The importance of Machine / Deep Learning within process analysis, monitoring and lifecycle prediction

Recent evolutions in microchip development are broadening the possible applications of machine learning (ML) and even deep learning (DL). Estimating hard-to-measure machine states has been a problem for engineers for decades. Research shows that, using state-of-the-art data-driven techniques, machine monitoring can be realized in a more efficient way. Citing a case study from the TETRA project WearAI, the added value of classification algortimes is clearly demonstrated.

I. THE RISE OF DATA-DRIVEN INSIGHTS

The application of advanced signal processing methods in the monitoring of essential processes is considered a very interesting area with great potential. Material costs are reduced by monitoring anomalies and more production is possible by adding the calculated RUL to maintenance schedules. The result is an increase in efficiency.

In the past, process modeling was a complex task of long duration. Personnel with deep process knowledge often start with the "first principles" method. By combining their knowledge and experience with physical laws, it is possible to describe characteristics to arrive at an acceptable model. In system identification, this is also known as a white-box model. The internal structure is fully known.

The process industry is constantly changing and systems are becoming more and more complex. More actuators, sensors, set points, limit values, etc. make the white-box approach time-consuming. Therefore, grey-box and black-box techniques have rightly gained a lot of ground in recent years. Both use captured data to gain information about the process to make the model more accurate. Grey-box models still have a white-box structure, based on physical laws, but the parameters in the equation are estimated and adjusted iteratively to minimize predictive inaccuracies. Black-box models do not have this predetermined structure and are entirely determined by machine learning techniques, for example. Both the structure and parameters are based on the information learned from the dataset. This data-driven approach significantly speeds up the design phase.

II. CHALLENGES FROM PRACTICE

The objective of the case study is to design a data-driven algorithm that can inform the machine operator about the state of components in a machine or a system as a whole. As an example, we cite a case study developed in the TETRA project WearAI. It involved the milling of laminate at a partner with more than 60 years of experience and among the largest flooring manufacturers in the world. The hard plastic top layer of laminate is the main contributor to wear on the cutting tool. High feed rates and RPMs ensure smooth production processes, but this also accelerates tool degradation, necessitating regular downtimes to either rotate or replace the chisels. Planned maintenance minimizes downtime due to low-quality cutting edges, but does not utilize the full life of cutting tools. Maintenance based on current tool condition is the next logical step in this process. Placing an external system at original machine generated data streams such as vibrations, acoustic emissions, currents and voltages. Valuable information for further analysis and training of machine learning algorithms. Due to the absence of labels, it is appropriate to investigate unsupervised learning methods, as here more specifically the DBSCAN algorithm.

III. CLASSIFICATIONS

After extracting the desired features from the captured data streams, feature selection is an interesting (often underestimated) intermediate step. Several filters are necessary to standardize outliers, noise, the data domain, etc. From the many generated features, in this case there were 1360, only the most valuable features are allowed to go on to the next round to eventually serve as training data. The original dataset corresponds to a run-to-failure experiment. Features over the entire tool life are thus present. By subdividing the set into 40 different bins, each bin contains information about about 2.5% of tool life. Using these labels, it is possible to train a simple linear regressor. Using the accuracy of the fully trained model as a benchmark, we systematically omit features from the training set and compare the score. In this way, the differences in accuracy give an indication of the proportion of a particular feature. In graph 1 we see all features (shown on the x-axis) and their "impact factor" on correctly identifying a bin.

Ideally, we are left with significant features that represent the different bins with unique properties. To visually represent this multidimensional dataset, t-SNE could be employed. By converting the original space to probabilities, it reduces 59 dimensions to only 2 dimensions. Which is ideal for graphical representation. Figure 2 shows 5 bins by their differences in color. Note that two bins are partially mixed, which may complicate later classification.

With only the most important parts of the dataset remaining, it was time to decide which model was appropriate. Due to its unsupervised nature paired with relatively few tuning parameters, Density Based Spatial Clustering Application with Noise (DBSCAN) was a logical choice. Sources from the literature give good results in a similar implementation of this technique for tool wear. An important adjustment value is the average distance between "normal" data points (ε). Generating the point cloud and some iterative calculations give us the following data.

Outliers and noise have larger distances from points belonging to a cluster. Therefore, we choose ε smaller than this peak which in this case means a value approximately equal to 1. By adding this in the DBSCAN algorithm, we are able to recognize clusters without giving any information about the process ourselves.

The DBSCAN model is clearly able to distinguish different clusters. In the lower right corner of Figure 4, two bins are merged into one cluster. Because the information from these data points was measured during the middle of the tool life, we can conclude that there is less variance between these bins with respect to bins during the break-in or end of tool life phase. A conclusion parallel to previous experiments and correct according to the machine operators.

IV. POTENTIAL

The consequences of implementing ML/DL in existing machines cannot be underestimated. This worked example does not include information about ML deployment, but clearly demonstrates that regions of break-in and accelerated wear can be distinguished using unsupervised machine learning algorithms. Working further on this are systems in an industrial context that can prevent catastrophic break-in. Which again is a direct factor in minimizing downtime, which in turn increases production. Increased insight into machine condition enables the move to conditional maintenance which again lowers downtime and extends operation time per tool. More production hours per tool translates to less stock and less associated inventory costs.

V. DECISION

Data-driven techniques are not just the latest new technology hobbyhorse. It is a shift in vision and perspective. It fits perfectly into the Industry 4.0 philosophy, and any organization that wants to follow this path must build a careful structure. It is not a singular project, but a transition to a data-centered approach. It will bring any company with a healthy data management system one step closer to predictive maintenance and smart data exchange.

By demonstrating that even relatively simple models are capable of detecting a fracture, for example, the clout and potential of this approach is clear. We are convinced of its potential and look forward to extending it to other applications.

VI. RECOGNITION

We would like to thank Dr. Tim Claeys and Prof. Jeroen Boydens for support and dedication in this project. Project collaborators Hans Naert, Pieter Ideler, Peter Vanbiervliet and Robin Loicq were indispensable in data capture and communication to industrial partners.