Ignacio Díaz

A new approach to exploratory analysis of system dynamics using SOM. Applications to industrial processes

The self-organizing map (SOM) constitutes a powerful method for exploratory analysis of process data that is based on the so-called dimension reduction approach. The SOM algorithm defines a smooth non-linear mapping from a high-dimensional input space onto a low-dimensional output space (typically 2D) that preserves the most significant information about the input data distribution. This mapping can be used to obtain 2D representations (component planes, u-matrix, etc.) of the process variables that reveal the main static relationships, allowing to exploit available data and process-related knowledge in an efficient way for supervision and optimization purposes. In this work we present a complementary methodology to represent also the process dynamics in the SOM visualization, using maps in which every point represents a local dynamical behavior of the process and that, in addition, are consistent with the component planes of the process variables. The proposed methodology allows in this way to find relationships between the process variables and the process dynamics, opening important ways for the exploratory analysis of the dynamic behavior in non-linear and non-stationary processes. Experimental results from real data of two different industrial processes are also described, showing the possibilities of the proposed approach.

Visual dynamic model based on self-organizing maps for supervision and fault detection in industrial processes

Visual data mining techniques have experienced a growing interest for processing and interpretation of the large amounts of multidimensional data available in current industrial processes. One of the approaches to visualize data is based on self-organizing maps (SOM), which define a projection of the input space onto a 2D or 3D space that can be used to obtain visual representations. Although these techniques have been usually applied to visualize static relations among the process variables, they have proven to be very useful to display dynamic features of the processes. In this work, an approach based on the SOM to model the dynamics of multivariable processes is presented. The proposed method identifies the process conditions (clusters) and the probabilities of transition among them, using the trajectory followed by the input data on the 2D visualization space. Furthermore, a new method of residual computation for fault detection and identification that uses the dynamic information provided by the model of transitions is proposed. The proposed method for modeling and fault identification has been applied to supervise a real industrial plant and the results are included.

Residual generation and visualization for understanding novel process conditions

We study the generation and visualization of residuals for detecting and identifying unseen faults using auto-associative models learned from process data. Least squares and kernel regression models are compared on the basis of their ability to describe the support of the data. Theoretical results show that kernel regression models are more appropriate in this sense. Moreover, experiments on vibration and current data from an asynchronous motor confirm the theory and yield more meaningful results.

Internet-based remote supervision of industrial processes using self-organizing maps

In upcoming years, the strategies for maintenance, traceability, management and operation of productive processes will demand the use of novel information and communication technologies. Supervisory systems in these new scenarios will have to be able to integrate large volumes of information and knowledge coming both from local and remote points of large processes. These systems will therefore require new tools for management and integration of information and knowledge. In this work, the authors present an internet-based remote supervision system of industrial processes that incorporates powerful data and knowledge visualization tools based on selforganizing maps (SOM). This architecture adds an intermediate layer (database) to the well-known client and server layers, that isolates the client part from the industrial process, allowing to incorporate the required data management and neural network processing tasks. Remote users have access to advanced information visualization tools based on SOM, including both static visualizations, such as component planes or distance maps, and dynamical ones, such as residuals and state trajectory, allowing the interpretation of knowledge extracted by the SOM as well as the analysis and detection of possible abnormal conditions. This architecture has been validated through the supervision of an industrial pilot plant. r 2006 Elsevier Ltd. All rights reserved.

Visual data mining and monitoring in steel processes

Steel processes are often of a complex nature and difficult to model. All information that we have at hand usually consists of more or less precise models of different parts of the process, some rules obtained on the basis of experience, and typically a great amount of high-dimensional data coming from numerous sensors and variables of process computers which convey a lot of information about the process state. We suggest in this paper the use of a continuous version of the self-organizing map (SOM) to project a high dimensional vector of process data on a 2D visualization space in which different process conditions are represented by different regions. Later, all sorts of information resulting from the fusion of knowledge obtained from data, mathematical models and fuzzy rules can be described in a graphical way in this visualization space.

Interactive feature space extension for multidimensional data projection

Projecting multi-dimensional data to a lower-dimensional visual display is a commonly used approach for identifying and analyzing patterns in data. Many dimensionality reduction techniques exist for generating visual embeddings, but it is often hard to avoid cluttered projections when the data is large in size and noisy. For many application users who are not machine learning experts, it is difficult to control the process in order to improve the “readability” of the projection and at the same time to understand their quality. In this paper, we propose a simple interactive feature transformation approach that allows the analyst to de-clutter the visualization by gradually transforming the original feature space based on existing class knowledge. By changing a single parameter, the user can easily decide the desired trade-off between structural preservation and the visual quality during the transforming process. The proposed approach integrates semi-interactive feature transformation techniques as well as a variety of quality measures to help analysts generate uncluttered projections and understand their quality.

A visual approach for fuzzy rule induction

Models are descriptions of real facts that serve us to think and reason. In building models, a compromise always exists between accuracy, that is, how precisely the model describes reality, and simplicity, without which the model would be useless. So, a good model must be simple and intuitive while being accurate enough. In this paper we propose a novel approach based on visual techniques aiming to help the human in fine-tuning fuzzy decision trees to enhance its interpretability and insightfulness with a minimal loss of accuracy. By involving the human in the design process, these techniques allow to include prior knowledge in the selection of membership functions as well as to assess the significance of rules in the model to help in the pruning stage.

Analysis of electricity consumption profiles in public buildings with dimensionality reduction techniques

The analysis of the daily electricity consumption profile of a building and its correlation with environmental factors makes it possible to examine and estimate its electricity demand. As an alternative to the traditional correlation analysis, a new approach is proposed to provide a detailed and visual analysis of the correlations between consumption and environmental variables. Since consumption profiles can be characterized by many components, the input space is high dimensional. For that reason, it is necessary to apply dimensionality reduction techniques that enable a projection of these data onto an easily interpretable 2D space. In this paper, several dimensionality reduction techniques are tested in order to determine the most appropriate one for the stated purpose. Later, the proposed approach uses the chosen algorithm to analyze the influence of the environmental variables on the electricity consumption in public buildings located at the University of Leon. Finally, electricity profiles from all buildings are compared with regard to two aspects, the magnitude and dynamics of the consumption.

Stability comparison of dimensionality reduction techniques attending to data and parameters variations

The analysis of the big volumes of data requires efficient and robust dimension reduction techniques to represent data into lower-dimensional spaces, which ease human understanding. This paper presents a study of the stability, robustness and performance of some of these dimension reduction algorithms with respect to algorithm and data parameters, which usually have a major influence in the resulting embeddings. This analysis includes the performance of a large panel of techniques on both artificial and real datasets, focusing on the geometrical variations experimented when changing different parameters. The results are presented by identifying the visual weaknesses of each technique, providing some suitable data-processing tasks to enhance the stability.

Visual analysis of a cold rolling process using a dimensionality reduction approach

The rolling process is a strategical industrial and economical activity that has a large impact among world-wide commercial markets. Typical operating conditions during the rolling process involve extreme mechanical situations, including large values of forces and tensions. In some cases, these scenarios can lead to several kinds of faults, which might result in large economic losses. Thereby, a proper assessment of the process condition is a key aspect, not only as a fault detection mechanism, but also as an economic saving system. In the rolling process, a remarkable kind of fault is the so-called chatter, a sudden powerful vibration that affects the quality of the rolled material. In this paper, we propose a visual approach for the analysis of the rolling process. According to physical principles, we characterize the exit thickness and the rolling forces by means of a large dimensional feature vector, that contains the energies at specific frequency bands. Afterwards, we use a dimensionality reduction technique, called t-SNE, to project all feature vectors on a visual 2D map that describes the vibrational states of the process. The proposed methodology provides a way for an exploratory analysis of the dynamic behaviors in the rolling process and allows to find relationships between these behaviors and the chatter fault. Experimental results from real data of a cold rolling mill are described, showing the application of the proposed approach.