Gregorio Sainz

Interpretable knowledge extraction from emergency call data based on fuzzy unsupervised decision tree

Nowadays, call centers are common in different areas of activity providing customer services, medical attention, security services, etc. Each type of call center has its own particularities but all call centers have to plan the availability of resources at each time period to support the incoming calls. The emergency call centers are a special case with extra restrictions. In this context, this work is devoted to providing support for the decision making about resource planning of an emergency call center in order to reach its mandatory quality of service. This is carried out by the extraction of interpretable knowledge from the activity data collected by an emergency call center. A linguistic prediction, categorization and description of the days based on the call center activity and information permits the workload for each category of day to be known. This has been generated by a fuzzy version of an unsupervised decision tree (FUDT), merging decision trees and clustering. This involves quality indexes to reach an adequate trade-off between the tree complexity and the category quality in order to guarantee interpretability and performance. This unsupervised approach deals correctly with the real management of this type of centers generating and preserving expert knowledge.

Comparison and design of interpretable linguistic vs. scatter FRBSs: Gm3m generalization and new rule meaning index for global assessment and local pseudo-linguistic representation

This work is devoted to defining more general interpretability indexes to be applied to any scatter or linguistic model implemented by any type of membership functions. They are based on metrics that should take into account the semantic and inference issues: the semantic issue in order to preserve the meaning of the linguistic labels and the inference issue since this can influence the behavior of the rules. On the other hand, these metrics have been designed to be intuitive in order to support the analysis or selection of a final model and to favor a low computational cost within an optimization process. In order to check their usefulness, a multi-objective evolutionary algorithm, simultaneously performing a rule selection and an adjustment of the fuzzy partitions, is guided by the proposed indexes on several benchmark data sets to obtain models with different degrees of accuracy and interpretability. In addition, using these metrics, a local analysis can be carried out between models of a different nature. This local analysis through the model components, gives support to the user to make the best choice from among the models.

Checking orthogonal transformations and genetic algorithms for selection of fuzzy rules based on interpretability-accuracy concepts

Fuzzy modeling is one of the most known and used techniques in different areas to emulate the behavior of systems and processes. In most cases, as in data-driven fuzzy modeling, these fuzzy models reach a high performance from the point of view of accuracy, but from other points of view, such as complexity or interpretability, the models can present a poor performance. Several approaches are found in the specialized literature to reduce the complexity and improve the interpretability of the fuzzy models. Here, a post-processing approach is taken into account via the definition of the rules selection criterion that aims to choose the most relevant rules according to the well-known accuracy-interpretability trade-off. This criterion is based on Orthogonal Transformations, here the QRP transformation is taking into consideration, and its parameters are tuned genetically. The main objective is to check the true significance, drawbacks and advantages the firing matrix of the rules, that is the foundation of the most usual approaches based on orthogonal transformations for the complexity reduction of the fuzzy models. A neuro-fuzzy system, FasArt (Fuzzy Adaptive System ART based), and several case studies, data sets from the KEEL Project Repository, are used to tune and check this approach. This neuro-fuzzy system generates Mamdani fuzzy rule based systems (FRBSs), each with its own particularities and complexities from the point of view of fuzzy sets and rule generation. NSGA-II is the MOEA tool used to tune the criterion parameters based on accuracy-interpretability ideas.

Interpretability-accuracy improvement in a neuro-fuzzy ART based model of a DC motor

Abstract The aim of this paper is to propose a general methodology applicable to any rule based fuzzy model generated by any precise or linguistic fuzzy algorithm to improve the linguistic-accuracy trade-off. Here, the neuro-fuzzy system FasArt (Fuzzy Adaptive System ART based) is used for its proven model capabilities, as shown in previous papers and works. If does, however, have the usual drawbacks, from the linguistic point of view, of most fuzzy modeling methods found in the scientific literature. A fuzzy model of a DC motor is generated by FasArt, whose performance is a good estimation of the motors behavior, then this performance is improved by a better interpretability of the knowledge attained and stored by this fuzzy model. The main idea behind this approach is to find a fuzzy model with enough accuracy and an adequate capacity of explanation or interpretability of its data acquired knowledge. The modeling process can thus be seen as knowledge extraction in human or linguistic terms: from a numeric level (data) to a symbolic one (linguistic fuzzy rules).

Design of residuals in a model-based Fault Detection and Isolation system using Statistical Process Control techniques

In the work presented in this paper Statistical Process Control (SPC) techniques are applied to a model-based Fault Detection and Isolation (FDI) approach. The residuals, produced as outputs from the FDI system, are manipulated with typical SPC charts to improve the overall diagnosis process. The charts explained in this work: Shewhart control chart, Cumulative Sum (CUSUM) control chart and Exponentially Weighted Moving Average (EWMA) charts are able to accurately determine significant deviations in the residuals. The integration of model-based tools with SPC supervision can be a step towards robustness and effectiveness in fault detection. This scheme reduces the number of false alarms, which is an important aspect in FDI tasks, and can reduce the fault isolation time. This approach has been applied to a laboratory plant with real data, obtaining interesting results.

Recurrent Neuro-Fuzzy Modeling of a Wastewater Treatment Plant

This paper deals with the use of a special kind of recurrent neuro-fuzzy model to represent complex systems. The neuro-fuzzy system, called RFasArt, has been used in this work to model a complex bio technological process: an activated sludge process taken from a real wastewater treatment plant. This network is based on the adaptive resonance theory (ART) but it also introduces formalisms from the fuzzy set theory and takes into account the available contextual information in its processing stage. Real data records taken from the plant were used to train this network. The results obtained with this recurrent fuzzy neural network have been compared with the ones obtained with a classical recurrent neural network, showing the advantageous behaviour of the RFasArt system. Apart from modelling, the RFasArt architecture provides a knowledge base of fuzzy rules containing information about the plant dynamic behaviour.

A study on the use of machine learning methods for incidence prediction in high-speed train tracks

In this paper a study of the application of methods based on Computational Intelligence (CI) procedures to a forecasting problem in railway maintenance is presented. Railway maintenance is an important and long-standing problem that is critical for safe, comfortable and economic transportation. With the advent of high-speed lines, the problem has even more importance nowadays. We have developed a study, applying forecasting procedures from Statistics and CI, to examine the feasibility of predicting one-month-ahead faults on two high-speed lines in Spain. The data are faults recorded by a measurement train which traverses the lines monthly. The results indicate that CI methods are competitive in this forecasting task against the Statistical regression methods, with e-support vector regression outperforming the other employed methods. So, application of CI methods is feasible in this forecasting task and it is useful in the planning process of track maintenance.

Knowledge extraction based on fuzzy unsupervised decision tree: application to an emergency call center

This paper describes the application of a fuzzy version of Unsupervised Decision Tree (UDT) to the problem of an emergency call center. The goal is to obtain a decision support system that helps in the resource planning, reaching a trade-off between efficiency and quality of service. To reach this objective, the different types of days have been characterized based on variables that permits available resources assignment in an easy and understandable way. In order to deal with availability of expert knowledge on the problem, an unsupervised methodology had to be used, so fuzzy UDT is a solution merging decision trees and clustering, providing the performance of both viewpoints. Quality indexes give criteria for the selection of a reasonable solution to the complexity, as well as interpretability of the trees and the quality of generated clusters, and also the type of days and the performance from the resources point of view.

Monitoring and fault detection in processes with multiple operating modes, transitory phases and start-ups using principal component analysis

This paper presents a global monitoring and fault detection approach considering the different operation points, start-ups and transitory states that can appear during plant operation.

Neuro-fuzzy control of a pH plant

Abstract This paper studies the control of a pH process by using a neuro fuzzy controller with gain scheduling. As the process to be controlled is highly non-linear the PI-type fuzzy controller that will be used generally is not able to control the system adequately. For this, a very simple feedforward neural network trained on-line, is put at the output of the PI-type fuzzy controller in order to calculate the gain of the controller. This neuro-fuzzy regulator has been tested in real-time on a bench plant. On-line results show that the designed control system allows the plant to operate in a range of pH values, despite perturbations and variations of the plant parameters, obtaining good performance at the desired workings points.