Self-Supervised Learning for Robust Fault Diagnosis in Electrical Motors

By Dr. Sampath LonkaReviewed by Susha Cheriyedath, M.Sc.Aug 3 2023

In a recent paper submitted to the arXiv* server, researchers introduced a groundbreaking framework for fault diagnosis of electrical motors, addressing the limitations of traditional approaches that assume uniform data distribution and require substantial labeled data. The proposed model leverages self-supervised learning and fine-tuning on a neural network-based backbone, enabling high-performance fault diagnosis with minimal labeled data.

*Important notice: arXiv publishes preliminary scientific reports that are not peer-reviewed and, therefore, should not be regarded as definitive, used to guide development decisions, or treated as established information in the field of artificial intelligence research.

Background

Electrical motors play a critical role in diverse industries, ensuring reliable operation in applications like aerospace, robotics, and defense. Nevertheless, these motors are prone to faults, which can impair performance and result in downtime. Traditional fault diagnosis methods often struggle with different fault types and severity levels due to their reliance on uniform data distribution. Moreover, manual data annotation is expensive and time-consuming.

To overcome these challenges, the researchers propose a new framework for the diagnosis of electrical motor faults. It consists of two main steps: first, training a neural network-based backbone using self-supervised learning to extract high-level features, and second, fine-tune the backbone with limited labeled data to capture finer details. This approach allows the model to adapt effectively to diverse fault diagnosis tasks, requiring fewer labeled samples than traditional supervised learning methods, making it practical for scenarios with scarce or expensive labeled data.

Related work

In the industrial landscape, machines continue to advance in precision, automation, efficiency, and complexity. However, this progress also entails a heightened risk of breakdowns and accidents. Consequently, fault diagnosis of electrical motors becomes crucial for ensuring equipment safety and reliability. Existing fault diagnosis methods can be categorized into three groups:

Signal processing techniques: These methods can identify fault types and locations, but they rely on specialized knowledge not always available to maintenance personnel, and their outcomes may be challenging to interpret.

Neural network-based methods: methods such as artificial neural networks (ANN), machine learning algorithms, and convolutional neural networks (CNNs) have been widely used but assume consistent data distribution, making them less robust in varying conditions.

Transfer learning: While transfer learning aids in knowledge acquisition from related scenarios, it may lose effectiveness with substantial differences between the source and target domains.

Methodology

Foundational models

Foundational models leverage transfer learning and scale. While foundational models have been widely adopted in natural language processing and computer vision tasks, their application in electrical motor fault diagnosis remains unexplored. The suggested approach for diagnosing faults in electrical motors introduces an innovative implementation of foundational models specific to this field. The framework for constructing this foundational model for electrical motor fault diagnosis comprises training andfine-tuning the backbone model.

Training the backbone model: The researchers trained a 1D CNN-based backbone model using a dataset that includes vibration, temperature, current, and acoustic data for various mechanical faults and healthy scenarios under constant and variable speeds. The model is comprised of convolutional layers, a global average pooling layer, and a fully connected output layer using softmax activation. The Adam optimizer and sparse categorical cross-entropy loss function are employed, contributing to the model's effectiveness and efficiency.

Defining target tasks: The target tasks are categorized into two groups: (a) evaluating the foundational model's performance within the same machine under different fault cases and operating conditions and (b) assessing its adaptability across different machines. These tasks are designed using a layered approach, beginning with the classification of samples into healthy and faulty categories, followed by the diagnosis of fault type, location, and severity. The tasks are executed at both constant and varying speeds. Additionally, the model's scalability and adaptability are tested using data from a different motor, and the impact of white Gaussian noise is examined to assess generalizability and robustness.

Fine-tuning the backbone model: Fine-tuning, a form of transfer learning, adapts the pre-trained neural network to new tasks or domains. The pre-trained model's initial and final dense layers are thawed, and then they undergo fine-tuning using a smaller dataset specifically tailored to the new tasks. This distinctive fine-tuning method significantly expands the foundational model's abilities, enabling it to achieve exceptional accuracy in diagnosing faults not only within a single machine but also across different machines, even when there are limited labeled samples available.

Results

The practical effectiveness of the proposed foundational model for diagnosing electrical motor faults was evaluated through a thorough assessment. Fine-tuning the base model for multiple tasks highlights scalability, expressivity, and generalizability. The model efficiently acquires and understands real-world data, achieving accurate fault diagnosis with limited labeled samples. It maintains accuracy levels above 90% for various tasks with just 5% labeled data, accommodating diverse machine datasets. Generalizability is demonstrated through successful fine-tuning with 10% white Gaussian noise, indicating the model's robustness and adaptability to varying conditions.

Conclusion

In conclusion, the researchers introduced a novel foundational model-based approach for fault diagnosis of electrical motors, overcoming the limitations of traditional methods. The proposed framework leverages self-supervised learning and fine-tuning on a neural network-based backbone, achieving remarkable classification performance with minimal labeled data. The proposed model's scalability and generalization capabilities make it a valuable tool for fault diagnosis in electrical motors under diverse operating conditions. With potential for further extension and incorporation of physics information, our proposed foundational model offers a promising solution for real-world applications.

*Important notice: arXiv publishes preliminary scientific reports that are not peer-reviewed and, therefore, should not be regarded as definitive, used to guide development decisions, or treated as established information in the field of artificial intelligence research.