Electric Compressor NVH Detection System Based on Acoustic-Vibration Signal Fusion

To address the vibration and noise generated during the pressure buildup process of an electric compressor at constant speed, a noise, vibration, and harshness (NVH) detection system was designed. Using cylinders and other actuating components, the product was clamped and sensors were mounted, while a soundproof enclosure was employed to reduce external noise interference. Based on the working principle of the electric compressor, a pneumatic system was designed to simulate real operating conditions. A real-time sound and vibration acquisition system was developed using an NI USB-4431 acquisition card, a triaxial accelerometer, and a microphone. Mel-frequency cepstral coefficient (MFCC) feature matrices were extracted from the sound signals, and short-time Fourier transform (STFT) was applied to the triaxial vibration signals to obtain feature matrices. A four-channel feature matrix was then constructed as input through acoustic-vibration fusion. Hyperparameter search was carried out using the Optuna framework, pruning, early stopping, and cross-validation to obtain the optimal parameter combination, and the results were compared with those of 1D-CNN, ResNet18, and MobileNetV3 models. Ablation experiments indicate that both multi-axis vibration fusion and acoustic-vibration fusion improve OK/NG classification accuracy over single-axis inputs. Based on the EfficientNet-B0 model, binary OK/NG classification under combined high-speed and low-speed conditions was performed, achieving a classification accuracy of 98.75%, with training time within 3 h and detection time within 9 s.