Ship intelligent fault diagnosis method and system based on open label space identification
By using a deep neural network model based on an open label space and a dynamic knowledge base expansion mechanism, the problem of insufficient identification of unknown faults under the traditional closed label space is solved, realizing the proactive perception and adaptive learning of unknown faults, and improving the reliability and safety of the ship fault diagnosis system.
Patent Information
- Application Number
- CN202511799904.X
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-12-02
- Publication Date
- 2026-05-05
AI Technical Summary
Existing ship fault diagnosis models based on deep learning or pattern recognition rely on the assumption of a closed label space, which cannot effectively identify unknown faults not included in the training samples, leading to misclassification and potential navigation safety risks.
An open label space-based identification method is adopted, which extracts discriminative embedding features from multi-source sensor signals through a deep neural network model. By combining the center loss function and the open set loss function, it can achieve accurate identification of known faults and adaptive detection of unknown faults. Furthermore, the fault knowledge base is dynamically expanded through DBSCAN clustering and incremental learning.
It enables fault identification in an open label space, improves the environmental adaptability and safety of the ship fault diagnosis system, reduces long-term maintenance costs, and enhances the ability to distinguish fault modes and the accuracy of diagnosis.
Smart Images

Figure CN121980326A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of ship intelligent monitoring and control technology, and in particular relates to a ship intelligent fault diagnosis method and system based on open tag space recognition. Background Technology
[0002] Green and intelligent ships have become an important direction for future ship development, and the safe and stable operation of their core systems (such as electric propulsion systems) is directly related to the ship's navigation reliability. Currently, ship fault diagnosis technology mainly relies on intelligent diagnostic models based on deep learning or pattern recognition. However, these methods are generally based on the assumption of a closed label space, meaning that all possible fault categories need to be pre-defined during the model training phase.
[0003] In actual operation, a ship's electromechanical equipment operates in a complex and ever-changing environment, and its operational status is easily affected by various dynamic factors such as load fluctuations, electromagnetic interference, mechanical wear, and temperature drift. Therefore, throughout the equipment's entire lifecycle, new and unknown fault types not included in the training samples are highly likely to occur. Traditional diagnostic models based on closure assumptions, when faced with these unknown faults, will force them to be classified as a known fault type, thus producing incorrect classification results. More seriously, the model may sometimes assign high confidence to such misclassifications, causing the system to fail to trigger effective alarms or produce false diagnoses. This not only significantly reduces maintenance efficiency but also poses potential risks to the ship's navigational safety. Summary of the Invention
[0004] In view of the above-mentioned shortcomings in the field of ship intelligent monitoring and control technology, the present invention provides a ship intelligent fault diagnosis method based on open label space recognition. This method can overcome the limitations of the closed set assumption and achieve intelligent diagnosis of known faults and adaptive detection of unknown faults in an open label space, thereby improving the environmental adaptability and safety assurance capabilities of ship fault diagnosis systems.
[0005] To achieve the above objectives, the embodiments of the present invention adopt the following technical solutions:
[0006] A ship intelligent fault diagnosis method based on open label space recognition includes:
[0007] Acquire time-series signals from multiple sources of sensors in the ship's system and construct a training sample set;
[0008] The training sample set is used to extract features to obtain the discriminative embedding features of the samples; during the training process, a center loss function is used to enhance the clustering characteristics of similar features in the embedding space.
[0009] Based on the feature extraction, a classifier is trained and an open set loss function is introduced as part of the overall objective function for model training; the overall objective function is composed of the open set loss function and the center loss function.
[0010] For the sample to be diagnosed, its feature embedding is extracted, and its distance to the feature centers of all known fault categories is calculated; if the minimum distance is greater than a preset threshold, it is determined to be an unknown fault.
[0011] If the fault is determined to be a known fault, the corresponding fault category is output; if the fault is determined to be an unknown fault, an alarm is triggered; and the fault knowledge base of the model is dynamically expanded and updated based on the accumulation of the unknown fault samples.
[0012] According to one aspect of the present invention, the acquisition of multi-source sensor timing signals of a ship system and the construction of training samples includes: collecting multi-source sensor signals from at least one system in the ship's electric propulsion system, engine room power equipment, energy conversion device, and auxiliary electromechanical system; segmenting the signals into sample segments of length L using a sliding time window mechanism; and normalizing the sample segments to obtain a training sample set with known fault labels.
[0013] According to one aspect of the invention, the multi-source sensor signals include current, voltage, rotational speed, vibration, and temperature.
[0014] According to one aspect of the present invention, the deep neural network model sequentially includes: a one-dimensional convolutional layer and a multi-layer residual convolutional module for extracting local time series features; a channel attention mechanism for recalibrating the channel weights of the convolutional features; a self-attention mechanism for modeling the global dependency relationship of the features in the time dimension; and multi-scale pooling and cascaded layers for fusing multi-scale features and outputting discriminative embedding features.
[0015] According to one aspect of the present invention, the feature clustering loss is a center loss function, which is calculated as follows:
[0016]
[0017] in, For feature embedding of samples, Indicates the first The feature center of the class.
[0018] According to one aspect of the present invention, the open set regularization term is the open set loss, which is calculated as follows:
[0019]
[0020] in, Cross-entropy loss; For open sets, a penalty term is applied when the sample confidence level is below a threshold. Punishment should be carried out at the appropriate time; This is the weighting factor.
[0021] According to one aspect of the invention, the distance metric used to calculate the distance to the feature centers of all known fault categories is Euclidean distance, and the calculation formula is as follows:
[0022]
[0023] According to one aspect of the present invention, the clustering analysis of samples with unknown faults specifically employs the DBSCAN density clustering algorithm.
[0024] According to one aspect of the present invention, the dynamic expansion and updating of the model's fault knowledge base is achieved by re-optimizing the model parameters through an incremental learning mechanism.
[0025] A ship intelligent fault diagnosis system based on open label space recognition includes:
[0026] The data collection module acquires time-series signals from multiple sources of sensors in the ship's system and constructs a training sample set.
[0027] The learning module uses a deep neural network model to extract features from the training sample set to obtain the discriminative embedding features of the samples; during the training process, a center loss function is used to enhance the clustering characteristics of similar features in the embedding space.
[0028] The training module, based on the feature extraction, trains a classifier and introduces an open set loss function as part of the overall objective function for model training; the overall objective function is composed of the open set loss function and the center loss function.
[0029] The diagnostic module extracts the feature embedding of the sample to be diagnosed and calculates the distance between it and the feature centers of all known fault categories; if the minimum distance is greater than a preset threshold, it is determined to be an unknown fault.
[0030] The extension module outputs the corresponding fault category if the fault is determined to be a known fault; if the fault is determined to be an unknown fault, it triggers an alarm; and dynamically expands and updates the model's fault knowledge base based on the accumulation of unknown fault samples.
[0031] The advantages of this invention are as follows: Through the above technical solution, a ship intelligent fault diagnosis system capable of adapting to real and complex environments has been successfully constructed. First, by introducing an open set loss function and a distance-based threshold judgment mechanism, the limitations of the traditional closed set assumption are effectively overcome, enabling accurate perception and early warning of unknown faults not present in the training set, fundamentally improving the system's reliability and safety. Second, by utilizing the center loss function to enhance the intra-class clustering and inter-class separation of features, the ability to distinguish fault modes and the classification accuracy of known faults are significantly improved. Third, through a dynamic expansion mechanism combining DBSCAN clustering and incremental learning, the system possesses continuous evolution capabilities, adaptively expanding its cognitive boundaries and effectively reducing long-term maintenance costs. Furthermore, by adopting a hybrid network structure integrating 1D-CNN, channel attention, and self-attention mechanisms, robust fusion of multi-source information and interpretable decision analysis are achieved. In summary, this invention upgrades the fault diagnosis paradigm from a static "closed world" to a dynamic "open world," achieving a leap from passive identification to active perception and adaptive learning, significantly improving the operational efficiency and navigation safety assurance capabilities of ship systems. Attached Figure Description
[0032] 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.
[0033] Figure 1 This is a flowchart illustrating a ship intelligent fault diagnosis method based on open label space recognition as described in this invention. Detailed Implementation
[0034] 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.
[0035] Example 1
[0036] like Figure 1 As shown, a ship intelligent fault diagnosis method based on open label space recognition includes:
[0037] Acquire time-series signals from multiple sources of sensors in the ship's system and construct a training sample set;
[0038] The training sample set is used to extract features to obtain the discriminative embedding features of the samples; during the training process, a center loss function is used to enhance the clustering characteristics of similar features in the embedding space.
[0039] Based on the feature extraction, a classifier is trained and an open set loss function is introduced as part of the overall objective function for model training; the overall objective function is composed of the open set loss function and the center loss function.
[0040] For the sample to be diagnosed, its feature embedding is extracted, and its distance to the feature centers of all known fault categories is calculated; if the minimum distance is greater than a preset threshold, it is determined to be an unknown fault.
[0041] If the fault is determined to be a known fault, the corresponding fault category is output; if the fault is determined to be an unknown fault, an alarm is triggered; and the fault knowledge base of the model is dynamically expanded and updated based on the accumulation of the unknown fault samples.
[0042] Step 1: Data Acquisition and Preprocessing
[0043] Assume the ship's electric propulsion system is at time... The multi-source signal input is:
[0044]
[0045] in, Indicates the first Sensor signals (such as current, voltage, speed, vibration, temperature, etc.) This represents the number of sensors.
[0046] By using a sliding time window mechanism, the signal is segmented into segments of length [length missing]. Sample fragments:
[0047]
[0048] After normalization, the training sample set is obtained:
[0049]
[0050] in , Given a set of known labels.
[0051] Step 2: Feature Extraction and Representation Learning
[0052] First, the acquired multimodal input signals The input is fed into a one-dimensional convolutional layer, and local time-series features are extracted through multiple residual convolutional modules to obtain intermediate feature representations after normalization and nonlinear activation. .
[0053] Temporal Feature Extraction in CNNs (1D Convolution + Residual)
[0054] Stem convolution:
[0055]
[0056] in For the number of output channels, kernel length, Step length, filling, This represents the length of the sequence after convolution. (Followed by...) .
[0057] Residual block (the first) layer):
[0058]
[0059] and after the residual Typical settings: Each residual block, core length .
[0060] Subsequently, a channel attention mechanism is introduced at the channel dimension, which obtains the channel description vector through global average pooling of features. Feature recalibration is achieved through a two-layer perceptron. Perform global temporal convergence:
[0061]
[0062] Channel weights are obtained through two layers of perceptrons:
[0063]
[0064] in For ReLU, It is Sigmoid. Compression ratio .
[0065] The weighted output is:
[0066]
[0067] In multi-sensor / multi-physical quantity fusion, the contribution of important sensor channels is enhanced while redundant / noisy channels are suppressed. Perform a linear mapping to obtain the query / key / value (expanded by the time dimension):
[0068]
[0069] in ,get .
[0070] Relative position encoding (lightweight sine / cosine):
[0071]
[0072] It can be predefined (sine) or learned.
[0073] Single-head attention:
[0074]
[0075] Multi-head fusion ( head):
[0076]
[0077] Then, the residuals are normalized via a feedforward network (time-by-time MLP):
[0078]
[0079] in It is a two-layer MLP (containing ReLU and Dropout) that is time-by-time.
[0080] Perform multi-scale pooling and concatenate the outputs of several layers:
[0081]
[0082] in Mapping to discriminative embedding and normalization:
[0083]
[0084] Employing a center loss function to enhance clustering:
[0085]
[0086] in, Indicates the first The feature center of the class.
[0087] Step 3: Training the Open Label Space Recognition Model
[0088] Traditional classifiers assume that the label space is closed, i.e. .
[0089] This invention introduces the open label space assumption:
[0090]
[0091] in, It is a set of unknown categories.
[0092] The classifier outputs the confidence score for each class as follows:
[0093]
[0094] To enhance open set recognition capabilities, an open-set loss function is introduced:
[0095]
[0096] Where: the first term is the cross-entropy loss; the second term is the open set penalty term, which is applied when the sample confidence is below a threshold. Punishment should be carried out at the appropriate time; This is the weighting factor.
[0097] The overall training objective function is:
[0098]
[0099] Step 4: Unknown Fault Detection and Dynamic Tag Expansion
[0100] During the fault detection phase, for the sample to be diagnosed Calculate its feature embedding:
[0101]
[0102] Calculate its distance to all known class centers:
[0103]
[0104] If the minimum distance is greater than the threshold If it is, then it is judged as an unknown fault sample:
[0105]
[0106] For the cumulative set of unknown samples Grouping is performed using density clustering (DBSCAN):
[0107]
[0108] Each cluster Defined as a new category, the corresponding tag set is expanded to:
[0109]
[0110] The model parameters are then re-optimized using an incremental learning mechanism.
[0111] Step 5: Diagnostic Decision and Outcome Output
[0112] Final output: Fault category prediction results
[0113]
[0114] And output its confidence level:
[0115]
[0116] like The system triggers an unknown fault alarm, records relevant signal characteristics, and uploads them to the control center.
[0117] The advantages of this invention are as follows: By collecting multi-source sensor signals (such as current, voltage, speed, vibration, and temperature) from the ship's electric propulsion system and preprocessing them using a sliding time window mechanism, a high-quality training sample set is constructed. In the feature extraction stage, a hybrid neural network structure integrating one-dimensional convolution, residual modules, channel attention mechanisms, and self-attention mechanisms is adopted to effectively extract multi-scale spatiotemporal features, and the intra-class clustering and inter-class separation of features are enhanced through a central loss function. An open set loss function is introduced into the model training, combined with a distance threshold judgment mechanism, to achieve accurate detection and early warning of unknown faults. At the same time, the DBSCAN clustering algorithm is used to adaptively group the accumulated unknown fault samples, and incremental learning technology is used to dynamically expand the model knowledge base, enabling the system to have continuous evolution capabilities. Ultimately, this method not only improves the diagnostic accuracy of known faults but also breaks through the limitations of the traditional closed set assumption, realizing active perception, adaptive learning, and real-time alarm of unknown faults, significantly enhancing the reliability, safety, and long-term maintenance efficiency of the ship fault diagnosis system in complex environments.
[0118] 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 ship intelligent fault diagnosis method based on open label space recognition, characterized in that, include: Acquire time-series signals from multiple sources of sensors in the ship's system and construct a training sample set; The training sample set is used to extract features to obtain the discriminative embedding features of the samples; during the training process, a center loss function is used to enhance the clustering characteristics of similar features in the embedding space. Based on the feature extraction, a classifier is trained and an open set loss function is introduced as part of the overall objective function for model training; the overall objective function is composed of the open set loss function and the center loss function. For the sample to be diagnosed, its feature embedding is extracted, and its distance to the feature centers of all known fault categories is calculated; if the minimum distance is greater than a preset threshold, it is determined to be an unknown fault. If the fault is determined to be a known fault, the corresponding fault category is output; if the fault is determined to be an unknown fault, an alarm is triggered; and the fault knowledge base of the model is dynamically expanded and updated based on the accumulation of the unknown fault samples.
2. The intelligent ship fault diagnosis method based on open tag space recognition according to claim 1, characterized in that, The process of acquiring multi-source sensor time-series signals of the ship system and constructing training samples includes: Collect multi-source sensor signals from at least one system in the ship's electric propulsion system, engine room power equipment, energy conversion device, and auxiliary electromechanical system; The signal is segmented into sample segments of length L using a sliding time window mechanism; Normalize the sample fragments to obtain a training sample set with known fault labels.
3. The intelligent ship fault diagnosis method based on open tag space recognition according to claim 2, characterized in that, The multi-source sensor signals include current, voltage, rotational speed, vibration, and temperature.
4. The intelligent ship fault diagnosis method based on open tag space recognition according to claim 1, characterized in that, The deep neural network model includes, in sequence: One-dimensional convolutional layers and multi-layer residual convolutional modules are used to extract local time series features; Channel attention mechanism is used to recalibrate the channel weights of convolutional features; Self-attention mechanism is used to model global dependencies of features in the time dimension; Multi-scale pooling and cascaded layers are used to fuse multi-scale features and output discriminative embedding features.
5. The intelligent ship fault diagnosis method based on open tag space recognition according to claim 1, characterized in that, The feature clustering loss is the center loss function, which is calculated as follows: in, For feature embedding of samples, Indicates the first The feature center of the class.
6. The intelligent ship fault diagnosis method based on open tag space recognition according to claim 1, characterized in that, The open set regularization term is the open set loss, and its calculation method is as follows: in, Cross-entropy loss; For open sets, a penalty term is applied when the sample confidence level is below a threshold. Punishment should be carried out at the appropriate time; This is the weighting factor.
7. The intelligent ship fault diagnosis method based on open tag space recognition according to claim 1, characterized in that, The distance between the fault and the feature centers of all known fault categories is calculated using Euclidean distance, and the formula is as follows:
8. The intelligent ship fault diagnosis method based on open tag space recognition according to claim 1, characterized in that, The clustering analysis of samples with unknown faults specifically employs the DBSCAN density clustering algorithm.
9. The intelligent ship fault diagnosis method based on open tag space recognition according to claim 1, characterized in that, The dynamic expansion and updating of the model's fault knowledge base is achieved by re-optimizing the model parameters through an incremental learning mechanism.
10. A ship intelligent fault diagnosis system based on open label spatial recognition, characterized in that, The intelligent ship fault diagnosis method based on open label space recognition according to any one of claims 1 to 9 includes: The data collection module acquires time-series signals from multiple sources of sensors in the ship's system and constructs a training sample set. The learning module uses a deep neural network model to extract features from the training sample set to obtain the discriminative embedding features of the samples; during the training process, a center loss function is used to enhance the clustering characteristics of similar features in the embedding space. The training module, based on the feature extraction, trains a classifier and introduces an open set loss function as part of the overall objective function for model training; the overall objective function is composed of the open set loss function and the center loss function. The diagnostic module extracts the feature embedding of the sample to be diagnosed and calculates the distance between it and the feature centers of all known fault categories; if the minimum distance is greater than a preset threshold, it is determined to be an unknown fault. The extension module outputs the corresponding fault category if the fault is determined to be a known fault; if the fault is determined to be an unknown fault, it triggers an alarm; and dynamically expands and updates the model's fault knowledge base based on the accumulation of unknown fault samples.
Citation Information
Cited By
A method for identifying open faults of a marine equipment driving motor
CN122283438A
Inland river new energy ship operation and maintenance fault prediction analysis system based on big data analysis
CN122346658A
