anomaly detection when chunks of historical data can be clearly labeled as normal or abnormal. Labeling is often done manually by engineers, who assign labels based on maintenance records or past observations. Supervised learning models learn relationships between patterns in the data and their labels by training on a labeled dataset. Tools like Classifier Learner in MATLAB help engineers experiment with multiple machine learning methods simultaneously to find the best model. Mondi Gronau, a global leader in packaging and paper products, has actually used supervised learning to predict potential failures in its plastics manufacturing machinery. The trained model can then predict whether a new chunk of sensor data is normal or abnormal.

○ Unsupervised learning
Many organizations do not have the 'labeled anomaly data' required for supervised learning approaches. This may be because the anomaly data is not stored, or because there are not enough anomalies to create a large training dataset. In most cases where the training data is normal, unsupervised learning is required.

In unsupervised learning approaches, the model is trained to understand the characteristics of normal data, and any new data that falls outside the normal range is labeled as anomaly. Unsupervised models can analyze sensor data to identify unusual patterns that could be signs of a problem, even if that type of failure has never occurred or been labeled in the past.

○ Feature Engineering
Some AI models are trained with raw sensor data, but it is often more effective to extract useful features from the data before training, through a process called feature engineering. Feature engineering is the process of extracting useful values from raw data, which helps AI models learn more efficiently from underlying patterns. Experienced engineers often already know the types of important features to extract from sensor data. Predictive Maintenance Toolbox provides an interactive tool for extracting and ranking the most relevant features from a dataset to enhance the performance of supervised or unsupervised AI models.

For some data types, such as images or text, you can take advantage of deep learning approaches that can automatically extract patterns without explicit feature extraction. IMCORP, a leader in underground cable life cycle condition assessment and performance, combined time series and image-based anomaly detection and used deep learning to identify faults in underground transmission lines. While powerful, these deep learning approaches require large training datasets and computational resources.

■ Verification and Testing

Validation and testing of AI models ensure reliability and robustness. Engineers usually divide data into three parts: training, validation, and test sets. Training and validation data are used to adjust model parameters during the training phase, while test data is used to judge the model's performance on data that the model has never seen before after it has been trained. Engineers can also use performance metrics such as precision and recall to evaluate and fine-tune models to meet the needs of specific anomaly detection problems.

■ Distribution and Integration

Once trained and tested, AI models become useful when they are deployed into production and begin making predictions on new data. Engineers consider factors such as compute requirements, latency, and scalability when choosing the right deployment environment. Deployment environments can range from edge devices located close to the manufacturing process to on-premises servers to cloud platforms with nearly unlimited compute power but high latency. Deployment tools such as MATLAB Compiler and MATLAB Coder enable engineers to generate standalone applications and code that can be integrated into other software systems. Aerzen Digital Systems, a provider of software and automation services for the industrial sector, deployed an integrated, cloud-based anomaly detection solution to detect problems in critical industrial facilities such as wastewater treatment plants.

Integration requires developing an API to access model predictions and setting up a data pipeline so that the model can receive preprocessed inputs in the appropriate format. This allows the model to work with other components of the application or system to provide full value.

■ Conclusion

AI-based anomaly detection is a major step forward in the journey toward manufacturing efficiency and cost effectiveness. AI, combined with the expertise of engineers and the latest technological advancements, enables manufacturers to prevent defects, optimize maintenance schedules, and improve overall productivity. Integrating AI into manufacturing processes can be a complex undertaking, but the potential benefits in terms of efficiency, cost savings, and competitive advantage are enormous. As the manufacturing industry evolves, AI’s role in driving innovation and operational excellence will continue to grow.

※ Contributor

Kim Young-woo, Executive Director of Mathworks Korea

Executive Director Kim Young-woo worked at Samsung Electronics for about 19 years and worked at Intel Korea R&D Center. He mainly worked on projects in the wireless communications field. He has been with MathWorks Korea for 17 years.