Symmetrized Dot Patterns and CNN-Based Acoustic Signal Analysis for Fault Diagnosis in Internal Combustion Engines

Area of Science:

• Mechanical Engineering
• Signal Processing
• Artificial Intelligence

Background:

• Early fault detection in internal combustion engines (ICEs) is crucial for operational reliability.
• Conventional diagnostic methods have practical limitations, necessitating advanced non-invasive techniques.
• Acoustic analysis offers a promising avenue for condition assessment.

Purpose of the Study:

• To develop and validate a novel pipeline for identifying injector and ignition failures in ICEs using acoustic signals.
• To explore the efficacy of Symmetrized Dot Pattern (SDP) transformations combined with convolutional neural networks (CNNs) for fault classification.
• To establish a low-cost, non-invasive method for acoustic fault diagnosis in multicylinder engines.

Main Methods:

• Acquisition of acoustic measurements using smartphone MEMS microphones in a controlled environment.
• Transformation of acoustic signals into Symmetrized Dot Pattern (SDP) grayscale images.
• Training and testing a custom convolutional neural network (CNN) architecture on SDP images for fault identification.

Main Results:

• The proposed SDP-CNN pipeline achieved an overall accuracy of 81.11% in independent verification.
• High F1 scores were obtained for Spark Plug 3, Spark Plug 4, Injector 4, and normal operation (F1 = 1.000 and 0.947 respectively).
• Moderate performance was observed for Injector 1, Injector 3, and Spark Plug 1 (F1 ≈ 0.70-0.78) due to spectral overlap.

Conclusions:

• The study successfully demonstrates the feasibility of detecting injection and ignition faults in ICEs using acoustic measurements and the SDP-CNN pipeline.
• The approach offers a non-invasive, low-cost alternative for engine fault analysis utilizing consumer-grade sensors.
• Future work should focus on expanding datasets and validating performance under real-world conditions for improved generalization.