Traditional motor fault diagnosis approaches frequently rely on domain expertise and face limitations in analyzing high-dimensional, complex signal data with high accuracy. To overcome these constraints, this study introduces a novel diagnostic method that integrates SDP (symmetric dot pattern) visualization with a MCFRes-CapsGNN (multi-channel residual fusion graph capsule network). This method leverages visual features from multi-source signals, creating a comprehensive and detailed representation of motor faults, which enhances diagnostic precision and robustness. A major advantage of this approach is its ability to capture spatial and hierarchical relationships within the signal data by structuring them into graph-based data, enabling a more reliable automated diagnosis across various fault types.

The proposed methodology was initiated by transforming one-dimensional time-domain vibration signals into two-dimensional SDP images via polar coordinate transformations. This conversion produced intricate "snowflake-like" patterns that can distinctly represent different motor fault types. While traditional frequency-domain analysis methods require complex feature extraction and manual interpretation, SDP visualization offered a more direct and interpretable depiction of signal features, allowing for easier fault classification.

Subsequently, we employed superpixel segmentation on the SDP images, which divided each image into smaller regions based on attributes such as texture, color, and spatial positioning. These regions, called superpixels, enhanced image segmentation accuracy by reducing data redundancy and maintaining the boundary information critical to effective fault analysis. The segmented regions were then represented as graph-structured data, where each segment or superpixel forms a node, and relationships between segments (based on spatial proximity or shared characteristics) were modeled as edges. This graph structure provided a more flexible and informative representation of the signal data, as it allows the model to capture spatial and textural nuances effectively.

This graph data was then processed through the MCFRes-CapsGNN model, a unique structure that combined graph neural networks with capsule networks to manage both hierarchical relationships and spatial patterns. Capsule networks, by employing dynamic routing mechanisms, allow for complex pattern recognition that was robust to variations in input signal patterns. Additionally, multi-channel fusion incorporated different sensor data into RGB channels, producing a color-enhanced SDP image that enriched feature representation by integrating diverse sensor signals into a single, cohesive image. This multi-sensor fusion captured subtle variations across different channels, making the model more sensitive to intricate fault features.

Experimental validation was conducted using a dataset containing eight motor fault types, including rotor imbalance, bearing failures, rotor misalignment, and other common faults. Results from these experiments indicate that the proposed approach achieves an accuracy rate of 98.12%, outperforming traditional diagnostic methods such as LSTM (long short-term memory) networks, CNNs (convolutional neural networks), and standard CapsNet (capsule networks). The superior diagnostic accuracy observed is attributed to the combined effects of SDP visualization, which provides clearer feature differentiation, and multi-channel fusion, which enriches the feature space available for analysis. Additionally, the model's performance in terms of precision and recall demonstrates its robustness in identifying and categorizing motor faults accurately.

Beyond accuracy, the computational efficiency of the MCFRes-CapsGNN model represents a significant advantage. Compared to other models that rely on pixel-based computations, the proposed method reduces computational demands by utilizing superpixel segmentation and graph-structured data, both of which streamline the analysis without compromising diagnostic accuracy. This efficiency makes the model more suitable for practical applications where computational resources may be limited or where real-time analysis is required.

Integrating SDP and MCFRes-CapsGNN presents a promising solution for enhancing motor fault diagnosis, particularly in environments with complex and high-dimensional signal data. By converting vibration signals into intuitive visual representations and leveraging the graph neural network’s capability to capture structural and spatial features, this method serves as a robust alternative to traditional approaches. The combination of multi-sensor data, color fusion, and structured graph representation enhances the clarity and depth of feature representation, establishing a strong foundation for highly accurate motor fault diagnostics.

The method demonstrates considerable potential for applications in real-time industrial monitoring, where quick, accurate identification of motor faults is crucial to maintaining operational efficiency and minimizing downtime. The approach could extend to other areas requiring similar precision and reliability in diagnostic tasks, such as the aerospace, transportation, and manufacturing sectors, providing a new, scalable tool for fault detection and predictive maintenance.