Oil drilling platform variable frequency motor end overvoltage prediction method based on IOOA-CNN-BiLSTM algorithm

Abstract

The invention relates to the technical field of motor insulation, in particular to an oil drilling platform variable frequency motor end overvoltage prediction method based on an IOOA-CNN-BiLSTM algorithm, and the method comprises the following steps: collecting voltage data of a motor; preprocessing the voltage data to obtain a voltage data set; adopting a population initialization mode, a convergence factor and a position updating formula of an improved OOA algorithm, and introducing an acceleration gradient in a development stage; optimizing the size of a convolution kernel, the size of a hidden layer and an initial learning rate of the CNN-BiLSTM model on the basis of IOOA, endowing the optimized parameters into the CNN-BiLSTM model again, and constructing an IOOA-CNN-BiLSTM prediction model; and inputting voltage data in the voltage data set into the IOOA-CNN-BiLSTM prediction model to obtain an overvoltage prediction result corresponding to the voltage data of the motor. According to the invention, through the improvement of the motor insulation algorithm, the parameters of the CNN-BiLSTM model can be optimized, and then the precision of motor insulation overvoltage prediction can be improved.

Description

Technical Field

[0001] The present application relates to the field of motor insulation technology, and in particular to a method for predicting overvoltage at the terminal of a variable frequency motor on an oil drilling platform based on an IOOA-CNN-BiLSTM algorithm. Background Art

[0002] As the power source for oil drilling operations, variable-frequency motors (VFDs) have a significant impact on drilling stability due to their operational reliability. However, insulation failure, a potential hazard in motor operation, is both insidious and slow to develop, making it a key factor impacting the motor's long-term stable operation.

[0003] In practical applications, the main drive motor of a drilling platform experiences rapid insulation degradation due to overvoltage spikes. The peak voltage on the motor side is twice the normal voltage. The steep rising and falling edges of the voltage waveform lead to uneven voltage distribution in the motor windings, and the first and last turns are subjected to high pulse voltages, causing insulation failure. This voltage phenomenon is caused by the high-speed IGBTs switching on and off when the inverter is connected to the motor via a long cable, generating rapid dv / dt. Furthermore, the uneven routing of the output cable causes reflections of the PWM pulse voltage waveform during transmission over the long cable, inducing transient overvoltages at the generator end. This, combined with the humid, hot, and salty environment, causes rapid degradation of the motor insulation, seriously impacting normal platform operations. Traditional monitoring methods struggle to meet the real-time and accuracy requirements. Therefore, a method that can accurately predict overvoltage at the motor end is urgently needed to ensure efficient drilling production.

[0004] Currently, maintenance of variable-frequency drive motors on drilling platforms still relies primarily on periodic inspections and manual judgment, lacking real-time, intelligent online insulation diagnostics. Traditional methods struggle to accurately extract key features in environments with strong interference, non-stationary conditions, and complex background signals, leading to problems such as missed detections and misjudgments. Therefore, there is an urgent need for a motor insulation diagnostic method suitable for the complex operating conditions of drilling platforms, featuring high accuracy, robustness, and intelligence. This approach can improve equipment reliability, reduce maintenance costs, and ensure the safe and stable operation of drilling production systems. To this end, a motor overvoltage prediction method based on IOOA-CNN-BiLSTM is urgently needed. Summary of the Invention

[0005] The embodiment of the present application provides a method for predicting overvoltage at the variable frequency motor end of an oil drilling platform based on the IOOA-CNN-BiLSTM algorithm. By improving the motor insulation algorithm, the parameters of the CNN-BiLSTM model can be optimized, thereby improving the accuracy of motor insulation overvoltage prediction.

[0006] To solve the above technical problems, an embodiment of the present application provides an overvoltage prediction method for a variable frequency motor end of an oil drilling platform based on the IOOA-CNN-BiLSTM algorithm, comprising the following steps: first, collecting the voltage data of the motor; then, preprocessing the voltage data to obtain a voltage data set; next, adopting the population initialization method, convergence factor and position update formula of the improved OOA algorithm, introducing an acceleration gradient in the development stage, and constructing a CNN-BiLSTM model; then, based on IOOA, optimizing the convolution kernel size, hidden layer size, and initial learning rate of the CNN-BiLSTM model, and re-assigning the optimized parameters to the CNN-BiLSTM model to construct an IOOA-CNN-BiLSTM prediction model; finally, inputting the voltage data in the voltage data set into the IOOA-CNN-BiLSTM prediction model to obtain an overvoltage prediction result corresponding to the voltage data of the motor.

[0007] In some exemplary embodiments, preprocessing includes data cleaning, outlier and missing value processing on the voltage data; preprocessing the voltage data to obtain a voltage data set includes: using a box plot method to identify outliers in the voltage data, setting the outliers to null values ​​and then performing linear filling; using a Lagrange interpolation method to fill the missing values ​​in the voltage data; wherein the formula for identifying outliers using the box plot method is as follows:

[0008] LB=Q1-1.5×IQR,UB=Q3+1.5×IQR

[0009] Where LB is the lower limit; UB is the upper limit; Q1 represents the first quartile; Q3 represents the third quartile; IQR represents the interquartile range; the Lagrange interpolation polynomial L is obtained by interpolation basis function n (x), the expression is:

[0010]

[0011] ω n+1 (x)=(x-x0)(x-x1)…(xx n )=(x i -x0)…(x i -x i-1 )(x i -x i+1 )…(x i -x n )

[0012] Among them, l i (x) is the Lagrange basis polynomial, ω n+1 (x) is the interpolation polynomial.

[0013] In some exemplary embodiments, the population initialization method, convergence factor, and position update formula of the improved OOA algorithm are adopted, and an acceleration gradient is introduced in the development stage to construct a CNN-BiLSTM model, including:

[0014] Step 1: Use Sine mapping to initialize the population and select the individuals with the best fitness as the initial population;

[0015] Step 2: Calculate the fitness of the individual osprey and record the current global optimal position;

[0016] Step 3: Update the convergence factor according to the linear decay formula as the number of iterations increases;

[0017] Step 4: Generate a new position vector for each osprey using the fly-circle-dive model;

[0018] Step 5: Perform gradient learning on the new population and perform boundary checking;

[0019] Step 6: Re-evaluate fitness and refresh the global optimal position;

[0020] Step 7: If there is no improvement in the global optimum for N consecutive generations, the worst individual in the population is replaced with a Sine random solution to restart the population.

[0021] Step 8: If the number of iterations reaches the upper limit, the optimal value and convergence curve are output; otherwise, return to step 3 to continue iteration.

[0022] In some exemplary embodiments, in step 1, the Sine mapping formula is:

[0023] x n+1 =μsin(πx n ),x n ∈(0,1),μ∈(0,1]

[0024] Among them, x n The chaotic value of the nth iteration, μ is the control parameter.

[0025] In some exemplary embodiments, the convergence factor in step 3 decreases linearly with the number of iterations, and the calculation formula is:

[0026]

[0027] Among them, α(t) is the convergence factor at the tth iteration, α initial Initial convergence factor, t Current number of iterations, T max Maximum number of iterations, gamma decay exponential.

[0028] In some exemplary embodiments, in step 4, the individual seagull position is updated using a flying-circling-diving model, and the formula is as follows:

[0029]

[0030] Among them, X i (t+1) is the position of the ith osprey at the t+1th iteration, X i (t) is the position of the ith osprey at the tth iteration, α(t) is the convergence factor at the tth iteration, D is the preset offset, sign is the sign function, F is the fitness function, and η is the learning rate used to control the step size of the gradient descent. is the objective function f in X i (t) The gradient at .

[0031] In some exemplary embodiments, in step seven, if the global optimal fitness does not improve in N=5 consecutive iterations, the global search strategy is triggered, new individuals are regenerated using Sine mapping and gradient learning is performed, and the updated individual positions are boundary checked.

[0032] In some exemplary embodiments, the CNN-BiLSTM prediction model includes a CNN module and a BiLSTM module; the CNN module includes an input layer, a convolution layer, and a maximum pooling layer; the BiLSTM module includes a forward LSTM subnetwork and a backward LSTM subnetwork, which can simultaneously capture the relationship between the previous and next elements in the sequence, dynamically adjust the weights of forward and backward propagation, and reduce errors in the propagation process.

[0033] In some exemplary embodiments, the calculation formula of the input layer of the CNN module is:

[0034]

[0035] Where C1 is the input matrix of the input layer, with a size of b×p×q, where the number of batches is b, the time series length is p, and the number of spatial features is q; X is the Xavier initialization method; L is the linear transformation method; T(2,3) represents the second and third dimensions of the transposed matrix; C2 is the input matrix of the convolutional layer, with a size of b×m×n;

[0036]

[0037] Among them, U is the uniform distribution in the interval range; n is the size of the input matrix;

[0038] L r (p,n)=X p×i W i×n +b m

[0039] Where: Wi×n is the parameter that needs to be learned in the neural network; b m is the vector bias;

[0040] The size of CNN input is b×m×n, which means there are batches of b m×n spatial matrices, where b represents the time series feature. After processing the spatial matrix, CNN extracts the spatial features to obtain a feature matrix of size b×x×y, which serves as the input of BiLSTM.

[0041] In some exemplary embodiments, in the process of building a CNN-BiLSTM prediction model, the convolutional layers and pooling layers in the CNN are first used to extract features from high-dimensional input data to identify important patterns in the data; then the L1 regularization method is used to eliminate features with less influence and reduce the overfitting phenomenon of the model; finally, the selected key features are input into the bidirectional LSTM to retain important time series information and generate feature vectors and prediction sequences;

[0042] The calculation process of BiLSTM is as follows:

[0043]

[0044] Among them, w is the weight of the forward LSTM; w' is the weight of the backward LSTM; b y is the bias vector of the output layer.

[0045] The technical solution provided by the embodiments of the present application has at least the following advantages:

[0046] An embodiment of the present application provides a method for predicting overvoltage at the variable frequency motor end of an oil drilling platform based on the IOOA-CNN-BiLSTM algorithm, the method comprising the following steps: first, collecting voltage data of the motor; then, preprocessing the voltage data to obtain a voltage data set; next, adopting the population initialization method, convergence factor and position update formula of the improved OOA algorithm, introducing an acceleration gradient in the development stage, and constructing a CNN-BiLSTM model; then, optimizing the convolution kernel size, hidden layer size, and initial learning rate of the CNN-BiLSTM model based on IOOA, re-assigning the optimized parameters to the CNN-BiLSTM model, and constructing the IOOA-CNN-BiLSTM prediction model; finally, inputting the voltage data in the voltage data set into the IOOA-CNN-BiLSTM prediction model to obtain an overvoltage prediction result corresponding to the voltage data of the motor.

[0047] First, this application introduces a Sine map for population initialization in the Osprey Optimization Algorithm (OOA). Combined with a dynamically updated convergence factor α and perturbation coefficient C, this significantly improves the algorithm's global search capability and convergence speed. By introducing an improved position update formula and an accelerated gradient mechanism, the algorithm effectively avoids falling into local optimality and improves parameter optimization efficiency.

[0048] Secondly, this application optimizes the key parameters of CNN-BiLSTM (convolution kernel size, hidden layer size and initial learning rate) through an improved IOOA algorithm, which solves the limitations of traditional manual parameter adjustment and enables the model to adaptively capture local features and timing dependencies in voltage signals, thereby improving the accuracy of overvoltage prediction.

[0049] Thirdly, this application introduces an accelerated gradient mechanism in the development stage of OOA, which further improves the local search efficiency of the algorithm; at the same time, it maintains population diversity through boundary checking and perturbation strategies to ensure that the optimized CNN-BiLSTM model has both high accuracy and generalization ability.

[0050] Furthermore, the proposed CNN-BiLSTM fusion architecture leverages the strengths of CNN in spatial feature extraction and BiLSTM in long-term time series modeling. Combined with the dynamic optimization capabilities of IOOA, the model can more accurately predict the amplitude and waveform characteristics of transient overvoltages, making it particularly suitable for analyzing non-stationary signals under the complex operating conditions of oil drilling platforms. BRIEF DESCRIPTION OF THE DRAWINGS

[0051] One or more embodiments are exemplarily described by the pictures in the corresponding drawings. These exemplifications do not constitute limitations on the embodiments. Unless otherwise stated, the pictures in the drawings do not constitute proportional limitations.

[0052] Figure 1 This is a flow chart of a method for predicting overvoltage at the variable frequency motor end of an oil drilling platform based on the IOOA-CNN-BiLSTM algorithm provided in one embodiment of the present application.

[0053] Figure 2 A simplified schematic diagram of a motor drive system provided in one embodiment of the present application.

[0054] Figure 3 This is a flow chart of a method for predicting overvoltage at the motor end of a motor drive system provided in one embodiment of the present application.

[0055] Figure 4 This is a CNN-BiLSTM combined prediction model provided in one embodiment of the present application.

[0056] Figure 5 This is a comparison chart of the test results of the 100A algorithm provided in one embodiment of the present application and the existing algorithm.

[0057] Figure 6 This is a graph of the output voltage waveform predicted by CNN-BiLSTM according to one embodiment of the present application.

[0058] Figure 7 This is a diagram of the output voltage waveform predicted by OOA-CNN-BiLSTM according to an embodiment of the present application.

[0059] Figure 8 This is a graph showing the output voltage waveform predicted by 100A-CNN according to an embodiment of the present application.

[0060] Figure 9 This is a graph of the output voltage waveform predicted by IOOA-CNN-BiLSTM according to one embodiment of the present application. DETAILED DESCRIPTION

[0061] As can be seen from the background, traditional methods struggle to accurately extract key features in environments with strong interference, non-stationary conditions, and complex background signals, leading to problems such as missed detections and misjudgments. Therefore, there is an urgent need for a motor insulation diagnosis method suitable for the complex operating conditions of drilling platforms, featuring high precision, robustness, and intelligence. This approach can improve equipment reliability, reduce maintenance costs, and ensure the safe and stable operation of drilling production systems.

[0062] CNN excels at capturing local features and spatial hierarchical structures, effectively extracting key features from voltage signals. This is particularly true when processing sequence data, enabling rapid identification of voltage-related features. BiLSTM, on the other hand, excels at capturing long-range dependencies in sequence data. Its bidirectional structure simultaneously considers both past and future information, enabling a better understanding of dynamic voltage trends. By integrating CNN and BiLSTM into a CNN-BiLSTM model, the advantages of both can be fully leveraged to achieve rapid and accurate overvoltage prediction.

[0063] Therefore, in order to solve the above technical problems, an embodiment of the present application provides an overvoltage prediction method for the variable frequency motor end of an oil drilling platform based on the IOOA-CNN-BiLSTM algorithm, the method comprising the following steps: first, collecting the voltage data of the motor; then, preprocessing the voltage data to obtain a voltage data set; next, adopting the population initialization method, convergence factor and position update formula of the improved OOA algorithm, introducing an acceleration gradient in the development stage, and constructing a CNN-BiLSTM model; then, based on IOOA, the convolution kernel size, hidden layer size and initial learning rate of the CNN-BiLSTM model are optimized, and the optimized parameters are re-assigned to the CNN-BiLSTM model to construct an IOOA-CNN-BiLSTM prediction model; finally, the voltage data in the voltage data set is input into the IOOA-CNN-BiLSTM prediction model to obtain the overvoltage prediction result corresponding to the voltage data of the motor. An embodiment of the present application provides a method for predicting overvoltage at the variable-frequency motor end of an oil drilling platform based on the IOOA-CNN-BiLSTM algorithm. By improving the motor insulation algorithm, the parameters of the CNN-BiLSTM model can be optimized. Moreover, the convolution kernel size, hidden layer size, and learning rate parameters of the CNN-BiLSTM model are optimized using the improved Osprey algorithm, so that the prediction method exhibits greater accuracy and robustness, thereby improving the prediction accuracy of motor overvoltage.

[0064] The following detailed description of the various embodiments of the present application is provided in conjunction with the accompanying drawings. However, those skilled in the art will appreciate that many technical details are provided in the various embodiments of the present application to facilitate a better understanding of the present application. However, even without these technical details and the various variations and modifications based on the following embodiments, the technical solutions claimed in the present application can still be implemented.

[0065] See Figure 1 The embodiment of the present application provides a method for predicting overvoltage at the variable frequency motor terminal of an oil drilling platform based on the IOOA-CNN-BiLSTM algorithm, comprising the following steps:

[0066] Step S1: Collect voltage data of the motor.

[0067] Step S2: pre-process the voltage data to obtain a voltage data set.

[0068] Step S3: Using the population initialization method, convergence factor, and position update formula of the improved OOA algorithm, introducing the acceleration gradient in the development stage, and constructing the CNN-BiLSTM model.

[0069] Step S4: Based on IOOA, the convolution kernel size, hidden layer size, and initial learning rate of the CNN-BiLSTM model are optimized, and the optimized parameters are re-assigned to the CNN-BiLSTM model to construct the IOOA-CNN-BiLSTM prediction model.

[0070] Step S5: Input the voltage data in the voltage data set into the 100A-CNN-BiLSTM prediction model to obtain the overvoltage prediction result corresponding to the voltage data of the motor.

[0071] Figure 2 A variable frequency drive system for oil drilling rigs is shown. Figure 2 As shown, on a drilling platform, because the inverter is far from the motor and is connected to the motor via a long cable, the high-speed IGBT is constantly switching on and off, forming a fast dv / dt. In addition, the output cable is unevenly laid, causing the PWM pulse voltage waveform to be reflected during the long cable transmission process, inducing transient overvoltage at the generator end. Because traditional methods have difficulty accurately extracting key features in environments with strong interference, non-stationary, and complex background signals, and there are problems such as missed detection and misjudgment, this application proposes a motor overvoltage prediction method based on IOOA-CNN-BiLSTM. By improving the motor insulation algorithm, the parameters of the CNN-BiLSTM model can be optimized, thereby improving the accuracy of motor insulation overvoltage prediction.

[0072] Figure 3 The flow chart of the method for predicting the overvoltage at the motor end of the motor drive system is shown in FIG. Figure 3 As shown, first, the voltage data of the motor is collected. Then, the collected voltage signal data is preprocessed; specifically, data cleaning and outlier processing are used as the voltage data set of the motor. Next, the population initialization method, convergence factor and position update formula of the OOA algorithm are improved, and the acceleration gradient is introduced in the development stage; then, based on IOOA, the convolution kernel size, hidden layer size and initial learning rate parameters of the CNN-BiLSTM model are optimized, and the optimized parameters are re-assigned to the CNN-BiLSTM model to construct the IOOA-CNN-BiLSTM prediction model; finally, the voltage data of the motor in step S2 is input into the IOOA-CNN-BiLSTM prediction model of step S4 to obtain the motor overvoltage result corresponding to the motor voltage data. This application combines the spatial feature extraction of CNN with the bidirectional time series modeling advantages of BiLSTM, and dynamically optimizes key parameters through IOOA, which significantly improves the prediction accuracy of overvoltage amplitude and waveform under complex working conditions, solves the problem of insufficient feature extraction of traditional methods in strong interference environments, and provides an efficient solution for motor insulation fault early warning.

[0073] In some embodiments, the preprocessing in step S2 includes data cleaning, outlier and missing value processing on the voltage data; step S2 preprocesses the voltage data to obtain a voltage data set, including the following steps:

[0074] Step S201: using a box plot method to identify abnormal values ​​in the voltage data, setting the abnormal values ​​to null values ​​and then performing linear filling.

[0075] Step S202: Using Lagrange interpolation method to fill missing values ​​in the voltage data.

[0076] The formula for identifying outliers using the box plot method in step S201 is as follows:

[0077] LB=Q1-1.5×IQR,UB=Q3+1.5×IQR

[0078] Where LB is the lower limit; UB is the upper limit; Q1 is the first quartile; Q3 is the third quartile; IQR is the interquartile range.

[0079] In step S202, the Lagrange interpolation polynomial L is obtained by interpolating the basis function. n (x), the expression is:

[0080]

[0081] ω n+1 (x)=(x-x0)(x-x1)…(xx n )=(x i -x0)…(x i -x i-1 )(x i -x i+1 )…(x i -x n )

[0082] Among them, l i (x) is the Lagrange basis polynomial, ω n+1 (x) is the interpolation polynomial.

[0083] In some embodiments, step S3 adopts the population initialization method, convergence factor, and position update formula of the improved OOA algorithm, introduces the acceleration gradient in the development stage, and constructs the CNN-BiLSTM model, including the following steps:

[0084] Step S301: Initialize the population using Sine mapping, and select individuals with the best fitness as the initial population;

[0085] Step S302: Calculate the fitness of the individual osprey and record the current global optimal position;

[0086] Step S303: updating the convergence factor according to the linear decay formula as the number of iterations increases;

[0087] Step S304: Generate a new position vector for each osprey using the flight-circling-dive model;

[0088] Step S305: perform gradient learning on the new population and perform boundary checking;

[0089] Step S306: Re-evaluate the fitness and refresh the global optimal position;

[0090] Step S307: If there is no improvement in the global optimum for N consecutive generations, the worst individual in the population is replaced with a Sine random solution to restart the population.

[0091] Step S308: If the number of iterations reaches the upper limit, the optimal value and the convergence curve are output; otherwise, the process returns to step S303 to continue iteration.

[0092] In some embodiments, in step S301, the mapping formula of Sine is:

[0093] x n+1 =μsin(πx n ),x n ∈(0,1),μ∈(0,1]

[0094] Among them, x n The chaotic value of the nth iteration, μ is the control parameter.

[0095] In some embodiments, the convergence factor in step S303 decreases linearly with the number of iterations, and the calculation formula is:

[0096]

[0097] Among them, α(t) is the convergence factor at the tth iteration, α initial Initial convergence factor, t Current number of iterations, T max Maximum number of iterations, gamma decay exponential.

[0098] In step S303, the convergence factor α and the disturbance coefficient C are updated. The updating formula of the disturbance coefficient C is:

[0099]

[0100] C(t) is the perturbation coefficient at the tth iteration, C initial Initial perturbation coefficient.

[0101] In some embodiments, in step S304, the individual seagull position is updated using a flying-circling-diving model, and the formula is as follows:

[0102]

[0103] Among them, X i (t+1) is the position of the ith osprey at the t+1th iteration, X i (t) is the position of the ith osprey at the tth iteration, α(t) is the convergence factor at the tth iteration, D is the preset offset, sign is the sign function, F is the fitness function, and η is the learning rate used to control the step size of the gradient descent. is the objective function f in X i (t) The gradient at .

[0104] The calculation formula for the flight phase is:

[0105]

[0106] Among them, X i (t+1) is the position of the ith osprey at the t+1th iteration, X i (t) is the position of the ith osprey at the tth iteration, X best is the current global optimal position, ω is the inertia weight, and α is the acceleration gradient parameter.

[0107] The calculation formula for the hovering phase is:

[0108] X i (t+1) =X best +βrand(X best -X i (t) )

[0109] Where β is a random factor and rand is a random number in the range of [-1,1].

[0110] The calculation formula for the subduction phase is:

[0111] X i (t+1) =X best +γrand(X best -X i (t) )

[0112] Among them, γ is a random factor.

[0113] Furthermore, in step S3, the gradient acceleration formula is introduced as follows:

[0114]

[0115] Among them, as mentioned above, X i (t+1) is the position of the ith osprey at the t+1th iteration, X i (t) is the position of the ith osprey at the tth iteration, η is the learning rate used to control the step size of the gradient descent, is the objective function f in X i (t) The gradient at .

[0116] In some embodiments, if the global optimal fitness does not improve in N=5 consecutive iterations in step S307, the global search strategy is triggered, a new individual is regenerated using Sine mapping and gradient learning is performed, and a boundary check is performed on the updated individual position.

[0117] like Figure 4 As shown, in some embodiments, the CNN-BiLSTM prediction model includes a CNN module and a BiLSTM module; the CNN module includes an input layer, a convolution layer, and a maximum pooling layer; the BiLSTM module includes a forward LSTM subnetwork and a reverse LSTM subnetwork, which can simultaneously capture the relationship between the previous and next elements in the sequence, dynamically adjust the weights of forward and reverse propagation, and reduce errors in the propagation process.

[0118] In some embodiments, the calculation formula of the input layer of the CNN module is:

[0119]

[0120] Where C1 is the input matrix of the input layer, with a size of b×p×q, where the number of batches is b, the time series length is p, and the number of spatial features is q; X is the Xavier initialization method; L is the linear transformation method; T(2,3) represents the second and third dimensions of the transposed matrix; C2 is the input matrix of the convolutional layer, with a size of b×m×n;

[0121]

[0122] Among them, U is the uniform distribution in the interval range; n is the size of the input matrix;

[0123] L r (p,n)=X p×i W i×n +b m

[0124] Where: W i×n is the parameter that needs to be learned in the neural network; b m is the vector bias;

[0125] The size of CNN input is b×m×n, which means there are batches of b m×n spatial matrices, where b represents the time series feature. After processing the spatial matrix, CNN extracts the spatial features to obtain a feature matrix of size b×x×y, which serves as the input of BiLSTM.

[0126] In some exemplary embodiments, in the process of building a CNN-BiLSTM prediction model, the convolutional layers and pooling layers in the CNN are first used to extract features from high-dimensional input data to identify important patterns in the data; then the L1 regularization method is used to eliminate features with less influence and reduce the overfitting phenomenon of the model; finally, the selected key features are input into the bidirectional LSTM to retain important time series information and generate feature vectors and prediction sequences;

[0127] The calculation process of BiLSTM is as follows:

[0128]

[0129] Among them, w is the weight of the forward LSTM; w' is the weight of the backward LSTM; b y is the bias vector of the output layer.

[0130] Therefore, this application significantly enhances the algorithm's global search capability and optimization accuracy by introducing innovative methods such as Sine initialization, dynamically updating the convergence factor and perturbation coefficient, adopting a flight-circling-dive model, and incorporating gradient learning. By optimizing the key parameters of the CNN-BiLSTM model using IOOA, the constructed IOOA-CNN-BiLSTM prediction model demonstrates higher accuracy and robustness in motor overvoltage prediction tasks, effectively improving prediction accuracy and providing strong support for the safe operation of the motor system.

[0131] Specifically, the parameter settings of the improved 100A algorithm are shown in Table 1.

[0132] Table 1 Parameters of the improved IOOA algorithm

[0133] parameter Setting value Population size 50 Maximum number of iterations 30 Initial convergence factor 0.9 Gradient learning rate 0.01

[0134] Specifically, in step S302, the prediction error is used as the fitness of the calculation algorithm, and the calculation formula is:

[0135]

[0136] Among them, n is the number of data in the test set; y i is the true value; is the predicted value, RMSE is the prediction accuracy of the model, MAE is the actual prediction error, and MAPE is the model accuracy. The smaller the values ​​of RMSE, MAE, and MAPE, the better the prediction effect of the model and the higher the prediction accuracy.

[0137] In this embodiment, in step S3, the convolution kernel size, hidden layer size, and initial learning rate parameters of the CNN-BiLSTM model are optimized by IOOA, the optimized parameters are reassigned to the CNN-BiLSTM model, and the CNN-BiLSTM model with updated parameters is retrained and tested.

[0138] It should be noted that, in order to verify the effectiveness of the motor fault diagnosis method, the designed IOOA algorithm, OOA, and SSA algorithms are optimized for the same test function. The race size is set to 50 and the number of iterations is 200. In order to see the performance comparison more intuitively, the fitness value is transformed by log10. It can be seen that the optimization and convergence capabilities of the IOOA algorithm are better than those of OOA and SSA. The experimental results are as follows: Figures 5 to 9 shown.

[0139] Table 2 Comparison of evaluation indicators of different models

[0140] Prediction Model RMSE MAE MAPE / % IOOA-CNN-BiLSTM 7.3931 5.0944 2.16 IOOA-CNN 9.8632 6.3792 27.3086 OOA-CNN-BiLSTM 146.3183 119.2843 49.74 CNN-BiLSTM 14.2401 9.6775 4.26

[0141] As shown in Table 2, the RMSE, MAE, and MAPE of the IOOA-CNN-BiLSTM model are 7.3931, 5.0944, and 2.16%, respectively, significantly outperforming the other models. In particular, the RMSE of the IOOA-CNN-BiLSTM model improves by 55.45%, 94.94%, and 48.08%, respectively, compared to the other three models.

[0142] In summary, this application combines the powerful advantages of the improved Osprey Optimization Algorithm (IOOA) and Convolutional Neural Network (CNN) with Bidirectional LSTM (BiLSTM) for motor fault diagnosis tasks. IOOA optimizes the hyperparameters of the CNN-BiLSTM model by introducing Sine chaos initialization and adding gradient learning techniques, thereby improving the global search capability and accuracy of the model in high-dimensional and complex data. CNN extracts the spatial features of the input signal, while BiLSTM effectively captures the long-term and short-term dependencies in the time series through its bidirectional learning mechanism. Combined with IOOA for hyperparameter optimization, the model shows better prediction ability and faster convergence speed than traditional optimization algorithms, significantly improving the accuracy of the prediction.

[0143] According to the following technical solution, an embodiment of the present application provides a method for predicting overvoltage at the variable frequency motor end of an oil drilling platform based on the IOOA-CNN-BiLSTM algorithm, the method comprising the following steps: first, collecting the voltage data of the motor; then, preprocessing the voltage data to obtain a voltage data set; next, adopting the population initialization method, convergence factor and position update formula of the improved OOA algorithm, introducing an acceleration gradient in the development stage, and constructing a CNN-BiLSTM model; then, based on IOOA, optimizing the convolution kernel size, hidden layer size, and initial learning rate of the CNN-BiLSTM model, and re-assigning the optimized parameters to the CNN-BiLSTM model to construct an IOOA-CNN-BiLSTM prediction model; finally, inputting the voltage data in the voltage data set into the IOOA-CNN-BiLSTM prediction model to obtain an overvoltage prediction result corresponding to the voltage data of the motor.

[0144] First, this application introduces a Sine map for population initialization in the Osprey Optimization Algorithm (OOA). Combined with a dynamically updated convergence factor α and perturbation coefficient C, this significantly improves the algorithm's global search capability and convergence speed. By introducing an improved position update formula and an accelerated gradient mechanism, the algorithm effectively avoids falling into local optimality and improves parameter optimization efficiency.

[0145] Secondly, this application optimizes the key parameters of CNN-BiLSTM (convolution kernel size, hidden layer size and initial learning rate) through an improved IOOA algorithm, which solves the limitations of traditional manual parameter adjustment and enables the model to adaptively capture local features and timing dependencies in voltage signals, thereby improving the accuracy of overvoltage prediction.

[0146] Thirdly, this application introduces an accelerated gradient mechanism in the development stage of OOA, which further improves the local search efficiency of the algorithm; at the same time, it maintains population diversity through boundary checking and perturbation strategies to ensure that the optimized CNN-BiLSTM model has both high accuracy and generalization ability.

[0147] Furthermore, the proposed CNN-BiLSTM fusion architecture leverages the strengths of CNN in spatial feature extraction and BiLSTM in long-term time series modeling. Combined with the dynamic optimization capabilities of IOOA, the model can more accurately predict the amplitude and waveform characteristics of transient overvoltages, making it particularly suitable for analyzing non-stationary signals under the complex operating conditions of oil drilling platforms.

[0148] Those skilled in the art will appreciate that the above-described embodiments are specific examples for implementing the present application, and that in actual applications, various changes in form and detail may be made thereto without departing from the spirit and scope of the present application. Any person skilled in the art may make changes and modifications without departing from the spirit and scope of the present application. Therefore, the scope of protection of the present application shall be subject to the scope defined in the claims.

Claims

1. A method for predicting overvoltage at the variable frequency motor terminal of an oil drilling platform based on the IOOA-CNN-BiLSTM algorithm, characterized in that: The following steps are involved: Collect the voltage data of the motor; Preprocessing the voltage data to obtain a voltage data set; The improved OOA algorithm's population initialization method, convergence factor, and position update formula were used, and accelerated gradients were introduced during the development phase to construct a CNN-BiLSTM model. Based on IOOA, the convolution kernel size, hidden layer size, and initial learning rate of the CNN-BiLSTM model are optimized, and the optimized parameters are re-assigned to the CNN-BiLSTM model to construct the IOOA-CNN-BiLSTM prediction model; The voltage data in the voltage data set is input into the IOOA-CNN-BiLSTM prediction model to obtain the overvoltage prediction result corresponding to the voltage data of the motor.

2. The method for predicting overvoltage at the terminal of a variable frequency motor on a petroleum drilling platform based on the 100A-CNN-BiLSTM algorithm according to claim 1, wherein: The preprocessing includes data cleaning, outlier and missing value processing on the voltage data; Preprocessing the voltage data to obtain a voltage data set includes: A box plot method is used to identify abnormal values ​​in the voltage data, and the abnormal values ​​are set as null values ​​and then linearly filled; Using Lagrange interpolation method to fill missing values ​​in the voltage data; Among them, the formula for identifying outliers using the box plot method is as follows: LB=Q1-1.5×IQR,UB=Q3+1.5×IQR Where LB is the lower limit; UB is the upper limit; Q1 is the first quartile; Q3 is the third quartile; IQR is the interquartile range; The Lagrange interpolation polynomial L is obtained by interpolation basis function n (x), the expression is: ω n+1 (x)=(x-x0)(x-x1)…(xx n )=(x i -x0)…(x i -x i-1 )(x i -x i+1 )…(x i -x n ) Among them, l i (x) is the Lagrange basis polynomial, ω n+1 (x) is the interpolation polynomial.

3. The method for predicting overvoltage at the terminal of a variable frequency motor of a petroleum drilling platform based on the 100A-CNN-BiLSTM algorithm according to claim 1, wherein: The improved OOA algorithm's population initialization method, convergence factor, and position update formula are adopted, and accelerated gradients are introduced during the development phase to construct a CNN-BiLSTM model, including: Step 1: Use Sine mapping to initialize the population and select the individuals with the best fitness as the initial population; Step 2: Calculate the fitness of the individual osprey and record the current global optimal position; Step 3: Update the convergence factor according to the linear decay formula as the number of iterations increases; Step 4: Generate a new position vector for each osprey using the fly-circle-dive model; Step 5: Perform gradient learning on the new population and perform boundary checking; Step 6: Re-evaluate fitness and refresh the global optimal position; Step 7: If there is no improvement in the global optimum for N consecutive generations, the worst individual in the population is replaced with a Sine random solution to restart the population. Step 8: If the number of iterations reaches the upper limit, the optimal value and convergence curve are output; otherwise, return to step 3 to continue iteration.

4. The method for predicting overvoltage at the terminal of a variable frequency motor of a petroleum drilling platform based on the 100A-CNN-BiLSTM algorithm according to claim 3, wherein: In step 1, the Sine mapping formula is: x n+1 =μsin(πx n ),x n ∈(0,1),μ∈(0,1] Among them, x n The chaotic value of the nth iteration, μ is the control parameter.

5. The method for predicting overvoltage at the terminal of a variable frequency motor of a petroleum drilling platform based on the 100A-CNN-BiLSTM algorithm according to claim 3, wherein: The convergence factor in step 3 decreases linearly with the number of iterations, and the calculation formula is: Among them, α(t) is the convergence factor at the tth iteration, α initial Initial convergence factor, t Current number of iterations, T max Maximum number of iterations, gamma decay exponential.

6. The method for predicting overvoltage at the terminal of a variable frequency motor of a petroleum drilling platform based on the 100A-CNN-BiLSTM algorithm according to claim 3, wherein: In step 4, the individual position of the seagull is updated using the flight-circling-dive model, and the formula is as follows: Among them, X i (t+1) is the position of the ith osprey at the t+1th iteration, X i (t) is the position of the ith osprey at the tth iteration, α(t) is the convergence factor at the tth iteration, D is the preset offset, sign is the sign function, F is the fitness function, and η is the learning rate used to control the step size of the gradient descent. is the objective function f in X i (t) The gradient at .

7. The method for predicting overvoltage at the terminal of a variable frequency motor of a petroleum drilling platform based on the 100A-CNN-BiLSTM algorithm according to claim 3, wherein: In step 7, if the global optimal fitness does not improve in N = 5 consecutive iterations, the global search strategy is triggered, and new individuals are regenerated using Sine mapping and gradient learning is performed, and the updated individual positions are boundary checked.

8. The method for predicting overvoltage at the terminal of a variable frequency motor of a petroleum drilling platform based on the 100A-CNN-BiLSTM algorithm according to claim 1, wherein: The CNN-BiLSTM prediction model includes a CNN module and a BiLSTM module; The CNN module includes input layer, convolution layer and maximum pooling layer; The BiLSTM module includes a forward LSTM sub-network and a backward LSTM sub-network, which can simultaneously capture the relationship between the previous and next elements in the sequence, dynamically adjust the weights of forward and backward propagation, and reduce errors in the propagation process.

9. The method for predicting overvoltage at the terminal of a variable frequency motor of a petroleum drilling platform based on the 100A-CNN-BiLSTM algorithm according to claim 8, wherein: The calculation formula of the input layer of the CNN module is: Where C1 is the input matrix of the input layer, with a size of b×p×q, where the number of batches is b, the time series length is p, and the number of spatial features is q; X is the Xavier initialization method; L is the linear transformation method; T(2,3) represents the second and third dimensions of the transposed matrix; C2 is the input matrix of the convolutional layer, with a size of b×m×n; Among them, U is the uniform distribution in the interval range; n is the size of the input matrix; L r (p,n)=X p×i W i×n +b m Where: W i×n is the parameter that needs to be learned in the neural network; b m is the vector bias; The size of CNN input is b×m×n, which means there are batches of b m×n spatial matrices, where b represents the time series feature. After processing the spatial matrix, CNN extracts the spatial features to obtain a feature matrix of size b×x×y, which serves as the input of BiLSTM.

10. The method for predicting overvoltage at the terminal of a variable frequency motor of a petroleum drilling platform based on the 100A-CNN-BiLSTM algorithm according to claim 1, wherein: In the process of building the CNN-BiLSTM prediction model, the convolutional and pooling layers in the CNN are first used to extract features from the high-dimensional input data and identify important patterns in the data. Then, the L1 regularization method is used to eliminate less influential features and reduce the overfitting phenomenon of the model. Finally, the selected key features are input into the bidirectional LSTM to retain important time series information and generate feature vectors and prediction sequences. The calculation process of BiLSTM is as follows: Among them, w is the weight of the forward LSTM; w' is the weight of the backward LSTM; b y is the bias vector of the output layer.

Images