Study On Method Of Fault Diagnosis Based On Acoustic Images Pattern Recognition

Though the fault diagnosis is usually based on vibration signals, the vibration sensors cannot be mounted on machinery easily at some situations. Therefore, the vibration based diagnosis technique has some restrictions, which can be solved by some non-contacted measurement based diagnosis techniques, such as the acoustic based diagnosis technique (ABD). As the result of vibration emission in air, machine sound signals carry ample information about the mechanical working conditions. The ABD technique possesses the advantage of non-contacted measurement and simple operation, which can partly take place of the vibration based fault diagnosis. The ABD technique has great potential, but also has some shortcomings, which make it develop slowly. For the traditional ABD technique, the sound signals are sampled by single channel and only the acoustic characteristics changing with the time and frequencies at one position can be obtained. Because the sound signals have poor anti-interference ability, especially in coherent condition, choosing measuring positions is another difficulty for ABD technique, which can influence the diagnosis effect severely and usually be determined by experience. The signals measured at inappropriate position may be insensitive to the mechanical faults, and then the diagnosis results would be not credible. Especially when the fault information is covered by the interference sources‘sound signals in coherent condition, the traditional ABD technique would be not applicable. The different mechanical faults can result in different mechanical sound fields, but not at any positions, at which the acoustic features may have similar changes in different working conditions. If the changes of whole sound field can be utilized, the fault diagnosis will become easily. And especially when some fault-related sound sources or parts are highlighted appropriately and the interference signals are weakened, the diagnosis effect will be further improved. Based on above thinking, an acoustic image based fault diagnosis technique is presented in this work. The new ABD technique combines with array measurement, acoustic imaging, spectrum analysis, image processing, feature extraction and pattern recognition techniques. It takes advantage of the information of source position and pressure or vibration distribution for fault diagnosis. The acoustic image based fault diagnosis widens the applications of acoustic images in the area of fault diagnosis and would be new choice for fault diagnosis. The beamforming based diagnosis technique, the near field acoustic holography (NAH) based diagnosis technique and the cyclostationary NAH (CYNAH) are developed in this work, which can be applied to different situations, such as the different frequency ranges, measuring distances and sound fields. A gearbox experiment is studied in this work to verify the acoustic image based diagnosis technique. The main contents of this dissertation can be summarized as follows.Firstly, the development of fault diagnosis is briefly introduced and the intelligent fault diagnosis that the direction of development and research in fault diagnosis is highlighted. The development of the acoustic based fault diagnosis is summarized, which is based on single channel measurement. Though the ABD technique has great potential, the poor anti-interference ability of acoustic signals makes it develop slowly. During the review of acoustic imaging technique, it is found that the acoustic imaging technique has a certain anti-interference ability. It employs the array measurement to sample acoustic signals and reconstructs the pressures or velocities on the source surface by some imaging algorithms. Based on the relation of vibration and sound, it is feasible to apply the changing information of pressure distribution in different working conditions to fault diagnosis, which can be taken as the research basis.After that, a fault diagnosis technique based on acoustic image pattern recognition is proposed, which is based on the relation of fault diagnosis and pattern recognition. The new diagnosis technique applies the techniques of image processing, feature extraction and pattern recognition to the acoustic images for fault diagnosis. It combines multidisciplinary research results, and therefore its research, development and improvement need to constantly add new researches of all fields, which shows the great potential for development of the new ABD technique.The acoustic images obtained by different imaging algorithms have different resolution and accuracies, which relates to the sensitivities of fault and the number of fault information. If the acoustic images are more sentive to the faults and include more abundant fault information, it will be easier to diagnose the weak mechanical faults.On the contrary, only the mechanical states can be judged. Therefore, the acoustic images obtained by different imaging techniques can be employed in different conditions. The acoustic images obtained by far-field test based imaging technique, which are limited by the?Rayleigh Criterion?, have low resolution. The missing of?evanescent wave?makes the far-field acoustic images has low ability to describe the details of texture information and not reconstructs the true amplitueds and phases of the sound pressures. Whereas, the near-field acoustic images have high resolution and can describe the subtle changes in the sound field, which can reconstruct the real amplitudes and phases of sound pressures. Therefore, the far-field acoustic images can be used to judge the working states of machinery in which the mechanical failure can change the sound field more obviously, the near-field acoustic images can not only be used for state judgments, but also diagnose the weak faults. Although the near-field acoustic images contain more fault information, the cost of its measurement and calculation is high, and its applicable frequency range is also restricted. On the contrary, the measurement and calculation of far-field imaging techniques are convenient. Therefore, the costs of measurement and calculation must be taken into account during the selection of acoustic images for fault diagnosis. In accordance to the diagnosing idea of the acoustic image based diagnosis method, the beamforming image based fault diagnosis technique is developed, which is preliminary implementation of the former. Beamforming is a simple acoustic imaging algorithm and has been applied broadly in engineering. It belongs to far-field measurement, so the resolution is less. The feasibility and effectiveness of beamforming based diagnosis technique is verified by the simulation and experiment of abnormal sound source identification. Beamforming has low resolution and is not suitable to identify the small and continuous sources emitting low-frequency sound. These shortcomings restrict the application of beamforming based diagnosis technique. To solve the problems, an NAH based diagnosis technique is developed. It can reconstruct the real vibration and sound distribution on the source surface, and has high resolution. The subtle changes of the sound field can be described by the NAH images, so the diagnosis of weak faults can be possible. In this work, the feasibility and effectiveness of NAH based diagnosis technique is tested by a coherent excitation experiment, in which one rib plate with stiffened boundary conditions is excited by several forces. One common single microphone measurement based diagnosis method is also carried out in the experiment for comparative purpose. The results show that NAH based diagnosis technique is superior in coherent condition.The SGLCM-based textural features and singular value features of the acoustic images and SVM are employed in the new ABD diagnosis, which are commonly used in their respective fields but not necessarily the best. The choice of the feature extraction technique and classification machine need to constantly test and can be considered as the section of innovation and improvement.From the acoustic images, the sound sources can be located and the changes of pressure distributions can be reflected. Based on the acoustic image‘s physical characteristics, a block feature extraction method is introduced, by which the source positions and the changes of local sound pressure distributions are highlighted and the diagnosis effect is effectively improved. For the common essential defect of singular values (SV), of which two arbitrary images have two different base spaces, an improved class estimated basis singular value decomposition (CSVD) technique is proposed and the diagnosis results are further improved. The improvement of feature extraction according to the essential characteristics of acoustic image provides the idea and direction for future research, and also shows the potential and development direction of acoustic image based diagnosis technique.To further verify the practical value of the new ABD technique, it is applied to the gearbox fault diagnosis. The gearbox experiment is carried out in a semi-anechoic room, by which the environment and fault simulation are very ideal and the random noise is reduced as much as possible, so nice diagnosis results are obtained. The validity of the method is further verified and its further research is prepared. Because the gearbox is typical rotating machinery, its radiation sound field is cyclostationary. The fault signals generated by pitting and partial broken teeth faults are special cyclostationary signals. Although the use of traditional stationary sound field based NAH has gotten good results, the spectrum or PSD distribution of sound pressure displayed in the original NAH hologram cannot show the statistical information of sound energy varying with time, which makes the original NAH contain incomplete sound field information. Therefore, for the gearbox‘s cyclostationary sound field, the CYNAH is introduced and the CYNAH based fault diagnosis technique is developed. Its feasibility is verified by the gearbox experiment. The CYNAH based fault diagnosis technique can be one new choice for the rotation machinery fault diagnosis.In this dissertation, several acoustic image based diagnosis techniques are developed by constantly introducing several acoustic imaging techniques, which can be applied to different situations, mechanical structures and frequency ranges. Their feasibilities and effectiveness are verified by simulations and experiments, all of which build the prototype of acoustic image based fault diagnosis firstly. The new ABD technique takes full use of the physical meaning of acoustic image and introduces the acoustic image into fault diagnosis. It provides a new way for ABD and further promotes the development of multi-field intersection.