A method for predicting the probability of electric valve failure

Abstract

The present application provides a method for predicting the probability of electric valve failure, including: collecting historical data of an electric valve failure monitoring system, the historical data including chip temperature, signal strength, tension and compression value, hydrogen sulfide concentration, hydrogen concentration, and valve switching time; calculating the historical failure probability based on the historical data; inputting the historical failure probability into a trained failure probability prediction model, and outputting the failure probability prediction value of the electric valve. After the model training is completed, the present application realizes a more accurate description of the physical quantity data such as the failure probability and gas concentration of the electric valve equipment in the next half hour through serialized prediction, so as to facilitate the early maintenance and inspection of the equipment and avoid the occurrence of safety accidents.

Description

Technical Field

[0001] The present application relates to the technical field of electric valve fault monitoring, and in particular to a method for predicting the probability of electric valve faults. Background Art

[0002] In the chemical industry, the key devices and equipment areas are usually prone to internal harmful media leakage. Since the pipelines are in a stable static position after installation, leakage is relatively rare within the scope of meeting standard specifications and service life. For transmission equipment, such as pressure pumps, control valves, etc., these equipment are in continuous or regular operation and switching actions, and the possibility of internal media leakage is extremely high. Control valves are generally installed at the front and rear ends of pressure pumps or key positions in the device to control the flow, pressure, temperature, etc. of these high-risk media in real time.

[0003] The current problem is that the maintenance cost of electric valves in the chemical industry is high and the efficiency is low, and it is impossible to intelligently and scientifically predict and prevent failures after a scheduled time. Summary of the invention

[0004] In view of this, the purpose of this application is to propose a method for predicting the probability of electric valve failure, which can solve the existing problems in a targeted manner.

[0005] Based on the above purpose, the present application proposes a method for predicting the probability of electric valve failure, including:

[0006] Collect historical data of the electric valve fault monitoring system, the historical data including chip temperature, signal strength, tension and compression value, hydrogen sulfide concentration, hydrogen concentration and valve switching time;

[0007] Calculate the historical failure probability according to the historical data;

[0008] The historical failure probability is input into a trained failure probability prediction model, and a failure probability prediction value of the electric valve is output.

[0009] Furthermore, the electric valve fault monitoring system includes: MCU, hydrogen sulfide sensor, hydrogen sensor, tension and pressure sensor, and actuator; wherein,

[0010] The tension and compression sensors are connected to the A / D conversion interface of the MCU through a signal amplifier. The hydrogen sulfide sensor and the hydrogen sensor are connected to the A / D conversion interface of the MCU through their respective signal transmitters. The D / A conversion interface of the MCU is connected to the opening control port of the actuator through a voltage-to-current transmitter.

[0011] Furthermore, the open position signal and the closed position signal of the actuator are connected to the GPIO input interface of the MCU, and the GPIO output interface of the MCU is connected to the control open port and the control close port of the actuator through the actuator relay respectively.

[0012] Furthermore, the system further includes a ME3616 module matching circuit and an STM32 hardware minimum system, which are respectively connected to the UART interface of the MCU. The STM32 hardware minimum system is also connected to a signal amplifier and a voltage-to-current transmitter through a 5V to 24V voltage amplifier.

[0013] Furthermore, the system further includes an NB-IoT module, which is connected to the MCU to send and receive data of the MCU between the MCU and an external server.

[0014] Furthermore, the training process of the trained fault probability prediction model includes:

[0015] Collect historical failure probability data of the electric valve fault monitoring system within a certain period of time;

[0016] Dividing the historical failure probability data into a training set and a test set;

[0017] Establishing an LSTM neural network model, training the LSTM neural network model using the training set, and testing the model using the test set;

[0018] The training and testing results of the model are evaluated, and parameters are adjusted according to the evaluation results until the model converges.

[0019] Further, including:

[0020] Collecting real-time data of the electric valve fault monitoring system, the real-time data including chip temperature, signal strength, hydrogen sulfide concentration, and hydrogen concentration;

[0021] The four items of real-time data are input into their respective trained LSTM models one by one to obtain the predicted value of each item of real-time data.

[0022] Furthermore, the fault probability prediction model is an LSTM model.

[0023] Furthermore, the specific parameters of the LSTM model are as follows: the model structure is an LSTM layer with 10 LSTM neural units and a fully connected layer, the time step is 1000, the training batch size is 256, the loss function is the mean square error function, the optimizer is the adaptive moment estimation, and the learning rate is dynamically determined by the Keras framework, where one round of training consists of multiple batches, and the training batch size represents the amount of data selected for a single batch.

[0024] Further, including:

[0025] When the predicted value of the fault probability is higher than the preset threshold, the physical quantity that needs to be alarmed is determined based on the predicted data of the chip temperature, signal strength, hydrogen sulfide concentration, and hydrogen concentration at the corresponding time.

[0026] In general, the advantages of this application and the experience it brings to users are: after the model training is completed, this application can achieve a more accurate description of the failure probability, gas concentration and other physical quantity data of the electric valve equipment in the next half hour through serialized prediction, so as to facilitate the early maintenance and inspection of the equipment and avoid the occurrence of safety accidents. BRIEF DESCRIPTION OF THE DRAWINGS

[0027] In the accompanying drawings, unless otherwise specified, the same reference numerals throughout the multiple drawings represent the same or similar parts or elements. These drawings are not necessarily drawn to scale. It should be understood that these drawings only depict some embodiments disclosed in the present application and should not be regarded as limiting the scope of the present application.

[0028] Figure 1 The following is a flowchart of the prediction model application of the present application.

[0029] Figure 2 The structural diagram of the electric valve remote fault monitoring system is shown.

[0030] Figure 3 A schematic diagram of the fault probability prediction model structure is shown.

[0031] Figure 4 A schematic diagram showing the results of a single prediction experiment.

[0032] Figure 5 A schematic diagram of the failure probability prediction of the present application is shown;

[0033] Figure 6 A schematic diagram comparing different step length predictions of the present application is shown.

[0034] Figure 7 A schematic diagram of the model prediction results for the same number of training rounds when the time step of the application is 1000 and the training batch sizes are 1, 128, 256 and 2000 respectively.

[0035] Figure 8 The fault prediction data interaction flow chart of the present application is shown. DETAILED DESCRIPTION

[0036] The present application will be further described in detail below in conjunction with the accompanying drawings and embodiments. It is to be understood that the specific embodiments described herein are only used to explain the relevant invention, rather than to limit the invention. It is also necessary to explain that, for ease of description, only the parts related to the relevant invention are shown in the accompanying drawings.

[0037] It should be noted that, in the absence of conflict, the embodiments and features in the embodiments of the present application can be combined with each other. The present application will be described in detail below with reference to the accompanying drawings and in combination with the embodiments.

[0038] Fault prediction is a prerequisite for predictive maintenance. With the rapid development of artificial intelligence, existing technologies have the ability to analyze and predict data in real time. In recent years, research on LSTM applications in equipment fault prediction and data prediction has achieved good results. This application uses LSTM neural network to predict the probability of equipment failure. It explains the research on system equipment failure probability prediction from three aspects: LSTM working principle, model construction process, and prediction model verification experiment. Finally, the preliminary practice of the fault probability prediction method is given.

[0039] 1 Principle of LSTM Neural Network

[0040] As a variant of the Recurrent Neural Network (RNN), LSTM has changed the status quo in the field of machine learning and neural network computing, and has also improved Google's speech recognition and text translation. In addition, LSTM is also widely used in time series prediction, such as stock prediction and weather prediction, and has achieved good results. This section first introduces the concept of RNN and the iteration problem, and then specifically explains the concept of LSTM and the LSTM model structure used in this application.

[0041] 1.1 Recurrent Neural Network (RNN)

[0042] RNN is different from traditional ANN. The RNN network structure is serial, that is, the input at the current moment will also be input to the neuron at the next moment, which can maintain the dependency in continuous data and is mainly used to process time series data. Although RNN is designed to learn long-term dependencies, standard RNN also has difficulty in long-term information preservation. Bengio et al. proposed that due to the iteration problem of RNN, it will bring about the phenomenon of "gradient disappearance" and "gradient explosion". Each output is determined by the current input and previous information. Sometimes only the latest information is needed as the input of the current prediction task, but sometimes more previous information is needed. It is not enough to refer to the latest information alone. However, if this information is called at a place far away from the current prediction task, such as more than 8 to 10 steps, that is, it has "long-term dependency", RNN may have the problem of "gradient disappearance" or "gradient explosion", that is, the predicted data only contains recent information, resulting in inaccurate prediction.

[0043] 1.2 Long Short-Term Memory Neural Network (LSTM)

[0044] In order to solve the problems of "gradient explosion" and "gradient vanishing" when learning long-term dependencies, Hochreiter et al. proposed the long short-term memory network to improve the traditional recurrent neural network model. LSTM limits the error signal to each unit by improving the original RNN unit into a constant error carousel (CEC). At the same time, each LSTM unit itself is also a cyclic structure. CEC expands it by adding input gates and output gates, making the model nonlinear and eventually forming a storage unit.

[0045] A common LSTM unit consists of a cell, an input gate, an output gate, and a forget gate. The forget gate did not exist in the initial version of LSTM and was later proposed by Gers et al. The LSTM unit memorizes the parameter value within a given arbitrary time interval, and the three gates are responsible for regulating the information flow associated with the LSTM unit. A complete LSTM architecture consists of a set of recurrently connected subnetworks, namely memory blocks. The memory block can maintain the cell state over time and regulate the information flow through nonlinear gate units. An LSTM unit structure contains gates, input signals, output signals, activation functions, and peephole connections.

[0046] To illustrate the working principle of the LSTM model, assume that an LSTM model consists of A blocks and B inputs, and its forward propagation process is as follows:

[0047] (1) Block information input. (t)The output y of the previous LSTM unit (t-1) Combined, the formula is expressed as:

[0048] z (t) =tanh(W z x (t) +R z y (t-1) +b z ). (5.1)

[0049] Where t represents the tth LSTM unit in the LSTM network, W z and R z x (t) and (t-1) The weight of b z is the bias vector.

[0050] (2) Input gate. In this step, the input x (t) , the output of the t-1th time step and the cell value c (t-1) Update the input gate with the following formula:

[0051] i (t) =Sigmoid(W i x (t) +R i y (t-1) +p i c (t-1) +b i ). (5.2)

[0052] W i , R i 、p i The input x is (t) , t-1 layer unit output y (t-1) , the cell state c at the t-1th time step (t-1) The weight of b i is the bias vector. After the block input and the update of the input gate are completed, LSTM can calculate the cell c that retains part of the information to the current t-layer unit (t) This information will be obtained from the result of the block input z (t) And the activation result i after the input gate is updated (t) Make your selection.

[0053] (3) Forget gate. Through the forget gate, the LSTM unit will decide from the t-1th cell state c (t-1) Which information to keep or delete. Based on the current input value x (t) , the output y of the t-1th layer unit (t-1) , the cell state c at the t-1th time step (t-1) , hole connection and the bias b of the forget gate f To calculate the forget gate activation value f at the tth time step (t) , the formula is as follows:

[0054] f (t) =Sigmoid(W f x (t) +R f y (t-1) +p f c (t-1) +b f ). (5.3)

[0055] (4) Cell state. The cell state value is determined by the block input z (t) , input gate activation value i (t) , forget gate activation value f (t) And the cell state of the previous t-1 time step is calculated, and the calculation process is expressed as:

[0056] c (t) =z (t) i (t) +c (t-1) f (t) . (5.4)

[0057] (5) Output gate. The output gate result is passed through the current input x (t) , the output y of the previous time step (t-1) And the cell state value c (t-1) The calculation is performed and the result is given as:

[0058] o (t) =Sigmoid(W o x (t) +R o y (t-1) +p o c (t-1) +b o ). (5.5)

[0059] (6) Block output. Finally, the output calculation result of the tth time step is:

[0060] y (t) =tanh(c (t) ) (t) . (5.6)

[0061] After the above six steps, we can get the block output and cell state of the t-th time step. The Sigmoid function formula is:

[0062]

[0063] 2 Construction of Failure Probability Prediction Model

[0064] This application uses LSTM to predict the failure probability of electric valves. The prediction model application process is as follows: Figure 1 As shown in the figure. After the amount of data collected by the terminal reaches a complete cycle, the historical data is analyzed to generate a fault probability prediction model. This section describes the prediction model construction process from three parts: data collection, preprocessing, and fault probability prediction model.

[0065] 2.1 Data Collection

[0066] This application divides terminal failures into operational failures and safety failures. Operational failures mainly occur during the operation of electric actuators, such as damage to actuator components, resulting in execution jams and other problems; safety failures mainly refer to problems such as internal / external leakage of valve pipeline media and excessive chip temperature. This application uses the chip temperature, signal strength, tension and pressure values, H2S concentration, H2 concentration, and valve switching time as data for failure probability prediction. The tension and pressure values ​​and valve switching time are nonlinear in their impact on the probability of equipment failure and have a high correlation with equipment operation failures, so they are used as operational failure probability prediction data, and the remaining data are safety failure probability prediction data.

[0067] The operating data of the electric actuator is as follows: Figure 2 The hardware terminal shown acquires data. Figure 2 In the embodiment, the hardware terminal includes an MCU, a hydrogen sulfide sensor, a hydrogen sensor, a tension and compression sensor, and an actuator; the tension and compression sensor is connected to the A / D conversion interface of the MCU through a signal amplifier, the hydrogen sulfide sensor and the hydrogen sensor are connected to the A / D conversion interface of the MCU through their respective signal transmitters, and the D / A conversion interface of the MCU is connected to the opening control port of the actuator through a voltage-to-current transmitter. The open position signal and the closed position signal of the actuator are connected to the GPIO input interface of the MCU, and the GPIO output interface of the MCU is connected to the control open port and control close port of the actuator through the actuator relay. It further includes a ME3616 module matching circuit and an STM32 hardware minimum system, which are respectively connected to the UART interface of the MCU, and the STM32 hardware minimum system is also connected to the signal amplifier and the voltage-to-current transmitter through a 5V to 24V voltage amplifier.

[0068] The data obtained by the MCU is sent to the SQL Server database on the server through NB-IoT. The terminal IMSI number is 460113003233760, and the time is from 2021-10-15 to 2021-10-21, with a total of 16,560 records.

[0069] 2.2 Data Preprocessing

[0070] The original data is the data collected by the equipment, and it is inevitable that the data is not standardized. It is necessary to standardize the format of the input data, standardize the data for time series prediction, and normalize the data to facilitate gradient calculation and speed up convergence.

[0071] Since occasional frame loss may occur during the communication between the terminal and the server, and the basic platform of the terminal and the server uses a "periodic reconnection mechanism" to ensure a stable connection, data may be lost at some point. Data compensation is required to fill in the gaps. The number of missing data frames is determined based on the data frame number, sending time, and sending interval. The formula used for the number of missing data frames is:

[0072]

[0073] Among them, time (n) Represents the sending time of the nth data, time (n+1 ) represents the sending time of the n+1th data, frequency (n) Represents the sending interval of the nth data. The time interval between two frames of data sent by the terminal is generally 30-60 seconds. The formula used to assign missing data is:

[0074]

[0075] Among them, value (t) Represents the data lost at time t, where t is in the range [n+1, n+count], value (n) Indicates the previous frame data, value (n+1 ) indicates the next frame of data.

[0076] The chip temperature range selected by the system is -40℃→85℃, and the safety range is set to -20℃→70℃; the signal strength is the indicated value calculated proportionally by the signal strength of the NB-IoT module. In order to ensure the stability of communication, the safety range is set to 20~100; the tension and compression value and the valve switching time are set to a safety value that is 1.5 times the data collected during stability; the lower explosive limit (LEL) of H2 is 4% by volume, and the warning value in the model is set to 25%LEL; the maximum allowable concentration of H2S is 10mg / m 3 , which is approximately equal to 6.59ppm. The warning value is accurate to 0.1ppm and is set to 6.6ppm. The other parameter ranges and warning values ​​used by the model are shown in Table 1. The warning value here is the value that reminds the user to perform maintenance.

[0077] Table 1. Failure probability prediction model parameters

[0078]

[0079] On this basis, the historical comprehensive fault probability is filled in, and the input parameters are [P1, P2, P3, P4, A1, A2], where P1-P4 are the data sets of chip temperature, signal strength, H2S, and H2 concentration, respectively, and A1 and A2 are the data sets of tension and pressure values ​​and valve switching time. Compare the physical quantity with the warning value in Table 1 to obtain the fault ratio within [0, 1], which is used as a parameter to fill in the comprehensive fault probability. The formula is:

[0080]

[0081] in is the highest failure rate of the safety failure probability prediction data, To predict the highest failure rate of the running failure probability data, the formula is:

[0082]

[0083]

[0084] Among them, P1 (n) Represents the nth data in the chip temperature data set, A1 (n) Represents the nth data item in the tension and compression value data set.

[0085] In order to ensure the accuracy of the prediction model construction, the original data may exceed the normal value range at some point after data filling. The outliers need to be processed and the data outside the normal value range should be filled with the median.

[0086] At the same time, in order to avoid special samples affecting the prediction results, the data set needs to be standardized. Standardized data is used for time series prediction, the format of input data is standardized, and data normalization is performed after generating historical failure probability prediction values. This application uses min-max normalization to limit the data range to (0, 1), and its formula is expressed as:

[0087]

[0088] Among them, X (n) is the nth data in the original data sequence X, min is the minimum value in the sequence X, and max is the maximum value in the sequence X. is the normalized data. After the model is trained using normalized data, the normalized output will be obtained during the inference process. The normalized output is denormalized to obtain the output result within the normal range. The denormalization formula is expressed as:

[0089]

[0090] 2.3 Failure probability prediction model

[0091] In the actual chemical production environment, if the predicted failure probability value is only calculated by the model, when the failure probability value is high, the management personnel cannot judge the authenticity of the predicted failure probability value in real time. Therefore, the system uses the LSTM neural network with full sequence input and output to make real-time predictions on the four safety failure probability prediction data and historical failure probabilities. At this time, the LSTM neural network model is as follows: Figure 3 shown.

[0092] In the fault probability prediction model, all the data detected by the system are filled with historical fault probabilities using formula (5.10), and then separate models are trained for the four safety fault probability prediction physical quantities monitored by the system. Since the historical fault probability is calculated from six types of data, and the tension and pressure values ​​and valve switching time can only collect data at the time of valve switching, it is difficult to predict the rules, and the predicted fault probability cannot be obtained from only four types of prediction data, so the system is also designed and trained separately for the fault probability. In the terminal main loop, this application judges and records the safety threshold of the tension and pressure values ​​and the valve switching time. Although the model structures corresponding to each type of data are similar, the single-dimensional input model of each sequence can better express the characteristics and rules of the corresponding data. The output of the fault probability prediction model is the predicted state data matrix of the physical quantities of the safety fault probability prediction data monitored by this application. The system will provide predicted state data and predicted fault probability at the same time. When the predicted fault probability is high, relevant personnel can manually verify the authenticity of the fault probability through the predicted state data.

[0093] 3 Fault Probability Prediction Model Training Experiment

[0094] This section experiments and analyzes the prediction of the failure probability of the prediction model and the results of each single prediction, and also compares the values ​​of key parameters in LSTM. The experimental evaluation platform is built using the keras deep learning framework in Python, and the CPU is used for calculation. The host CPU model is i7-9750H, and the experimental data uses the data collected in Section 2.1.

[0095] 3.1 Model training experiment

[0096] In the fault probability prediction model of this application, the system designs models for four types of safety fault probability prediction data: chip temperature, signal strength, H2S and H2 concentrations, as well as comprehensive fault probability prediction. 90% of the data in the dataset is used for training, and 10% of the data is used for prediction. The specific parameters of the LSTM model in this section are as follows: the model structure is an LSTM layer with 10 LSTM neural units and a fully connected layer, the time step is 1000, the training batch size is 256, the loss function is the mean square error function, the optimizer is Adaptive Moment Estimation (Adam), and the learning rate is dynamically determined by the Keras framework. One round of training consists of multiple batches, and the training batch size indicates the amount of data selected for a single batch. The experimental results of each single prediction of the fault prediction model are shown in the figure. Figure 4 As shown, the four individual prediction results can clearly demonstrate the accuracy of the fault probability prediction model in the time series prediction task.

[0097] The original data will be normalized before entering the model. After the model training is completed, the output result data range is consistent with the original data after the denormalization calculation is completed. In the H2S and H2 prediction result graphs with low data complexity, the predicted curve and the true curve have a high overlap rate, and the model only has wrong predictions in very few time series. Although the root mean square error (RMSE) in the H2 prediction result reaches 251.46, its original data range is large, so it is still possible to train a good H2 prediction model. In the signal strength prediction experiment with medium data complexity, the prediction curve is still stable and good, and only the overall predicted data is slightly lower. For the complex data set of chip temperature prediction, the overlap between the model prediction curve and the true curve is reduced, but the overall trend of the two is similar, and the amplitude of the prediction curve is relatively small.

[0098] In the case of single-dimensional input, the prediction curve of comprehensive failure probability in training is as follows: Figure 5 As shown in the figure. The RMSE in the comprehensive fault probability training is 0.03, and the fault probability prediction data curve can well conform to the real data law. The fault probability prediction data curve is predicted by the historical fault probability data. When the fault probability prediction result is high, the management personnel can view the specific prediction data of the corresponding time and judge the specific alarm physical quantity and the corresponding prediction value.

[0099] 3.2 Time step performance comparison

[0100] In the design of LSTM model, the time step is one of the most critical hyperparameters. The LSTM model will use the amount of data in a time step to predict the next data, and at the same time affect the model input data dimension. The size of the time step is closely related to the data set. The model will integrate and analyze the rules corresponding to the data of a time step to make model predictions more accurately. The time step is usually different in different tasks, and it needs to be determined in combination with the size of the data set and the rules of data collection and distribution. In order to determine the appropriate time step, this application predicts and analyzes the data at different time steps. The prediction results at each time step are as follows: Figure 6 shown.

[0101] When the time step is 1000, the RMSE of model training is only 0.03, which is the best among the four step sizes. In terms of image performance, when the time step is 1000, the predicted data curve basically coincides with the real data. Another model with a high degree of overlap is the model with a time step of 2000, which has excellent prediction results in the peak. However, when the time step is 2000, multiple training processes often result in RMSE of 0.15, overfitting of the model, and straight line prediction of the data. When the time step is 200 and 500, the model's prediction ability for the prediction set decreases.

[0102] 3.3 Training Batch Size Performance Comparison

[0103] When the training batch size is not set, that is, when the network directly inputs all data into the network in each training, the model will calculate the gradient of all data for back propagation. Since the gradient is calculated from all data, the resulting gradient direction is more accurate. However, in this case, the calculated gradient values ​​vary greatly, and using only a global learning rate will make the model more difficult to fit. At the same time, inputting all data into the network at once will increase the memory burden. In this section, an experiment is conducted on the effect of the training batch size on the model performance to determine the specific parameters of the model. The experimental results are shown in the figure. Figure 7 shown.

[0104] Figure 7The model prediction results for the same number of training rounds when the time step is 1000 and the training batch sizes are 1, 128, 256 and 2000 respectively are shown. When the training batch size is 1, the single-round training time of the model is 13 seconds, which is much longer than the other experimental results, and more rounds of training are required to complete the model fitting. When the training batch size is increased, the system can call parallel computing to greatly accelerate the training. When the training batch size is 128, the time required for a single round of training is only 1106 milliseconds. The larger the training batch size, the more accurate the gradient value obtained by the model in a single batch. However, too large a training batch size will cause the model to over-optimize the model parameters in the gradient calculation, as shown in the prediction result figure when the training batch size is 2000. When the training batch size is 128 and 256, the model training performance is basically the same, and the training batch size value finally selected in this application is 256.

[0105] 3.4 Experimental Conclusion

[0106] In summary, for the prediction target and data set of this application, the model can achieve the best training and prediction results when the data from 2021-10-15 to 2021-10-21 is selected, the time step of the fault probability prediction model is 1000, and the training batch size value is 256. As the system deployment time increases, the available data set will also increase, and the prediction accuracy of the model can be improved by retraining the model.

[0107] 4 Preliminary practice of fault probability prediction program

[0108] After training and verification of the fault probability prediction model, the model can predict the fault probability trend on the original data set. In order to verify the prediction accuracy of the model during the use of the equipment, the model will be tested using a new data set, and an interactive interface between the fault prediction program and the cloud listening program will be designed.

[0109] 4.1 Prediction model validation

[0110] The LSTM model can use the data of the previous time step to predict the next data, and it can output multiple data at one time through the test set. In the real-time processing of the actual system, the system cannot have future data, so the LSTM model can only predict one piece of future data through the existing historical data. To address this problem, the fault prediction model uses serialized output to predict data for a period of time in the future. Serialized output allows the model to use the current predicted data as known data in the next prediction process during the prediction process, so that data prediction can be controlled and continuous. Using this method, 200 data predictions were serialized and output based on the data from January 21, 2022 to January 22, 2022. The experimental results show that the fault probability prediction model is basically correct in judging the trend of future data, and the RMSE calculation is performed by taking the average of the prediction results for 5 times. As shown in Table 2, when 66 items are predicted, the model has the highest accuracy, followed by when 60 items are predicted. In this application, the initial number of predictions is set to 60, corresponding to the situation in the next 30 minutes. Since the fault probability prediction model can only perform trend prediction on future data, the specific number of predictions needs to be determined based on project requirements and model parameters.

[0111] Table 2 Comparison of RMSE of prediction data

[0112]

[0113] 4.2 Fault Probability Prediction and Cloud Monitoring Data Interaction

[0114] Since the fault prediction program is written in Python, and the prediction results need to be used in the cloud listener program and user application written in C#, this application uses the flask library in Python to implement a simple API service in the fault prediction program. Then, the API is called in the cloud listener program and user application to obtain the result of Python's fault probability prediction of the data. The cloud listener broadcasts the corresponding IMSI number and prediction information to the user application and user device. The cloud listener program performs fault prediction data interaction process as follows Figure 8 shown.

[0115] The external API interfaces provided in the fault prediction program include: fault probability prediction, incremental model training, and independent model training. The specific contents are shown in Table 3.

[0116] Table 3 Fault prediction program interface

[0117]

[0118] 4.3 Failure probability prediction result processing

[0119] The fault prediction program does not process the prediction results. The data is sent to the cloud listening program for analysis and processing of the fault probability data. Since the fault prediction program calculates the fault probability as data within [0,1], this application divides the faults into four levels based on experience values ​​and historical operating status, which are good, normal, general, and warning, respectively. When the fault probability prediction level is good and normal, no user notification is made. When the fault probability prediction level is general, the equipment maintenance personnel are reminded to perform equipment maintenance inspections. When the fault probability prediction level is warning, the equipment maintenance personnel are reminded to immediately check the equipment and perform repairs. When there is data in the prediction result sequence whose fault level reaches the warning level, the cloud listening program saves the device information corresponding to the warning data to the fault warning table for storage, and issues a device alarm notification at the same time.

[0120] 5. Application Summary

[0121] This application uses historical data collected by terminal devices, first designs an electric valve fault probability prediction model based on LSTM, and gives the LSTM model construction process. Then, the fault probability prediction model is experimentally compared and analyzed from three aspects: model training, time step, and training batch size. After the model training is completed, a more accurate description of the physical quantity data such as the failure probability and gas concentration of the electric valve equipment in the next half hour is achieved through serialized prediction, so as to facilitate the early maintenance and inspection of the equipment and avoid the occurrence of safety accidents. Finally, a data interaction interface between the fault prediction program and the cloud listening program is designed, so that data can be interacted between the cloud listening program and the fault prediction program, and the real-time forwarding and processing of the prediction data can be realized.

[0122] It should be noted that:

[0123] The algorithms and displays provided herein are not inherently related to any particular computer, virtual system or other device. Various general purpose systems may also be used together with the teachings based thereon. It is apparent from the above description that the structure required for constructing such systems is not intended for any particular programming language. In addition, the present application is not intended for any particular programming language either. It should be understood that the content of the present application described herein may be implemented using various programming languages, and the description of the particular language above is intended to disclose the best mode of implementation of the present application.

[0124] In the description provided herein, a large number of specific details are described. However, it is understood that the embodiments of the present application can be practiced without these specific details. In some instances, well-known methods, structures and techniques are not shown in detail so as not to obscure the understanding of this description.

[0125] Similarly, it should be understood that in order to streamline the present application and help understand one or more of the various inventive aspects, in the above description of the exemplary embodiments of the present application, the various features of the present application are sometimes grouped together into a single embodiment, figure, or description thereof. However, the disclosed method should not be interpreted as reflecting the following intention: the claimed application requires more features than the features clearly stated in each claim. More specifically, as reflected in the claims below, the inventive aspects are less than all the features of the single embodiment disclosed above. Therefore, the claims following the specific embodiment are hereby expressly incorporated into the specific embodiment, wherein each claim itself serves as a separate embodiment of the present application.

[0126] Those skilled in the art will appreciate that the modules in the devices in the embodiments may be adaptively changed and arranged in one or more devices different from the embodiments. The modules or units or components in the embodiments may be combined into one module or unit or component, and in addition they may be divided into a plurality of submodules or subunits or subcomponents. Except that at least some of such features and / or processes or units are mutually exclusive, all features disclosed in this specification (including the accompanying claims, abstracts and drawings) and all processes or units of any method or device disclosed in this manner may be combined in any combination. Unless otherwise expressly stated, each feature disclosed in this specification (including the accompanying claims, abstracts and drawings) may be replaced by an alternative feature providing the same, equivalent or similar purpose.

[0127] In addition, those skilled in the art will appreciate that, although some embodiments described herein include certain features included in other embodiments but not other features, the combination of features of different embodiments is meant to be within the scope of the present application and form different embodiments. For example, in the claims below, any one of the claimed embodiments may be used in any combination.

[0128] The various component embodiments of the present application can be implemented in hardware, or implemented in software modules running on one or more processors, or implemented in a combination thereof. It should be understood by those skilled in the art that a microprocessor or digital signal processor (DSP) can be used in practice to implement some or all functions of some or all components in the creation system of the virtual machine according to the embodiment of the present application. The application can also be implemented as a device or system program (for example, a computer program and a computer program product) for performing a part or all of the methods described herein. Such a program implementing the present application can be stored on a computer-readable medium, or can have the form of one or more signals. Such a signal can be downloaded from an Internet website, or provided on a carrier signal, or provided in any other form.

[0129] It should be noted that the above embodiments illustrate the present application rather than limit the present application, and that those skilled in the art may design alternative embodiments without departing from the scope of the appended claims. In the claims, any reference symbol between brackets should not be constructed as a limitation to the claims. The word "comprising" does not exclude the presence of elements or steps not listed in the claims. The word "one" or "an" preceding an element does not exclude the presence of multiple such elements. The present application may be implemented by means of hardware including several different elements and by means of appropriately programmed computers. In a unit claim that lists several systems, several of these systems may be embodied by the same hardware item. The use of the words first, second, and third, etc. does not indicate any order. These words may be interpreted as names.

[0130] The above is only a specific implementation of the present application, but the protection scope of the present application is not limited thereto. Any technician familiar with the technical field can easily think of various changes or substitutions within the technical scope disclosed in the present application, which should be included in the protection scope of the present application. Therefore, the protection scope of the present application should be based on the protection scope of the claims.

Claims

1. A method for predicting the probability of electric valve failure, characterized in that: include: Collect historical data of the electric valve fault monitoring system, the historical data including chip temperature, signal strength, tension and compression value, hydrogen sulfide concentration, hydrogen concentration and valve switching time; Calculate the historical failure probability according to the historical data; Inputting the historical failure probability into a trained failure probability prediction model, and outputting a failure probability prediction value of the electric valve; The electric valve fault monitoring system includes: MCU, hydrogen sulfide sensor, hydrogen sensor, tension and pressure sensor, and actuator; wherein, The tension and compression sensor is connected to the A / D conversion interface of the MCU through a signal amplifier, the hydrogen sulfide sensor and the hydrogen sensor are connected to the A / D conversion interface of the MCU through their respective signal transmitters, and the D / A conversion interface of the MCU is connected to the opening control port of the actuator through a voltage-to-current transmitter; The open position signal and the closed position signal of the actuator are connected to the GPIO input interface of the MCU, and the GPIO output interface of the MCU is connected to the control open port and the control close port of the actuator through the actuator relay respectively; The system further includes a ME3616 module matching circuit and an STM32 hardware minimum system, which are respectively connected to the UART interface of the MCU, and the STM32 hardware minimum system is also connected to the signal amplifier and the voltage-to-current transmitter through a 5V to 24V voltage amplifier; The system further includes an NB-IoT module connected to the MCU to send and receive data of the MCU between the MCU and an external server; When the predicted value of the fault probability is higher than the preset threshold, the physical quantity that needs to be alarmed is determined based on the predicted data of the chip temperature, signal strength, hydrogen sulfide concentration, and hydrogen concentration at the corresponding time.

2. The method according to claim 1, characterized in that The training process of the trained fault probability prediction model includes: Collect historical failure probability data of the electric valve fault monitoring system within a certain period of time; Dividing the historical failure probability data into a training set and a test set; Establishing an LSTM neural network model, training the LSTM neural network model using the training set, and testing the model using the test set; The training and testing results of the model are evaluated, and parameters are adjusted according to the evaluation results until the model converges.

3. The method according to claim 1, characterized in that Further including: Collecting real-time data of the electric valve fault monitoring system, the real-time data including chip temperature, signal strength, hydrogen sulfide concentration, and hydrogen concentration; The four items of real-time data are input into their respective trained LSTM models one by one to obtain the predicted value of each item of real-time data.

4. The method according to claim 1, characterized in that: The fault probability prediction model is an LSTM model.

5. The method according to claim 3 or 4, characterized in that: The specific parameters of the LSTM model are as follows: the model structure is an LSTM layer with 10 LSTM neural units and a fully connected layer, the time step is 1000, the training batch size is 256, the loss function is the mean square error function, the optimizer is the adaptive moment estimation, and the learning rate is dynamically determined by the Keras framework. One round of training consists of multiple batches, and the training batch size represents the amount of data selected for a single batch.

Images