Diagnostics of failures of technological equipment of chemical industries using artificial intelligence

Zubov Dmitrii Vladimirovich ORCID: 0000-0002-0703-1577 PhD in Technical Science Associate Professor; Department of Information Computer Technologies; D.I. Mendeleev Russian University of Chemical Technology 123580, Russia, Moscow, Geroyev Panfilovtsev str., 20, room 127 dvzubov@gmail.com

Lebedev Danila Aleksandrovich ORCID: 0009-0007-2873-2341 Postgraduate student; Department of Information Computer Technologies; D.I. Mendeleev Russian University of Chemical Technology 119331, Russia, Moscow region, Moscow, Maria Ulyanova str., 16, sq. 188 lebedev.d.a@muctr.ru

Modern information technologies greatly facilitate the management of complex chemical and oil refining industries: SCADA systems allow operators to see current and archived information from sensors at their workplace, remotely control actuators and receive prompts during the technological process. Modern automated control systems have made it possible to reduce the number of employees engaged in routine operations, but it is not possible to completely abandon operators who carry out dispatching control of the technological process. Significant difficulties for management are the processes of starting and stopping technological lines, switching to new raw materials, changing technological modes, etc. The processes of starting new production lines and starting after repair are particularly difficult, but difficulties may also arise in continuous installations operating in stationary mode, for example, as a result of accidents.

Emergency situations are manifested by changes in sensor readings and the most important task of the operator is the timely correct assessment of the situation and taking the necessary actions. A significant help to operators can be a plan for emergency response, which outlines the signs of an accident, possible causes of its occurrence and recommended actions. Nevertheless, timely detection of accidents is a difficult task for operators, especially in the early stages, when the warning and emergency alarms of technological quantities have not yet been triggered.

Computer simulators are a useful tool for training operators, in particular, computer simulators based on the RTsim platform ^[1], combining a dynamic model of a specific production, a simulator of a SCADA system, a training complex from a set of exercises of varying complexity and a system for assessing the level of operator training. The set of instructions includes tasks from the simplest (maintaining a stationary continuous technological process on a separate device) to very complex (independent start-up and shutdown of an oil refining production site, recognition and elimination of typical emergency situations).

The advantage of the training complex is its compliance with real production, including according to the indications of indicators on mnemonic circuits, trends, etc., which the automated control system receives and which the operator sees ^[2], i.e. there is a screen of the "real world" – a model of the technological process on which you can see the condition and control local equipment (in particular – manual valves) and, if necessary, you can enable the display mode for all model state variables, including those that are not controlled by sensors and therefore are not visible on the mnemonic circuits of the SCADA system. A fragment of the screen of the process model (TP model) is shown in Fig. 1.

Fig. 1. A fragment of the screen of the TP model of the RTsim training complex

The screen of the distributed control system (DCS) shows the interface of the SCADA system, and in some cases the trainee can choose the system used, since they differ in the type of interface elements, etc. Figure 2 shows the same section of the processing line as in Figure 1, but as it is visible in the SCADA system based on the Yokogawa software package.

Fig. 2. A fragment of the RTsim simulator DCS screen

An example of displaying an accident of the type "Complete destruction of column C-1" is shown in Fig. 3. The operator must identify the type of incident based on indicator readings, trends and warning messages and take action to stop the technological process.

Fig. 3. A fragment of the RCS screen of the RTsim training complex when simulating an accident of the type "Complete destruction of column C-1"

Preparing operators to work in emergency situations reduces their reaction time to an already familiar situation, but the set of emergencies that are considered in the training process is limited, because there may be several types of possible accidents with each device of the technological chain, as a result, operators will be ready for already familiar emergencies that they saw during training and they will be less prepared for situations that they have not encountered, which will increase the likelihood of performing suboptimal actions. A logical step is to create an automated emergency recognition system.

In ^[3], it was proposed to build a failure tree of a technological line and, taking into account the probabilities of the occurrence of initial events and the expected harm from the consequences of accidents, choose the actions of the operator that will minimize the most likely damage. Unfortunately, the problem remains obtaining the probabilities of the initial events (leakage of the device, blockage of the pipeline, valve failure, etc.) and therefore this approach is applicable rather at the design stage.

The paper ^[4] provides an overview of publications on the use of artificial intelligence for the diagnosis of accidents in industry. Most solutions are focused on use in transport and mechanical engineering, but there are also publications on the use of artificial intelligence in the chemical industry: in ^[5] it was proposed to use a Bayesian network to detect the causes of accidents in chemical industries, in ^[6] it was proposed to use a special case of a recurrent neural network – LSTM network - for early diagnosis of accidents in the forced oxidation system during wet flue gas desulfurization. In ^[7], it was proposed to use a neural network to diagnose safety valve failures. In all the cases considered, data from real objects was used to train the network, which limited the amount of dataset used.

We propose to use an automated accident recognition system based on a neural network trained using data obtained from a specific production model during the operator's work.

In ^[8], an approach was proposed to simulate typical accidents of technological devices using chemical process modeling systems such as Aspen One, UniSim and others. The use of modeling systems or digital doubles ^[9] will greatly facilitate the collection of model data, but the proposed approach may well be used on simulators.

The RTsim training complex simulates process lines up to for manual and safety valves, which allows you to simulate a wide range of emergency situations. Let's look at some examples:

– simulation of stable depressurization of the device – opening of the shut–off valve for drainage (in the case considered in Fig. 1-3 - the UV-007 valve), which is completely closed in normal mode and fully opens at the moment of the beginning of the simulation of an accident;

– simulation of the developing depressurization of the device – the pre-cut-off valve for drainage (UV-007) is open, but the manual valve (Z-024) is closed, which begins to open smoothly at the moment of the simulation of an accident;

– the failure of the centrifugal pump is simulated by the abrupt closure of the manual valve on the discharge line, which is fully open in normal mode.

Using the proposed approach allows you to simulate a large number of accidents and get trends in technological variables, lock triggers, etc., which can be used to train a neural network.

To train a neural network, it is necessary to form a dataset, the structure of which may be different, but it is obvious that the target features in it will be the code of the type of accident, and the predictors will be the values of technological variables. For certainty, we will assume that each type of accident is encoded with the value 1 in the corresponding column, i.e. there are Na columns of target features, Na is the number of encoded types of accidents, and 0 means that an accident of this type did not occur. Let's assume that the second column of target features encodes the type of accident "pump failure H-1A", and the third column – the type of accident "pump failure H-1B", then the appearance in the cell of the second column 1 will mean the failure of the pump H-1A, and 0 – means its serviceable condition (including when it is turned off).

The number of predictor columns will depend on the number of information signals in the automated control system, they can be divided into groups:

– technological variables with an analog or digital signal (data from sensors of temperature, pressure, level, composition, etc.). Variables of this group are normalized in the range from 0 to 255, i.e. 0 – will correspond to the lower limit of the measuring range of the transducer, 255 – the upper limit. At the same time, there is some coarsening of the signal, which performs the function of a zero-order digital filter, which somewhat saves from accidental fluctuations of technological magnitude. If for some reason this is undesirable, at the stage of system development, you can set the range of normalization in accordance with the opinion of the developer.

– process variables with a discrete signal (signal from the lower and upper level alarms, pressure switches, etc.) – encoded as 0 (the corresponding alarm did not work) and 1 (the alarm worked);

– digital control variables (degree of opening of the control valve, heater power, etc.) – normalizes in the range from 0 (valve is completely closed/heater is completely turned off) up to 255 (valve is fully open/heater power is maximum);

– discrete control variables (shut–off valves, switching on the centrifugal pump, etc.) - encoded as 0 (valve closed, pump off) and 1 (valve open, pump on).

In most cases, a distributed control system within a single technological section is controlled by a single industrial controller, which operates in a cyclic manner: at the beginning of the cycle, the values of technological variables are entered from the input blocks into the corresponding memory area and they are considered unchanged until the end of this controller cycle. Based on them, the controller calculates the values of the control actions, enters them into the output data area and they are simultaneously output to the output blocks, after which a new cycle of operation of the controller begins. A departure from this scheme is possible, but dramatically complicates the computational complexity of the control system and is relatively rare. Similarly, process modeling systems usually operate with the concept of a clock cycle, in which the data calculated at the previous iteration is considered unchanged and updated at the same time at the end of the clock cycle. A set of data on the state of the system – that is, all values of technological quantities, valve positions, etc. are called a snapshot of the system.

One snapshot of the system could be used to assess its condition if the system did not have "memory" and the control channels did not have a delay time. In fact, to control inertial objects (capacitive devices, heat exchangers, pipelines), it is necessary to mean not only their current state, but also a certain number of previous ones. For certainty, let's assume that the duration of one sensor polling period in the automated control system and one cycle in the modeling system is 1 second, and the value of the largest of the delay times in the considered fragment of the processing line is equal to ?, then it can be assumed that 1.5 – 2 ? images are sufficient for correct recognition of all system states. Probably, the exact minimum amount of data sufficient for training a neural network will depend on the topology of the flows inside the considered fragment of the processing line.

To illustrate the proposed method, consider an extremely simplified example: there is a centrifugal pump pumping coolant, the pump condition is described by the variable S, S=0 means that the pump is in good condition, S=1 means that the pump is out of order. The pump control is described by the variable H, H=1 means that the pump is energized, it must be turned on, H=0 means that the pump is not energized, it must be turned off. The coolant flow rate on the pump discharge line is denoted by the variable F, F=0 indicates a lack of flow, F= 255 is the maximum possible flow rate. Let's assume that the duration of the simulation cycle in the system is 1 second, and the delay time on the control channel is 2 seconds. Then the dataset fragment will have the form shown in Table 1.

Table 1. Example of a dataset fragment for training a neural network

In the dataset fragment shown in Table 1, we see the following scenario:

line i is the initial state, the command to turn on the sediment has not been given, the pump is not working, the liquid flow is absent, it is assumed that the pump is in good condition;

line i+1 – a command has been given to turn on the sediment, but while it has not yet managed to turn on, there is no liquid flow, it is assumed that the pump is in good condition;

line i+2 – a command has been given to turn on the sediment, the pump is already working, but the flow sensor does not notice it yet, it is assumed that the pump is in good condition;

line i+3 – a command was given to turn on the sediment, the pump is running, the flow sensor measured the flow rate at 10/255 of the maximum, it is assumed that the pump is in good condition;

line i+4 – a command has been given to turn on the sediment, the pump is running, the flow sensor measured the flow rate at 50/255 of the maximum, it is assumed that the pump is in good condition;

line i+5 – a command was given to turn on the sediment, the pump is running, the flow sensor measured the flow rate at 100/255 of the maximum, it is assumed that the pump is in good condition;

line i+6 – a command was given to turn on the sediment, the pump is running, the flow sensor measured the flow rate at 10/255 of the maximum – less than it was, this is a sign of an accident.

A variant of the neural network structure for emergency recognition is shown in Fig. 4.

Fig. 4. The structure of the neural network for emergency recognition

It is advisable to choose the number of neurons in the input layer equal to the number of input variables in one snapshot of the system multiplied by the number of snapshots, i.e. in our example with pump diagnostics, we get 2 variables in a snapshot for 3 snapshots = 6 neurons in the input layer. In the output layer, we select the number of neurons equal to the number of recognized types of accidents (m), increased by 1 (state S0 of good operation), in our example, the number of types of accidents was 1, which means that 2 neurons need to be provided in the output layer. Since there cannot be a state of good operation and an accident at the same time, then for the output layer we will choose an activation function of the Softmax type (a multi-variable logistic function). To select the number of hidden layers and activation functions in them, it is necessary to consider the dataset in more detail, but for the considered example with a pump, two hidden layers with the ReLU activation function are quite enough. 

The RTsim simulator includes a model and a number of typical scenarios for the butane separation shop, in particular, it is possible to start, stop the column, monitor the stationary mode, switch from mode to mode. Among the typical accidents, the simulator kit includes scenarios of complete destruction of the C-1 column, failure of the H-1A pump and failure of the power supply system. All complete scenarios (start, stop, regular operation, three types of accidents) were used to form a training dataset, additionally data was generated based on the results of simulation of the H-1B pump accident.

The resulting dataset was used to train a neural network with two hidden layers and it was possible to find a set of neuron weights that provides recognition of these four types of accidents. With a further set of simulation data, it is planned to train the network to recognize more types of equipment failures.

Conclusion. A method for recognizing failures of devices of chemical and technological systems using a neural network trained on a model of a specific production is proposed. The structure of a neural network that recognizes single accidents is selected. In the future, the created system can be improved to recognize not only single, but multiple and dependent failures.