Introduction:
Modeling the manufacturing operations has always been a challenging task due to several factors leading visibility in real time and lack of critical data sometimes. Cement manufacturing process has been holding some of the complex operations including fossil fuels, alternative fuels, mix of feed and extensively high temperature to melt the feed for effective reactions to take place. This industry has been at the forefront of technology adaption since long but still has not been able to reach its desired state of operations due to the complex operations. Several manufacturing plant operations are running with advanced process controllers, inferentials and data reconciliation models in place. But with the severity of complexity with cement plants pose to the technology, existing solutions are not themselves sufficient to support the complete requirements. Thanks to the technology, which is continuously evolving thriving to meet the needs of industry supporting them to sustain and improvement further. In this article we are going to focus on how advanced AI techniques can enhance your current practices including 1st principles or non-linear inferential models. One such AI technique known by reinforcement learning, or deep learning in simple words has shown tremendous potential to enhance the current modeling strategies.
Reinforcement learning:
With bunch of technologies and so much of advancement it has now been possible to address the undressed. There are some subtle advantages which is important to make to a note of, while can be quite significant while addressing the model improvement scenarios. Lets start having a look at some of the benefits with AI ca bring over the traditional modeling techniques, and above it what reinforcement learning can cover up. Please note, the scope of this article is limited to cement manufacturing processes only.
- Complex nonlinear process dynamics: One of the most important aspects of cement manufacturing is to capture the dynamics of the underlying process precisely. By far the best way has always been to utilize first principles to understand the complex dynamics among multiple variables in the processes. First principles doesn’t account for all the factors which are measured or unmeasured. Reinforcement learning has shown incremental success over the data driven technique by penalizing the dynamics which were wrongly captured. This technique can be leveraged to simulate several scenarios and understand the process much better by analyzing the scenarios which never happened in the history. This allows for significant benefit over the traditional approaches where the mathematical models rely on the standard data points only, and not covering the complete space of decision variables which are critical enough to make decisions.
- Process disturbances: Several traditional techniques suffer from the issue of frequent model management, fine tuning and re development efforts due to continuously changing processes and unforeseen factors known to us as noises. Noises may include seasonality, feed quality variations, AFR quality and many others which makes it difficult to account for using the traditional approaches, due its internal structure. AI on the other hand tries to capture the complex patterns in the timeseries data which could also uncover the characteristics of the other factors which were not captured. With this AI tries to delineate and capture the influences of noises on the process targets independently. Reinforcement learning being an advanced version tries to further enhance the inefficacies of AI models by rewarding and penalizing the actions. The actions would typically include the scenarios of the formulated optimization function where for every deviation in the target function, variable deviations, the model penalizes the objective function to correct its course of actions. Whereas on the other hand for every correct actions the model rewards the objective function which registered as the action to be utilized for future advantages.
- What if Scenarios: With reinforcement learning it is possible to simulate the scenarios by considering the possible changes in the input parameters which had never been observed in the historical dataset. With this potential the model allows the operator to identify the opportunities for enhancement of the processes. Beyond and above, this capabilities also enables the operators to simulate the scenarios where the detrimental effects of the process disturbances are compensated by manipulating the levers which can be modified within the existing limits of operations to ensure the optimal operations are in place.
The above-mentioned benefits are some of the key aspects where the advanced modeling techniques can be leveraged to enhance the current models. To understand the details of reinforcement learning please reach out to us and learn how it can potentially help your organization to improve various business levers.
