Intelligent monitoring method for stern seal device state of pod propeller

Abstract

The present application relates to a kind of intelligent monitoring method of tail seal device state of pod propeller, utilize the feature of minimum structural risk of support vector machine, even in the case of few fault samples, accurate classification hyperplane can be established, the problem of insufficient fault data is solved;Fusion acoustic emission, vibration and oil multi-source characteristics, and through "one-to-one" multi-classification support vector machine, can automatically distinguish the multiple fault modes of tail seal, reduce the dependence of artificial interpretation;Through the model parameter of particle swarm optimization algorithm and acoustic emission high-frequency signal extraction, the early micro-cracks, dry friction and other abnormal states of tail seal end face are more sensitive, can be early warning before leakage occurs;Combined with the remaining life prediction function of support vector machine regression, the degradation trend of tail seal can be quantitatively evaluated, to provide direct decision basis for condition-based maintenance, help to extend the maintenance interval of pod tail seal, reduce the life cycle cost.

Description

Technical Field

[0001] This invention relates to the field of marine propulsion system technology, and in particular to an intelligent monitoring method for the status of the stern sealing device of a podded propulsion unit. Background Technology

[0002] Podded propulsion, as a new type of propulsion device, has gained popularity among ship owners and research institutions due to its high efficiency and low noise. A podded propulsion device is a novel type of special propulsion device, belonging to the high-tech field of marine electric propulsion. Its propulsion device adopts a podded structure, breaking through the conventional design of "diesel engine plus open transmission shaft system," bringing a revolutionary change to electric propulsion. Using podded electric propulsion reduces water pollution caused by fuel emissions. The podded propulsion device integrates the propulsion motor and propeller into a podded housing, and this assembly (propulsion motor, podded housing, propeller) is "suspended" at the stern by a rotating base. The hull only needs to provide the power required by the propulsion motor. The podded propulsion device can change the ship's direction of travel by rotating 360°.

[0003] The stern seal of a podded propulsion system is a core component ensuring the safe operation of the entire propulsion system. Its primary function is to prevent seawater intrusion, avoiding a series of problems ranging from bearing corrosion to motor burnout. Since the underwater structure of a podded propulsion system is in constant contact with seawater, the seal must prevent seawater from entering the precision components such as motors and bearings along the stern shaft, preventing corrosion, insulation failure, or burnout. Secondly, it balances internal and external pressure. The underwater structure of a podded propulsion system is subjected to varying water pressure. The sealing system is typically equipped with a gravity oil tank or pressure compensation unit to ensure that the lubricating oil pressure is always slightly higher than the external water pressure. Finally, the stern seal isolates impurities such as silt and fishing net debris from the water, protecting internal precision surfaces; for icebreakers, its special design can also resist ice debris compression. Stern seal devices have complex structures and many classifications, the most common being mechanical face seals and pressure-compensated seals. Regardless of the type, the basic principle is the same. The main components of a mechanical seal include a rotating ring, a stationary ring, a bushing, an elastic element, a rotating ring seal ring, a stationary ring seal ring, and a gland. Among them, the dynamic and static rings form a friction pair, which is the most important component of the entire mechanical seal and determines the two main performance characteristics of the entire mechanical seal: sealing performance and friction performance.

[0004] Existing stern seal condition monitoring can be broadly categorized into two types: invasive direct measurement and non-invasive indirect monitoring. Invasive direct monitoring technology enters the sealing cavity through a pre-designed interface. This involves periodically sampling and analyzing the oil's viscosity, moisture content, particle size, and oxidation level; or embedding thermocouples or resistance thermometers near the sealing end face, where abnormal temperature increases are key signals of dry friction or overload on the sealing surface; or monitoring the rate of change in the liquid level in the gravity oil tank, or installing a flow meter on the return oil pipe to accurately measure the leakage rate. Non-invasive indirect monitoring technology requires no modification to the sealing structure. It employs acoustic emission technology, using acoustic emission sensors attached to the sealing shell to capture high-frequency stress waves generated by friction, cracks, or cavitation. This is highly sensitive to early wear and poor lubrication, and has a high signal-to-noise ratio. Vibration analysis methods are used, with acceleration sensors placed near the bearing housing to monitor characteristic frequency changes caused by the sealing friction surface. However, due to the complex structure of the podded propulsion system, the signal is susceptible to interference, making it technically challenging. Ultrasonic detection technology is commonly used for routine inspections. Ultrasonic probes detect characteristic signals generated by cavitation or leakage on the sealing surface.

[0005] The current trend is to introduce artificial intelligence and build digital twin models. By integrating multi-source data such as power, speed, and draft, the model can predict the service status and remaining service life of the seals before a failure occurs. Meanwhile, new self-powered wireless sensors have solved the problem of wiring difficulties in the rotating parts of the pod's propulsion motor. Existing stern seal monitoring relies heavily on single-threshold alarms, making it difficult to detect progressive wear early and accurately distinguish between different fault modes such as seal lip aging, spring failure, or net entanglement. Traditional vibration analysis is susceptible to interference from the complex transmission paths inside the pod, resulting in a low signal-to-noise ratio. Summary of the Invention

[0006] To address the challenges of early fault warning, low fault mode recognition rate, and reliance on large sample training in existing stern seal monitoring technologies, this paper proposes an intelligent monitoring method for the stern seal device of a podded propulsion unit. This method transforms post-maintenance into condition monitoring, enabling dynamic monitoring of the stern seal's operational status. When there is a risk of leakage or wear, early warnings can be issued to avoid downtime and unnecessary economic losses.

[0007] The technical solution of this invention is: a method for intelligent monitoring of the status of the stern sealing device of a podded propulsion unit, comprising the following steps: Step S1: Simultaneously collect acoustic emission signals, vibration signals, and oil data of the stern sealing device, extract features from each, align them with the time scale, and construct a hybrid feature vector characterizing the operating state of the stern seal. Step S2: Input the hybrid feature vector into the pre-trained support vector machine classification model and output the current stern seal failure mode identification result; Step S3: When the fault mode identification result is a preset degradation type fault, calculate the Mahalanobis distance between the currently acquired real-time hybrid feature vector and the preset health benchmark vector, and use the Mahalanobis distance as a health indicator. Step S4: Obtain the time series of the health indicators, use the support vector regression model to predict future health indicator values, and combine the degradation curve model with the preset failure threshold to assess the remaining service life. When the remaining service life is lower than the preset service life threshold, issue a maintenance warning.

[0008] Furthermore, the specific process of feature extraction in step S1 is as follows: The acoustic emission signal is decomposed by wavelet packet analysis to extract the energy ratio, peak-to-peak value, ring count and center frequency of each frequency band, thus forming an acoustic emission feature vector. The vibration signal is subjected to time-domain statistical features and characteristic frequency amplitudes to form a vibration feature vector; The concentration of ferromagnetic particles and the water saturation of the oil data are extracted to form an oil feature vector; The acoustic emission feature vector, vibration feature vector, and oil feature vector are concatenated and normalized to obtain the mixed feature vector.

[0009] Furthermore, the specific strategy for time scale alignment in step S1 is as follows: Set the time analysis window for fast-changing signals and the data refresh cycle for slow-changing signals; For slowly changing oil data, a zero-order hold strategy is adopted to map the slowly changing oil feature values ​​onto the time axis of the fast-changing signal, so that a complete mixed feature vector containing fast-changing features and aligned slow-changing features is obtained within any time analysis window.

[0010] Furthermore, the training process of the support vector machine classification model in step S2 includes: An improved particle swarm optimization algorithm is used to globally optimize the penalty factor and kernel parameters of the support vector machine. In the improved particle swarm optimization algorithm, an adaptive inertia weight adjustment strategy is introduced, which makes the inertia weight change dynamically with the number of iterations. This can maintain global search capability in the early stage of iteration and enhance local fine search capability in the later stage of iteration.

[0011] Furthermore, the method for obtaining the preset health baseline vector in step S3 is as follows: After the break-in period following the sealing installation, acoustic emission, vibration, and oil signals were collected under the set benchmark conditions, and multiple sets of mixed feature vectors under healthy conditions were extracted. Calculate the mean center of the mixed feature vectors under the multiple health states, and use it as the health baseline vector; at the same time, calculate the covariance matrix between the mixed feature vectors under the multiple health states, and use it to calculate the Mahalanobis distance.

[0012] Furthermore, the set reference operating conditions include: straight-line operation at rated speed and small rudder angle turning operation; The formula for calculating the Mahalanobis distance is: ,in, V 当前 For real-time blending of feature vectors, V 基准 As a health baseline vector, S -1 The inverse of the covariance matrix is ​​given. The Mahalanobis distance is mapped to a wear index of 0-100% as the health indicator.

[0013] Furthermore, the specific process for assessing the remaining useful life in step S4 is as follows: Construct a time series sample by taking the actual monitored health indicator values ​​of N consecutive time points in the past, input it into a trained support vector regression model, and predict the health indicator prediction value sequence for F time points in the future. The predicted value sequence of the health indicators was fitted using an exponential degradation model to obtain the degradation curve parameters; Based on the degradation curve parameters and the failure threshold, the time required for the health index to reach the failure threshold is calculated, which is taken as the remaining service life.

[0014] Furthermore, the method for determining the failure threshold is as follows: In the bench test, the health index values ​​were recorded when the measured leakage rate first exceeded the preset leakage threshold, when the vibration intensity caused by abnormal friction exceeded the preset alarm line, and when the preset abnormal noise occurred during operation. The average of the health index values ​​obtained from the above multiple sets of experiments is taken as the final failure threshold.

[0015] Furthermore, the specific strategy for issuing maintenance warnings in step S4 is as follows: When the fault mode identification result is a normal state, the normal state indication is maintained. When the fault mode identification result is a degradation-type fault, a first-level warning is triggered; When the remaining service life is lower than the preset service life threshold, the first-level warning is upgraded to the second-level warning, and a prompt message containing the suggested maintenance time is output.

[0016] The beneficial effects of this invention are as follows: The intelligent monitoring method for the stern seal of the podded propulsion device of this invention offers high precision with small sample sizes: Utilizing the risk minimization characteristic of support vector machines, an accurate classification hyperplane can be established even when fault samples are scarce, solving the problem of insufficient fault data; Automatic multi-mode identification: By integrating acoustic emission, vibration, and oil multi-source features, and through a "one-to-one" multi-classification support vector machine, it can automatically distinguish various fault modes of the stern seal, reducing reliance on manual interpretation; Strong early warning capability: Through the model parameters of the particle swarm optimization algorithm and the extraction of high-frequency acoustic emission signals, it is highly sensitive to early abnormal states such as micro-cracks and dry friction on the stern seal end face, enabling early warning before leakage occurs; Enables condition-based maintenance: Combining the remaining life prediction function of support vector machine regression, it can quantitatively assess the deterioration trend of the stern seal, providing direct decision-making basis for condition-based maintenance, helping to extend the maintenance interval of the podded stern seal and reduce the total life cycle cost. Attached Figure Description

[0017] Figure 1 This is a schematic diagram of the arrangement of the stern sealing device of the present invention; Figure 2 This is a diagram of the multi-dimensional data fusion architecture of the present invention. Detailed Implementation

[0018] The present invention will now be described in detail with reference to the accompanying drawings and specific embodiments. These embodiments are based on the technical solution of the present invention and provide detailed implementation methods and specific operating procedures. However, the scope of protection of the present invention is not limited to the following embodiments.

[0019] The present invention provides an intelligent monitoring method for the stern seal status of a podded propulsion unit. Acoustic emission, vibration, and oil level sensors are installed on the stern seal base, and signals are transmitted to an internal control panel via a slip ring device. The operator can observe the operating status of the stern seal on the control panel. A support vector machine algorithm is introduced to perform pattern recognition on different states of the stern seal, such as "health," "wear," and "leakage," based on the monitoring signals.

[0020] To achieve the above objectives, the specific details of the invention are as follows: like Figure 1As shown, acoustic emission sensors, vibration acceleration sensors, and online oil particle size sensors are arranged on the stern sealing base and adjacent structures of the podded propulsion unit. Acoustic emission waveform signals, vibration acceleration signals, and oil contamination data reflecting the stern sealing state are synchronously acquired via a synchronous data acquisition card at a preset sampling frequency. The acquired signals undergo preprocessing, specifically including: for the acquired acoustic emission signals, noise reduction is performed using wavelet packet transform, and the energy ratio, peak-to-peak value, ring count, and center frequency of each frequency band are extracted to form an acoustic emission feature vector; for the vibration signals, the mean, variance, peak factor, kurtosis, and amplitude of the shaft frequency and blade passing frequency are extracted to form a vibration feature vector; for the acquired oil data, the ferromagnetic particle concentration (ppm) and water saturation (%RH) are directly extracted to form an oil feature vector.

[0021] like Figure 2 The three sets of feature vectors are concatenated to form a high-dimensional hybrid feature vector representing the stern seal state, and then subjected to deviation standardization to map it to the [0, 1] interval. A classification fault diagnosis model based on support vector machine is constructed. The standardized hybrid feature vector is input into the pre-trained support vector machine model, and the identification result of the current stern seal state is output. The identification result includes one or more of "normal state", "lip wear", "spring failure", "rotary seal ring aging" and "seawater leakage". The support vector machine is improved and trained and its parameters are optimized. The hybrid feature vectors of the podded propeller stern seal under different identification results in historical operation and bench tests are collected to form a small sample training set.

[0022] Penalty factor for support vector machines using an improved particle swarm optimization (IPSO) algorithm C and kernel parameters g Global optimization is performed. Specifically, an adaptive inertia weight adjustment strategy is introduced into the particle swarm optimization algorithm to avoid getting trapped in local optima.

[0023] (1), (2), In the formula, t This represents the current iteration number. i For the first i One particle, r 1. r 2 is a random number between 0 and 1. c 1. c 2 represents the learning factor; For the first i Individual particles t In the next iteration, the individual optimal position refers to the best position found by the particle in its own history; Gbest t yest The global optimum in the next iteration refers to the best position found by the entire particle swarm so far, representing shared intelligence. X It represents a position, indicating a candidate solution to the problem, corresponding to the value of the input variable; V Speed ​​determines the direction and rate of particle movement. It is the inertial part, which maintains the original flight trend of the particle and has global search capability; This is the cognitive part; the particle moves closer to its historical optimal position based on its own memory. In the social component, the particle reference group's experience tends to converge towards the globally optimal position.

[0024] When the diagnostic model outputs degradation-type faults such as "lip wear" or "spring failure", a support vector machine algorithm model is further used, with the mixed feature vector of the continuous time series as input and the wear rate as output, to establish a prediction curve for the remaining service life of the seal; when the predicted remaining service life is lower than a preset threshold, a maintenance warning is issued.

[0025] After the break-in period following the sealing installation, data was collected under different operating conditions to establish a health baseline model. Acoustic emission, vibration, and hydraulic signals were collected during operation at rated speed under straight-line conditions and at small rudder angles (-15°~15°), and a 19-dimensional feature vector was extracted. Generally, the pod does not operate at large rudder angles. Most of the time, it rotates slowly at small rudder angles. The mean center of these multiple vector sets was calculated as the health baseline vector. V 基准 Simultaneously, the covariance matrix S among these healthy samples is calculated for subsequent Mahalanobis distance calculation (because Mahalanobis distance is dimensionless and can eliminate interference from inter-feature correlation). After the monitoring system is running, the real-time 19-dimensional hybrid feature vector obtained from each sampling is... V 当前 Calculate its comparison with health benchmarks V 基准 The characteristic distance between them is used as the basis for defining the health index HI. The Mahalanobis distance is calculated as follows: Equation (3), In the formula, S -1 It is the inverse matrix of the covariance matrix of the health samples.

[0026] In an initial healthy state HI The value is close to 0. As the seal wears, this value increases monotonically. For ease of observation, HI The value is uniformly mapped to 0-100%, and is called the wear index. Tests are conducted on a bench to determine the point at which the sealing function fails. HI critical value HI阈值 The criteria for judgment are: 1) When the measured leakage rate first exceeds 5 ml / 24h (this reference standard is based on historical data), record the value at this point. HI value; 2) When the vibration intensity caused by abnormal friction exceeds the alarm threshold in the ISO 10816-3 standard, record the vibration intensity at this point. HI value; 3) If any obvious abnormal noise occurs during operation (by listening), record the sound. HI value.

[0027] The results obtained from the above multiple sets of experiments HI The mean of the values ​​is used as the final failure threshold. HI 阈值 .

[0028] When a time series is obtained HI After setting the values, support vector regression (SVR) is used to predict trends and forecast future trends. HI Time series. A degradation curve model is established and fitted using an exponential model: Equation (4), in, t It is runtime. a , b These are the degenerate parameters to be fitted. e It is the base of the index.

[0029] Take the actual monitoring data from N consecutive time points in the past ( t 1, HI 1), ( t 2, HI 2), ..., construct time series samples. Train the current optimal exponential curve parameters, and use the trained SVR model to predict the next F time points. HI value sequence HI HI HI HI HI HI HI HI HI HI HI HI HI HI HI HI HI HI HI HI HI 预测 ( t ).

[0030] This invention provides an improved method for monitoring the condition of the stern sealing device of a podded propulsion system. It employs a support vector machine (SVM) algorithm model and utilizes an improved particle swarm optimization (IPSO) algorithm with a penalty factor applied to the SVM. C and kernel parameters g Global optimization can quantitatively assess the degradation trend of the stern seal, providing a direct basis for decision-making regarding condition-based maintenance. This helps extend the maintenance interval of the podded stern seal and reduce the total life cycle cost. The specific implementation plan is as follows: First, the hardware configuration. A wideband sensor with a resonant frequency of 150kHz was selected and mechanically mounted on the radial surface of the stern sealing flange. This location is closest to the friction pair, resulting in minimal signal attenuation, and a sampling frequency of 2MHz was used. A triaxial accelerometer was selected as the vibration sensor and attached radially to the bearing housing of the stern sealing cavity shell. The focus was on acquiring axial vibration signals at a frequency of 10.24kHz, with an analysis bandwidth of approximately 4kHz. The oil sensor was positioned on the return oil management system between the drain port at the bottom of the sealing cavity and the gravity oil tank, with a sampling interval of 30 seconds per acquisition. A synchronous data acquisition unit with GPS timestamps ensured synchronized triggering of acoustic emission and vibration signals. The oil data acquisition timestamp was aligned with the vibration signal. The sample length for each acquisition was set to 0.5 seconds.

[0031] As mentioned earlier, the acoustic emission sensor and vibration acceleration sensor are hardware-level synchronized via the same synchronous data acquisition card, ensuring microsecond-level precise correspondence between sound and vibration signals. The online oil sensor, due to its detection principle, has a slow data refresh cycle, set to output ferromagnetic particle concentration (ppm) and water content (%RH) values ​​every 30 seconds. A feature extraction time window is set, with 0.5 seconds as the basic time analysis window for condition monitoring. This window is chosen because it is sufficient to cover multiple complete cycles of the tail shaft frequency (typically 5-10Hz) and the complete envelope of acoustic emission transient events. Every 0.5 seconds, the system processor extracts the raw acoustic emission and vibration waveform signals and extracts the corresponding 11-dimensional acoustic emission features and 6-dimensional vibration features as described in the scheme. This means that two sets are generated per second, and 120 sets of rapidly changing feature vectors are generated per minute. Since the changes in oil particle size and moisture are relatively slow physicochemical processes, abrupt changes typically do not occur within a 30-second cycle. Therefore, a zero-order hold strategy is adopted to map the slowly changing oil data onto a fast-changing time axis with a resolution of 0.5 seconds.

[0032] Let the oil sensor be at time points T0, T 30 ,T 60 Output data; within the time interval [T0, T 30 Each 0.5-second rapid change feature vector generated within a given timeframe has its oil characteristic value determined by the sensor. T The measured value at time 0. This means that the oil contamination level in the sealed cavity remains relatively stable within this 30-second period. Using the above strategy, the system can obtain a complete, time-aligned 19-dimensional hybrid feature vector within any 0.5-second time window. , This vector can be directly input into a support vector machine classifier for fault identification, or stored in a historical database for calculating health indicators (Mahaviron distance) and predicting remaining lifespan as described in subsequent embodiments. This strategy effectively solves the problem of fusing multi-source heterogeneous sensor data streams in the time dimension, ensuring the integrity of the feature vector at every moment, and laying a unified data foundation for all subsequent time-series-based degradation analyses.

[0033] Secondly, the fault feature vector was extracted and constructed. In the bench test, four typical states were set, with 50 samples collected for each state. Normal state: newly installed seal, operating under rated speed and straight-line conditions, and operating under small rudder angle (-15°~15°) conditions; Lip wear: laser etching was performed on the stationary ring end face to simulate micron-level wear after approximately 1000 hours of operation; Spring failure: one small spring was cut to simulate elasticity failure due to corrosion; Rotary seal aging: the rotary seal ring was immersed in high-temperature oil and then installed to simulate hardening failure; Seawater leakage: a small bypass valve was opened to inject salt water into the sealing cavity to simulate trace seawater intrusion. After setting the above states, data with an internal resistance of 0.5 seconds were processed. For acoustic emission signal sources, energy ratio, peak-to-peak value, ring count, and center frequency of eight frequency bands after three-layer wavelet packet decomposition are extracted, totaling 11 dimensions. For vibration signal sources, the mean, variance, peak factor, kurtosis, and amplitude of the axial signal, as well as the axial frequency (e.g., 5Hz) and blade passing frequency (e.g., 25Hz) are extracted, totaling 6 dimensions. For oil signal sources, ferromagnetic particle concentration (ppm) and water saturation (%RH) are extracted, totaling 2 dimensions, for a total of 19 dimensions. All feature vectors are then normalized to the [0, 1] interval. 20 samples are randomly selected from 50 samples for each state as the training set, and the remaining 30 samples are used as the test set.

[0034] A sample specifically refers to a complete 19-dimensional hybrid feature vector generated within a single 0.5-second time window of continuous acquisition, after feature extraction and time-scale alignment. The time length is precisely the 0.5-second physical signal duration corresponding to one sample. It integrates 17-dimensional rapidly changing features within the 0.5-second window, as well as 2-dimensional slowly changing oil features aligned with its time. This single sample corresponds to one operating state of the stern seal at that 0.5-second instant and is assigned only one state label (such as "normal" or "lip wear"). Taking the construction of a "normal state" sample set as an example: In bench testing, the seal is allowed to operate stably for 30 minutes under rated speed, straight-line conditions, and small rudder angle (-15°~15°). Sampling is performed according to a strategy of acquiring samples every 10 minutes, with each acquisition lasting 5 seconds (i.e., containing 10 consecutive non-overlapping 0.5-second samples). From these 30 minutes, 3 × 10 samples / acquisition = 30 independent "normal" samples are obtained. If the operating conditions are changed and this process is repeated, hundreds of 19-dimensional feature vector samples with the "normal" label can be quickly accumulated through several trials.

[0035] Furthermore, the Particle Swarm Optimization (IPSO) algorithm is used to optimize the training and optimization of the Support Vector Regression (SVR) model. 20 particles are initialized, each representing a set of parameters (C, g). Search range: An adaptive inertia weighting strategy is adopted, which allows the algorithm to focus on global search in the early stages and on fine-grained local search in the later stages. After 100 iterations, the optimal parameter combination is obtained.

[0036] Next, the trained "one-to-one" classification model is deployed to the host computer monitoring software. The host computer automatically captures a segment of real-time data every 10 minutes, extracts 19-dimensional features, and inputs it into the model. All binary classifiers vote, and the category with the most votes becomes the current state. If the output is "normal," the green light remains on; if the output is "lip wear" or "spring failure," the system triggers a yellow warning. In bench testing, when the model identifies "lip wear" three consecutive times, it means the stern seal is already in a degradation stage.

[0037] Finally, remaining service life prediction. When the system triggers a yellow alert for "lip wear," the remaining service life prediction module is automatically activated. It selects 30 consecutive days of online monitoring data and calculates the distance between the 19-dimensional feature vectors each day as the wear index. This index typically shows an accelerating upward trend as failure approaches. When the remaining service life is less than 720 hours (approximately 30 days), the system automatically upgrades the yellow alert to a red alert and displays a prompt on the human-machine interface in the cab: "It is recommended to schedule a dry docking for seal replacement within 30 days," thus achieving predictive maintenance.

[0038] The embodiments described above merely illustrate specific implementations of the present invention, and while the descriptions are detailed, they should not be construed as limiting the scope of the invention patent. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the present invention, and these all fall within the protection scope of the present invention. Therefore, the protection scope of this invention patent should be determined by the appended claims.

Claims

1. A method for intelligent monitoring of the status of the stern sealing device of a podded propulsion unit, characterized in that, Includes the following steps: Step S1: Simultaneously collect acoustic emission signals, vibration signals, and oil data of the stern sealing device, extract features from each, align them with the time scale, and construct a hybrid feature vector characterizing the operating state of the stern seal. Step S2: Input the hybrid feature vector into a pre-trained support vector machine classification model and output the current stern seal failure mode identification result; Step S3: When the failure mode identification result is a preset degradation type failure, calculate the Mahalanobis distance between the currently acquired hybrid feature vector and the preset health benchmark vector, and use the Mahalanobis distance as a health indicator; Step S4: Obtain the time series of the health indicator, use a support vector regression model to predict future health indicator values, and combine the degradation curve model with a preset failure threshold to assess the remaining service life. When the remaining service life is lower than the preset service life threshold, issue a maintenance warning.

2. The intelligent monitoring method for the status of the podded propulsion stern sealing device according to claim 1, characterized in that, The specific process of feature extraction in step S1 is as follows: wavelet packet decomposition is performed on the acoustic emission signal to extract the energy ratio, peak-to-peak value, ring count, and center frequency of each frequency band, forming an acoustic emission feature vector; time-domain statistical features and characteristic frequency amplitude are extracted from the vibration signal to form a vibration feature vector; ferromagnetic particle concentration and water saturation are extracted from the oil data to form an oil feature vector; the acoustic emission feature vector, vibration feature vector, and oil feature vector are concatenated and normalized to obtain the mixed feature vector.

3. The intelligent monitoring method for the status of the podded propulsion stern sealing device according to claim 1, characterized in that, The specific strategy for time scale alignment in step S1 is as follows: set the time analysis window for fast-changing signals and the data refresh cycle for slow-changing signals; for slowly changing oil data, adopt a zero-order hold strategy to map the slowly changing oil feature values ​​onto the time axis of the fast-changing signals, so that in any time analysis window, a complete mixed feature vector containing fast-changing features and aligned slow-changing features is obtained.

4. The intelligent monitoring method for the status of the podded propulsion stern sealing device according to claim 1, characterized in that, The training process of the support vector machine classification model in step S2 includes: using an improved particle swarm optimization algorithm to globally optimize the penalty factor and kernel parameters of the support vector machine; in the improved particle swarm optimization algorithm, an adaptive inertia weight adjustment strategy is introduced, so that the inertia weight changes dynamically with the number of iterations, which can maintain the global search capability in the early stage of iteration and enhance the local fine search capability in the later stage of iteration.

5. The intelligent monitoring method for the status of the podded propulsion stern sealing device according to claim 1, characterized in that, The method for obtaining the preset health benchmark vector in step S3 is as follows: after the break-in period following the sealing installation, acoustic emission, vibration, and oil signals are collected under the set benchmark operating conditions, and multiple sets of mixed feature vectors under the health state are extracted. Calculate the mean center of the mixed feature vectors under the multiple health states, and use it as the health baseline vector; at the same time, calculate the covariance matrix between the mixed feature vectors under the multiple health states, and use it to calculate the Mahalanobis distance.

6. The intelligent monitoring method for the status of the podded propulsion stern sealing device according to claim 5, characterized in that, The established benchmark operating conditions include: straight-line operation at rated speed and small rudder angle turning operation; the formula for calculating the Mahalanobis distance is: ,in, V 当前 For real-time blending of feature vectors, V 基准 As a health baseline vector, S -1 The covariance matrix is ​​the inverse of the covariance matrix; the Mahalanobis distance is mapped to a wear index of 0-100% as the health indicator.

7. The intelligent monitoring method for the status of the podded propulsion stern sealing device according to claim 1, characterized in that, The specific process of assessing the remaining lifespan in step S4 is as follows: take the actual monitored health indicator values ​​of the past N consecutive time points to construct a time series sample, input it into the trained support vector regression model, and predict the health indicator prediction value sequence for the next F time points. The predicted value sequence of the health indicators is fitted using an exponential degradation model to obtain degradation curve parameters; based on the degradation curve parameters and the failure threshold, the time required for the health indicators to reach the failure threshold is calculated, which is taken as the remaining service life.

8. The intelligent monitoring method for the status of the podded propulsion stern sealing device according to claim 1, characterized in that, The failure threshold is determined as follows: In bench tests, health index values ​​are recorded when the measured leakage rate first exceeds the preset leakage threshold, when the vibration intensity caused by abnormal friction exceeds the preset alarm line, and when a preset abnormal noise occurs during operation. The average of the health index values ​​obtained from the above multiple tests is taken as the final failure threshold.

9. The intelligent monitoring method for the status of the podded propulsion stern sealing device according to claim 1, characterized in that, The specific strategy for issuing maintenance warnings in step S4 is as follows: when the fault mode identification result is a normal state, maintain the normal state indication; when the fault mode identification result is a degradation fault, trigger the first-level warning. When the remaining service life is lower than the preset service life threshold, the first-level warning is upgraded to the second-level warning, and a prompt message containing the suggested maintenance time is output.

Images