Version française
BlueKaizen Tool Condition
BlueKaizen designed BK Tool Condition, a generic algorithmic approach to meet the challenging needs involved in the monitoring, control and predictive maintenance of manufacturing equipment.
BlueKaizen Tool Condition uses the BlueKaizen Machine Learning Toolkit to provide statistical learning solutions to problems belonging to the traditional manufacturing industry, such as the characterization of the normal state of a water pump, and to problems faced by more recent industries, such as the telecommunication and microelectronics industries.
In the last decades, the fast evolution of manufacturing and non-intrusive measurement technologies, together with the exponential growth in the volumes of data to acquire, transmit and store, affected the problems of equipment monitoring in several ways:
- The solution to many traditional problems such as, for example, the control of the temperature of a blast furnace, calls upon an accurate physical model of the relevant parts of the process. In some industries, like the microelectronics industry, such a model is often very difficult (or expensive) to determine completely. In such a situation, a good solution seems to consist in building a statistical model from the measured parameters.
- Being able to measure a large number of equipment parameters has several advantages, but also a major drawback: the set of correlations determined by the global behavior of the equipment parameters can become extremely difficult to analyze and interpret, because statistical interactions between the various measurements appear as well at the inter-parameter level as at the level of temporal intra-parameter auto-correlation. This explains why data-mining approaches can be valuable in helping engineers focus their analyses on the most interesting parts of the data.
- The large volume of data to be analyzed, and the large number of parameters to be modeled, makes most traditional approaches obsolete whereby mono-parameter models are built and their outputs are analyzed by hand or using a simple rule-based-system (e.g. traditional control charts being the most significant example of this situation). Although the traditional approach is effective when one has less than ten parameters to supervise, it becomes quite difficult to apply when a hundred parameters are involved. In the latter case, multivariate algorithms will be more effective, and the corresponding models easier to generate and maintain.
Applications
BlueKaizen Tool Condition can be applied to various types of problems:
- Statistical process control (SPC)
The principal advantage compared to standard SPC lays in the ability to simultaneously model several parameters (up to several hundreds) and to define rejection and acceptance regions, which take into account the correlations between these parameters (multivariate SPC). Multivariate SPC calls in particular upon density estimation and quantile estimation algorithms.
For more information about our software solution for the microelectronics and semiconductor industries, we invite you to have a look at BlueKaizen WaferProfile.
- Fault detection and classification (FDC)
The modeling of a tool's normal state makes it possible to detect abnormal behaviors, of which there are generally various types. BK Tool Condition provides algorithms to classify the detected faults, either in an unsupervised way, or in a supervised way (i.e. using a reference base of known defects). Multivariate FDC calls upon density estimation algorithms for fault detection, clustering and commonality search algorithms for unsupervised fault characterization or supervised classification algorithms.
For more information about our software solution for the microelectronics and semiconductor industries, we invite you to have a look at BlueKaizen Tool Pattern.
- Predictive maintenance
Regression and density estimation algorithms make it possible to generate a model of equipment ageing and use it for predictive maintenance, whenever a complete physical model of the equipment is not available. The implementation of the models calls upon time series analysis and non-linear, multivariate regression algorithms. The generation of alarms calls upon multivariate density estimation algorithms.