A method and system for detecting anomalies in a marine main engine exhaust gas turbocharger

Abstract

The application discloses a kind of marine main engine exhaust turbocharger anomaly detection method and system, it is related to marine health detection field, it includes obtaining normal operation data when normal operation of exhaust turbocharger;Normal operation data is preprocessed, and the normal operation data after preprocessing is obtained;Reconstruction error threshold is obtained by constructing and training self-associative neural network based on the normal operation data after preprocessing;Real-time operation data of exhaust turbocharger is collected in real time, and the trained self-associative neural network is input after preprocessing real-time operation data, and real-time reconstruction error is obtained;The size relationship of reconstruction error and reconstruction error threshold is compared to detect whether abnormality appears;If reconstruction error is greater than reconstruction error threshold, then determine that exhaust turbocharger appears abnormality.The application has the effect of strong adaptability.

Description

Technical Field

[0001] This application relates to the field of ship health monitoring, and in particular to a method and system for detecting abnormalities in a ship's main engine exhaust turbocharger. Background Technology

[0002] The exhaust gas turbocharger of a marine main engine is a critical component of the marine propulsion system, and its operating status directly affects the performance and reliability of the main engine. Due to the complex and variable operating environment of ships, exhaust gas turbochargers are prone to various faults, such as rotor imbalance, blade damage, and gas seal leakage. If these faults are not detected and addressed in a timely manner, they may lead to decreased main engine performance, increased fuel consumption, and even serious safety accidents. Currently, fault detection methods for marine main engine exhaust gas turbochargers mainly include traditional methods based on sensor monitoring and intelligent detection methods based on machine learning. Traditional methods typically rely on data collected by one or more sensors, using fixed thresholds to determine whether the equipment is operating normally. However, this method has limited ability to detect complex fault modes and is difficult to adapt to changes in the marine operating environment. While machine learning-based intelligent detection methods can handle complex nonlinear relationships, they often require a large amount of labeled fault data for training, which is difficult to obtain in practice. Therefore, developing a method that can effectively detect anomalies in marine main engine exhaust gas turbochargers without requiring a large amount of labeled fault data is of significant practical importance. Summary of the Invention

[0003] To improve the poor adaptability of traditional methods, this application provides a method and system for detecting abnormalities in marine main engine exhaust turbochargers.

[0004] In the first aspect, the method for abnormal detection of a marine main engine exhaust gas turbocharger provided in this application adopts the following technical solution: S1, acquiring normal operation data of the exhaust gas turbocharger during normal operation; S2. Preprocess the normal operation data to obtain preprocessed normal operation data; S3. Construct and train an autoassociative neural network based on the preprocessed normal operation data to obtain the reconstruction error threshold; S4. Real-time operating data of the exhaust gas turbocharger is collected, and the pre-processed real-time operating data is input into the trained autoassociative neural network to obtain the real-time reconstruction error. S5. Compare the magnitude of the reconstruction error with the reconstruction error threshold to detect whether any abnormalities occur. S6. If the reconstruction error is greater than the reconstruction error threshold, the exhaust gas turbocharger is determined to be abnormal.

[0005] By adopting the above technical solution, normal operating data of the exhaust gas turbocharger is acquired and preprocessed, and then an autoassociative neural network is trained to determine the reconstruction error threshold. This simplifies the process of establishing the autoassociative neural network model for anomaly detection, eliminating the need for additional sensor calibration or manual intervention. Real-time operating data is collected, preprocessed, and then input into the trained autoassociative neural network to obtain the real-time reconstruction error, which is then compared with the reconstruction error threshold, enabling real-time monitoring and anomaly detection of the exhaust gas turbocharger. This solution, trained based on normal operating data, ensures the robustness and accuracy of the network model, improves detection efficiency, and avoids false alarms and missed alarms. Through automated real-time monitoring and threshold comparison, the risk of equipment failure and maintenance costs are reduced, thereby improving the operational safety and reliability of the exhaust gas turbocharger and extending its service life.

[0006] Optionally, S2 includes cleaning the normal operation data to remove erroneous or missing data; normalizing the cleaned normal operation data to generate preprocessed normal operation data.

[0007] By employing the aforementioned technical solution, normal operation data is cleaned to remove erroneous or missing data, ensuring the integrity and accuracy of the normal operation dataset and avoiding interference from noise or outliers in subsequent analysis. Subsequent normalization processing standardizes the normal operation data to a uniform scale, eliminating the influence of differences in feature dimensions and facilitating the training and convergence of the autoassociative neural network. This preprocessing process requires no additional sensor calibration or manual intervention, simplifying the data preparation process and improving processing efficiency. By improving data quality, the robustness and generalization ability of the trained model are enhanced, making the determination of the reconstruction error threshold more reliable. This directly optimizes the accuracy of anomaly detection and reduces the risk of false positives and false negatives.

[0008] Optionally, constructing and training an autoassociative neural network based on preprocessed normal operating data includes constructing an autoassociative neural network comprising an input layer, at least one hidden layer, and an output layer; inputting the preprocessed normal operating data from the input layer into the autoassociative neural network; adjusting the network weights through a backpropagation algorithm to enable the hidden layer to learn the feature representation of the input normal operating data; then minimizing the reconstruction error between the input normal operating data and the reconstructed data output by the output layer; repeating the iteration until the reconstruction error converges, wherein the reconstruction error is defined as the mean square error between the input normal operating data and the reconstructed data output by the output layer.

[0009] By adopting the above technical solution, an autoassociative neural network comprising an input layer, hidden layers, and an output layer is constructed. Preprocessed normal operation data is input into this network, and the backpropagation algorithm is used to adjust the network weights, enabling the hidden layers to effectively learn the feature representations of the normal operation data. Simultaneously, the reconstruction error between the input data and the reconstructed data is minimized, and the process is iterated until the error converges, thus achieving an efficient and robust model training process. This solution significantly improves the neural network's ability to capture normal patterns of ship equipment, enhances the accuracy of anomaly detection, and reduces the risk of false alarms. It also simplifies the training process, ensures rapid convergence of the network model under limited data conditions, improves the reliability and stability of the system, and ultimately provides a solid technical foundation for anomaly detection in exhaust gas turbochargers, significantly optimizing overall detection performance.

[0010] Optionally, training an associative neural network may also include dividing the preprocessed normal operating data into a training subset and a validation subset; continuously monitoring the reconstruction error of the validation subset during training; and stopping training and saving the network weights when the reconstruction error of the validation subset fails to improve after multiple iterations.

[0011] By employing the above technical solution, the preprocessed normal operating data is divided into training and validation subsets, achieving simultaneous optimization of network model training and generalization capabilities, thus avoiding overfitting risks. During training, the reconstruction error of the validation subset is continuously monitored, providing an objective evaluation basis for the network model's convergence state. Training automatically stops when the reconstruction error fails to improve after multiple consecutive calculations, effectively preventing wasted computational resources caused by ineffective iterations and significantly improving training efficiency. This solution dynamically controls the training cycle without manual intervention, ensuring the network model achieves optimal generalization performance. By balancing training intensity and model robustness, the accuracy of reconstruction error threshold discrimination is enhanced, thereby improving the reliability of anomaly detection. Simultaneously, hardware resource consumption and maintenance complexity are reduced, strengthening the real-time performance and long-term operational stability of the exhaust gas turbocharger monitoring system and reducing the risk of equipment failure and downtime.

[0012] Optionally, obtaining the reconstruction error threshold in S3 includes using preprocessed normal operating data to perform multiple autoassociative neural network reconstruction experiments and recording the reconstruction error of each reconstruction experiment; then statistically analyzing the distribution of the multiple reconstruction errors, and selecting a reconstruction error threshold based on this distribution, so that the reconstruction error values ​​of the normal operating data within a preset range are all lower than the reconstruction error threshold.

[0013] By adopting the above technical solution, the reconstruction error threshold is determined based on the distribution of reconstruction errors statistically analyzed from the reconstruction experiment results. This ensures the scientific rigor and objectivity of the reconstruction error threshold, eliminating the need for reliance on human experience or preset parameters. This reconstruction error threshold strictly covers the upper limit of reconstruction errors for all normally operating data, naturally forming a dynamic probability boundary for normal conditions and significantly reducing the risk of misjudgment. Furthermore, distribution modeling fully considers data fluctuation characteristics, enhancing the adaptability of the reconstruction error threshold to changes in normal equipment operating conditions and avoiding false triggering due to environmental fluctuations.

[0014] Optionally, S5 includes real-time acquisition of the ship's environmental state parameters and main engine operating condition parameters; generation of dynamic compensation factors based on a coupling relationship model of the environmental state parameters and operating condition parameters; wherein the coupling relationship model adopts an ensemble learning method based on gradient boosting; adjustment of the reconstruction error threshold according to the dynamic compensation factor; and comparison of the real-time reconstruction error with the adjusted reconstruction error threshold to detect anomalies.

[0015] By adopting the above technical solution, environmental state parameters and main engine operating condition parameters of the ship are collected in real time, and a coupling relationship model based on gradient boosting ensemble learning is constructed to dynamically generate compensation factors that accurately match changes in environmental conditions. Based on these compensation factors, the reconstruction error threshold is adaptively adjusted, effectively overcoming the misjudgment problem caused by interference from wind, waves, temperature, humidity, and sudden load changes in complex navigation environments caused by traditional fixed thresholds. This dynamic compensation mechanism significantly improves the anomaly detection system's ability to perceive the coupling effects of multiple variables, enhances the robustness of the model under extreme conditions, and ensures that the anomaly judgment results of the exhaust gas turbocharger are not affected by fluctuations in the navigation environment, greatly reducing the false alarm rate and the missed alarm rate, and providing highly reliable real-time early warning protection for the safe operation of the ship's power system.

[0016] Optionally, the generation of dynamic compensation factors based on the coupling relationship model of environmental state parameters and operating condition parameters includes constructing an environmental and operating condition coupling feature extractor and a multilayer perceptron neural network; inputting the environmental state parameters and operating condition parameters into the multilayer perceptron neural network, wherein the multilayer perceptron neural network includes an input layer, a feature fusion layer, and an output layer; calculating the interaction effect tensor between the environmental state parameters and the operating condition parameters through the feature fusion layer; inputting the interaction effect tensor into the dynamic compensation factor generation layer to generate initial values ​​for the dynamic compensation factors; mapping the initial values ​​of the dynamic compensation factors using a nonlinear activation function to output the final dynamic compensation factor, which is then used to adjust the reconstruction error threshold.

[0017] By adopting the above technical solution, a multilayer perceptron neural network architecture is constructed. The input layer synchronously receives ship environmental state parameters and main engine operating condition parameters. A feature fusion layer deeply analyzes the nonlinear interaction effects between these two types of parameters, generating a high-dimensional interaction effect tensor. This tensor is transformed into an initial compensation value by a dynamic compensation factor generation layer, and then outputs a final dynamic compensation factor with clear physical meaning through a nonlinear activation function mapping. This scheme achieves accurate quantitative modeling of the coupling effect between environmental disturbances and operating conditions, making the dynamic adjustment of the reconstruction error threshold mathematically interpretable. It significantly improves the system's robustness to complex disturbances such as sudden temperature and humidity changes, wave impacts, and load fluctuations, effectively suppressing false triggers caused by environmental noise. Simultaneously, by constraining the range of the compensation factor through the activation function, excessive deviation of the reconstruction error threshold is avoided, ensuring the reliability of anomaly detection results across the entire operating range of the ship's power system.

[0018] Secondly, this application provides a system for detecting abnormalities in a marine main engine exhaust turbocharger, including a data acquisition module, a preprocessing module, a training module, a real-time detection module, and an anomaly determination module. The data acquisition module is equipped with an exhaust gas turbocharger sensor network to acquire historical operating data of the exhaust gas turbocharger during normal operation, as well as real-time operating data of the exhaust gas turbocharger. The preprocessing module is connected to the data acquisition module and is used to perform data cleaning on historical and real-time running data, and to perform normalization processing on the cleaned data. The training module is connected to the preprocessing module and includes an autoassociative neural network architecture; it is used to build and train the autoassociative neural network and then generate a reconstruction error threshold based on the reconstruction error distribution of historical data. The real-time detection module connects the preprocessing module and the neural network training module. It is used to input the preprocessed real-time running data into the trained autoassociative neural network to generate real-time reconstruction error. The anomaly detection module connects the real-time detection module and the training module to compare the real-time reconstruction error with the reconstruction error threshold. When the real-time reconstruction error is greater than the reconstruction error threshold, the exhaust gas turbocharger is determined to be abnormal.

[0019] Understandably, the system for detecting abnormalities in the marine main engine exhaust turbocharger provided in the second aspect above is used to perform 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.

[0020] In summary, this application includes at least one of the following beneficial technical effects: 1. By using preprocessed normal operation data to train an autoassociative neural network and based on a reconstruction error comparison mechanism, the system can automatically learn the normal operation mode characteristics of the exhaust gas turbocharger, effectively solving the problem of false alarms and missed alarms caused by the reliance on manually set thresholds in traditional methods. This enables real-time intelligent identification of abnormalities in the ship's main engine exhaust gas turbocharger, significantly improving the accuracy of fault detection and response time.

[0021] 2. Because the threshold boundary is dynamically determined based on the reconstruction error distribution characteristics of historical normal operation data, ensuring that all normal data errors are below the threshold, the system can adapt to changes in equipment operating status and data fluctuations. This effectively solves the failure risk caused by fixed thresholds due to environmental drift or equipment aging, thereby ensuring the long-term robustness and generalization ability of the anomaly detection model and reducing operation and maintenance costs.

[0022] 3. Due to the adoption of a modular system architecture, which integrates the collaborative operation of data acquisition, preprocessing, training, real-time detection and anomaly detection modules, the system can handle the entire process from data acquisition to anomaly output from end to end. It effectively solves the problems of difficulty in integrating multi-source sensor data and real-time processing bottlenecks, thereby achieving lightweight deployment and millisecond-level response, significantly reducing ship operation and maintenance costs and extending the life of key equipment. Attached Figure Description

[0023] 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.

[0024] Figure 1 This is a flowchart illustrating the method for detecting abnormalities in a marine main engine exhaust turbocharger provided in Embodiment 1 of this application; Figure 2 This is a schematic diagram of the structure of the marine main engine exhaust turbocharger anomaly detection system provided in the embodiments of this application; Figure 3 This is another flowchart illustrating the method for detecting abnormalities in the exhaust gas turbocharger of a ship's main engine provided in Embodiment 2 of this application. Detailed Implementation

[0025] The following is in conjunction with the appendix Figure 1-3 This application will be described in further detail.

[0026] This application discloses a method and system for detecting abnormalities in a marine main engine exhaust turbocharger.

[0027] Reference Figure 1A method for detecting abnormalities in a marine main engine exhaust gas turbocharger includes S1, acquiring normal operation data of the exhaust gas turbocharger during normal operation.

[0028] The normal operating data refers to the operating data collected when the exhaust gas turbocharger is in normal operating condition, including main engine load, turbocharger speed, average cylinder burst pressure, main engine turbocharger back pressure, main engine scavenging pressure, main engine scavenging temperature, main engine turbocharger air inlet and outlet temperature difference, main engine turbocharger exhaust gas inlet and outlet temperature difference, and main engine turbocharger lubricating oil outlet temperature.

[0029] Specifically, the system continuously monitors and captures the real-time data stream of the aforementioned data through a sensor array pre-deployed at key locations on the ship's main engine. The sensor array includes pressure sensors, temperature sensors, and speed sensors, ensuring that raw, normal operating data is collected during the stable operation phase of the exhaust gas turbocharger, avoiding data deviations caused by human interference or abnormal operating conditions. The sensor output signals are converted into digital format and initially buffered to reduce transmission delays. Simultaneously, a timestamp recording function is integrated to track the temporal sequence of data, ensuring the integrity of the normal operating data for subsequent preprocessing stages.

[0030] S2. Preprocess the normal operation data to obtain preprocessed normal operation data.

[0031] Data cleaning refers to the system automatically identifying and removing erroneous or missing data points from normally operating datasets. Erroneous data points include outliers that exceed the reasonable range of physical data, such as temperature or pressure values ​​that do not conform to common sense in equipment operation. Normalization processing refers to converting the cleaned normally operating data into a uniform numerical range, eliminating the dimensional differences between different data, and ensuring that data such as main engine load, turbocharger speed, cylinder average burst pressure, main engine turbocharger back pressure, main engine scavenging pressure, main engine scavenging temperature, main engine turbocharger air inlet and outlet temperature difference, main engine turbocharger exhaust gas inlet and outlet temperature difference, and main engine turbocharger lubricating oil outlet temperature are compared and processed under the same scale.

[0032] Specifically, the data stream is first scanned sequentially to detect anomalies or missing items. Data points that exceed a reasonable physical range are identified and removed. Temporarily missing data is filled by interpolation between adjacent time points. Then, normalization is performed. Specifically, each data value is mapped to between zero and one using the maximum-minimum linear transformation formula to eliminate numerical scale bias caused by differences in data units. This ensures that all data participate in the training and computation of the autoassociative neural network with the same weight, and finally outputs uniform normal operating data, providing high-quality input for the subsequent training of the autoassociative neural network.

[0033] S3. Construct and train an autoassociative neural network based on the preprocessed normal operating data to obtain the reconstruction error threshold.

[0034] The autoassociative neural network consists of an input layer, hidden layers, and an output layer. The number of neurons in the input layer matches the dimension of the preprocessed normal operating data, and the number of neurons in the output layer is the same as that in the input layer. The reconstruction error threshold is a pre-determined critical value used to distinguish between normal and abnormal states. The reconstruction error is quantified by the mean square error function to determine the degree of deviation between the input vector corresponding to the normal operating data and the vector of the output reconstructed data; let the input vector be... The output vector of the reconstructed data is Where n is a positive integer, and all instances of n in the following text refer to this definition, without further elaboration. The formula for calculating the reconstruction error is: Other error measures, such as mean absolute error, can be used in specific implementation scenarios, but mean squared error is the optimal solution.

[0035] Specifically, the autoassociative neural network architecture is first initialized. The number of neurons in the input and output layers is set according to the dimensions of the preprocessed normal operating data, and the hidden layer structure is configured according to network design principles. The number of neurons in the input layer strictly matches the dimensions of the preprocessed normal operating data, and the number of neurons in the output layer synchronously replicates the size of the input layer to ensure dimensionality consistency. The hidden layer neurons compress the input features using a non-linear activation function. The dataset is then divided into a training subset and a validation subset based on temporal continuity. The former drives iterative updates of the network weights, while the latter monitors generalization ability. The training subset adjusts the network weights using backpropagation, forcing the hidden layers to learn the dense distribution pattern of the normal operating data in the feature space, simultaneously minimizing the reconstruction error between the input vector and the reconstructed output vector. The validation subset calculates the trend of the reconstruction error in each iteration in real time. When the reconstruction error fails to decrease after multiple iterations, training is immediately terminated and the network weights are fixed, achieving model convergence and avoiding overfitting. Then, a batch reconstruction experiment is performed using all the preprocessed normal operating data. After each normal operating data point is input into the trained autoassociative neural network, its reconstructed output vector is recorded, and the mean square error between this vector and the input vector is calculated as the reconstruction error value. After accumulating all reconstruction error values, a non-parametric density estimation algorithm is used to construct an error probability distribution model, which captures the central tendency and discrete characteristics of the error values. Based on this distribution model, the error distribution pattern is analyzed, and a reconstruction error threshold boundary is selected. This boundary must ensure that all reconstruction error values ​​within the preset coverage area of ​​normal operating data are strictly below this reconstruction error threshold; for example, 95% of the reconstruction errors can be preset to be below this threshold. This process relies on statistical decision rules to identify the smooth decay region of the distribution curve, avoiding misjudgment of outliers. Simultaneously, the reconstruction error threshold position is automatically adjusted based on preset coverage requirements to reliably encompass normal operating data. The finally generated reconstruction error threshold serves as the benchmark for subsequent detection, supporting the entire logical closed loop of real-time anomaly judgment.

[0036] It should be noted that the training termination condition in this embodiment is triggered based on the failure of the validation subset reconstruction error to show an improvement trend after multiple consecutive iterations. This judgment mechanism uses a preset fixed iteration number threshold as an objective judgment benchmark. For example, if the decrease in validation error does not exceed one ten-thousandth after five consecutive complete data iterations, the training process is forcibly terminated and the current optimal network weights are saved. This avoids the risk of overfitting due to overtraining and ensures the feasibility of patent implementation through objective quantitative standards.

[0037] S4. Real-time operating data of the exhaust gas turbocharger is collected, and after preprocessing the real-time operating data, it is input into the trained autoassociative neural network to obtain the real-time reconstruction error.

[0038] Real-time operational data refers to the multi-dimensional real-time data set continuously acquired by the main engine sensors during the ship's real-time operation. This includes main engine load, turbocharger speed, average cylinder combustion pressure, main engine turbocharger back pressure, main engine scavenging pressure, main engine scavenging temperature, main engine turbocharger air inlet and outlet temperature difference, main engine turbocharger exhaust gas inlet and outlet temperature difference, and main engine turbocharger lubricating oil outlet temperature. Preprocessing refers to performing data cleaning and normalization operations on the raw sensor data. Data cleaning removes erroneous and missing values ​​caused by sensor malfunctions or transmission interruptions. Normalization uses linear transformation to unify the data to the same dimensional range to eliminate differences in numerical ranges. The autoassociative neural network refers to a network model trained using historical normal operation data. Real-time reconstruction error refers to the mean square error calculated between the reconstructed data output from the trained autoassociative neural network and the original input real-time operational data.

[0039] Specifically, the system continuously receives real-time operational data streams from sensors. First, it identifies and removes data points exceeding reasonable physical limits. Temporarily missing data is filled using interpolation between adjacent time points. Then, a minimum-maximum normalization algorithm is used to unify the numerical range of each data point. The processed data vector is immediately input into a trained autoassociative neural network. This network compresses and reconstructs the input features through hidden layers, and the output layer generates reconstructed data. The system calculates the mean squared difference of each dimension between the output vector of the reconstructed data and the input vector of the real-time operational data. The resulting real-time mean squared error is the real-time reconstruction error, which is then passed to the subsequent anomaly detection module for comparison with reconstruction error thresholds.

[0040] S5. Compare the reconstruction error with the reconstruction error threshold to detect any anomalies.

[0041] S6. If the reconstruction error is greater than the reconstruction error threshold, the exhaust gas turbocharger is determined to be abnormal.

[0042] Specifically, the system first directly compares the obtained real-time reconstruction error value with the reconstruction error threshold. Based on strict logical judgment rules, if the reconstruction error is greater than the threshold, the exhaust gas turbocharger is determined to be in an abnormal state, and an alarm mechanism is triggered to send an alarm command to the ship's monitoring system. If the reconstruction error is less than or equal to the threshold, the exhaust gas turbocharger is determined to be operating normally. This process relies on the reconstruction capability of an associative neural network. After learning the characteristics of normal data, the neural network shows a significant increase in reconstruction error under abnormal conditions, thus enabling efficient identification of anomalies through simple numerical comparison, avoiding reliance on complex models. Simultaneously, the system uses optimization algorithms during the comparison phase to ensure the real-time performance and accuracy of the calculations. For example, it employs a memory-efficient data processing mechanism to reduce latency and uses a threshold comparison module to achieve rapid decision-making, supporting immediate responses to ensure ship operational safety.

[0043] Reference Figure 2 This is a schematic diagram of a system structure for detecting abnormalities in a marine main engine exhaust turbocharger, as described in an embodiment of this application.

[0044] A system for detecting anomalies in a marine main engine exhaust turbocharger includes a data acquisition module 1, a preprocessing module 2, a training module 3, a real-time detection module 4, and an anomaly determination module 5.

[0045] The data acquisition module 1 is deployed in the ship's engine room and includes a network of various types of industrial sensors 11, such as temperature sensors, pressure sensors, and speed sensors. It is directly connected to key measuring points of the exhaust gas turbocharger via physical cables or industrial bus protocols to collect nine types of operating parameters in real time, including main engine load and turbocharger speed. The historical and real-time operating data are then transmitted to the preprocessing module 2.

[0046] The preprocessing module 2 is equipped with a data cleaning unit 21 and a normalization processing unit 22. It receives raw data through an Ethernet interface, performs error data removal, missing data imputation and parameter normalization operations, and the processed data is transmitted to the training module 3 via an internal bus.

[0047] Training module 3 acquires preprocessed historical data through memory channel to complete network training, and generates reconstruction error threshold based on reconstruction error statistical analysis. This reconstruction error threshold is transmitted to anomaly detection module 5 through shared storage area 31.

[0048] The real-time detection module 4 receives the real-time running data preprocessed by the preprocessing module 2 through a high-speed data interface, calls the trained autoassociative neural network model to perform real-time reconstruction data calculation, generates real-time reconstruction error values, and pushes them to the anomaly judgment module 5 via the internal communication bus.

[0049] The anomaly detection module 5 is integrated into the ship's central monitoring system. It is equipped with a threshold comparison unit 51 and an alarm triggering unit 52. It receives the reconstruction error value and compares it with the reconstruction error threshold in real time. When the reconstruction error value of multiple consecutive detection cycles exceeds the threshold, it sends an abnormal command to the ship's alarm device via industrial Ethernet to trigger an audible and visual alarm.

[0050] Reference Figure 3 The following is a description of another embodiment of the method and system provided in this implementation.

[0051] Based on the method for detecting abnormalities in the exhaust gas turbocharger of a ship's main engine described in Embodiment 1, this Embodiment 2 adds some specific implementation methods.

[0052] In this embodiment, step S205 includes: real-time acquisition of the ship's environmental state parameters and main engine operating condition parameters; generation of a dynamic compensation factor based on a coupling relationship model of the environmental state parameters and operating condition parameters; wherein the coupling relationship model adopts an ensemble learning method based on gradient boosting; adjustment of the reconstruction error threshold according to the dynamic compensation factor; and comparison of the real-time reconstruction error with the adjusted reconstruction error threshold to detect anomalies.

[0053] The process of generating a dynamic compensation factor based on a coupling relationship model of environmental state parameters and operating condition parameters includes constructing an environmental and operating condition coupling feature extractor and a multilayer perceptron neural network. The environmental state parameters and operating condition parameters are input into the multilayer perceptron neural network, which includes an input layer, a feature fusion layer, and an output layer. The interaction effect tensor between the environmental state parameters and operating condition parameters is calculated through the feature fusion layer. The interaction effect tensor is input into the dynamic compensation factor generation layer to generate initial values ​​for the dynamic compensation factor. The initial values ​​of the dynamic compensation factor are mapped using a nonlinear activation function to output the final dynamic compensation factor, which is then used to adjust the reconstruction error threshold.

[0054] Environmental state parameters refer to real-time acquired external environmental factors of the ship, including wind direction and speed, wave height, and temperature and humidity, used to characterize the impact of the navigation environment on the main engine equipment. Operating condition parameters refer to synchronously acquired internal operating conditions of the main engine, including main engine load values, turbocharger speed fluctuations, and cylinder pressure changes, reflecting the current operating status of the exhaust gas turbocharger. The dynamic compensation factor is a real-time modulation factor generated through a coupling relationship model, used to adaptively adjust the reconstruction error threshold to adapt to environmental changes. The reconstruction error threshold is a mean square error threshold determined by the self-associative neural network after learning from historical normal operating data, used to distinguish abnormal equipment states. The coupling relationship model employs a machine learning framework built using an ensemble learning method based on gradient boosting, aiming to capture the interaction between environmental and operating condition parameters. The multilayer perceptron neural network is the core structure of this model, containing an input layer, a feature fusion layer, and an output layer, responsible for processing the input signal; the feature fusion layer is an intermediate layer of the multilayer perceptron neural network, used to calculate the nonlinear correlation mapping between environmental state parameters and operating condition parameters. The interaction effect tensor is the output data structure of the feature fusion layer, which quantifies the multidimensional representation of the coupling effect between the environment and operating parameters; the initial value of the dynamic compensation factor is an intermediate calculation result, which is then processed to generate the final modulation factor.

[0055] Specifically, environmental state parameters and operating condition parameters are collected in real time through a shipborne sensor network. These parameters are then input into a coupling relationship model, which integrates a multilayer perceptron neural network architecture. This model includes an input layer that receives dual-channel feature vectors of environmental state parameters and operating condition parameters, and a feature fusion layer that calculates an interaction effect tensor using activation function operations and a weighted fusion matrix of environmental features and operating condition signals to characterize the dynamic coupling relationship. The interaction effect tensor is then passed to a dynamic compensation factor generation layer, where a nonlinear activation function maps the tensor to an initial value for the dynamic compensation factor and performs normalization to generate the final dynamic compensation factor. The anomaly detection module then uses this final dynamic compensation factor to scale the initial reconstruction error threshold, forming a dynamically adjusted reconstruction error threshold to eliminate detection bias caused by environmental disturbances. Simultaneously, the real-time detection module sends pre-processed exhaust gas turbocharger operating data to a trained autoassociative neural network to calculate the real-time reconstruction error and performs a direct comparison between the reconstruction error value and the dynamically adjusted threshold. If the reconstruction error exceeds the adjusted threshold, the system immediately triggers an anomaly alarm signal to notify maintenance personnel; otherwise, it maintains normal equipment operation. This technology enables the model to adaptively adjust its detection sensitivity in a variable marine environment. By fully learning the interaction between the environment and operating conditions through the feature fusion layer, it avoids the risk of false alarms caused by fixed thresholds, and significantly improves the robustness and reliability of exhaust gas turbocharger anomaly detection in harsh navigation scenarios.

[0056] It should be noted that this embodiment uses the Sigmoid function as the nonlinear activation function because its output is limited to a continuous interval between zero and one. This facilitates subsequent linear transformation to control the dynamic compensation factor within the range of 0.8 to 1.2, ensuring its adaptability to fluctuations in the ship's environment without threshold failure. The specific transformation formula is that the final dynamic compensation factor equals the scaling factor multiplied by the Sigmoid output value plus the offset. The scaling factor and offset were determined through repeated experiments to cover typical operating conditions. The threshold adjustment operation is achieved through a multiplication formula: the adjusted reconstruction error threshold equals the final dynamic compensation factor multiplied by the initial reconstruction error threshold. This suppresses false alarms under adverse sea conditions and maintains detection reliability.

[0057] 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.

[0058] 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 detecting abnormalities in a marine main engine exhaust gas turbocharger, characterized in that, include: S1. Obtain normal operating data of the exhaust gas turbocharger during normal operation; S2. Preprocess the normal operation data to obtain preprocessed normal operation data; S3. Construct and train an autoassociative neural network based on the preprocessed normal operation data to obtain the reconstruction error threshold; S4. Real-time operating data of the exhaust gas turbocharger is collected, and the pre-processed real-time operating data is input into the trained autoassociative neural network to obtain the real-time reconstruction error. S5. Compare the magnitude of the reconstruction error with the reconstruction error threshold to detect whether any abnormalities occur. S6. If the reconstruction error is greater than the reconstruction error threshold, the exhaust gas turbocharger is determined to be abnormal.

2. The method for detecting abnormalities in a marine main engine exhaust turbocharger according to claim 1, characterized in that: S2 includes cleaning the normal operation data, removing erroneous or missing data; then normalizing the cleaned normal operation data to generate preprocessed normal operation data.

3. The method for detecting abnormalities in a marine main engine exhaust turbocharger according to claim 1, characterized in that: Constructing and training an autoassociative neural network based on preprocessed normal operating data includes constructing an autoassociative neural network comprising an input layer, at least one hidden layer, and an output layer; inputting the preprocessed normal operating data from the input layer into the autoassociative neural network; adjusting the network weights through a backpropagation algorithm to enable the hidden layer to learn the feature representation of the input normal operating data; then minimizing the reconstruction error between the input normal operating data and the reconstructed data output by the output layer; repeating the iteration until the reconstruction error converges, wherein the reconstruction error is defined as the mean square error between the input normal operating data and the reconstructed data output by the output layer.

4. The method for detecting abnormalities in a marine main engine exhaust turbocharger according to claim 3, characterized in that: Training an associative neural network also includes dividing the preprocessed normal operating data into a training subset and a validation subset; continuously monitoring the reconstruction error of the validation subset during training; and stopping training and saving the network weights when the reconstruction error of the validation subset fails to improve after multiple iterations.

5. The method for detecting abnormalities in a marine main engine exhaust turbocharger according to claim 4, characterized in that: The reconstruction error threshold obtained in S3 includes using preprocessed normal operating data to conduct multiple autoassociative neural network reconstruction experiments and recording the reconstruction error of each reconstruction experiment; then, the distribution of the multiple reconstruction errors is statistically analyzed, and a reconstruction error threshold is selected based on this distribution, so that the reconstruction error value of the normal operating data within a preset range is lower than the reconstruction error threshold.

6. The method for detecting abnormalities in a marine main engine exhaust turbocharger according to claim 1, characterized in that: S5 includes real-time acquisition of the ship's environmental state parameters and main engine operating condition parameters; generation of dynamic compensation factors based on the coupling relationship model of environmental state parameters and operating condition parameters; wherein the coupling relationship model adopts an ensemble learning method based on gradient boosting; adjustment of the reconstruction error threshold according to the dynamic compensation factor; and comparison of the real-time reconstruction error with the adjusted reconstruction error threshold to detect anomalies.

7. The method for detecting abnormalities in a marine main engine exhaust turbocharger according to claim 6, characterized in that: The dynamic compensation factor is generated by the coupling relationship model based on environmental state parameters and operating condition parameters. This includes constructing an environmental and operating condition coupling feature extractor and a multilayer perceptron neural network. The environmental state parameters and operating condition parameters are input into the multilayer perceptron neural network, which includes an input layer, a feature fusion layer and an output layer. The interaction effect tensor between environmental state parameters and operating condition parameters is calculated through the feature fusion layer; the interaction effect tensor is then input into the dynamic compensation factor generation layer to generate the initial value of the dynamic compensation factor. The initial value of the dynamic compensation factor is mapped using a nonlinear activation function, and the final dynamic compensation factor is output and used to adjust the reconstruction error threshold.

8. A system for detecting abnormalities in a marine main engine exhaust turbocharger, characterized in that: It includes a data acquisition module, a preprocessing module, a training module, a real-time detection module, and an anomaly detection module; The data acquisition module is equipped with an exhaust gas turbocharger sensor network to acquire normal operation data of the exhaust gas turbocharger during normal operation, as well as real-time operation data of the exhaust gas turbocharger. The preprocessing module is connected to the data acquisition module and is used to perform data cleaning on normal operation data and real-time operation data, and to perform normalization processing on the cleaned data; The training module is connected to the preprocessing module and includes an autoassociative neural network architecture; Used to generate a reconstruction error threshold based on the reconstruction error distribution of normal operating data after constructing and training an associative neural network; The real-time detection module connects the preprocessing module and the neural network training module. It is used to input the preprocessed real-time running data into the trained autoassociative neural network to generate real-time reconstruction error. The anomaly detection module connects the real-time detection module and the training module to compare the real-time reconstruction error with the reconstruction error threshold. When the real-time reconstruction error is greater than the reconstruction error threshold, the exhaust gas turbocharger is determined to be abnormal.

Images