Gearbox Ferrography Wear Debris identification based on faster CNN

Gearbox has the advantages of compact structure and high transmission efficiency, which is the most widely used transmission component at this stage. Gearbox failure may lead to industrial production stagnation, and bring high maintenance costs, and even cause casualties, which is not conducive to the safe and economic operation of mechanical equipment. Therefore, it is very important to carry out the fault warning of gearbox.

At present, the main methods of gearbox fault warning are vibration signal analysis and oil analysis. The vibration signal analysis method can locate the fault parts and diagnose the fault types by analyzing the vibration acceleration signal in time domain and frequency domain. However, the vibration signal analysis method mainly aims at the abnormal vibration caused by parts deformation and other reasons, which can not make accurate early warning for the initial state of gearbox component failure. Oil analysis includes spectral analysis, ferrography analysis, particle technology and so on. Ferrography analysis method evaluates the wear status of gearbox from qualitative and quantitative perspectives by analyzing the type and quantity of wear particles in the oil. It can find the early unobvious faults and judge the parts of the equipment that cause wear and failure. Therefore, this method is widely used.

In recent years, wear debris image recognition technology has become the research focus of ferrography analysis. The wear status of gear box is evaluated by wear particle image preprocessing, feature extraction and wear particle classification. Many experts and scholars have done in-depth research and put forward many practical wear particle identification methods. A wear particle recognition system based on radial basis function (RBF) neural network is proposed. The shape, color and texture of wear particles are extracted by mathematical method, and the wear particles are classified by RBF neural network. Canny operator is used to detect the edge of wear particle image, and support vector machine is used to complete the classification of wear particles. Mining

The gray level co-occurrence matrix is used to extract the features of wear particles, and the BP neural network model improved by genetic algorithm is used to classify them. The accuracy of the above algorithms can reach 85%, but there are still some problems

(1) the efficiency of manual feature extraction is generally low, and if there are many kinds of wear particles in the image, it will increase the difficulty of feature extraction;

(2) when multiple wear particles appear in ferrography image, the system can not recognize all wear particles at the same time, only one of them can be recognized;

(3) the above algorithm can only identify the type of wear particles, but can not count the number of wear particles.

Deep learning method has the characteristics of automatic feature extraction, high accuracy and fast calculation speed, and has good performance in the field of target recognition. In view of the problems existing in the above Ferrography Wear Particle identification algorithm, ZHY gear introduces deep learning method to carry out ferrography analysis, and realizes the fault warning of gearbox. In this paper, fast CNN algorithm is used to identify ferrographic wear particles. Res net-34 network is used to automatically extract the characteristics of wear particles, which improves the detection speed and accuracy; RPN (region proposal networks) network and region of interest pooling network are used to accurately determine the location of wear particles, which is convenient for counting the number of various kinds of wear particles; The target regression network and the boundary classification network are used to classify the ferrographic wear particles, and the multi class cross recognition is realized with high accuracy.

Scroll to Top