|
Description:
The aim of this project was to develop a fault diagnosis program for the Outokumpu copper flash smelting process using Kohonen Self-Organizing Maps in conjunction with heuristics rules. The purpose of the fault diagnosis system was to detect abnormal process states and inform the plant operator accordingly. A process deviation from the normal operating range is often caused by equipment malfunctions or an incorrect control strategy. Early detection of undesirable process states enables the execution of correcting actions, thus minimising the damage caused by process malfunctions.
The implemented fault diagnosis system consists of a process interface, data preprocessing, data controllers, and symptom generation objects and rules. The inferencing results are stored in a database, which can be used to perform a statistical analysis of the faults. The structure of the system is presented in the following picture.
The monitoring and classification are performed by using Self-Organizing Maps (SOM). The SOM is a neural network algorithm which is used to form a neural network model of an unknown system based only on the data received from the system. The nature of the phenomena monitored with SOMs differ, some of them being closely linked with process control and process disturbances, while the rest monitor the states of the process. One aim was to train a SOM for the quality of the feed material, which is the most important phenomenon that affects process control in copper flash smelting process.
The overall state of the process was monitored in order to detect undesired process states because a deviation from the desired state cannot always be considered as a direct process disturbance. From an economic point –of –view, it is just as important to detect a decrease in the quality of the product as to detect an equipment failure; both lead to a reduction in productivity and profitability. The monitored process states were viscosity of the slag and the matte, temperature of the waste heat boiler, and the sooting, i.e. cleaning, of the boiler or the gooseneck.
The following actual process disturbances were searched for: flooding of the feed material, aggregation of feed material in the concentrate burner, formation of dust aggregations inside the boiler or the gooseneck, and malfunctions of the fields of the electrostatic precipitator and gas blower. The most important part of a rule-based fault diagnosis system is the rules that are applied. With well functioning rules a very effective system can be built while, on the other hand, a program cannot predict phenomena correctly if its rules are false or feeble.
In the system the rules are always connected to a specific piece of equipment, and can be formed using only those variables that are known for that equipment. This prevents the formation of illogical rules, that may be the reason why systems show unexpected behaviour. Rules that are based on equipment also make it possible to know which rules are affected if certain changes are made in the process equipment. The fault diagnosis system has been giving promising results in both process monitoring and detection of the condition of process equipment failures.
|
