Recall that the output 0. If we multiply the error in output —0. Recall that the sum of the weighted inputs of the output neuron 1. To change this sum S o by —0. So, the weights between the hidden neurons and the output neuron become:. After adjusting the weights between the hidden layer neurons and the output neuron Figure 13 , we repeat the process and similarly adjust the weights between the input and hidden layer neurons. This is done by first calculating the gradient at the input coming into each neuron in the hidden layer.
For example, the gradient at X h 3 is: 0. A backpropagation algorithm is used to adjust the weightings between the hidden layer neurons and the output neurons, so that the output is closer to the target value 0.
The proposed change in the sum of weighted inputs of X h 3 i. Note that we use the original value of W 9 0. This is because although we are working one step at a time, we are trying to search the entire space of possible weight combinations and change them in the right direction toward the bottom of the hill.
- Software Configuration Management Strategies and IBM (R) Rational (R) ClearCase (R): A Practical Introduction!
- Harry Potter and the Prisoner of Azkaban (Piano Solos)?
- Artificial Intelligence in Petroleum Industry?
- Education and Social Inequality in the Global Culture!
In each iteration, we propagate the output error through original weights, leading to new weights for the iteration. Once you know the proposed change in the weighted sum of inputs of each neuron S 1 , S 2 , S 3 , you can change the weights leading to the neuron W 1 through W 6 proportional to the output from the previous neuron. Thus, W 6 changes from 0. Upon repeating this process for all weights, the new output in this example becomes 0. By performing just one such iteration of forward and back propagation, the network is already learning!
A small neural network like the one in this example will typically learn to produce correct outputs after a few hundred such iterations of weight adjustments. Even with all the amazing progress in AI, such as self-driving cars, the technology is still very narrow in its accomplishments and far from autonomous. And, people need to supply and fine-tune the appropriate algorithms. All of this relies on manual labor. Thus far, we have looked at neural networks that learn from data.
This approach is called supervised learning. An error is generated if there is a difference between the actual output and the target output and the weights are adjusted based on this error until the actual output matches the desired output. Supervised learning relies on manual human labor for collecting, preparing, and labeling a large amount of training data. In Part 2 of this series, we will delve into two other approaches that are more autonomous: unsupervised learning and reinforcement learning.
Unsupervised learning does not depend on target outputs for learning. Instead, inputs of a similar type are combined to form clusters. When a new input pattern is applied, the neural network gives an output indicating the class to which the input pattern belongs. Reinforcement learning involves learning by trial and error, solely from rewards or punishments. Such neural networks construct and learn their own knowledge directly from raw inputs, such as vision, without any hand-engineered features or domain heuristics.
AlphaGo Zero, the successor to AlphaGo, is based on reinforcement learning. Unlike AlphaGo, which was initially trained on thousands of human games to learn how to play Go, AlphaGo Zero learned to play simply by playing games against itself. Although it began with completely random play, it eventually surpassed human level of play and defeated the previous version of AlphaGo by games to 0.
Last but not least, we will discuss social and ethical aspects, as the recent explosion of progress in AI has created fear that it will evolve from being a benefit to human society to taking control. Shell has been collecting real-time data across its operations for decades. More than 10 million operational variables per minute are presently collected, streamed, archived, and integrated with operational control systems.
There is enormous potential to exploit these data further. Predictive analytics and machine learning algorithms could make it possible to avoid unexpected failures and unnecessary maintenance, which would save millions of dollars per year in optimized maintenance and deferment avoidance. In the studies, a team organized by Shell used unlabeled historical process control data to develop a digital twin algorithm that predicts valve failures. Experiments for the use case were performed on multiple control valves.
The aim was to verify whether machine-learning methods are capable of distinguishing between normal and abnormal valve behavior. The difference between the predicted and measured system output i. Teammate and study coauthor Sander Suursalu developed multiple solutions based on artificial neural networks and statistical approaches to model the normal behavior of the monitored systems at the Pernis site.
Mismatches with predictions of the modeled systems were then used to predict failures. The artificial neural networks were able to predict failure up to a month in advance in some cases. The team found that four-layer gated recurrent units GRUs with tanh activation functions and an input sequence length of four samples produced the best results. Furthermore, this approach enabled highly accurate failure prediction. These systems could output deviations five times larger than the deviations present during normal operation.
They tested this model using a traditional mass balance, with meters upstream and downstream of a target flow element. The model verified that the mass balance approach provided acceptable meter accuracies and it allowed engineers to track predicted flowmeter performance vs. Peter Kwaspen and Bruce Lam were subject matter experts for this work.
Bringing in the right expertise related to the problem in question is crucial for success of a machine-learning project. Going forward, these models should enable equipment deterioration analysis and be the catalytic step-change toward predictive maintenance. Solutions like the ones presented have tremendous potential for replication in thousands of valves both upstream and downstream.
The machine-learning-based solution should have the potential to generate exponential value for the business.
- Structural impact.
- Information About!
- Danish Directors: Dialogues on a Contemporary National Cinema.
He has worked for AIChE for more than 9 years, incrementally assuming responsibilities for web, IT, database, and mobile solutions. Previously, he held positions at Altria and Philip Morris International. Would you like to reuse content from CEP Magazine? Simply click here to connect instantly to licensing services, where you can choose from a list of options regarding how you would like to reuse the desired content and complete the transaction.
Skip to main content. Log in About Join Contact. Introduction to Deep Learning: Part 1. Amit Gupta. What is deep learning? Predicting Valve Failure with Machine Learning Shell has been collecting real-time data across its operations for decades. Son, H. Hardesty, L. Coley, C. Diagnostic Methods 6. Expert System Development 7.
Discussion and Summary Acknowledgments References 4. Characteristics of Artificial Neural Networks 3. ZNL Architecture 4. Fault Detection and Diagnosis Examples 5. Conclusion References 5. Shallow Versus Deep Knowledge 3. The Model-Based Approach 4. Multiple Faults Diagnosis 5. Other Approaches 6. Divide and Conquer MFD2 7. Conclusions References 6. Overview 3. The Scenario 4. The Plant Model 6.
chapter and author info
The Diagnostic Expert System 7. Conclusions and Future Work Acknowledgments References 7. Control System Design Process 3. Expert Process Modeling 5.
CL 688 Artificial Intelligence in Process Engineering
Expert Controller Design 6. Conclusion Acknowledgments References 8. Review 3. An Expert System Approach 5. Associated Conventional Control Strategies 6. Using AI-based methods, inherently qualitative variables in chemical processes like catalyst deactivation in reactors can also be considered in the model, while these types of variables are not possible to implement in mechanistic models.
The most common methods of AI for modeling purposes in chemical engineering are ANN and fuzzy logic, which sometimes are hybridized with evolutionary algorithms [ 4 , 5 , 6 , 7 ]. In addition to ANN and fuzzy logic methods, their hybrid scheme named adaptive-network-based fuzzy inference system ANFIS which is actually a fuzzy inference system implemented in the framework of adaptive networks has also been applied for modeling purposes in chemical engineering.
Afterward, according to the experimental data or the knowledge of the governing phenomena, the model is developed. The parameters that characterize the AI-based model like the number of fuzzy sets when using fuzzy logic , the number and the transfer functions of hidden layers when using the ANN method depend on the complexity and nonlinearity of the system and the types of variables affecting the process.
Among the types of ANN structures, the multi-layer perceptron MLP neural network which has a feed-forward scheme is believed as the most useful topology for system modeling [ 8 ]. Moreover, the recurrent ANN model which is a mapping of past inputs and outputs to the future outputs can be used for dynamic processes. Mamdani Fuzzy that differs in the way the information and rules are presented has several superiorities over the TS approach for the modeling of chemical processes. First, qualitative experience and knowledge of the experts who are dealing with the process are incorporated in the development of the model [ 12 ].
In addition, there is no need for data in order to build the Mamdani fuzzy model. Consequently, a Mamdani fuzzy model is more intuitive, transparent and interpretable [ 13 ]. In contrast, each TS-type model is a local approximator and the predictability of the model is valid for the specific operating condition of the process under which the model was developed and tested [ 14 ]. Accordingly, it can hardly be applied for analyzing the process behavior and cannot be scaled up or down and therefore is less useful for industrial practice.
Despite the capabilities of Mamdani method, it is worth underlying that a Mamdani fuzzy model suffers from the large number of rules when dealing with the processes with large number of variables. Genetic algorithm GA can be used to optimize the performance of a fuzzy model. The schematic of this algorithm is shown in Figure 2.
In the first step, the output variables determining the behavior of the system are defined, given that, the input variables which affect the selected output variables are determined. Afterward, a base fuzzy model is defined, characterized by the number and types of fuzzy sets of variables and the production rules presenting the behavior of the process based on the knowledge and expertise of the experts who have been working with the system. This model is used as the start-up version of the model which has to be tuned. In the second step, GA is formulated for optimization of parameters that characterize model, such as membership function parameters, membership function types and so on.
Chemical process optimization has its origins in linear programming at the beginning of the s [ 19 ]. This problem is finding the best solution from a variety of efficient alternatives of design or operating variables in order to minimize or maximize a desired objective function. In a general way, the objective function can be the minimization of the operating costs and the undesired material production or the maximization of energy efficiency, the yields and operation productivity, the profitability, safety and the reliability of the plant.
Most chemical processes are nonlinear and complex, so there are many solutions in some cases becoming endless in the optimization problems. Such problems are often too complex to be solved through gradient-based optimization approaches. Evolutionary algorithms EAs like GA [ 20 ], harmony search [ 21 ], particle swarm optimization [ 22 ] and so on categorized in the AI-based method that is a generic population-based metaheuristic optimization algorithm are capable of efficiently finding an optimal solution in complex problems, such as optimization of chemical processes.
The process control strategies have been developed to improve the performance of the process, reduce energy consumption and ensure high safety and environmental goals. The conventional controllers cannot show satisfactory responses in many industrial chemical processes with high nonlinear dynamics and parameter uncertainties, whereas AI approaches can be effectively controlled for a number of complex and nonlinear processes [ 23 ]. Because of their high potential for handling nonlinear relationships and self-learning capabilities, there has been considerable interest in the use of neural networks for the control in different fields of chemical processes such as thermal processes [ 24 ], reaction processes [ 25 ] and separation and purification [ 26 , 27 ].
One of the algorithms based on neural network control is the inverse model control.
Artificial intelligence in process engineering - PDF Free Download
In this approach, it is assumed that the input vector for neural network is the required future or reference output together with the past inputs and the past output variables; the approach can help to make for better performance the controlled variables when the unmeasured disturbance is present.
The manipulated variable of the controlled plant is the output of the neural network controller [ 23 ]. Therefore, the controller predicts the control action, as shown in Figure 3 , by having current and past values of the process model state variables and the past control action. Fuzzy systems have been used in different applications for controlling chemical processes [ 28 , 29 , 30 ].
Researchers also used fuzzy logic controller coupled with an optimal control in an exothermic chemical reaction [ 31 ], a batch polymerization reactor [ 32 ] and polymerization processes [ 33 ]. In addition, since time delay can be often seen in many industrial chemical processes, a possible alternative is the fuzzy model predictive control FMPC which has been proposed [ 34 , 35 ]. In systems with uncertainties of the system model, the choice of type-1 fuzzy may not always be the appropriate solution for a control problem [ 36 ].
In these cases, the type-2 fuzzy logic control has been represented in many fields of chemical processes [ 37 , 38 ]. Hybrid controller based on AI strategies combines two or more AI techniques in order to improve control performance of the chemical process. This approach is a hybrid intelligent system which uses the learning ability of the neural network with the knowledge representation of the fuzzy logic [ 39 ]. As shown in Figure 4 , the ANFIS architecture contains five layers of feed-forward neural network which are explained as follows:.
This layer is named as an input layer. Each neuron in this layer saves the parameters of the membership function and crisp inputs are converted to membership degree values which change between 0 and 1. Each neuron of this layer performs a connective operation i. The normalized firing strength is multiplied by a linear combination of the inputs i.
The last layer of the network is the weighted average of the outputs of the fourth layer. The application of ANFIS in the process control of chemical plants was seen in the distillation column [ 40 ], biodiesel reactor [ 41 ]. A fault is defined as a deviation from an acceptable range of an observable variable or calculated parameter that is referred to as a failure.
The abnormal conditions in a plant can result in financial losses. Therefore, in the chemical processes, fault detection and diagnosis have been the focal point of many researches and various fault detection and diagnosis strategies have been presented in the literature. The fault diagnostic systems should possess desirable characteristics such as quick detection, isolability, robustness and multiple fault identifiability [ 42 ]. One of the intelligent fault diagnosing techniques is neural network systems. Because of their high potential for capturing nonlinear relationships, neural networks represent a powerful tool for fault diagnosis [ 43 , 44 , 45 , 46 , 47 ].
In fault detection based on neural networks, the number of neurons in the input and output layers are equal to the number of measured variables and the number of potential faults in the process, respectively. The outputs of the neural diagnoser are binary variables representing the occurrence of a fault if the corresponding value is 1 or the lack of fault occurrence if the corresponding value is 0 [ 47 ]. Another fault diagnosis approach of AI techniques is fuzzy logic, which is applied in chemical processes [ 48 , 49 , 50 , 51 ]. In fault diagnosis based on fuzzy logic, the fuzzy relations between faults and symptoms are assumed to be from one to many i.