Cementing pump fault detection method based on contrastive self-supervised random configuration network

Abstract

The application discloses a cementing pump fault detection method based on a contrast self-supervised random configuration network, belongs to the technical field of cementing equipment safety monitoring and fault diagnosis, and is used for cementing pump fault detection and comprises the following steps: collecting normal working condition data of the cementing pump, wherein the normal working condition data comprises vibration signals and rotating speed signals, and the vibration signals are divided into data windows according to strokes; calculating time domain statistical features and frequency domain statistical features of the vibration signals in each data window to form a statistical feature vector; establishing a contrast self-supervised random configuration network model; and constructing a cementing pump working condition monitoring index. The application realizes a monitoring model by establishing a fault detection framework based on self-supervised learning, improves an existing random configuration network optimization objective function by using a contrast learning theory, and improves fault detection sensitivity.

Description

TECHNICAL FIELD

[0001] The application discloses a cementing pump fault detection method based on a contrast self-supervised stochastic configuration network and belongs to the technical field of cementing equipment safety monitoring and fault diagnosis. BACKGROUND

[0002] Cementing is an important link in the oil well exploitation process of an oil field. As the "heart" of a cementing construction device, a cementing pump may cause the cementing operation of the entire oil well to fail, result in expensive non-production time and economic losses, and even cause serious safety accidents such as well control out of control if a fault occurs in the high-pressure and continuous operation process. The traditional fault judgment mode depending on manual experience has hysteresis and subjectivity and is difficult to timely warn in the fault germination period. An automatic fault detection technology system can realize the transformation from "after-maintenance" to "pre-warning" by monitoring the operating parameters (such as vibration signals) of the cementing pump in real time and using a machine learning method for abnormal detection. This can not only minimize unexpected downtime and equipment damage but also ensure the safety of operating personnel and lay a solid foundation for the long-term stable production of oil and gas wells.

[0003] Typical machine learning methods in the field of industrial process fault detection include principal component analysis (PCA) and kernel principal component analysis (KPCA). Although these traditional methods are applied more frequently, their performance is difficult to satisfy in some complex data analysis scenarios. In recent years, as a new type of intelligent machine learning model, a stochastic configuration network (SCN) has advantages such as provable model convergence, fast training speed, and strong nonlinear feature learning capability, which provides a new means for modeling the operating data of a cementing pump. However, the cementing pump fault detection technology based on the stochastic configuration network faces two challenging problems. The traditional stochastic configuration network is a supervised learning model and needs multi-class fault data for model training. In actual industrial production processes, fault data are relatively few, and therefore how to establish a cementing pump operating condition monitoring model based on the stochastic configuration network only by using normal operating condition data is a key problem. In the early stage of a cementing pump fault, the inherent fault features of the vibration signal are weak and have low discrimination from normal data, and therefore how to improve the stochastic configuration network model to improve the fault detection sensitivity is another problem worthy of research. SUMMARY

[0004] The application aims to provide a cementing pump fault detection method based on a contrast self-supervised stochastic configuration network to solve the problems in the prior art that the traditional stochastic configuration network model depends on multi-class fault data and has a low fault detection rate.

[0005] The cementing pump fault detection method based on the contrast self-supervised stochastic configuration network comprises the following steps.

[0006] S1, collect normal working condition data of the cementing pump, the normal working condition data including vibration signals and rotation speed signals, the vibration signals are divided into data windows according to strokes; time domain statistical features and frequency domain statistical features of the vibration signals in each data window are calculated to form a statistical feature vector;

[0007] S2, taking the statistical feature vector as input, a contrastive self-supervised random configuration network is constructed, a method combining random configuration and contrastive learning is adopted for neural network training, an optimization objective function of fusion reconstruction error and contrastive loss is used to constrain the random configuration process of the hidden layer nodes and determine the weight calculation of the output layer, after the training is completed, the contrastive self-supervised random configuration network model is output;

[0008] S3, according to the reconstruction error and the hidden layer feature of the contrastive self-supervised random configuration network model, a reconstruction error monitoring index and a hidden layer feature monitoring index are constructed, the method of kernel density estimation is used to calculate the reconstruction error monitoring index threshold and the hidden layer feature monitoring index threshold based on the normal working condition data, and the reconstruction error monitoring index threshold and the hidden layer feature monitoring index threshold are used to judge the occurrence of the cementing pump fault.

[0009] S1 includes S1.1, the vibration signal is a pump body vibration signal of a hydraulic end, and the rotation speed signal is a rotation speed signal of a crankshaft of the hydraulic end;

[0010] The pump body vibration signal of the hydraulic end of the cementing pump under the normal running condition is collected to form a vibration signal amplitude set The rotation speed signal of the crankshaft of the hydraulic end of the cementing pump under the normal running condition is collected to form a rotation speed signal set , is the total number of global sampling points.

[0011] S1 includes S1.2, the vibration signal is divided into data windows according to strokes, is the data window index, each data window is divided into sampling points , is time, is the starting time of the i-th data window, is the total number of vibration signal sampling points of the i-th data window, and is the sampling serial number of the right dead point of the plunger movement of the cementing pump, is the ending time of the i-th data window is:

[0012] ;​​

[0013] In the formula, for The rotational speed value at any given time. The data sampling frequency is given, with the rotational speed measured in revolutions per minute (rpm) and the sampling frequency measured in Hertz (Hz).

[0014] set up For sampling point index, , For the first The first data window The amplitude of the vibration signal at each sampling point.

[0015] S1 includes, S1.3, and... Calculate vibration signal The statistical feature vector consists of time-domain statistical features and frequency-domain statistical features. , .

[0016] S2 includes a contrastive self-supervised randomized configuration network framework, which includes a randomized configuration network, an input contrastive learning sample generation module, an expected output module, a hidden layer node constraint module, a model optimization objective function module, and a prediction error module.

[0017] Will Input the random configuration network, the input contrastive learning sample generation module, and the expected output module respectively;

[0018] The randomly configured network consists of an input layer, hidden layers, and an output layer. The number of nodes in the input and output layers is consistent with the dimension of the statistical feature vector, while the number of hidden layer nodes grows dynamically during the construction process. The input layer is mapped to the hidden layer via input parameters, and then mapped to the output layer via output parameters to produce the reconstructed output. ;

[0019] The learning sample generation module receives Positive and negative samples are generated, and the generated positive and negative samples are respectively transmitted to the hidden layer node constraint module and the model optimization objective function module.

[0020] Expected output module to receive and output itself;

[0021] The prediction error module will output the expected output of the module. and the output of the randomly configured network Perform subtraction to calculate Prediction error ,Will Feedback is sent to the hidden layer node constraint module and the model optimization objective function module;

[0022] The hidden layer node constraint condition module receives positive samples and negative samples, establishes a constraint condition to constrain randomly generated input parameters, and transmits input parameters meeting the constraint condition into the random configuration network;

[0023] The model optimization objective function module receives positive samples and negative samples, constructs a network optimization objective function based on contrast learning, directly calculates the final output parameters by an analytical method, and transmits the output parameters into the random configuration network.

[0024] In the random configuration network, the network model expression with hidden layer nodes is:

[0025] ;

[0026] In the formula, y is the corresponding network model output, is an activation function; is a weight, is a bias, is a weight vector connecting the input layer to the i-th hidden layer node, is a bias term of the i-th hidden layer node, and constitute input parameters; is a weight vector, is an output weight vector connecting the i-th hidden layer node to all nodes of the output layer, which constitutes output parameters alone; is a transpose symbol, is the number of hidden layers, is the index of the hidden layer, ; In the prediction error module, the prediction error formula of y is: ;

[0027] The overall prediction error matrix is:

[0028] ;

[0029] The positive sample of y is defined as the nearest neighbor in the model input sample set:

[0030] ;

[0031] The negative sample of y is defined as the farthest neighbor in the model input sample set:

[0032] ;​​​​

[0033] wherein, to find the value of the input variable that makes the function attain a minimum value, is a norm;

[0034] negative samples of are:

[0035] ;

[0036] || ||;

[0037] wherein, is the center point of the sample, is the distance from to , is a negative sample distance parameter, , is the set of .

[0038] In the hidden layer node constraint condition module, the input parameters of the first hidden layer node of the network are configured randomly , by the hidden layer node constraint condition, and the constraint condition expression is:

[0039] ;

[0040] wherein, is the column index of , , is the first column of , is the first hidden layer node vector, is the corresponding hidden layer node vector, is the corresponding hidden layer node vector, is the corresponding hidden layer node vector, is a regularization parameter;

[0041] , and expressions are:

[0042] ;

[0043] ;

[0044] ;​

[0045] ;

[0046] ;

[0047] ;

[0048] wherein, is the i-th hidden layer node vector of the j-th sample, is the i-th hidden layer node vector of the j-th sample, is the output value of the i-th hidden layer node to the j-th positive sample, is the output value of the i-th hidden layer node to the j-th negative sample, is the weight vector connecting the input layer to the i-th hidden layer node, is the bias term of the i-th hidden layer node; In the model optimization objective function module, a network optimization objective function based on contrastive learning is established : ; wherein,

[0049] is the input sample matrix, , is the hidden layer node matrix,

[0050] , is the hidden layer node matrix corresponding to all positive samples, is the hidden layer node matrix corresponding to all negative samples,

[0051] is the weight parameter matrix, , is the Frobenius norm; Solving the network optimization objective function obtains the optimal weight parameter matrix , is the solution that makes the objective function minimum, . Based on

[0052] obtains the model expression containing hidden layer nodes: ;

[0053] Based on obtains the model expression containing hidden layer nodes:

[0054] ;

[0055] wherein, is the input feature vector,​​​ The final output of the contrast self-supervised random configuration network model is compared.

[0056] S3 includes S3.1, constructing a reconstruction error monitoring index , constructing a hidden layer feature monitoring index , for , and The expression is:

[0057] ;

[0058] ;

[0059] In the formula, is the hidden layer feature vector of , , is the average value of , is the covariance matrix of , is the covariance matrix of ;

[0060] Based on normal working condition data, the kernel density estimation method is used to calculate the reconstruction error monitoring index threshold and the hidden layer feature monitoring index threshold .

[0061] S3 includes S3.2, after the data to be detected is processed by steps S1 to S2, the statistical feature vector of the data to be detected is calculated , and The input of the contrast self-supervised random configuration network model outputs the reconstruction error monitoring index and the hidden layer feature monitoring index of the data to be detected;

[0062] If or , it is determined that the cementing pump state corresponding to the data to be detected is faulty.

[0063] Compared with the prior art, the present application has the following beneficial effects: the present application realizes a monitoring model by establishing a fault detection framework based on self-supervised learning, improves the existing random configuration network optimization objective function using the contrast learning theory, and improves the fault detection sensitivity. BRIEF DESCRIPTION OF DRAWINGS

[0064] Figure 1 is the technical flowchart of the present application;

[0065] Figure 2 is the contrast self-supervised random configuration network framework diagram;

[0066] Figure 3 is a principal component analysis method hidden layer feature monitoring index analysis chart;

[0067] Figure 4 is a principal component analysis method reconstruction error monitoring index analysis chart;

[0068] Figure 5 is a kernel principal component analysis method hidden layer feature monitoring index analysis chart;

[0069] Figure 6 is a kernel principal component analysis method reconstruction error monitoring index analysis chart;

[0070] Figure 7 is a self-supervised stochastic configuration network hidden layer feature monitoring index analysis chart;

[0071] Figure 8 is a self-supervised stochastic configuration network reconstruction error monitoring index analysis chart;

[0072] Figure 9 is a contrastive self-supervised stochastic configuration network hidden layer feature monitoring index analysis chart;

[0073] Figure 10 is a contrastive self-supervised stochastic configuration network reconstruction error monitoring index analysis chart. DETAILED DESCRIPTION

[0074] In order to make the objects, technical solutions and advantages of the present application clearer, the technical solutions in the present application will be described clearly and completely below. Obviously, the described embodiments are part of the embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor fall within the scope of protection of the present application.

[0075] The cementing pump fault detection method based on the contrastive self-supervised stochastic configuration network comprises the following steps:

[0076] S1, collecting normal working condition data of the cementing pump, the normal working condition data comprising vibration signals and rotation speed signals, dividing the vibration signals into data windows according to the stroke; calculating time domain statistical features and frequency domain statistical features of the vibration signals in each data window to form a statistical feature vector;

[0077] S2, taking the statistical feature vector as input, constructing a contrastive self-supervised stochastic configuration network, adopting a method combining random configuration and contrastive learning to train the neural network, constraining the random configuration process of the hidden layer nodes through an optimization objective function fusing reconstruction error and contrastive loss, and determining the weight calculation of the output layer, after the training is completed, outputting a contrastive self-supervised stochastic configuration network model;

[0078] S3. Based on the reconstruction error and hidden layer characteristics of the self-supervised random configuration network model, construct the reconstruction error monitoring index and the hidden layer characteristic monitoring index. Using the kernel density estimation method, calculate the threshold of the reconstruction error monitoring index and the threshold of the hidden layer characteristic monitoring index based on normal operating data. Use the threshold of the reconstruction error monitoring index and the threshold of the hidden layer characteristic monitoring index to determine the occurrence of cementing pump failure.

[0079] S1 includes, S1.1, the vibration signal is the vibration signal of the hydraulic end pump body, and the rotation speed signal is the rotation speed signal of the hydraulic end crankshaft;

[0080] Vibration signals of the hydraulic end pump body under normal operating conditions of the cementing pump are collected to form a set of vibration signal amplitudes. The rotational speed signal of the hydraulic end crankshaft under normal operating conditions of the cementing pump is collected to form a rotational speed signal set. , This represents the total number of global sampling points.

[0081] S1 includes, S1.2, dividing the vibration signal according to the stroke. Data window ,set up Indexing the data window, Based on the rotational speed signal, each data window Divided into One sampling point, , For time, For the first The start time of each data window For the first The total number of vibration signal sampling points for each data window, let's assume... The sampling sequence number of the cementing pump plunger when it reaches the right dead center, number 1. The end time of each data window for:

[0082] ;

[0083] In the formula, for The rotational speed value at any given time. The data sampling frequency is given, with the rotational speed measured in revolutions per minute (rpm) and the sampling frequency measured in Hertz (Hz).

[0084] set up For sampling point index, , For the first The first data window The amplitude of the vibration signal at each sampling point.

[0085] S1 includes, S1.3, and... Calculate vibration signal The statistical feature vector consists of time-domain statistical features and frequency-domain statistical features. , .

[0086] S2 includes a contrastive self-supervised randomized configuration network framework, which includes a randomized configuration network, an input contrastive learning sample generation module, an expected output module, a hidden layer node constraint module, a model optimization objective function module, and a prediction error module.

[0087] Will Input the random configuration network, the input contrastive learning sample generation module, and the expected output module respectively;

[0088] The randomly configured network consists of an input layer, hidden layers, and an output layer. The number of nodes in the input and output layers is consistent with the dimension of the statistical feature vector, while the number of hidden layer nodes grows dynamically during the construction process. The input layer is mapped to the hidden layer via input parameters, and then mapped to the output layer via output parameters to produce the reconstructed output. ;

[0089] The learning sample generation module receives Positive and negative samples are generated, and the generated positive and negative samples are respectively transmitted to the hidden layer node constraint module and the model optimization objective function module.

[0090] Expected output module to receive and output itself;

[0091] The prediction error module will output the expected output of the module. and the output of the randomly configured network Perform subtraction to calculate Prediction error ,Will Feedback is sent to the hidden layer node constraint module and the model optimization objective function module;

[0092] Hidden layer node constraint module receives Positive and negative samples are used to establish constraints on randomly generated input parameters, and input parameters that satisfy the constraints are passed into a random configuration network.

[0093] Model optimization objective function module receives Positive and negative samples are used to construct a network optimization objective function based on contrastive learning. The final output parameters are directly calculated using analytical methods and then fed into a randomly configured network.

[0094] In a randomly configured network, having The network model expression for hidden layer nodes is:

[0095] ;

[0096] wherein, is the corresponding network model output, is an activation function; is a weight, is a bias, is a weight vector connecting the input layer to the th hidden layer node, is a bias term of the th hidden layer node, and constitute input parameters; is a weight vector, is an output weight vector connecting the th hidden layer node to all nodes of the output layer, constitute output parameters individually; is a transpose symbol, is the number of hidden layers, is the index of the hidden layer, ;

[0097] In the prediction error module, the prediction error formula is:

[0098] ;

[0099] The overall prediction error matrix is:

[0100] ;

[0101] The positive sample of is defined as the nearest neighbor in the model input sample set:

[0102] ;

[0103] wherein, is the value of the input variable that finds the minimum of the function, is a norm;

[0104] The negative sample of is:

[0105] ;

[0106] || ||;

[0107] wherein, As the center point of the sample, for arrive distance, For negative sample distance parameters, , for A set of.

[0108] In the hidden layer node constraint module, the first random configuration network Input parameters of each hidden node , The constraints are configured by the hidden node constraints, and the constraint expressions are as follows:

[0109] ;

[0110] In the formula, for column index, , for The List, For the first A vector of hidden layer nodes for The corresponding hidden layer node vector, for The corresponding hidden layer node vector, For regularization parameters;

[0111] , and The expression is:

[0112] ;

[0113] ;

[0114] ;

[0115] ;

[0116] ;

[0117] ;

[0118] In the formula, For the first The first sample A vector of hidden layer nodes For the first The hidden layer node pairs with the first The output value of each positive sample. No. output value of the i-th hidden layer node for the j-th negative sample, is a weight vector connecting the input layer to the i-th hidden layer node, is a bias term of the i-th hidden layer node;

[0119] In the model optimization objective function module, a network optimization objective function based on contrastive learning is established :

[0120] ;

[0121] wherein, is an input sample matrix, , is a hidden layer node matrix, , is a hidden layer node matrix corresponding to all positive samples, is a hidden layer node matrix corresponding to all negative samples, is a weight parameter matrix, , is a Frobenius norm;

[0122] The network optimization objective function is solved to obtain an optimal weight parameter matrix , is a solution that makes the objective function minimum, .

[0123] Based on , a model expression containing i hidden layer nodes is obtained:

[0124] ;

[0125] wherein, is an input feature vector, is a final output of the contrastive self-supervised random configuration network model.

[0126] S3 includes, S3.1, constructing a reconstruction error monitoring index , constructing a hidden layer feature monitoring index , for , and , the expressions are:

[0127] ;

[0128] ;

[0129] wherein,​​​​ is the average value of the hidden layer feature vector, , is the average value of the hidden layer feature vector, is the covariance matrix of the hidden layer feature vector, is the covariance matrix of the hidden layer feature vector;

[0130] Based on normal working condition data, the kernel density estimation method is used to calculate the reconstruction error monitoring index threshold and the hidden layer feature monitoring index threshold .

[0131] S3 includes, S3.2, after the data to be detected is processed through steps S1 to S2, the statistical feature vector of the data to be detected is calculated , the input comparison self-supervised random configuration network model, output the reconstruction error monitoring index of the data to be detected and the hidden layer feature monitoring index ;

[0132] If or , it is determined that the cementing pump state corresponding to the data to be detected is faulty.

[0133] Further description will be made below in combination with the drawings. The technical process of the present application is shown in Figure 1 , first, normal working condition data of the cementing pump is collected, the data window sample is divided according to the stroke, the time domain and frequency domain statistical features of each data window are calculated, then the monitoring model based on the comparison self-supervised random configuration network is established, finally the working condition monitoring index is constructed, and the monitoring index is used for fault detection.

[0134] The comparison self-supervised random configuration network framework of the present application is shown in Figure 2 , the normal working condition data ( ) is respectively input into the random configuration network, the input comparison learning sample generation module and the expected output module; the random configuration network includes an input layer, a hidden layer and an output layer, the node number of the input layer ​​​​The hidden layer nodes dynamically grow with the construction process. The self-supervised random configuration network framework comprises a random configuration network, an input contrast learning sample generation module, an expected output module, a hidden layer node constraint condition module, a model optimization objective function module and a prediction error module. Input data is mapped from an input layer to a hidden layer through input parameters, and then mapped to an output layer through output parameters to generate a reconstructed output. The learning sample generation module receives input data to generate positive samples and negative samples. The generated positive samples and negative samples are transmitted to the hidden layer node constraint condition module and the model optimization objective function module respectively. The expected output module outputs the input data itself. The prediction error module takes the output of the expected output module and the output of the random configuration network as input, performs subtraction operation, calculates the prediction error of the model, and feeds back the prediction error to the hidden layer node constraint condition module and the model optimization objective function module respectively. The hidden layer node constraint condition module receives the prediction error, the positive sample and the negative sample, evaluates the randomly generated candidate input parameters, and transmits the input parameters meeting the constraint condition to the random configuration network. The model optimization objective function module receives the prediction error, the positive sample and the negative sample, evaluates the randomly generated candidate input parameters, constructs and solves the complex optimization objective function, and directly calculates the final output parameters after the hidden layer nodes are determined, and transmits the final output parameters to the random configuration network. Among them, is a set composed of the application results of , , The components of are generated by acting on the net input at each node.

[0135] During the operation of the cementing pump, a group of vibration data and rotation speed data under normal working conditions are collected. After data preprocessing, 1600 statistical feature samples are obtained for constructing the working condition monitoring model of the cementing pump. At the same time, in order to test the performance of the fault detection algorithm, a group of fault data of the cementing pump under the condition of valve rubber wear is collected. Let , calculate 19 time domain and frequency domain statistical features of the vibration signal to form a statistical feature vector [ ] . The 19 features are maximum value, maximum absolute value, minimum value, mean value, peak-to-peak value, absolute mean value, root mean square value, root amplitude value, standard deviation, kurtosis, skewness, excess factor, Waveform factor, Pulse factor, Peak factor, Center of gravity frequency, Mean frequency, Root mean square frequency, Frequency standard deviation. After pretreatment, 2200 statistical feature samples of fault data were obtained. Rubber wear is a typical cementing pump failure, which can cause the sealing of the cementing pump valve to deteriorate, thereby causing fluid leakage.

[0136] Four methods were used to analyze the normal working condition data and establish monitoring models, including principal component analysis (PCA), kernel principal component analysis (KPCA), self-supervised stochastic configuration network (S-SCN), and comparative self-supervised stochastic configuration network (CS-SCN). Among them, principal component analysis (PCA) and kernel principal component analysis (KPCA) are classic data-driven anomaly detection methods, which are used for method comparison here.

[0137] The monitoring chart of the cementing pump valve rubber wear fault is shown in FIGS. Figure 3 、 Figure 4 、 Figure 5 、 Figure 6 、 Figure 7 、 Figure 8 、 Figure 9 and Figure 10 , the first 200 samples are normal samples, the last 2200 samples are statistical feature samples of fault data, and the dashed line represents the detection threshold. Figure 3 、 Figure 4 are the hidden feature monitoring index and reconstruction error monitoring index curves of principal component analysis, respectively, and it can be seen that the principal component analysis has poor detection effect on the 600th to 1500th samples and cannot effectively detect anomalies. Figure 5 、 Figure 6 are the hidden feature monitoring index and reconstruction error monitoring index curves of kernel principal component analysis, Figure 5 perform poorly, but Figure 6 the reconstruction error index in Figure 7 , Figure 8 are the hidden feature monitoring index and reconstruction error monitoring index curves of self-supervised stochastic configuration network. Figure 7 The hidden layer feature index of the self-supervised stochastic configuration network in Figure 8 can effectively detect the anomalies of the 600th to 1500th samples, Figure 9 the reconstruction error monitoring index in Figure 10The figure is the curve of the hidden feature monitoring index and the reconstruction error monitoring index of the comparative self-supervised random configuration network. It can be seen that the two monitoring indexes of the comparative self-supervised random configuration network can effectively detect the fault samples. The performance index analysis result is shown in Table 1.

[0138] Table 1, fault detection rate of four kinds of models

[0139] ;

[0140] It can be seen that the fault detection rates of the two monitoring indexes of the principal component analysis method and the kernel principal component analysis method model are both below 70%, the self-supervised random configuration network improves the fault detection rate of the hidden layer feature monitoring index by more than 90%, and the fault detection rates of the reconstruction error monitoring index and the hidden layer feature monitoring index of the comparative self-supervised random configuration network model are 95.95% and 91.86% respectively, which shows the best fault detection performance and fully reflects the performance advantage of the method.

[0141] The above examples are only used to illustrate the technical solutions of the present application, but not to limit it. Although the present application has been described in detail with reference to the foregoing examples, it should be understood by those skilled in the art that the technical solutions recorded in the foregoing examples can be modified, or some or all of the technical features can be replaced by equivalents, and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the scope of the technical solutions of the embodiments of the present application.

Claims

1. A cementing pump fault detection method based on a comparative self-supervised randomized configuration network, characterized in that, include: S1. Collect normal operating data of the cementing pump. Normal operating data includes vibration and speed signals. The vibration signals are divided according to the stroke. Each data window is used to calculate the time-domain and frequency-domain statistical characteristics of the vibration signal in each data window, forming a statistical feature vector. S2. Using statistical feature vectors as input, construct a contrastive self-supervised stochastic configuration network. Use a combination of stochastic configuration and contrastive learning to train the neural network. By fusing the optimization objective function of reconstruction error and contrastive loss, constrain the stochastic configuration process of hidden layer nodes and determine the weight calculation of the output layer. After training, output the contrastive self-supervised stochastic configuration network model. S3. Based on the reconstruction error and hidden layer characteristics of the self-supervised random configuration network model, construct the reconstruction error monitoring index and the hidden layer characteristic monitoring index. Using the kernel density estimation method, calculate the threshold of the reconstruction error monitoring index and the threshold of the hidden layer characteristic monitoring index based on normal operating data. Use the threshold of the reconstruction error monitoring index and the threshold of the hidden layer characteristic monitoring index to determine the occurrence of cementing pump failure. S2 includes a contrastive self-supervised randomized configuration network framework, which includes a randomized configuration network, an input contrastive learning sample generation module, an expected output module, a hidden layer node constraint module, a model optimization objective function module, and a prediction error module. Will Input the random configuration network, the input contrastive learning sample generation module, and the expected output module respectively; The randomly configured network consists of an input layer, hidden layers, and an output layer. The number of nodes in the input and output layers is consistent with the dimension of the statistical feature vector, while the number of hidden layer nodes grows dynamically during the construction process. The input layer is mapped to the hidden layer via input parameters, and then mapped to the output layer via output parameters to produce the reconstructed output. ; The learning sample generation module receives Positive and negative samples are generated, and the generated positive and negative samples are respectively transmitted to the hidden layer node constraint module and the model optimization objective function module. Expected output module to receive and output itself; The prediction error module will output the expected output of the module. and the output of the randomly configured network Perform subtraction to calculate Prediction error ,Will Feedback is sent to the hidden layer node constraint module and the model optimization objective function module; Hidden layer node constraint module receives Positive and negative samples are used to establish constraints on randomly generated input parameters, and input parameters that satisfy the constraints are passed into a random configuration network. Model optimization objective function module receives Positive and negative samples are used to construct a network optimization objective function based on contrastive learning. The final output parameters are directly calculated using analytical methods and then fed into a randomly configured network. In a randomly configured network, having The network model expression for hidden layer nodes is: ； In the formula, for The corresponding network model output, For activation functions; As weight, For bias, It connects the input layer to the first... The weight vector of each hidden layer node. For the first The bias term of each hidden layer node. and Constitutes the input parameters; For the weight vector, To connect the first The output weight vector from each hidden layer node to all nodes in the output layer. Individually constitute the output parameters; It is the transpose symbol. This represents the number of hidden layers. For the hidden layer index, ; In the prediction error module, The prediction error formula is: ； Overall prediction error matrix for: ； Positive samples Defined as the nearest neighbor in the model's input sample set: ； In the formula, To find the value of the input variable that makes the function reach its minimum value, It is a norm; negative samples for: ； || ||； In the formula, As the center point of the sample, for arrive distance, For negative sample distance parameters, , for A set; In the hidden layer node constraint module, the first random configuration network Input parameters of each hidden node , The constraints are configured by the hidden node constraints, and the constraint expressions are as follows: ； In the formula, for column index, , for The List, For the first A vector of hidden layer nodes for The corresponding hidden layer node vector, for The corresponding hidden layer node vector, For regularization parameters; , and The expression is: ； ； ； ； ； ； In the formula, For the first The first sample A vector of hidden layer nodes For the first The hidden layer node pairs with the first The output value of each positive sample No. The hidden layer node pairs with the first The output value of each negative sample. It connects the input layer to the first... The weight vector of each hidden layer node. For the first The bias term of each hidden layer node; In the model optimization objective function module, a network optimization objective function based on contrastive learning is established. : ； In the formula, For the input sample matrix, , The hidden layer node matrix, , This is the hidden layer node matrix corresponding to all positive samples. The hidden layer node matrix is ​​the matrix corresponding to all negative samples. The weight parameter matrix, , It is the Frobenius norm; Solving the network optimization objective function yields the optimal weight parameter matrix. , It is to make the objective function The solution that obtains the minimum value ; based on Get included The model expression for each hidden node: ； In the formula, For the input feature vector, To compare the final output of the self-supervised randomized network model.

2. The cementing pump fault detection method based on a comparative self-supervised random configuration network according to claim 1, characterized in that, S1 includes, S1.1, the vibration signal is the vibration signal of the hydraulic end pump body, and the rotation speed signal is the rotation speed signal of the hydraulic end crankshaft; Vibration signals of the hydraulic end pump body under normal operating conditions of the cementing pump are collected to form a set of vibration signal amplitudes. The rotational speed signal of the hydraulic end crankshaft under normal operating conditions of the cementing pump is collected to form a rotational speed signal set. , This represents the total number of global sampling points.

3. The cementing pump fault detection method based on a comparative self-supervised random configuration network according to claim 2, characterized in that, S1 includes, S1.2, dividing the vibration signal according to the stroke. Data window ,set up Indexing the data window, Based on the rotational speed signal, each data window Divided into One sampling point, , For time, For the first The start time of each data window For the first The total number of vibration signal sampling points for each data window, let's assume... The sampling sequence number of the cementing pump plunger when it reaches the right dead center, number 1. The end time of each data window for: ； In the formula, for The rotational speed value at any given time. The data sampling frequency is given, with the rotational speed measured in revolutions per minute (rpm) and the sampling frequency measured in Hertz (Hz). set up For sampling point index, , For the first The first data window The amplitude of the vibration signal at each sampling point.

4. The cementing pump fault detection method based on a comparative self-supervised random configuration network according to claim 3, characterized in that, S1 includes, S1.3, and... Calculate vibration signal The statistical feature vector consists of time-domain statistical features and frequency-domain statistical features. , .

5. The cementing pump fault detection method based on a comparative self-supervised random configuration network according to claim 4, characterized in that, S3 includes S3.1, and the construction of reconstruction error monitoring indicators. Construct hidden layer feature monitoring indicators ,for , and The expression is: ； ； In the formula, for The hidden layer feature vectors , for The average value, for The covariance matrix, for The covariance matrix; Based on normal operating condition data, the threshold of the reconstruction error monitoring index is calculated using the kernel density estimation method. and hidden layer feature monitoring index threshold .

6. The cementing pump fault detection method based on a comparative self-supervised random configuration network according to claim 5, characterized in that, S3 includes S3.2, which involves processing the data to be detected through steps S1 to S2 and then calculating the statistical feature vector of the data to be detected. ,Will Input a self-supervised randomized network model and output a reconstruction error monitoring index for the data to be detected. and hidden layer feature monitoring indicators ; like or The cementing pump status corresponding to the data to be tested is determined to be faulty.

Images