Fault monitoring system and monitoring method for centrifugal pump unit of deep sea oil production platform

Abstract

The invention belongs to the technical field of intelligent operation and maintenance of deep-sea equipment, and discloses a fault monitoring system and method for a centrifugal pump unit of a deep-sea oil production platform. The main control module is connected with the deep-sea pressure-resistant optical fiber current sensing module, the data storage unit, the man-machine interaction module and the power supply module; the data storage unit is connected with the man-machine interaction module and used for sending and storing an SCBA model data set; the man-machine interaction module is connected with the power supply module and the data storage unit and is used for displaying the monitoring result and inputting the rotating speed of the centrifugal pump; and the power supply module is connected with the deep-sea pressure-resistant optical fiber current sensing module and the man-machine interaction module and provides electric energy. According to the method, a multi-mode hyper-parameter collaborative optimization framework is constructed based on the Sabot swarm optimization algorithm, effective recognition of the centrifugal pump impeller fault signal is achieved, and the accuracy of feature recognition of the centrifugal pump impeller fault signal is improved.

Description

Technical Field

[0001] The present invention relates to the technical field of intelligent operation and maintenance of deep-sea equipment, and particularly to a fault monitoring system and method for a centrifugal pump unit of a deep-sea oil production platform. Background Art

[0002] As the core power equipment of the oil and gas transportation system of a deep-sea oil production platform, a centrifugal pump undertakes key tasks such as crude oil lifting and water injection pressurization. Compared with the land industrial scenario, the deep-sea centrifugal pump unit faces extreme environmental challenges for a long time, including: 1. The continuous deep-sea pressure is much higher than that on land, resulting in the saturation failure of the magnetic core of traditional electromagnetic sensors; 2. Electrochemical corrosion and microbial attachment caused by the penetration of high-salinity seawater accelerate the wear of mechanical components; 3. Unsteady load fluctuations caused by complex seabed flow patterns exacerbate the risks of impeller cavitation and fatigue fracture. An important source of deep-sea centrifugal pump failures is the failure of the impeller system, such as blade fracture, surface pitting, and dynamic balance instability. Its sudden failure can lead to a large reduction in the daily production of a single well, and the deep-sea maintenance cost is much higher than that of land operations.

[0003] Based on the motor current signal analysis technology (MCSA), by analyzing the motor stator current signal to extract characteristic indicators, the fault diagnosis of the driven equipment can be realized. When using the MCSA technology to detect impeller breakage faults, the current signal transmission path is less, there is no need to open the pump shell, and it is not easily affected by the surrounding environment of the operating equipment. It has the advantages of non-invasive, low cost, and easy implementation.

[0004] Chinese invention patent: Publication number "CN117743962A", titled "A Fault Diagnosis Method for a Train Traction Motor Bearing Based on Stator Current Signal", discloses a fault diagnosis method for a train traction motor bearing based on stator current signal. First, obtain the motor bearing temperature, speed, and motor stator current signal; when the motor bearing temperature reaches the alarm threshold, directly perform bearing fault alarm; otherwise, perform motor speed and current monitoring; when the speed reaches the set value, perform data preprocessing, variational mode decomposition, and optimal component reconstruction on the original data of the motor current signal in sequence, perform secondary filtering noise reduction based on the maximum spectral kurtosis and Hilbert envelope demodulation on the reconstructed signal to extract weak fault characteristics, and finally complete fault early warning and diagnosis according to threshold dynamic learning. This technical solution improves the traditional MCSA technology to extract weak fault characteristics and effectively diagnose early bearing faults. However, this technical solution is limited to the conventional industrial environment and does not involve core problems such as deep-sea high-pressure compensation and multi-physical field coupling modeling. Summary of the Invention

[0005] In order to solve the problem that the technical solutions in the above-mentioned prior art are limited to conventional industrial environments and do not involve problems such as deep-sea high-pressure compensation and multi-physical field coupling modeling, the present invention proposes a deep-sea oil production platform centrifugal pump unit fault monitoring system and monitoring method.

[0006] The present invention is achieved through the following technical solutions:

[0007] Deep-sea compression-resistant optical fiber current sensing module, used to collect current signals in real time and transmit them to the multi-modal signal processing host;

[0008] The multimodal signal processing host is connected to the deep-sea compressive optical fiber current sensing module, the data storage unit, the human-computer interaction module and the power module, and is used to receive and pre-process the current signal, and then perform feature extraction to form an SCBA data set to train the SCBA model and obtain monitoring results;

[0009] A data storage unit connected to the human-computer interaction module and used for sending and storing the SCBA model data set;

[0010] A human-computer interaction module, connected to the power module and the data storage unit, for displaying monitoring results and inputting the rotation speed of the centrifugal pump;

[0011] The power module is connected to the deep-sea pressure-resistant optical fiber current sensing module and the human-computer interaction module and provides electrical energy.

[0012] Furthermore, the multimodal signal processing host includes a multi-factor fault feature extraction module and a fault intelligent identification module; the multi-factor fault feature extraction module is respectively connected to the power supply module, the fault intelligent identification module, the deep-sea compressive optical fiber current sensing module and the data storage unit, and the fault intelligent identification module is also connected to the power supply module, the human-computer interaction module and the data storage module.

[0013] Furthermore, the human-computer interaction module includes a display unit and a touch control unit; the display unit is respectively connected to the multimodal signal processing host, the data storage unit and the touch control unit, and the touch control unit is respectively connected to the data storage unit and the power module.

[0014] Furthermore, the deep-sea pressure-resistant optical fiber current sensing module includes a titanium alloy sealed cabin and a deep-sea pressure-resistant optical fiber current sensor encapsulated inside the titanium alloy sealed cabin, and the two ends of the titanium alloy sealed cabin are matched by threaded fasteners and threaded engagement structures, and are sealed to form a closed cavity; the live wire insulation layer of the pump group cable passes through the titanium alloy sealed cabin, and the contact part between the live wire insulation layer of the pump group cable and the titanium alloy sealed cabin is sealed, and the deep-sea pressure-resistant optical fiber current sensor is non-invasively wrapped around the live wire insulation layer of the pump group cable through an annular Faraday sensing optical path.

[0015] Furthermore, the pressure resistance level of the titanium alloy sealed cabin is ≥50 MPa.

[0016] Furthermore, the touch control unit is an LCD touch control module.

[0017] The present invention also provides a monitoring method applicable to the centrifugal pump unit fault monitoring system of the deep-sea oil production platform described in the present invention, which includes the following steps:

[0018] S1. The user sets the rotation speed n of the centrifugal pump through the touch control unit and sends it to the multi-fault feature extraction module through the data storage unit;

[0019] S2. The deep-sea pressure-resistant optical fiber current sensing module real-time collects the single-phase motor stator current signal of the centrifugal pump and sends it to the multi-fault feature extraction module;

[0020] S3. The multi-fault feature extraction module decomposes the single-phase motor stator current signal of the centrifugal pump transmitted by the deep-sea pressure-resistant optical fiber current sensing module and preprocesses the signal. The multi-fault feature extraction module uses the cyclic autocorrelation function algorithm to obtain a series of cyclic autocorrelation feature components;

[0021] S4. The multi-fault feature extraction module decomposes the single-phase motor stator current signal of the centrifugal pump transmitted by the deep-sea pressure-resistant optical fiber current sensing module and preprocesses the signal. The multi-fault feature extraction module uses the empirical mode decomposition algorithm to obtain a series of rotor imbalance feature components;

[0022] S5. The multi-fault feature extraction module decomposes the single-phase motor stator current signal of the centrifugal pump transmitted by the deep-sea pressure-resistant optical fiber current sensing module and preprocesses the signal. The multi-fault feature extraction module solves the root mean square value to obtain the energy-related feature components;

[0023] S6. The multi-fault feature extraction module constructs an SCBA model data set composed of cyclic autocorrelation feature components, rotor imbalance feature components, and energy-related feature components, and sends the SCBA model data set to the data storage unit for storage. At the same time, the multi-fault feature extraction module sends the feature indicators corresponding to the faults to the fault intelligent identification module;

[0024] S7. The data storage unit sends the SCBA model data set to the fault intelligent identification module. The SCBA model built in the fault intelligent identification module trains the SCBA model data set to obtain the trained SCBA model;

[0025] S8. The fault intelligent identification module uses the trained SCBA model to monitor the impeller breakage state for the feature indicators corresponding to the faults sent by the multi-fault feature extraction module to obtain the monitoring results;

[0026] S9. Send the monitoring results obtained in step S8 to the data storage unit for storage by the fault intelligent recognition module;

[0027] S10. Send the monitoring results obtained in step S8 to the display unit for display by the fault intelligent recognition module, and provide the operation information of the centrifugal pump unit of the deep - sea oil production platform to the user.

[0028] Further, the specific steps of steps S3, S4, and S5 are as follows:

[0029] S31. Use the power - frequency modulation effect elimination model based on singular value decomposition for the single - phase motor stator current signal of the centrifugal pump to isolate the power - frequency - dominant modulation component in the current signal;

[0030] S32. Use the cyclic autocorrelation function algorithm to extract cyclic autocorrelation features, and obtain a series of cyclic autocorrelation feature components. The formula is as follows:

[0031] R x (t,τ)=E{x(t + τ / 2)x * (t - τ / 2)}

[0032] Among them, R x (t,τ) is the autocorrelation feature function; * represents the conjugate variable; t is time; τ is the time delay factor, and E{·} represents calculating the average;

[0033] S41. Use the power - frequency modulation effect elimination model based on singular value decomposition for the single - phase motor stator current signal of the centrifugal pump to isolate the power - frequency - dominant modulation component in the current signal;

[0034] S42. Use the empirical mode decomposition algorithm to decompose the pre - processed current signal and extract a series of rotor imbalance feature components;

[0035] S51. Use the power - frequency modulation effect elimination model based on singular value decomposition for the single - phase motor stator current signal of the centrifugal pump to isolate the power - frequency - dominant modulation component in the current signal. This step is the same as steps S31 and S41.

[0036] S52. Solve the root - mean - square value to obtain the energy - related feature components. The formula is as follows:

[0037]

[0038] The above formula is the root - mean - square value of the single - phase motor stator current signal of the centrifugal pump within the time period 0 - T.

[0039] Further, the specific steps of step S7 are as follows:

[0040] S71. Based on the sand cat swarm optimization algorithm and the convolutional neural network - bidirectional long short - term memory network - attention mechanism model, a multi - modal hyperparameter collaborative optimization framework is constructed, and the formula is as follows:

[0041]

[0042] Among them, η is the learning rate of the model, h is the number of nodes in the optimal hidden layer, and l is the optimal L2 regularization coefficient.

[0043] S72. The sand cat swarm optimization algorithm and the SCBA model training set are used to optimize the multi - modal hyperparameter collaborative optimization framework, and the trained SCBA model is obtained.

[0044] Furthermore, the SCBA model is composed of a sequence input layer, a sequence folding layer, a convolutional layer, a ReLU activation layer, a sequence unfolding layer, a BiLSTM layer, an Attention layer, a fully - connected layer, and a Softmax layer.

[0045] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0046] 1. The fiber optic current sensor encapsulated in a titanium alloy sealed cabin is adopted in the present invention. Based on the MCSA technology, the non - destructive assessment of the impeller state is realized. During the fault monitoring process, there is no need to open the pump shell. This technology is non - invasive, has a significantly lower price compared with traditional methods, and is easy to implement. By breaking through the high - voltage failure bottleneck of traditional Hall sensors through a non - invasive fiber optic sensing architecture, and integrating environmental coupling characteristics and a dynamic optimization model, a high - reliability state monitoring solution for deep - sea pump sets is provided.

[0047] 2. Based on the circular auto - correlation function algorithm (CAF algorithm for short) and empirical mode decomposition (EMD for short) and time - domain signal processing methods, the present invention analyzes the impeller damage characteristics composed of energy - related characteristics, auto - correlation characteristics, and rotor imbalance characteristics from the original current signal, and constructs a high - dimensional impeller fault feature matrix, solving the mode mixing problem of traditional single - domain feature extraction under strong non - stationary signals.

[0048] 3. Based on the sand cat swarm optimization algorithm (SCSO algorithm for short), the convolutional neural network - bidirectional long short - term memory network - attention mechanism model (CNN - BiLSTM - Attention) is dynamically adjusted in the present invention, and a multi - modal hyperparameter collaborative optimization framework is constructed, realizing the effective identification of centrifugal pump impeller fault signals and improving the accuracy of centrifugal pump impeller fault signal feature recognition. Description of the Drawings

[0049] Figure 1 This is the overall structural block diagram of the fault monitoring system of the present invention.

[0050] Figure 2 This is the schematic diagram of the hardware structure of the deep-sea pressure-resistant optical fiber current sensing module of the present invention.

[0051] Figure 3 This is the overall process schematic diagram of the fault monitoring method of the present invention.

[0052] Figure 4 This is the schematic diagram of the SCBA model architecture of the present invention.

[0053] Markings in the figure:

[0054] 1. Threaded fastener; 2. Threaded meshing structure; 3. Insulation layer of the live wire of the pump group cable; 4. Deep-sea pressure-resistant optical fiber current sensor; 5. Ring-shaped Faraday sensing optical path; 6. Titanium alloy sealed cabin. Specific embodiments

[0055] The advantages and features of the present invention will be illustrated and explained by the following non-restrictive description of preferred embodiments, which are given only by way of example with reference to the accompanying drawings.

[0056] As Figure 1 shown, the present invention provides a centrifugal pump unit fault monitoring system for a deep-sea oil production platform, which includes a deep-sea pressure-resistant optical fiber current sensing module, a power supply module, a multi-modal signal processing host, a data storage unit, and a human-computer interaction module; the multi-modal signal processing host includes a multi-fault feature extraction module and a fault intelligent recognition module; the human-computer interaction module includes a display unit and a touch control unit; the multi-modal signal processing host is respectively connected to the deep-sea pressure-resistant optical fiber current sensing module, the data storage unit, and the display unit; the display unit is respectively connected to the multi-modal signal processing host, the data storage unit, and the touch control unit. The power supply module is connected to the deep-sea pressure-resistant optical fiber current sensing module, the multi-modal signal processing host, and the human-computer interaction module for supplying electrical energy to them.

[0057] As Figure 2The figure shows a schematic diagram of the hardware structure of the deep-sea pressure-resistant optical fiber current sensing module of the present invention. The deep-sea pressure-resistant optical fiber current sensing module includes a titanium alloy sealed cabin 6 and a deep-sea pressure-resistant optical fiber current sensor 4 encapsulated inside the titanium alloy sealed cabin 6. The pressure resistance level of the titanium alloy sealed cabin 6 is ≥50MPa. The two ends of the titanium alloy sealed cabin 6 are matched by threaded fasteners 1 and threaded engagement structures 2, and are sealed to form a closed cavity. The live wire insulation layer 3 of the pump group cable passes through the titanium alloy sealed cabin 6, and the contact part of the live wire insulation layer 3 of the pump group cable and the titanium alloy sealed cabin 6 is sealed. The sealing process is a prior art and will not be repeated here. The deep-sea pressure-resistant optical fiber current sensor 4 is non-invasively wound on the live wire insulation layer 3 of the pump group cable through an annular Faraday sensing optical path, and the stator current signal of the single-phase motor of the centrifugal pump is collected in real time based on the magneto-optical rotation effect. The live wire insulation layer 3 of the pump group cable is connected to the drive motor in the centrifugal pump assembly. The monitoring method and monitoring system described in the present invention mainly determine the fault classification of the centrifugal pump by monitoring the current signal of the drive motor in the centrifugal pump assembly. The drive motor is a three-phase asynchronous motor.

[0058] The multivariate fault feature extraction module is connected to the deep-sea pressure-resistant optical fiber current sensing module, the data storage unit, the power supply module and the fault intelligent identification module, and is used to receive the current signal of the deep-sea pressure-resistant optical fiber current sensing module, pre-process the current signal, eliminate the modulation effect of the power frequency on the feature, and then extract the feature of the pre-processed current signal to obtain the characteristic index corresponding to the fault. The characteristic index corresponding to the fault includes a cyclic autocorrelation characteristic component, a rotor imbalance characteristic component and an energy-related characteristic component. The multivariate fault feature extraction module also extracts the characteristic index corresponding to the fault to form a data set of the SCBA model, and transmits the data set of the SCBA model and the label information corresponding to the fault to the data storage unit. The SCBA model is a sand cat group-convolutional neural network-bidirectional long short-term memory network-attention mechanism model (SCSO-CNN-BiLSTM-Attention, referred to as SCBA model) designed by the present invention.

[0059] The fault intelligent identification module has a built-in SCBA model, is connected to the data storage unit, the display unit and the power module, and is used to receive the data set of the SCBA model and the label information corresponding to the fault sent by the data storage unit, and input the data set of the SCBA model and the label information corresponding to the fault into the built-in SCBA model for training to obtain a trained SCBA model. The fault intelligent identification module uses the trained SCBA model to identify and judge the characteristic indicators corresponding to the fault extracted by the multivariate fault feature extraction module, and realizes the classification of different impeller damage faults of the centrifugal pump.

[0060] The data storage unit is connected to the multi-fault feature extraction module, the fault intelligent recognition module, and the touch control unit, and is used to send and store the data set of the SCBA model and the label information corresponding to the faults. The touch control unit is connected to the data storage unit and the power module. The touch control unit is an LCD touch control module, which is used for the user to input the rotation speed of the centrifugal pump, and sends the rotation speed of the centrifugal pump input by the user to the multi-fault feature extraction module through the data storage unit, so as to provide known parameters for the algorithms in the multi-fault feature extraction module. The display unit is connected to the fault intelligent recognition module and the power module, and is used to display the signal analysis results, historical data, monitoring results, user suggestions, and usage instructions.

[0061] As Figure 3 shown, the present invention also designs a method for fault monitoring using the deep-sea oil production platform centrifugal pump unit fault monitoring system of the present invention, including the following steps:

[0062] S1. The human-computer interaction module performs parameter setting: The user sets the rotation speed n of the centrifugal pump through the touch control unit and sends it to the multi-fault feature extraction module through the data storage unit;

[0063] S2. The deep-sea pressure-resistant optical fiber current sensing module real-time collects the single-phase motor stator current signal of the centrifugal pump and sends it to the multi-fault feature extraction module.

[0064] The acquisition frequency of the deep-sea pressure-resistant optical fiber current sensing module when obtaining the single-phase motor stator current signal of the centrifugal pump is 1 kHz to 10 kHz.

[0065] S3. The multi-fault feature extraction module decomposes the single-phase motor stator current signal of the centrifugal pump transmitted by the deep-sea pressure-resistant optical fiber current sensing module and preprocesses the signal. The multi-fault feature extraction module uses the cyclic autocorrelation function algorithm to obtain a series of cyclic autocorrelation feature components.

[0066] S31. Use the power frequency modulation effect elimination model based on singular value decomposition for the single-phase motor stator current signal of the centrifugal pump to isolate the power frequency-dominated modulation component in the current signal.

[0067] The preprocessing is to construct a power frequency modulation effect elimination model based on singular value decomposition (SVD, singular value decomposition for short) to eliminate the modulation effect of power frequency on the fault features and isolate the power frequency-dominated modulation component in the original signal. The power frequency is 50 Hz.

[0068] S32. Use the cyclic autocorrelation function algorithm to extract cyclic autocorrelation features to obtain a series of cyclic autocorrelation feature components. The formula is as follows:

[0069] R xR(t,τ) = E{x(t + τ / 2)x * (t - τ / 2)}

[0070] where R x (t,τ) is the autocorrelation characteristic function; * represents the conjugate variable; t is time; τ is the time delay factor, and E{·} represents calculating the average.

[0071] The singular value decomposition (SVD) described in the present invention is a linear algebra method. After performing a fast Fourier transform, the single-phase motor stator current signal of the centrifugal pump is reconstructed into a new matrix A as follows:

[0072]

[0073] where X NM is the element in the Nth row and Mth column of matrix A. Each component in matrix A is decomposed into a series of orthogonal subspaces to reveal the signal characteristics, and the formula is as follows:

[0074]

[0075] where ∑ is a matrix, and the elements in matrix ∑ are arranged in descending order, that is, σi is greater than σi+1; k is the rank of the matrix; σi is the singular value of the matrix. Due to the presence of noise in matrix A, all diagonal elements are non-zero. Most of the energy corresponding to the power frequency is contained in the first term of the summation formula, which is also the main component of matrix A, thereby eliminating the modulation effect of the power frequency on the fault characteristics. Compared with traditional filters, this method preserves the integrity of the fault frequency band while eliminating the power frequency modulation of the single-phase motor stator current signal of the centrifugal pump.

[0076] In the case of damage to the impeller of the centrifugal pump unit on the deep-sea oil production platform, the single-phase motor stator current signal of the centrifugal pump collected by the deep-sea pressure-resistant optical fiber current sensing module has the characteristics of non-stationarity and periodicity. Therefore, aiming at the instantaneous inherent characteristics of dynamic rotational imbalance, the circular auto-correlation function algorithm (CAF algorithm for short) is used to extract its circular auto-correlation characteristics.

[0077] S4. The multi-fault feature extraction module decomposes the single-phase motor stator current signal of the centrifugal pump transmitted by the deep-sea pressure-resistant optical fiber current sensing module and preprocesses the signal. The multi-fault feature extraction module uses the empirical mode decomposition algorithm to obtain a series of rotor imbalance feature components;

[0078] S41. Apply the power frequency modulation effect elimination model based on singular value decomposition to the single-phase motor stator current signal of the centrifugal pump to isolate the modulation component dominated by the power frequency in the current signal. This step is the same as step S31.

[0079] S42. Use the empirical mode decomposition algorithm to decompose the preprocessed current signal, and extract a series of rotor imbalance characteristic components, that is, a series of IMF components as the characteristic indicators corresponding to the faults.

[0080] After performing the preprocessing operation on the stator current signal of the centrifugal pump single-phase motor, use the empirical mode decomposition algorithm to perform adaptive time-frequency analysis on the preprocessed current signal. The core of the empirical mode decomposition algorithm (Empirical Mode Decomposition, abbreviated as EMD algorithm) technology lies in decomposing the non-stationary signal into a finite number of intrinsic mode functions and a residual term. The EMD algorithm extracts the IMF through an iterative sifting process. Each IMF represents the characteristics of different frequencies and scales in the signal, and a series of rotor imbalance characteristic components can be extracted for the fault frequency characteristics of the centrifugal pump unit on the deep-sea oil production platform.

[0081] For the signal x(t), the sifting conditions for the EMD algorithm to iteratively screen and extract the IMF components are as follows:

[0082] (1) The number of extreme points is equal to or at most differs by 1 from the number of zero-crossing points;

[0083] (2) The mean value of the upper and lower envelope lines of the signal is zero.

[0084] The decomposition expression of the signal x(t) is as follows:

[0085]

[0086] In the formula, C i (t) is the i-th order IMF component, and r n (t) is the residual component.

[0087] For the fault frequency characteristics of the centrifugal pump unit on the deep-sea oil production platform, special attention can be paid to the entropy values of IMF1 and IMF4 as the characteristic indicators corresponding to the faults.

[0088] S5. The multi-fault feature extraction module decomposes the stator current signal of the centrifugal pump single-phase motor transmitted by the deep-sea pressure-resistant optical fiber current sensing module and preprocesses the signal. The multi-fault feature extraction module solves the root mean square value to obtain the energy-related characteristic components;

[0089] S51. Use the power frequency modulation effect elimination model based on singular value decomposition for the stator current signal of the centrifugal pump single-phase motor to isolate the power frequency-dominated modulation components in the current signal. This step is the same as steps S31 and S41.

[0090] S52. Solve the root mean square value to obtain the energy-related characteristic components. The formula is as follows:

[0091]

[0092] The above formula is the root mean square value of the single-phase motor stator current signal of the centrifugal pump within the time period 0 - T.

[0093] The energy-related characteristics of the motor stator current signal are analyzed in the time domain. The root mean square value (RMS value) gradually increases as the signal instability intensifies due to the impeller damage of the centrifugal pump, and can be used as an energy-related feature component to quantify the energy content in the signal.

[0094] S6. The multi-fault feature extraction module constructs an SCBA model dataset composed of cyclic autocorrelation feature components, rotor imbalance feature components, and energy-related feature components, and sends the SCBA model dataset to the data storage unit for storage. At the same time, the multi-fault feature extraction module sends the feature indicators corresponding to the faults to the fault intelligent recognition module.

[0095] S7. The data storage unit sends the SCBA model dataset to the fault intelligent recognition module. The SCBA model built into the fault intelligent recognition module trains the SCBA model dataset and outputs the trained SCBA model.

[0096] S71. Based on the sand cat swarm optimization algorithm (SCSO) and the convolutional neural network - bidirectional long short-term memory network - attention mechanism model (CNN - BiLSTM - Attention), a multi-modal hyperparameter collaborative optimization framework is constructed, and the formula is as follows:

[0097]

[0098] Among them, η is the learning rate of the model, h is the optimal number of hidden layer nodes, and l is the optimal L2 regularization coefficient.

[0099] S72. The sand cat swarm optimization algorithm (SCSO) and the SCBA model training set are used to optimize the multi-modal hyperparameter collaborative optimization framework to obtain the trained SCBA model. The specific steps are as follows:

[0100] S721. When |R| > 1, the sand cat is in the exploration stage, and the sand cat searches for prey, and the formula is as follows:

[0101]

[0102] r = r G ×rand(0, 1)

[0103] Among them, R is a parameter to judge which stage the sand cat is in; r G represents the auditory range of the sand cat; s M is to simulate the hearing of the sand cat, and its value is set to 2, iter c is the current iteration, iter maxis the maximum iteration; r is the sand cat position update parameter.

[0104] The search of the sand cat for prey depends on the release of low-frequency noise. It is assumed that the sensitivity range of the sand cat is from 0 - 2 kHz.

[0105] The sand cat position update is controlled by the following formula:

[0106] Pos(t + 1) = r·(Pos bc (t) - rand(0, 1)·Pos c (t))

[0107] where Pos bc represents the best candidate position, Pos c is the current position, and each sand cat updates its position according to the best candidate position Pos bc and the current position Pos c and its sensitivity range r. Therefore, the sand cat can find other possible best prey positions.

[0108] S722. When |R| ≤ 1, the sand cat enters the exploitation stage;

[0109] S723. Use the best position Pos bc and the current position Pos c to generate a random position. The judgment formula for the sand cat to attack the prey is as follows:

[0110] Pos rnd = |rand(0, 1)·Pos b (t) - Pos c (t)|

[0111]

[0112] Assume that the sensitivity range of the sand cat is a circle. Use the roulette wheel method to randomly select an angle θ for each sand cat, and finally implement the attack on the prey through the above formula. The random position can ensure that the sand cat approaches the prey, and the random angle can avoid the algorithm falling into a local optimum.

[0113] The fault intelligent identification module is based on the SCSO algorithm. Through the dynamic inertia weight and neighborhood search strategy, it realizes the global-local collaborative optimization of the hyperparameter space. Set the three hyperparameters of the model based on the adaptive strategy of the dynamic inertia weight: learning rate η, the best number of hidden layer nodes h, and the best L2 regularization coefficient l.

[0114] The search space can be expressed as (η, h, l). The SCSO algorithm can effectively search in this high-dimensional space.

[0115] Among them, the larger the learning rate η of the model, the faster the convergence in the initial stage of acceleration. However, at the same time, the model is prone to loss oscillation or divergence, and even unable to converge. The smaller the learning rate η of the model, the more refined the parameter update, and it may find a better solution. However, at the same time, the model is prone to too slow convergence speed and falling into local optimum.

[0116] Among them, the number of hidden layer nodes h of the model determines the complexity of the model. The more the number of hidden layer nodes h of the model, the higher the model capacity can be improved to fit complex functions. However, at the same time, it is easy to have the problem of overfitting the training data and poor generalization ability. The smaller the number of hidden layer nodes h of the model, the lower the computational cost and the risk of overfitting can be reduced. However, at the same time, it is easy to have the phenomenon of underfitting and unable to capture data features.

[0117] Among them, the larger the optimal L2 regularization coefficient l of the model, the lower the model complexity and the overfitting phenomenon can be alleviated. However, at the same time, it is easy to have the problem of underfitting of the model. The smaller the optimal L2 regularization coefficient l of the model, the better the generalization ability can be improved and the weight distribution can be made smoother. However, at the same time, the regularization effect is weak. In actual work, the inflection point where the verification loss first decreases and then increases is often verified as the optimal value.

[0118] The SCBA model is established based on the sand cat swarm optimization algorithm SCSO and the convolutional neural network - bidirectional long short-term memory network - attention mechanism model CNN - BiLSTM - Attention. The sand cat swarm optimization algorithm SCSO can adapt to complex and high-dimensional problems. Its efficient exploration and development mechanism, adaptive parameter adjustment, ability to handle non-linear and multi-modal problems, strong adaptability and robustness, and ability to handle high-dimensional search spaces make the sand cat swarm optimization algorithm SCSO have significant advantages in neural network optimization.

[0119] As Figure 4 shown, the SCBA model described in the present invention is composed of a sequence input layer (Sequence Input Layer), a sequence folding layer, a convolutional layer, a ReLU activation layer, a sequence unfolding layer, a BiLSTM layer, an Attention layer, a fully connected layer, and a Softmax layer.

[0120] The sequence folding layer folds the input sequence data and converts the one-dimensional sequence data into a two-dimensional matrix form suitable for processing by the convolutional layer.

[0121] The convolutional layer uses a 3×1 convolutional kernel for two-dimensional convolutional operation to extract features in the sequence.

[0122] The ReLU activation layer applies the ReLU activation function to introduce non-linearity and increase the non-linear characteristics of the network, so that the model can learn and represent more complex functions.

[0123] The sequence expansion layer expands the data after convolution and activation operations back into a sequence form for subsequent processing.

[0124] The BiLSTM layer uses a bidirectional long short-term memory network to process sequence data, which helps capture long-term dependencies in the sequence.

[0125] The Attention layer introduces an attention mechanism that can enhance the network's attention to important parts of the input data, improving the model's discriminative ability by assigning different weights to different input parts.

[0126] The Softmax layer applies the Softmax activation function to convert the output into a probability distribution for the classification of centrifugal pump impeller breakage faults.

[0127] S8. The fault intelligent recognition module uses the trained SCBA model to monitor the impeller breakage state for the characteristic indicators corresponding to the faults sent by the multi-fault feature extraction module, and obtains the monitoring results.

[0128] S9. The monitoring results obtained in step S8 are sent by the fault intelligent recognition module to the data storage unit for storage.

[0129] S10. The monitoring results obtained in step S8 are sent by the fault intelligent recognition module to the display unit for display, providing the user with the operation information of the centrifugal pump unit on the deep-sea oil production platform.

[0130] The SCBA model described in the present invention can also be referred to as an impeller breakage data model. Therefore, the SCBA model dataset can also be referred to as an impeller breakage dataset.

[0131] In addition to the above embodiments, the present invention can also have other implementation manners. Any technical solutions formed by equivalent replacement or equivalent transformation fall within the protection scope required by the present invention.

Claims

1. A deep-sea oil production platform centrifugal pump unit fault monitoring system, characterized in that: include Deep-sea compression-resistant optical fiber current sensing module, used to collect current signals in real time and transmit them to the multi-modal signal processing host; The multimodal signal processing host is connected to the deep-sea compressive optical fiber current sensing module, the data storage unit, the human-computer interaction module and the power module, and is used to receive and pre-process the current signal, and then perform feature extraction to form an SCBA data set to train the SCBA model and obtain monitoring results; A data storage unit connected to the human-computer interaction module and used for sending and storing the SCBA model data set; A human-computer interaction module, connected to the power module and the data storage unit, for displaying monitoring results and inputting the rotation speed of the centrifugal pump; The power module is connected to the deep-sea pressure-resistant optical fiber current sensing module and the human-computer interaction module and provides electrical energy.

2. The deep-sea oil production platform centrifugal pump unit fault monitoring system according to claim 1 is characterized by: The multimodal signal processing host includes a multi-factor fault feature extraction module and a fault intelligent identification module; the multi-factor fault feature extraction module is respectively connected to the power supply module, the fault intelligent identification module, the deep-sea compressive optical fiber current sensing module and the data storage unit, and the fault intelligent identification module is also connected to the power supply module, the human-computer interaction module and the data storage module.

3. The deep-sea oil production platform centrifugal pump unit fault monitoring system according to claim 2 is characterized in that: The human-computer interaction module includes a display unit and a touch control unit; the display unit is respectively connected to the multimodal signal processing host, the data storage unit and the touch control unit, and the touch control unit is respectively connected to the data storage unit and the power module.

4. The deep-sea oil production platform centrifugal pump unit fault monitoring system according to claim 3 is characterized by: The deep-sea pressure-resistant optical fiber current sensing module comprises a titanium alloy sealed cabin (6) and a deep-sea pressure-resistant optical fiber current sensor (4) encapsulated inside the titanium alloy sealed cabin (6); the two ends of the titanium alloy sealed cabin (6) are matched by threaded fasteners (1) and threaded engagement structures (2), and are sealed to form a closed cavity; the live wire insulation layer (3) of the pump group cable passes through the titanium alloy sealed cabin (6), and the contact portion between the live wire insulation layer (3) of the pump group cable and the titanium alloy sealed cabin (6) is sealed; the deep-sea pressure-resistant optical fiber current sensor (4) is non-invasively wound on the live wire insulation layer (3) of the pump group cable through an annular Faraday sensing optical path.

5. The deep-sea oil production platform centrifugal pump unit fault monitoring system according to claim 4 is characterized in that: The pressure resistance level of the titanium alloy sealed cabin (6) is ≥50 MPa.

6. The deep-sea oil production platform centrifugal pump unit fault monitoring system according to claim 4, characterized in that: The touch control unit is an LCD touch control module.

7. A monitoring method applicable to the deep-sea oil production platform centrifugal pump unit fault monitoring system according to any one of claims 4 to 6, characterized in that: The following steps are included: S1. The user sets the speed n of the centrifugal pump by touching the control unit and sends it to the multivariate fault feature extraction module through the data storage unit; S2, deep-sea pressure-resistant optical fiber current sensing module collects the stator current signal of the centrifugal pump single-phase motor in real time and sends it to the multi-factor fault feature extraction module; S3, the multivariate fault feature extraction module decomposes the centrifugal pump single-phase motor stator current signal transmitted by the deep-sea pressure-resistant optical fiber current sensor module and pre-processes the signal. The multivariate fault feature extraction module uses a cyclic autocorrelation function algorithm to obtain a series of cyclic autocorrelation feature components; S4, the multivariate fault feature extraction module decomposes the stator current signal of the centrifugal pump single-phase motor transmitted by the deep-sea pressure-resistant optical fiber current sensor module and pre-processes the signal. The multivariate fault feature extraction module uses the empirical mode decomposition algorithm to obtain a series of rotor imbalance feature components; S5, the multivariate fault feature extraction module decomposes the centrifugal pump single-phase motor stator current signal transmitted by the deep-sea pressure-resistant optical fiber current sensor module and pre-processes the signal, and the multivariate fault feature extraction module solves the root mean square value to obtain energy-related feature components; S6, the multivariate fault feature extraction module constructs a SCBA model data set consisting of a cyclic autocorrelation feature component, a rotor imbalance feature component, and an energy-related feature component, and sends the SCBA model data set to a data storage unit for storage. At the same time, the multivariate fault feature extraction module sends a feature index corresponding to the fault to the fault intelligent identification module; S7, the data storage unit sends the SCBA model data set to the fault intelligent identification module, and the SCBA model built in the fault intelligent identification module trains the SCBA model data set to obtain a trained SCBA model; S8, the fault intelligent identification module uses the trained SCBA model to monitor the impeller damage state based on the characteristic indicators corresponding to the fault sent by the multi-factor fault feature extraction module to obtain the monitoring results; S9, sending the monitoring result obtained in step S8 to the data storage unit for storage by the fault intelligent identification module; S10, sending the monitoring result obtained in step S8 to the display unit by the fault intelligent identification module for display, and providing the user with the operation information of the centrifugal pump unit of the deep-sea oil production platform.

8. The monitoring method of the deep-sea oil production platform centrifugal pump unit fault monitoring system according to claim 7, characterized in that: The specific steps of steps S3, S4 and S5 are as follows: S31. A power frequency modulation effect elimination model based on singular value decomposition is used for the stator current signal of a single-phase motor of a centrifugal pump to isolate the modulation component dominated by the power frequency in the current signal; S32, using a cyclic autocorrelation function algorithm to extract cyclic autocorrelation features, to obtain a series of cyclic autocorrelation feature components, the formula is as follows: R x (t,τ)=E{x(t+τ / 2)x * (t-τ / 2)} Among them, R x (t,τ) is the autocorrelation characteristic function; * indicates the conjugate variable; t is time; τ is the time delay factor, and E{·} indicates the calculated average; S41. A power frequency modulation effect elimination model based on singular value decomposition is used for the stator current signal of a single-phase motor of a centrifugal pump to isolate the modulation component dominated by the power frequency in the current signal; S42, using an empirical mode decomposition algorithm to decompose the preprocessed current signal and extract a series of rotor imbalance characteristic components; S51, using a power frequency modulation effect elimination model based on singular value decomposition to isolate the power frequency-dominated modulation component in the current signal for the stator current signal of the single-phase motor of the centrifugal pump. This step is the same as steps S31 and S41. S52. Solve the root mean square value to obtain the energy-related characteristic component. The formula is as follows: The above formula is the RMS value of the stator current signal of the centrifugal pump single-phase motor in the time period 0-T.

9. The monitoring method of the deep-sea oil production platform centrifugal pump unit fault monitoring system according to claim 8, characterized in that: The specific steps of step S7 are as follows: S71. Based on the sand cat swarm optimization algorithm and the convolutional neural network-bidirectional long short-term memory network-attention mechanism model, a multimodal hyperparameter collaborative optimization framework is constructed. The formula is as follows: Among them, η is the learning rate of the model, h is the optimal number of hidden layer nodes, and l is the optimal L2 regularization coefficient. S72. Use the sand cat swarm optimization algorithm and the SCBA model training set to optimize the multimodal hyperparameter collaborative optimization framework to obtain the trained SCBA model.

10. The monitoring method of the deep-sea oil production platform centrifugal pump unit fault monitoring system according to claim 9, characterized in that: The SCBA model consists of a sequence input layer, a sequence folding layer, a convolution layer, a ReLU activation layer, a sequence unfolding layer, a BiLSTM layer, an Attention layer, a fully connected layer and a Softmax layer.

Images