Study On Theory And Methods Of Intelligent Fault Diagnosis Based On Kernel Algorithm

The new technique of fault diagnosis and monitoring of equipments is developing and perfecting continuously. It plays an important role in the modern duplicate productions with the characteristics that safeguards the safety production and prevents from the accidents and saves the maintenance costs. However, the more complex structures of the facilities and its closer inner connection increase the difficulties in diagnosing fault and monitoring the running state of the equipments. The new theories and methods have to be investigated in order to solve the problems encountered in reality. Since 1960s, researchers represented by Vapnik have devoted themselves to the study on statistic learning theory. They established a new type of learning algorithm, support vector machine (SVM), based on the statistic learning theory. It is the successful application of kernel function to SVM that the study on learning algorithm based on kernel functions or kernel algorithm for simplification has attracted great interest. Applying the kernel algorithm to fault diagnosis will solve the non-linear, imprecise and uncertain problems. This provides a completely new and feasible approach in the domain. Many problems are worth deeply studying and discussing about the practice of the approach for the technique of intelligent fault diagnosis, based on kernel algorithm, is a brand new field in the world.This paper provides the theoretical foundations for the applications of kernel algorithm to fault diagnoses though the deep and systematical study on the application of kernel algorithm to intelligent fault diagnosis, the processing of the uncertain information in the diagnosis, the real-time realization of fault diagnoses, the choice of kernel function and parameter optimization, multiple classes of fault diagnoses, and incipient fault diagnosis, and the sample data compaction. Thus, it promotes the development of fault diagnoses technique. The main tasks and the innovations works are as the follows.A posterior probability algorithm is presented based on the geometric distance to solve the problem that the miscarriage of justice in two classes causes the different loss in the fault diagnosis, Furthermore, a SVM method on the base of the minimum risk is proposed by combining the SVM with the Bayesian decision theory after the definition of the degree of diagnosis confidence. Finally, the method is validated by applying it to the practical fault diagnosis of electro-hydraulic servo valve.Two theorems about the radial basis function on the parameter condition of s→0 or s→∞are presented and proved aiming at the characteristics that the one-class of samples is trained by the one-class SVM. This paper explores the relation between the two types of support vectors (boundary support vectors and non-boundary support vectors) and the recognition rate of object; proposes an improved method of the model parameter choice of"leave one out"; which dramatically decreases the time of model parameter choice in the precondition of generalizing performance of classifier, so that the recognition rates of the objects and the non-objects are determined on purpose; presents a new one-class SVM learning algorithm based on time–rolling window for the fault diagnosis of dynamic system, which will contribute to the practical application of one-class SVM. In addition, two methods are presented though which the one class SVM is extended into multiple faults diagnoses. If the methods are applied to the fiducially database and the hydraulic pressure pump respectively, we can solve the problem existing in the method of the available SVM multi-class classification that the object does not belong to any class or the object belongs to more than one class simultaneously and speed up the training and decision making of the algorithm.Aiming at the difficulty of choosing the parameters of support vector regression (SVR) model, an automatically optimized method of SVR parameter is presented based on the genetic algorithm after the influence of each SVR parameter on SVR performance. In addition, incipient fault diagnosis and a method of weak information retrieval in the background of heavy chaos are created by using the predictive SVR model. Simulation shows that the method has a more stable performance and a more general characteristic.Finally, the boundary theorem of the residual error estimation is presented after discussing the problem of the dimensional reductions of kernel matrices in detail. With the consideration of data's correlation and minimal residual norm, the heuristic algorithm, which is the greedy algorithm, is proposed for the dimensional reductions of the kernel matrices. Also, a kind of sparse regression algorithm is presented based on the greedy algorithm in the reproducing kernel Hilbert space.