Research on Fault Diagnosis System of IOT for Oil Well Pump Based on Machine Learning

The measurement and control of oil well pump data is the foundation of digital intelligent oilfield construction. Its goal is to detect the operating status and parameters of various components of oil pump equipment in real time, and transmit the collected operating data to the Internet of Things through intelligent instrument communication nodes and dedicated 5G networks. Control center machine [14]. On the one hand, the central computer conducts analysis and processing through the upper-level intelligent control software, makes decision-making control when necessary, and then outputs it to the equipment control mechanism through a dedicated 5G network; on the other hand, the central computer can convert and standardize the collected data through algorithms and store them to the database server, and at the same time serve mobile terminal or desktop terminal users through the public 5G network or Internet network [15]. The principle composition structure of the system equipment based on the Internet of Things architecture is shown in Figure 1. The system adopts a layered design method to realize real-time data collection and monitoring of oil pipeline pressure, flow and temperature, and has low cost and strong automatic monitoring capabilities. It mainly consists of three layers: perception layer, network layer and application layer.

Figure 1.

The sensing signals of this system for the measurement and control of oil delivery pumps in oilfield well groups mainly include: valve switches of various components, oil pump inlet and outlet pressure, oil pump inlet and outlet temperature, oil pump output flow, oil pump body temperature and other sensors. The sensing layer is composed of ZigBee wireless subnets of oil well pumps located in several areas, and the ZigBee wireless subnet is composed of sensor nodes with ZigBee communication modules and aggregation nodes (ZigBee well group control communicators, also known as gateway nodes). The convergence node in this system is the transfer station of each oil well’s external communication, and its core component is the ZigBee wireless transceiver module. The aggregation node is a relatively independent processing unit. If the external network fails to contact the IoT control center, it can automatically adjust the operating status of the local oil pump according to the node’s internal operation logic rule base. In addition, it can also store the local data for a period of time. The measurement and control data can be queried locally to display the real-time or historical operation data of each sensor.

The network layer mainly completes the reliable transmission of the information collected by the oil pump or the control signal of the Internet of Things center [16-17]. It is essentially based on the wireless local area network (WLAN) or the wireless mesh protocol to form a larger network of multiple aggregation nodes to monitor the distribution of multiple areas. Oil well pumps. Specifically, its function is to receive the information from the aggregation node of the perception layer, which can be encrypted and transmitted to the control application layer through appropriate algorithms as needed [18]. In addition, the information and control output of the control application layer are reversely transmitted to the aggregation node of the perception layer. The network layer mainly acts as a bridge. Since the oil wells of oilfield companies are mostly located in uninhabited mountain areas, these places often do not have public communication 5G networks, and a dedicated wireless communication network needs to be set up to serve this system.

The application layer is the highest layer in the functional structure diagram of the system, and it is the remote control center of the system. It realizes the automatic safety control of the oil pump on the basis of receiving and analyzing the sensory information of each oil well [19-21]. The control application layer is mainly composed of IoT central computer, database server, desktop terminal, mobile terminal and so on. The central computer of the Internet of Things generally has a wireless transceiver module, which can receive the equipment information transmitted by the dedicated 5G network base station of the transport layer in real time, analyze and process it according to the requirements of the oil pump operation process parameters through the software system, and store the data periodically in the database server [22-23]. Oil wells with different production volumes are equipped with different oil pump powers and different operating parameters. Various equipment parameters can be set through the man-machine interface of the control application layer. Spare. The data received by the control application layer from the network layer generally undergoes the steps of verification, unpacking, and anti-encryption conversion [24-25]. It can dynamically display data such as oil well temperature, pressure, and flow rate through a powerful host computer intelligent graphics software system, and it can monitor the abnormal operation of the equipment. It can predict in advance and make corresponding output actions quickly according to the equipment process. Mobile terminals such as mobile phones and handheld PDAs can access the web interface of the control system through the public 5G network, and can access the operating curves and data of the equipment under the permission of the authority; in addition, various departments of the oilfield enterprise can also access and browse the operating status of the system through the Internet Or remote control, which greatly facilitates the efficient management of oil pumps by oilfield companies.

The main function of the sensor node is to collect current and voltage data, perform A/D conversion on the data and undergo a digital filtering process. These sensors are all equipped with wireless ZigBee communication modules. They can upload data to the sink node and accept commands from the sink node at the same time. Make the necessary adjustments to the output. A typical oil pump sensor node is usually composed of a sensor module, an analog filter module, an A/D conversion module, a D/A output module, an I/O control module, a node processor CPU, a wireless RFID module, a node storage module, and a power management module, etc [26]. Composition, the block diagram of its hardware composition is shown in Figure 2.

Figure 2.

The pressure sensor of this system adopts Rosemount sensor with a measurement accuracy of 0.1%, the measurement range is 0-10 Mpa, the DC power supply voltage is DC24V, the output sensing signal is 4-20mA, and the medium temperature is -20-90°C. The flow sensor adopts FD-M100AT sensor, the detection distance is 4-150 mm, the switching frequency is 1000 Hz, the output mode is NPN, and the response time is less than 0.5 ms. It can realize the detection of liquid flow in harsh environments, and is very suitable for oil pipelines in oil field enterprises, field arrangement environment. The model used for the temperature sensor is DS18B20, and its range is - 55 to 125°C. It can sample and quantify the temperature data, and the resolution can generally reach 0.0625. The working voltage is 3.0 to 5.5V. The format only needs a maximum of 740ms, which is suitable for the low power consumption of sensor nodes.

Convergence node (ZigBee well group control communicator, also known as gateway node) is responsible for collecting local sensor signals and interacting data with the IoT central computer through the network layer [27]. When the central computer of the Internet of Things sends a control command through data calculation, the sink node immediately sends it to the sensor node through wireless communication in real time after receiving it. After the sensor node unpacks the communication command, it specifically performs D/A output or I/O control. The aggregation node is mainly composed of the lower wireless communication module, the upper wireless communication module, the node processor CPU, the node LCD display module, the node key control module, the node storage module and the power management module [28-29]. The block diagram of its hardware composition is shown in the figure 3. The sink node itself is a relatively independent local oil pump control instrument. It has two modes: Remote and Local. Normally, it is in the remote mode. Once the communication delay or communication failure occurs, the sink node will automatically switch to the local mode according to the instrument backup rule base Automatic regulation of the oil pump.

Figure 3.

The lower layer wireless communication module selects the CC1100 communication module, which can directly receive the information collected by the perception layer. The upper wireless communication module selects the SRWF508A wireless module, and the transmission distance between the node and the network layer base station can reach 3.5km. It is easy to realize long-distance transmission of collected information by building a wireless network. The node storage module uses the AT 24C256 chip, and the storage capacity can be easily expanded through multi-chip interconnection. The node LCD display module can display the oil pump operation collection information and equipment status transmitted by the current sensor node, which is convenient for users to directly view on site. The Key button control module is convenient to use direct local manual input to run the control rule library, set the working mode of the node, or directly issue control commands to start and stop the oil pump equipment, etc., and it uses the matrix scanning method to work.

The node program is mainly based on the design of the single-chip microcomputer program. In this system, the sensor node and the sink node have many similarities in their working principles, and their programming methods are similar. Due to space limitations, only the workflow of the sink node single-chip microcomputer program is given below, as shown in the figure As shown in figure 4, the node program uses a modularized subroutine design mode, which is composed of multiple relatively independent subroutines called by the main control loop program. Each round of cycle through these subroutines mainly completes the lower sensor data communication acquisition, data storage and processing, Key button processing, upper-layer wireless communication data submission, node data display and other tasks, set the corresponding maximum running cycle time for each subroutine during the main loop running, so that if a node occurs during the running of a loop subroutine For exceptions, the system uses the stack to record the exception numbers in sequence, and unify to the last step to focus on popping the stack to execute the exception handling subroutine processing, which can improve the program operation efficiency and the timeliness of peripheral wireless communication response.

Figure 4.

The single-chip microcomputer program of the aggregation node is written in an MCU development environment compatible with C language syntax, and the code of the main loop program is as follows:

   void main() /*main loop function definition*/

  {

   GatherGaugeLoad();/*Gathering node

power-on parameter initialization*/

       while (1) /*Infinite loop until the node is powered off*/

     {

   DownWirelessComm();/*lower layer

communication, collect oil pump data*/

     DataProcessSave();/*data normalization processing and saving*/

     KeyInService();/*User Key button processing*/

     UpWirelessComm();/*upper layer

communication, execute remote command*/

     KeyInService();/*Node instrument LCD displays required data*/

     if (RunError()!=0) /*There is an exception in this round*/

       Exceptionhand();/*Unified processing cycle exception*/

  }

  }

The wireless network communication between the sensor node and the aggregation node of this system is based on the traditional MAC protocol. It mainly realizes the establishment and maintenance of wireless data links between the devices of each layer of communication. The MAC data frame adopts the time slot CSMA/CA mechanism. It uses 3 the three parameters are the back-off index NB, the collision window CW and the back-off index BE to achieve reliable data transmission [32].

According to the actual needs of oil field enterprises for remote monitoring of well group oil pumps, the IoT central computer monitoring system software is mainly composed of 7 modules, as shown in Figure 5, including the system user management module, network communication processing module, data record processing module, process screen A display module, a dynamic curve display module, a sensor node parameter module, and a convergence node parameter module.

Figure 5.

The system user management module is used to control the user’s authority to use the software, including remote desktop terminal and mobile terminal users, and can display or maintain basic information records of logged-in users. The network communication processing module is the key core module of the system. On the one hand, it performs real-time data exchange with multiple convergence nodes of the system through the dedicated network layer network, and on the other hand, it provides data services for enterprise remote mobile terminals and desktop terminals through the public 5G network and the Internet.. The system control center adopts the C/S design mode and adopts the TCP/IP asynchronous communication mode. It handles network-related work through proxy callback functions, so that there is no need to block or suspend threads when performing network operations. The system provides a standard Modbus TCP RTU-based interface for remote desktop or mobile terminal web-based page access.

The main function of the data record processing module is to normalize the data of all oil pumps in the oil station monitored by the system, and then periodically (the time can be set, usually 5 minutes) record it to the database server [33-34]. The process screen display module is the most important man-machine interface of the system. It uses virtual instrument technology to display the operation status of each component of the oil pump and the corresponding position of the sensor parameters in real time, and can prompt whether the system has an alarm. The general refresh time of the screen is 5S. The function of the dynamic curve display module will display each sensor parameter in the form of a curve for each oil well. The dynamic curve has historical data curve display and a short-term operation trend graph in the near future. The built-in fault diagnosis algorithm of the system can predict whether the future will there will be a possibility of failure, which allows users to prepare in advance and prevent problems before they happen [35]. The sensor node parameter module can set a series of parameters such as the type of the system oil pump sensor, the working mode of the sensor, and the communication mode of the sensor network. The aggregation node parameter module can set a series of parameters such as the working mode of the aggregation nodes of each oil station in the system, the backup logic rule base, and the node network communication mode.

Combining the operation characteristics of the oil well oil pump process and the user’s fault diagnosis discovery and processing requirements, the fault diagnosis expert system structure shown in Figure 6 is designed.

Figure 6.

It consists of 7 parts, including: oil well pump man-machine interface, system interpretation mechanism, system knowledge acquisition mechanism, fault tree structure knowledge base, LSTM machine learning reasoning mechanism, system evaluation learning optimization mechanism and oil pump skid status information monitoring mechanism. The description of each part is as follows:

Man-machine interface of oil well oil pump: The terminal for displaying various information for the system operation and the operation interface for the user to interact with the system. On the one hand, the user inputs the original device fault diagnosis rules, logic rules and other knowledge into the system through this interface; on the other hand, the system outputs the diagnosis results through this interface and asks or answers reasoning questions to the user. In addition, the operator can input the manual intervention parameters of the diagnosis system through this interface at any time according to experience and knowledge, so as to achieve real-time manual correction of the diagnosis system, so that it can quickly enter the correct diagnosis track.

This system plans to use the LSTM algorithm, and its core is the BP neural network, so it is necessary to introduce the principle of the neural network in advance [45].

The BP neural network module used in the bottom layer of machine learning in this oil well oil pump fault diagnosis system is divided into four layers: input layer, fuzzy layer, fuzzy reasoning layer and output layer. The BP neural network input layer nodes correspond to the parameters that need to be controlled by the oil pump system, and the BP neural network output layer nodes correspond to the future measurement data predicted by the sensor. The function of the BP neural network module is to train and learn a large number of historical data of oil well pump sensors. The BP neural network controller can continuously optimize the real-time parameters of the machine learning fault diagnosis of the oil well oil pump by adjusting the weight coefficients of each layer, so as to realize the self-adaptive adjustment of fault diagnosis, prediction and alarm. The structure diagram of BP neural network is shown in Figure 7.

Figure 7.

In this network diagram, the first layer is the input layer of the BP neural network module. Each neuron node in this layer represents an input variable of the system such as oil well pressure, flow rate, temperature and other information. Since the system needs to control the oil well pump Inlet oil pressure, outlet oil pressure and output oil flow, so the neural network module of the input layer fuzzy neural network LSTM machine learning control is composed of three neurons, that is, three nodes, and the main function of this layer is that the neurons will input the variable The value is passed to the neurons in the fault diagnosis and prediction layer through the action function, and the input and output in this layer are as follows: 1 Ii(1)=xi 2 Oij(2)=xi

The second layer in the BP neural network is the fuzzy layer of fault diagnosis prediction. Each neuron in this layer is used to simulate the membership function of the input variables of fault diagnosis prediction. The input and output of fault diagnosis prediction in this layer are expressed as: 3 Iij(2)=−(xi−cij)2bij2 4 Oij(2)=exp(Iij(2))

LSTM is short for Long Short-Term Memory Neural Network. It is a variant of Recurrent Neural Networks (RNN) that excels at processing sequential data. LSTM was proposed by Hochreiter and Schmidhuber in 1997. LSTMs work by using gates to control the flow of information in the network. A gate is a structure in a neural network that controls the flow of information. LSTM has three gates: input gate, forget gate and output gate. Input gates control the entry of new information into the network. Forget gates control the deletion of old information from the network. The output gate controls the information output by the network. The structural principle of the LSTM machine learning model is shown in Figure 8.

Figure 8.

LSTMs process sequence data through the following steps:

The input information is fed into the LSTM network.

LSTM has been widely used in natural language processing, machine translation, speech recognition, time series prediction and other fields. LSTM is a very powerful neural network model that has achieved excellent performance in many tasks.

Here are some advantages of LSTMs:

Can handle long sequence data

Many parameters require a large amount of data for training, easy to overfit, Overall, LSTM is a very powerful neural network model that achieves excellent performance in many tasks.

The data basis formula adopted by the LSTM algorithm is as follows: 5 ft=σ(Ufht−1+Wfxt+bf) 6 it=σ(Uiht−1+Wixi+bi) 7 ćt=tanh(Ucht−1+Wcxt+bc)$${_t} = tanh({U_c}{h_{t - 1}} + {W_c}{x_t} + {b_c})$$ 8 ct=ftct−1+itćt$${c_t} = {f_t}{c_{t - 1}} + {i_t}{_t}$$ 9 ot=σ(Uoht−1+Woxt+bo) 10 ht=ottanh(ct)

In Figure 8, the cell state ct – 1 and hidden state ht – 1 at the previous time step is tuned by the forget gate switch to generate the corresponding states ct and ht at the current time step, where the forget gate vector ft is expressed in Formula 5.

In Formula 5, σ denotes the sigmoid function layer which determines the parameters to be updated, U and W denote the weight matrices to be applied to the hidden state vector ht – 1 at the previous time step and input vector xt at the current time step, respectively, and b denotes the bias vector to reduce inaccuracy during training.

In order to update the cell state, the input gate utilizes the sigmoid function σ and tanh function to process the cell state vector ct – 1 at the previous time step to generate the cell state vector ct at the current time step. The working expression of the sigmoid function σ and tanh function are illustrated in Formula 6 and Formula 7, respectively.

Then, the cell state ct – 1 at the previous time step is processed by forget gate, and the product of tanh layer output and sigmoid layer output are added together to update to the cell state ct at current time step, as illustrated in Formula 8.

Finally, after being processed by the same sigmoid function and tanh function in the output gate to determine the output parameters and the weighted value in Formula 9, the hidden state ht at the current time step is output by an LSTM cell to transfer to the next one or use as the final output in Formula 10.

This system builds a data set based on a large amount of historical data such as temperature, pressure, and valve opening monitored online by all sensors of the oil well oil pump. Through the LSTM machine learning neural network prediction model and effective training, the future of the oil well oil pump can be obtained more accurately. The occurrence temperature, pressure, valve opening trend curve and the probability of failure data, once the abnormality is predicted by the fault diagnosis machine learning reasoning mechanism, an alarm will be issued immediately.

In the process of system experiment, the construction of back-end server is a crucial step. The back-end server is responsible for receiving and processing front-end requests, performing data processing, implementing business logic, and returning the results to the front-end. In this design, we need to build a reliable, efficient, and secure back-end server to support various functions and requirements of the system. We chose Python as the back-end development language, combined with the development of the Flask framework. Python is a concise and powerful programming language with rich library and framework support, suitable for rapid development and deployment. Flask is a lightweight web framework with flexible routing management and template rendering functions, easy to learn and use.

LSTM normalization refers to normalizing the input and output in the LSTM network to improve the training speed and performance of the network. Commonly used normalization methods include, but are not limited to: Batch Normalization, Layer Normalization, and Instance Normalization, etc.

Among them, Batch Normalization is a commonly used normalization method, which can normalize the input on each mini-batch to reduce the internal covariate displacement, thereby improving the training speed and performance of the network. Layer Normalization is another commonly used normalization method, which can normalize the input on each layer to reduce the internal covariate displacement, thereby improving the training speed and performance of the network. Instance Normalization is a normalization method for image processing tasks, which can normalize the input on each channel to reduce the internal covariate displacement, thereby improving the training speed and performance of the network.

The following takes oil well oil pump pressure as an example to carry out fault diagnosis and prediction. The algorithm steps are as follows: Step 1:

Obtain the historical data of oil well pump pressure to form a training data set;

The algorithm formula used for normalization in this system is as follows: 11 xstandard=x−xminxmax−xmin 12 xscaled=xstandard*(max?min)+min

In Formula 11 and Formula 12, max represents the maximum value, and min represents the minimum value.

In the LSTM prediction experiment of the oil well pump pressure in this system, the system uses the normalized root mean square error (n-RMSE) to characterize the pressure prediction performance of the oil well pump for different models given different lengths of input windows and prediction windows. In order to obtain the indicator data n-RMSE in Python, it is necessary to sum the deviation squares between the predicted value of the oil well pump pressure and the actual value of the oil well pump pressure, and then divide the sum by the size of the oil well pump pressure test sample to calculate the MSE. Then, the indicator RMSE of oil well delivery pump pressure machine learning can be calculated by taking the root of MSE. Finally, the metric nRMSE for machine learning of oil well pump pressure is calculated by dividing the RMSE by the mean value of the real dataset. Usually, a lower index nRMSE of oil well oil pump pressure machine learning will indicate that the model prediction performance is better, and at this time it can be used for real-time diagnosis of oil well oil pump pressure faults.

The main technical index parameters of the LSTM machine learning algorithm for oil well pump pressure failure are shown in formula 13, formula 14 and formula 15:

MSE (Mean Squared Error): 13 MSE=1N∑i=1N(yi−y^i)2

RMSE (Root Mean Squared Error): 14 RMSE=1N∑i=1N(yi−y^i)2

n-RMSE (Normalized Root Mean Squared Error): 15 nRMSE=RMSEy¯

This experiment is based on the Python development environment. The oil well pump pressure sensor recorded in the project is used as the data set, and the data is cleaned, and then normalized. There are more than 3,000 pieces of data in this dataset, and the first 20 pieces of data are shown in Table 1.

Then, the LSTM machine learning model is used for training to obtain the effective parameters of the model, and finally the prediction is made. The error of the training collection and distribution point is shown in Figure 9, the error of the testing collection and distribution point is shown in Figure 10, and the actual curve and predicted curve of the oil well pump are shown in Figure 11.

Figure 9.

Figure 10.

Figure 11.

It can be seen from Figure 11 that the prediction accuracy of the LSMT machine learning algorithm is very high, which fully meets the needs of the oil well oil delivery pump for late pressure over-limit fault diagnosis and alarm.

According to the aforementioned evaluation indicators, the data obtained by machine learning operations are shown in Table 2.

It can be seen from Table 2 that the machine learning model proposed in this paper has a very small error in the fault diagnosis and prediction of oil well pumps.

In the actual use of the system, when it is predicted that the pressure of the oil pump in the oil well will exceed the safety limit in the future, the oil pump will be diagnosed in advance and the engineer will be reminded to deal with it on site, so as to truly prevent problems before they happen.