A method, device, and medium for detecting a fault type of a marine power system

Abstract

This invention discloses a method, equipment, and medium for detecting fault types in marine propulsion systems, belonging to the field of marine fault detection technology. Specifically, it includes: synchronously acquiring ship environmental and operational data, as well as multi-source sensor signals from the propulsion system, to generate real-time operating condition excitation vectors and real-time component state feature sets, respectively; constructing and utilizing a dynamic mapping model to map the real-time operating condition excitation vectors to the expected state feature range that the system should have under the current specific external load conditions; calculating the multidimensional deviation between the real-time component state feature set and this dynamic expected range; matching diagnostic samples containing complete operating condition and state information with a pre-stored fault mode feature library to output specific fault type codes; and using confirmed diagnostic samples to update the fault mode feature library and recalibrate the dynamic mapping model. This invention achieves accurate and adaptive detection of faults in marine propulsion systems by establishing a dynamic correlation between operating conditions and states.

Description

Technical Field

[0001] This invention relates to the field of ship fault detection technology, specifically to a method, equipment, and medium for detecting fault types in a ship's power system. Background Technology

[0002] Fault detection in marine propulsion systems is a key technology for ensuring navigation safety and operational economy. Current mainstream detection methods rely on deploying sensor networks on critical components of the propulsion system to continuously collect state signals such as vibration, temperature, and pressure, and then identifying abnormal patterns by analyzing the time-frequency domain characteristics of these signals. These methods are generally based on a core assumption that specific faults will exhibit unique and stable signal characteristics.

[0003] However, in actual navigation, the loads and operating states of a ship's propulsion system are strongly coupled with complex and ever-changing sea conditions, navigation operation commands, and ship loading conditions. The normal operating signal of the same propulsion component in calm sea conditions may exhibit similar characteristics to its signal response under severe external excitation in adverse sea conditions; conversely, weak abnormal signals caused by early component damage are easily drowned out by strong broadband vibration noise caused by external environmental factors such as high winds and waves. Existing technologies typically distinguish between normal and abnormal conditions by setting fixed alarm thresholds or establishing static health benchmark models. Essentially, this treats the propulsion system as an isolated object operating in a constant or quasi-static environment, failing to incorporate real-time, multi-dimensional navigation environment and operating conditions as core variables into the logical judgment process of fault diagnosis.

[0004] This limitation in the detection logic makes it difficult for the system to reliably distinguish between "benign high-load state caused by harsh environment" and "malicious performance degradation caused by component damage" when facing highly dynamic navigation missions in the real world. This results in a high false alarm rate and insufficient sensitivity to early and slowly changing faults. Summary of the Invention

[0005] The purpose of this invention is to provide a method, equipment, and medium for detecting fault types in a marine propulsion system, thereby solving the following technical problems:

[0006] Existing methods for detecting faults in marine propulsion systems fail to incorporate real-time, ever-changing navigation environments and operating conditions as core variables into the diagnostic logic. This makes it difficult to accurately distinguish between benign high-load responses caused by adverse external stimuli and severe performance degradation signals caused by damage to internal components.

[0007] The objective of this invention can be achieved through the following technical solutions:

[0008] A method for detecting fault types in a marine propulsion system includes the following steps:

[0009] S1: Collect environmental data and operation command data during the ship's navigation process to form a navigation condition data stream, and simultaneously collect multi-channel sensor signals of rotating mechanical components in the power system to form a condition monitoring data stream;

[0010] S2: Based on the ship's motion mechanics model, the navigation condition data stream is analyzed to obtain the real-time condition excitation vector. At the same time, the state monitoring data stream is transformed in the time and frequency domain and statistical features are extracted to generate a real-time component state feature set.

[0011] S3: Construct a dynamic mapping model based on the correspondence between the operating condition excitation vector and the component state feature set in historical data. Input the real-time operating condition excitation vector into the dynamic mapping model to obtain the expected state feature range.

[0012] S4: Compare the generated real-time component state feature set with the expected state feature range obtained by mapping through dynamic correlation, and calculate the multidimensional deviation value between the two.

[0013] S5: When the multidimensional deviation value exceeds the preset dynamic threshold, a diagnostic sample is formed by combining the corresponding real-time operating condition excitation vector and the real-time component state feature set. The diagnostic sample is matched and retrieved with the pre-stored fault mode feature library, and the fault type code associated with the retrieval result is output.

[0014] S6: The confirmed diagnostic samples and their matching fault type codes are added as new records and incorporated into the fault mode feature library, and the parameters of the dynamic correlation are recalibrated based on the updated fault mode feature library.

[0015] As a further aspect of the present invention: in step S2, the specific process for generating the real-time operating condition excitation vector is as follows:

[0016] The heading angle, speed setpoint, and main engine speed command from the operation command data are extracted as basic control parameters. Simultaneously, wave height, wave direction, period, wind speed, and wind direction from the environmental data are extracted as environmental input parameters. A ship motion response model is established based on the ship's main dimensions, hull lines, and mass distribution data. The basic control parameters and environmental input parameters are input into the model to calculate the heave, pitch, and roll amplitudes of the ship in the waves, as well as the wave-induced vertical and horizontal bending moments of the hull beams. Combining the design parameters of the ship's propulsion system and the real-time speed, the flow field characteristics at the stern of the hull are calculated based on the ship's hull line data and speed to determine the wake fraction and its circumferential non-uniformity at the propeller disk.

[0017] Simultaneously, based on real-time wind speed, wind direction, and the windward area of ​​the ship's superstructure, the additional resistance and turning moment generated by wind load on the ship are calculated. The vertical bending moment and horizontal bending moment of the hull beam, the wake fraction of the propeller disk, the circumferential non-uniformity of the wake, the wind-induced additional resistance, and the wind-induced turning moment are used as standardized indicators, arranged and normalized according to a predefined vector structure to form a standardized real-time operating condition excitation vector.

[0018] As a further aspect of the present invention: the specific process for generating the real-time component state feature set in S2 is as follows:

[0019] Raw signals from acceleration sensors, velocity sensors, temperature sensors, and pressure sensors deployed on various rotating components of the power system are collected. The raw signals undergo preprocessing corresponding to their signal types: bandpass filtering is performed on vibration-related raw signals, and moving average filtering is performed on temperature and pressure-related raw signals. The preprocessed vibration signals are segmented, and time-domain statistical characteristics and frequency-domain spectral characteristics are calculated for each data segment. The time-domain statistical characteristics include RMS value, peak value, skewness, and kurtosis; the frequency-domain spectral characteristics include fundamental frequency and harmonic amplitude, as well as the energy proportion of multiple preset frequency bands.

[0020] For the preprocessed temperature and pressure signals, calculate their mean, extreme values, and trend slope within an analysis window; combine the time-domain statistical features, frequency-domain spectral features, mean, extreme values, and trend slope extracted from all sensors of the same component into a feature sub-vector of the component; concatenate the feature sub-vectors of all components of the power system into a comprehensive vector in a predetermined order, and normalize each dimension of the comprehensive vector to generate a standardized real-time component state feature set.

[0021] As a further aspect of the present invention: in S3, the specific construction process of the dynamic mapping model is as follows:

[0022] Extract historical operating condition excitation vectors and corresponding historical component state feature sets from the historical database, establish a pairing relationship between the two to form a training sample set; calculate the mean and standard deviation of each dimension of the historical component state feature set in the training sample set, and use the mean and standard deviation to standardize the historical component state feature set; divide the standardized training sample set into a training subset and a validation subset in chronological order.

[0023] Construct a neural network whose input layer dimension matches the dimension of the historical operating condition stimulus vectors, and whose output layer dimension matches the dimension of the standardized historical component state feature set. Use the historical operating condition stimulus vectors from the training subset as input and the corresponding standardized historical component state feature set as the target output, and train the neural network using the backpropagation algorithm. Evaluate the trained neural network using a validation subset, and stop training when the prediction error on the validation subset no longer decreases within consecutive training cycles. Save the connection weight parameters, mean, and standard deviation of the neural network at this point to form a dynamic mapping model.

[0024] As a further aspect of the present invention: the specific calculation process of the multidimensional deviation value in S4 is as follows:

[0025] Obtain the value of each dimension of the real-time component state feature set, and simultaneously obtain the upper and lower limits of each dimension corresponding to the expected state feature range output by the dynamic mapping model; for each feature dimension, compare the value in the real-time component state feature set with the upper and lower limits of the expected state feature range of that dimension, and calculate the difference between the real-time value and the range boundary.

[0026] When the real-time value is within the expected state feature range, the difference in that dimension is zero; when the real-time value exceeds the expected state feature range, the difference in that dimension is the absolute distance between the real-time value and the nearest range boundary value. A predefined weight coefficient vector is used, and each component of the weight coefficient vector is multiplied by the difference in the corresponding feature dimension to obtain the weighted deviation of each dimension. The sum of squares of the weighted deviations of all feature dimensions is calculated, and the square root of the sum of squares is taken as the multidimensional deviation value.

[0027] As a further aspect of the present invention: the specific construction process of the pre-stored fault mode feature library in step S5 is as follows:

[0028] From the historical maintenance records and abnormal event reports of the ship's power system, extract confirmed fault events and their corresponding fault type identifiers; based on the occurrence time of the fault events, retrieve the complete historical operating condition excitation vectors and historical component state feature sets synchronously recorded within that time period from the historical database; combine each retrieved set of historical operating condition excitation vectors and historical component state feature sets, and associate them with the corresponding fault type identifiers to form an initial fault mode sample.

[0029] The historical component state feature sets in all initial fault mode samples are subjected to dimensionless standardization. The parameters used in the standardization process are consistent with the mean and standard deviation calculated during the construction of the dynamic mapping model. The standardized historical operating condition excitation vectors and the standardized historical component state feature sets are recombined to generate standardized fault feature vectors. For each type of fault, a preset number of standardized fault feature vectors are selected as typical patterns for that type of fault. All typical patterns and their corresponding fault type identifiers are stored in a structured manner. The structured storage data set constitutes the pre-stored fault mode feature library.

[0030] As a further aspect of the present invention: the specific process of parameter recalibration in S6 is as follows:

[0031] Obtain diagnostic samples confirmed by fault type coding matching, and standardize the real-time component status feature set in the diagnostic samples; add the standardized diagnostic samples to the fault mode feature library; extract historical operating condition excitation vectors and corresponding historical component status feature sets from the updated fault mode feature library and historical operation database to form a new training sample set; recalculate the mean and standard deviation of each dimension of the historical component status feature set in the new training sample set, and use the recalculated mean and standard deviation parameters to standardize the new training sample set.

[0032] The standardized new training sample set is input into the neural network structure contained in the dynamic mapping model, and the connection weight parameters of the neural network are adjusted using the backpropagation algorithm. When the prediction error of the neural network on the independent validation dataset reaches the convergence state, the adjusted neural network connection weight parameters and the mean and standard deviation of each dimension of the recalculated historical component state feature set are saved to complete the parameter recalibration of the dynamic mapping model.

[0033] The present invention also includes a computer device, comprising a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the fault type detection method for a ship propulsion system as described in any one of the preceding claims.

[0034] The present invention also includes a computer-readable storage medium storing a computer program, which, when executed by a processor, implements the fault type detection method for a marine propulsion system as described in any one of the preceding claims.

[0035] The beneficial effects of this invention are:

[0036] 1) By constructing a dynamic mapping relationship between operating conditions and states, a health benchmark that adaptively adjusts with environmental loads is established, effectively suppressing false alarms caused by strong external excitations. This method does not compare the collected vibration, temperature, and other component state signals in isolation with a fixed threshold or model. Instead, it first uses a ship kinematics model to analyze and fuse real-time multi-dimensional environmental and operational data, such as waves, wind, speed, and maneuvering commands, into a "operating condition excitation vector" that quantifies the comprehensive external loads currently experienced by the ship. This vector transforms abstract and complex environmental influences into a set of standardized indicators with clear engineering physical meaning (such as hull beam bending moment, wake non-uniformity, and wind resistance moment). Subsequently, through a "dynamic mapping model" (such as a neural network) trained based on historical data, this real-time operating condition excitation vector is directly and continuously mapped to the corresponding "expected state characteristic range." The physical essence of this mapping process is establishing a dynamic functional relationship between "external excitation input" and "system normal response output." Therefore, the benchmark used to determine whether the current state is abnormal is no longer static, but a theoretically normal response range that is calculated in real time based on the actual physical load borne by the ship. When severe sea conditions such as high winds and waves cause a general increase in vibration of the power system, the expected state range calculated by this method will also widen synchronously, thus encompassing this "benign high-load response" caused by known external factors and conforming to the laws of mechanics within the normal range. An alarm is only triggered when the signal characteristics deviate significantly from the expectation based on the current physical situation, thus overcoming the inherent defect of the traditional static threshold method with a high false alarm rate under varying operating conditions.

[0037] 2) This invention, by introducing the core principles of "operating condition normalization" and "correlation residual analysis," achieves enhanced perception of early fault symptoms at the principle level. Firstly, this invention utilizes a dynamic mapping model based on ship motion mechanics to transform the real-time complex navigation conditions (waves, wind, speed, and operations) into a quantified "expected normal response spectrum." This expected response spectrum essentially characterizes the reasonable theoretical range of all vibration, temperature, and other signals that a healthy ship propulsion system should produce under currently measurable external physical excitations. Subsequently, the actually collected multidimensional state feature set is compared with this dynamic "theoretical health baseline." The calculated "comprehensive deviation" is no longer, in a physical sense, the absolute strength of the signal itself, but rather the "residual" between the actual system response and the theoretical response of the healthy system. It can be understood that the most important characteristic of an early fault (such as initial pitting of a bearing) is often not an increase in overall vibration (which may be masked by severe sea conditions), but rather a change in the inherent "excitation-response" transfer function, causing the actual response to begin systematically deviating from the theoretical prediction based on the health model at certain frequencies or directions. For example, under specific wave frequency excitation, a fault may cause the amplitude growth pattern of a certain high-frequency harmonic component to deviate from the predicted curve of the mechanical model, even if its absolute value does not exceed the traditional static threshold. This method captures such subtle variations in the transfer function by calculating the coordinated deviation of multidimensional features on a dynamic benchmark. Subsequent matching with the fault mode feature library essentially compares this "abnormal transfer function variation pattern" with a known fault case library. Utilizing the principle that different faults have different "frequency-condition" characteristics in their impact on the system's transfer function, accurate identification is achieved. Therefore, more sensitive and specific early warnings for early faults are realized.

[0038] 3) A closed-loop parameter recalibration mechanism is designed to enable the diagnostic system to continuously learn and adapt, ensuring accuracy and robustness during long-term operation. This invention introduces a key self-updating cycle. It is understood that the "health" baseline of a ship's propulsion system will slowly change with natural equipment aging, maintenance and replacement, or long-term performance drift, and may encounter new operating conditions not covered in historical data. Therefore, the method sets up a feedback loop: for each confirmed fault diagnosis, its complete "diagnostic sample" (including the real-time operating condition excitation vector that triggered the alarm, the component state feature set, and the finally verified fault type code) is standardized and incorporated into the "fault mode feature library" as a new knowledge sample. This makes the feature library a continuously expanding case library, capable of covering a wider range of fault-operating condition combinations. By periodically utilizing the updated feature library and recent historical normal operation data, training samples are re-extracted, and the "dynamic mapping model" is "parameter recalibrated." By refitting the function model describing the "excitation-response" relationship with data containing the latest state information, the dynamic mapping model can continuously learn and track the evolution trajectory of the "normal response" caused by the slow performance degradation of the ship's power system throughout its entire life cycle. This allows it to dynamically adjust the "expected state characteristic range" of its output, ensuring that it always matches the ship's current actual health status and capabilities, thereby maintaining the accuracy and reliability of long-term monitoring. Attached Figure Description

[0039] The invention will now be further described with reference to the accompanying drawings.

[0040] Figure 1 This is a schematic diagram of a fault type detection method for a ship propulsion system according to the present invention. Detailed Implementation

[0041] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0042] Please see Figure 1 As shown, the present invention is a method for detecting fault types in a ship's propulsion system, comprising the following steps:

[0043] S1: Collect environmental data and operation command data during the ship's navigation process to form a navigation condition data stream, and simultaneously collect multi-channel sensor signals of rotating mechanical components in the power system to form a condition monitoring data stream;

[0044] First, environmental and operational data, including wind speed, wind direction, significant wave height, average wave period, wave direction, ship speed, heading angle, main engine speed, and output torque commands, are simultaneously collected through the ship's integrated weather station, motion sensors (MRU), GPS / BDS navigation receiver, and main engine control system, forming a navigation condition data stream. This step ensures real-time and comprehensive capture of the external marine environment and internal operational status faced by the ship. Simultaneously, accelerometers, speed sensors, temperature sensors, and hydraulic sensors deployed on key rotating mechanical components such as the main engine, bearings, gearboxes, and generators collect multi-channel high-frequency time-series signals of vibration, temperature, and pressure, forming a condition monitoring data stream. The collected raw sensor signals undergo preprocessing, including moving average filtering and trend term removal, to eliminate measurement noise and signal drift, laying the foundation for subsequent accurate feature analysis.

[0045] S2: Based on the ship's motion mechanics model, the navigation condition data stream is analyzed to obtain the real-time condition excitation vector. At the same time, the state monitoring data stream is transformed in the time and frequency domain and statistical features are extracted to generate a real-time component state feature set.

[0046] Based on the ship's principal dimensional parameters, hull lines, mass distribution, and loading conditions, a six-degree-of-freedom ship motion frequency domain or time domain response model is established. Real-time acquired wave height, direction, and period sequences are used as input wave spectra. Combined with the current speed and heading, the amplitude operators for the ship's heave, pitch, and roll motion responses in waves are calculated. This allows for the determination of the short-term statistical characteristic values ​​of the wave-induced vertical bending moment, horizontal bending moment, and torque of the hull beams. Simultaneously, based on real-time wind speed, wind direction, and the windward and current-receiving areas of the ship's superstructure and underwater hull, the additional drag and turning moment generated by wind and current loads are calculated. Furthermore, combining the propeller design parameters, stern hull lines, and real-time speed, the wake fraction, wake circumferential non-uniformity, and thrust reduction factor are calculated. The wave-induced bending moment and torque, wind / flow load moment, and wake characteristic parameters obtained above are normalized according to a predefined vector structure and dimensions to form a standardized real-time operating condition excitation vector that comprehensively represents the main mechanical and fluid loads on the ship and propulsion system under the current navigation environment.

[0047] Meanwhile, the specific process of performing time-frequency domain transformation and statistical feature extraction on the state monitoring data stream to generate a real-time component state feature set is as follows: The preprocessed vibration acceleration signal is segmented, windowed, and subjected to Fast Fourier Transform (FFT) to calculate the power spectral density of each frequency band, and the amplitude and phase information of the fundamental frequency, harmonics, and sideband components related to the rotor speed are extracted. The time-domain statistical characteristics of the vibration velocity signal are calculated, including root mean square value, peak value, impulse factor, and margin factor. For temperature and pressure signals, their average, maximum, minimum, standard deviation, and linear and nonlinear trend coefficients within a fixed time window are calculated. All feature parameters extracted from different sensors and different analysis dimensions of the same component are sorted and concatenated according to component identification and feature category to form a high-dimensional feature sub-vector at the component level. Finally, the feature sub-vectors of all monitored components of the power system are integrated according to the global component order, and each dimension of the integrated comprehensive feature vector is standardized based on historical data distribution to generate a standardized real-time component state feature set that can be used for model comparison.

[0048] S3: Construct a dynamic mapping model based on the correspondence between the operating condition excitation vector and the component state feature set in historical data. Input the real-time operating condition excitation vector into the dynamic mapping model to obtain the expected state feature range.

[0049] Historical navigation condition data covering various sea states and operating modes, along with synchronously recorded historical component state feature data, are extracted from the ship's historical data center to form historical sample pairs. The historical component state feature data is cleaned and aligned to ensure accurate matching with the timestamps of the navigation condition data. A sliding time window is used to divide the continuous time-series data into multiple independent training samples, each containing a historical navigation condition stimulus vector and its corresponding historical component state feature set. Principal component analysis or an autoencoder is used to reduce the dimensionality and denoise the high-dimensional historical component state feature set, retaining the main variation information. Subsequently, a machine learning model suitable for handling high-dimensional nonlinear relationships is selected, using the previously processed historical navigation condition stimulus vector as input and the dimensionality-reduced historical component state feature set as the output target, for supervised learning on the training set. The model hyperparameters are adjusted through cross-validation, and the model performance is evaluated using an independent test set to ensure it can accurately capture the mapping relationship between complex navigation conditions and the system's normal state response. The trained and validated model, along with its required feature preprocessing parameters (such as the dimensionality reduction matrix and standardization coefficients), is saved, thus forming the dynamic mapping model.

[0050] S4: Compare the generated real-time component state feature set with the expected state feature range obtained by mapping through dynamic correlation, and calculate the multidimensional deviation value between the two.

[0051] For the current moment, the real-time operating condition excitation vector generated in step S2 is input into the constructed dynamic mapping model. This model outputs a center point of the expected state characteristics that matches the current operating condition, along with its reasonable fluctuation range (an interval consisting of the predicted value plus or minus a certain number of standard deviations). Subsequently, the real-time component state feature set generated in step S2 undergoes the same dimensionality reduction and standardization transformation as historical data, projecting it onto the same feature subspace as the model output. In this subspace, the Euclidean distance between the projected real-time feature points and the boundary of the expected feature range output by the model is calculated; this distance is the multidimensional deviation value. This process essentially quantifies the difference between the currently observed system state and the health baseline state predicted based on current external conditions.

[0052] S5: When the multidimensional deviation value exceeds the preset dynamic threshold, a diagnostic sample is formed by combining the corresponding real-time operating condition excitation vector and the real-time component state feature set. The diagnostic sample is matched and retrieved with the pre-stored fault mode feature library, and the fault type code associated with the retrieval result is output.

[0053] When the calculated multidimensional deviation exceeds a dynamic threshold set based on model confidence and historical anomaly statistics, an anomaly is identified. The real-time operating condition excitation vector and the complete real-time component state feature set (including the original high-dimensional features) of the current time slice are frozen and packaged into a diagnostic sample. This diagnostic sample is then sent to the fault mode feature library for matching retrieval. The fault mode feature library is constructed from historically confirmed fault cases. Each sample in the library contains the operating condition excitation vector at the time of the fault, the component state feature set, and the finally determined fault type code (e.g., "bearing #1 outer ring wear", "gearbox tooth surface pitting"). Matching retrieval can employ methods such as distance-based nearest neighbor search, support vector machine classification, or deep learning matching networks. Under the constraint of operating condition similarity, several historical fault samples most similar to the current diagnostic sample are found in the state feature space. Through voting or probability fusion mechanisms, the most probable fault type code is output.

[0054] S6: The confirmed diagnostic samples and their matching fault type codes are added as new records and incorporated into the fault mode feature library, and the parameters of the dynamic correlation are recalibrated based on the updated fault mode feature library.

[0055] The diagnostic conclusions (fault type codes) confirmed by engineers are fed back to the system. The correct diagnostic samples (including operating conditions, status characteristics, and confirmed fault codes) are standardized and added as new knowledge entries to the fault mode feature library, enriching its coverage. Simultaneously, the dynamic mapping model is incrementally learned or its parameters recalibrated periodically (e.g., monthly or quarterly) using a subset of normal samples from the expanded fault mode feature library and recently added normal navigation data. The recalibration process updates the model's weight parameters, enabling it to adapt to system characteristic drift caused by slow aging, maintenance, or modifications, ensuring the accuracy of health benchmarks under long-term operation, and thus giving the entire fault detection system the ability to continuously self-optimize and evolve.

[0056] In a preferred embodiment of the present invention, the specific process for generating the real-time operating condition excitation vector in step S2 is as follows:

[0057] The basic heading angle and speed setpoints are obtained by analyzing the gyrocompass heading signal and log speed signal output from the ship's integrated navigation system. Simultaneously, the current engine telegraph commands and corresponding engine speed setpoints are read from the digital interface of the main engine remote control system, forming the basic control parameters describing the ship's maneuvering intentions. In parallel, measured wind speed and direction data are obtained from the ship's weather station, and effective wave height, main wave direction, and average wave period data, derived through spectral analysis, are obtained from the ship's motion reference unit (MRU) or wave radar. These together constitute the environmental input parameters. Subsequently, based on the pre-entered ship general arrangement diagram, lines diagram, and mass distribution data under different load conditions, a frequency domain response model of the ship's motion and loads in regular waves is established using slice theory or three-dimensional potential flow theory. Real-time speed, wave direction, and wave period parameters are input into this model. Heave, pitch, and roll motion amplitude response operators at the current encountered frequency are obtained through querying or real-time calculation. Furthermore, characteristic values ​​of these motion amplitudes are synthesized using short-term forecasting methods. Simultaneously, characteristic values ​​of the vertical and horizontal wave moments acting on the hull beams are calculated. Regarding the propulsion system, the ship's propeller design drawings and stern hull lines are accessed. Based on the real-time speed, the average wake fraction at the propeller disk and the harmonic component amplitude representing circumferential non-uniformity are interpolated using empirical formulas or a database generated by computational fluid dynamics pre-calculation. For wind loads, based on real-time wind speed and direction, combined with the windward area and centroid position data of various parts of the ship's superstructure, and according to nautical wind force calculation standards, the additional drag generated by wind pressure and the turning moment on the ship's center of gravity are calculated. Finally, the calculated hull beam vertical bending moment, horizontal bending moment, propeller wake fraction, wake circumferential non-uniformity, wind-induced additional drag, and wind-induced turning moment—indicators with different physical dimensions—are normalized by dividing them by their respective design allowable values ​​or typical operating condition values. These are then arranged into a multi-dimensional vector according to a fixed order that reflects the load coupling logic, thus forming a standardized real-time operating condition excitation vector that comprehensively and quantitatively represents the current state of all major external mechanical and fluid loads.

[0058] This method transforms the complex, ambiguous, and difficult-to-use macroscopic navigation environment (such as "very rough seas") into a structured and computable quantitative description of physical loads directly related to the mechanical damage mechanism of the ship's propulsion system. Its direct purpose is to provide a stable and meaningful "operating condition context" for subsequent intelligent diagnostic models. Traditional methods often ignore these external excitations or treat them merely as simple classification labels (such as "severe sea conditions"), making it impossible to distinguish whether severe vibrations originate from equally severe, reasonable external excitations or from internal damage. This method, by systematically generating this excitation vector, effectively establishes a dynamic "reference frame" for fault diagnosis. It enables the system to answer a key question: "Under the current specific combination of wind, waves, speed, and maneuvering loads, what state range should a healthy propulsion system exhibit?" This elevates fault detection from blindly comparing static thresholds to analyzing the residual between the "actual system response" and the "theoretical healthy response under the current external inputs." Therefore, physical mechanisms are embedded from the source of data input, so that subsequent anomaly detection is no longer about finding absolute "abnormal signals", but about finding "signals that cannot be explained under specific physical conditions". This greatly improves the accuracy and reliability of distinguishing between benign high-load responses and early signs of malignant failures in the variable and harsh real navigation environment, laying a crucial data foundation for achieving the ultimate goal of reducing false alarms and providing early warnings under complex operating conditions.

[0059] In another preferred embodiment of the present invention, the specific process for generating the real-time component state feature set in step S2 is as follows:

[0060] The raw vibration acceleration waveform is acquired using piezoelectric accelerometers placed on key mechanical interfaces such as the main engine crankcase, thrust bearing housing, intermediate bearing housing, reduction gearbox housing, and generator bearing housing. Vibration velocity signals are acquired using magnetoelectric velocity sensors, and raw readings of oil temperature and pressure at each lubrication point are collected using platinum resistance temperature sensors and piezoresistive pressure transmitters. For the acquired vibration acceleration and velocity signals, a bandpass filter centered on the fundamental frequency and covering several harmonics of the measured component (e.g., the fundamental frequency corresponding to the main engine's rated speed) is used for filtering to retain frequency bands directly related to mechanical motion while suppressing high-frequency electrical noise and extremely low-frequency interference from the ship's rigid body motion. For signals such as temperature and pressure that change relatively slowly, a sliding window with a time width matching the system's thermal inertia or pressure fluctuation period is used for averaging filtering to smooth random fluctuations and retain their macroscopic trends. For the filtered vibration signal, the data is segmented at fixed durations (e.g., covering dozens of rotation cycles). For each segment, the root mean square value is calculated in the time domain to characterize the average energy, the peak value is calculated to capture the maximum instantaneous impact, the skewness is calculated to assess the asymmetry of the signal distribution (which may indicate impact failure), and the kurtosis is calculated to measure the sharpness of the signal distribution (sensitive to early surface damage such as pitting corrosion). In the frequency domain, a fast Fourier transform is performed on the data segment to obtain the spectrum, accurately identifying and extracting the amplitude at the current shaft fundamental frequency and its second, third, and other harmonic frequencies provided by the speed tracking module. Simultaneously, the percentage of vibration energy in several pre-defined frequency bands based on the fault mechanism (e.g., the high-frequency band related to bearing raceway defects) is calculated as the percentage of total energy. For the smoothed temperature and pressure signals, their average values ​​are calculated within a synchronized analysis time window to characterize the steady-state level, their maximum and minimum values ​​are recorded to understand the fluctuation range, and their change curves are fitted using linear regression to obtain the trend slope, thereby quantifying the rate of increase or decrease. Next, all time-domain features, frequency-domain features, means, extreme values, and trend slopes calculated by all sensors belonging to the same physical component (e.g., the same bearing) are arranged in a fixed, predefined order (e.g., vibration features before temperature and pressure features, time-domain before frequency-domain) to form the component's feature sub-vector. Then, the feature sub-vectors of all detected components, such as the main unit, bearings, gearbox, and generator, are concatenated in a global, predetermined order reflecting the power system's transmission chain sequence to form a very long composite vector. Finally, the mean and standard deviation of each dimension of this vector, calculated during long-term normal operation and stored in the historical database, are used to standardize each value of the current composite vector by subtracting the mean and dividing by the standard deviation, thereby generating a standardized real-time component state feature set with uniform dimensions and relatively stable numerical range.

[0061] Through systematic multi-source information fusion and multi-level feature engineering, a high-dimensional "digital fingerprint" capable of comprehensively and multi-scale characterizing the microscopic health status of a power system is constructed from noisy sensor readings. It's understandable that if only a single signal is observed, the faint signs of early faults are easily drowned out by background noise and normal fluctuations. This method, however, captures subtle changes in signal distribution patterns by extracting time-domain statistical features (such as kurtosis's sensitivity to shocks); detects periodic shocks caused by local defects by extracting frequency-domain features (such as specific harmonics or sideband amplitudes); and perceives parameter drifts that develop slowly but may foreshadow significant problems by calculating trend slopes. Combining multi-sensor features of the same component achieves a leap from single-point measurement to overall component status assessment; concatenating the feature sub-vectors of all components in the entire system constructs a global state view. The final standardization process is crucial, eliminating differences in the physical dimensions and orders of magnitude of different features, allowing subsequent models to measure changes across various dimensions equally and effectively. Therefore, the ultimate benefit of this approach is that it provides the entire fault diagnosis system with a high-quality input that is both rich in detail (sensitive to early faults) and well-normalized (facilitating model computation). This carefully constructed feature set enables the system to detect subtle, cooperative anomaly patterns across sensors and feature dimensions that are imperceptible at the level of a single raw signal, thus laying a solid data foundation for achieving the leap from "perceiving obvious faults" to "anticipating potential risks."

[0062] In another preferred embodiment of the present invention, the specific construction process of the dynamic mapping model in step S3 is as follows:

[0063] First, from the ship's historical data center, based on a unified timestamp index, all historical operating condition excitation vector archives stored over the past few years (each vector corresponds to a standardized external load description at a past moment) and historical component state feature sets synchronously recorded and archived by the monitoring system (each feature set corresponds to a standardized multidimensional state fingerprint at the same moment) are extracted. This ensures that the load vector at each historical moment is precisely aligned with the state feature set, thus forming tens of thousands of well-paired "operating condition-state" training samples. Subsequently, the entire training sample set is traversed, and the arithmetic mean and standard deviation of each dimension of all historical component state feature sets (e.g., main engine bearing vibration kurtosis, gearbox energy proportion in a certain frequency band, thrust bearing temperature trend slope, etc.) are calculated over the entire historical data range. Then, using the formula "the original value of the current dimension minus the mean of that dimension, then divided by the standard deviation of that dimension", a dimension-by-dimensional standardization transformation is performed on all historical component state feature sets in the entire dataset, so that the numerical distribution center of each feature dimension is zero and the scale is consistent. Then, following chronological order, the first 80% of the standardized samples are divided into a training subset and the last 20% into a validation subset to simulate the scenario where the model uses past data to predict future states in practical applications. Next, a fully connected deep neural network is designed, with the number of neurons in its input layer strictly equal to the length of the historical operating condition excitation vector (e.g., if it includes 10 indicators such as wave bending moment, wake fraction, and wind drag moment, then the input layer has 10 neurons), and the number of neurons in its output layer strictly equal to the dimension of the standardized historical component state feature set (e.g., if it includes 200 features extracted from the entire system, then the output layer has 200 neurons). The number of layers and neurons in the intermediate hidden layers are determined through hyperparameter adjustment. During training, batches of data are sequentially taken from the training subset, and the historical operating condition excitation vector is input into the network. The output generated by the network is compared with the paired, standardized historical component state feature set as the target, and the mean squared error loss is calculated. Then, the connection weight parameters of each layer in the network are adjusted through the backpropagation algorithm and an optimizer (such as Adam) to minimize this error. Simultaneously, after each training cycle, the performance of the current network model is evaluated using a validation subset, and its prediction error on unseen data is calculated. The changes in this validation error are monitored, and when it stops decreasing or even begins to increase over multiple consecutive training cycles (e.g., 10 cycles), it is determined that the model has sufficiently learned and may be overfitting, thus halting the training process. Finally, the system packages and saves all connection weights of the neural network at the time of training termination, along with the previously calculated means and standard deviations for state feature standardization, into a loadable model file. This complete model package constitutes a "dynamic mapping model" capable of dynamically predicting a health status benchmark based on input conditions.

[0064] By leveraging a large amount of historical normal data, the neural network learns a complex, non-linear mapping between "observable external conditions" (operating condition excitation vectors) and "observable internal responses" (component state feature sets). Instead of attempting to set static thresholds for every possible sea state combination, it automatically summarizes and generalizes a unified dynamic prediction function through a data-driven approach. When deployed in real-time, for any unprecedentedly complex operating condition combination acquired in real-time (e.g., a specific combination of wind, waves, and speed never before recorded), the model can, based on its learned generalization rules, calculate a matching range of state characteristics that theoretically "should be present in a healthy system." This allows subsequent anomaly detection steps to no longer simply compare real-time signals with a fixed, potentially unsuitable, old benchmark, but rather with an intelligent prediction benchmark "tailor-made for this specific situation." Therefore, it fundamentally improves the situational awareness and adaptive capabilities of the fault diagnosis system, making it possible to distinguish between "normal severe response under harsh conditions" and "abnormal equipment damage under any conditions". This directly solves the problem of high false alarms and false alarms caused by rigid references in traditional methods, and is the key technological cornerstone for achieving accurate and reliable early warning.

[0065] In another preferred embodiment of the present invention, the specific calculation process of the multidimensional deviation value in step S4 is as follows:

[0066] First, the standardized real-time component state feature set generated at the current moment is read from memory. This is an array containing hundreds of specific values, each corresponding to a specific state monitoring dimension, such as "the amplitude of the vibration of the #1 cylinder liner at the second harmonic" or "the current hourly linear trend slope of the lubricating oil temperature of the intermediate thrust bearing". Simultaneously, the real-time operating condition excitation vector at the current moment is input into the loaded dynamic mapping model. The model outputs an array of expected state feature ranges corresponding to the dimensions of the input vector. This range array provides two values ​​for each state feature dimension: a reasonable upper limit and a reasonable lower limit, which together constitute the theoretical healthy fluctuation range of that dimension under the current specific external load conditions. Then, an element-wise iterative comparison process is initiated: for each dimension in the feature set, such as the "second harmonic vibration amplitude" mentioned above, the program compares its specific real-time value (e.g., a normalized value of 0.85) with the expected upper limit (e.g., 0.75) and lower limit (e.g., 0.10) obtained from the model for that dimension. The program logic determines whether the value of 0.85 falls within or outside the interval [0.10, 0.75]. Since 0.85 is greater than the upper limit of 0.75, the program calculates its absolute distance from the nearest boundary (i.e., the upper limit of 0.75), obtaining a difference of 0.10. If the real-time value of a certain dimension falls within the expected range, the difference for that dimension is recorded as 0, indicating no deviation. After calculating the raw differences for all hundreds of feature dimensions, a predefined weighted coefficient vector is invoked. Each weighted component in this vector corresponds to a feature dimension, and its magnitude reflects the relative importance of that feature in the overall health assessment (for example, the weight of the "vibration peak" feature, which directly reflects mechanical shock, is greater than the weight of the "average oil temperature," which reflects slow changes). The program multiplies the raw difference for each dimension by its corresponding weighted component to obtain the weighted deviation for that dimension. This step amplifies the contribution of key abnormal signals and suppresses the influence of minor fluctuations. Finally, the squares of all weighted deviations are summed, and the square root of this sum is taken to obtain a single non-negative scalar value, namely the multidimensional deviation value. Mathematically, this value can be regarded as the comprehensive distance of the current overall state from the theoretical healthy range "hypersphere" in weighted Euclidean space.

[0067] In another preferred embodiment of the present invention, the specific construction process of the pre-stored fault mode feature library in step S5 is as follows:

[0068] First, maintenance engineers sift through the ship's management system's electronic logs to identify all verified and closed-loop fault work orders. For example, a maintenance report confirming "poor lubrication of the starboard intermediate bearing leading to excessive temperature" is used to extract a clear "fault type identifier" code, such as "BB-01". The system program then automatically parses the precise start and end timestamps of the fault event and uses this as an index to retrieve and extract all "historical operating condition excitation vectors" (recording loads such as wave, wind, and speed at the time) and "historical component state feature sets" (recording the overall system state, including the bearing temperature and vibration) synchronously recorded and archived by the data acquisition system during the fault's occurrence and development period. The program automatically pairs and packages the load vectors and state feature sets at the same moment and binds them to the fault code "BB-01," thereby generating a complete information package containing "when, under what load conditions, what abnormal state occurred, and what fault was ultimately identified"—an initial fault mode sample. This process traverses all historical fault records, accumulating to form an initial sample library covering various types of faults. Next, to eliminate feature dimensions and ensure all samples use the same benchmark as the current operational monitoring system, the program calls the global "mean" and "standard deviation" parameters calculated and saved during the construction of the dynamic mapping model, performing a completely consistent standardization transformation on the "historical component state feature set" portion of all initial samples. Subsequently, the standardized state feature set is recombinated with its original corresponding "historical operating condition excitation vector" (this vector itself is already in a standardized format) to generate a completely new "standardized fault feature vector" with all dimensions within the same numerical range. Based on this, to build an efficient and representative feature library, the system performs cluster analysis on all samples under each fault code, or selects several of the most representative samples based on time and severity; for example, for faults like "BB-01," 20 typical instances occurring in different seasons and sea states are selected. Finally, the program structurally stores these selected, standardized "fault feature vectors" and their corresponding "fault type identifiers" into a dedicated database or feature vector index file. This high-speed queryable knowledge base constitutes the "pre-stored fault mode feature library" upon which the system relies for intelligent diagnosis.

[0069] Traditional methods often only set alarm thresholds but fail to answer the question of "what exactly is the fault in the current anomaly?" By systematically building this feature library, we essentially precipitate each confirmed fault and its complete context (load conditions + system state) as standard cases. Standardization using parameters consistent with the dynamic mapping model ensures that real-time monitoring data and historical case data are within the same fully comparable standard space, a prerequisite for accurate matching. Selecting multiple typical patterns for each type of fault fully considers the potential state differences of the same type of fault under different external conditions, improving the coverage and robustness of pattern matching. This allows the diagnostic process to go beyond simply "discovering anomalies" to further "identifying anomalies." When a new anomaly is detected in real time, its features can be quickly compared with the vast number of known patterns in the feature library to find the most similar historical cases. This provides a specific and understandable fault type prediction for the current alarm, such as "this vibration and temperature anomaly is highly similar to historical cases 15 and 38 (both bearing lubrication problems)." This greatly enhances the operability and guiding value of the diagnostic results, completing the final step from anomaly monitoring to intelligent diagnosis.

[0070] In another preferred embodiment of the present invention, the specific process of parameter recalibration in step S6 is as follows:

[0071] Obtain diagnostic samples confirmed by fault type coding matching, and standardize the real-time component status feature set in the diagnostic samples; add the standardized diagnostic samples to the fault mode feature library; extract historical operating condition excitation vectors and corresponding historical component status feature sets from the updated fault mode feature library and historical operation database to form a new training sample set; recalculate the mean and standard deviation of each dimension of the historical component status feature set in the new training sample set, and use the recalculated mean and standard deviation parameters to standardize the new training sample set.

[0072] The standardized new training sample set is input into the neural network structure contained in the dynamic mapping model, and the connection weight parameters of the neural network are adjusted using the backpropagation algorithm. When the prediction error of the neural network on the independent validation dataset reaches the convergence state, the adjusted neural network connection weight parameters and the mean and standard deviation of each dimension of the recalculated historical component state feature set are saved to complete the parameter recalibration of the dynamic mapping model.

[0073] The present invention also includes a computer device comprising a memory and a processor, the memory storing a computer program, the processor executing the computer program to implement a fault type detection method for a ship propulsion system as described in any of the above claims.

[0074] The present invention also includes a computer-readable storage medium storing a computer program that, when executed by a processor, implements a method for detecting fault types in a marine propulsion system as described in any one of the above claims.

[0075] The foregoing has provided a detailed description of one embodiment of the present invention, but this description is merely a preferred embodiment and should not be construed as limiting the scope of the invention. All equivalent variations and modifications made within the scope of the claims of this invention should still fall within the patent coverage of this invention.

Claims

1. A method for detecting fault types in a marine propulsion system, characterized in that, Includes the following steps: S1: Collect environmental data and operation command data during the ship's navigation process to form a navigation condition data stream, and simultaneously collect multi-channel sensor signals of rotating mechanical components in the power system to form a condition monitoring data stream; S2: Based on the ship's motion mechanics model, the navigation condition data stream is analyzed to obtain the real-time condition excitation vector. At the same time, the state monitoring data stream is transformed in the time and frequency domain and statistical features are extracted to generate a real-time component state feature set. S3: Construct a dynamic mapping model based on the correspondence between the operating condition excitation vector and the component state feature set in historical data. Input the real-time operating condition excitation vector into the dynamic mapping model to obtain the expected state feature range. S4: Compare the generated real-time component state feature set with the expected state feature range obtained by mapping through dynamic correlation, and calculate the multidimensional deviation value between the two. S5: When the multidimensional deviation value exceeds the preset dynamic threshold, a diagnostic sample is formed by combining the corresponding real-time operating condition excitation vector and the real-time component state feature set. The diagnostic sample is matched and retrieved with the pre-stored fault mode feature library, and the fault type code associated with the retrieval result is output. S6: The confirmed diagnostic samples and their matching fault type codes are added as new records and incorporated into the fault mode feature library, and the parameters of the dynamic correlation are recalibrated based on the updated fault mode feature library.

2. The method for detecting fault types in a ship propulsion system according to claim 1, characterized in that, In S2, the specific process for generating the real-time operating condition excitation vector is as follows: The heading angle, speed setpoint, and main engine speed command from the operation command data are extracted as basic control parameters. Simultaneously, wave height, wave direction, period, wind speed, and wind direction from the environmental data are extracted as environmental input parameters. A ship motion response model is established based on the ship's main dimensions, hull lines, and mass distribution data. The basic control parameters and environmental input parameters are input into the model to calculate the heave, pitch, and roll amplitudes of the ship in the waves, as well as the wave-induced vertical and horizontal bending moments of the hull beams. Combining the design parameters of the ship's propulsion system and the real-time speed, the flow field characteristics at the stern of the hull are calculated based on the ship's hull line data and speed to determine the wake fraction and its circumferential non-uniformity at the propeller disk. Simultaneously, based on real-time wind speed, wind direction, and the windward area of ​​the ship's superstructure, the additional resistance and turning moment generated by wind load on the ship are calculated. The vertical bending moment and horizontal bending moment of the hull beam, the wake fraction of the propeller disk, the circumferential non-uniformity of the wake, the wind-induced additional resistance, and the wind-induced turning moment are used as standardized indicators, arranged and normalized according to a predefined vector structure to form a standardized real-time operating condition excitation vector.

3. The method for detecting fault types in a ship propulsion system according to claim 1, characterized in that, In S2, the specific process for generating the real-time component state feature set is as follows: The system collects raw signals from acceleration sensors, velocity sensors, temperature sensors, and pressure sensors deployed on various rotating components of the power system; it performs preprocessing on the raw signals according to their signal types, performs bandpass filtering on vibration-type raw signals, and performs moving average filtering on temperature and pressure-type raw signals. The preprocessed vibration signal is segmented, and time-domain statistical features and frequency-domain spectral features are calculated for each data segment. The time-domain statistical features include effective value, peak value, skewness and kurtosis, and the frequency-domain spectral features include fundamental frequency and harmonic amplitude and energy proportion of multiple preset frequency bands. For the preprocessed temperature and pressure signals, calculate their mean, extreme values, and trend slope within an analysis window; combine the time-domain statistical features, frequency-domain spectral features, mean, extreme values, and trend slope extracted from all sensors of the same component into a feature sub-vector of the component; concatenate the feature sub-vectors of all components of the power system into a comprehensive vector in a predetermined order, and normalize each dimension of the comprehensive vector to generate a standardized real-time component state feature set.

4. The method for detecting fault types in a ship propulsion system according to claim 1, characterized in that, In S3, the specific construction process of the dynamic mapping model is as follows: Extract historical operating condition excitation vectors and corresponding historical component state feature sets from the historical database, establish a pairing relationship between the two to form a training sample set; calculate the mean and standard deviation of each dimension of the historical component state feature set in the training sample set, and use the mean and standard deviation to standardize the historical component state feature set; divide the standardized training sample set into a training subset and a validation subset in chronological order. Construct a neural network whose input layer dimension matches the dimension of the historical operating condition stimulus vectors, and whose output layer dimension matches the dimension of the standardized historical component state feature set. Use the historical operating condition stimulus vectors from the training subset as input and the corresponding standardized historical component state feature set as the target output, and train the neural network using the backpropagation algorithm. Evaluate the trained neural network using a validation subset, and stop training when the prediction error on the validation subset no longer decreases within consecutive training cycles. Save the connection weight parameters, mean, and standard deviation of the neural network at this point to form a dynamic mapping model.

5. The method for detecting fault types in a ship propulsion system according to claim 1, characterized in that, In S4, the specific calculation process for the multidimensional deviation value is as follows: Obtain the value of each dimension of the real-time component state feature set, and simultaneously obtain the upper and lower limits of each dimension corresponding to the expected state feature range output by the dynamic mapping model; for each feature dimension, compare the value in the real-time component state feature set with the upper and lower limits of the expected state feature range of that dimension, and calculate the difference between the real-time value and the range boundary. When the real-time value is within the expected state feature range, the difference in that dimension is zero; when the real-time value exceeds the expected state feature range, the difference in that dimension is the absolute distance between the real-time value and the nearest range boundary value. A predefined weight coefficient vector is used, and each component of the weight coefficient vector is multiplied by the difference in the corresponding feature dimension to obtain the weighted deviation of each dimension. The sum of squares of the weighted deviations of all feature dimensions is calculated, and the square root of the sum of squares is taken as the multidimensional deviation value.

6. The method for detecting fault types in a ship propulsion system according to claim 1, characterized in that, In S5, the specific construction process of the pre-stored fault mode feature library is as follows: From the historical maintenance records and abnormal event reports of the ship's power system, extract confirmed fault events and their corresponding fault type identifiers; based on the occurrence time of the fault events, retrieve the complete historical operating condition excitation vectors and historical component state feature sets synchronously recorded within that time period from the historical database; combine each retrieved set of historical operating condition excitation vectors and historical component state feature sets, and associate them with the corresponding fault type identifiers to form an initial fault mode sample. The historical component state feature set in all initial fault mode samples is subjected to dimensionless standardization. The parameters used in the standardization process are consistent with the mean and standard deviation calculated during the construction of the dynamic mapping model. The standardized historical operating condition excitation vector and the standardized historical component state feature set are then recombined to generate a standardized fault feature vector. For each type of fault, a preset number of standardized fault feature vectors are selected as typical patterns for that type of fault. All typical patterns and their corresponding fault type identifiers are stored in a structured manner, and the structured storage data set constitutes the pre-stored fault pattern feature library.

7. The method for detecting fault types in a ship propulsion system according to claim 1, characterized in that, In step S6, the specific process of parameter recalibration is as follows: Obtain diagnostic samples confirmed by fault type coding matching, and standardize the real-time component status feature set in the diagnostic samples; add the standardized diagnostic samples to the fault mode feature library; extract historical operating condition excitation vectors and corresponding historical component status feature sets from the updated fault mode feature library and historical operation database to form a new training sample set; recalculate the mean and standard deviation of each dimension of the historical component status feature set in the new training sample set, and use the recalculated mean and standard deviation parameters to standardize the new training sample set. The standardized new training sample set is input into the neural network structure contained in the dynamic mapping model, and the connection weight parameters of the neural network are adjusted using the backpropagation algorithm. When the prediction error of the neural network on the independent validation dataset reaches the convergence state, the adjusted neural network connection weight parameters and the mean and standard deviation of each dimension of the recalculated historical component state feature set are saved to complete the parameter recalibration of the dynamic mapping model.

8. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements a fault type detection method for a ship propulsion system as described in any one of claims 1-7.

9. A computer-readable storage medium storing a computer program, characterized in that, When the computer program is executed by the processor, it implements a fault type detection method for a ship propulsion system as described in any one of claims 1-7.

Images