Method for condition analysis of bearing assembly for marine engineering cable pulling

Abstract

The application relates to a state analysis method of a bearing assembly for a marine engineering cable pulling device and relates to the technical field of marine engineering equipment state monitoring and analysis. First, a physical information neural network is used to solve bearing degradation physical quantities and residual life by embedding corrosion-abrasion coupling degradation physical constraints; meanwhile, vibration signals are adaptively decomposed and a dynamic function atlas is constructed, early degradation features are extracted by using a graph neural network; a mapping layer is constructed to convert the degradation features into compensation items of the physical information neural network; a generative adversarial network is used to generate an expanded training set with a small amount of target domain samples and to perform joint model fine-tuning; finally, the joint model is deployed on an edge end to perform real-time collaborative reasoning and output state evaluation results; the application has the advantages of strong physical consistency, sensitive early degradation perception and high robustness under a small sample condition and is suitable for degradation monitoring and life prediction of key components such as marine platform winch bearings and guide pulley bearings.

Description

Technical Field

[0001] This invention belongs to the field of marine engineering equipment condition monitoring and analysis technology, specifically relating to a condition analysis method for bearing assemblies used in marine engineering cable traction. Background Technology

[0002] Offshore platform cable traction bearing assemblies (such as winch bearings and guide pulley bearings) operate long-term in harsh environments characterized by high salt spray, high humidity, and continuous wave impact. The annual salt spray concentration can reach 3.5 mg / m³. 3 As described above, seawater splashing and condensation cause a corrosive thin layer to easily form on the bearing surface. Simultaneously, the platform continuously generates low-frequency oscillations (dominant frequency approximately 0.1~0.3Hz) under wave action, subjecting the bearing to alternating loads and slight oscillations, making mechanical wear inevitable. Corrosion roughens the surface, accelerating material peeling; wear, in turn, continuously removes the passivation layer, exposing fresh metal substrate and further intensifying corrosion. This strong positive feedback coupling between salt spray corrosion and mechanical wear constitutes the unique and complex degradation mechanism of this type of bearing.

[0003] Traditional condition monitoring methods primarily rely on vibration signal spectrum analysis. Faults are identified by capturing amplitude changes at bearing characteristic frequencies through Fast Fourier Transform (FFT) or envelope spectrum analysis of the vibration signals. However, in offshore platform conditions characterized by strong low-frequency wave vibrations, frequent equipment start-ups and shutdowns, and frequent load changes, environmental vibrations and degradation signals are severely mixed. In the early degradation stage, the energy of fault characteristic frequencies is extremely weak, easily masked or misjudged, resulting in very low early warning sensitivity.

[0004] In recent years, deep learning methods have been widely used in bearing fault diagnosis. These methods are driven by a large amount of labeled historical fault data, automatically learning the mapping from raw signals to fault categories. However, in the context of offshore platforms, due to extremely high requirements for continuous operation and very short maintenance windows, bearings are rarely replaced only after complete failure, resulting in a scarcity of real fault samples (possibly only a few records of minor anomalies before replacement). Purely data-driven models cannot be effectively trained under these conditions, easily producing predictions that violate physical laws, such as falsely reporting severe fracture when only minor wear occurs, leading to an excessively high false alarm rate and unnecessary downtime for inspection. Meanwhile, while existing generative adversarial network (GAN) data augmentation methods can synthesize fault samples, they typically use laboratory data for training directly, failing to consider the unique background noise and sensor characteristics of the target platform. This results in significant inter-domain distribution shifts in the generated samples, and direct use of these samples actually reduces model accuracy.

[0005] Therefore, there is an urgent need for a bearing condition analysis method that integrates degradation physical mechanisms, data-driven representation, and adaptive enhancement in a few-sample domain. This method should not only ensure the physical consistency of the prediction results, but also be able to keenly detect early degradation from strong noise, and complete model training and robust deployment even when the target platform has extremely scarce fault samples. Summary of the Invention

[0006] The purpose of this invention is to overcome the problems of poor physical consistency in corrosion and wear coupled degradation prediction, low sensitivity in early degradation perception, and difficulty in model training due to the scarcity of fault samples in the target domain in existing technologies, and to provide a state analysis method for bearing assemblies used in marine engineering cable traction.

[0007] The technical solution of the present invention is as follows:

[0008] A condition analysis method for bearing assemblies used in marine engineering cable traction includes the following steps:

[0009] Step S1: Obtain the physical and chemical monitoring data of the bearing, establish a coupled degradation model of salt spray corrosion and mechanical wear, construct a physical information neural network, embed the coupled degradation model into the loss function of the physical information neural network in the form of physical constraint terms, train the physical information neural network using the physical and chemical monitoring data, and output the predicted value of degradation physical quantity and remaining life.

[0010] Step S2: Obtain the vibration signal of the bearing, perform adaptive decomposition on the vibration signal to obtain multiple signal components, select the target component that is sensitive to degradation from the multiple signal components, construct a dynamic spectrum representing the coupling relationship between modes based on the target component, and extract degradation features from the dynamic spectrum using a graph neural network.

[0011] Step S3: Construct a mapping layer to map the degradation features to the model compensation terms of the physical information neural network. The output of the mapping layer is equipped with a physical boundary constraint unit to ensure that the parameter correction does not violate physical laws. The physical equation parameters of the physical information neural network are dynamically corrected using the model compensation terms.

[0012] Step S4: Construct a generative adversarial network, pre-train the generative adversarial network using source domain fault data, and perform domain adaptation on the generative adversarial network using a small number of abnormal samples from the target domain to generate expanded training samples in the style of the target domain.

[0013] Step S5: Use the expanded training samples to jointly fine-tune the joint model composed of the physical information neural network, graph neural network, and mapping layer;

[0014] Step S6: Deploy the fine-tuned joint model at the edge, and use the joint model to perform collaborative reasoning on the real-time collected physical and chemical monitoring data and vibration signals to output the bearing condition assessment results.

[0015] Furthermore, the coupled degradation model in step S1 includes a positive feedback coupling relationship between mechanical wear and salt spray corrosion; the predicted degradation physical quantities include predicted wear volume and predicted corrosion layer thickness.

[0016] Furthermore, the physicochemical monitoring data in step S1 includes oil ferromagnetic particle concentration data and bearing total clearance change data. The oil ferromagnetic particle concentration data is converted to obtain the measured wear volume value, and the bearing total clearance change data is converted to obtain the measured corrosion layer thickness value.

[0017] Furthermore, the physical constraints in step S1 include: a data loss term, used to constrain the consistency between the output of the physical information neural network and the measured data; a physical residual loss term, used to constrain the output of the physical information neural network to satisfy the coupling degradation model; and a boundary condition loss term, used to constrain the output of the physical information neural network to satisfy the physical initial state at the initial moment.

[0018] Furthermore, in step S2, the adaptive decomposition uses a particle swarm optimization algorithm to adaptively determine the decomposition parameters and then performs variational mode decomposition. The basis for selecting the target components sensitive to degradation is: calculating the dominant frequency and degradation sensitivity index of each signal component, removing signal components with a dominant frequency lower than a set threshold and a degradation sensitivity index lower than a set threshold as interference components, and retaining the rest as target components.

[0019] Furthermore, the method for constructing the dynamic spectrum in step S2 is as follows: each target component is used as a node, a node feature matrix is ​​constructed using the multidimensional time-frequency micro-features of each target component, an adjacency matrix is ​​constructed using the weighted sum of the phase synchronization index and coherence coefficient between each target component, and the node feature matrix and adjacency matrix are assembled into a frame of dynamic spectrum; a dynamic spectrum sequence is formed by sliding with the time window.

[0020] Furthermore, the multidimensional time-frequency micro-features include normalized energy, spectral centroid, sample entropy, kurtosis, and envelope spectral kurtosis; the graph neural network includes a graph convolutional layer, a graph state update layer based on gated recurrent units, a graph attention pooling layer, and a fully connected readout layer connected in sequence.

[0021] Furthermore, the physical boundary constraint unit in step S3 achieves parameter correction in the following way: the unconstrained output of the mapping layer is constrained within a finite interval by the hyperbolic tangent function, and then multiplied by the preset maximum allowable correction ratio coefficient to obtain the parameter correction amount; the total parameter value after correction is always greater than zero.

[0022] Furthermore, the generative adversarial network mentioned in step S4 is a domain-adaptive auxiliary classification Wasserstein generative adversarial network, whose discriminator includes a true / false discrimination branch, a fault category classification branch, and a domain classification branch; during the domain adaptation process, the domain classification branch is used to extract the noise distribution features and sensor frequency response features of the target domain, driving the generator to complete style transfer.

[0023] Furthermore, in the joint fine-tuning process described in step S5, two data flows are executed simultaneously for each training batch: the first flow inputs the time-frequency map of the expanded training samples into the graph neural network to extract degradation features; the second flow parses the corresponding physical degradation amount label from the expanded training samples and calculates the loss with the output of the physical information neural network; the two data flows maintain the same sample order in the batch dimension to ensure cross-modal label alignment; the state assessment results described in step S6 include remaining usable lifetime, wear amount, corrosion depth, and degradation stage.

[0024] Compared with the prior art, the beneficial effects of the present invention are:

[0025] 1. This invention embeds the differential equations of salt spray corrosion and mechanical wear into the PINN loss function in the form of physical residual constraints, restricting the network solution space to the physical feasible region. Even under the condition of zero real fault samples, the remaining lifetime prediction still follows objective physical laws, and the extrapolation reliability is much higher than that of pure data-driven methods.

[0026] 2. The present invention uses IVMD adaptive decomposition and screening of low-frequency wave modes. The graph neural network captures degradation precursors from microscopic changes in phase synchronization and coherence between sensitive modes. It is naturally robust to wave vibration and variable operating condition disturbances and can issue early warnings before the bearing enters the accelerated degradation stage.

[0027] 3. The domain adaptive AC-WGAN-GP of the present invention requires only a very small number (5~8 images) of abnormal state samples of the target platform. It can transfer rich laboratory fault data to generate a realistic target domain fault time-frequency map through domain adversarial training, and maintains strict alignment of cross-modal degradation labels through physical state isomorphic mapping.

[0028] 4. The online inference process only involves forward propagation collaborative computation of each sub-network, without any backward iterative optimization or adversarial generation. Actual measurements show that the overall time for a single state assessment can be as low as 62ms (ensuring completion within 65ms), fully meeting the online dynamic real-time monitoring needs of marine platforms under complex and variable operating conditions.

[0029] 5. PINN provides physical feasibility constraints, GNN extracts early weak precursors from vibration mode coupling, and the two are deeply coupled through a physical boundary compensation layer. GAN solves the problem of data scarcity. The three are not simply spliced ​​together, but form a positive feedback loop of "physical constraints - sensitive perception - data completion". The combined effect far exceeds the sum of the performance of each module used individually. Attached Figure Description

[0030] Figure 1 This is a flowchart of the method of the present invention. Detailed Implementation

[0031] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0032] like Figure 1 As shown in the figure, this embodiment provides a state analysis method for bearing assemblies used for cable traction in marine engineering, including the following contents.

[0033] Step S1: Construct a physical information neural network for corrosion-wear coupled degradation prediction.

[0034] This embodiment first establishes a coupled degradation model of salt spray corrosion and mechanical wear. Bearing degradation is caused by the combined effects of mechanical wear and salt spray corrosion, which exhibit a positive feedback coupling effect: corrosion roughens the surface, thus accelerating wear, and wear strips away the surface passivation layer, thus accelerating corrosion. The mathematical expression of this coupled degradation model is as follows:

[0035] ;

[0036] ;

[0037] In the formula: Indicates wear volume, unit: ; This represents the normalized value of the corrosion layer thickness, dimensionless, with a range of values. , This indicates that the surface is completely covered by a corrosion layer; Indicates contact load, in units of ; Represents sliding speed, in units of ; express Salt spray concentration time series at time, in units of ; The wear coefficient is... The corrosion rate coefficient is... For load index, For speed index, The coefficient of acceleration of corrosion on wear. The wear coefficient is the coefficient of corrosion.

[0038] in , , , , , The initial value was determined using the following calibration method: Standard pin disc specimens were prepared using the same bearing material (GCr15) on a friction and wear testing machine. Accelerated tests were conducted simulating pure wear conditions (no salt spray) and pure corrosion conditions (no contact load). The wear volume and corrosion layer thickness were measured using a weighing method and a surface profilometer. Parameter fitting was performed on the ODE to obtain the initial value in this embodiment: , , , , , .

[0039] Construct a Physical Information Neural Network (PINN), where the backbone of the PINN adopts a fully connected feedforward neural network, and the input layer has 4 neurons, corresponding to the input vector. The hidden layers consist of 6 layers, each with 128 neurons, using the Swish function as the activation function; the output layer has 2 neurons, outputting... , where represents the predicted wear volume and the predicted corrosion layer thickness normalized value, respectively.

[0040] PINN's total loss function It consists of a weighted sum of three terms:

[0041] ;

[0042] in, Data loss is defined as the mean square error (MSE) between the output value of the physical information neural network and the actual monitoring data.

[0043] ;

[0044] The concentration of ferromagnetic particles in the oil monitored online was converted into the measured value of wear volume. The specific calculation formula is as follows:

[0045] ;

[0046] In the formula, for The mass concentration of ferromagnetic particles measured at any time, This refers to the lubricating oil circulation flow rate. For the density of lubricating oil, For the density of bearing steel, This represents the oil sensor's capture efficiency coefficient. This indicates that the time variable τ is from Integral summation is performed up to t.

[0047] The oil sensor capture efficiency coefficient The system calibration constants are used to reflect the flow characteristics of the pipeline, the flow rate of the lubricating oil, and the geometric capture capability of the sensor measuring chamber.

[0048] In this embodiment, the coefficient is determined through an offline benchmark comparison calibration experiment: before being put into operation or during the standard maintenance window, the bearing assembly is run for 100 hours under preset rated operating conditions, the concentration is recorded, and after the operation, the return oil pipeline is thoroughly cleaned to collect all precipitated ferromagnetic particles. The true absolute value of physical wear volume is reconstructed using a high-precision ferrography analyzer. ,Will The capture efficiency coefficient is calculated by comparing the total wear amount calculated from the online sensor integration during that period with the theoretical total wear amount. (Its empirical value usually ranges from 0.85 to 0.95).

[0049] The total bearing clearance variation was estimated based on the measured corrosion layer thickness. The specific method is: change in total bearing clearance. Radial thinning caused by mechanical wear It is composed of the geometric change caused by the change in the thickness of the corrosion layer.

[0050] First, based on the calculated wear volume... Based on the bearing raceway contact geometry, a geometric transformation function is used. Calculate the radial thinning caused by wear Furthermore, the amount of interstitial change caused by pure corrosion can be separated. Finally, the normalized measured value of the corrosion layer thickness was obtained:

[0051] ;

[0052] In the formula, This is the preset maximum allowable critical corrosion depth.

[0053] The measured value of wear volume Converted to radial thinning caused by wear Geometric transformation function Its specific mathematical formula is constructed based on the simplified assumption of uniform wear on the raceway contact surface, and the specific calculation formula is as follows:

[0054] ;

[0055] In the formula: This represents the total number of rolling elements in the bearing. The pitch circle radius of the bearing (in units of...) ); Effective contact neck width between rolling elements and inner / outer raceways (unit: ); This is the nominal initial contact angle (in radians) of the bearing. For the bearing model (SKF 22220 EK) in this embodiment, the above parameters are all known manufacturing geometric constants.

[0056] Among them, the change in total bearing clearance Displacement is acquired in real time using an eddy current displacement sensor mounted radially on the bearing housing. The sensor probe is fixed to the bearing housing housing, with its target surface facing the outer wall of the bearing outer ring, and outputs a displacement signal. The initial displacement calibration value of the sensor is recorded when the bearing assembly is newly installed and not yet in operation. During service, the displacement value monitored in real time by the sensor is... The total change in gap at the current moment is calculated using the following formula: ;

[0057] It should be noted that the above geometric transformation function is based on the assumption of uniform wear. If non-uniform wear such as localized pitting or edge wear occurs in actual operation, the function can be adjusted according to the actual contact stress distribution. Weighted corrections are applied, but in this embodiment, the bearing load is relatively stable, and the assumption of uniform wear is sufficient to meet the engineering accuracy requirements.

[0058] To address the physical residual loss, an automatic differentiation mechanism is used to evaluate the output of the physical information neural network. and Find the time relationship respectively The derivative of the physical information neural network is used to force the output of the physical information to satisfy the degenerate physics:

[0059] ;

[0060] For boundary condition loss, forced and At the initial moment Satisfies the physical initial state:

[0061] ;

[0062] It should be noted that the total loss function weight , , Employing a dynamic adjustment strategy: In the initial stage of training , Larger values ​​allow the network to prioritize learning physical laws and boundary conditions; in the later stages of training... Gradually increase the size of the network to better fit the measured data.

[0063] In this embodiment, the specific hyperparameters for PINN training are: the optimizer uses Adam, and the initial learning rate is... The training time decreases to 0.9 times every 100 epochs, with a total of 500 training epochs and a batch size of 64.

[0064] The dynamic weighting strategy is: take the first 200 rounds... , , ; the last 300 rounds , , .

[0065] Remaining lifetime (RUL) is calculated using the following formula:

[0066] ;

[0067] in, This is the maximum allowable wear volume threshold for the bearing, which is preset based on the bearing model, rated dynamic load, and on-site operating conditions.

[0068] Step S2: Adaptive Variational Mode Decomposition and Dynamic Functional Graph Feature Extraction

[0069] The Particle Swarm Optimization (PSO) algorithm is used, with the fitness function being the minimization of the average envelope entropy of each mode, to adaptively search for the optimal parameter set. .

[0070] The PSO algorithm is specifically configured as follows: particle swarm size 20, maximum number of iterations 30, inertia weight decreasing linearly from 0.9 to 0.4, and acceleration constant... Search scope , .

[0071] Then the vibration signal Perform variational mode decomposition to obtain Each intrinsic mode function (IMF) is denoted as . .

[0072] For any eigenmode function component obtained from demodulation decomposition First, the power spectral density (PSD) function is calculated using the periodogram method. Then, the global maximum value of the power spectral density function is searched on the frequency axis, and the frequency point corresponding to the maximum power spectral density function is defined as the dominant frequency of the IMF component. :

[0073] ;

[0074] The amplitude of the envelope spectrum of each IMF at the bearing inner ring fault characteristic frequency BPFI, outer ring fault characteristic frequency BPFO, rolling element fault characteristic frequency BSF and its first three harmonics is calculated, and the proportion of the amplitude of the envelope spectrum of the IMF component to the total amplitude of the entire frequency band is used as the degradation sensitivity index.

[0075] If the degradation sensitivity index of a certain IMF component is lower than 0.15, and the dominant frequency calculated by the above formula is... Physically, the component is determined to be an interference mode dominated by low-frequency swaying caused by ocean waves or rigid body vibration and is therefore removed; the remaining components are retained as sensitive modes, for a total of [number missing] retained. One, recorded as .

[0076] To capture the time-varying characteristics of the coupling relationships between sensitive modes, the sensitive modes and their interrelationships within each time window are organized into a dynamic functional graph. The evolution of the graph structure is used to characterize the micro-dynamics of the degradation process. The specific generation rules and time axis alignment methods are as follows:

[0077] The online vibration signal is continuously divided into overlapping time windows, each 10 seconds long, with a sliding step of 5 seconds between adjacent windows. For any given 10-second time window, the following calculations are performed in parallel within that same window:

[0078] Node feature construction: Each retained sensitive mode is considered as a graph. Each node is used to extract the sensitive modal components within the current 10 seconds of data, calculate their 5-dimensional time-frequency micro-features within the current window, and thus construct the node feature matrix for the current window. (dimension) ).

[0079] The 5-dimensional time-frequency micro-features include dimension 1: normalized energy; dimension 2: spectral centroid; and dimension 3: sample entropy, which measures signal complexity according to the embedding dimension. Similarity tolerance Standard deviation condition calculation; Dimension 4: Kurtosis; Dimension 5: Envelope spectrum kurtosis.

[0080] Adjacency matrix construction: Using only the waveforms of each sensitive modal component within the current 10-second data window, calculate the pairwise phase synchronization index (PLV, extracted by Hilbert transform and phase difference consistency) and coherence coefficient (Coh, the average coherence value within the bearing characteristic frequency band of 10Hz~2000Hz). Obtain the adjacency matrix for the current window using a weighted sum formula. :

[0081] ;

[0082] In the formula, the weighting coefficients satisfy By performing a grid search optimization on the validation set, the value of the fixed constant was determined. , .

[0083] The node feature matrix calculated in the current window and adjacency matrix Assembly, i.e., generating an independent function graph frame corresponding to the current window. The timestamp of this function graph frame. The timeframe is strongly bound to the end of the current 10-second time window. As the time window slides forward in 5-second increments, a continuous sequence of dynamic function graphs is generated. .

[0084] To automatically learn the temporal evolution of intermodal coupling relationships from dynamic function graph sequences and extract low-dimensional feature vectors sensitive to degradation states, this invention uses each frame of the function graph... The input graph neural network is subjected to hierarchical encoding and pooling.

[0085] The structure of a Graph Neural Network (GNN) is as follows: The first layer is a graph convolutional layer (GCNConv) with 5 input channels and 32 output channels; the second layer is a graph state update layer based on a gated recurrent unit (GRU) with 64 hidden states, used to capture the temporal evolution of modal coupling relationships; the third layer is a graph attention pooling layer, which performs attention-weighted pooling on the output node features to obtain a 64-dimensional graph-level vector; the fourth layer is a fully connected readout layer with 64-dimensional input and 16-dimensional output, using ReLU activation function to obtain the final early degradation feature vector. .

[0086] Step S3: Feature fusion compensation based on physical boundary operators

[0087] Construct a fully connected mapping layer, with the early degradation feature vector obtained in step S2 as input. Output 2D parameter correction amount .

[0088] To ensure that the compensation term output by the fully connected mapping layer does not violate the physical feasible region of PINN, an activation operator based on physical boundary constraints is set at the output of the fully connected mapping layer, with the specific formula as follows:

[0089] ;

[0090] ;

[0091] In the formula, and These are the initial phenomenological physical constants in the coupled degradation model in step S1; , This represents the weight matrix of the fully connected mapping layer. , This represents the bias vector of the fully connected mapping layer. Limit the unconstrained output of the fully connected layer to between; and The maximum allowable correction ratio (in this embodiment, the value is 0.2, meaning the parameter is allowed to be within...) (fluctuation within a range).

[0092] By activating the operator at this boundary, the corrected total wear coefficient is ensured. and total corrosion acceleration coefficient Strictly always greater than zero, the parameters of the physical equations are dynamically corrected during the forward propagation of the physical information neural network, eliminating the risk that negative parameters in the network output will cause the physical laws to fail.

[0093] Step S4: Sample generation enhancement based on few-shot domain adaptive adversarial network

[0094] A domain-adaptive auxiliary classification Wasserstein generative adversarial network (AC-WGAN-GP) is constructed. The goal of this network is to transfer the time-frequency map of faults in the laboratory source domain to the style of the target domain, in order to address the problem of extremely scarce fault samples for the target platform.

[0095] AC-WGAN-GP employs an adversarial training approach between a generator and a discriminator to achieve domain adaptation. The generator produces time-frequency maps of the target domain style, while the discriminator identifies genuine and fake data, classifies fault types, and distinguishes domain affiliation. The specific structure is as follows:

[0096] The generator G takes a 100-dimensional random noise vector as input. With fault category labels The concatenation of 100-dimensional noise and class labels is progressively upsampled through a series of transposed convolutional layers to output a 224×224×3 time-frequency map. The generator's specific architecture is as follows: the input 100-dimensional noise is concatenated with the class labels and then mapped to a 4×4×512 feature map through a fully connected layer. Subsequently, it passes through 5 transposed convolutional blocks (upsampled to 8×8, 16×16, 32×32, 64×64, 128×128, and 224×224 respectively). Each transposed convolutional block is followed by a batch normalization layer and a ReLU activation function.

[0097] The discriminator D has three outputs: the first output is 1D, used for Wasserstein distance authenticity determination; the second output is a softmax of the category dimension, used to assist in fault category classification; and the third output is a 2D softmax, used for domain classification (source domain / target domain), which is enabled during the target domain adaptation phase.

[0098] The discriminator uses a 5-layer convolutional neural network with channels of 64, 128, 256, 512, and 512 respectively. The kernel size is 4×4, the stride is 2, and the end outputs three fully connected outputs.

[0099] In the pre-training phase, a large number of labeled fault time-frequency images collected from a laboratory bearing fault simulation test bench were used as the source domain dataset to train AC-WGAN-GP. In the target domain adaptation phase, 5-8 abnormal state time-frequency images collected on the target platform were introduced, and the domain classification branch was enabled to perform domain adversarial training.

[0100] Based on the aforementioned AC-WGAN-GP structure, the generator and discriminator are jointly trained through a triple game involving Wasserstein adversarial testing, auxiliary classification, and domain adversarial testing. During this process, the discriminator's domain classification branch forces the generator to extract first- and second-order statistical distribution features of the target domain (such as environmental noise power spectral density and sensor transfer function amplitude-frequency response), rather than simply memorizing fault patterns. Therefore, only 5–8 abnormal state samples from the target domain are needed to drive the generator to complete style transfer.

[0101] When the generated samples are saved, their filenames directly inherit the physical state fields (wear volume W and corrosion depth C) of the source domain image. Therefore, each generated target domain fault time-frequency map comes with a physical degradation label and can be directly used for cross-modal joint fine-tuning in step S5.

[0102] Step S5: Cross-modal label alignment and fine-tuning of end-to-end joint training

[0103] The physical information neural network constructed in step S1, the graph neural network constructed in step S2, and the fully connected mapping layer constructed in step S3 are combined into a joint model. Using the target domain fault time-frequency map generated in step S4 and the physical degradation label carried in its file name, a cross-modal balanced training set is constructed, and the joint model is fine-tuned end-to-end.

[0104] During training, two data flows are executed simultaneously in each batch:

[0105] Data flow to A (graph network branch): The pixel matrix of the time-frequency graph is input into the graph neural network to extract early degradation feature vectors. ;

[0106] Data flow to B (physical constraint branch): The W and C field values ​​parsed from the same time-frequency plot file name are used as the physical degradation label for that sample. , and the predicted value of the PINN forward propagation output Calculate the mean squared error loss.

[0107] The two data streams maintain the same sample order in the batch dimension, ensuring that the vibration time-frequency features and physical degradation labels are strictly aligned during training.

[0108] The joint model was fine-tuned for 20 epochs using the aforementioned cross-modal balanced training set, and the joint loss function also included the PINN total loss. Graph network classification cross-entropy loss and The regularization term optimizes the parameters of the three modules synchronously through end-to-end backpropagation.

[0109] Step S6: Online Edge Inference and Intelligent State Assessment

[0110] The finely tuned joint model was exported and deployed on the NVIDIA Jetson Orin edge computing platform of the marine platform. During online inference, only forward computation needs to be performed, with data from three sensors being acquired in parallel as input.

[0111] Vibration signals acquired by vibration sensors are fed into an online IVMD module in real time. Sensitive modes are extracted and a dynamic function graph sequence is constructed. Early degradation feature vectors are then obtained through graph neural network feature extraction. Then it is converted into parameter correction values ​​through a fully connected mapping layer. Inject physical information into the neural network.

[0112] The data collected by the online ferromagnetic particle concentration sensor and the eddy current displacement sensor are converted into measured wear volume values ​​using the integral formula and gap separation method described in step S1, respectively. Measured values ​​of corrosion layer thickness Input physical information neural network.

[0113] The physical information neural network uses the corrected parameters to predict values. and Perform collaborative forward propagation reasoning.

[0114] The actual test conditions for the overall time consumption of a single state assessment were: NVIDIA Jetson Orin 64GB version, CPU frequency 2.2GHz, GPU frequency 1GHz, vibration signal sampling rate 25.6kHz, and window length 10 seconds (i.e., 256,000 sampling points). Under these conditions, the overall time consumption of a single assessment can be as low as 62ms, which can fully meet the requirements of online dynamic real-time monitoring.

[0115] The system interface displays in real time: current remaining usable lifespan (hours), cumulative predicted wear amount ( ), Predicted corrosion depth percentage (%), and degradation stage (normal / early warning / dangerous);

[0116] The rule for determining the degradation stage is as follows: when and The time was judged as normal;

[0117] When 0.3≤ <0.8 or 0.3≤ A value less than 0.8 is considered a warning.

[0118] when ≥0.8 or A value ≥0.8 is considered dangerous.

[0119] Example Effect Verification

[0120] To verify the superiority of the method of the present invention, the main traction bearing of a cable winch of a certain offshore oil drilling platform (model SKF 22220 EK, capable of withstanding a traction load of approximately 50kN, a speed range of 200~800rpm, and an annual average salt spray concentration of approximately 3.5mg / m) was used. 3 Historical operational lifecycle data (with a dominant wave frequency of approximately 0.1~0.3Hz) were used for retrospective comparative testing. The specific implementation details and parameter settings of the three existing technologies involved in the comparative test are disclosed below:

[0121] The envelope spectrum thresholding method is implemented as follows: First, the original vibration signal is bandpass filtered. Based on the finite element modal analysis of the SKF 22220EK bearing, its first natural frequency is approximately 3.8kHz. Therefore, the filtering frequency band is selected in the bearing resonance high-frequency region (3kHz~5kHz) to eliminate vibration interference from the low-frequency environment of ocean waves. Next, the filtered signal is subjected to Hilbert transform to achieve envelope demodulation, and a fast Fourier transform is performed to obtain the envelope spectrum. The spectral amplitudes at the bearing inner ring characteristic frequency BPFI, outer ring characteristic frequency BPFO, and rolling element characteristic frequency BSF are accurately captured and extracted from the envelope spectrum. The alarm threshold is determined using 30 days of data collected during the platform's initial normal operation. The method is based on statistical principles. An anomaly warning is triggered when the envelope amplitude at any characteristic frequency exceeds a threshold five times consecutively. However, because this method cannot establish a time-series evolution model, it cannot provide continuous predictions of remaining usable lifetime.

[0122] The pure data-driven CNN (one-dimensional ResNet18) is implemented by directly using raw one-dimensional temporal vibration signals collected by sensors as network input, with each sample set having a length of 2048 sampling points. The network adopts the standard one-dimensional ResNet18 architecture, including one input convolutional layer, eight one-dimensional residual blocks, and one global average pooling layer. The network ends with two parallel output heads: a classification head using Softmax to output the fault category, and a regression head using linear activation to output the RUL. To achieve basic generalization diagnostic performance, the model training phase must use at least 200 sets of labeled fault samples that have been run on the target platform until failure throughout their entire lifespan. Due to the lack of physical mechanism constraints, its extrapolated prediction values ​​are highly susceptible to severe non-physical monotonic oscillations caused by sudden environmental changes.

[0123] The VMD+Random Forest implementation specifically uses fixed empirical parameters (preset number of modes). Punishment factor A conventional variational mode decomposition (VMD) is performed on the one-dimensional vibration signal. Subsequently, five temporal macroscopic statistical features—root mean square (RMS), peak factor, kurtosis, skewness, and waveform factor—are calculated for each IMF component and combined into a 25-dimensional feature vector, which is then input into the downstream model. Downstream, a random forest regressor and classifier composed of 100 decision trees are constructed, with a maximum tree depth of 12. However, this method suffers from poor lifetime prediction accuracy due to its fixed VMD parameters, inability to adaptively isolate mode aliasing caused by transient wave impacts, and the lack of temporal evolution representation capabilities in the random forest.

[0124] Parallel backtesting was conducted using historical full lifecycle service data of the platform (actual replacement time was taken as the true life value). The comprehensive performance evaluation results of the whole method are shown in Table 1.

[0125] method RUL predicts RMSE (hours) Fault classification accuracy (%) False alarm rate (%) Number of real fault samples required in the target domain (images) Envelope spectral thresholding Unable to predict continuously 72.3 18.5 0 (Depends on expert rules) Pure data-driven CNN 168.4 89.1 12.7 ≥200 VMD+ Random Forest 132.7 92.5 9.3 ≥100 Method of the present invention 28.6 98.7 2.1 6

[0126] As shown in Table 1, the test results of this invention reduce the root mean square error (RMSE) of remaining lifetime prediction to 28.6 hours under the stringent condition of extremely scarce real fault samples in the target domain (requiring only 6 abnormal samples), which is 4 to 6 times more accurate than pure data-driven and fixed feature extraction methods. Simultaneously, the fault classification accuracy reaches 98.7%, and the false alarm rate is significantly reduced to 2.1%. The predicted trajectory comparison also clearly demonstrates that the degradation prediction trajectory given by this invention strictly satisfies temporal monotonicity and mechanistic causality throughout the entire life cycle. This effectively solves the monitoring challenges of marine engineering cable traction bearing assemblies caused by strong coupling of corrosion and wear and the scarcity of samples, demonstrating extremely high engineering application value.

[0127] Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for condition analysis of bearing assemblies used in marine engineering cable traction, characterized in that: Includes the following steps: Step S1: Obtain the physical and chemical monitoring data of the bearing, establish a coupled degradation model of salt spray corrosion and mechanical wear, construct a physical information neural network, embed the coupled degradation model into the loss function of the physical information neural network in the form of physical constraint terms, train the physical information neural network using the physical and chemical monitoring data, and output the predicted value of degradation physical quantity and remaining life. Step S2: Obtain the vibration signal of the bearing, perform adaptive decomposition on the vibration signal to obtain multiple signal components, select the target component that is sensitive to degradation from the multiple signal components, construct a dynamic spectrum representing the coupling relationship between modes based on the target component, and extract degradation features from the dynamic spectrum using a graph neural network. Step S3: Construct a mapping layer to map the degradation features to the model compensation terms of the physical information neural network. The output of the mapping layer is equipped with a physical boundary constraint unit to ensure that the parameter correction does not violate physical laws. The physical equation parameters of the physical information neural network are dynamically corrected using the model compensation terms. Step S4: Construct a generative adversarial network, pre-train the generative adversarial network using source domain fault data, and perform domain adaptation on the generative adversarial network using a small number of abnormal samples from the target domain to generate expanded training samples in the style of the target domain. Step S5: Use the expanded training samples to jointly fine-tune the joint model composed of the physical information neural network, graph neural network, and mapping layer; Step S6: Deploy the fine-tuned joint model at the edge, and use the joint model to perform collaborative reasoning on the real-time collected physical and chemical monitoring data and vibration signals to output the bearing condition assessment results.

2. The method for condition analysis of bearing assemblies for traction of marine engineering cables according to claim 1, characterized in that: The coupled degradation model in step S1 includes a positive feedback coupling relationship between mechanical wear and salt spray corrosion; the predicted degradation physical quantities include predicted wear volume and predicted corrosion layer thickness.

3. The condition analysis method for bearing assemblies used for traction of marine engineering cables according to claim 1, characterized in that: The physicochemical monitoring data mentioned in step S1 includes oil ferromagnetic particle concentration data and bearing total clearance change data. The oil ferromagnetic particle concentration data is converted to obtain the measured wear volume value, and the bearing total clearance change data is converted to obtain the measured corrosion layer thickness value.

4. The method for condition analysis of bearing assemblies for traction of marine engineering cables according to claim 1, characterized in that: The physical constraints mentioned in step S1 include: a data loss term, used to constrain the consistency between the output of the physical information neural network and the measured data; a physical residual loss term, used to constrain the output of the physical information neural network to satisfy the coupling degradation model; and a boundary condition loss term, used to constrain the output of the physical information neural network to satisfy the physical initial state at the initial moment.

5. The condition analysis method for bearing assemblies used for traction of marine engineering cables according to claim 1, characterized in that: In step S2, the adaptive decomposition uses a particle swarm optimization algorithm to adaptively determine the decomposition parameters and then performs variational mode decomposition. The basis for selecting the target components that are sensitive to degradation is: calculating the dominant frequency and degradation sensitivity index of each signal component, removing signal components whose dominant frequency is lower than a set threshold and whose degradation sensitivity index is lower than a set threshold as interference components, and retaining the rest as target components.

6. The method for condition analysis of bearing assemblies for traction of marine engineering cables according to claim 5, characterized in that: The method for constructing the dynamic spectrum in step S2 is as follows: each target component is used as a node, a node feature matrix is ​​constructed using the multidimensional time-frequency micro-features of each target component, an adjacency matrix is ​​constructed using the weighted sum of the phase synchronization index and coherence coefficient between each target component, and the node feature matrix and adjacency matrix are assembled into a frame of dynamic spectrum; a dynamic spectrum sequence is formed by sliding with the time window.

7. The condition analysis method for bearing assemblies used for traction of marine engineering cables according to claim 6, characterized in that: The multidimensional time-frequency micro-features include normalized energy, spectral centroid, sample entropy, kurtosis, and envelope spectral kurtosis; the graph neural network includes a graph convolutional layer, a graph state update layer based on gated recurrent units, a graph attention pooling layer, and a fully connected readout layer connected in sequence.

8. The method for condition analysis of bearing assemblies for traction of marine engineering cables according to claim 1, characterized in that: The physical boundary constraint unit in step S3 achieves parameter correction in the following way: the unconstrained output of the mapping layer is constrained within a finite interval by the hyperbolic tangent function, and then multiplied by the preset maximum allowable correction ratio coefficient to obtain the parameter correction amount; The corrected total parameter value is always greater than zero.

9. The condition analysis method for bearing assemblies used for traction of marine engineering cables according to claim 1, characterized in that: The generative adversarial network mentioned in step S4 is a domain-adaptive auxiliary classification Wasserstein generative adversarial network, whose discriminator includes a true / false discrimination branch, a fault category classification branch, and a domain classification branch; during the domain adaptation process, the domain classification branch is used to extract the noise distribution features and sensor frequency response features of the target domain, driving the generator to complete style transfer.

10. The method for condition analysis of bearing assemblies for traction of marine engineering cables according to claim 1, characterized in that: In the joint fine-tuning process described in step S5, two data flows are executed synchronously for each training batch: the first flow inputs the time-frequency map of the expanded training samples into the graph neural network to extract degradation features; the second flow parses the corresponding physical degradation amount label from the expanded training samples and calculates the loss with the output of the physical information neural network; the two data flows maintain the consistency of sample order in the batch dimension to ensure cross-modal label alignment.

Images