![]() ![]() If the training data from phase one differs greatly from the data of the new request in phase two, the AI results may also be unreliable.If the training data for model development in phase one is poor, the model prediction in phase two will also be poor.The reason for this lies in the training data for the model: no comparably large harness was present during training.ĭuring model development and use, we became aware of a few aspects concerning AI: Only for one harness with more than 400 minutes of true labor time, the prediction was off by more than 20%. In a separate test run with data from a new request with eight harnesses (average labor time 237 minutes), the mean percentage error was about 10%.For small harnesses, on the other hand, the model still needs to be improved, since the percentage error increases for small harnesses. For large harnesses, this is already quite good. For the validation dataset, the AI model predicted labor times that had a mean absolute error of 4.1 minutes (the mean true labor time was about 58 minutes).In this validation phase, the model made predictions solely based on the learned training data and without human intervention. We then fed this model with data from about 50 other past wiring harness projects and then tested its accuracy. In our case, we used the data from about 200 harnesses from the past during the training phase to select the best model from a variety of models. How well does the predictive model actually work? All that is needed are the influencing factors. If the model is accurate enough, it can give a recommendation for answering a new request. This simply means that different mathematical models (ML algorithms) are tried on the training data to find the best model that minimizes the error between model prediction and actual labor time. During machine learning, the relationships (model parameters) between the features (influencing factors) and the target (labor time) are then "learned". After selecting and collecting these features from old projects, we now needed to add the correct labor times from the normal (full) calculation processes to complete the training data. As influencing factors, we defined data about the harness that was relevant to predicting labor time and was available quickly – such as the number and length of wires, the number of connectors, etc. To solve such a supervised machine learning regression problem, we needed two data inputs for the model creation: the labor time (target) and its influencing factors (features, cf. In our pilot project, we focused only on one cost factor that was time-consuming to get, namely the labor time to produce a harness. If we do not submit an offer, we are already eliminated at the beginning of the offer process. Today, the customer's deadlines for submitting an initial indicative offer are much shorter. In the past, a particular customer would send us a request with a drawing and a wire list of the desired harness and we would calculate the costs within some weeks. ![]() Without AI, we'd be out - Responding faster to the customer's request This is no wonder, as there are different subfields within AI with different areas of application. Moreover, people do not always have the same conceptions about AI. For instance, sales forecasts are based on numbers, quality control relies on images, and certain automation processes handle text. For example, it works with many different file formats as input data, which means that it often looks very different in the application. Only when you understand AI and its limits and also recognize your own application possibilities, then AI turns from an unreflected "hype" into a realistic "hope". And lo and behold, some realistic application possibilities emerged.īefore describing our first AI pilot application, it makes sense to take a look at AI itself. At a cross-departmental workshop, we explained the basics of AI to LEONI employees and discussed their ideas and potential with them. Once the RPA service was established, we launched a second AI investigation balloon.
0 Comments
Leave a Reply. |
AuthorWrite something about yourself. No need to be fancy, just an overview. ArchivesCategories |