Data Science


Data Science


Controller Performance assessment- a data analysis perspective   by Mohd Faheem Ullah


Controller performance assessment is the field of study where we assess and monitor the health of a control loop so that we can come up with strategies that will help us to improve the performance of control loop. Basically, the loop which consists of the sensor, the actuator, the information of the process under control and the feedback mechanism is referred to as control loop. This control loop is an integral part of large-scale industries. Poorly performing control loops maylead to oscillation in the process variable. The consequences of poor performance of these control loops include degradation of product quality, higher material and energy consumption, low profitability and poor performance of the process. The control loops are tuned to make sure that the system tracks the desired set-point correctly, quickly and remains stable. During the setting up of the control loop, a particular tuning is employed to get an optimal operation. The need for assessment is on the high considering, the current scenario in industries, where typically only about a third of the control loops perform satisfactorily. We looked into three different aspect of controller performance assessment- detection of oscillation, root cause analysis and detection and diagnosing of stiction.

Hardware setup for CPA

Detection of oscillation:

Oscillation is a phenomenon observed very commonly in the wide range of the systems. The challenges like non-stationarity, intermittency, measurement noise and presence of multiple oscillations at the time etc. need to be addressed. We approach the problem by developing an algorithm that robustly identifies the dominant frequencies in the data and localizing the region of oscillation for identified dominant frequencies.

Root cause analysis

In industries, we deal with many interacting loops. Once, the disturbance or oscillation enters through any of the control loops, it will propagate to all other interacting loops. Detecting the root cause in such scenarios is very cumbersome especially when the number of loops involved is large. We develop a sophisticated algorithm to detect the root node generating the oscillation from simulated data.

Interacting control loops

Detection and diagnosis of stiction

Stiction is basically static friction in control valve which results in sticking and slipping of control valve movement. Stiction exists in control valve due to ageing, excessive tightening of the control valve. My current work involves detection and diagnosis of stiction from data using one or (and) two parameter models.

 

References:

  • Srinivasan, Ranganathan, et al. “Control loop performance assessment. 2. Hammerstein model approach for stiction diagnosis.” Industrial & engineering chemistry research 44.17 (2005): 6719-6728.
  • Srinivasan, B., and Raghunathan Rengaswamy. “Automatic oscillation detection and characterization in closed-loop systems.” Control Engineering Practice 20.8 (2012): 733-746.
  • Babji, S., U. Nallasivam, and R. Rengaswamy. “Root cause analysis of linear closed-loop oscillatory chemical process systems.” Industrial & engineering chemistry research 51.42 (2012): 13712-13731.

Controller design for fast implementation by Manikandan S


Model predictive controllers (MPC) find applications across wide range of engineering domains that require continuous monitoring and control. The main advantage of MPC is the use of system model to predict future trajectory and compute the input values based on the chosen objective. Since MPC is an optimization framework it can handle constraints that are present in the system with ease when compared to a PID based controller, where we use thresholding when the constraints are violated. This formulation also helps us keep track of the overall objective of the system by means of the objective function value.

The system model is the heart of any MPC formulation. Identifying such model in industrial applications requires careful planning and execution of input moves while collecting the data. This step is the most time consuming step in MPC implementation. MPC implementation time can be accelerated if system model can be assembled faster. This can be done using

  1. Easy to construct directional model
  2. Historical data to assemble the model
  3. Reinforcement learning based formulation in place of a model

Each of these methods poses various requirements on the controller formulation. We try and address these constraints and formulate a stable and optimal controller.