Thruster Fault Diagnosis and Predictive Maintenance Methods and Systems

Abstract

This application relates to the field of data processing technology and discloses a method and system for thruster fault diagnosis and predictive maintenance. The method includes: collecting thruster current, temperature, vibration, and pressure signals to extract feature parameters and construct a time-series dataset; mapping the feature parameters to nodes and establishing a multi-physics coupled hyperedge to construct a time-series correlation matrix; inputting the data into an encoder to extract degradation state feature vectors and mapping them to health indicators; calculating the degradation rate function in conjunction with operating parameters; and integrating the degradation rate function to predict the remaining service life. This application captures the high-order coupling relationships and time-series evolution characteristics between multiple thruster parameters through a multi-physics coupled hypergraph time encoder, and quantifies the impact of different operating conditions on degradation by combining an operating condition-sensitive degradation rate model. This solves the problem of being unable to predict fault evolution trajectories and remaining service life, providing an accurate time window reference for predictive maintenance.

Description

Technical Field

[0001] This application relates to the field of data processing technology, and in particular to a method and system for thruster fault diagnosis and predictive maintenance. Background Technology

[0002] Thruster fault diagnosis and predictive maintenance technologies are crucial for ensuring the safe operation of autonomous underwater vehicles (AUVs). Traditional thruster maintenance relies primarily on periodic servicing and reactive repair, involving fixed-cycle replacement of bearings and inspection of sealing systems. However, this approach fails to dynamically adjust maintenance timing based on the thruster's actual degradation state and operating conditions, leading to over-maintenance resulting in resource waste or untimely maintenance causing sudden failures. In recent years, data-driven fault diagnosis methods have been increasingly applied to thruster health monitoring. By collecting operating parameters such as current, temperature, and vibration, machine learning algorithms are used to identify the current fault type of the thruster. For example, neural networks can be used to classify fault modes such as bearing wear or seal leakage, or support vector machines can be used to determine the thruster's health status level based on vibration spectrum characteristics. Some studies have introduced graph neural network technology to handle the correlations between sensor data. For instance, a semi-supervised fault diagnosis method based on parallel hypergraph attention networks and graph convolutional networks accurately identifies thruster fault types in small sample and interference environments by constructing high-order relationships between nodes and learning dynamic and static features in parallel, solving the problems of insufficient labeled training samples and marine environmental interference.

[0003] However, existing thruster fault diagnosis methods mainly focus on instantaneous state classification, i.e., determining the current fault state of the thruster, but lack the ability to model the evolution process of a fault from its inception to failure. When a thruster operates underwater, degradation processes such as magnetic coupling transmission torque attenuation, increased bearing clearance, and pressure imbalance in the oil-filled sealed chamber exhibit progressive evolution characteristics. Furthermore, they are subject to complex influences from various operating conditions, including water depth pressure gradients, load fluctuations, and start-stop shocks, resulting in significant differences in degradation rates under different operating conditions. While existing diagnostic methods based on static graph networks can capture the instantaneous correlations between sensor parameters, they cannot track the dynamic changes of these correlations over time, nor can they establish the temporal evolution law of the coupled degradation between multiple physical field parameters such as current harmonic distortion, temperature rise rate, vibration characteristics, and pressure response. In addition, key components of the thruster, such as the bearings mentioned in the Haoye thruster manual, have a life cycle of 2,000 hours. However, the actual remaining life is affected by factors such as accelerated aging of seals under deep water pressure, accelerated magnetic coupling wear under high load, and cumulative damage from frequent start-stop cycles. Life prediction based solely on operating hours cannot reflect the modulating effect of real working conditions on the degradation rate.

[0004] Because existing methods lack temporal evolution modeling, they cannot extrapolate future degradation trends from current fault diagnosis results, failing to answer key questions such as "how will the fault develop" and "when is maintenance required." Specifically, existing methods do not establish temporal correlation models of the coupling relationships between multi-physics parameters, and cannot capture cross-physics time-varying coupling mechanisms such as changes in oil filling pressure caused by motor heating and the co-evolution of vibration characteristics caused by bearing wear. Therefore, it is difficult to identify the bifurcation characteristics of different degradation modes on the time axis, and it is impossible to distinguish between different fault evolution modes such as bearing wear-dominated degradation, seal failure-dominated degradation, and motor insulation aging-dominated degradation. In addition, existing methods do not establish condition-sensitive degradation rate models, and cannot quantify the contribution of water depth, speed, load, and start-stop frequency to the acceleration of propeller degradation. This results in remaining life predictions being out of touch with the actual operating environment, with low prediction accuracy and an inability to support maintenance decision optimization. More critically, there is a technical gap between existing instantaneous diagnostic methods and predictive maintenance decisions; diagnostic results cannot be translated into specific maintenance time windows and condition adjustment strategies. Summary of the Invention

[0005] This application provides a method and system for thruster fault diagnosis and predictive maintenance. It is used to capture the high-order coupling relationship and temporal evolution characteristics between parameters such as thruster current, temperature, vibration, and pressure by constructing a multi-physics coupled hypergraph time encoder. Combined with a condition-sensitive degradation rate model, it quantifies the impact of water depth, speed, load, and start-stop on degradation acceleration, thus solving the problem that existing technologies cannot establish fault evolution trajectories and predict remaining life.

[0006] In a first aspect, this application provides a method for thruster fault diagnosis and predictive maintenance, the method comprising:

[0007] Step S1: Collect the current signal, temperature signal, vibration signal and pressure signal of the thruster, extract the current harmonic distortion rate, temperature rise rate, vibration peak factor, vibration kurtosis factor and pressure response time, and construct a multi-dimensional time series dataset;

[0008] Step S2: Map the feature parameters in the multidimensional time series dataset to nodes, establish electromechanical coupling hyperedges, thermal-fluid coupling hyperedges, mechanical degradation hyperedges and sealing failure hyperedges, and construct the hyperedge time series correlation matrix;

[0009] Step S3: Input the hyperedge temporal correlation matrix into the hypergraph convolutional layer and temporal attention layer of the multiphysics coupled hypergraph temporal encoder for forward propagation calculation, and output the degenerate state feature vector;

[0010] Step S4: Map the degradation state feature vector to a health index, perform time-series differentiation on the health index to obtain the degradation rate benchmark function, calculate the operating condition acceleration factor based on water depth, speed, load torque and number of start-stop cycles, and obtain the degradation rate function.

[0011] Step S5: Integrate the reciprocal of the degradation rate function over time. When the health index reaches the failure threshold, terminate the integration to obtain the predicted remaining service life.

[0012] Secondly, this application provides a thruster fault diagnosis and predictive maintenance system, the thruster fault diagnosis and predictive maintenance system comprising:

[0013] The module is used to collect current, temperature, vibration and pressure signals from the thruster, extract current harmonic distortion rate, temperature rise rate, vibration peak factor, vibration kurtosis factor and pressure response time, and construct a multidimensional time series dataset.

[0014] The mapping module is used to map the feature parameters in the multidimensional time series dataset to nodes, establish electromechanical coupling hyperedges, thermal-fluid coupling hyperedges, mechanical degradation hyperedges and sealing failure hyperedges, and construct the hyperedge time series correlation matrix;

[0015] The input module is used to input the hyperedge temporal correlation matrix into the hypergraph convolutional layer and temporal attention layer of the multiphysics coupled hypergraph temporal encoder for forward propagation calculation, and output the degradation state feature vector.

[0016] The output module is used to map the degradation state feature vector into a health index, perform time-series differentiation on the health index to obtain a degradation rate benchmark function, and calculate the operating condition acceleration factor based on water depth, speed, load torque and number of start-stop cycles to obtain the degradation rate function.

[0017] The calculation module is used to perform an integral operation on time with respect to the reciprocal of the degradation rate function. When the health indicator reaches the failure threshold, the integration is terminated to obtain the predicted value of the remaining service life.

[0018] Thirdly, a thruster fault diagnosis and predictive maintenance device is provided, comprising: a memory and at least one processor, wherein the memory stores instructions; the at least one processor invokes the instructions in the memory to cause the thruster fault diagnosis and predictive maintenance device to perform the aforementioned thruster fault diagnosis and predictive maintenance method.

[0019] Fourthly, a computer-readable storage medium is provided, wherein instructions are stored therein, which, when executed on a computer, cause the computer to perform the aforementioned thruster fault diagnosis and predictive maintenance method.

[0020] The technical solution provided in this application collects current, temperature, vibration, and pressure signals from the thruster, and extracts current harmonic distortion rate, temperature rise rate, vibration peak factor, vibration kurtosis factor, and pressure response time to construct a multi-dimensional time-series dataset. This overcomes the limitations of existing technologies that only collect data from a single type of sensor or instantaneous state parameters, and achieves a comprehensive characterization of the multi-physics degradation characteristics of the thruster. The current harmonic distortion rate reflects the electromagnetic state of the magnetic coupling drive and motor windings; the temperature rise rate reflects the dynamic changes in the motor's thermal management capability; the vibration peak factor and kurtosis factor capture the bearing impact characteristics and wear distribution characteristics, respectively; and the pressure response time reflects the dynamic compensation capability of the oil-filled sealing system. These parameters together constitute a feature system covering the multi-physics domains of electromechanical heat flux. This method maps feature parameters in a multidimensional time-series dataset to nodes and establishes electromechanical coupling hyperedges, thermal-fluid coupling hyperedges, mechanical degradation hyperedges, and sealing failure hyperedges. By connecting multiple nodes through these hyperedges, a high-order relational structure is established, overcoming the limitation of traditional methods that only establish second-order relations connecting pairs of nodes. This allows for explicit characterization of multi-physics coupling mechanisms, such as motor heating caused by current distortion, the impact of temperature rise on oil filling pressure, and the co-evolution of vibration peak value and kurtosis. When constructing the hyperedge time-series correlation matrix, a time dimension is introduced, and the elements of the correlation matrix are multiplied by an exponentially decaying time weight function, giving higher weight to historical data closer to the current time. This solves the problem that existing static graph networks cannot capture the time-varying characteristics of topological relationships, laying a data structure foundation for subsequent time-series evolution modeling. The hyperedge temporal correlation matrix is ​​input into the hypergraph convolutional layer and temporal attention layer of the multiphysics coupled hypergraph temporal encoder for forward propagation calculation, outputting a degradation state feature vector. The hypergraph convolutional layer achieves high-order propagation aggregation of node features on the hypergraph topology through multiple matrix multiplication operations of node degree normalization and hyperedge degree normalization, capturing the spatial relationship of mutual influence of parameters such as current, temperature, vibration, and pressure through physical coupling mechanism. The temporal attention layer calculates the query matrix key matrix and value matrix and performs attention weighted aggregation, automatically identifying the most critical historical moment for judging the current degradation state within the time window and assigning it high weight, suppressing the interference of unimportant moments. This solves the problem of key degradation information being submerged due to the equal weighting of all historical data in existing methods. The degradation state feature vector simultaneously integrates multiphysics spatial coupling information and temporal evolution trend information, providing a high-quality feature representation for accurately predicting fault evolution trajectory.

[0021] By mapping the feature vector of the degradation state to a health index and performing time-series differentiation on the health index to obtain the degradation rate benchmark function, a mapping relationship from the abstract feature space to an interpretable health metric is established. The health index quantifies the current degree of degradation of the thruster in the range of 0 to 1, and the degradation rate benchmark function reflects the rate of natural degradation, providing a benchmark reference for remaining life prediction. When calculating the acceleration factor based on water depth, speed, load torque, and number of start-stop cycles, the following terms are used: the square term for water depth reflects the nonlinear accelerated aging of the sealing system due to high pressure in deep water; the linear term for speed reflects the linear promotion of bearing wear by hydrodynamic pressure; the 1.5 power term for load torque reflects the superlinear degradation of the magnetic coupling drive under heavy load; and the logarithmic term for the number of start-stop cycles reflects the diminishing marginal effect of cumulative damage from start-stop impacts. This physics-driven acceleration factor design breaks through the limitations of existing methods that assume a constant degradation rate or simple linear weighting. It can accurately quantify the differentiated contributions of different operating conditions to propeller degradation. By multiplying the degradation rate benchmark function with the acceleration factor, the degradation rate function is obtained, realizing dynamic adjustment of the degradation rate according to the operating conditions. This solves the problem that existing technologies cannot correlate the actual operating environment with the degradation rate. By integrating the reciprocal of the degradation rate function over time and terminating the integration when the health index reaches the failure threshold, the remaining service life prediction is obtained. The trapezoidal rule numerical integration ensures computational accuracy and efficiency. By linking the integration termination condition with the failure threshold, a direct mapping from the degradation trajectory to the maintenance timing is achieved, solving the fundamental defect of existing methods that can only diagnose the current state and cannot predict future failure moments. The remaining service life prediction comprehensively considers the degradation trend coupled by multi-physics fields and the modulation effect of operating conditions, providing an accurate time window reference for predictive maintenance decisions. This represents a technological leap from passive fault diagnosis to active life prediction, significantly reducing the risk of sudden thruster failures and optimizing maintenance resource allocation. Attached Figure Description

[0022] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0023] Figure 1 This is a schematic diagram of one embodiment of the thruster fault diagnosis and predictive maintenance method in this application.

[0024] Figure 2 This is a schematic diagram of one embodiment of the thruster fault diagnosis and predictive maintenance system in this application.

[0025] Figure 3 This is a schematic block diagram of the propeller fault diagnosis and predictive maintenance device in an embodiment of the present invention. Detailed Implementation

[0026] This application provides a method and system for thruster fault diagnosis and predictive maintenance. The terms "first," "second," "third," "fourth," etc. (if present) in the specification, claims, and accompanying drawings are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments described herein can be implemented in a sequence other than that illustrated or described herein. Furthermore, the terms "comprising" or "having" and any variations thereof are intended to cover a non-exclusive inclusion; for example, a process, method, system, product, or device that includes a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or devices.

[0027] For ease of understanding, the specific process of the embodiments of this application is described below. Please refer to [link / reference]. Figure 1 One embodiment of the thruster fault diagnosis and predictive maintenance method in this application includes:

[0028] Step S1: Collect the current signal, temperature signal, vibration signal and pressure signal of the thruster, extract the current harmonic distortion rate, temperature rise rate, vibration peak factor, vibration kurtosis factor and pressure response time, and construct a multi-dimensional time series dataset;

[0029] Step S2: Map the feature parameters in the multidimensional time series dataset to nodes, establish electromechanical coupling hyperedges, thermal-fluid coupling hyperedges, mechanical degradation hyperedges, and sealing failure hyperedges, and construct the hyperedge time series correlation matrix;

[0030] Step S3: Input the hyperedge temporal correlation matrix into the hypergraph convolutional layer and temporal attention layer of the multiphysics coupled hypergraph temporal encoder for forward propagation calculation, and output the degenerate state feature vector;

[0031] Step S4: Map the degradation state feature vector to a health index, perform time-series differentiation on the health index to obtain the degradation rate benchmark function, calculate the operating condition acceleration factor based on water depth, speed, load torque and number of start-stop cycles, and obtain the degradation rate function.

[0032] Step S5: Integrate the time with the reciprocal of the degradation rate function. When the health index reaches the failure threshold, terminate the integration to obtain the predicted remaining useful life.

[0033] It is understood that the executing entity of this application can be a thruster fault diagnosis and predictive maintenance system, or it can be a terminal or a server; the specific implementation is not limited here. This application's embodiments use a server as an example for illustration.

[0034] Specifically, operational data is collected through built-in and external sensors in the thruster. A current sensor acquires three-phase current waveforms at a sampling frequency of 1000Hz. The waveforms are then subjected to a Parker transform to convert the three-phase stationary coordinate system into a two-phase rotating coordinate system, yielding the d-axis and q-axis current components. A Clarke transform further converts the three-phase system into a two-phase orthogonal system. The total harmonic distortion rate (THD) is calculated by taking the square root of the sum of the squares of the ratios of the amplitudes of each harmonic to the fundamental amplitude in the d-axis and q-axis current components. This parameter reflects the motor winding condition and the efficiency of the magnetic coupling transmission. A temperature sensor collects the motor winding temperature and calculates the temperature rise rate by dividing the difference between the current and previous temperatures within a 10-second window by the time interval. A vibration accelerometer collects radial vibration data of the magnetic coupling transmission mechanism and performs a Fast Fourier Transform to convert the time-domain signal to a frequency-domain signal. The vibration amplitude at characteristic frequencies is extracted, and the ratio of the maximum value to the root mean square value of this amplitude sequence is calculated to obtain the vibration peak factor. The vibration kurtosis factor, reflecting the bearing wear degree, is obtained by dividing the fourth-order central moment by the square of the second-order central moment. Pressure sensors synchronously collect the internal pressure of the oil-filled sealed chamber and the external water pressure. After differential calculation, the time interval from the initial change in pressure difference to reaching a stable value is recorded as the pressure response time, reflecting the compensation capability of the sealing system. These parameters are arranged in a time series, and the water depth, speed, load torque, and cumulative number of starts and stops at each moment are recorded to form a multi-dimensional time-series dataset.

[0035] Five feature parameters of the multidimensional time-series dataset are mapped to nodes in a graph structure: current harmonic distortion rate is mapped to the first node, temperature rise rate to the second node, vibration peak factor to the third node, vibration kurtosis factor to the fourth node, and pressure response time to the fifth node. Hyperedges are established to connect multiple nodes to represent multi-physics coupling relationships. The electromechanical coupling hyperedge connecting the first and second nodes represents the coupling process of current distortion leading to motor heating; the thermal-fluid coupling hyperedge connecting the second and fifth nodes represents the coupling mechanism of temperature change affecting oil filling pressure compensation; the mechanical degradation hyperedge connecting the third and fourth nodes represents the co-evolution of vibration characteristics caused by bearing wear; and the sealing failure hyperedge also connecting the second and fifth nodes represents the combined effect of temperature and pressure on sealing performance. When constructing the correlation matrix, if a node belongs to a hyperedge, its corresponding element is set to 1; otherwise, it is set to 0. After introducing the time dimension, each element of the correlation matrix is ​​multiplied by an exponentially decaying time weight function, so that historical data closer to the current time has a higher weight, resulting in a hyperedge time-series correlation matrix. This matrix contains both the spatial coupling topology between nodes and temporal evolution information.

[0036] The hyperedge temporal correlation matrix is ​​input into a multiphysics coupled hypergraph temporal encoder for feature extraction. This encoder consists of three modules: a hypergraph convolutional layer, a temporal attention layer, and a feature fusion layer. The hypergraph convolutional layer first calculates the total number of hyperedges connected to each node in the correlation matrix, forming the diagonal elements of the node degree matrix. Then, it calculates the total number of nodes connected to each hyperedge, forming the diagonal elements of the hyperedge degree matrix. The node degree matrix is ​​then raised to a power of -2 to obtain the node degree normalization matrix, and the hyperedge degree matrix is ​​raised to a power of -1 to obtain the hyperedge degree normalization matrix. The hypergraph convolution operation is performed by sequentially multiplying the node degree normalization matrix, the hyperedge temporal correlation matrix, the hyperedge weight matrix, the hyperedge degree normalization matrix, the transpose of the hyperedge temporal correlation matrix, the node degree normalization matrix, the initial node feature matrix, and the trainable parameter matrix. The result is then activated by a ReLU function to obtain the convolutional feature matrix, capturing the higher-order spatial coupling relationships between nodes. The temporal attention layer multiplies the convolutional feature matrix by the query weight matrix, key weight matrix, and value weight matrix to obtain the query matrix, key matrix, and value matrix, respectively. The matrix product of the query matrix and the transpose of the key matrix is ​​calculated, divided by the square root of the feature dimension, and then normalized using the softmax function to obtain the attention weight matrix. This weight matrix is ​​multiplied by the value matrix to perform weighted aggregation, resulting in the attention feature matrix that captures the correlation of features at different times within the time window. The feature fusion layer concatenates the convolutional feature matrix and the attention feature matrix along the feature dimension and inputs them into the fully connected layer. Dimensionality reduction is achieved through matrix multiplication and bias addition, outputting a 128-dimensional degenerate state feature vector. This vector simultaneously contains multi-physics spatial coupling information and temporal evolution trend information.

[0037] The degradation state feature vector is input into a mapping layer. This layer includes a linear transformation that reduces the 128-dimensional vector to a 1-dimensional scalar. This scalar is nonlinearly mapped to the 0-1 interval using a sigmoid function to obtain a health index. A health index value of 1 indicates a fully healthy state, while 0 indicates a completely failed state. The health index time series is numerically differentiated, and the difference between the current health index and the previous health index is calculated and divided by the time interval to obtain the degradation rate benchmark function. This function reflects the natural degradation rate of the thruster under the current operating conditions. Operating parameters recorded synchronously are extracted from the multi-dimensional time-series dataset. The following terms are calculated: the square of the ratio of water depth to the maximum operating water depth of 6000 meters, the linear term of the ratio of speed to maximum speed, the 1.5 power of the ratio of load torque to rated torque, and the logarithm of the ratio of the number of starts and stops to the threshold of 1000. These four terms are multiplied by weighting coefficients of 0.3, 0.2, 0.4, and 0.1, respectively, summed, and then 1 is added to obtain the operating condition acceleration factor. The square term for water depth reflects the nonlinear accelerated aging of the sealing system due to high pressure in deep water; the linear term for speed reflects the promoting effect of hydrodynamic pressure on bearing wear; the 1.5 power term for load torque reflects the accelerated degradation of the magnetically coupled transmission under heavy load; and the logarithmic term for start-stop frequency reflects the cumulative damage from start-stop shocks. The degradation rate baseline function is multiplied by the operating condition acceleration factor to obtain the degradation rate function, which dynamically adjusts with operating conditions to reflect the true degradation rate of the propeller under actual operating conditions.

[0038] A health indicator failure threshold of 0.15 is set. When the health indicator drops to this value, the thruster is deemed to require maintenance. Starting from the current moment, the reciprocal of the degradation rate function is used to obtain the degradation time function, which represents the time required for the health indicator to decrease by one unit value. The degradation time function is numerically integrated along the time axis using the trapezoidal rule. The trapezoidal rule divides the integration interval into multiple smaller intervals, each with a step size of 1 hour. Within each smaller interval, the integral value is approximated by the area of ​​the trapezoid, which is the arithmetic mean of the function values ​​at both ends of the interval multiplied by the interval width. The sum of the trapezoidal areas of all smaller intervals is then used to obtain the integral value. The health indicator value is continuously monitored during the integration process. The integration operation terminates when the health indicator first drops to the 0.15 failure threshold. The time difference between the termination of integration and the current moment is the predicted remaining service life. For example, a current health index of 0.82 for an underwater thruster indicates it is in good condition. The health index value is obtained by mapping the degradation state feature vector extracted through a hypergraph convolutional layer and a temporal attention layer. The degradation rate baseline function is calculated by differentiating the health index sequence over the past 100 hours, resulting in a decrease of 0.003 per hour. The current operating conditions are: water depth 4500 meters, speed 80% of rated value, load torque 90% of rated value, and 800 cumulative start-stop cycles. Substituting these values ​​into the formula... The water depth term is 0.3 multiplied by 4500 divided by 6000 squared, which equals 0.169; the speed term is 0.2 multiplied by 0.8, which equals 0.16; the load term is 0.4 multiplied by 0.9 to the power of 1.5, which equals 0.343; the start-stop term is 0.1 multiplied by the natural logarithm 1.8, which equals 0.059. Adding these four terms together and then adding 1 gives the acceleration factor for the operating condition, which is 1.731. The degradation rate function is 0.003 multiplied by 1.731, which equals a decrease of 0.0052 per hour. Taking the reciprocal of the degradation rate function, it is found that it takes 1.923 hours for the health index to decrease by 0.01. From the current integral of 0.82 to the failure threshold of 0.15, the health index needs to decrease by 0.67. Using the trapezoidal rule to accumulate in 1-hour increments, the remaining service life is calculated to be approximately 129 hours. This prediction comprehensively considers the degradation trend of multi-physics field coupling, including current harmonics, temperature rise rate, vibration characteristics, and pressure response, as well as the accelerating effect of actual working conditions on the degradation rate.

[0039] In one specific embodiment, step S1 includes:

[0040] The three-phase current waveform data of the thruster is collected by a current sensor at a sampling frequency of 1000Hz. The Parker transform and Clarke transform are performed on the three-phase current waveform data to extract the d-axis current component and the q-axis current component. The total harmonic distortion rate of the d-axis current component and the q-axis current component is calculated to obtain the current harmonic distortion rate.

[0041] Temperature data of the propeller motor windings are collected by a temperature sensor, and the temperature rise rate is obtained by performing time-series differential calculation on the temperature data within a 10-second time window.

[0042] Radial vibration acceleration data of the magnetic coupling transmission mechanism of the propeller is collected by a vibration accelerometer. The vibration amplitude at the characteristic frequency is extracted by fast Fourier transform of the vibration acceleration data. The vibration peak factor is obtained by calculating the ratio of the peak value to the root mean square value of the vibration amplitude. The vibration kurtosis factor is obtained by calculating the ratio of the fourth moment to the square of the second moment of the vibration amplitude.

[0043] The pressure sensor collects the internal pressure signal and external water pressure signal of the oil-filled sealing chamber of the thruster. The internal pressure signal and the external water pressure signal are differentially calculated to calculate the time interval from the pressure difference to the stable end point, and the pressure response time is obtained.

[0044] The current harmonic distortion rate, temperature rise rate, vibration peak factor, vibration kurtosis factor, and pressure response time are organized according to time series, and the water depth, speed, load torque, and cumulative start-stop times at corresponding moments are recorded simultaneously to construct a multi-dimensional time series dataset.

[0045] Specifically, the current sensor continuously acquires the instantaneous values ​​of the three-phase current of the propeller motor at a sampling frequency of 1000 times per second, obtaining a three-phase current waveform data sequence. The Parker transform converts the current components of the three-phase stationary coordinate system to a two-phase rotating coordinate system. Specifically, it maps the A-phase, B-phase, and C-phase currents to d-axis and q-axis current components through a coordinate transformation matrix, where the d-axis is aligned with the magnetic field direction and the q-axis is perpendicular to the magnetic field direction. The Clarke transform converts the three-phase coordinate system to a two-phase orthogonal stationary coordinate system. These two transformations simplify the three-phase current signal into two-phase components for easier subsequent analysis. Fourier decomposition is performed on the d-axis and q-axis current components to extract each harmonic component. The total harmonic distortion rate (THD) is obtained by calculating the sum of the squares of all harmonic amplitudes (excluding the fundamental frequency), dividing by the square of the fundamental frequency amplitude, and then taking the square root. A higher THD value indicates less stable motor operation and lower magnetic coupling transmission efficiency. A temperature sensor continuously collects the surface temperature of the motor windings, recording the temperature values ​​at the beginning and end of a 10-second time window. Subtracting the two values ​​yields the temperature change, which, when divided by the 10-second time interval, gives the temperature rise rate. This rate reflects the motor's heating speed and heat dissipation capacity. A vibration accelerometer collects the time-domain signal of the radial vibration acceleration of the magnetic coupling transmission mechanism's housing. A Fast Fourier Transform converts the time-domain signal to a frequency-domain signal, obtaining the vibration amplitude corresponding to different frequencies. The vibration amplitudes at characteristic frequencies corresponding to the bearing's rotational frequency and its harmonics are extracted to form an amplitude sequence. The vibration peak factor is calculated by dividing the largest amplitude in this sequence by the root mean square (RMS) value of the amplitude sequence. The RMS value is the square root of the sum of the squares of all amplitudes divided by the number of amplitudes. A larger peak factor indicates the presence of significant impact vibration. The vibration kurtosis factor is calculated by dividing the fourth central moment of the amplitude sequence by the square of the second central moment. The fourth central moment is the average of the fourth power of each amplitude minus the average value, and the second central moment is the average of the squares of each amplitude minus the average value, i.e., the variance. A larger kurtosis factor indicates a sharper amplitude distribution and more severe bearing wear. Pressure sensors simultaneously collect the internal pressure of the oil-filled sealed chamber and the external water pressure of the thruster. The subtraction of the two pressure signals yields a pressure difference sequence. When the external water pressure changes, the internal oil-filled system must automatically compensate for the pressure to maintain balance. The time interval from the initial deviation of the pressure difference from the equilibrium point to the re-establishment of equilibrium is recorded as the pressure response time. A longer response time indicates poorer performance of the oil-filled pressure compensation system and a higher risk of seal failure.

[0046] The five characteristic parameters mentioned above are arranged in the order of acquisition time to form a time series. At each time point, the values ​​of current harmonic distortion rate, temperature rise rate, vibration peak factor, vibration kurtosis factor, and pressure response time are recorded. Simultaneously, the operating parameters of the thruster are recorded at each time point: water depth is measured by a depth gauge to obtain the current underwater depth of the thruster; speed is calculated from the thruster speed sensor and propeller parameters; load torque is calculated from the motor current and voltage to obtain the motor output torque; and the cumulative number of start-stop cycles is recorded by a counter to record the total number of start-stop cycles from the start of use to the current time point. The five characteristic parameter sequences and four operating parameter sequences are organized chronologically to form a multi-dimensional time-series dataset. Each row of the dataset corresponds to a sampling time point, and each column corresponds to a parameter dimension, forming a state-operating condition associated dataset with nine columns of data.

[0047] In one specific embodiment, step S2 includes:

[0048] The current harmonic distortion rate, temperature rise rate, vibration peak factor, vibration kurtosis factor, and pressure response time in the multidimensional time series dataset are mapped to the first node, the second node, the third node, the fourth node, and the fifth node, respectively, to construct a node set;

[0049] Establish an electromechanical coupling superedge connecting the first and second nodes, establish a thermal-fluid coupling superedge connecting the second and fifth nodes, establish a mechanical degradation superedge connecting the third and fourth nodes, establish a sealing failure superedge connecting the second and fifth nodes, and construct a set of superedges.

[0050] Construct an association matrix based on the set of nodes and the set of hyperedges, where if a node belongs to a hyperedge, the corresponding element of the association matrix is ​​1, otherwise it is 0.

[0051] By introducing a time dimension extension to the correlation matrix, setting a preset time window length, and multiplying each element of the correlation matrix by a time weighting function, which is an exponential decay function, a hyperedge temporal correlation matrix is ​​obtained.

[0052] Specifically, the multidimensional time-series dataset contains time series of five feature parameters. The current harmonic distortion rate sequence is mapped to the first node representing the electromagnetic state of the motor; the temperature rise rate sequence is mapped to the second node representing the thermal state of the motor; the vibration peak factor sequence is mapped to the third node representing vibration impact characteristics; the vibration kurtosis factor sequence is mapped to the fourth node representing vibration distribution characteristics; and the pressure response time series is mapped to the fifth node representing the state of the sealing system. These five nodes constitute a node set. Hyperedges are high-order relational structures connecting multiple nodes. The electromechanical coupling hyperedge connecting the first and second nodes represents the physical coupling relationship of current distortion causing motor heating; the thermal-fluid coupling hyperedge connecting the second and fifth nodes represents the coupling relationship of temperature rise leading to oil expansion affecting pressure compensation; the mechanical degradation hyperedge connecting the third and fourth nodes represents the coupling relationship of bearing wear simultaneously affecting vibration peak and kurtosis; and the seal failure hyperedge connecting the second and fifth nodes represents the coupling relationship of temperature and pressure acting together on the aging of the sealing material. These four hyperedges constitute a hyperedge set.

[0053] The incidence matrix is ​​a two-dimensional matrix describing the connection relationships between nodes and hyperedges. It has five rows (equal to the number of nodes) and four columns (equal to the number of hyperedges). By iterating through each combination of a node and a hyperedge, we determine whether the node is connected to that hyperedge. If the first node belongs to an electromechanical coupling hyperedge, the element in the first row and first column of the incidence matrix is ​​1; otherwise, the element in the first row and second column is 0. This process is repeated to fill all elements of the matrix, resulting in the complete incidence matrix. The first node belongs to the electromechanical coupling hyperedge with an element of 1 and does not belong to the other three hyperedges with an element of 0. The second node belongs to the electromechanical coupling hyperedge, the thermal-fluid coupling hyperedge, and the sealing failure hyperedge with an element of 1 and does not belong to the mechanical degradation hyperedge with an element of 0. The third node belongs to the mechanical degradation hyperedge with an element of 1 and does not belong to the other three hyperedges with an element of 0. The fourth node belongs to the mechanical degradation hyperedge with an element of 1 and does not belong to the other three hyperedges with an element of 0. The fifth node belongs to the thermal-fluid coupling hyperedge and the sealing failure hyperedge with an element of 1 and does not belong to the other two hyperedges with an element of 0. This forms a sparse correlation matrix that reflects the topological connection pattern between nodes and hyperedges.

[0054] Introducing a time dimension expands the two-dimensional association matrix into a three-dimensional hyperedge temporal association matrix. The third dimension is the time axis, with a length equal to the preset time window length. The time weighting function adopts an exponential decay function, where the function value decreases as the time interval increases. Historical data closer to the current time has a higher weight, while historical data farther away has a lower weight. Each element of the association matrix is ​​multiplied by the corresponding time weighting function value. Elements at the current time are multiplied by a weight of 1 to retain their original value, elements at previous times are multiplied by a weight less than 1 to reduce their value, and elements at even earlier times are multiplied by an even smaller weight to further reduce their value. Iterating through all times within the time window, all elements of the association matrix at each time are multiplied by the corresponding time weighting function value. The weighted association matrices are then stacked in chronological order to form a three-dimensional hyperedge temporal association matrix. The first dimension of the matrix is ​​the node dimension, the second dimension is the hyperedge dimension, and the third dimension is the time dimension. The matrix element values ​​reflect both the connection relationship between nodes and hyperedges and the weight relationship based on time proximity.

[0055] In one specific embodiment, step S3 includes:

[0056] The hyperedge temporal correlation matrix is ​​input into the hypergraph convolutional layer of the multiphysics coupled hypergraph temporal encoder to calculate the node degree matrix and the hyperedge degree matrix. The hypergraph convolution operation is then performed on the hyperedge temporal correlation matrix to obtain the convolutional feature matrix.

[0057] The convolutional feature matrix is ​​input into the temporal attention layer of the multiphysics coupled hypergraph temporal encoder. The query matrix, key matrix, and value matrix at different times within the time window are calculated. The attention weights are calculated based on the query matrix and key matrix. The value matrix is ​​weighted and aggregated to obtain the attention feature matrix.

[0058] The convolutional feature matrix and the attention feature matrix are concatenated and input into the feature fusion layer of the multiphysics coupled hypergraph temporal encoder for dimensionality reduction, and the degenerate state feature vector is output.

[0059] Specifically, after the hyperedge temporal correlation matrix is ​​input into the hypergraph convolutional layer of the multiphysics coupled hypergraph temporal encoder, the node degree matrix and hyperedge degree matrix are first calculated to reflect the topological characteristics of the graph structure. The node degree matrix is ​​a diagonal matrix, where each element on the diagonal represents the total number of hyperedges connected to the corresponding node. Iterating through each row of the hyperedge temporal correlation matrix, all non-zero elements in that row are summed to obtain the degree value of the node, which is then filled into the diagonal of the node degree matrix. The hyperedge degree matrix is ​​also a diagonal matrix, where each element on the diagonal represents the total number of nodes connected to the corresponding hyperedge. Iterating through each column of the hyperedge temporal correlation matrix, all non-zero elements in that column are summed to obtain the degree value of the hyperedge, which is then filled into the diagonal of the hyperedge degree matrix. The node degree normalization matrix is ​​obtained by raising each diagonal element of the node degree matrix to a negative 1 / 2 power, and the hyperedge degree normalization matrix is ​​obtained by raising each diagonal element of the hyperedge degree matrix to a negative 1 / 2 power. The hypergraph convolution operation left-multiplies the node degree normalization matrix by the hyperedge temporal correlation matrix, then right-multiplies the result by the hyperedge weight matrix. The hyperedge weight matrix is ​​a diagonal matrix, with diagonal elements representing the importance weights of each hyperedge. This is then right-multiplied by the hyperedge degree normalization matrix, and the result is right-multiplied by the transpose of the hyperedge temporal correlation matrix. The transpose matrix swaps rows and columns, and the result is right-multiplied by the node degree normalization matrix. This result is then right-multiplied by the initial node feature matrix, where each row corresponds to a node and each column corresponds to a feature dimension. Finally, the result is right-multiplied by the trainable parameter matrix to complete the linear transformation. The parameter matrix is ​​obtained through training. All matrix multiplication results are input into the ReLU activation function for non-linear transformation. The ReLU function sets negative values ​​to zero and retains positive values. The output convolutional feature matrix contains the feature representations propagated between nodes through hyperedges.

[0060] After the convolutional feature matrix is ​​input into the temporal attention layer of the multiphysics coupled hypergraph temporal encoder, it is multiplied by three trainable weight matrices to generate a query matrix, a key matrix, and a value matrix. The query matrix is ​​obtained by right-multiplying the convolutional feature matrix by the query weight matrix, the key matrix by right-multiplying the convolutional feature matrix by the key weight matrix, and the value matrix by right-multiplying the convolutional feature matrix by the value matrix. The three matrices have the same dimension but different contents, representing query information, key information, and value information in the attention mechanism, respectively. The matrix product of the query matrix and the transpose of the key matrix is ​​calculated. Each element of the product matrix represents the similarity of features between two time steps. Each element of the product matrix is ​​divided by the square root of the feature dimension to scale the matrix and avoid excessively large values. The scaled matrix is ​​then input into the softmax function for normalization. The softmax function performs an exponential operation on each row of the matrix and divides it by the sum of the exponents of all elements in that row, making the sum of the elements in each row equal to 1. The normalized matrix is ​​the attention weight matrix, where each element represents the attention weight assigned to another time step at a certain time step. The attention weight matrix is ​​multiplied by the value matrix to complete the weighted aggregation. Each row of the weight matrix is ​​multiplied by the value matrix to obtain the aggregated feature vector. The aggregation results of all time points form the attention feature matrix, which highlights the features of important time points within the time window and suppresses the features of unimportant time points.

[0061] The convolutional feature matrix and the attention feature matrix are concatenated along the feature dimension, merging the two matrices column-wise. If each row of the convolutional feature matrix and the attention feature matrix has 64 features, then each row of the concatenated matrix has 128 features. The concatenated matrix contains both spatial coupling information and temporal correlation information. The concatenated matrix is ​​input to the feature fusion layer of the multiphysics coupled hypergraph temporal encoder. The fusion layer contains a fully connected layer that performs dimensionality reduction. The fully connected layer performs a linear transformation by right-multiplying the concatenated matrix by the dimensionality reduction weight matrix and adding a bias vector. The dimensionality reduction weight matrix has fewer columns than rows to compress the feature dimension. After the linear transformation, the degenerate state feature vector is output. The feature vector dimension is fixed at 128 dimensions. Each element of the vector is an abstract representation extracted from the original multidimensional features. The feature vector contains both the high-order coupling relationship between multiphysics parameters such as current, temperature, vibration, and pressure, and the evolution trend information of these parameters on the time axis.

[0062] In one specific embodiment, the hyperedge temporal correlation matrix is ​​input into the hypergraph convolutional layer of the multiphysics coupled hypergraph temporal encoder, the node degree matrix and the hyperedge degree matrix are calculated, and hypergraph convolution operation is performed on the hyperedge temporal correlation matrix to obtain the convolutional feature matrix, including:

[0063] The hyperedge temporal correlation matrix is ​​input into the hypergraph convolutional layer of the multiphysics coupled hypergraph temporal encoder;

[0064] Count the number of hyperedges connected to each node in the hyperedge temporal correlation matrix, and construct the diagonal elements of the node degree matrix;

[0065] Count the number of nodes connected by each hyperedge in the hyperedge temporal correlation matrix, and construct the diagonal elements of the hyperedge degree matrix;

[0066] The node degree matrix is ​​obtained by performing a negative 1 / 2 power operation on the node degree matrix, and the hyperedge degree matrix is ​​obtained by performing a negative 1 power operation on the hyperedge degree matrix.

[0067] The node degree normalization matrix, hyperedge temporal correlation matrix, hyperedge weight matrix, hyperedge degree normalization matrix, transpose of hyperedge temporal correlation matrix, node degree normalization matrix, initial node feature matrix, and trainable parameter matrix are sequentially multiplied by matrix multiplication. The ReLU activation function is then applied to the result to obtain the convolution feature matrix.

[0068] Specifically, after the hyperedge temporal correlation matrix is ​​input into the hypergraph convolutional layer of the multiphysics coupled hypergraph temporal encoder, the preparation stage for hypergraph convolution operation begins. Each row of the hyperedge temporal correlation matrix is ​​traversed, with each row corresponding to a node. The number of elements with a value of 1 in that row is counted, representing the number of hyperedges connected to that node. The first node's row has only one electromechanical coupling hyperedge column with elements of 1, and the rest are 0, therefore the first node's degree value is 1. The second node's row has three columns of electromechanical coupling, thermal-fluid coupling, and sealing failure hyperedges with elements of 1, therefore the second node's degree value is 3. The third node's row has only one column of mechanical degradation hyperedges with elements of 1, therefore the degree value is 1. The fourth node's row has only one column of mechanical degradation hyperedges with elements of 1, therefore the degree value is 1. The fifth node's row has two columns of thermal-fluid coupling and sealing failure hyperedges with elements of 1, therefore the degree value is 2. These five degree values ​​are then sequentially filled into the diagonal of the node degree matrix to form a diagonal matrix, with the remaining elements being 0. Traverse each column of the hyperedge temporal correlation matrix. Each column corresponds to a hyperedge. Count the number of elements with a value of 1 in the column, which is the number of nodes connected by the hyperedge. For electromechanical coupling hyperedges, the first and second node rows have elements with a value of 1, so the degree value is 2. For thermal-fluid coupling hyperedges, the second and fifth node rows have elements with a value of 1, so the degree value is 2. For mechanical degradation hyperedges, the third and fourth node rows have elements with a value of 1, so the degree value is 2. For sealing failure hyperedges, the second and fifth node rows have elements with a value of 1, so the degree value is 2. Fill the four degree values ​​into the diagonal of the hyperedge degree matrix to form a diagonal matrix. The remaining elements are 0.

[0069] Perform a power-of-half operation on each diagonal element of the node degree matrix. The power-of-half of the first node degree value 1 equals 1, the power-of-half of the second node degree value 3 equals 1 / 2 square root of 3, the power-of-half of the third node degree value 1 equals 1, the power-of-half of the fourth node degree value 1 equals 1, and the power-of-half of the fifth node degree value 2 equals 1 / 2 square root of 2. Fill the result into the diagonal of the node degree normalization matrix to obtain the normalized matrix. Perform a power-of-half operation on each diagonal element of the hyperedge degree matrix, i.e., take the reciprocal. The power-of-half of the electromechanical coupling hyperedge degree value 2 equals 1 / 2, the power-of-half of the thermal-fluid coupling hyperedge degree value 2 equals 1 / 2, the power-of-half of the mechanical degradation hyperedge degree value 2 equals 1 / 2, and the power-of-half of the sealing failure hyperedge degree value 2 equals 1 / 2. Fill the result into the diagonal of the hyperedge degree normalization matrix to obtain the normalized matrix. Normalization reduces the weight of nodes and hyperedges with large degree values ​​and increases the weight of nodes and hyperedges with small degree values, balancing the contributions of nodes and hyperedges in different topological positions.

[0070] The matrix multiplication process is performed sequentially. First, the node degree normalization matrix is ​​left-multiplied by the hyperedge temporal incidence matrix. Matrix multiplication follows the row-column dot product rule. Left-multiplying the node degree normalization matrix (a diagonal matrix) is equivalent to scaling the corresponding rows of the incidence matrix using diagonal elements. The first row is multiplied by 1 to maintain its integrity, while the second row is multiplied by the square root of 3 to reduce its overall size, resulting in the first intermediate result matrix. Next, the first intermediate result matrix is ​​right-multiplied by the hyperedge weight matrix. The hyperedge weight matrix is ​​also a diagonal matrix, with diagonal elements representing the importance weights of each hyperedge. Right-multiplying is equivalent to scaling the corresponding columns of the intermediate result matrix using diagonal elements, adjusting each column according to the hyperedge weights to obtain the second intermediate result matrix. Finally, the second intermediate result matrix is ​​right-multiplied by the hyperedge degree normalization matrix. The diagonal matrix is ​​right-multiplied by the scaling column elements, and each column is multiplied by the reciprocal of its corresponding hyperedge degree to complete the hyperedge-side normalization, resulting in the third intermediate result matrix. The third intermediate result matrix is ​​right-multiplied by the transpose of the hyperedge temporal correlation matrix. The transpose matrix interchanges the rows and columns of the original matrix. This step realizes the aggregation process of information back from the hyperedge to the nodes. Each column of the third intermediate result matrix represents the information weighted by the hyperedge. After multiplying by the transpose matrix, each node aggregates information from different hyperedges to obtain the fourth intermediate result matrix. The fourth intermediate result matrix is ​​right-multiplied by the node degree normalization matrix. The diagonal matrix is ​​right-multiplied by the scaling column elements to complete the node-side normalization, resulting in the fifth intermediate result matrix. The fifth intermediate result matrix is ​​right-multiplied by the initial node feature matrix. Each row of the initial node feature matrix is ​​the original feature vector of a node, which is composed of parameters such as current harmonic distortion rate and temperature rise rate after normalization. Matrix multiplication applies topology propagation weights to the node features to obtain the sixth intermediate result matrix. The sixth intermediate result matrix is ​​right-multiplied by the trainable parameter matrix. The trainable parameter matrix is ​​learned from the training data through the backpropagation algorithm. Matrix multiplication completes the linear transformation of the feature space to map to the new representation space, resulting in the final matrix multiplication result.

[0071] ReLU activation is applied to each element of the final matrix multiplication result. The ReLU function is defined as outputting zero when the input is less than zero and outputting equal to the input when the input is greater than or equal to zero. All elements of the result matrix are traversed. If the element value is negative, the element is set to zero; if the element value is positive, the element remains unchanged. ReLU activation introduces nonlinear transformation capability, enabling the network to learn complex nonlinear mapping relationships. After ReLU processing, a convolutional feature matrix is ​​obtained. Each row of the matrix corresponds to a node, and each column corresponds to a feature dimension. The matrix element values ​​reflect the representation of the node in that feature dimension after hypergraph topology propagation and nonlinear transformation. The convolutional feature matrix integrates the node's own features and the features transmitted by neighboring nodes through hyperedges, capturing high-order feature representations of the relationships between multiple physical field parameters such as current, temperature, vibration, and pressure through physical mechanisms such as electromechanical coupling and thermal-fluid coupling.

[0072] In one specific embodiment, step S4 includes:

[0073] The degraded state feature vector is input into the mapping layer, and a non-linear mapping is performed through the sigmoid function to obtain the health index;

[0074] Numerical differentiation is performed on the time series of health indicators to calculate the ratio of the difference between health indicators at adjacent times to the time interval, thus obtaining the degradation rate benchmark function.

[0075] Based on the water depth, speed, load torque, and cumulative number of starts and stops recorded in the multidimensional time series data, the square term of the ratio of water depth to maximum working water depth, the linear term of the ratio of speed to maximum speed, the 1.5 power term of the ratio of load torque to rated torque, and the logarithmic term of the ratio of number of starts and stops to threshold are calculated. The terms are then weighted and summed, and 1 is added to obtain the working condition acceleration factor.

[0076] The degradation rate function is obtained by multiplying the degradation rate reference function by the operating condition acceleration factor.

[0077] Specifically, the degradation state feature vector is a 128-dimensional vector containing multi-physics coupling degradation information of the thruster. After being input into the mapping layer, the 128-dimensional vector is first projected into a 1-dimensional scalar through a linear transformation. The linear transformation is completed by multiplying the vector with the weight vector and then adding the bias value. The weight vector and bias value are obtained through training. The range of the obtained scalar value is from negative infinity to positive infinity. This scalar is input into the sigmoid function for nonlinear mapping. The expression of the sigmoid function is that the output is equal to 1 divided by 1 plus the negative power of the natural constant e. When the input is negative infinity, the output approaches 0; when the input is 0, the output is equal to 0.5; and when the input is positive infinity, the output approaches 1. The sigmoid function maps any real number to the interval between 0 and 1 to obtain a health index. A health index value of 1 indicates that the thruster is in a completely healthy state, 0 indicates that it is in a completely failed state, and intermediate values ​​represent different degrees of degradation.

[0078] The degradation rate is calculated by performing numerical differentiation on the time series of health indicators. The health indicator series records the health status of the thruster at different times in chronological order. The health indicator at two adjacent times is the current health indicator and the health indicator at the previous time, respectively. Subtracting the two gives the change in health indicator. Subtracting the previous time gives the time interval. Dividing the change in health indicator by the time interval gives the degradation rate value at that time. The same calculation is performed for each time point throughout the entire time series to obtain the degradation rate sequence, which serves as the degradation rate benchmark function. The degradation rate benchmark function reflects the natural degradation rate of the thruster under actual operating conditions. A negative value indicates a decline in health indicators and thruster degradation. The larger the absolute value, the faster the degradation. This benchmark function does not consider differences in operating conditions and assumes a constant degradation rate.

[0079] The accelerated effect of operating conditions on the degradation rate is calculated based on the synchronously recorded operating parameters in the multidimensional time-series dataset. The water depth value at the current moment is extracted and divided by the maximum operating water depth to obtain the water depth ratio. The water depth ratio is then squared to obtain the water depth square term. The maximum operating water depth of the Haoye propeller is 6000 meters. The water depth square term reflects the nonlinear accelerated aging effect of deep-water high pressure on the sealing system; the deeper the water, the greater the pressure, and the faster the sealing material ages. The square relationship indicates that the acceleration effect increases nonlinearly with depth. The speed value is extracted and divided by the maximum speed to obtain the speed ratio as the speed linear term. The speed linear term reflects the promoting effect of hydrodynamic pressure on bearing wear; the higher the speed, the faster the propeller speed, the greater the load on the bearing, and the faster the wear. The linear relationship indicates that the wear rate is proportional to the speed. The load ratio is obtained by dividing the load torque value by the rated torque. The load ratio is then raised to the power of 1.5 to obtain the load 1.5 term. This term reflects the accelerated degradation characteristics of the magnetic coupling drive under heavy load; the larger the load torque, the greater the torque the magnetic coupling mechanism experiences, and the faster the transmission efficiency decreases. The power of 1.5 indicates that the degradation rate increases superlinearly with the load. The start-stop ratio is obtained by dividing the cumulative number of start-stop cycles by a threshold of 1000. Adding 1 to the start-stop ratio and taking its natural logarithm yields the start-stop logarithm term. This term reflects the cumulative damage effect of start-stop shocks; the more start-stop cycles, the more frequently the motor and magnetic coupling mechanism experience start-stop shocks, and the more severe the cumulative damage. The logarithmic relationship indicates diminishing marginal damage effects. Multiply the square term of water depth by a weighting factor of 0.3, the linear term of speed by a weighting factor of 0.2, the term of load raised to the power of 1.5 by a weighting factor of 0.4, and the logarithmic term of start-stop by a weighting factor of 0.1. The weighted sum of these four terms plus 1 yields the acceleration factor under operating conditions. The addition of 1 ensures that the acceleration factor has a minimum value of 1, indicating no acceleration under rated operating conditions, and a value greater than 1, indicating degradation acceleration under adverse operating conditions. The weighting factors reflect the relative contribution of different operating conditions to the degradation of the thruster.

[0080] The degradation rate baseline function is obtained by multiplying the operating condition acceleration factor at each time step. The baseline function provides the natural degradation rate, while the acceleration factor provides the operating condition modulation coefficient. Multiplying the two yields the actual degradation rate. The value of the degradation rate function at each time step equals the baseline degradation rate at that time multiplied by the operating condition acceleration factor at that time. When the thruster operates under harsh conditions such as high load and frequent start-stop in deep water, a large operating condition acceleration factor and a large degradation rate function value indicate rapid degradation. When the thruster operates under favorable conditions such as stable operation under low load in shallow water, a working condition acceleration factor close to 1 and a degradation rate function value close to the baseline value indicate slow degradation. In a certain deep-sea operation, the characteristic vector of the thruster's degradation state was transformed by a linear transformation of the mapping layer and mapped by the sigmoid function to obtain a health index. During continuous operation, the health index gradually decreased from a high value, reflecting thruster degradation. Differentiating the health index time series, it was found that the health index at the current time step was lower than that at the previous time step. Dividing the difference between the two by the time interval yields a negative degradation rate baseline function, indicating a downward trend in the health index. Simultaneously, extracting operating parameters revealed that the thruster was located in deep water, where the water depth was higher than the rated water depth. Squaring the water depth by the maximum operating water depth yielded a water depth square term greater than the rated operating condition. The speed approaching the maximum speed yielded a large speed linear term. The load torque exceeding the rated torque yielded a larger load 1.5 power term. The numerous cumulative start-stop cycles yielded a start-stop logarithmic term of a certain value. The weighted sum of these four terms plus 1 yielded an operating condition acceleration factor greater than 1, indicating that the current adverse operating conditions accelerate the thruster degradation. Multiplying the degradation rate benchmark function by the operating condition acceleration factor yielded a degradation rate function with a larger absolute value, reflecting that the thruster's degradation rate under the current deep-water, high-load operating conditions is significantly faster than under normal operating conditions.

[0081] In one specific embodiment, step S5 includes:

[0082] The failure threshold is set to 0.15. When the health index drops to the failure threshold, the thruster is determined to need maintenance.

[0083] Starting from the current moment, the degradation rate function is inversely calculated to obtain the degradation time function;

[0084] The degradation time function is integrated along the time axis, and numerical integration is performed using the trapezoidal rule, with the integration step size set to 1 hour.

[0085] When the health indicator reaches the failure threshold, the integration calculation is terminated, and the time difference between the termination time and the current time is used as the predicted value of the remaining service life.

[0086] Specifically, a failure threshold of 0.15 is set as the critical value for the health indicators that the thruster needs to maintain. This threshold is determined based on statistical analysis of historical failure data of Haoye thrusters. When the health indicator drops below 0.15, the degradation of key thruster components, such as bearing wear, seal failure, and decreased magnetic coupling transmission efficiency, reaches a level that affects safe operation, necessitating shutdown for maintenance and component replacement. The health indicator is continuously monitored from the current moment, with a value calculated at each sampling time. When the health indicator first falls below 0.15 at a certain moment, a maintenance decision is triggered.

[0087] Starting from the current moment, the degradation rate function is inversely calculated to obtain the degradation time function. Each moment value of the degradation rate function represents the rate at which the health indicator declines. A negative degradation rate indicates a decline in the health indicator, and a larger absolute value of the degradation rate indicates a faster decline. The inverse calculation takes the reciprocal of each moment value of the degradation rate function. Since a negative degradation rate remains negative after the inverse, its physical meaning is converted into the time required for the health indicator to decrease by a unit value. A large absolute value of the degradation rate and a small absolute value of the reciprocal indicate a faster decline and shorter time required; conversely, a small absolute value of the degradation rate and a large absolute value of the reciprocal indicate a slower decline and longer time required. By iterating through the degradation rate function sequence and performing the inverse calculation on each moment value, a degradation time function sequence is obtained. The degradation time function reflects the rate at which the thruster degrades from its current healthy state to a failed state at different moments.

[0088] The total time consumption is calculated by integrating the degradation time function along the time axis. The area enclosed by the function curve and the time axis is calculated. When the degradation time function is negative, the integral value is negative, indicating the accumulated time in the direction of the decline in health indicators. The trapezoidal rule is used for numerical integration. The trapezoidal rule divides the integration interval into multiple smaller intervals, with the integration step size set to 1 hour, meaning each smaller interval is 1 hour wide. The area of ​​each smaller interval is approximated by the area of ​​a trapezoid. The trapezoidal area is calculated by taking the degradation time function values ​​at the beginning and end of the interval, adding the two function values, and dividing by 2 to obtain the average height of the trapezoid. The average height of the trapezoid is multiplied by the interval width of 1 hour to obtain the trapezoidal area of ​​that smaller interval. Starting from the current moment and moving forward along the time axis, the first smaller interval is calculated from the current moment plus 1 hour. The degradation time function values ​​at the beginning and end of this interval are calculated, averaged, and multiplied by 1 hour to obtain the first trapezoidal area. The second smaller interval is calculated from the current moment plus 1 hour to the current moment plus 2 hours, and the second trapezoidal area is calculated. The trapezoidal area of ​​each smaller interval is calculated and accumulated.

[0089] During integration, the health index value corresponding to each moment is calculated simultaneously. The initial value of the health index is the health index at the current moment. For each small interval advanced, the health index decreases by the degradation amount corresponding to that interval. The degradation amount equals the degradation rate of that interval multiplied by the time interval of 1 hour. The cumulative degradation amount continuously increases, causing the health index to continuously decline. After each small interval is completed, it is checked whether the health index at that moment is below the failure threshold of 0.15. If the health index is still above 0.15, the integration continues to the next small interval. If the health index drops to 0.15 or below for the first time, the integration operation is terminated, and the termination moment is the predicted failure moment. The time difference between the integration termination moment and the current moment is calculated. The time difference equals the number of small intervals completed multiplied by the integration step size of 1 hour. This time difference serves as the predicted remaining service life value, representing the remaining usable time of the thruster from the current moment to the moment when maintenance is required. A certain thruster's current health index is higher than the failure threshold. The degradation rate function is negative, reflecting a continuous decline in the health index. Taking the reciprocal of the degradation rate function yields a degradation time function, which is also negative. Using the trapezoidal rule, the area of ​​a trapezoid is calculated every hour starting from the current moment. In the first interval, the degradation time function value is smaller and the absolute value is larger, indicating a slower decline. In the second interval, due to changes in operating conditions, the degradation time function value is larger and the absolute value is smaller, indicating a faster decline. The trapezoidal areas are accumulated while the health index is decreased. Integration continues until the health index drops to 0.15 at a certain moment, at which point the integration ends. The time difference between this moment and the current moment is the predicted remaining service life, reflecting the remaining usable time window of the thruster under the current degradation trend and operating conditions.

[0090] The above describes the thruster fault diagnosis and predictive maintenance method in the embodiments of this application. The following describes the thruster fault diagnosis and predictive maintenance system in the embodiments of this application. Please refer to [link / reference]. Figure 2 One embodiment of the thruster fault diagnosis and predictive maintenance system in this application includes:

[0091] The module is used to collect current, temperature, vibration and pressure signals from the thruster, extract current harmonic distortion rate, temperature rise rate, vibration peak factor, vibration kurtosis factor and pressure response time, and construct a multidimensional time series dataset.

[0092] The mapping module is used to map the feature parameters in the multidimensional time series dataset to nodes, establish electromechanical coupling hyperedges, thermal-fluid coupling hyperedges, mechanical degradation hyperedges and sealing failure hyperedges, and construct the hyperedge time series correlation matrix;

[0093] The input module is used to input the hyperedge temporal correlation matrix into the hypergraph convolutional layer and temporal attention layer of the multiphysics coupled hypergraph temporal encoder for forward propagation calculation, and output the degradation state feature vector.

[0094] The output module is used to map the degradation state feature vector into a health index, perform time-series differentiation on the health index to obtain a degradation rate benchmark function, and calculate the operating condition acceleration factor based on water depth, speed, load torque and number of start-stop cycles to obtain the degradation rate function.

[0095] The calculation module is used to perform an integral operation on time with respect to the reciprocal of the degradation rate function. When the health indicator reaches the failure threshold, the integration is terminated to obtain the predicted value of the remaining service life.

[0096] above Figure 2 The propeller fault diagnosis and predictive maintenance system in this embodiment of the invention is described in detail from the perspective of modular functional entities. The propeller fault diagnosis and predictive maintenance device in this embodiment of the invention is described in detail from the perspective of hardware processing.

[0097] Reference Figure 3 This invention also provides a thruster fault diagnosis and predictive maintenance device, which can be a server, and its internal structure can be as follows: Figure 3 As shown. The thruster fault diagnosis and predictive maintenance device includes a processor, memory, display screen, input device, network interface, and database connected via a system bus. The processor, designed as a computer, provides computing and control capabilities. The memory of the thruster fault diagnosis and predictive maintenance device includes a non-volatile storage medium and internal memory. The non-volatile storage medium stores the operating system, computer programs, and database. The internal memory provides an environment for the operation of the operating system and computer programs in the non-volatile storage medium. The database of the thruster fault diagnosis and predictive maintenance device stores the data corresponding to this embodiment. The network interface of the thruster fault diagnosis and predictive maintenance device is used for communication with external terminals via a network connection. When the computer program is executed by the processor, it implements the above-described method.

[0098] Those skilled in the art will understand that Figure 3 The structure shown is merely a block diagram of a portion of the structure related to the present invention and does not constitute a limitation on the thruster fault diagnosis and predictive maintenance equipment to which the present invention is applied.

[0099] The present invention also provides a computer-readable storage medium, which may be a non-volatile computer-readable storage medium or a volatile computer-readable storage medium, wherein the computer-readable storage medium stores instructions that, when executed on a computer, cause the computer to perform the steps of the thruster fault diagnosis and predictive maintenance method.

[0100] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working processes of the systems and units described above can be referred to the corresponding processes in the foregoing method embodiments, and will not be repeated here.

[0101] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a thruster fault diagnosis and predictive maintenance device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0102] The above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method of propeller fault diagnosis and predictive maintenance, characterized by, The method comprises: Step S1: collecting current signals, temperature signals, vibration signals and pressure signals of the propeller, extracting current harmonic distortion rate, temperature rise rate, vibration peak factor, vibration kurtosis factor and pressure response time, and constructing a multi-dimensional time series data set; Step S2: mapping the feature parameters in the multi-dimensional time series data set as nodes, establishing electromechanical coupling super-edge, heat flow coupling super-edge, mechanical degradation super-edge and sealing failure super-edge, and constructing a super-edge time series correlation matrix; Step S3: inputting the super-edge time series correlation matrix into the supergraph convolution layer and time series attention layer of the multi-physical field coupling supergraph time series encoder for forward propagation calculation, and outputting a degradation state feature vector; Step S4: mapping the degradation state feature vector to a health index, performing time series differentiation on the health index to obtain a degradation rate reference function, calculating a working condition acceleration factor according to water depth, speed, load torque and start-stop times, and obtaining a degradation rate function; Step S5: integrating the reciprocal of the degradation rate function with respect to time, and terminating the integration when the health index reaches the failure threshold to obtain a remaining useful life prediction value.

2. The method of claim 1, wherein, The step S1 comprises: Collecting three-phase current waveform data of the propeller by a current sensor at a sampling frequency of 1000 Hz, performing Park transformation and Clark transformation on the three-phase current waveform data, extracting d-axis current component and q-axis current component, calculating total harmonic distortion rate of the d-axis current component and the q-axis current component, and obtaining the current harmonic distortion rate; Collecting temperature data of the propeller motor winding by a temperature sensor, performing time series differentiation operation on the temperature data within a 10-second time window to obtain the temperature rise rate; Collecting radial vibration acceleration data of the propeller magnetic coupling transmission mechanism by a vibration accelerometer, extracting vibration amplitude at the characteristic frequency by performing fast Fourier transform on the vibration acceleration data, calculating the ratio of the peak value to the root mean square value of the vibration amplitude to obtain the vibration peak factor, and calculating the ratio of the fourth moment to the square of the second moment of the vibration amplitude to obtain the vibration kurtosis factor; Collecting internal pressure signal and external water pressure signal of the propeller oil-filled sealed cabin by a pressure sensor, performing differential operation on the internal pressure signal and the external water pressure signal, calculating the time interval from the starting point to the stable end point of the pressure difference value, and obtaining the pressure response time; Organizing the current harmonic distortion rate, the temperature rise rate, the vibration peak factor, the vibration kurtosis factor and the pressure response time in time sequence, synchronously recording water depth, speed, load torque and cumulative start-stop times at corresponding time, and constructing the multi-dimensional time series data set.

3. The method of claim 1, wherein, The step S2 comprises: Mapping the current harmonic distortion rate, the temperature rise rate, the vibration peak factor, the vibration kurtosis factor and the pressure response time in the multi-dimensional time series data set to first node, second node, third node, fourth node and fifth node respectively, and constructing a node set; establishing a mechanical-electric coupling hyperedge connecting the first node and the second node, establishing a heat flow coupling hyperedge connecting the second node and the fifth node, establishing a mechanical degradation hyperedge connecting the third node and the fourth node, establishing a sealing failure hyperedge connecting the second node and the fifth node, and constructing a hyperedge set; constructing a correlation matrix according to the node set and the hyperedge set, wherein if a node belongs to a hyperedge, a corresponding element of the correlation matrix is 1, otherwise, the corresponding element of the correlation matrix is 0; introducing a time dimension extension to the correlation matrix, setting a preset time window length, multiplying each element of the correlation matrix by a time weight function, the time weight function being an exponential decay function, and obtaining a hyperedge time sequence correlation matrix.

4. The method of claim 1, wherein, The step S3 comprises: inputting the hyperedge time sequence correlation matrix into a hypergraph convolution layer of a multi-physical field coupling hypergraph time sequence encoder, calculating a node degree matrix and a hyperedge degree matrix, performing hypergraph convolution operation on the hyperedge time sequence correlation matrix, and obtaining a convolution feature matrix; inputting the convolution feature matrix into a time sequence attention layer of the multi-physical field coupling hypergraph time sequence encoder, calculating a query matrix, a key matrix and a value matrix at different time points within a time window, calculating an attention weight according to the query matrix and the key matrix, performing weighted aggregation on the value matrix, and obtaining an attention feature matrix; splicing the convolution feature matrix and the attention feature matrix, inputting the spliced matrix into a feature fusion layer of the multi-physical field coupling hypergraph time sequence encoder for dimension reduction processing, and outputting the degradation state feature vector.

5. The method of claim 4, wherein, The step S3 comprises: inputting the hyperedge time sequence correlation matrix into a hypergraph convolution layer of a multi-physical field coupling hypergraph time sequence encoder; counting the number of hyperedges connected to each node in the hyperedge time sequence correlation matrix to construct diagonal elements of a node degree matrix; counting the number of nodes connected to each hyperedge in the hyperedge time sequence correlation matrix to construct diagonal elements of a hyperedge degree matrix; performing a negative two power operation on the node degree matrix to obtain a node degree normalization matrix, and performing a negative one power operation on the hyperedge degree matrix to obtain a hyperedge degree normalization matrix; performing matrix multiplication operation on the node degree normalization matrix, the hyperedge time sequence correlation matrix, a hyperedge weight matrix, the hyperedge degree normalization matrix, a transpose matrix of the hyperedge time sequence correlation matrix, the node degree normalization matrix, an initial node feature matrix and a trainable parameter matrix in sequence, and applying a ReLU activation function to an operation result to obtain the convolution feature matrix.

6. The method of claim 1, wherein, The step S4 comprises: inputting the degradation state feature vector into a mapping layer to perform nonlinear mapping through a sigmoid function to obtain the health index; performing numerical differentiation operation on a time sequence of the health index to calculate a ratio of a difference value of adjacent time points of the health index to a time interval to obtain the degradation rate reference function; According to the water depth, the speed, the load torque and the cumulative start-stop times recorded in the multi-dimensional time series data set, a square term of the ratio of the water depth to the maximum working water depth, a linear term of the ratio of the speed to the maximum speed, a 1.5 power term of the ratio of the load torque to the rated torque, and a logarithmic term of the ratio of the start-stop times to a threshold value are calculated, the terms are weighted and summed, and 1 is added to obtain the working condition acceleration factor; The degradation rate function is obtained by multiplying the degradation rate reference function by the working condition acceleration factor.

7. The method of claim 1, wherein, The step S5 comprises: The failure threshold value is set to 0.15, and when the health index decreases to the failure threshold value, it is determined that the propeller needs maintenance; The degradation time function is obtained by performing inverse operation on the degradation rate function from the current time; The integral operation is performed on the degradation time function along the time axis, and the trapezoidal rule is used for numerical integration, and the integral step is set to 1 hour; When the health index reaches the failure threshold value, the integral operation is terminated, and the time difference between the integral termination time and the current time is taken as the remaining service life prediction value.

8. A propeller fault diagnosis and prognostic maintenance system characterized by, The propeller fault diagnosis and predictive maintenance system comprises: A construction module is configured to collect current signals, temperature signals, vibration signals and pressure signals of the propeller, extract current harmonic distortion rate, temperature rise rate, vibration peak factor, vibration kurtosis factor and pressure response time, and construct a multi-dimensional time series data set; A mapping module is configured to map feature parameters in the multi-dimensional time series data set as nodes, establish electromechanical coupling hyper-edges, thermal flow coupling hyper-edges, mechanical degradation hyper-edges and sealing failure hyper-edges, and construct a hyper-edge time series correlation matrix; An input module is configured to input the hyper-edge time series correlation matrix into a hypergraph convolution layer and a time series attention layer of a multi-physical field coupling hypergraph time series encoder to perform forward propagation calculation, and output a degradation state feature vector; An output module is configured to map the degradation state feature vector to a health index, perform time series differentiation on the health index to obtain a degradation rate reference function, calculate a working condition acceleration factor according to the water depth, the speed, the load torque and the start-stop times, and obtain a degradation rate function; An operation module is configured to perform integral operation on time with the inverse of the degradation rate function, and terminate the integral when the health index reaches a failure threshold value to obtain a remaining service life prediction value.

9. A propeller fault diagnosis and predictive maintenance device, characterized by The computer program causes the processor to execute the propeller fault diagnosis and predictive maintenance method according to any one of claims 1 to 7 when the computer program is executed by the processor.

10. A computer-readable storage medium having stored thereon a computer program, characterized in that, The computer program causes the processor to execute the propeller fault diagnosis and predictive maintenance method according to any one of claims 1 to 7 when the computer program is executed by the processor.

Images