Industry transformation with AI-based anomaly detection

Industrial processes and machines must act predictably and precisely. Anomalies, i.e. unexpected patterns in sensor data, can indicate problems such as faulty components or sensors. Anomaly detection based on artificial intelligence (AI) helps engineers to identify potential problems at an early stage, optimize maintenance plans and thus increase the efficiency of processes. Along with this, 86% of manufacturing executives believe that smart factories will determine their competitiveness in the next five years and AI will play an important role in manufacturing.

In the past, engineers and technicians relied on manual data inspections or automated alarms when sensor values exceeded defined thresholds. However, engineers can't analyze thousands of sensors simultaneously - this can cause them to miss anomalies that form complex hidden patterns across numerous sensors. Due to the increasing complexity of machines in modern factories, traditional methods of anomaly detection are no longer sufficient.

Faced with these challenges, engineers in the manufacturing industry are now using AI to improve the scope and accuracy of anomaly detection. AI algorithms can be trained with vast amounts of data from thousands of sensors to detect complex anomalies that the human eye alone cannot detect. By combining the amount of data that AI can analyze and the contextual expertise of engineers, manufacturing companies can develop a comprehensive anomaly detection solution.

Developing an AI-based solution for anomaly detection

Developing an AI-based anomaly detection solution is an extensive process, from planning and data collection to deployment and integration. To develop a solution that can effectively identify potential problems, engineers need a deep understanding of algorithm development as well as the operating environment.

Planning and data collection

The process of developing an AI-based anomaly detection system begins with defining the problem. This involves assessing the available sensor data, the components or processes, and the types of anomalies that could occur. For companies that are just beginning to use AI, it is important to start with a limited feasibility study, the success of which will bring clear benefits to the company before proceeding with larger initiatives.

Due to the increasing complexity of machines in modern factories, traditional methods of anomaly detection are often no longer sufficient

High quality data is crucial for AI systems. Engineers must therefore first define what constitutes an anomaly and under what conditions data is categorized as abnormal. Data acquisition involves the use of sensors to continuously monitor equipment and processes, as well as manual checks to ensure data accuracy.

Investigation and pre-processing of data

Most data for industrial anomaly detection comes from sensors that collect time series data such as temperature, pressure, vibration, voltage, layer thickness, pH and other values collected over time. Associated values can also be included, such as environmental data, maintenance logs and operational parameters. The first step in developing an anomaly detection algorithm is to sort and pre-process this data to make it suitable for analysis. This includes reformatting and restructuring the data, extracting portions of the data that are relevant to the problem, handling missing values and removing outliers. The next step is to select a technique for detecting anomalies. For this, the characteristics of the data, the type of anomalies and the available computing resources must be evaluated.

Model selection and training

It is important to experiment with different training approaches for an AI model to determine the optimal solution for a particular data set. At a higher level, AI techniques can be categorized into supervised and unsupervised learning approaches depending on the type of data available.

Supervised learning

Supervised learning is used for anomaly detection when parts of the historical data can be clearly labeled as normal or abnormal. The corresponding labeling of the data is often done manually by engineers and compared with maintenance logs or historical observations. By training on this labeled data set, the supervised learning model learns the relationships between the patterns in the data and their corresponding labels. Tools such as the Classification Learner in Matlab software help engineers experiment with multiple machine learning methods simultaneously to determine which model performs best. The trained model can then predict whether a new set of sensor data is normal or abnormal. In this way, for example, the plastic products manufacturer Mondi Gronau from Westphalia was able to predict potential machine failures for plastics production.

Unsupervised learning

Many companies do not have labeled anomalous data that is needed for a supervised learning approach. This may be because anomalous data has not been archived or because anomalies do not occur often enough for a large training data set. If most or all of the training data is normal, unsupervised learning is used for model training.

Unexpected patterns in sensor data, known as anomalies, can indicate problems such as faulty components or sensors

In an unsupervised learning approach, the model is trained to understand the characteristics of normal data and flags any new data that falls outside the normal range as anomalous. Unsupervised models can analyze sensor data to identify unusual patterns that may indicate a problem, even if this type of error has never occurred or been flagged before.

Feature exploration and extraction (feature engineering)

Although some AI models are trained with raw data from sensors, it is often more effective to extract useful features from this data before training. This process is known as "feature exploration and extraction" or "feature engineering". AI models can then learn more efficiently from the underlying patterns. Experienced engineers may already know the types of important features that should be extracted from the sensor data. Interactive tools for extracting and ranking the most relevant features in a dataset can help them to improve the performance of supervised and unsupervised AI models.

Some types of data, such as images or text, benefit from deep learning approaches that can automatically extract patterns without the need for explicit feature extraction. For example, the US company Imcorp combined time series and image-based anomaly detection to identify faults in underground power cables using deep learning. Although such deep learning approaches are powerful, they also require larger training data sets and more computing resources.

Validation and testing

Validation and testing ensures the reliability and robustness of AI models. Engineers usually divide the data into three subsets: Training, Validation and Test datasets. The training and validation data is used to fine-tune the model parameters during the training phase and the test data is used after model training to determine its performance with unknown data. In addition, engineers can evaluate the model based on performance metrics, such as accuracy and hit rate, and fine-tune it to meet the requirements of the specific anomaly detection problem.

Deployment and integration

A trained and tested model comes into its own when it is deployed in the field and makes predictions based on new data. Engineers consider factors such as computing requirements, latency and scalability when selecting a suitable deployment environment. Applications range from edge devices located close to the manufacturing process to local servers and cloud platforms with almost unlimited computing power, but also higher latency. Aerzen Digital Systems, for example, provides an integrated, cloud-based anomaly detection solution that detects problems in critical industrial facilities such as wastewater treatment plants.

The integration of anomaly detection requires the use of APIs (application programming interfaces) to access the model's predictions and build data pipelines. The latter ensure that the model receives correctly formatted and pre-processed inputs. This ensures that the model works together with other components of the application or system and achieves its full benefit.

First publication in Elektronik Praxis

Photos: Mathworks