Hunting for anomalies: Automated fault detection in electrical motors using ML

ABB has developed a highly effective and scalable capacity for automated fault detection using unsupervised lightweight models that retrain themselves. By linking parallelizing, anomaly detection, and post-processing into a pipeline ABB makes sure this solution isreliable, reproducible as well as scalable.

Subscribe to ABB innovation and technology news

_{Ugur Aydin, ABB Motion Services Helsinki, Finland, ugur.aydin@fi.abb.com}

According to a survey conducted by ABB with 3,215 organizations [1], industrial businesses experience a typical cost of about $125,000 per hour due to unplanned downtime. It was also found that 69 percent of the surveyed plants experience unplanned downtimes at least once a month. Clearly, equipment such as electrical motors is essential for the smooth running of many industries, and reliability is key to minimizing failures and the negative financial impact of downtime.

Nevertheless, despite their prevalence – more than 300 million operational motors in the world [2] – ensuring this reliability is an ongoing challenge.

Condition monitoring that utilizes advanced algorithms on operational data from motors could mitigate unplanned downtime by detecting anomalies and alerting operators to issues before they escalate into more serious problems.

Challenges for motor anomaly detection

The fault types in electrical machines can be divided into two main categories: electrical and mechanical faults. The main electrical and mechanical faults are summarized in →01. The physics behind these faults differs. Thanks to advanced signal processing, the symptoms can be observed by analyzing the vibration, stray flux, acoustics and electrical current characteristics of the motors. [3]. For instance, considering mechanical faults, these can be observed in specific harmonics of the motor’s vibration. 
Such harmonics can be revealed by performing a FFT on the measured vibration signals. On the other hand, wavelet and envelope analysis of the vibrations can also be used to observe various faults in a low-intervention manner. [4]. In earlier days, the monitoring of these harmonics was done manually or by setting simple thresholds on the amplitudes. These methods are neither effective nor practical, considering the number of assets that need monitoring, lack of universal thresholds and the influence of factors like machine characteristics, operating speed and load.

These issues have been partly addressed in recent years, with advances in the AI domain. For instance, existing literature suggests that supervised ML methods could be effective in detecting and classifying faults in electrical machines [5]. Such methods however, require labeled data from the studied machines: models developed for one machine may not generalize to others of different sizes, applications, or operating conditions. Consider the measurement of radial vibration signals for two identical healthy motors that are connected to different load types →02. Even though the motors run at their rated power and speed, the vibration patterns are significantly different. Therefore, the models trained on vibration patterns of motor 1 will most likely produce incorrect predictions for motor 2.

To address this challenge, one option is to collect more labels from various machines and explore additional fault cases to train larger or multiple models. However, this process is impractical, costly, and time-consuming. Alternatively, unsupervised methods could be used to detect anomalies caused by deteriorating machine health conditions, since labeled data is not required for training the model in these cases.

And yet, it is extremely difficult to train a single, generalizable unsupervised anomaly detector if a fleet of motors is being monitored, as is often the case. Hurdles exist, including factors such as motor design, size, mounting type, instantaneous torque, and speed, that impact the measured motor vibration and magnetic flux density patterns. Such challenges can be circumvented by employing dedicated unsupervised models for each motor, which is precisely what ABB has done.

Having selected this approach, two relevant criteria had to be in focus. First, to monitor a wide array of motors, a traditional lightweight ML-based model that does not compromise performance should be chosen, rather than the computationally heavy deep neural network (DNN) method. Second, a consideration of the proper infrastructure was necessary in order to support the deployment and maintenance of models across a vast number of motors.

Considering these criteria, ABB’s data scientists have developed a highly effective and scalable capacity for automated fault detection for motors. The detection feature incorporates both pattern-recognition-based detection of quickly developing serious faults and an unsupervised machine learning (ML) approach for identifying early-stage faults. This paper focuses on the ML-based approach.

Developing unsupervised anomaly detection for electrical motors

The complexity of developing reliable, reproducible and scalable ML solutions is troubling for many organizations; this creates an obstacle for companies to take advantage of these solutions in real-life cases [6,7]. To tackle these challenges, ABB benefitted from best practices in the disciple of ML Operations (MLOps) in developing their anomaly detector solution →03.

Although the unsupervised model training does not require labeled data, ABB collected data from various motors with known healthy and faulty periods to assist with model validation, hyperparameter tuning and model selection. Once selected, the model only requires unlabeled data to be trained. Analysis of the labeled data showed that the data is highly imbalanced, with the faulty class in the minority. All faults in the studied cases can be detected effectively using vibration measurements.

First, the data has been preprocessed by removing the missing values, splitting the data into train and test sets, removing the outliers from the training data and standardizing the data. Afterward, during the feature engineering step, new features were created from existing data to enhance anomaly detection performance. Here, a categorical feature has been generated to group measured data points based on their deviation from the rated speed or slip. For variable speed drive motors, speed was used, while slip was used for direct online motors. This new feature helps to effectively consider the operating characteristics of the motor.

The final development step involved experimentation for model selection. The model has been trained using a predefined initial number of healthy data points. As evaluation metrics for such a binary classification problem with highly imbalanced data, precision-recall area under curve (PR-AUC) and Fβ-score with β=0.5 were selected to give more weight to precision and reduce false positives. PR-AUC and Fβ-score were used for model hyperparameter tuning and anomaly threshold optimization, respectively. In addition to these evaluation metrics, a constraint was employed to consider the computational burden of the calculations. One-Class Support Vector Machines (OCSVM), Isolation Forest, Minimum Covariance Determinant, Robust Random Cut Forest and Local Outlier Factor algorithms were studied; OCSVM was determined to be the best algorithm in terms of evaluation metrics and computational effort →04.

Model deployment

The deployment of the model was completed by connecting all the development steps into a pipeline. Considering the infrastructure design for deployment, one of the most difficult challenges is to design an infrastructure that can train and serve thousands of models in a reasonable amount of time. To this end, ABB utilized the parallel processing capabilities of Microsoft Azure Machine Learning Studio, where the job executions can be distributed to parallel compute clusters and nodes [8]. This can be scaled by increasing the number of compute-clusters or nodes, which comes with a cost. Thanks to the light computational burden of the OCSVM algorithm, only affordable CPU clusters were used.

The pipeline consists of three key steps: parallelizing, anomaly detection, and post-processing. The parallelizing step checks the available data for a specific date, reads the corresponding connected asset IDs, and divides them for parallel processing. The anomaly detection step is the core of the solution, encompassing data extraction, validation, preprocessing, feature engineering, model training, and prediction. Model training is automatically performed using the first pre-determined amount of data from the motor, assumed to be from a healthy operation. The trained models are versioned and registered in the model registry for use in future predictions.

The post-processing step calculates the importance factor for anomalies, enabling appropriate user notifications. Finally, the results are presented in a user interface (UI) for further analysis.

One of the important aspects of the pipeline is the automated model retraining that is done regularly on a schedule. The selection of the training data for updating a model for a particular motor is done by random sampling from the historical healthy data points determined by the model, with higher weight given to the newest data points. Such model retraining helps the models to consider the environmental influences that could appear due to, eg, variation in seasonal temperature, humidity fluctuations and the addition of newly installed equipment within the vicinity that might impact the vibration measurements.

UI for the motor anomaly detection solution

The user interface (UI) for the motor anomaly detection pipeline offers a comprehensive set of features that include a table that provides a summary of information regarding motors with detected anomalies →06. Users, such as data analysts, can interactively apply filters to access the specific data they require. Additionally, the UI includes a line plot that displays the maximum RMS vibration along the 3-axis, highlighting any anomalies that might be present →07.

Furthermore, the UI is equipped with a data annotation feature, which allows users to provide feedback to the model by annotating the data. These annotations play a crucial role in monitoring the performance of the model. As more labels are collected, it becomes possible to explore the potential of supervised approaches. In addition, users have the option to request retraining of the model to obtain a new version. This control over model retraining is useful. For example, when a motor is repaired or replaced, the vibration characteristics might change completely compared to the past vibration features. Users can also revert the model to a previous version, if they observe better performance in one of the previous versions.

A case study

The effectiveness of ABB’s anomaly detection function has been demonstrated in a specific industrial case. An early anomaly detection flag was raised for a motor driving a fan with a belt. Thanks to the notification, further analysis of the raw vibration data by an ABB expert using advanced signal processing tools was performed and revealed the occurrence of elevated harmonics related to the driven equipment →07. Upon physical inspection, it was confirmed that the bearing on the driven fan and the belt in between were damaged.

This case study demonstrated that the anomaly detector is even capable of successfully identifying a problem that originated from the application rather than from the motor itself, thereby preventing potential unplanned downtime. In this way, ABB shows that the anomaly detector, using the ML-based approach, is ready to help users in the future resolve fault issues early, thereby diminsihing downtime.

_{Acknowledgements
The author would like to extend his gratitude to the data scientists Ignacio Rodriguez Burgos and Dominik Suszalski and software developer Piotr Drabik in ABB’s Motion Ser­vices Analytics team, who have contributed massively to the development of the AI-based detection feature. Additionally, thanks are extended to Matthew Gladden for his meticulous review and editing of the article.}

_{References
[1] ABB, “Value of Reliability: ABB Survey Report” ABB press release website, 2023, Available: https://search.abb.com/library/Download.aspx?DocumentID=9AKK108468A6878&LanguageCode=en&DocumentPartId=&Action-=Launch [Accessed: Dec. 11, 2023.]
[2] ABB, “ABB Inverter Duty Motors”, ABB Technology website, 2023, Available: https://global.abb/group/en/technology/did-you-know/inverter-duty-motors [Accessed: Dec. 11, 2023.]
[3] Gundewar, et al., “Condition Monitoring and Fault Diagnosis of Induction Motor”, Journal of Vibration Engineering & Technologies, vol. 9(4), 2021, pp. 643 – 674, Available: https://doi.org/10.1007/s42417-020-00253-y
[4] Tse, P. W., Peng, Y. H., & Yam, R. “Wavelet Analysis and Envelope Detection for Rolling Element Bearing Fault Diagnosis—Their Effectiveness and Flexibilities”, Journal of Vibration and Acoustics, vol. 123(3), 2001, pp. 303 – 310, Available: https://doi.org/10.1115/1.1379745
[5] Surucu O., et al., “Condition Monitoring using Machine Learning: A Review of Theory, Applications, and Recent Advances”, Expert Systems with Applications, Vol 221, 2023, pp. 119738, Available: https://doi.org/10.1016/j.eswa.2023.119738, [Accessed Nov 21, 2023.].
[6] Baier, L., et al., “Challenges in the Deployment and Operation of Machine Learning in Practice”, 27th European Conference on Information Systems (ECIS), Stockholm & Uppsala, Sweden, June 8 – 14, 2019.
[7] Paleyes, A., et al., “Challenges in Deploying Machine Learning: A Survey of Case Studies”, ACM Computing Surveys, vol. 55(6), 2023 pp. 1 – 29, Available: https://doi.org/10.1145/3533378
[8] Microsoft web site, “What is Azure Machine Learning?”, 2024, Available: https://learn.microsoft.com/en-us/azure/machine-learning/overview-what-is-azure-machine-learning?view=azureml-api-2 [Accessed 29 Feb. 2024.]}

Explore ABB Review