Real-time diagnosis method for mechanical failure of twin-screw oil transfer pump on-line monitoring

Abstract

The application provides a kind of twin-screw oil pump on-line monitoring mechanical fault real-time diagnosis method, comprising: obtaining the fault data of twin-screw oil pump to be trained;Data preprocessing is carried out for each parameter attribute related to the fault;Based on the corresponding input attribute of the corresponding node position of the neural network, the fault related data to be processed is displayed;Get the initial network model trained;Align the fault data node parameters in the target optimization parameters;After the weight is optimized using L-BFGS algorithm, supervised training is carried out, and the BP neural network model trained is obtained;The completed BP neural network model is used to train the fault related data in real-time database in batches, and finally the result of accurate prediction is achieved.The twin-screw oil pump on-line monitoring mechanical fault real-time diagnosis method can greatly reduce the production accident rate of twin-screw oil pump, improve the economic efficiency of enterprise production, and protect the life safety of enterprise workers.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of mechanical fault diagnosis, in particular to a double-screw oil pump online monitoring mechanical fault real-time diagnosis method. BACKGROUND

[0002] The double-screw oil pump is the basic equipment for enterprise production. Whether the mechanical equipment can run safely is directly related to the production efficiency and economic benefit of the enterprise. The technical requirements for the mechanical equipment are also increasing, which puts higher requirements on the safety of the equipment. If the mechanical fault cannot be diagnosed in time, many safety hazards will be left. Therefore, real-time fault diagnosis of mechanical equipment and effective measures to prevent accidents are of great significance to ensure the safety of workers and improve the economic benefit of enterprises.

[0003] In recent years, artificial intelligence technology has developed rapidly, so the core algorithm of artificial intelligence control can also be applied to mechanical equipment fault diagnosis and monitoring. For example, intelligent decision-making algorithms such as genetic algorithm, fuzzy control, and rough set theory have great advantages in fault diagnosis.

[0004] At present, the state monitoring and fault diagnosis system of single-screw oil pumps has been increasingly improved and successfully applied to field production operations. However, due to the complex profile of double-screw oil pumps, the mixing of vibration parameters, process parameters, and electrical parameters, and the many coupled and concurrent faults in the field, there is no complete state monitoring and real-time fault diagnosis system, which is a serious technical gap.

[0005] After searching, Chinese patent application No. CN201811007052.6, published on August 31, 2018, is entitled "Mechanical fault diagnosis method based on improved fruit fly-bat algorithm". The application relates to a mechanical fault diagnosis method based on an improved fruit fly-bat algorithm, which includes the following steps: A. Extracting time domain statistical features and frequency domain statistical features from the collected mechanical running state signals; B. Training a support vector machine using a mechanical fault diagnosis training sample set; C. Taking the fruit fly algorithm as the framework, integrating the echolocation idea of the bat algorithm, designing an improved fruit fly-bat parameter optimization method, and using the method to find the global optimal parameters of the support vector machine; D. Substituting the obtained global optimal parameters into the support vector machine to complete the construction of the fault diagnosis model based on the support vector machine.

[0006] The application proposes a mechanical fault diagnosis technical scheme based on an improved fruit fly-bat algorithm, but it has the following shortcomings: (1) the basic principle of the fruit fly algorithm is to determine the position by iterating the taste concentration, and the iteration process is extremely low in efficiency, and although the improved fruit fly-bat algorithm adds the echo positioning principle of the bat algorithm, it still cannot overcome this shortcoming; (2) the mechanical equipment fault diagnosis requires high diagnosis accuracy, and the fruit fly-bat algorithm has the advantage of being easy to implement, but the accuracy is low, and it is difficult to accurately predict the mechanical equipment fault.

[0007] In the Chinese patent application with the application number CN201310175532.4, a fault diagnosis system and method for an oil delivery pump are involved. The system mainly consists of four piezoelectric acceleration sensors, five temperature sensors, a signal conditioner, a temperature measurement module, a data acquisition card and an industrial control computer. The four piezoelectric acceleration sensors are connected with the signal conditioner, and the signal conditioner, data acquisition card and industrial control computer are connected in sequence. The five temperature sensors are connected with the temperature measurement module, and the temperature measurement module, data acquisition card and industrial control computer are connected in sequence. The method includes: 1) vibration signal acquisition step; 2) signal conditioning step; 3) data acquisition step; and 4) signal analysis step.

[0008] In the Chinese patent application with the application number CN201810246460.0, a fault diagnosis method for an oil delivery pump bearing is involved. The method includes the following steps: (1) using an acceleration sensor to collect bearing vibration signal data, performing EMD empirical mode decomposition and intrinsic mode function IMF envelope spectrum analysis to obtain sample data characteristic physical quantities; (2) processing the sample data characteristic physical quantities based on a bearing fault feature matrix to obtain BPA values as BPA evidence; and (3) improving the D-S evidence reasoning rule to fuse the BPA evidence to obtain a bearing fault diagnosis result.

[0009] The above prior arts are quite different from the present application and cannot solve the technical problems we want to solve. Therefore, we have invented a new online monitoring mechanical fault real-time diagnosis method for a double screw oil delivery pump. SUMMARY

[0010] The purpose of the present application is to provide an online monitoring mechanical fault real-time diagnosis method for a double screw oil delivery pump, which can accurately judge the fault of the double screw oil delivery pump in real time and has high generalization ability.

[0011] The purpose of the present application can be achieved by the following technical measures: the online monitoring mechanical fault real-time diagnosis method for a double screw oil delivery pump, which includes the following steps:

[0012] Step 1, obtaining the fault data of the twin-screw oil pump to be trained;

[0013] Step 2, data preprocessing is performed on the fault-related parameter attributes;

[0014] Step 3, an association between the to-be-processed fault-related data and an input format is established, and the to-be-processed fault-related data is displayed in the corresponding input attributes at the node position of the neural network based on the association;

[0015] Step 4, the obtained preprocessed mechanical fault-related data is used to supervise the training of the BP neural network, and an initial network model trained is obtained;

[0016] Step 5, the fault data node parameters in the target optimization parameters are aligned and confirmed;

[0017] Step 6, the L-BFGS algorithm is used to optimize the weight, and then the supervision training is performed, so as to obtain the BP neural network model trained;

[0018] Step 7, the completed BP neural network model is used to perform batch training on the fault-related data in the real-time database, and finally the precise prediction result is obtained.

[0019] The object of the application can also be achieved through the following technical measures:

[0020] In step 1, the real-time running data and historical record data of the twin-screw oil pump unit are collected as the mechanical fault data for training, and the running state data groups of the same equipment at different time points and in different running modes are obtained.

[0021] In step 1, the collected real-time running data includes nine mechanical running-related parameters, such as transverse displacement, longitudinal displacement, main shaft speed, load, temperature, pressure range, power, dynamic current and running efficiency; the historical database includes vibration state, cumulative running time, lubrication state and motor running state data; a measurement point database is established according to the measurement definition, collection definition and diagnosis parameter definition of the unit, including machine number, motor speed, bearing model, machine speed, gear transmission driving gear and driven gear tooth number.

[0022] In step 1, the real-time running data and historical record data of the twin-screw oil pump unit are collected as the mechanical fault data for training, and the input attributes corresponding to the to-be-trained data are set, the data of the same equipment in different modes are combined first, and the parameter classification is performed by using the clustering classification method.

[0023] In step 1, different fault evaluation indexes are formulated according to different types of data such as vibration parameters, process parameters and electrical parameters, including vibration evaluation indexes, flow evaluation indexes, temperature evaluation indexes, dynamic current evaluation indexes, running time evaluation indexes, lubrication evaluation indexes and motor state evaluation indexes; equipment managers draw diagnostic conclusions of machine faults, including unbalance, misalignment, rubbing, looseness, resonance and bearing faults.

[0024] In step 2, data preprocessing is performed on each parameter attribute related to the fault of the double-screw oil pump, including data migration, data screening, data combination, data completion and data denoising, and the processed vibration displacement signal is subjected to fast Fourier transform to obtain fault frequency characteristics.

[0025] In step 2, the data processing content includes data migration, data screening, data combination, data completion and data denoising; the selected clean data is subjected to enhancement processing, i.e. solving the problems of missing values, redundancy, inconsistency of data, inconsistency of data definition and outdated data, and also including sorting and merging of all data to maintain the integrity and correctness of the data; the original signal of the machine is acquired by the sensor, and the FFT technology is used for preprocessing the original vibration signal, which is used as the acquisition method of the original input data of the neural network; the selection of the mechanical fault related parameters adopts the correlation coefficient method; the calculation formula is as follows: Where r xy represents the sample correlation coefficient, S xy represents the sample covariance, S x represents the sample standard deviation of X, S y represents the sample standard deviation of Y.

[0026] In step 3, different evaluation indexes are formulated by combining the real-time running data and historical record data with the running states of different components of the double-screw oil pump unit, which are used to evaluate the normal and fault states;

[0027] The health and fault state database of the double-screw oil pump unit is constructed by using the real-time running data and historical record data, which is used to measure the health and fault states under different evaluation indexes; the database management regularly stores the diagnostic data uploaded by the offline system in the database and keeps it in the database;

[0028] The preprocessed data is analyzed, the corresponding generalized data batch is set according to the input attributes in the input parameters; the association relationship between the mechanical fault related data to be processed and the input format is established; and the mechanical fault related data to be processed is displayed in the node position of the corresponding neural network according to the corresponding input attributes based on the association relationship;

[0029] The empirical formula for determining the number of hidden layer nodes is Where k is the number of nodes in the hidden layer, m, n represent the number of nodes in the output layer and the input layer respectively, and a = 1-10; the first layer transfer function used in the design of the mechanical failure prediction neural network example model is the tangent S-shaped function tansig(n), and the second layer transfer function is logsig.

[0030] In step 5, the first initialization is performed using the two-loop recursion algorithm, and b1, b2 are set to 0, b1 is the bias between the input layer and the hidden layer, b2 is the bias between the hidden layer and the output layer, b1, b2 are the bias matrices of the neural network, the neural network model weight matrix is generated, the initial value of the weight matrix is obtained, and the mechanical failure data node parameters in the target optimization parameters are aligned and confirmed.

[0031] In step 6, the aligned mechanical failure data node parameters and the network model after initial training are used for incremental supervised training, and the L-BFGS algorithm is used to optimize the weight, and a small amount of data of the previous m iterations is stored to replace the previous Hessian matrix, and the trained BP neural network model is obtained.

[0032] In step 6, the L-BFGS algorithm is repeatedly executed through step iteration. The calculation formula is:

[0033] Let s (k) = x (k+1) -x (k) ;

[0034] Let

[0035] It can be obtained that

[0036] Taking the first m terms, we can get

[0037]

[0038] In the formula

[0039] s—step length;

[0040] y—gradient difference;

[0041] H—Hessian matrix;

[0042] ν, ρ—quantities defined for convenient operation;

[0043] Since the variables ρ, v, s, y can be ultimately calculated by the two vectors of step and gradient difference, only the last m times of step and gradient difference vectors need to be stored for iteration, wherein the determination of step needs to use one-dimensional search, and a total of 2*m+1 N-dimensional vectors need to be stored, and when m is 6, the data to be stored is far less than the Hessian matrix.

[0044] In step 6, the specific process of generating the mechanical fault prediction neural network model is that: based on the error between the mechanical fault historical database instance and the result obtained by the neural network instance model, the weight value of the neural network parameter is updated by the back propagation algorithm, and the neural network is modified according to the weight value, and the model prediction value approaches the target value in continuous modification until a certain accuracy or training times are reached.

[0045] In step 7, whether the monitoring parameter is in the safe range is judged according to the prediction result, if higher than the safety warning line, it is judged that the corresponding parameter index of the double screw oil pump unit is in a fault state, and an alarm is given.

[0046] In step 7, the alarm mode and alarm value of each unit are established according to the initial operation record and historical data accumulation of the equipment; in this way, the system can automatically check and judge the running state of the unit from a large amount of collected data, and the unit with abnormality or defects will automatically give an alarm information; the multi-dimensional alarm function including amplitude alarm, statistical alarm, various alarm modes and alarm value setting methods are adopted, so that the equipment management personnel can quickly and accurately extract the equipment abnormality or fault information from a large number of equipment, and distinguish whether the machine has a problem or not; the multi-clock format report function provides various standard report formats and allows custom report; through the analysis of the neural network and the comparison with the database data, the corresponding fault prompt is given, the state monitoring of the equipment is understood in time through the host computer, the current running state of the equipment, the prediction trend of state change, the diagnosis of fault causes, the inspection and maintenance effect and the active prompt of fault on the host computer are understood, so that the maintenance time for preventive maintenance is won.

[0047] The double screw oil pump online monitoring mechanical fault real-time diagnosis method in the application provides a convenient, easy-to-implement mechanical real-time fault diagnosis and instant prediction method by performing neural network data mining on the pretreated mechanical detection data set. The double screw oil pump online monitoring mechanical fault real-time diagnosis method in the application can accurately judge the double screw oil pump fault in real time and has high generalization ability, so as to assist the enterprise production according to the detection of the double screw oil pump fault, thereby greatly reducing the production accident rate of the double screw oil pump, improving the economic efficiency of enterprise production, and protecting the life safety of workers. BRIEF DESCRIPTION OF DRAWINGS

[0048] Figure 1A data processing block diagram in a specific embodiment of the present invention;

[0049] Figure 2 This is a diagram of the structure of an artificial neural network in a specific embodiment of the present invention;

[0050] Figure 3 This is a flowchart of L-BFGS algorithm optimization in a specific embodiment of the present invention;

[0051] Figure 4 The flowchart is a specific embodiment of the method for online monitoring and real-time diagnosis of mechanical faults of a twin-screw oil pump according to the present invention. DETAILED DESCRIPTION

[0052] It should be noted that the following detailed descriptions are exemplary and intended to provide further explanation of the present invention. Unless otherwise specified, all technical and scientific terms used herein have the same meaning as commonly understood by those skilled in the art to which the present invention belongs.

[0053] It should be noted that the terms used herein are only for describing specific embodiments and are not intended to limit the exemplary embodiments according to the present invention. As used herein, unless the context clearly indicates otherwise, the singular form is intended to include the plural form. In addition, it should be understood that when the terms "comprise" and / or "include" are used in this specification, they indicate the presence of features, steps, operations and / or combinations thereof.

[0054] like Figure 4 As shown, Figure 4 The figure is a flow chart of the method for online monitoring and real-time diagnosis of mechanical faults of a twin-screw oil pump according to the present invention.

[0055] Step 1: Obtain the fault data of the twin-screw oil pump to be trained;

[0056] Collect the real-time operation data and historical record data of the twin-screw oil pump unit as mechanical failure data for training, and obtain the operating status data group of the same equipment at different time points and different operating modes.

[0057] Collected data parameters primarily include nine machine operation-related parameters: lateral displacement, longitudinal displacement, spindle speed, load, temperature, pressure range, power, dynamic current, and operating efficiency. The historical database includes data on vibration status, cumulative operating time, lubrication status, and motor operating status. A measurement point database is established based on the unit's measurement and collection definitions, as well as diagnostic parameter definitions. These include machine number, motor speed, bearing model, machine speed, and the number of teeth on the driving and driven gears of the gear drive. A data input interface for the offline equipment fault diagnosis device is reserved to receive data from the precision inspection system.

[0058] Collect real-time operation data and historical record data of a twin-screw oil pump unit as mechanical fault data to be trained, set input attributes corresponding to the data to be trained, combine data of the same device in different modes, and classify parameters by using a clustering classification method.

[0059] Different fault evaluation indexes are formulated according to different types of data such as vibration parameters, process parameters and electrical parameters, including vibration evaluation indexes, flow evaluation indexes, temperature evaluation indexes, dynamic current evaluation indexes, running time evaluation indexes, lubrication evaluation indexes and motor state evaluation indexes. Equipment managers can obtain diagnostic conclusions of machine faults, such as imbalance, misalignment, rubbing, looseness, resonance, bearing failure and the like.

[0060] Step 2, data preprocessing is performed on each parameter attribute related to the fault;

[0061] Data preprocessing is performed on each parameter attribute related to the fault of the twin-screw oil pump, and the data preprocessing content includes data migration, data screening, data combination, data completion and data denoising. The processed vibration displacement signal is subjected to fast Fourier transform to obtain fault frequency characteristics. The preprocessing method is software preprocessing, and the software uses a programming language combined with a specific module to obtain the result.

[0062] In an embodiment, the data processing content includes data migration, data screening, data combination, data completion and data denoising. This process is to overcome the limitations of current data mining tools, and is a process of enhancing the selected clean data, i.e. solving the problems of missing values, redundancy, inconsistency of data, inconsistency of data definition, outdated data and the like in the data, and also includes sorting and merging of all data to maintain the integrity and correctness of the data. The original signal of the machine is obtained by a sensor, and FFT technology is used to preprocess the original vibration signal, which is used as the acquisition method of the original input data of the neural network. The selection of the mechanical fault related parameters uses the correlation coefficient method. The calculation formula is: Where r xy represents the sample correlation coefficient, S xy represents the sample covariance, S x represents the sample standard deviation of X, S y represents the sample standard deviation of Y.

[0063] Step 3, an association relationship between the fault related data to be processed and the input format is established, and the fault related data to be processed is displayed in the node position corresponding to the neural network in the corresponding input attribute based on the association relationship;

[0064] Through real-time running data and historical record data, combined with the running state of different components of the twin-screw oil pump unit, different evaluation indexes are formulated for evaluating normal and fault states;

[0065] Through real-time running data and historical record data, a health and fault state database of the twin-screw oil pump unit is constructed for measuring the health and fault states under different evaluation indexes, and the database management can regularly store the diagnostic data uploaded by the offline system in the database;

[0066] The preprocessed data is analyzed, the corresponding generalized data batch is set according to the input attribute in the input parameter, the association between the mechanical fault related data to be processed and the input format is established, and the mechanical fault related data to be processed is displayed in the node position of the corresponding neural network with the corresponding input attribute based on the association;

[0067] There is no general theory to determine the number of hidden layer nodes at present, so an empirical formula can be used Where k is the number of hidden layer nodes, m and n represent the number of nodes in the output layer and the input layer respectively, and a = 1-10. The first layer transfer function used in the design of the mechanical fault prediction neural network example model is the tangent S-shaped function tansig(n), and the second layer transfer function is logsig.

[0068] Step 4, using the obtained preprocessed mechanical fault related data to supervise the training of the BP neural network, and obtaining the initial network model after training;

[0069] Step 5, aligning and confirming the fault data node parameters in the target optimization parameters;

[0070] The two-loop recursion algorithm is used for the first initialization, and b1 and b2 are set to 0 (b1 is the bias between the input layer and the hidden layer, b2 is the bias between the hidden layer and the output layer, and b1 and b2 are the bias matrices of the neural network), the neural network model weight matrix is generated, the initial value of the weight matrix is obtained, and the mechanical fault data node parameters in the target optimization parameters are aligned and confirmed;

[0071] Step 6, using the L-BFGS algorithm to optimize the weight, and then performing supervised training to obtain the trained BP neural network model;

[0072] The aligned and confirmed mechanical fault data node parameters and the network model after initial training are used for incremental supervised training, and the L-BFGS algorithm is used to optimize the weight, and a small amount of data of the previous m iterations is stored to replace the previous Hessian matrix, and the trained BP neural network model is obtained;

[0073] The L-BFGS algorithm is repeatedly executed by step iteration. Its calculation formula is:

[0074] Let s (k) = x (k+1) -x (k) ;

[0075] Let

[0076] It can be obtained that

[0077] Taking the approximation of the first m terms, it can be obtained that

[0078]

[0079] In the formula

[0080] s - step length;

[0081] y - gradient difference;

[0082] H - Hessian matrix;

[0083] v, p - quantities defined for convenient operation;

[0084] Since the variables p, v, s, and y can ultimately be calculated from the step length and the gradient difference vectors, we only need to store the last m step length and gradient difference vectors to perform iteration, wherein the determination of the step length requires the use of one-dimensional search, and a total of 2*m+1 N-dimensional vectors need to be stored. When m is 6, the data to be stored is much smaller than the Hessian matrix.

[0085] The specific process of generating a mechanical fault prediction neural network model is as follows: based on the error between the mechanical fault historical database instance and the result obtained by the neural network instance model, the back propagation algorithm updates the weight value of the neural network parameter accordingly, and modifies the neural network according to the weight value. The model prediction value approaches the target value in continuous modification until a certain accuracy or training times are reached.

[0086] Step 7, use the completed BP neural network model to batch train the fault related data in the real-time database, and finally achieve the result of accurate prediction.

[0087] According to the prediction result, it is judged whether the monitoring parameter is within the safe range. If it is higher than the safety warning line, it is judged that the corresponding parameter index of the twin-screw oil pump unit is in a fault state, and an alarm is given.

[0088] According to the initial operation record and historical data accumulation of the device, the alarm mode and alarm value of each unit are established. In this way, the system can automatically check and judge the unit operation condition from a large amount of collected data, and the unit with abnormality or defect will automatically give an alarm information. The multi-dimensional alarm function includes amplitude alarm, statistical alarm and other alarm modes and alarm value setting methods, so that the device management personnel can quickly and accurately extract the device abnormality or fault information from a large number of devices, and distinguish whether the machine has a problem. The multi-clock format report function provides a variety of standard report formats and allows custom reports. Through the analysis of the neural network and the comparison with the database data, the corresponding fault prompt is given, and the host computer can timely understand the state monitoring of the device, understand the current operation condition of the device, predict the state change trend, diagnose the fault reason, test the maintenance effect, and actively prompt the fault on the host computer, so as to win the maintenance time for preventive maintenance.

[0089] By adopting the prediction method introduced in the present application, the fault of the double-screw oil pump unit can be judged in real time and accurately, and has high generalization ability, and can be widely applied to the state monitoring and fault diagnosis of the double-screw oil pump. Therefore, according to the detection of the double-screw oil pump fault, auxiliary decision of enterprise production can be made, so that the production accident rate of the double-screw oil pump unit can be greatly reduced, the economic efficiency of enterprise production can be improved, and the life safety of enterprise workers can be ensured.

[0090] In a specific embodiment 1 of the application, the double-screw oil pump online monitoring mechanical fault real-time diagnosis method includes the following steps:

[0091] 1) Collecting real-time operation data and historical record data of the double-screw oil pump unit as training mechanical fault data, and obtaining operation state data groups of the same device at different time points and in different operation modes;

[0092] 2) Data preprocessing is performed on each parameter attribute related to mechanical faults, and the data preprocessing content includes data migration, data screening, data combination, data completion, and data denoising. The processed vibration displacement signal is subjected to fast Fourier transform to obtain fault frequency characteristics. The preprocessing method is software preprocessing, and the software uses a programming language combined with a specific module method to obtain;

[0093] 3) Different evaluation indexes are formulated for evaluating normal and fault states by combining the real-time operation data and historical record data with the operation states of different components of the double-screw oil pump unit itself structure;

[0094] 4) A health and fault state database of the double-screw oil pump unit is constructed by using the real-time operation data and historical record data, which is used to measure the health and fault states under different evaluation indexes. The database management can regularly store the diagnosis data uploaded by the offline system in the database.

[0095] 5) analyzing the preprocessed data, setting corresponding generalized data batches according to input attribute settings in the input parameters; establishing an association between the mechanical fault related data to be processed and the input format; and displaying the mechanical fault related data to be processed in the corresponding input attribute based on the association at the node position of the corresponding neural network;

[0096] 6) using the obtained preprocessed mechanical fault related data to perform supervised training on the BP neural network to obtain an initial network model after training is completed;

[0097] 7) performing first initialization using a two-loop recursion algorithm, setting b1 and b2 to 0 (b1 is a bias between an input layer and a hidden layer, b2 is a bias between the hidden layer and an output layer, b1 and b2 are bias matrices of the neural network), generating a neural network model weight matrix, obtaining initial values of the weight matrix, and aligning and confirming mechanical fault data node parameters in the target optimization parameters;

[0098] 8) using the aligned and confirmed mechanical fault data node parameters and the network model after initial training to perform incremental supervised training, using an L-BFGS algorithm to optimize weights, and replacing a previous Hessian matrix by storing a small amount of data of the previous m iterations to obtain a trained BP neural network model;

[0099] 9) using the completed BP neural network model to perform batch training on mechanical fault related data in a real-time database, and finally achieving a precise prediction result;

[0100] 10) judging whether the monitoring parameters are within a safe range according to the prediction result, and if the parameters are higher than a safety warning line, judging that corresponding parameter indicators of the double screw oil pump unit are in a fault state and performing alarm.

[0101] The technical solution further has the following features:

[0102] 2-1) data migration: introducing a mechanical fault related data source into a historical and real-time database;

[0103] 2-2) data screening: screening out data useful for data mining in the mechanical fault historical database;

[0104] 2-3) data combination: splicing and combining the screened mechanical fault related data to form a new data table;

[0105] 2-4) data completion: checking and supplementing the completed data;

[0106] 2-5) Data denoising: remove some obviously flawed mechanical failure related data;

[0107] Further features of the above technical solution are that in step 6), the specific process of obtaining the trained initial network model is:

[0108] 6-1) Divide the historical data in the mechanical failure historical database, wherein 70% of the data is divided into a training set S, and 30% of the data is divided into a test set T;

[0109] 6-2) Set the input parameter S_InputDate of the training set, and use the preprocessed mechanical failure related parameters in 2) as the data for neural network training;

[0110] 6-3) Set the output standard value OutputDate of the training set, and the standard value is the corresponding mechanical failure of the historical data;

[0111] 6-4) Establish a BP neural network, set the number of nodes of the input layer, the hidden layer and the output layer, set the transfer function, and use the L-BFGS algorithm to perform unconstrained nonlinear optimization on the model.

[0112] Further features of the above technical solution are that in step 8), the first layer transfer function used in the mechanical failure prediction neural network instance model design is the tangent S-shaped function tansig(n), and the second layer transfer function is logsig.

[0113] Further features of the above technical solution are that in step 8), the specific process of generating the mechanical failure prediction neural network model is:

[0114] Based on the error between the results obtained from the mechanical failure historical database instance and the neural network instance model, the back propagation algorithm updates the weight value of the neural network parameters accordingly, and modifies the neural network according to the weight value. The model prediction value approaches the target value in continuous modification until a certain accuracy or training times are reached.

[0115] Further features of the above technical solution are that in step 9), the result of real-time mechanical failure prediction is:

[0116] The prediction method introduced in the present application can accurately judge the real-time failure of the double screw oil pump, has high generalization ability, and can assist in decision-making for enterprise production according to the detection of the double screw oil pump failure, so as to greatly reduce the production accident rate of the double screw oil pump, improve the economic efficiency of enterprise production, and protect the life safety of enterprise workers.

[0117] In a specific embodiment 2 of the application,Figure 1 is the data collection and processing block diagram in the technical scheme of the present application. The collected data parameters of the double screw oil pump mainly include 9 parameters such as transverse displacement, longitudinal displacement, main shaft speed, load, temperature, pressure range, power, dynamic current and running efficiency, including vibration parameters, process parameters and electrical parameters. The data processing content includes data migration, data screening, data combination, data completion and data denoising. The process is to overcome the limitations of the current data mining tools, and to enhance the process of clean data selected, that is, to solve the problems of missing values, redundancy, inconsistency of data, inconsistency of data definition, outdated data and other problems, and also includes the arrangement and merging of all data, so as to maintain the integrity and correctness of the data.

[0118] Figure 2 is the artificial neural network structure diagram. The original data in the historical database is divided into two groups, 70% of which is used as learning samples and 30% as test samples. In order to eliminate the mutual influence between input data, the numpy and matplotlib modules in python programming are called to implement normalization processing on learning samples and test samples, and 9 input parameters are converted between [-1, 1]. The mathematical formula of the normalization function is as follows:

[0119]

[0120] In the formula, X i — the processed value;

[0121] x i — the input data before processing;

[0122] x imin — represents the minimum value before processing;

[0123] x imax — represents the maximum value before processing.

[0124] After data processing, learning training is carried out according to the corresponding expected output. The transfer function of the hidden layer is selected as tansig(n), and the transfer function of the output layer is selected as logsig. The mathematical formula is as follows:

[0125]

[0126]

[0127] The total step is set to 1000, and the error accuracy is E=0.001.

[0128] Figure 3is the L-BFGS algorithm optimization flow chart, the L-BFGS method is one of the most advanced data fast optimization methods at present, the method is widely used to solve the unconstrained nonlinear programming problem, and has the advantages of fast convergence speed, less memory overhead, and the combination with neural network can achieve the effect of complementation.The method is limited optimization based on the BFGS method, and the basic idea is: by storing the vector sequence in the calculation process to replace the matrix H k , that is, when the matrix H k needs to be used, the calculation of the vector sequence is used to replace it without affecting the calculation precision, so that the occupied space size is reduced from O (n 2 ) to O (m*n).The specific process of optimizing the neural network by the L-BFGS algorithm is introduced in Figure 4 .

[0129] In a specific embodiment 3 of the application, the double screw oil pump online monitoring mechanical fault real-time diagnosis method comprises the following steps:

[0130] 1) Obtain the mechanical fault related data to be trained, set the input attributes corresponding to the data to be trained, first combine the data of the same device in different modes, and use clustering classification method for parameter classification;

[0131] 2) Data preprocessing is performed on each parameter attribute related to mechanical fault, and the data processing content includes data migration, data screening, data combination, data completion and data denoising, and the preprocessing method is software preprocessing, and the software uses the method of programming language combined with specific module to obtain.

[0132] 3) Analyze the data obtained in step 2), set the corresponding generalized data batch according to the input attribute in the input parameter; establish the association relationship between the mechanical fault related data to be processed and the input format; display the mechanical fault related data to be processed with the corresponding input attribute at the node position corresponding to the neural network based on the association relationship.

[0133] 4) Supervised training of BP neural network is performed by using the processed mechanical fault related data obtained in step 2), and an initial network model is obtained after training;

[0134] 5) The two-loop recursion algorithm is used for the first initialization, and b1 and b2 are 0 (b1 is the bias between the input layer and the hidden layer, b2 is the bias between the hidden layer and the output layer, b1 and b2 are the bias matrix of the neural network), then some methods are used to generate the neural network model weight matrix, the initial value of the weight matrix is obtained, and the mechanical fault data node parameters in the target optimization parameters are aligned;

[0135] 6) Incremental supervised training is performed using the aligned mechanical failure data node parameters and the network model after initial training, and the L-BFGS algorithm is used to optimize the weights, to obtain the trained BP neural network model;

[0136] 7) The completed BP neural network model is used to batch train the double screw oil pump failure related data in the real-time database, and finally the precise prediction result is achieved.

[0137] The further features of the above technical solution are that in step 1), the decomposition process using the clustering classification method is as follows:

[0138] 1-1) According to the multiple measurement indexes of mechanical failure data, some statistical quantities that can measure the similarity between parameters or indexes are found out, and these statistical quantities are used as the basis for dividing types.

[0139] 1-2) Some parameters with large similarity are aggregated into a class, and some other parameters with large similarity between each other are aggregated into another class, until all parameters are aggregated.

[0140] Through this idea, the double screw oil pump failure related data parameters can be divided into vibration class parameters, process class parameters and electrical class parameters, etc.

[0141] The further features of the above technical solution are that in step 2), the software data preprocessing mainly audits from two aspects of integrity and accuracy, and the specific process is as follows:

[0142] 2-1) The integrity audit mainly checks whether the parameters or individual data to be investigated are missing, and whether all parameter indexes are filled in completely.

[0143] 2-2) The accuracy audit mainly checks two aspects:

[0144] First, check whether the data information truly reflects the objective actual situation, and whether the content is consistent with the actual situation;

[0145] Second, check whether the data has errors, and whether the calculation is correct, etc. The methods for auditing data accuracy mainly include logical check and calculation check.

[0146] Logical check mainly checks whether the data is consistent with logic, whether the content is reasonable, and whether there are any contradictions between items or numbers, which is mainly suitable for auditing qualitative (quality) data. Calculation check is to check whether there are errors in the measurement results and calculation methods of each item of the questionnaire, which is mainly used for auditing quantitative (numerical) data.

[0147] The further features of the above technical solution are that in the step 3), the to-be-processed mechanical fault related data parameters are displayed in the corresponding input attributes at the node positions of the neural network based on the association relationship, and the specific process is as follows:

[0148] 3-1) Determine the input parameters according to the selected BP neural network prediction mechanical fault characteristic quantity, including 9 parameters such as lateral displacement, longitudinal displacement, main shaft speed, load, temperature, pressure range, power, dynamic current, and operation efficiency, representing 9 neuron nodes;

[0149] 3-2) Set 1 output neuron node, representing an output neuron, i.e., a fault prediction result.

[0150] The further features of the above technical solution are that in the step 5), the neural network model weight matrix is generated by using some methods, and there are three specific methods:

[0151] 5-1) Randomly generate weights and thresholds based on the Nguyen-Widrow algorithm, and set the thresholds b1 and b2 of the hidden layer and the output layer to 0;

[0152] 5-2) Set all the weight parameters lw from the hidden layer to the output layer to 1;

[0153] 5-3) Randomly generate the weight parameters 1w from the hidden layer to the output layer by using the random module of python:

[0154] The further features of the above technical solution are that in the step 6), the process of optimizing the weight matrix by using the L-BFGS algorithm is as follows:

[0155] 6-1) Define the initial point and the allowed error, store the latest iteration number m=6, and let k=0, H0=I, where H0 is temporarily set as the unit matrix, and f(x0) is initially set as the hyperbolic tangent function;

[0156] 6-2) If is established, return the optimal solution x k+1 , otherwise go to the next step;

[0157] 6-3) Determine the iteration direction pk and the step size a k , and start calculation by using the following formula:

[0158]

[0159] 6-4) Update the weight x, and keep the m times of vectors when k>m:

[0160] x k+1 = x k + a k pk

[0161] 6-5) Calculate the following formula and save:

[0162] s k = x k+1 - x k

[0163]

[0164] 6-6) Use two-loop recursion algorithm to get: After the superposition number, go to step (3).

[0165] Further features of the above technical solution are that in step 7), the result of real-time prediction of the double-screw oil pump failure is:

[0166] By using the prediction method introduced in the present application, the double-screw oil pump failure can be judged in real time and accurately, and has high generalization ability, so as to assist the enterprise production according to the detection of mechanical failure, thus the production accident rate of the double-screw oil pump mechanical equipment can be greatly reduced, the economic efficiency of enterprise production is improved, and the life safety of enterprise workers is ensured.

[0167] The present application is not limited to the above-mentioned embodiments. The above description of the specific embodiments is intended to describe and illustrate the technical solutions of the present application, and the above-mentioned embodiments are only illustrative and not restrictive. Without departing from the basic idea and purpose of the present application, many other specific transformations of other ways inspired by the present application all belong to the protection scope of the present application.

[0168] In addition, it should be noted that each specific technical feature described in the above specific embodiments can be combined and replaced by any suitable means without contradiction. In order to avoid unnecessary repetition, the present application will not further describe various possible combinations and replacements.

Claims

1. A method for real-time diagnosis of mechanical faults in on-line monitoring of twin screw oil transfer pumps, characterized in that, The real-time diagnosis method for mechanical fault of the twin-screw oil pump online monitoring comprises: Step 1, obtaining the twin-screw oil pump fault data to be trained; Step 2, data preprocessing is performed on the parameters related to the fault; Step 3, an association between the fault-related data to be processed and an input format is established, and the fault-related data to be processed is displayed in the node position of the neural network in the corresponding input attribute based on the association; Step 4, the obtained preprocessed mechanical fault-related data is used to supervise the training of the BP neural network, and an initial network model trained is obtained; Step 5, the fault data node parameters in the target optimization parameters are aligned and confirmed; Step 6, the L-BFGS algorithm is used to optimize the weights, and then the supervision training is performed, and a BP neural network model trained is obtained; Step 7, the completed BP neural network model is used to batch train the fault-related data in the real-time database, and finally the precise prediction result is achieved; In step 3, different evaluation indexes are formulated for evaluating the normal and fault states by combining the real-time running data and the historical record data with the running states of different components of the twin-screw oil pump unit; The health and fault state database of the twin-screw oil pump unit is constructed by using the real-time running data and the historical record data, which is used to measure the health and fault states under different evaluation indexes, and the diagnosis data uploaded by the offline system is regularly stored in and kept in the database; The preprocessed data is analyzed, the corresponding generalized data batch is set according to the input attribute in the input parameter, the association between the mechanical fault-related data to be processed and the input format is established, and the mechanical fault-related data to be processed is displayed in the node position of the neural network in the corresponding input attribute based on the association; The empirical formula for determining the number of hidden layer nodes is where k is the number of hidden layer nodes, m and n represent the number of output layer and input layer nodes respectively, and a = 1-10; the first layer transfer function used in the design of the mechanical fault prediction neural network example model is the tangent S-shaped function tansig(n), and the second layer transfer function is logsig. In step 5, the two-loop recursion algorithm is used for the first initialization, b1 and b2 are set to 0, b1 is the bias between the input layer and the hidden layer, b2 is the bias between the hidden layer and the output layer, b1 and b2 are the bias matrices of the neural network, the neural network model weight matrix is generated, the initial value of the weight matrix is obtained, and the mechanical fault data node parameters in the target optimization parameters are aligned and confirmed.

2. The method according to claim 1, wherein the method is characterized by, In step 1, the real-time running data and the historical record data of the twin-screw oil pump unit are collected as the mechanical fault data for training, and the running state data groups of the same equipment at different time points and in different running modes are obtained.

3. The method according to claim 2, wherein the method is characterized by, In step 1, the real-time running data collected includes 9 mechanical running-related parameters, such as transverse displacement, longitudinal displacement, main shaft speed, load, temperature, pressure range, power, dynamic current and running efficiency; the historical database includes vibration state, cumulative running time, lubrication state and motor running state; the measurement point database is established according to the measurement definition, collection definition and diagnosis parameter definition of the unit, including machine number, motor speed, bearing model, machine speed, gear ratio of the driving gear and the driven gear.

4. The method according to claim 3, wherein the method is characterized by, In step 1, the real-time operation data and historical record data of the twin-screw oil pump unit are collected as the training mechanical fault data, and the input attributes corresponding to the data to be trained are set. First, the data of the same device in different modes are combined, and the parameter classification is performed by using the clustering classification method.

5. The method according to claim 4, wherein the method is characterized by, In step 1, different fault evaluation indexes are formulated according to different types of data such as vibration parameters, process parameters and electrical parameters, including vibration evaluation indexes, flow evaluation indexes, temperature evaluation indexes, dynamic current evaluation indexes, running time evaluation indexes, lubrication evaluation indexes and motor state evaluation indexes. The equipment management personnel draw the diagnosis conclusion of the machine fault, including unbalance, misalignment, rubbing, looseness, resonance and bearing fault.

6. The method according to claim 1, wherein the method is characterized by, In step 2, the data preprocessing is performed on each parameter attribute related to the twin-screw oil pump fault, and the data preprocessing content includes data migration, data screening, data combination, data completion and data denoising. The processed vibration displacement signal is subjected to fast Fourier transform to obtain the fault frequency characteristics.

7. The method according to claim 6, wherein, In step 2, the data processing content includes data migration, data screening, data combination, data completion and data denoising. The selected clean data is subjected to enhancement processing, i.e. solving the problems of missing values, redundancy, inconsistency of data, inconsistency of data definition and outdated data. The data is also sorted and merged to maintain the integrity and correctness of the data. The original signal of the machine is obtained by the sensor, and the FFT technology is used for preprocessing the original vibration signal, which is used as the method for obtaining the original input data of the neural network. The selection of the mechanical fault related parameters adopts the correlation coefficient method. The calculation formula is: Wherein r xy represents the sample correlation coefficient, S xy represents the sample covariance, S x represents the sample standard deviation of X, S y represents the sample standard deviation of Y.

8. The method according to claim 1, wherein the method is characterized by, In step 6, the aligned mechanical fault data node parameters and the network model subjected to initial training are subjected to incremental supervised training, and the L-BFGS algorithm is used to optimize the weight. The last m iterations of a small amount of data are stored to replace the previous Hessian matrix, and the trained BP neural network model is obtained.

9. The method according to claim 8, wherein the method is characterized by, In step 6, the L-BFGS algorithm is repeatedly executed by step iteration. The calculation formula is as follows: Let Let Available, Taking the first m terms, we have In the formula s---step length; y---gradient difference; H---Hessian matrix; v, p---quantities defined for convenient operation; Since the variables p, v, s and y can be finally calculated from the step length and the gradient difference vectors, only the last m step length and gradient difference vectors need to be stored for iteration. The determination of the step length requires the use of one-dimensional search, and a total of 2*m+1 N-dimensional vectors need to be stored. Taking m as 6, the data to be stored is much smaller than the Hessian matrix.

10. The method according to claim 8, wherein the method is characterized by, In step 6, the specific process of generating the mechanical fault prediction neural network model is as follows: based on the error between the results obtained from the mechanical fault historical database instance and the neural network instance model, the back propagation algorithm updates the weight value of the neural network parameters, and modifies the neural network according to the weight value. The model prediction value approaches the target value in continuous modification until a certain accuracy or training times are reached.

11. The method for online monitoring and real-time diagnosis of mechanical faults of a twin-screw oil pump according to claim 1, characterized in that: In step 7, according to the prediction result, it is judged whether the monitoring parameter is in the safety range, if higher than the safety warning line, it is judged that the corresponding parameter index of the twin-screw oil pump unit is in a fault state, and an alarm is given.

12. The method according to claim 11, wherein the method is characterized by, In step 7, the alarm mode and alarm value of each unit are established according to the initial operation record and historical data accumulation of the equipment; in this way, the system can automatically check and judge the operation condition of the unit from a large amount of collected data, and the unit with abnormalities or defects will automatically give an alarm information; the multi-dimensional alarm function including amplitude alarm, statistical alarm, various alarm modes and alarm value setting methods enables the equipment management personnel to quickly and accurately extract the equipment abnormality or fault information from numerous equipment, and distinguish whether the machine has a problem; the multi-clock format report function provides various standard report formats and allows custom report; through the analysis of the neural network and the comparison with the database data, the corresponding fault prompt is given, the state monitoring of the equipment is understood in time through the upper computer, the current operation condition of the equipment is understood, the state change trend is predicted, the cause of the fault is diagnosed, the effect of inspection and maintenance is tested, and the fault is prompted on the upper computer, so that the maintenance time for preventive maintenance is won.

Images