A method and system for identifying a fault of a piston ring of a marine air compressor

Abstract

The application discloses a kind of ship air compressor piston ring fault identification method and system, it is related to ship fault identification field, it includes pre-acquisition piston ring at different types of operating state vibration signal;Wherein, different types of operating state include normal state and multiple types of fault state;The vibration signal is pre-processed to obtain pre-processing signal, and the pre-processing signal is converted into time-frequency image sample;Based on the time-frequency image sample, a Transformer model is constructed and trained to obtain a trained model;Collect the vibration signal of the piston ring to be identified when running, input the trained model after the same pre-processing and conversion, analyze the probability of each type of operating state of the piston ring to be identified, and output the probability distribution of all operating states;If the probability of any one type of fault state reaches the preset threshold, the piston ring fault is determined and the fault type is determined.The application has the effect of high identification accuracy.

Description

Technical Field

[0001] This application relates to the field of ship fault identification, and in particular to a method and system for identifying piston ring faults in ship air compressors. Background Technology

[0002] As a key component of a ship's propulsion system, the piston rings of a marine air compressor are subjected to high-frequency reciprocating motion for extended periods, making them susceptible to various types of failures. These failures can lead to decreased compression efficiency, increased engine vibration, and even system shutdowns. Existing fault identification methods primarily rely on vibration signal analysis, extracting only single-dimensional features such as those in the time or frequency domains. This approach fails to comprehensively capture the nonlinear characteristics of the faults; for example, the high-frequency impact caused by piston ring fracture or the low-frequency energy attenuation due to wear is often overlooked. When multiple faults occur concurrently, such as piston ring fracture accompanied by loose bolts, existing methods cannot effectively distinguish between fault types. Furthermore, the identification models are weak in recognizing early, minor faults and struggle to adapt to complex operating conditions such as load fluctuations and environmental disturbances during ship navigation. Although recent research has converted vibration signals into images and combined them with deep learning models for diagnosis, problems remain, including insufficient utilization of the global correlation of fault features and a lack of optimization for specific piston ring faults in marine air compressors. Therefore, there is an urgent need for an intelligent fault identification method that can effectively extract fault features, achieve high accuracy, and possess strong robustness to solve the challenge of identifying piston ring faults under complex operating conditions. Summary of the Invention

[0003] To improve the problems of low adaptability and poor identification accuracy of traditional identification methods, this application provides a method and system for identifying piston ring faults in marine air compressors.

[0004] Firstly, the method for identifying piston ring faults in a marine air compressor provided in this application adopts the following technical solution: S1, pre-collecting vibration signals of the piston ring when it is in different types of operating states; wherein, the different types of operating states include normal state and various types of fault states; S2. The vibration signal is preprocessed to obtain a preprocessed signal, and the preprocessed signal is converted into a time-frequency image sample; S3. Construct and train a Transformer model based on the time-frequency image samples to obtain a trained Transformer model; S4. Collect the vibration signal of the piston ring to be identified during operation, and input it into the trained Transformer model after the same preprocessing and transformation. Analyze the probability of the piston ring to be identified being in each type of operating state, and output the probability distribution of all operating states; where the sum of the probabilities of all operating states is 1. S5. If the probability of any type of fault state reaches a preset threshold, then the piston ring to be identified is determined and the fault type is determined.

[0005] By employing the aforementioned technical solution, vibration signals of piston rings under normal and various fault conditions are pre-collected, pre-processed, and converted to time-frequency images. A Transformer model is then constructed and trained, enabling automated intelligent identification of piston ring operating states. This solution automatically collects vibration signals from the piston rings to be identified, performs the same pre-processing and transformation, and inputs them into the model. It outputs the probability distribution of all operating states and accurately determines the fault type when the probability of any fault state reaches a preset threshold. This eliminates the reliance on traditional manual inspection, improves diagnostic accuracy and efficiency through deep learning models, covers multiple fault scenarios, and reduces the risk of misjudgment and missed detection. Simultaneously, the probabilistic output enhances the interpretability and reliability of the results, aiding in early fault warning, reducing maintenance costs and equipment downtime, thereby optimizing the piston ring fault identification process.

[0006] Optionally, the preprocessing of the vibration signal includes using the db4 wavelet basis function to perform wavelet packet denoising on the vibration signal to remove environmental noise and electromagnetic interference, while retaining the fault characteristic frequency band in the vibration signal.

[0007] By adopting the above technical solution, wavelet packet denoising using the db4 wavelet basis function can efficiently filter out environmental noise and electromagnetic interference in vibration signals, while accurately preserving key fault characteristic frequency band information, significantly improving the signal-to-noise ratio and purity of the vibration signal. This provides a high-quality input foundation for subsequent time-frequency image conversion, avoids feature distortion caused by noise, and enhances the stability and recognition accuracy of Transformer model training. This solution effectively suppresses the risk of misjudgment caused by external interference, improves the robustness of piston ring operating status diagnosis, and especially reduces the false negative and false positive rates of fault identification in complex industrial environments, thereby improving the efficiency of the entire fault identification process.

[0008] Optionally, converting the preprocessed signal into time-frequency image samples includes performing a continuous wavelet transform on the preprocessed signal using the Morlet wavelet function as a basis function; generating a time-frequency matrix and normalizing it; and adjusting the normalized time-frequency matrix into a time-frequency image sample of standard size.

[0009] By employing the above technical solution and using the Morlet wavelet function for continuous wavelet transform, the non-stationary characteristics of the preprocessed signal can be accurately captured, generating a high-resolution time-frequency matrix that comprehensively preserves the key fault features of the vibration signal in both time and frequency dimensions. Normalization of the time-frequency matrix effectively eliminates amplitude scale differences. Furthermore, the matrix is ​​adjusted to standard-sized time-frequency image samples, providing a unified and standardized input format for the Transformer model, significantly improving the convergence speed and generalization ability of model training. This solution significantly optimizes the visual representation of fault features, enabling the Transformer model to more sensitively identify subtle differences in fault states, reducing recognition errors caused by inconsistent input data, and thus improving the overall accuracy of piston ring operating state identification.

[0010] Optionally, the constructed Transformer model includes a feature extraction module, a location encoding module, a global feature encoding module, and a classification output layer. The feature extraction module uses a deformable convolutional network to adaptively capture fault feature frequency bands in the time-frequency image samples. The location encoding module uses a GRU recurrent neural network to preserve the temporal dependencies of the time-frequency image samples. The global feature encoding module uses a ViT encoder to capture global feature dependencies through a multi-head self-attention mechanism. The classification output layer uses a fully connected layer and a Softmax function to output the probability distribution of each type of operating state.

[0011] By adopting the above technical solutions, the feature extraction module uses a deformable convolutional network to adaptively capture fault feature frequency bands in time-frequency image samples, significantly improving the sensitivity and feature extraction accuracy for subtle fault states. The position encoding module utilizes a GRU recurrent neural network to effectively preserve the temporal dependencies of vibration signals, ensuring the integrity of time-series features and preventing the loss of key information. The global feature encoding module, based on a ViT encoder, captures the complex dependencies between global features through a multi-head self-attention mechanism, enhancing the model's ability to identify overall fault modes. The classification output layer combines a fully connected layer and a Softmax function to output the probability distribution of each type of operating state, enhancing the interpretability and reliability of the recognition results. This solution significantly improves the accuracy, robustness, and automation level of piston ring fault identification, reduces the risk of misjudgment and missed detection, and optimizes the model's adaptability to different industrial scenarios.

[0012] Optionally, training the Transformer model includes performing data augmentation and sample balancing on the time-frequency image samples to ensure that the number of time-frequency image samples in all types of running states is consistent; configuring training hyperparameters and inputting the time-frequency image samples into the Transformer model for training until the Transformer model converges and training is terminated to obtain the trained Transformer model.

[0013] By employing the above technical solutions, data augmentation and sample balancing are performed on time-frequency image samples to ensure a consistent number of samples across all operating conditions. This effectively eliminates model bias caused by imbalanced time-frequency image samples and significantly improves the ability to identify different types of fault states. Simultaneously, data augmentation increases the diversity of training data, enhancing the model's generalization ability to different operating conditions and preventing overfitting. By rationally configuring training hyperparameters, the model's learning efficiency is optimized, ensuring a stable and efficient training process. Training continues until model convergence, fully capturing the inherent patterns of the input training data and guaranteeing the robustness and accuracy of the final model. This solution significantly improves the performance of the Transformer model in piston ring operating state identification, reducing the risk of false positives and false negatives.

[0014] Optionally, configuring training hyperparameters includes setting the AdamW optimizer for model optimization; and using the cross-entropy loss function to calculate the error in predicting the probability distribution of each type of running state to guide model weight updates.

[0015] By adopting the above technical solution and setting the AdamW optimizer for model optimization, the adaptive learning rate adjustment and weight decay mechanism are effectively combined, significantly accelerating the model convergence process while suppressing the risk of overfitting and improving the model's generalization ability. Furthermore, the cross-entropy loss function is used to accurately calculate the error between the predicted probability distribution of each type of running state and the actual label, providing a clear optimization direction for model weight updates and ensuring efficient and stable training. This solution optimizes the training efficiency and final performance of the Transformer model, enhances the accuracy and robustness of piston ring running state diagnosis, and reduces the risk of misjudgment and missed detection.

[0016] Optionally, training the Transformer model further includes dividing the time-frequency image samples into a training set, a validation set, and a test set; using the training set as input for training the Transformer model, and simultaneously applying gradient clipping and early stopping strategies; after each training epoch, calculating the model accuracy using the validation set; and terminating training when the accuracy does not improve for N consecutive training epochs, where N is a preset positive integer whose value range can be determined according to the model training convergence speed; after training termination, evaluating the final performance index of the Transformer model using the test set to complete the training of the Transformer model.

[0017] By employing the above technical solution, the time-frequency image samples input to the model are divided into training, validation, and test sets, scientifically ensuring the generalization and accuracy of model training. Gradient clipping is applied synchronously during training to effectively prevent gradient explosion and ensure the stability of the training process. An early stopping strategy is combined with real-time monitoring of the validation set accuracy, automatically terminating training when the accuracy fails to improve for a preset number of consecutive times, significantly avoiding overfitting risks and optimizing computational resource consumption. After training termination, the final performance metrics are evaluated using an independent test set, providing reliable verification of the model's capabilities. This solution comprehensively optimizes the training efficiency of the Transformer model, improves the recognition accuracy and robustness for various piston ring operating states, reduces false negatives and misjudgments, and simultaneously lowers computational costs and training time.

[0018] Optionally, step S5 includes extracting the probability of each type of operating state from the probability distribution output by the trained Transformer model; comparing the probability value of each type of operating state with a preset threshold; if the probability value of the normal state reaches the preset threshold, then the piston ring is determined to be in a normal state; if the probability value of any fault type reaches the preset threshold, then the piston ring is determined to be in a fault state and identified as that fault type.

[0019] By employing the above technical solution, probability values ​​for each type of operating state are extracted from the probability distribution output by the trained Transformer model, and each value is compared with a preset threshold. This enables rapid and automated identification of the piston ring's operating state. When the probability of a normal state reaches the preset threshold, it is determined to be normal; when the probability of any fault type reaches the preset threshold, a fault is identified and the fault type is precisely located. This solution significantly improves the accuracy and response efficiency of fault identification, eliminates the bias of subjective human judgment, and ensures reliable identification of potential faults in the early stages. Simultaneously, the threshold comparison mechanism enhances sensitivity to high-probability faults, reducing the risk of misjudgment and missed detection.

[0020] Secondly, this application provides a system for identifying piston ring faults in a marine air compressor, comprising a data acquisition module, a signal processing module, a model training module, a fault analysis module, and a decision output module; The data acquisition module is used to pre-acquire vibration signals of piston rings when they are in different types of operating states, as well as to acquire vibration signals of the piston ring to be identified during operation. The signal processing module is used to preprocess the vibration signal, obtain the preprocessed signal, and convert the preprocessed signal into a time-frequency image sample; The model training module is used to build a Transformer model based on time-frequency image samples and train the Transformer model to obtain a trained Transformer model. The fault analysis module is used to input the time-frequency image into the trained Transformer model and output the probability distribution of each type of piston ring operating state; The decision output module is used to determine the fault and identify the fault type if the probability value of any fault type exceeds a preset threshold.

[0021] Understandably, the system for identifying piston ring faults in marine air compressors provided in the second aspect above is used to execute the method provided in this application. Therefore, the beneficial effects it can achieve can be referred to the beneficial effects in the corresponding method, and will not be repeated here.

[0022] In summary, this application includes at least one of the following beneficial technical effects: 1. By adopting the Transformer model construction method based on vibration signal preprocessing and time-frequency image conversion, combined with the full probability distribution output mechanism, the automated intelligent identification of piston ring operating status is realized. This effectively solves the problems of traditional manual inspection relying on experience and having a high rate of missed detection and misjudgment, thereby significantly improving the accuracy and efficiency of fault identification, covering multiple fault types and reducing maintenance costs.

[0023] 2. By adopting a hybrid Transformer model architecture that combines deformable convolutional networks, GRU temporal coding, and ViT global feature coding, the model adaptively captures fault feature frequency bands and preserves temporal dependencies, thus solving the technical difficulties of insufficient fault feature extraction and loss of temporal information in complex industrial environments. This enhances the model's sensitivity to different types of fault states and the robustness of the identification results, and optimizes the generalization ability under different working conditions.

[0024] 3. By adopting a sample equalization training strategy and a gradient pruning and early stopping strategy in synergistic optimization mechanism, combined with the hyperparameter configuration of the AdamW optimizer and cross-entropy loss function, the model bias and overfitting risk caused by imbalanced training data are solved. At the same time, the probability threshold decision mechanism enables accurate determination of fault type, thereby improving the reliability of the system in real-time monitoring scenarios and providing intelligent fault identification guarantee for the safe operation of ship air compressors. Attached Figure Description

[0025] To more clearly illustrate the technical solutions in the embodiments of this application, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0026] Figure 1 This is an overall flowchart of a method for identifying piston ring faults in a marine air compressor provided in an embodiment of this application; Figure 2 This is an overall structural diagram of a system for identifying piston ring faults in a marine air compressor, provided in an embodiment of this application. Detailed Implementation

[0027] This application discloses a method and system for identifying piston ring faults in marine air compressors, which are described below in conjunction with the appendix. Figure 1 - Appendix Figure 2 This application will be described in further detail.

[0028] Reference Figure 1 A method for identifying piston ring faults in a marine air compressor, comprising: S1, pre-collecting vibration signals of the piston ring under different operating conditions; wherein, the different operating conditions include normal conditions and various types of fault conditions.

[0029] In this context, the normal state represents the piston ring's normal operation without faults. The fault states in this embodiment mainly refer to three typical fault states: piston ring breakage, piston ring seizure, and excessive piston ring wear. The vibration signal is a mechanical dynamic response acquired through sensors, reflecting the piston ring's operating characteristics.

[0030] Specifically, a piezoelectric accelerometer was selected as the core sensing device for the data acquisition process. This piezoelectric accelerometer is suitable for capturing the high-frequency reciprocating motion characteristics of piston rings. In actual installation, the sensor was deployed vertically on the top of the cylinder head and horizontally on the left side of the ship's air compressor to simultaneously monitor vibration characteristics in both vertical and horizontal dimensions. This dual-point measurement method ensured comprehensive coverage of the piston ring's motion patterns. The vibration signal acquisition duration was set to 20 seconds to ensure signal coverage of at least 500 operating cycles, fully capturing the periodic characteristics of faults. The acquired signal states included four types: normal state signal, piston ring fracture state signal, piston ring seizure state signal, and piston ring excessive wear state signal. The normal state signal served as a baseline reference. The fracture state signal exhibited high-frequency impact characteristics, the seizure state signal showed a lack of characteristic frequencies, and the excessive wear state signal showed low-frequency amplitude attenuation. The data acquisition process was conducted in the actual operating environment of the ship's air compressor to simulate real-world conditions, such as high-frequency reciprocating motion and complex interference conditions, ensuring the representativeness and reliability of the acquired vibration signals. Ultimately, the collected vibration signals will be converted into digital format and stored to provide input for subsequent preprocessing and model training, supporting the entire fault identification process.

[0031] S2. The vibration signal is preprocessed to obtain a preprocessed signal, and the preprocessed signal is converted into a time-frequency image sample.

[0032] Preprocessing refers to the process of eliminating environmental noise and electromagnetic interference through signal denoising technology. The preprocessed signal is the signal that retains the key fault characteristic frequency bands after denoising. Time-frequency image samples are a two-dimensional visual representation. Time-frequency transformation technology maps one-dimensional vibration signals into an image format that can be parsed by deep learning models.

[0033] Specifically, the preprocessing stage employs a wavelet packet denoising algorithm based on the db4 wavelet basis function. The db4 wavelet basis function is a Daubechies wavelet with compact support and orthogonality. It reconstructs the vibration signal through a 5-level decomposition and uses a thresholding method to filter out noise components outside the 10-1000Hz range, ensuring the integrity of the fault characteristic frequency band. The core of the transformation stage is the Continuous Wavelet Transform (CWT), which uses the Morlet wavelet function as its basis function. The Morlet wavelet function is a complex exponential wavelet with a Gaussian envelope, suitable for capturing transient features. The scale parameter is set from 1 to 64 to cover the entire frequency band, and the time window length is synchronized with the air compressor's working cycle. The calculated wavelet coefficients form a time-frequency matrix, which is normalized to the [0,255] grayscale range and then adjusted to a standard time-frequency image sample of 224×224 pixels. During this process, different types of fault states exhibit different patterns in the time-frequency image samples. Piston ring fracture is characterized by an increase of more than 52% in the amplitude of the high-frequency impact zone, piston ring seizure shows a lack of characteristic frequency bands, and excessive wear exhibits low-frequency energy attenuation characteristics. The final generated time-frequency image samples combine temporal dynamics with frequency domain resolution, providing highly discriminative input data for the Transformer model.

[0034] S3. Construct and train a Transformer model based on the time-frequency image samples to obtain a trained Transformer model.

[0035] The constructed Transformer model includes a feature extraction module, a location encoding module, a global feature encoding module, and a classification output layer. The feature extraction module employs a deformable convolutional network to adaptively capture fault feature frequency bands in the time-frequency image samples. The location encoding module uses a GRU recurrent neural network to preserve the temporal dependencies of the time-frequency image samples. The global feature encoding module uses a ViT encoder to capture global feature dependencies through a multi-head self-attention mechanism. The classification output layer uses a fully connected layer and a Softmax function to output the probability distribution of each type of operating state.

[0036] Specifically, the feature extraction module of the constructed Transformer model employs a four-layer deformable convolutional network, with each layer configured with a 3×3 convolutional kernel and dilation factors of 1, 2, 4, and 8 respectively. It adaptively captures irregularly distributed local fault features in time-frequency image samples by dynamically learning the spatial deformation offset of the convolutional kernels, such as high-frequency impact stripes caused by piston ring fracture or frequency band breakage regions caused by piston ring failure. This overcomes the geometric constraints of traditional fixed convolutional kernels and significantly improves the adaptability to the deformation of fault areas, especially considering the non-stationary characteristics of vibration signals under complex ship operating conditions. The position encoding module uses a gated recurrent unit (GRU) network to convert the image block sequence after feature map segmentation of the time-frequency image samples extracted by the feature extraction module into a temporal vector. Utilizing the update and reset gate mechanisms of the GRU, the cyclic dependence of the vibration signal in the time dimension is preserved. This effectively retains the temporal context of fault feature occurrence, avoiding the shortcomings of standard Transformer position encoding in modeling the periodic characteristics of mechanical vibration. The global feature encoding module adopts the ViT encoder structure, which is a 12-layer stacked Transformer encoder with 8 attention heads deployed in each layer. It decomposes the features of the input time-frequency image samples into 14×14 sequences of 16×16 pixel image patches. A multi-head self-attention mechanism establishes global dependencies across these patches, such as associating high-energy spots from a broken piston ring with low-energy regions in the normal background. This allows the model to synthesize long-range feature context, addressing the feature association loss problem caused by the limited receptive field of traditional CNNs, and improving the model's feature decoupling ability for multiple concurrent faults, such as piston ring breakage accompanied by loose fuselage bolts. The classification output layer connects a fully connected layer with an input dimension of 768 and an output dimension of 4 to a Softmax function, mapping the encoded features to probability distributions of four operating states. Probability normalization ensures that the sum of the probabilities of all states is 1.

[0037] During model training, data augmentation and sample balancing were first implemented. Time-frequency image samples were horizontally flipped and randomly cropped, for example, to a size of 200×200, to simulate the installation angle deviation and vibration offset of piezoelectric accelerometers during ship navigation, thus enhancing the realism of the time-frequency image samples. Simultaneously, SMOTE oversampling technology was used to amplify time-frequency image samples of piston ring fracture, piston ring jamming, and piston excessive wear, ensuring that the number of time-frequency image samples for each fault state strictly matched the number of samples in the normal state. Furthermore, all operating states required ≥1000 time-frequency image samples to eliminate classifier bias caused by time-frequency image sample imbalance. Training hyperparameter settings included a fixed batch size of 32 and an initial learning rate of 0.0001; the optimization algorithm used was AdamW with fused weight decay, and the loss function was the cross-entropy loss function to measure probability distribution error. The gradient clipping constraint threshold is set to 0.1 to suppress training divergence, the maximum number of training epochs is set to 50, and an early stopping strategy is implemented, i.e., training is forcibly terminated when the accuracy on the validation set does not improve for 10 consecutive epochs to avoid overfitting. Furthermore, before training begins, all time-frequency image sample sets of all running types can be divided into training, validation, and test sets in a 7:2:1 ratio. The Transformer model constructed using the training set updates the deformable convolution kernel offset parameters and attention weights. The validation set monitors the generalization performance of the model training in real time and triggers early stopping to prevent overfitting. The test set independently evaluates the final model performance until convergence generates the trained Transformer model.

[0038] It should be noted that the four-layer architecture of this Transformer model solves the sensitivity problem of traditional CNNs to vibration and impact morphological changes through the local geometric adaptation characteristics of deformable convolution. The temporal dependencies preserved by the GRU position encoding can improve the capture efficiency of early minor wear features. The global feature association network constructed by the multi-head self-attention mechanism significantly improves the ability to identify coupled faults. Moreover, this Transformer model can be directly deployed on the edge computing equipment on ships to realize real-time online diagnosis and early warning of air compressor piston ring faults.

[0039] S4. Collect the vibration signal of the piston ring to be identified during operation, and input it into the trained Transformer model after the same preprocessing and transformation. Analyze the probability of the piston ring to be identified being in each type of operating state, and output the probability distribution of all operating states; where the sum of the probabilities of all operating states is 1.

[0040] Among them, the piston ring to be identified is obtained by real-time measurement and acquisition by piezoelectric acceleration sensors installed on the top of the cylinder head of the ship's air compressor in the vertical direction and on the left side of the fuselage in the horizontal direction; the same preprocessing and conversion is performed in step S2, firstly the vibration signal is preprocessed by wavelet packets, and then the time-frequency image sample is obtained after continuous wavelet transform conversion; the trained Transformer model is the Transformer model constructed and trained in step S3.

[0041] Specifically, piezoelectric accelerometers installed vertically on the top of the cylinder head and horizontally on the left side of the ship's air compressor are used to collect vibration signals of the piston rings to be identified during real-time operation. Then, the vibration signal is subjected to the same preprocessing operation as in step S2, that is, wavelet packet denoising based on the db4 wavelet basis function is combined with threshold filtering to remove environmental noise and electromagnetic interference, retaining only the key fault characteristic frequency band of 10-1000Hz, and generating an optimized preprocessed signal. Next, continuous wavelet transform is applied to the preprocessed signal for image conversion. The scale parameter range is set from 1 to 64 using the Morlet wavelet as the basis function, the wavelet coefficients are calculated to construct the time-frequency matrix, and the matrix is ​​normalized to the gray range of [0,255] and adjusted to a standard time-frequency image sample of 224×224 pixels. Then, the time-frequency image sample is input into the Transformer model trained in step S3. The trained Transformer model, with the help of its internal structure, uses a deformable convolutional network to adaptively extract local fault features, a GRU recurrent neural network to preserve temporal dependencies, and a ViT encoder to capture global feature dependencies through a multi-head self-attention mechanism. After processing by a fully connected layer and a Softmax function, the probability distribution of the piston ring to be identified as being in a normal state, a broken state, a bulging state, and an excessively worn state is obtained. The final output is a four-dimensional probability vector in which the sum of the probability values ​​of each type of operating state is 1.

[0042] It should be noted that the sum of the probabilities of all types of operating states being 1 is due to the mathematical properties and design purpose of the Softmax function used. In the classification output layer of the Transformer model, the Softmax function performs nonlinear normalization on the four-dimensional vector output by the fully connected layer, converts the feature score of each state into a probability value through exponential operation, and ensures that the sum of all probability values ​​is strictly equal to 1, thereby providing a standardized probability distribution output. This not only avoids inconsistencies in probability distribution but also enhances the interpretability of fault identification decisions. For example, when the probability of a certain type of operating state is ≥0.95, a diagnostic result can be directly output; otherwise, an early warning mechanism is triggered, improving the reliability and practicality of fault identification for ship air compressors.

[0043] S5. If the probability of any type of fault state reaches a preset threshold, then the piston ring to be identified is determined and the fault type is determined.

[0044] Specifically, the algorithm first reads the four-dimensional probability distribution vector output by the Transformer model and extracts the probability values ​​corresponding to the four operating states. Then, it compares the probability values ​​of the four operating states with a preset threshold of 0.95. The preset threshold is a pre-defined critical value for judging the operating state. If any fault state is detected, in this embodiment, the probability value of piston ring breakage, piston seizure, or excessive piston wear is greater than or equal to 0.95, the piston ring is determined to be faulty, and the fault type is output as the identification conclusion. If the probability of all types of fault states is less than 0.95, but the probability value of a certain type of fault state is between 0.85 and 0.95, a fault warning is triggered, and a re-inspection suggestion is output to avoid misjudgment. If the probability of normal state is greater than or equal to 0.95, it is determined to be a normal operating state, and the operating state of the piston ring is continuously monitored. The logic of the probability threshold comparison algorithm is to traverse the fault state components in the probability distribution. When any fault probability component is detected to exceed the threshold, the traversal is immediately terminated, and the fault type is locked and output. If no fault exceeds the threshold, the diagnostic confidence level is further evaluated, and a low confidence result triggers the warning mechanism.

[0045] It should be noted that this threshold setting is based on ROC curve analysis, ensuring a false alarm rate of less than 2% while maintaining a 95% fault recall rate. The lower limit of the threshold is also determined by considering the sensitivity of the ship's online monitoring scenario to the cost of unintended shutdowns. The fault early warning mechanism addresses situations where the probability of an early, minor fault is higher than the background noise but has not reached the preset threshold. It improves system reliability by activating a manual re-inspection process. The final output includes confirmation of normal status, specific fault type markings, or graded early warning signals, directly supporting maintenance decisions for the ship's air compressor.

[0046] Reference Figure 2 This is a schematic diagram of a system for identifying piston ring faults in a marine air compressor, as described in an embodiment of this application.

[0047] A ship inspection system based on federated learning includes a data acquisition module 1, a signal processing module 2, a model training module 3, a fault analysis module 4, and a decision output module 5.

[0048] The data acquisition module 1 is deployed in the field control cabinet and includes sensors 11 that are vertically installed on the top of the air compressor cylinder head and horizontally installed on the left side of the machine body. The sensors 11 are connected to the signal conditioning circuit 12 and the analog-to-digital converter 13 via shielded cables to convert the analog vibration signal into a digital signal and transmit it to the signal processing module 2.

[0049] The signal processing module 2 is deployed in the field control cabinet and includes a digital signal processor 21 and a storage unit 22. The digital signal processor 21 performs preprocessing by using a specific wavelet basis function to perform wavelet packet noise reduction, retains the fault characteristic frequency band and continuous wavelet transform image operation, and temporarily stores the normalized time-frequency image in the storage unit 22.

[0050] The model training module 3 can run independently on an offline server and includes a graphics processor 31, high-speed memory 32 and storage device 33. The graphics processor 31 reads the time-frequency image samples transmitted by the signal processing module 2, performs data augmentation operations, and constructs a hybrid model including deformable convolutional layers, gated recurrent units and Transformer encoders. After training by optimization algorithms, the model weights are stored in the storage device 33.

[0051] The fault analysis module 4 is integrated in the ship's central control room and includes a central processing unit 41 and a model loading unit 42. The central processing unit 41 calls the pre-trained model weights stored in the model training module 3, inputs the time-frequency image output by the signal processing module 2 into the model for feature extraction and global attention calculation, and outputs the fault probability distribution of each type of piston ring state.

[0052] The decision output module 5 is integrated in the ship's central control room and includes a programmable logic controller 51 and a human-machine interface 52. The human-machine interface 52 includes a display screen and alarm indicator lights. The programmable logic controller 51 receives the probability data output by the fault analysis module 4. When the probability of a single fault type exceeds a preset threshold, the corresponding alarm indicator light is triggered and the diagnosis conclusion is output on the display screen. Otherwise, a warning prompt and a re-inspection instruction are generated.

[0053] Unless otherwise defined, the technical or scientific terms used in this application shall have the ordinary meaning understood by one of ordinary skill in the art to which this application pertains. The terms "first," "second," "third," and similar terms used in this application specification and claims do not indicate any order, quantity, or importance, but are merely used to distinguish different components. The terms "an" or "a" and similar terms do not indicate a quantity limitation, but rather indicate the presence of at least one. The terms "comprising" or "including" and similar terms mean that the elements or objects preceding "comprising" or "including" encompass the elements or objects listed following "comprising" or "including" and their equivalents, and do not exclude other elements or objects. "Above," "below," "left," "right," etc., are used only to indicate relative positional relationships; when the absolute position of the described object changes, the relative positional relationship may also change accordingly.

[0054] The above are all preferred embodiments of this application, and are not intended to limit the scope of protection of this application. Therefore, all equivalent changes made in accordance with the structure, shape and principle of this application should be covered within the scope of protection of this application.

Claims

1. A method for identifying piston ring faults in a marine air compressor, characterized in that: This includes S1, which involves pre-collecting vibration signals of the piston rings under different operating conditions; these different operating conditions include normal conditions and various types of fault conditions. The data acquisition process uses a piezoelectric accelerometer, which is deployed vertically on the top of the cylinder head of the marine air compressor and horizontally on the left side of the fuselage to simultaneously monitor vibration characteristics in both vertical and horizontal dimensions and collect vibration signals of the piston rings at two points. The specific signal status types collected include normal status signals, piston ring fracture status signals, piston ring seizure status signals, and piston ring excessive wear status signals. S2. The vibration signal is preprocessed to obtain a preprocessed signal, and the preprocessed signal is converted into a time-frequency image sample; The preprocessing of the vibration signal includes using the db4 wavelet basis function to perform wavelet packet denoising on the vibration signal to remove environmental noise and electromagnetic interference, reconstructing the vibration signal through 5-level decomposition, and filtering out noise components outside the 10-1000Hz range using a threshold method while retaining the fault characteristic frequency band in the vibration signal. The step of converting the preprocessed signal into time-frequency image samples includes performing a continuous wavelet transform on the preprocessed signal using the Morlet wavelet function as the basis function; generating a time-frequency matrix and normalizing it; and adjusting the normalized time-frequency matrix into a time-frequency image sample of standard size. The calculated wavelet coefficients form a time-frequency matrix, which is normalized to the gray range of [0, 255] and then adjusted to a standard time-frequency image sample of 224×224 pixels. S3. Construct and train a Transformer model based on the time-frequency image samples to obtain a trained Transformer model; The constructed Transformer model includes a feature extraction module, a positional encoding module, a global feature encoding module, and a classification output layer. The feature extraction module employs a deformable convolutional network to adaptively capture fault feature frequency bands in the time-frequency image samples. It uses a four-layer deformable convolutional network, with each layer configured with a 3×3 convolutional kernel and dilation factors set to 1, 2, 4, and 8 respectively. The spatial deformation offset of the convolutional kernel is dynamically learned. The location encoding module uses a GRU recurrent neural network to convert the image block sequence after feature map segmentation of the time-frequency image samples extracted by the feature extraction module into a time-series vector. The update gate and reset gate mechanism of GRU are used to preserve the temporal dependency of the time-frequency image samples. The global feature encoding module uses the ViT encoder to capture global feature dependencies through a multi-head self-attention mechanism. The ViT encoder is a Transformer encoder with 12 stacked layers, with 8 attention heads deployed in each layer. The classification output layer uses a fully connected layer and a Softmax function to output the probability distribution of each type of running state. S4. Collect the vibration signal of the piston ring to be identified during operation, and input it into the trained Transformer model after the same preprocessing and transformation. Analyze the probability of the piston ring to be identified being in each type of operating state, and output the probability distribution of all operating states; where the sum of the probabilities of all operating states is 1. S5. If the probability of any type of fault state reaches the preset threshold, then the piston ring to be identified is determined and the fault type is determined. Step S5 includes extracting the probability of each type of operating state from the probability distribution output by the trained Transformer model; comparing the probability value of each type of operating state with a preset threshold; if the probability value of the normal state reaches the preset threshold, then the piston ring is determined to be in a normal state; if the probability value of any fault type reaches the preset threshold, then the piston ring is determined to be in a fault state and identified as that fault type.

2. The method for identifying piston ring faults in a marine air compressor according to claim 1, characterized in that: Training the Transformer model includes performing data augmentation and sample balancing on the time-frequency image samples to ensure that the number of time-frequency image samples in all types of running states is consistent; configuring training hyperparameters and inputting the time-frequency image samples into the Transformer model for training until the Transformer model converges and training is terminated to obtain the trained Transformer model.

3. The method for identifying piston ring faults in a marine air compressor according to claim 2, characterized in that: The configuration of training hyperparameters includes setting the AdamW optimizer for model optimization; and using the cross-entropy loss function to calculate the error in predicting the probability distribution of each type of running state to guide model weight updates.

4. The method for identifying piston ring faults in a marine air compressor according to claim 2, characterized in that: Training the Transformer model also includes dividing the time-frequency image samples into a training set, a validation set, and a test set; using the training set to input the Transformer model for training, and simultaneously applying gradient clipping and early stopping strategies; after each training epoch, calculating the model accuracy using the validation set; and terminating training when the accuracy does not improve for N consecutive training epochs, where N is a preset positive integer whose value range can be determined according to the model training convergence speed. After training is terminated, the final performance metrics of the Transformer model are evaluated using the test set to complete the training of the Transformer model.

5. A system for identifying piston ring faults in a marine air compressor, characterized in that: It includes a data acquisition module, a signal processing module, a model training module, a fault analysis module, and a decision output module; The data acquisition module is used to pre-acquire vibration signals of piston rings when they are in different types of operating states, as well as to acquire vibration signals of the piston ring to be identified during operation. The data acquisition process uses a piezoelectric accelerometer, which is deployed vertically on the top of the cylinder head of the marine air compressor and horizontally on the left side of the fuselage to simultaneously monitor vibration characteristics in both vertical and horizontal dimensions and collect vibration signals of the piston rings at two points. The specific signal status types collected include normal status signals, piston ring fracture status signals, piston ring seizure status signals, and piston ring excessive wear status signals. The signal processing module is used to preprocess the vibration signal, obtain the preprocessed signal, and convert the preprocessed signal into a time-frequency image sample; The preprocessing of the vibration signal includes using the db4 wavelet basis function to perform wavelet packet denoising on the vibration signal to remove environmental noise and electromagnetic interference, reconstructing the vibration signal through 5-level decomposition, and filtering out noise components outside the 10-1000Hz range using a threshold method while retaining the fault characteristic frequency band in the vibration signal. The step of converting the preprocessed signal into time-frequency image samples includes performing a continuous wavelet transform on the preprocessed signal using the Morlet wavelet function as the basis function; generating a time-frequency matrix and normalizing it; and adjusting the normalized time-frequency matrix into a time-frequency image sample of standard size. The calculated wavelet coefficients form a time-frequency matrix, which is normalized to the gray range of [0,255] and then adjusted to a standard time-frequency image sample of 224×224 pixels. The model training module is used to build a Transformer model based on time-frequency image samples and train the Transformer model to obtain a trained Transformer model. The constructed Transformer model includes a feature extraction module, a positional encoding module, a global feature encoding module, and a classification output layer. The feature extraction module employs a deformable convolutional network to adaptively capture fault feature frequency bands in the time-frequency image samples. It uses a four-layer deformable convolutional network, with each layer configured with a 3×3 convolutional kernel and dilation factors set to 1, 2, 4, and 8 respectively. The spatial deformation offset of the convolutional kernel is dynamically learned. The location encoding module uses a GRU recurrent neural network to convert the image block sequence after feature map segmentation of the time-frequency image samples extracted by the feature extraction module into a time-series vector. The update gate and reset gate mechanism of GRU are used to preserve the temporal dependency of the time-frequency image samples. The global feature encoding module uses the ViT encoder to capture global feature dependencies through a multi-head self-attention mechanism. The ViT encoder is a Transformer encoder with 12 stacked layers, with 8 attention heads deployed in each layer. The classification output layer uses a fully connected layer and a Softmax function to output the probability distribution of each type of running state. The fault analysis module is used to input the time-frequency image into the trained Transformer model and output the probability distribution of each type of piston ring operating state; The decision output module is used to determine the fault and identify the fault type if the probability value of any fault type exceeds a preset threshold. This includes extracting the probabilities of each type of operating state from the probability distribution output by the trained Transformer model; comparing the probability values ​​of each type of operating state with a preset threshold; if the probability value of the normal state reaches the preset threshold, then the piston ring is determined to be in a normal state; if the probability value of any fault type reaches the preset threshold, then the piston ring is determined to be in a fault state and identified as that fault type.

Images