A ship propulsion motor rotor fault diagnosis method and system
By introducing a joint loss function of class-weighted focus loss, adversarial domain adaptive loss, and class center constraint loss, along with a pseudo-label self-training mechanism, into the fault diagnosis of ship propulsion motor rotors, the problems of domain adaptation and class imbalance are solved, achieving accurate and robust fault diagnosis. This method is suitable for intelligent monitoring and fault early warning of ship propulsion motor status.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- HENAN JIAOTONG PORT & SHIPPING CO LTD
- Filing Date
- 2025-12-02
- Publication Date
- 2026-06-19
AI Technical Summary
Existing technologies for diagnosing rotor faults in marine propulsion motors suffer from domain adaptation and class imbalance issues, which lead to a decline in the diagnostic performance of the model in practical applications and make it unable to meet the accurate identification requirements in complex real-world scenarios.
A joint loss function, including class-weighted focus loss, adversarial domain adaptive loss, and class center constraint loss, is adopted. Through an adversarial training mechanism, the model learns domain-invariant and fault category-sensitive feature representations. Pseudo-label self-training and online adaptive mechanisms are introduced to optimize the performance of the diagnostic model.
It achieves accurate and robust diagnosis in complex real-world scenarios, improves the ability to identify rare faults, ensures the long-term adaptability and reliability of the diagnostic model, and is suitable for intelligent monitoring and fault early warning of ship propulsion motor status.
Smart Images
Figure CN122241330A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of ship electric propulsion and intelligent monitoring technology, and in particular to a method and system for diagnosing rotor faults in ship propulsion motors. Background Technology
[0002] With the rapid development of green and intelligent ships, electric propulsion systems have become the core power unit of modern ships due to their advantages such as flexible layout, high propulsion efficiency, and low vibration and noise. As a key power output device in electric propulsion systems, the operating status of marine propulsion motors directly affects the ship's navigation safety and energy efficiency management. However, during long-term service, propulsion motors are continuously subjected to complex marine environmental disturbances, dynamic load fluctuations, temperature rise effects, and mechanical vibration shocks, making them highly susceptible to typical faults such as rotor bar breakage, air gap eccentricity, rotor wear, and axial imbalance. If these rotor-related faults are not detected and diagnosed in a timely manner, they may trigger a chain reaction, leading to equipment shutdown and even serious safety accidents. Therefore, intelligent, accurate, and reliable fault diagnosis of marine propulsion motor rotor conditions is of paramount practical significance.
[0003] Current motor fault diagnosis methods can be broadly categorized into two types: signal processing-based methods and machine learning-based methods. Signal processing-based methods (such as Fast Fourier Transform, wavelet analysis, and Empirical Mode Decomposition) rely on expert knowledge, identifying faults by analyzing the spectral characteristics and envelope demodulation components of signals like current and vibration. While intuitive, these methods are sensitive to environmental noise, their feature extraction quality heavily depends on human experience, and they exhibit poor adaptability under varying operating conditions.
[0004] Machine learning-based methods, especially deep learning models (such as convolutional neural networks and long short-term memory networks), can automatically learn hierarchical features from raw data, reducing reliance on manual features to some extent and achieving good results on standard laboratory datasets. However, these methods typically rely on a key assumption: that the data used in the model training phase (source domain) and the data collected in the actual application scenario (target domain) follow independent and identically distributed (i.i.d.) distributions. In actual ship operations, this assumption is often difficult to uphold. Significant differences exist between different ships in terms of propulsion motor models, sensor installation methods, navigation conditions (such as heavy load, light load, and wind and wave interference), and equipment aging, leading to severe distribution shifts between the source and target domains. Directly applying a model trained under one operating condition or on another results in a sharp decline in diagnostic performance. This "domain difference" problem severely restricts the engineering deployment and widespread adoption of data-driven intelligent diagnostic models.
[0005] Furthermore, a particularly prominent but often overlooked challenge is class imbalance. During normal operation, motors are in a healthy state for the vast majority of the time, resulting in an extremely limited number of fault samples available for training (especially severe but rare faults such as rotor bar breakage), with the ratio to normal samples potentially as low as 1:50 or even 1:100. This severe class imbalance causes traditional accuracy-oriented classification models to be dominated by the majority class (normal samples) during training. The model tends to predict all samples as normal, significantly reducing the recognition rate of critical minority class faults (i.e., the faults we most want to identify), leading to a risk of missed detections and failing to meet the stringent requirements of early fault warning. Summary of the Invention
[0006] In view of the above-mentioned shortcomings in the field of ship electric propulsion and intelligent monitoring technology, the present invention provides a method and system for diagnosing ship propulsion motor rotor faults. This innovative fault diagnosis method can effectively solve both domain adaptation and class imbalance problems, so as to achieve accurate and robust identification of ship propulsion motor rotor faults in complex real-world scenarios.
[0007] To achieve the above objectives, the embodiments of the present invention adopt the following technical solutions: A method for diagnosing rotor faults in a ship propulsion motor includes the following steps: Acquire tagged source domain ship propulsion motor operating data and untagged target domain ship propulsion motor operating data; Construct a diagnostic model that includes a feature extractor, a classifier, and a domain discriminator; The diagnostic model is trained with the goal of minimizing the joint loss function. Through adversarial training between the domain discriminator and the feature extractor, the model learns a domain-invariant feature representation that is sensitive to the fault category. The joint loss function includes at least class-weighted focus loss, adversarial domain adaptive loss, and class center constraint loss. Input the target domain data to be diagnosed into the trained diagnostic model, and output fault category diagnostic information.
[0008] According to one aspect of the invention, the operating data includes one or more of stator current signals, terminal voltage signals, and vibration signals.
[0009] According to one aspect of the invention, the adversarial domain adaptive loss is implemented through a gradient inversion layer; during the forward propagation of model training, the gradient inversion layer directly transmits the input features, and during the backward propagation, the gradient inversion layer multiplies the gradient from the domain discriminator by a negative coefficient before backpropagating.
[0010] According to one aspect of the present invention, the formula for calculating the class-weighted focus loss is:
[0011] in, Focus adjustment factor, used to emphasize difficult samples; Balance the weights for each category; yic is the probability that the (i)th sample is predicted to be of class (c); yic is the one-hot encoding of the true label of the (i)th sample; This is the focusing factor.
[0012] According to one aspect of the present invention, the class center constraint loss is used to minimize the distance between a depth feature and its corresponding class center, and the formula for calculating the constraint loss is:
[0013] in, Number of source domain samples; For feature extractors; Input the (i)th source domain sample; Let be the feature center of the true class label of the (i)th sample.
[0014] According to one aspect of the invention, after training the diagnostic model with the objective of minimizing the joint loss function, a pseudo-label self-training step is further included: For the target domain data, calculate the predicted probability distribution of the model output and take the maximum probability value pmax; If pmax is greater than the preset confidence threshold τ, then a pseudo-label is assigned to the target domain sample and added to the training set; The diagnostic model is further optimized using an expanded training set that includes pseudo-labeled target domain samples.
[0015] According to one aspect of the invention, the confidence threshold Linear attenuation is used, and the specific formula is as follows:
[0016] in, For the current round, This is the maximum number of training rounds.
[0017] According to one aspect of the invention, after outputting the fault category diagnostic information, an online adaptive step is further included: Real-time calculation of the difference between the feature distribution of newly input target domain data and the feature distribution of existing target domain data. ; When the difference Exceeding the preset threshold When this occurs, a lightweight retraining of the diagnostic model using new data is triggered.
[0018] According to one aspect of the invention, the feature extractor is composed of a one-dimensional convolutional neural network for extracting deep features from raw time-series signals or time-frequency images.
[0019] A fault diagnosis system for a marine propulsion motor rotor includes: The acquisition module acquires tagged source domain ship propulsion motor operating data and untagged target domain ship propulsion motor operating data. The model module constructs a diagnostic model that includes a feature extractor, a classifier, and a domain discriminator. The training module trains the diagnostic model with the goal of minimizing the joint loss function. Through adversarial training between the domain discriminator and the feature extractor, the model learns a domain-invariant feature representation that is sensitive to the fault category. The joint loss function includes at least class-weighted focal loss, adversarial domain adaptive loss, and class center constraint loss. The diagnostic module takes the target domain data to be diagnosed and inputs it into the trained diagnostic model, then outputs fault category diagnostic information.
[0020] The advantages of this invention are as follows: Through the above technical solution, it innovatively and collaboratively addresses the two key challenges of "domain adaptation" and "class imbalance," thereby achieving accurate and robust diagnosis in complex real-world scenarios. This method designs a joint loss function comprising class-weighted focus loss, adversarial domain adaptation loss, and class center constraint loss, and leverages an adversarial training mechanism to drive the model to learn a "domain-invariant" feature representation that is highly sensitive to fault categories. This not only effectively eliminates data distribution differences caused by different operating conditions, ships, or sensors, but also significantly improves the ability to identify rare sample categories, ensuring the diagnostic model's excellent generalization performance and reliability. By introducing a pseudo-label self-training strategy (combined with a linearly decaying confidence threshold) and an online adaptive triggering mechanism, the system can continuously utilize unlabeled target domain data after model deployment. This design allows the diagnostic model to not only gradually improve its performance on specific target ships through self-training, but also to proactively perform lightweight retraining when significant data distribution drift is detected, thereby achieving long-term, intelligent tracking and adaptation of the ship's propulsion motor status, greatly enhancing the engineering practical value and lifecycle of this method. Attached Figure Description
[0021] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.
[0022] Figure 1This is a flowchart illustrating a method and system for diagnosing rotor faults in a marine propulsion motor, as described in this invention. Detailed Implementation
[0023] 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.
[0024] like Figure 1 As shown, a method for diagnosing rotor faults in a ship propulsion motor includes the following steps: Step 10: Obtain the tagged source domain ship propulsion motor operation data and the untagged target domain ship propulsion motor operation data; Step 20: Construct a diagnostic model that includes a feature extractor, a classifier, and a domain discriminator; Step 30: The diagnostic model is trained with the goal of minimizing the joint loss function. Through adversarial training between the domain discriminator and the feature extractor, the model learns a domain-invariant feature representation that is sensitive to the fault category. The joint loss function includes at least class-weighted focus loss, adversarial domain adaptive loss, and class center constraint loss. Step 40: Input the target domain data to be diagnosed into the trained diagnostic model and output fault category diagnostic information.
[0025] Step 10 specifically includes the following steps: Step 11: Signal Acquisition and Preprocessing
[0026] Collect multi-condition signals (stator current, terminal voltage, vibration signal) from the propulsion motor, and define the source and target domain datasets as follows:
[0027] in: ● : indicates a length of The number of channels is Timing signals; ● The category label for the source domain sample (e.g., normal, broken rotor bar, eccentric, etc.); ● Total number of categories.
[0028] After bandpass filtering and sliding window segmentation, the input matrix is constructed:
[0029] in and These are the mean and standard deviation, respectively, used to achieve normalization. Step 12: Time-Frequency Transformation and Feature Enhancement
[0030] Perform a short-time Fourier transform (STFT) on the signal to obtain a time-frequency domain representation:
[0031] in The window function is used to obtain the characteristic matrix. .
[0032] To alleviate sample imbalance, a mixup augmentation method is used for minority class data:
[0033] Step 20 specifically includes the following steps: deep feature extraction and classification. Building a feature extractor With classifier .
[0034] Network structure is defined as:
[0035] in: ●: Convolutional layer weights and biases; ●: Convolution operation; ●: : Non-linear activation function (ReLU).
[0036] The classifier outputs the predicted probability:
[0037] in For the first Logarithmic odds of a class.
[0038] Step 30 specifically includes the following steps: Class-Imbalance Adversarial Domain Adaptive Optimization The goal of the model is to simultaneously minimize classification error and domain distribution differences, thereby achieving cross-domain feature alignment.
[0039] The overall loss function is:
[0040] The following are the definitions of each item.
[0041] (1) Weighted Focal Loss
[0042] Used to address class imbalance. Its expression is:
[0043] in:
[0044] Parameter definition: ● Focus adjustment factor, used to emphasize difficult samples; ● Balance the weights for each category; ● Let be the probability that the (i)th sample is predicted to be of class (c); ● Encode the true label (one-hot) of the (i)th sample; ● This is the focusing factor.
[0045] This allows minority classes (such as rotor bars) to receive higher weights in gradient updates.
[0046] (2) Domain Adversarial Loss To reduce the distribution difference between the source and target domains, a domain discriminator is introduced. .
[0047] Its output is the probability that a sample belongs to the source domain:
[0048] Adversarial loss is defined as:
[0049] The training objective employs adversarial game-like optimization:
[0050] Gradient sign flipping is achieved by adding a gradient inversion layer (GRL) between the feature extractor and the domain discriminator:
[0051] This forces We learned the domain-invariant feature representation.
[0052] To enhance intra-class clustering, a feature center vector is defined:
[0053] The constraint loss is:
[0054] in, Number of source domain samples; For feature extractors; Input the (i)th source domain sample; Let be the feature center of the true category label of the (i)th sample.
[0055] Its gradient update rule is as follows:
[0056] (3) Overall optimization objectives The final joint optimization objective is:
[0057] Parameter updates use the Adam optimizer:
[0058] Steps 31-33: Pseudo-labels and semi-supervised self-training Within the target domain, calculate the class confidence score for each sample:
[0059] like Then assign a pseudo-label:
[0060] Constructing a pseudo-labeled dataset:
[0061] And update the overall training set:
[0062] threshold Linear attenuation is used:
[0063] in For the current round, This is the maximum number of training rounds.
[0064] Step 40 specifically includes the following steps: online diagnosis and drift detection. After deploying the model in the ship's propulsion system, signals are input in real time. Calculate the characteristic distribution difference index (K-S distance):
[0065] when Lightweight retraining is triggered when the threshold is typically set to 0.1–0.2.
[0066] This ensures the model's long-term adaptability.
[0067] Steps 41-42: Diagnostic Output and Decision Making Model outputs fault category probability Smooth output through a voting mechanism:
[0068] The rotor condition diagnosis results are obtained, enabling online intelligent fault identification and early warning.
[0069] The advantages of this invention are as follows: The above-described scheme successfully achieves a highly efficient intelligent diagnostic effect that can simultaneously address domain distribution differences and class imbalance. Its core advantage lies in introducing a joint optimization objective that integrates class-weighted focus loss, adversarial domain adaptive loss, and class center constraint loss. This fundamentally and collaboratively solves two key problems that lead to degraded model performance: domain shift and class imbalance. This mechanism not only effectively aligns feature distributions under different operating conditions and equipment, endowing the model with strong cross-domain generalization capabilities, but also significantly improves the sensitivity and recall rate for identifying rare faults such as rotor bar breakage, thereby achieving accurate diagnosis of various fault states, especially minority class faults.
[0070] Furthermore, this invention possesses significant engineering practical value and long-term adaptability. The introduction of pseudo-label self-training and online distribution drift detection mechanisms transforms the model from static to an intelligent system capable of continuous learning and self-optimization after actual deployment. This effectively overcomes the challenge of slow data distribution changes (concept drift) caused by long-term equipment aging, ensuring the long-term reliability of the diagnostic system in complex real-world scenarios. In summary, this invention provides a complete solution that is accurate, robust, and adaptive, effectively overcoming the limitations of existing technologies and demonstrating significant application prospects in the field of intelligent condition monitoring and fault early warning for ship electric propulsion systems.
[0071] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.
Claims
1. A method for diagnosing rotor faults in a marine propulsion motor, characterized in that, Includes the following steps: Acquire tagged source domain ship propulsion motor operating data and untagged target domain ship propulsion motor operating data; Construct a diagnostic model that includes a feature extractor, a classifier, and a domain discriminator; The diagnostic model is trained with the goal of minimizing the joint loss function. Through adversarial training between the domain discriminator and the feature extractor, the model learns a domain-invariant feature representation that is sensitive to the fault category. The joint loss function includes at least class-weighted focus loss, adversarial domain adaptive loss, and class center constraint loss. Input the target domain data to be diagnosed into the trained diagnostic model, and output fault category diagnostic information.
2. The method for diagnosing rotor faults in a ship propulsion motor according to claim 1, characterized in that, The operating data includes one or more of the following: stator current signal, terminal voltage signal, and vibration signal.
3. The method for diagnosing rotor faults in a ship propulsion motor according to claim 1, characterized in that, The adversarial domain adaptive loss is implemented through a gradient inversion layer. During the forward propagation of model training, the gradient inversion layer directly transmits the input features. During the backward propagation, the gradient inversion layer multiplies the gradient from the domain discriminator by a negative coefficient and then propagates it back.
4. The method for diagnosing rotor faults in a ship propulsion motor according to claim 1, characterized in that, The formula for calculating the weighted focus loss is as follows: in, Focus adjustment factor, used to emphasize difficult samples; Balance the weights for each category; yic is the probability that the (i)th sample is predicted to be of class (c); yic is the one-hot encoding of the true label of the (i)th sample; This is the focusing factor.
5. The method for diagnosing rotor faults in a ship propulsion motor according to claim 1, characterized in that, The class center constraint loss is used to minimize the distance between deep features and their corresponding class centers. The formula for calculating the constraint loss is as follows: in, Number of source domain samples; For feature extractors; Input the (i)th source domain sample; Let be the feature center of the true class label of the (i)th sample.
6. The method for diagnosing rotor faults in a ship propulsion motor according to claim 1, characterized in that, After training the diagnostic model with the objective of minimizing the joint loss function, a pseudo-label self-training step is also included: For the target domain data, calculate the predicted probability distribution output by the model and take the maximum probability value. ; like Greater than the preset confidence threshold Then, a pseudo-label is assigned to the sample in the target domain and added to the training set; The diagnostic model is further optimized using an expanded training set that includes pseudo-labeled target domain samples.
7. The method for diagnosing rotor faults in a ship propulsion motor according to claim 6, characterized in that, The confidence threshold Linear attenuation is used, and the specific formula is as follows: in, For the current round, This is the maximum number of training rounds.
8. The method for diagnosing rotor faults in a ship propulsion motor according to claim 1, characterized in that, Following the output of fault category diagnostic information, an online adaptive step is also included: Real-time calculation of the difference between the feature distribution of newly input target domain data and the feature distribution of existing target domain data. ; When the difference Exceeding the preset threshold When this occurs, a lightweight retraining of the diagnostic model using new data is triggered.
9. The method for diagnosing rotor faults in a ship propulsion motor according to claim 1, characterized in that, The feature extractor is composed of a one-dimensional convolutional neural network and is used to extract deep features from the original time-series signal or time-frequency image.
10. A fault diagnosis system for a marine propulsion motor rotor, characterized in that, A method for diagnosing rotor faults in a ship propulsion motor according to any one of claims 1 to 9 includes: The acquisition module acquires tagged source domain ship propulsion motor operating data and untagged target domain ship propulsion motor operating data. The model module constructs a diagnostic model that includes a feature extractor, a classifier, and a domain discriminator. The training module trains the diagnostic model with the goal of minimizing the joint loss function. Through adversarial training between the domain discriminator and the feature extractor, the model learns a domain-invariant feature representation that is sensitive to the fault category. The joint loss function includes at least class-weighted focal loss, adversarial domain adaptive loss, and class center constraint loss. The diagnostic module takes the target domain data to be diagnosed and inputs it into the trained diagnostic model, then outputs fault category diagnostic information.