A new method of early fault detection and diagnosi

A new method of early fault detection and diagnosis in production process

classification number: tp277 document identification code: a

article number: (2000) a novel method of early fault detection and diagnosis

for technical process Yan Jun Wang Pei Hong Lu Zhen Zhong Chen Lai Jiu

(Power Engineering Department, Southwest University, Nanjing 210096, China) Abstract:based on the analysis of characteristics of two classes of technical process fault and defects of two family fault detection and diagnosis domestic consumption of plastic shopping bags is about 700000 tons s methods, this paper introduced an early fault detection and diagnosis method for technical process Taking advantage of quatitative redundant relationship of some dependent variables or symptoms in the process system after fault happened , The method avoid the lag which often appeared in family fault detection and diagnosis m reduces the vehicle weight of Cd level fusion to a level equivalent to that of B-level "Fiesta". Methods by utilization of dynamic road information of the variables or symptoms which is fused through artificial neural network Theoretical analysis and simulation examples demonstrated that this method could achieve early fault detection and diagnosis for technical process.

KEY WORDS：technical process； fault detection and diagnosis； Learning ▲ 1 Introduction production process faults can be basically divided into two categories [1]: one is that when the fault occurs, the process system gradually transitions from a normal stable state to another abnormal stable state. For example, in the regenerative system of the power station, when a high-pressure heater causes feedwater leakage due to the rupture of a steel pipe, due to the dynamic characteristics of the object itself and the role of the control system, the unit will gradually transition to another stable state and still be able to maintain operation; The other is that after the fault occurs, the system state will gradually deteriorate until it crashes. If the water-cooled wall leakage fault of the utility boiler is not handled in time, it will inevitably lead to the gradual deterioration of the unit operation until the furnace flameout or the unit forced to stop. The common point of the above two types of faults is that when the fault occurs, the process system will experience a specific dynamic process represented by multiple variables until the system stabilizes or collapses again

the fault detection and diagnosis methods of production process can be basically divided into fault diagnosis based on object mechanism model and qualitative fault diagnosis independent of object model [2]. The former uses observers or filters to reconstruct the state and parameters of the process system, and compares the model output with the actual output of the object to form a residual sequence, which is used for fault detection and diagnosis; The latter is essentially a process of more accurate identification and matching of data of typical fault modes, including fault diagnosis expert system, fault tree analysis, artificial neural network method, etc. this kind of diagnosis method often requires some special sensor detection devices to obtain the direct symptoms of faults. Generally, the above two kinds of fault diagnosis methods, whether residual sequence analysis or symptom signal matching, are based on the state of the system at a certain time. Only when the fault occurs to a certain extent, and the system state has a large deviation from the normal state, can the fault be diagnosed. Therefore, there is a lag in fault detection and diagnosis

the early fault detection and diagnosis method in the production process proposed in this paper qualitatively uses the analytical redundancy relationship between multiple related variables or symptom information in the system after the fault occurs, and effectively fuses their dynamic trend information using artificial neural networks, which avoids the shortcomings of the above conventional fault diagnosis methods to a certain extent and reduces the lag of fault detection and diagnosis, It has won the most fault handling time and initiative for operators. 2 basic principle of early fault detection and diagnosis in the production process the early fault detection and diagnosis method in the production process proposed in this paper realizes the early detection and diagnosis of faults in the dynamic change process of faults according to the dynamic change trend information of multiple variables. The fault diagnosis strategy is shown in Figure 1. Figure 1 early fault detection and diagnosis strategy in the production process

fig.1 early fault detection and diagnosis

algorithm for t at present, the sensors for tension machines in the market have small force values. Generally, S-type sensors are used. Because of their good generalization ability and fitting ability to any nonlinear function, as well as potential parallel processing, self-learning ability and fault tolerance, It has been widely studied and applied in fault detection and diagnosis [3, 4]. In this paper, the sliding data window is used to dynamically filter the original data, and then the dynamic trend information extraction of measurement data or symptoms and fault detection and diagnosis are completed through two-level and multi-layer feedforward neural networks, which greatly improves the robustness of fault diagnosis [5]. At the same time, due to the unique adaptive and self-learning functions of neural networks, the fault diagnosis system can always maintain a low missed detection rate and false alarm rate

2.1 sliding data window filtering

as a data filtering method, sliding data window [6] can effectively track the dynamic trend of variables, capture the gradual state of faults, and realize the normalization of input data. Figure 2 shows the basic principle of sliding data window filtering. Figure 2 sliding data window filtering principle

fig.2 principle of