Deep learning algorithms have many applications in Manufacturing analytics domain. The core of Smart factories, in my perspective, is embedded in the right and optimal use of Machine learning algorithms.
In this article, I will explore three areas within manufacturing where we can leverage Deep learning to our advantage. Those three areas are:
- Job shop scheduling
- Quality control
- New Product Development
Now let us start exploring these in detail.
One way in which a Neural Network can be used in job shop scheduling is to first train the network by providing multiple different combinations of inputs (setup times, deadlines) and outputs (processing order for each machine). These outputs may be derived in numerous ways, for example by using a Job Shop Manufacturing (JSM) heuristics.
The Neural Network can be taught to learn the underlying non linear function that maps the output to the input, which is not susceptible to any analytical formula. Subsequently, for any new order of jobs and related completion deadlines, the actual output can be obtained from the Neural Network itself.
Note that the optimal size of training data depends on the optimization heuristics (like JSM) used to generate the data. I have read many papers that suggest that JSM heuristics are not as efficient as a direct Branch and Bound solution to the optimal sequence. You will have to do some experimentation to determine a good input generation method.
Deep learning can allow us to create automated or semiautomated testing procedures and quality control. A research actually suggests that a pattern recognition software for quality control has, on an average, has less than 1% error rate, as compared to 14% average error rate for humans.
The processs of creating a Deep Learning algorithm in this case starts with capturing as many measurements about the part as possible, as it moves in the manufacturing process, from raw material to semi-finished good to finished product. You must, however, ensure that the measurement process is consistent and repeatable.
Example of this can be that is RGB cameras are being used, the orientation of the camera with respect to the object and the illumination should all be fixed.
During the training of the Neural Network, it is important to accurately label the output as being acceptable or defective (in case of simple Binary classification). This labeled data must also be asigned to the measurement data from multiple steps of the manufacturing process. All this data is used to train the Neural Network whose output could be the measured characteristics of the work-in-progress after any particular processing step, and the output is the desired classification label.
Once the training is complete, these Neural Networks can be used for real-time quality control. As an example, if right after a particular manufacturing step, the measured characteristics is indicative of a defective part, that particular part cabn be removed from the process, without investing any additional resources on a aprt that will eventually be deemed defective.
New Product Development
If you are familiar with New Product design best practices, you know that eliminating non-value add complexity is a significant profit creator, as precious expert machinist time is saved for value-add machining instead of setup waste (just from a manufacturing perspective).
If properly designed and trained, this task of effective new product creation can also be solved using a Neural Network. This Neural Network will have as inputs a set of existing part numbers and the order in which the various tools are used to produce the parts.
By feeding the Neural Network plenty of such examples of processing tools and job orders for existing tools, it can learn from the function that maps them to the output. That way, when a new part is encountered, the Neural Network will check which part in the existing set is closest to the new product and then try to start with that order to produce the new part.
For example, if part A is an old part and part B is new, similar to A but new, then the set that was used to build the portion of A that is similar to B, can also be used to build that portion of B. Continuing the logic, for other parts of B, the Neural Network can select which of the oter parts are similar and choose sets accordingly.
Views and Perspectives my Own.