Ramesh Rayudu & Sandhya Samarasinghe.

Artificial Intelligence (AI) systems are computer programs that are designed to reason and act in some ways like human beings or living organisms. Over the years, AI techniques have been developed and refined to the present stage where many techniques can be applied to practical situations. These techniques include Expert Systems (ES), Artificial Neural Networks (ANN), Model based Reasoning Systems (MBR), Genetic Algorithms (GA), Machine Learning (ML) and others (simulated annealing, case based reasoning, causal reasoning, hypothetical and cognitive reasoning).

Expert systems use domain knowledge in the form of if..then rules which are processed by an inference engine to find a solution. Expert systems can be goal driven in that for given observations, conditions or symptoms, rules are fired until a goal (eg. diagnosis)

is achieved; or event driven in that appropriate rules are fired without any goal as soon as conditions given in one or more rules are met (eg. open or close valves depending on change in pressures).

Artificial neural networks were inspired by the way our brain processes information. Input signals enter nerve cells (nodes) which process them and determine the response. In an artificial neural network, a large number of nodes interact together with many inputs to produce an output(s). Each node contains an activation function which determines how to process the incoming signals and produce an output. Neural networks are particularly useful for situations where the relationship between inputs and outputs are complex.

Genetic algorithms are computational optimisation techniques

based on the evolutionary concept of survival of the fittest. Algorithms mimic natural selection by evolving possible solutions that compete to survive by interacting and exchanging information until the systems converge towards a solution. Machine learning uses algorithms that automatically extract logically complex relationships existing in data or information into rules. Some algorithms are able to further generate generalised rules from the learnt knowledge.

Each of these techniques have been used to develop practical AI systems. For example, expert systems are used in control and maintenance of INTELSAT communications satellite, and many major companies such as AT&T, Bell, Unisys and IBM use them for several purposes ranging from asset management to fault diagnosis; neural networks are used in applications such as automatic control (smart washing machines and audio systems), classification and hand writing recognition; MBR is used in nuclear power plant control and medical diagnosis; GAs are used in operations research and ML systems are used to extract concept analogies in software engineering and database management systems.

Although there are many practical applications, each technique stated above has its own advantages and disadvantages and modellers who use a particular technique have to reach a compromise. For example, expert systems purely rely on human expert knowledge and sometimes their solving power is restricted to human experts' knowledge. They cannot learn from their experience and need extensive maintenance in upgrading. On the other hand, MBR systems can be more versatile than ES but they are slow and hard to model. ANNs are fast and can withstand noise in data but require longer training time and are error prone. In the same way, all other techniques have their own limitations that hinder their application to real world problems. A simple solution to overcome the limitations is to combine two or more AI techniques. The systems thus developed are called hybrid systems. The purpose of this article is to discuss some issues related to the development of hybrid systems for practical situations.

Alarm overload

The problem domain of this case study is fault diagnosis in power transmission networks. In New Zealand, the power network owned by Trans Power is controlled by two control centres: one in the North Island and the other in the South Island. These control centres control the power system using high powered computers and a real time data acquisition software SCADA (Supervisory Control And Data Acquisition). Whenever there is a fault on any equipment of the network, the protective devices operate making SCADA send an alarm (or list of alarms) to the control centre. The system controllers who control the network then analyse the alarm or alarms set, determine the problem and take necessary actions to restore the power. The problem with this fault diagnostic process becomes immediately apparent when a big event like lightning strikes the network and many protective devices operate as a result sending a long list of alarms to the control centre. This big list confuses the controller and adds to his already stressful situation of having to restore the power quickly. The solution to this problem is to filter the incoming alarms, analyse the filtered alarms and generate one or two alarms summarising what had happened. To achieve this objective, the following requirements are considered:

• Processing of alarms involve controllers' knowledge, the system has to be based on heuristic (experiential) and temporal (time) reasoning

• The system should also be able to handle problems that are not encountered by system controllers. Hence it should also include generate-and-test or reasoning based on simulation and analysis (model based reasoning)

• SCADA and protective equipment are prone to errors, the system has to withstand incorrect and missing alarm information

• It is hard to extract knowledge from human experts, the system should be able to learn heuristics from the output of simulation and analysis module.

We developed our intelligent fault diagnostic system for the power network as a hybrid system since all the above specified requirements cannot be met by just a single AI technique. Following is a general description of our hybrid system.

Since we need heuristic, temporal and model based reasoning, we chose to develop an architecture which incorporates all three reasoning mechanisms. The model, shown in Figure 1, has in its core a qualitative model of the power system represented by heuristic and qualitative rules. Specifically, all the components and their interactions are modelled by rules. A model based algorithm then uses these qualitative models to create generate-and-test scenarios to arrive at a solution.

Power system network is a big domain, we have therefore also divided the complete problem solving architecture into four levels. The first level does the reasoning at `component' level (circuit breakers, protective relays etc.), the second level at `switching groups' (connections between protection relays, isolators and circuit breakers), the third level at `clusters' (a cluster includes two adjacent circuit breakers and the connecting component between them), and the final level at the whole power system. The decisions based on reasoning done at lower levels are passed onto upper levels thereby relieving the upper levels from depth-based reasoning.

To handle missing and inaccurate information, we chose a back-propagation neural network as it is popular and known to be reliable and accurate for this kind of problem. Since neural nets cannot adapt to changing power system topology and are prone to longer training times and greater errors with increase in input and training data, we have assigned one neural net for every `cluster'. The `cluster level' is the level where the topology does not change and the decisions made at this level are prone to errors due to missing or inaccurate information. In other words, neural nets occupy level two of the hierarchy.

To learn heuristic knowledge from model based reasoning, we have developed a machine learning (ML) algorithm which essentially records a solution, generalises it, deletes (or forgets) the unnecessary details and stores the solution for future use.

The complete algorithm thus generated works like a cooperative system. All modules, the four hierarchies (including the neural network module) and the machine learning module interact with each other while solving a problem.