Enhancing Hydraulic System Reliability: AI-driven Fault Detection with ResNet-18

By Dr. Sampath LonkaReviewed by Susha Cheriyedath, M.Sc.Aug 15 2023

Hydraulic systems find use in diverse industries such as mills, manufacturing, robotics, and ports due to their advantages over electrical and mechanical systems. The demand for hydraulic systems is on the rise. Artificial intelligence (AI)-based methods can predict and classify faults, thus preventing breakdowns.

In a recently published article in the journal Sensors, researchers proposed a novel approach using a residual network (ResNet-18) that achieved about 95% accuracy in classifying cooling system faults in a hydraulic test rig, enhancing sustainable operations.

Background

As per Technavio, a prominent global market analysis company, the hydraulic equipment market is projected to grow at a compound annual growth rate of 4.71% from 2020 to 2025. The market recorded 3.23% growth in 2021, with the Asia-Pacific region contributing 47% to this expansion. The estimated incremental growth during this period is approximately USD 15.50 billion.

Hydraulic systems have broad applications across industries such as construction, manufacturing, and robotics, chosen for their efficiency, cost-effectiveness, and adaptability. They surpass electrical, mechanical, and pneumatic systems in load handling and power. While hydraulic systems are indispensable, their failures, often due to pollution, temperature, and fluid issues, result in costly downtimes.

Integration of deep learning for effective fault detection

Extensive research has been conducted on fault classification in diverse systems, including hydraulic setups. Hierarchical strategies, reinforcement learning, and linear discriminant analysis (LDA) have been employed for improved accuracy in prognostic health monitoring (PHM). Convolutional neural networks (CNNs) were utilized for bearing fault categorization, while supervised LDA with automated feature extraction aided in fault diagnosis.

Cloud-edge fault recognition using GA-based feature selection and long short-term memory (LSTM)-autoencoders was proposed alongside CNNs to address data acquisition challenges in hydraulic systems. Deep learning, supervised learning, and LSTM autoencoders were combined for hydraulic fault diagnosis. Comparative studies evaluated deep learning and traditional algorithms for fault prediction.

Correlation-based clustering and generative adversarial networks (GANs) were introduced for innovative fault detection and identification. These innovative techniques, spanning deep learning, reinforcement learning, CNNs, and GANs, are vital in enhancing fault detection accuracy and overall system performance in hydraulic systems.

Exploring hydraulic system datasets and methodologies

The hydraulic system dataset is publicly available from the machine learning repository at the University of California, Irvine (UCI). This dataset was prepared using a test rig equipped with multiple sensors, including pressure, volume flow, and temperature sensors, and others monitoring motor power, cooling efficiency, cooling power vibration, and system efficiency. The test rig operated through repeated 60-second constant load cycles to measure process values such as pressures, temperatures, and volume flows.

Furthermore, the dataset encompasses failure scenarios, portraying fault conditions of four primary components: cooler, internal pump leakage, valve, and accumulator. This dataset, consisting of 43,680 attributes and 2205 instances, is a valuable resource for training and evaluating predictive models for hydraulic systems.

Spectrograms depicting signal strength over time at various frequencies were generated. These visual representations employed the Fast Fourier transform to divide digitally sampled data into blocks, transforming them into vertical lines in an image. A total of 440 spectrograms were produced for different cooler fault conditions.

AI automates human intellectual tasks, finding applications in finance, criminal justice, search engines, robotics, and more. Machine learning (ML), a subset of AI, involves algorithms enabling computers to improve through experience. Deep learning (DL), an advancement of ML, utilizes neural networks to mimic human decision-making. CNNs are designed for object detection and image processing. ResNet-18, a CNN with 18 layers, improves accuracy and performance.

The current study employed MATLAB for model training with the ResNet-18 architecture, utilizing evaluation metrics such as accuracy, precision, recall, and F-score. These metrics offer insights into a model's performance, aiding in optimization.

Insights from spectrogram classification and model performance

In the results analysis, the generated spectrograms were classified into three categories: close to failure, reduced efficiency, and full efficiency, denoted as Class Output 3, 20, and 100, respectively. Of the 440 spectrograms, 146 were Class Output 3, 148 were Class Output 20, and 146 were Class Output 100. These spectrograms visually represent radiation power at different frequencies and times, indicating hydraulic system components and fault conditions.

During training, the model exhibited high accuracy, effectively discerning conditions. Precision and recall for diverse situations showed excellence, reflected in F-scores. Validation showcased notable accuracy and precision, resulting in considerable F-scores. Likewise, testing revealed a general accuracy rate, especially for specific scenarios. Corresponding recall rates and F-scores were in harmony with precision for each case.

Conclusion

The researchers utilized data from the UCI ML repository, including failure scenarios for primary hydraulic test rig components. The focus was on cooler fault conditions using spectrograms derived from raw text data. Employing the ResNet-18 architecture, the model achieved 96% accuracy with 95% precision, recall, and F-score. This work signifies the potential for accurate fault classification using this approach. Future direction includes identifying more effective methods yielding higher accuracy. Comparing different techniques' impacts and results can provide valuable insights. The proposed approach could extend to classifying other hydraulic system faults, enabling valuable comparisons.