Q-learning Based Maintenance Scheduling For Multi-yield Deteriorating Machines

The advancement and degree of automation of production machines are becoming anadvantage of these manufactories. Meanwhile, the reliability of machines maintained theprofit of industries since cost of maintenance makes up the most of total production cost.During production, machines deteriorate due to the aging, wearing or some other reasons;therefore, proper maintenance policies are paramount in improving production efficiencyand reducing production cost.In actual production, people usually preventively maintain the machines according tocertain production time or fixed number of products, or repair after machine breakdown.They seldom make maintenance plans on the base of the whole system which haseconomic and functional relations among the machines. A production system usuallycontains two or more machines both of which works as different roles. Especially incellular manufacturing system, raw materials become products after several steps ondifferent machines. The deterioration due to aging or corrosion of these machines alsoleads to quality reduction of products. Machines in a system do not work independently, ifnot to be considered, large costs will be wasted and breakdown time will increase. In orderto obtain proper maintenance policies for production system, this study covers the internalrelationships and the deteriorating states of machines. A single machine system was firstlyconsidered and then expanded to a two-machine-one-buffer flow line system, in which thequality of products is of multi-yield form and Q-learning method is adopted since it avoidsthe dimension disaster of Markov decision models. To verify the effectiveness ofQ-learning, a simulation comparison between the preventive maintenance in this study anda maintenance policy with fixed time interval has been made. The application of thetheoretical results can effectively solve maintenance problems in reality.