Thesis SISTEMA EN LÍNEA DE DIAGNÓSTICO DE ESTADO DE OPERACIÓN Y DETECCIÓN DE FALLAS EN UNA PLANTA PILOTO DE FLOTACIÓN ROUGHER
Loading...
Date
2017
Authors
BRAVO RAIN, GIAN LEONARDO
Journal Title
Journal ISSN
Volume Title
Publisher
Abstract
El presente trabajo tiene como objetivo analizar la capacidad de un modelo PCA para detectar en línea operaciones anormales y fallas en una planta piloto de un circuito de flotación rougher de operación híbrida ubicado en el Laboratorio de Control de Procesos de la Universidad Técnica Federico Santa María. Para esto se desarrolló un sistema modular , en el software Matlab 2016, que integra un detector de estado estacionario, un simulador metalúrgico, un sistema de almacenamiento de datos y, un sistema de diagnóstico del estado de la operación y detección de fallas. Además, se utiliza una interfaz gráfica en el software Intouch HMI 8.0 de Wonderware, el cual permite monitorear y controlar la operación de la planta piloto y, además, todas las funcionalidades del sistema creado en Matlab 2016. La planta piloto consiste en 3 celdas de flotación en serie, siendo operadas con una mezcla de agua y espumante, utilizándose el simulador metalúrgico para simular la presencia de los minerales cobre, fierro y molibdeno. Este simulador permite estimar resultados metalúrgicos tales como la recuperación y la ley de concentrado y cola del circuito rougher.Para la construcción de los 2 modelos PCA implementados, se utilizó el software PLS_Toolbox 8.5, el cual posee una interfaz gráfica a la cual se le ingresa los datos experimentales y se obtiene el modelo PCA, en donde se debe definir el método de preprocesamiento de los datos experimentales y además, se debe definir el número de componentes principales.El primer modelo PCA se construyó considerando 21 variables, las cuales corresponden a variables operacionales y metalúrgicas, considerándose tanto sensores como actuadores. El modelo se construye a partir de un conjunto de operaciones en estado estacionario, conformado por 339.993 operaciones diferentes, desde donde se seleccionaron aquellas que correspondían a operaciones normales, obteniéndose 83.079. Una operación normal es considerada como aquella que posee una recuperación de cobre del circuito dentro del rango entre 90 [%] y 91 [%]. Se aplicó el método de autoescalamiento a los datos experimentales para así normalizarlos. Finalmente, se escogieron 9 componentes principales, los cuales logran explicar el 90,12 [%] de la variabilidad del conjunto de datos de construcción y, además, todos estos componentes poseen valores propios superiores o cercanos a 1. Además, se define que los límites de los estadígrafos T2 y Q Residual corresponden a 21,67 y 7,00, respectivamente.El modelo PCA obtenido, es utilizado por el sistema de diagnóstico de la operación y detección de fallas, desarrollado en Matlab 2016, permitiendo detectar en línea, el estado de la operación.Para analizar el comportamiento del sistema para la detección de condiciones anormales de operación utilizando el primer modelo PCA, se desarrollaron 5 experiencias, en donde se tenían diferentes condiciones para los niveles de espuma de cada celda, ley de cobre de alimentación, porcentaje de sólidos de alimentación, flujo de alimentación y recuperación del circuito. Cada una de estas experiencias se iniciaba con una condición anormal de operación y, luego de alcanzar el estado estacionario, el sistema detectó correctamente todas las operaciones anormales iniciales. Luego, el sistema recomendaba realizar cambios en los niveles de espuma de cada celda y porcentaje de sólidos de alimentación (ya que son las que normalmente se pueden variar en una operación real) en donde, luego de realizar el cambio sugerido en la planta y, de alcanzar el nuevo estado estacionario, la operación alcanzaba un estado de operación normal en todas las experiencias realizadas. Cabe destacar que en las 5 experiencias realizadas, la operación normal se alcanzó luego de cambiar solo una vez las condiciones sugeridas por el sistema.Para analizar el comportamiento del sistema para la detección de fallas utilizando el primer modelo PCA, se realizaron 6 experiencias, en donde cada una de ellas correspondía a una condición diferente de operación. Cada experiencia se iniciaba con una operación normal (detectada correctamente por el sistema) y luego, se le simulaba una falla en algún sensor o actuador, analizando el comportamiento del sistema. Luego de simulada la falla, se eliminaba ésta y se retornaba a una operación normal nuevamente. Se simularon fallas en los niveles de espuma de cada celda, flujo de bomba de alimentación, porcentaje de sólidos de alimentación y ley de cobre de alimentación. En las 6 experiencias realizadas, el sistema detectó correctamente la falla. El mínimo error necesario para que el sistema detecte la falla varía entre 24 [%] a 65 [%], en donde el promedio corresponde a 37,8 [%]. El máximo valor de error está asociado al nivel de espuma de la celda 1 y, el mínimo, está asociado al flujo de bomba de alimentación.Luego, se construye el segundo modelo PCA, en el cual se excluyen variables que no poseen variabilidad en el conjunto de datos construcción, por lo que inicialmente se consideran 15 variables. Luego, aplicando PCA, se obtienen 7 componentes principales. Los límites de los estadígrafos T2 y Q residual corresponden a 18,48 y 6,383 respectivamente. Se realiza una prueba para diagnóstico de operación detectándose correctamente la operación anormal y normal, entregándose recomendaciones que permitieron retornar a una operación normal. Para el caso de detección de fallas, el sistema detecta correctamente la falla generada en el sensor. Además, el valor del mínimo error detectable no posee variación con respecto al primer modelo PCA.Finalmente, se realizan pruebas agregando ruido a las variables metalúrgicas de entrada al simulador, obteniéndose un diagnóstico correcto del estado de operación anormal y normal, además de entregarse recomendaciones correctas que permitieron retornar a operación normal. Además, para el caso de detección de fallas, se obtuvo una correcta identificación del sensor por parte del sistema. Además el valor del mínimo error detectable no se modificó.Finalmente, se concluye que el sistema implementado (tanto con el primer y segundo modelo PCA) permite un diagnóstico correcto del estado de la operación y, además, permite la detección de fallas generadas en la planta piloto de circuito de flotación rougher.
The objective of this paper is to analyze the ability of a PCA model to detect abnormal operations and failures online in a pilot plant of a hybrid operation rougher flotation circuit located in the Process Control Laboratory of the Universidad Técnica Federico Santa María. For this, a modular system was developed in the Matlab 2016 software, which integrates a steady-state detector, a metallurgical simulator, a data storage system and a diagnostic system for the state of the operation and fault detection. In addition, a graphic interface is used in Wonderware Intouch HMI 8.0 software, which allows monitoring and controlling the operation of the pilot plant and, in addition, all the functionalities of the system created in Matlab 2016. The pilot plant consists of 3 cells of floatation in series, being operated with a mixture of water and foaming, using the metallurgical simulator to simulate the presence of copper, iron and molybdenum minerals. This simulator allows to estimate metallurgical results such as the recovery and the law of concentrate and tail of the rougher circuit.For the construction of the 2 PCA models implemented, the software PLS_Toolbox 8.5 was used, which has a graphical interface to which the experimental data is entered and the PCA model is obtained, where the preprocessing method of the experimental data and in addition, the number of main components must be defined.The first PCA model was constructed considering 21 variables, which correspond to operational and metallurgical variables, considering both sensors and actuators. The model is constructed from a set of operations in steady state, consisting of 339,993 different operations, from which those that corresponded to normal operations were selected, obtaining 83,079. A normal operation is considered to be one that has a circuit copper recovery within the range between 90 [%] and 91 [%]. The self-scaling method was applied to the experimental data to normalize them. Finally, 9 main components were chosen, which explain 90.12 [%] of the variability of the construction data set and, in addition, all these components have eigenvalues greater than or close to 1. In addition, it is defined that The limits of the T2 and Q Residual statisticians correspond to 21.67 and 7.00, respectively.The obtained PCA model is used by the operation diagnosis and fault detection system, developed in Matlab 2016, allowing to detect online the status of the operation.To analyze the behavior of the system for the detection of abnormal operating conditions using the first PCA model, 5 experiences were developed, where different conditions were had for the foam levels of each cell, copper grade of feeding, percentage of solids of power, feed flow and circuit recovery. Each of these experiences started with an abnormal condition of operation and, after reaching the steady state, the system correctly detected all the initial abnormal operations. Then, the system recommended making changes in the foam levels of each cell and percentage of feed solids (since they are the ones that can normally be varied in a real operation) where, after making the suggested change in the plant and, After reaching the new stationary state, the operation reached a normal state of operation in all the experiences carried out. It should be noted that in the 5 experiences carried out, the normal operation was reached after changing only once the conditions suggested by the system.To analyze the behavior of the system for the detection of faults using the first PCA model, 6 experiences were carried out, where each of them corresponded to a different operation condition. Each experience started with a normal operation (correctly detected by the system) and then, it simulated a failure in some sensor or actuator, analyzing the behavior of the system. After the failure was simulated, it was eliminated and returned to a normal operation again. Faults were simulated in the foam levels of each cell, feed pump flow, percentage of feed solids and copper grade of feed. In the 6 experiences carried out, the system correctly detected the fault. The minimum error necessary for the system to detect the fault varies between 24 [%] to 65 [%], where the average corresponds to 37.8 [%]. The maximum error value is associated with the foam level of cell 1 and, the minimum, is associated with the supply pump flow.Then, the second PCA model is constructed, in which variables that do not have variability in the construction data set are excluded, so initially 15 variables are considered. Then, applying PCA, 7 main components are obtained. The limits of the residual statistics T2 and Q correspond to 18.48 and 6.383 respectively. An operation diagnostic test is performed, detecting the abnormal and normal operation correctly, delivering recommendations that allowed to return to normal operation. For the case of fault detection, the system correctly detects the fault generated in the sensor. In addition, the value of the minimum detectable error has no variation with respect to the first PCA model.Finally, tests are carried out adding noise to the metallurgical variables entering the simulator, obtaining a correct diagnosis of the abnormal and normal operation state, in addition to delivering correct recommendations that allowed to return to normal operation. In addition, for the case of fault detection, a correct identification of the sensor by the system was obtained. In addition, the value of the minimum detectable error was not modified.Finally, it is concluded that the implemented system (both with the first and second PCA model) allows a correct diagnosis of the state of the operation and, in addition, allows the detection of faults generated in the rougher flotation circuit pilot plant.
The objective of this paper is to analyze the ability of a PCA model to detect abnormal operations and failures online in a pilot plant of a hybrid operation rougher flotation circuit located in the Process Control Laboratory of the Universidad Técnica Federico Santa María. For this, a modular system was developed in the Matlab 2016 software, which integrates a steady-state detector, a metallurgical simulator, a data storage system and a diagnostic system for the state of the operation and fault detection. In addition, a graphic interface is used in Wonderware Intouch HMI 8.0 software, which allows monitoring and controlling the operation of the pilot plant and, in addition, all the functionalities of the system created in Matlab 2016. The pilot plant consists of 3 cells of floatation in series, being operated with a mixture of water and foaming, using the metallurgical simulator to simulate the presence of copper, iron and molybdenum minerals. This simulator allows to estimate metallurgical results such as the recovery and the law of concentrate and tail of the rougher circuit.For the construction of the 2 PCA models implemented, the software PLS_Toolbox 8.5 was used, which has a graphical interface to which the experimental data is entered and the PCA model is obtained, where the preprocessing method of the experimental data and in addition, the number of main components must be defined.The first PCA model was constructed considering 21 variables, which correspond to operational and metallurgical variables, considering both sensors and actuators. The model is constructed from a set of operations in steady state, consisting of 339,993 different operations, from which those that corresponded to normal operations were selected, obtaining 83,079. A normal operation is considered to be one that has a circuit copper recovery within the range between 90 [%] and 91 [%]. The self-scaling method was applied to the experimental data to normalize them. Finally, 9 main components were chosen, which explain 90.12 [%] of the variability of the construction data set and, in addition, all these components have eigenvalues greater than or close to 1. In addition, it is defined that The limits of the T2 and Q Residual statisticians correspond to 21.67 and 7.00, respectively.The obtained PCA model is used by the operation diagnosis and fault detection system, developed in Matlab 2016, allowing to detect online the status of the operation.To analyze the behavior of the system for the detection of abnormal operating conditions using the first PCA model, 5 experiences were developed, where different conditions were had for the foam levels of each cell, copper grade of feeding, percentage of solids of power, feed flow and circuit recovery. Each of these experiences started with an abnormal condition of operation and, after reaching the steady state, the system correctly detected all the initial abnormal operations. Then, the system recommended making changes in the foam levels of each cell and percentage of feed solids (since they are the ones that can normally be varied in a real operation) where, after making the suggested change in the plant and, After reaching the new stationary state, the operation reached a normal state of operation in all the experiences carried out. It should be noted that in the 5 experiences carried out, the normal operation was reached after changing only once the conditions suggested by the system.To analyze the behavior of the system for the detection of faults using the first PCA model, 6 experiences were carried out, where each of them corresponded to a different operation condition. Each experience started with a normal operation (correctly detected by the system) and then, it simulated a failure in some sensor or actuator, analyzing the behavior of the system. After the failure was simulated, it was eliminated and returned to a normal operation again. Faults were simulated in the foam levels of each cell, feed pump flow, percentage of feed solids and copper grade of feed. In the 6 experiences carried out, the system correctly detected the fault. The minimum error necessary for the system to detect the fault varies between 24 [%] to 65 [%], where the average corresponds to 37.8 [%]. The maximum error value is associated with the foam level of cell 1 and, the minimum, is associated with the supply pump flow.Then, the second PCA model is constructed, in which variables that do not have variability in the construction data set are excluded, so initially 15 variables are considered. Then, applying PCA, 7 main components are obtained. The limits of the residual statistics T2 and Q correspond to 18.48 and 6.383 respectively. An operation diagnostic test is performed, detecting the abnormal and normal operation correctly, delivering recommendations that allowed to return to normal operation. For the case of fault detection, the system correctly detects the fault generated in the sensor. In addition, the value of the minimum detectable error has no variation with respect to the first PCA model.Finally, tests are carried out adding noise to the metallurgical variables entering the simulator, obtaining a correct diagnosis of the abnormal and normal operation state, in addition to delivering correct recommendations that allowed to return to normal operation. In addition, for the case of fault detection, a correct identification of the sensor by the system was obtained. In addition, the value of the minimum detectable error was not modified.Finally, it is concluded that the implemented system (both with the first and second PCA model) allows a correct diagnosis of the state of the operation and, in addition, allows the detection of faults generated in the rougher flotation circuit pilot plant.
Description
Catalogado desde la version PDF de la tesis.
Keywords
DETECCION DE FALLAS , METODOS DE PROYECCION , MODELO PCA