A multi-agent, adaptive FDD framework is presented in this study. The aim is to dynamically evaluate the confidence of a diagnosis agent in detecting certain faults and use the most reliable agent for the current fault by automatically switching between agents when the faulty conditions change. The framework consists of multiple agents that use partial least squares discriminant analysis (PLSDA) and Fisher discriminant analysis (FDA) and contribution plots techniques and effectively diagnoses all potential faults a complex interconnected CSTR network can experience. The performances of agents are evaluated under different fault scenarios and are recorded in a performance history. The FDD agents use the historical performance space to rank and select the agents with high performances for the current fault signature realized on the process variables. The multi-agent framework creates the environment for improved fault diagnosis via a performance-based consensus criterion as opposed to a voting-based criterion and enables further improvement in diagnosis in time via agent adaptation.
Multivariate statistical process control (SPC) methodologies are used to determine the variation thresholds of the normal operating conditions (NO) of a process that yield acceptable in-control products and to detect the variations from NO due to process abnormalities. Chemical processes experience abnormalities due to equipment malfunctions, process disturbances and environmental factors. In complex chemical processes, timely detection and diagnosis of process upsets is required to be able to take corrective action before the abnormal conditions spread and start affecting the entire process.
Statistical techniques such as principal component analysis (PCA), multi-block PCA (MBPCA) and dynamic PCA (DPCA) and their monitoring statistics, T2 and SPE are effective in detecting the process abnormalities; however, they do not provide information about the potential fault in effect. Combined fault detection and diagnosis (FDD) approaches have been proposed to timely detect the abnormalities and correctly diagnose the root causes. Only after fault diagnosis, it is possible to take the required corrective action. Contribution plots identify the variables that have been contributing to the inflation of the monitoring statistics. Linear classification techniques such as partial least squares discriminant analysis (PLSDA) and Fisher discriminant analysis (FDA) utilize clusters of data, each cluster representing a different fault type, and classify a faulty observation looking at its similarity to the existing fault clusters. These techniques have been used extensively in literature for various processes. In the design phase of FDD, it is difficult to predict which method would give correct diagnosis results for all possible faults a single process can experience.
An intelligent hierarchical framework for monitoring, analysis, diagnosis and control with agent-based systems, MADCABS, have been developed at IIT. The FDD agents that are embedded in MADCABS, agent cooperation and communications, performance management capabilities, effectiveness of a performance-based diagnosis scheme and agent adaptation will be presented in case studies in a simulated CSTR network.
The CSTR network is a (4 by 5) rectangular grid network with interconnected reactors. The network produces three product grades with a 30:30:40 production ratio using the same resource that is continuously fed to each reactor. Possible abnormalities that can occur in this network are a decrease in the feed flow rate, an increase in the outgoing and incoming interconnection flow rates to and from the neighboring reactors, and an impurity or one of the products mixing in from the feed.
Distributed monitoring and fault detection agents using PCA, DPCA and MBPCA are capable of effectively detecting the abnormalites in the system. In the object-oriented environment of MADCABS, every reactor in the network is monitored locally by all three agents simultaneously. The variables for each CSTR that are continuously monitored are the reactor concentrations, the feed flow rate and outgoing interconnection flow rates from each reactor. Every reactor in the network has diagnosis agents that are triggered when an abnormality is detected in the system. The FDD and mainly diagnosis activities are emphasized in this study.
A diagnosis agent for each reactor cooperates simultaneously with different discriminant agents that use PLSDA, FDA and contribution plots for diagnosis. In the training phase, the discriminant analysis models are built and each agent's diagnosis performance is recorded in a performance history space. The historical space is a map of the agent's performance for different fault signatures. In the testing phase, the historical space is used to predict an agent's reliability for a potential fault and keeps updated on an ongoing basis. MADCABS utilizes the performance space and online performance evaluations to determine the confidence of an agent for a potential fault and prioritize the best agents in a performance-based consensus for fault diagnosis. This enables the dynamic switching from one agent to another on the basis of their reliabilities for the current fault signature. The agents that provide false classifications can adapt by updating their models with the new faulty data and improve their reliability in consensus. The performance-based consensus yields fewer misclassifications than a voting-based consensus.
The multi-agent FDD framework, performance management mechanism, and performance-based consensus criterion will be presented and the improvement in the classification performance via performance-based consensus and agent adaptation in time will be illustrated for various faults.