Anomaly detection in industrial machines using explainable ai and acoustic signals

Industrial systems play a critical role in modern manufacturing, where efficiency, reliability, and sustainability are essential. Failures in these systems not only disrupt production and increase operational costs but also pose safety risks and lead to customer dissatisfaction. Early detection of anomalies and faults is vital for ensuring process continuity, improving product quality, and identifying potential issues before they escalate. This thesis focuses on the development of lightweight and effective supervised and unsupervised learning methods for detecting anomalies and fault conditions in industrial machines using audio signals collected from machine operations. By leveraging these methods, the study aims to enhance production efficiency and reliability while minimizing the risks associated with unexpected system failures. The thesis involves processing audio signals from sensors with feature engineering processes and testing them using various machine learning and deep learning models. Initially, 17 audio features were directly extracted from the raw audio signals. After these features were extracted, statistical feature extraction techniques were applied, such as mean, variance, standard deviation, skewness, kurtosis, maximum, and median absolute deviation autocorrelation coefficients, in order to derive additional attributes. As a result of the feature enhancement process, the initial set of 17 audio features was significantly expanded, leading to the creation of a robust and comprehensive dataset encompassing 110 distinct attributes for each instance. This expanded feature set allowed for a more in-depth and nuanced analysis of the acoustic characteristics of the signals, ultimately enabling a deeper understanding of the underlying machine behaviors and facilitating more accurate anomaly detection. These features were tested using supervised learning models, including Random Forest (RF), Extreme Gradient Boosting (XGBoost), Explainable Boosting Machines (EBM), and Deep Neural Networks (DNN), as well as unsupervised learning models like Autoencoder (AE) and Isolation Forest (IF). Supervised models demonstrated remarkable performance, achieving an average accuracy of 99.78% and an AUC of 99.95% on the MIMII dataset, which includes four distinct machine types. It is noteworthy that the study found that no single supervised model consistently outperformed others across all machine types. However, DNN and XGBoost generally exhibited superior performance across most machine types. Among the unsupervised models, the AE demonstrated superior performance compared to the IF across most machine types, except for the 'Pump 00' machine, where the IF model showed a slight edge. Specifically, the best average AUC values achieved by unsupervised models for each machine type were as follows: pumps (99.11%), fans (99.54%), slide rails (98.85%), and valves (98.12%). The thesis further substantiated the efficacy of its proposed methodologies by rigorously validating them on a variety of diverse datasets. Among these, the ToyADMOS dataset, comprising toy car audio signals, and a real-world case study involving acoustic signals from electric motors of white goods appliances, were employed. The proposed methods demonstrated remarkable performance, achieving competitive AUC values of 95.62% and 99.38%, respectively. These results not only highlight the robustness of the approach but also position it as a superior alternative, outperforming several state-of-the-art techniques documented in the current literature. In addition to evaluating performance, the thesis also explores the use of Explainable Artificial Intelligence (XAI) techniques to improve the interpretability of the machine learning models. Six different XAI techniques—SHAP, LIME, RF, IG, EBM, and LASSO—were integrated into the proposed framework to identify critical features specific to different machine types. This integration allowed for a better understanding of which features played the most significant role in detecting anomalies in pumps, fans, slide rails, and valves. By reducing the number of features to just 20–50 important attributes, the models not only maintained but also improved their accuracy, achieving a balance between computational efficiency and model effectiveness. The feature reduction process, particularly when applied to unsupervised models, significantly improved their classification performance. These findings suggest that focusing on the most relevant features is key to achieving better anomaly detection outcomes. The study identified several key features critical for detecting anomalies in machine behavior. Notably, the Standard Deviation of the Magnitudes, Standard Deviation of Instantaneous Amplitude, Kurtosis of the Signal, Standard Deviation of the RMS Energy, Maximum Spectral Bandwidth, and Maximum Spectral Rolloff emerged as crucial indicators across all machine types. These features, derived from acoustic signals, effectively captured subtle variations in the machine's operational characteristics, which were essential for distinguishing between normal and faulty conditions. A systematic comparison between supervised and unsupervised methods revealed a notable difference in performance. Supervised models achieved a higher average AUC score of 99.95%, compared to 98.91% for unsupervised methods, emphasizing the importance of labeled data in achieving reliable and accurate anomaly detection. Despite this, the unsupervised models still showed promising results, especially when coupled with feature selection techniques and XAI methods. One of the most significant contributions of this thesis is the development of lightweight models that are suitable for deployment in real-world industrial settings. Given the limitations of computational power and energy consumption in many industrial applications, the proposed models are designed to be computationally efficient without sacrificing performance. The lightweight nature of the models, which use architectures like DNN with just 3–4 hidden layers and Autoencoders, makes them particularly well-suited for edge devices. These models can be deployed on devices with limited resources, such as industrial sensors and controllers, thus addressing key challenges in practical applications. By focusing on models that are both interpretable and efficient, this thesis makes a substantial contribution to the field of industrial machine monitoring and fault detection. In conclusion, this thesis proposes a robust solution for anomaly detection in industrial machines, leveraging both supervised and unsupervised machine learning techniques. The proposed methods achieve high accuracy, strong interpretability, and computational efficiency, making them suitable for real-time monitoring and predictive maintenance applications in industrial settings. Moreover, the integration of XAI techniques enhances the transparency of the models, enabling actionable insights for maintenance teams and improving decision-making processes. The lightweight nature of the proposed models further ensures that they can be deployed on edge devices, offering a scalable and energy-efficient solution for industrial anomaly detection. Ultimately, this work not only addresses the current needs of industrial systems but also paves the way for future research in the field of explainable, scalable, and efficient fault detection systems for industrial applications.