Parts, that is, the predictive component replacement algorithm, is really something that everyone is particularly concerned about in industrial maintenance. Can you know which part is not good in advance? It is best to replace it early to avoid the machine suddenly crashing and delaying major tasks. It is not like before, it can only be repaired when it is broken. When it is repaired, it is in a hurry to find the accessories, or it can be replaced according to the deadly time, and it will be replaced even if it is broken. In the end, it can be used to analyze the data and figure out when the parts may be closed!
The first thing to talk about is data collection : don’t look at how mysterious it is to predict it, first of all, you have to eat something for it, that is, you have to collect a large amount of data to feed it to make it smart. For example, you may have to collect the speed of the parts when the machine runs on its own how high the temperature is? Is there any vibration that is very good when collisions with other parts? There are many sources of data; it is possible that even if the machine reacts after pressing a button, it will be recorded as information and wait for judgment; sometimes, whether there is anyone on the scene who is idle and knocks on the machine may be… It doesn't care much, it mainly depends on the valid data!
Then there is the selection and debugging of algorithm models : the mainstream is just a few regression analysis in the field of machine learning, which can roughly calculate how many days the part can last; there is also the decision tree, looking at the stupid tree, it will judge the rules one by one, such as first looking at whether the temperature of a part exceeds 80 degrees. If it exceeds the temperature, then look at whether the vibration amplitude exceeds the range, and guess step by step; right! Recently, deep learning models have become more used. After all, when there are a lot of parts, a simple model cannot turn around and calculate such a complex relationship.
Whether the model is accurate or not is reliable, it has to be taught carefully with historical data to let it know, "The temperature of this part last time soared to 90 degrees and the vibration lasted for three days, but it broke on the fourth day." Such lessons can be learned and be smart enough to predict the right trend.
Provide global procurement services for weak current intelligent products! When using it in practice, if the prediction is not accurate and the parts are broken, then it will be a waste of money if they are replaced in advance; if they say they will not be broken, they will stop production in a blink of an eye, and the losses will be even greater. Therefore, the model has to be continuously optimized, and the latest data is constantly used to readjust the parameters that are very troublesome to calculate; sometimes it has to be combined with the reliable experience of the master touching the machine vibration, listening to the sounds and abnormal sounds, and these people's reliable experiences.
What exactly do you look at the key indicators ? Anyway, you have to look at two numbers. One is the "accuracy" of this algorithm, which means whether there are 99 times in a hundred predictions. The other important number is the "false alarm rate", which is the most annoying. The part is obviously strong and can be used for several months. As a result, the algorithm is shocked and said that it is too fast to replace it immediately. If there are too many false alarms, everyone will think that this thing is unreliable and gradually they will not like to pay attention to it.
Q: Does this algorithm only estimate the approximate time period of the parts may have problems?
If the machine is stable, it can basically be determined in a few weeks, or in a few days or hours. If the working conditions are messy and the force that parts may also change… Then you may only give a vague reminder "Recently, please pay attention to the status of the parts!", but it is much better than waiting for it to be damaged completely; it is much better than silly and blindly, and it is much better than silly eyes and ignoring nothing.
Q: Will this be very expensive or troublesome when using an ordinary small factory?
Actually, it doesn’t need it… There are packed software on the market now. Even if you enter some data on your mobile app or install a small control on your mobile app, it won’t cost much money; and as mentioned earlier, after entering data enough and teaching it for a long time, the algorithm will become smarter and more reliable as an experienced old repairman! When it can really help the factory avoid huge losses from shutting down one or two stops, then the little money invested at the beginning will be… and you will have already recovered your money without a shadow!
Leave a Reply