Fault Diagnosis Algorithms Of Chemical Process Based On Improved Total Projection To Latent Structure

As various advanced control strategies and control methods have been adopted in modern industrial process, workers who directly participate in the control system are becoming less and less, which will improve equipment utilization greatly and promote the growth of national economy rapidly. This is a symbol of high level automation. The industrial process may convert to non-normal status due to harsh working environment, electromagnetic interference, aging equipment, human error operation and the other factors. In order to improve the control system’s automation level, real-time monitoring for the process is necessary. With the conclusion obtained by analyzing the non-normal process data, results and related operations must be carried out automatically and accurately. Fault diagnosis based on data-driven technique does not need to establish an accurate mathematical model. These motheds only need historical normal operation data and real-time sampling data to diagnose fault. This model has good practicality and versatility for the process which is difficult to establish precise mathematical model especially for the complex chemical process.The main purpose of actual industry process is producing target production with satisfied quality. Being different from the principal component analysis method which monitors the whole process, partial least squares method is suitable for diagnosing the faults which would have an impact on target production’s quality. PLS uses quality data to regress process data and it belongs to biased estimate. What’s more, PLS has been widely used when the final production’s quality is the main purpose of monitoring.This article focused on the strong nonlinear problems between process variables and quality variables in chemical process. In order to obtain better fault diagnosis result, total projection to letant structure (T-PLS) algorithm which has a high fault detection rate when detecting quality-relevant fault has been improved. The main works of this article are as follows:1. T-PLS model is a kind of linear algorithm essentially. According to the shortcoming of T-PLS model that it has a higher false alarm rate and a higher missing report rate when it is used to detect fault in the nonlinear process, kernel T-PLS algorithm has been proposed. This algorithm uses the kernel function to realize linearization between process data and quality data by projecting them from low dimensional process space onto high dimensional feature space through nonlinear function first. And then constructs a linear T-PLS model with kernel matrix in feature space. The validity and effectiveness of proposed algorithm have been verified by TE benchmark process. Simulation results of various fault modes show that KT-PLS model has higher fault detection rate than that of T-PLS when monitoring the strong nonlinear process.2. According to the disadvantage of single kernel function which is difficult to capture the process’global and local characteristics at the same time, hybrid kernel T-PLS model combined with priori knowledge of the actual process is proposed in this paper. This model uses radial basis function which owns local characteristic and polynomial function which owns global characteristic composing the hybrid kernel function to linearize the nonlinear process data. Hybrid kernel T-PLS model adjusts the weighted parameter of each function according to the actual process’prior knowledge, so as to obtain higher fault detection rate. Simulation results of TE process show that hybrid kernel T-PLS model can give a higher fault detection rate attributing to the hybrid kernel function used in the nonlinear process.3. According to the shortcoming that traditional contribution plot would lead to inaccurate fault identification results due to process data smearing effect caused by closed loop control strategy, generalized reconstruction-based contribution plot has been proposed in this paper. This method uses the priori knowledge of actual process to determine the subset of candidate faults after the traditional contribution plot giving out each variable’s contribution. Then reconstructs the fault data along the direction of candidate subset, uses the reconstruction data computing monitoring statistics. The actual fault can be determined if the monitoring statistics fall down their respective control limits after the fault data being reconstructed along a particular fault direction. Generalized reconstruction-based contribution plot is applied to TE process, and simulation results show that generalized RBC is able to identify the process’ fault accurately, which verifies the validity of this method.4. It is difficult to obtain complete real time measurement data in batch process, what’s more, raw materials of each batch process always change and the changeable operating condition may result in working point and control limit drift. According to these problems, multi-way kernel T-PLS algorithm is proposed in this paper. This method translates the three dimensional process data into two dimensional data matrix according to the sampling order first, and then uses kernel function to solve the nonlinear problem among process data, finally establishes linear T-PLS model in the high-dimensional feature space to diagnose fault. Statistics’control limits need to adjust in time according to the difference among each batch process data so as to make sure the latest batch’s fault diagnosis accuracy. Simulation results of penicillin fermentation process show that multi-way kernel T-PLS algorithm has a higher fault detection rate than multi-way kernel PLS. What’s more, this method has a universal applicability for batch or semi-batch process.