Research And Implementation Of Device Anomaly Detection Method Based On Mobile Microphone Array

Currently,the majority of equipment’s anomaly detection relies on the vibration signals of mechanical devices.However,this approach poses challenges in terms of installation and maintenance.When equipment malfunctions,it generates a substantial amount of abnormal sound information.Leveraging mature signal processing techniques and algorithms for device anomaly detection proves to be highly feasible.Compared to fixed sensors,mobile MEMS microphone arrays offer greater flexibility and maintainability.Therefore,an intelligent anomaly inspection system was presented in this thesis,which integrates a two-degree-offreedom gimbal,MEMS microphone arrays,and device anomaly detection technology.The system employs a mobile car equipped with MEMS microphone arrays.Utilizing the gimbal,the arrays are directed towards the sound source of the target machine equipment.By employing a multi-layer autoencoder,useful information was extracted from the sound signals,enabling the detection of the sound state during machine equipment operation.This system finds application in the industrial sector,facilitating intelligent anomaly inspections of machine equipment based on sound signals.(1)To enhance the signal-to-noise ratio in the equipment workshop and the flexibility of the MEMS microphone arrays,a well-designed scheme was devised for detecting abnormal sound during equipment operation.This scheme ingeniously integrates the MEMS microphone arrays with the two-degree-of-freedom gimbal technology.Simultaneously,the autonomous navigation capability of the inspection car was utilized to approach the target equipment,ensuring effective collection of sound signals during its operation and thereby improving the accuracy of the anomaly detection system.(2)A machine noise collection device,incorporating a gimbal-mounted MEMS microphone array,was designed to rotate and align the MEMS microphone array with the sound source of the target machine equipment.This enhances the intensity of the target sound signals while reducing the influence of environmental noise.To address interference from coherent signals in real-world scenarios,an array virtual translation-based decoupling method combined with the RB-MUSIC algorithm was employed to estimate the positions and intensities of multiple targets with coherent signals at arbitrary angles.Experimental results demonstrate that this method achieves minimal spatial angle error and exhibits high stability when dealing with coherent signals.When the signal-to-noise ratio(SNR)exceeds 9d B,the average estimation success rate for the two-dimensional spatial angles of coherent signals reaches 90%.(3)Addressing the characteristics of strong nonlinear equipment noise and the difficulty in collecting abnormal samples,a sound anomaly detection method based on DCSSC-AE was investigated.This method linearly combines the time-frequency features of two-channel sound signals and introduces SSC to enhance the learning capability of DC-AE for normal sample time-frequency features,thereby improving the recognition rate of equipment anomaly detection.The experiments employ four types of machine sound samples from the MIMII dataset.The results indicate that DCSSC-AE achieves an average AUC of 0.848 across the four machine datasets,demonstrating a 20.4% improvement compared to DAE and a 35.9%improvement compared to OCSVM.Furthermore,DCSSC-AE exhibits higher accuracy in detecting anomalies in non-stationary sound datasets such as sliders and valves compared to stationary sound.(4)Real-world software and hardware implementations are conducted in the fan equipment workshop of a cigarette factory,testing the autonomous navigation of the car in the workshop,estimation of the target sound source direction,and detection of machine operation status.The test results demonstrate the system’s high reliability in intelligent anomaly inspections,thus providing preliminary validation of the effectiveness of the system described in this thesis.