A method and system for marine common rail fuel injector health assessment

Abstract

This application discloses a method and system for health assessment of a common rail injector in a ship, relating to the field of ship equipment health monitoring. The method includes: acquiring the normal pressure signal of the common rail when the injector is operating normally; preprocessing the normal pressure signal to extract normal time-domain features, and normalizing these features; constructing and training a self-organizing map neural network using the normalized normal time-domain features to obtain the optimal matching unit and a reference distance threshold; wherein the optimal matching unit is the node in the output layer of the self-organizing map neural network with the smallest Euclidean distance to the input normal time-domain features; processing the real-time pressure signal acquired in real-time according to step S2 to obtain real-time time-domain features, calculating the Euclidean distance between the real-time time-domain features and the optimal matching unit, and comparing its magnitude with the reference distance threshold; when the reference distance threshold is larger, the injector is determined to be in a healthy state. This application has a high real-time performance.

Description

Technical Field

[0001] This application relates to the field of ship equipment health monitoring, and in particular to a method and system for health assessment of ship common rail injectors. Background Technology

[0002] As a key component of a ship's propulsion system, the common rail injector of a marine main engine directly affects the combustion efficiency, power output, and emissions performance of the main engine. However, due to the long-term operation of injectors in harsh environments such as high temperatures, they are prone to wear, blockage, and other malfunctions, which in turn affect the safe and stable operation of the ship. Currently, the health assessment of marine main engine common rail injectors mainly relies on manual experience and periodic maintenance. This method is not only inefficient but also fails to accurately assess the health status of the injectors in real time, making it difficult to detect potential faults in a timely manner. This may lead to sudden malfunctions during navigation, causing serious economic losses or even safety accidents. Therefore, developing a method and system that can accurately assess the health status of marine main engine common rail injectors in real time is of significant practical importance. Summary of the Invention

[0003] To improve the problems of low efficiency and poor real-time performance of traditional assessment methods, this application provides a method and system for health assessment of common rail injectors for ships.

[0004] In the first aspect, the method for health assessment of a common rail injector provided in this application adopts the following technical solution: S1, obtaining the normal pressure signal of the common rail pipe when the injector is working normally; S2. After preprocessing the normal pressure signal, extract the normal time domain features and normalize the normal time domain features. S3. Construct and train a self-organizing map neural network using the normalized normal time-domain features to obtain the best matching unit and the reference distance threshold. Among them, the best matching unit is the node in the output layer of the self-organizing map neural network that has the smallest Euclidean distance to the normal temporal features of the input; S4. Process the real-time pressure signal acquired in real time according to step S2 and obtain the real-time time domain features. Calculate the Euclidean distance between the real-time time domain features and the best matching unit, and compare its magnitude with the reference distance threshold. S5. When the Euclidean distance between the real-time temporal feature and the best matching unit is less than the reference distance threshold, the injector is determined to be in a healthy state.

[0005] By employing the above technical solution, the common rail pressure signal during normal injector operation is obtained as a benchmark. After preprocessing, normal time-domain features are extracted and normalized. These features are then used to train a self-organizing map neural network to obtain the optimal matching unit and determine a reference distance threshold. The acquired pressure signal is processed in real-time to obtain real-time time-domain features. The Euclidean distance between these features and the optimal matching unit in the neural network output layer is calculated and compared with the reference distance threshold. When this Euclidean distance is less than the reference distance threshold, the injector's health status is automatically determined. This solution utilizes feature extraction from pressure signals and a neural network model to achieve real-time online monitoring of injector faults. Through normalization processing and feature training, the robustness and accuracy of the model are enhanced. It eliminates the need for manual intervention, simplifies the diagnostic process, improves monitoring efficiency and reliability, and reduces hardware dependence and maintenance costs.

[0006] Optionally, S5 also includes determining that the injector is currently in a fault state when the Euclidean distance between the real-time time-domain feature and the best matching unit is greater than a reference distance threshold.

[0007] By adopting the above technical solution, a fault status judgment function is added to the existing health status judgment. When the Euclidean distance between the real-time time domain characteristics and the optimal matching unit is greater than a reference distance threshold, the injector is automatically determined to be in a fault state, thus forming a closed-loop fault monitoring mechanism. This solution does not rely on additional hardware or manual intervention. By comparing distance thresholds, it can quickly capture abnormal characteristics and promptly detect potential problems with the injector.

[0008] Optionally, S2 specifically includes cleaning the normal pressure signal to remove outliers and noise; extracting normal time-domain features from the cleaned normal pressure signal; and normalizing the normal time-domain features to a predetermined interval.

[0009] By employing the above technical solution, normal pressure signals are cleaned to remove outliers and noise, and normal time-domain features are extracted from the cleaned signals. These features are then normalized to a predetermined range, significantly improving the accuracy and reliability of feature extraction. Data cleaning ensures the purity of the input signal, reducing noise interference and the impact of outliers, providing a high-quality foundation for subsequent analysis. Time-domain feature extraction focuses on key time-domain information, enhancing the network model's ability to represent the injector's health status. Normalization standardizes the range of time-domain feature values, facilitating efficient learning and convergence of the neural network. This solution simplifies the data preprocessing process, reduces the need for manual correction, and improves the robustness and accuracy of fault diagnosis.

[0010] Optionally, S3 specifically includes constructing a self-organizing map neural network; inputting normalized normal temporal features into the self-organizing map neural network; calculating the Euclidean distance between the normalized normal temporal features and the weights of each network output layer node, and obtaining the output layer node with the smallest Euclidean distance as the best matching unit; updating the weights according to the learning rate based on the positions of the best matching unit and its neighboring nodes; repeating the iteration until the self-organizing map neural network converges, and automatically generating a reference distance threshold based on the kernel density probability model.

[0011] By adopting the above technical solution, a self-organizing mapping neural network is constructed, facilitating spatial mapping and neighborhood operations. The network is trained using normalized features, Euclidean distance is calculated to identify the best matching unit, and neighborhood weights are updated based on its position. Finally, a kernel density probability model is applied to automatically generate a reference distance threshold, achieving objectivity and adaptive determination of the reference distance threshold. This solution improves the accuracy and robustness of fault diagnosis. The kernel density model replaces manual threshold setting, reducing subjective bias and debugging costs. The weight update mechanism enhances the self-organizing mapping neural network's ability to represent normal patterns. Overall, through fully automated training and reference distance threshold generation, the deployment process of the self-organizing mapping neural network is simplified, maintenance requirements are reduced, and the reliability and real-time performance of injector health monitoring are improved.

[0012] Optionally, updating the weights according to the learning rate based on the positions of the best matching unit and its neighboring nodes includes calculating the topological distance weights of the neighboring nodes based on the position coordinates of the best matching unit in the output layer; generating dynamic decay update coefficients by combining the learning rate and the topological distance weights; and adjusting the weight vector of the best matching unit and its neighboring nodes using the dynamic decay update coefficients.

[0013] By employing the above technical solution, topological distance weights are calculated based on the position coordinates of the best matching unit. Dynamic decay update coefficients are generated by combining the learning rate, and these coefficients are used to adjust the weight vectors of the best matching unit and its neighboring nodes, achieving adaptive optimization of neural network weights. This solution utilizes topological distance weights to accurately quantify the spatial correlation of neighboring nodes, ensuring that neighboring nodes receive higher updated weights and strengthening the ability to preserve the local topological structure of the feature mapping. The dynamic decay update coefficients, by fusing the learning rate and spatial weights, intelligently regulate the weight iteration amplitude, effectively balancing the convergence speed and stability of the neural network. Through fine-tuning of the weight vectors, the spatial representation accuracy and generalization ability of the neural network for the characteristics of normal operating conditions of the fuel injector are significantly improved. This solution eliminates the need for manual intervention in weight parameter tuning, reducing model debugging complexity. The dynamic decay strategy reduces the risk of training oscillations and shortens the neural network convergence time. Collaborative updating of neighboring weights enhances the ability to capture nonlinear features, providing a more stable feature space distribution basis for health assessment.

[0014] Optionally, the step of calculating the topological distance weight of the neighboring nodes based on the position coordinates of the best matching unit in the output layer includes calculating the Euclidean distance between the position coordinates of the neighboring nodes and the best matching unit; and inputting the Euclidean distance into the radial basis function to generate the topological distance weight.

[0015] By employing the aforementioned technical solution, the Euclidean distance between the position coordinates of neighboring nodes and the best matching unit is quantified, and this Euclidean distance is input into the radial basis function to generate topological distance weights, thus refining the weight update process. This scheme uses Euclidean distance to accurately characterize the spatial distribution relationship of neighboring nodes, and the radial basis function nonlinearly transforms linear distance into weight coefficients, strengthening the expression of local correlation in the topological structure. Dynamically adjusting the neighborhood influence decay gradient through function mapping ensures that nodes closer to the best matching unit receive higher update weights, effectively maintaining the topological continuity of the feature mapping. This scheme avoids the blindness and limitations of manually setting the neighborhood range, and the smoothness of the radial basis function suppresses the risk of weight mutations, improving the stability and convergence efficiency of the self-organizing map neural network training. Through adaptive weight allocation, the self-organizing map neural network's ability to learn deep spatial patterns of injector normal operating conditions is enhanced, establishing a highly robust feature space benchmark for health assessment, ultimately optimizing the sensitivity and accuracy of state recognition, while reducing the parameter tuning complexity and overfitting risk of the self-organizing map neural network.

[0016] Optionally, the automatic generation of the reference distance threshold based on the kernel density probability model includes constructing a kernel density probability model of normal time-domain features based on the output layer of a trained self-organizing map neural network; and automatically generating the reference distance threshold according to a preset anomaly confidence level.

[0017] By adopting the above technical solution, a kernel density probability model of normal time-domain features is constructed based on the output layer of a trained self-organizing map neural network. A reference distance threshold is automatically generated according to a preset anomaly confidence level, achieving intelligent and objective setting of the reference distance threshold. This solution utilizes the kernel density probability model to perform nonparametric probability estimation of the feature space distribution, accurately acquiring the feature clustering characteristics of normal operating conditions and avoiding subjective bias from manually setting the reference distance threshold. Furthermore, by dynamically adjusting the anomaly judgment boundary through preset confidence parameters, it flexibly adapts to the fault tolerance requirements of different operating conditions, significantly improving the adaptability of fault diagnosis. The kernel density model effectively captures nonlinear patterns and edge distributions in the feature space, enhancing the generalization ability to complex operating conditions. The automatic generation mechanism significantly reduces the model debugging cycle and manual parameter tuning costs, ensuring that the reference distance threshold strictly matches the statistical characteristics of the actual operating state of the injector.

[0018] Optionally, this includes collecting the mapping positions of all normalized normal temporal features in the output layer of a self-organizing map neural network; applying a kernel density estimation algorithm to establish a kernel density probability model of the normal temporal features; locating the boundary position corresponding to a preset anomaly confidence level in the kernel density probability model; and calculating the Euclidean distance from the boundary position to the cluster center of the normalized normal temporal features as a reference distance threshold.

[0019] By adopting the above technical solution, the mapping positions of all normalized normal time-domain features in the output layer of the self-organizing map neural network are collected, and a kernel density estimation algorithm is applied to construct a probabilistic model of the feature space distribution, accurately quantifying the clustering characteristics of normal operating conditions. Furthermore, based on a preset anomaly confidence level, the probability boundary is located, and a reference distance threshold is automatically generated by calculating the Euclidean distance from the boundary to the feature cluster center, achieving high precision and adaptability in threshold setting. This solution utilizes the kernel density model to perform non-parametric probability estimation of the feature space, comprehensively capturing the distribution patterns and marginal characteristics of normal data, avoiding threshold deviations caused by human experience intervention. The confidence level parameter dynamically adjusts the anomaly judgment boundary, flexibly adapting to different fault tolerance requirements, significantly improving the robustness and scenario universality of fault diagnosis. A physically meaningful threshold benchmark is established through the cluster center distance metric, ensuring a strict match between the health assessment and the statistical characteristics of the actual operating state of the injector.

[0020] Optionally, the cluster center of the normalized normal temporal features is the arithmetic mean of the mapping positions of all normalized normal temporal features in the output layer of the self-organizing map neural network.

[0021] By adopting the above technical solution, the arithmetic mean of all mapping positions of the normalized normal time-domain features in the output layer of the neural network is defined as the cluster center, significantly simplifying the process of determining the core benchmark of the feature space. This solution eliminates the need for complex iterative calculations or additional clustering algorithms, directly utilizing the linear properties of the arithmetic mean to efficiently generate center points, greatly reducing computational complexity. The calculation of the arithmetic mean ensures that the center positions have clear mathematical meaning and physical interpretability, avoiding convergence bias problems that may be caused by nonlinear optimization.

[0022] Secondly, this application provides a system for health assessment of a common rail injector for ships, including a data acquisition module, a data processing module, a neural network training module, and a health assessment module; The data acquisition module is used to acquire the normal pressure signal of the common rail when the injector is working properly, as well as the real-time pressure signal acquired in real time. The data processing module is used to preprocess normal / real-time pressure signals to remove outliers and noise data, extract time-domain features from the cleaned pressure signals, and normalize the normal / real-time time-domain features to a predetermined range. The neural network training module is used to train a self-organizing map neural network using normalized normal temporal features to obtain the best matching unit and obtain the reference distance threshold. The health assessment module is used to calculate the Euclidean distance between the real-time temporal features and the best matching unit in the output layer of the self-organizing map neural network, and compare the relationship between the real-time temporal features and the best matching unit and the reference distance threshold. When the Euclidean distance between the real-time temporal features and the best matching unit is less than the reference distance threshold, the injector is judged to be in a healthy state. When the Euclidean distance between the real-time temporal features and the best matching unit is greater than the reference distance threshold, the injector is judged to be in a fault state.

[0023] Understandably, the system for assessing the health of common rail injectors for ships 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.

[0024] In summary, this application includes at least one of the following beneficial technical effects: 1. By using the time-domain features of preprocessed normal pressure signals to train a self-organizing map neural network and based on the Euclidean distance comparison mechanism of the best matching unit, the system can accurately capture subtle abnormal features of the injector pressure signal, effectively solving the problem that the traditional threshold method relies on human experience and cannot adapt to dynamic working conditions. This enables intelligent real-time identification of the health status of ship common rail injectors and significantly improves the accuracy of fault warning.

[0025] 2. Because the reference distance threshold is automatically generated based on the kernel density distribution of normal time domain characteristics, and the boundary positioning is optimized by combining the preset anomaly confidence level, it can adaptively learn the data distribution characteristics of the normal operation of the injector, effectively solving the risk of misjudgment caused by equipment aging or environmental changes due to fixed thresholds, thereby ensuring the long-term robustness and generalization ability of the health assessment model and reducing maintenance costs.

[0026] 3. By integrating the complete process of data cleaning, feature extraction, neural network training and real-time evaluation, and designing a dynamic weight update mechanism to optimize network convergence efficiency, it can simultaneously realize injector health status diagnosis and model self-optimization, effectively solving the problems of lag in traditional timed monitoring and low efficiency of manual intervention, thereby providing highly reliable real-time monitoring for ship power systems and ensuring navigation safety and equipment lifespan. Attached Figure Description

[0027] 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.

[0028] Figure 1 This is a flowchart illustrating the method for health assessment of marine common rail injectors provided in the embodiments of this application; Figure 2 This is a schematic diagram of the system for health assessment of common rail injectors for ships provided in the embodiments of this application. Detailed Implementation

[0029] The following is in conjunction with the appendix Figure 1 —2 provides further detailed description of this application.

[0030] This application discloses a method and system for health assessment of common rail injectors for ships.

[0031] Reference Figure 1 A method for health assessment of a common rail injector for ships includes S1, acquiring the normal pressure signal of the common rail pipe when the injector is operating normally.

[0032] The normal pressure signal refers to the dynamic change data of fuel pressure in the common rail fuel system when the injectors are in a healthy state and operating under calibrated conditions. This data characterizes the standard pressure fluctuation characteristics of the system under fault-free conditions. The common rail, as the fuel storage and distribution unit, has its internal pressure established by the fuel pump and controlled by the pressure regulating valve, providing a stable injection back pressure for the injectors. Normal injector operation means that the injectors accurately complete fuel metering, injection timing, and injection pattern according to the instructions of the engine control unit, without internal leakage, needle valve sticking, solenoid valve failure, or other malfunctions.

[0033] Specifically, a piezoelectric pressure sensor mounted on the common rail continuously measures the pressure at preset calibrated speeds and load conditions. This pressure sensor has a sufficient dynamic response frequency to capture the details of pressure fluctuations caused by the periodic opening and closing of the injectors in the common rail. The measurement process must cover the engine's stable operating cycle under calibrated conditions to ensure that the collected normal pressure signal contains a sufficient number of complete injection cycles to fully reflect the pressure dynamics under healthy conditions. The collected raw analog voltage signal is then converted into discrete time-series data by an analog-to-digital converter, which constitutes the normal pressure signal for subsequent processing.

[0034] It should be noted that, to ensure the acquired normal pressure signal accurately represents the injector's health status, data acquisition must be conducted under strictly controlled test conditions, including stable fuel temperature, coolant temperature, oil pressure, and environmental conditions. Simultaneously, it is necessary to confirm beforehand, through professional diagnostic equipment or rigorous physical inspection, that the injectors themselves and related critical components such as the common rail, fuel pump, and pressure regulating valve are in good working order. Furthermore, while acquiring the normal pressure signal, the real-time engine speed, load command, and injector control pulse width are simultaneously recorded to provide necessary background information for subsequent correlation analysis. The final dataset of normal pressure signals serves as the raw input for the entire health assessment method and the benchmark for determining health status.

[0035] S2. After preprocessing the normal pressure signal, extract the normal time domain features and normalize the normal time domain features.

[0036] In this embodiment, S2 specifically includes cleaning the normal pressure signal to remove outliers and noise data; extracting normal time-domain features from the cleaned normal pressure signal; and normalizing the normal time-domain features to a predetermined interval.

[0037] Preprocessing refers to the process of eliminating non-ideal interference components in normal pressure signals using specific algorithms to ensure the reliability of subsequent feature extraction. This includes data cleaning, outlier removal, and noise suppression. Outliers are isolated data points that significantly deviate from the true physical range of the pressure signal due to momentary sensor malfunctions or external electromagnetic interference. Noise data refers to high-frequency random fluctuations superimposed on the true pressure waveform, mainly originating from mechanical vibration, circuit thermal noise, and signal transmission crosstalk. Normal time-domain features are statistical and morphological descriptive quantities calculated directly from the cleaned pressure signal time series, used to quantify the dynamic pressure behavior under injector health conditions. Normalization involves mapping normal time-domain features of different dimensions or magnitudes to a unified predetermined numerical range through linear transformation, eliminating the negative impact of differences in the numerical range of the normal time-domain features themselves on subsequent network model training.

[0038] Specifically, the normal pressure signal is first cleaned by using a sliding window combined with a standard deviation threshold method to identify and remove transient outliers. Simultaneously, wavelet packet transform is used to decompose the signal into multiple scales, suppressing high-frequency noise by reconstructing low-frequency coefficients and preserving the pressure fluctuation essence characterizing the injection event. Then, key normal time-domain features are extracted from the cleaned normal pressure signal, including but not limited to signal mean, variance, peak value, trough value, peak-to-peak value, root mean square, waveform factor, impulse factor, margin factor, and zero-crossing rate. These normal time-domain features represent the static offset, fluctuation intensity, extreme amplitude, energy level, and periodicity of the pressure signal, respectively. The extracted multivariate normal time-domain features are then mapped to a predetermined interval using a max-min normalization algorithm, linearly transforming the values ​​of all normal time-domain features to the same scale.

[0039] It should be noted that the adaptive sliding window width used in the data cleaning stage must cover at least one complete injection cycle to ensure that outlier detection does not disrupt the continuity of the pressure waveform. The normal time-domain feature extraction process is strictly limited to the stable operating range of the main unit to avoid non-steady-state interference introduced by changes in main unit speed or load. The predetermined interval used for normalization must cover the theoretical distribution range of all normal time-domain feature values ​​and allow for margins for potentially unknown data. Furthermore, the normalized normal time-domain feature vector, while preserving the relative relationships of the original normal pressure signals, significantly improves the convergence speed and stability of the self-organizing map neural network in subsequent training stages. The final generated standard normal time-domain feature set serves as the basic input for training the self-organizing map neural network and a digital representation of the health benchmark.

[0040] S3. Construct and train a self-organizing map neural network using the normalized normal temporal features to obtain the best matching unit and obtain the reference distance threshold; where the best matching unit is the node in the output layer of the self-organizing map neural network with the smallest Euclidean distance to the input normal temporal features.

[0041] In this embodiment, S3 specifically includes setting the number of input layer nodes of the self-organizing map neural network to be the same as the number of normalized normal temporal features; setting the output layer to a two-dimensional planar structure and initializing the network weights. The self-organizing map neural network is trained using the normalized normal temporal features; the Euclidean distance between the normalized normal temporal features and the weights of each output layer node is calculated, and the output layer node with the smallest Euclidean distance is selected as the best matching unit. The Euclidean distance between the position coordinates of neighboring nodes and the best matching unit is calculated, and this Euclidean distance is input into the radial basis function to generate topological distance weights. A dynamic decay update coefficient is generated by combining the learning rate and the topological distance weights, and the weight vectors of the best matching unit and its neighboring nodes are adjusted using the dynamic decay update coefficient. Then, the mapping positions of all normalized normal temporal features in the output layer of the self-organizing map neural network are collected, and a kernel density estimation algorithm is applied to establish a kernel density probability model of the normal temporal features; the boundary position corresponding to the preset anomaly confidence level is located in the kernel density probability model, and the Euclidean distance from the boundary position to the cluster center of the normalized normal temporal features is calculated as a reference distance threshold. The cluster center of the normalized normal temporal features is the arithmetic mean of the mapping positions of all normalized normal temporal features in the output layer of the self-organizing map neural network.

[0042] The self-organizing map neural network (SOM) is an unsupervised neural network that maps high-dimensional input data to a low-dimensional discrete space through competitive learning. The spatial topology of its output layer nodes preserves the feature distribution of the input data. The reference distance threshold is the boundary distance determined based on the statistical distribution of the health status data (i.e., normal pressure signal) mapped to the output layer, used to determine abnormal injector conditions. The cluster center of the normalized normal time-domain features is the arithmetic mean of the coordinates of all normalized normal time-domain features mapped to the output layer of the SOM, i.e., the geometric center point determined by the arithmetic mean of the coordinates of all mapped points in each dimension.

[0043] Specifically, firstly, the number of nodes in the input layer of the self-organizing map neural network is set to be equal to the number of normalized normal temporal features. The output layer is constructed as a two-dimensional planar grid structure, and the node weight vectors are randomly initialized. Then, the normalized normal temporal features are input into the network one by one, and the Euclidean distance between each feature vector and the weights of all output layer nodes is calculated. The node with the smallest Euclidean distance is selected as the best matching unit. Next, the Euclidean distance between each node in the neighborhood of the best matching unit and its position coordinates in the output layer grid is calculated. This Euclidean distance is input into the radial basis function, thus mapping the Euclidean distance to a topological distance weight in the [0,1] interval. Because the radial basis function is sensitive to local features and has a smooth decay characteristic, it is particularly suitable for signal processing in ship vibration environments and can effectively adapt to the nonlinear topological relationship generated when the common rail pressure signal is disturbed by complex engine vibration. Combined with a preset learning rate and topological distance weights, a dynamic update coefficient is generated that decays with the number of network iterations and decreases with the increase of the topological distance. This coefficient is used to synchronously adjust the weight vectors of the best matching unit and its neighboring nodes. Accordingly, the weight vectors of the best matching unit and its neighboring nodes are adjusted synchronously, allowing the network to gradually learn the spatial distribution of health status features. After training, the coordinates of the mapping positions of all normalized normal temporal features in the output layer are collected, and their arithmetic mean is calculated as the cluster center. The calculation formula is as follows: Let the set of the mapping position coordinates of each normalized normal temporal feature in the output layer be {P1, P2, ..., P...}. n}, P i = (x i y i Where n and i are both positive integers and i ≤ n; then the cluster center C i =(C x C y )satisfy , , where N is a positive integer and N≤n. Then, a kernel density estimation algorithm is applied to construct a kernel density probability model of the mapping location, locate the probability boundary position corresponding to the preset anomaly confidence level, and calculate the Euclidean distance from the probability boundary position to the cluster center as a reference distance threshold.

[0044] It should be noted that the radial basis function adopts a Gaussian kernel form, and its scaling parameter is adaptively adjusted according to the typical vibration spectrum characteristics of the ship's common rail system, ensuring accurate characterization of the topological correlation between nodes even under strong background vibration noise during ship operation. The design of the dynamic attenuation update coefficient in the weight update allows for a larger adjustment range for nearest neighbor nodes, conforming to the principle of spatial continuity of the state characteristics of the ship's propulsion system. The kernel density probability model employs an adaptive bandwidth algorithm to avoid density estimation bias caused by the non-stationary characteristics of the vibration signal; the final obtained reference distance threshold mathematically represents the confidence boundary of the health state characteristic distribution, and engineeringally defines a quantitative criterion for injector performance degradation and failure.

[0045] S4. Process the real-time pressure signal acquired in real time according to step S2 and obtain the real-time time domain features. Calculate the Euclidean distance between the real-time time domain features and the best matching unit, and compare its magnitude with the reference distance threshold.

[0046] The real-time pressure signal refers to the dynamic sequence of injector operating pressure continuously collected by the common rail pressure sensor during ship navigation, directly reflecting the current actual operating status of the injectors. Real-time temporal features refer to the standardized real-time temporal features generated after processing the real-time pressure signal according to a data cleaning, feature extraction, and normalization process completely consistent with the normal pressure signal. The optimal matching unit is the node in the output layer of the trained self-organizing map neural network whose weight vector has the smallest Euclidean distance to the normal temporal feature vector; its spatial position represents the nearest neighbor of the current feature in the health state mapping topology. Here, the Euclidean distance is defined as the square root of the sum of the squares of the differences in each dimension between the real-time temporal feature vector and the optimal matching unit weight vector, quantifying the degree of deviation between the current state and the normal health state benchmark.

[0047] Specifically, the data processing module is first invoked to perform a preprocessing operation on the real-time pressure signal that is completely isomorphic to step S2, and the real-time time-domain features are mapped to the same predetermined interval based on the normalization parameters determined during the training phase; then the normalized real-time time-domain features are input into the trained self-organizing map neural network, and the Euclidean distance between the real-time time-domain features and the weight vector of the best matching unit is output.

[0048] It should be noted that the search process for the best matching unit employs a full traversal strategy rather than an approximation algorithm, ensuring accurate location of the nearest neighbor node even under complex ship operating conditions. The calculation of the Euclidean distance strictly preserves the physical meaning of each dimension of the normal time-domain feature vector, avoiding the loss of state information due to dimensionality reduction. This Euclidean distance value, as a core indicator, is essentially the local spatial deviation between the current injector pressure dynamic characteristics and the cluster center of the health status characteristics, providing direct input for subsequent comparison of reference distance thresholds. The entire processing has a defined execution time, meeting the real-time requirements of online health monitoring of the ship's power system. The final generated Euclidean distance value constitutes the core input for the quantitative criterion of the injector's real-time health status.

[0049] Among them, the size relationship refers to the numerical comparison result between the Euclidean distance value calculated in real time and the reference distance threshold generated in step S3. This result directly determines the qualitative judgment of the current state of the injector.

[0050] Specifically, the health assessment module receives the Euclidean distance value between real-time temporal features and the best matching unit, while simultaneously loading a pre-stored reference distance threshold. It then performs a scalar comparison operation through a calculation engine to determine whether the real-time Euclidean distance is less than or equal to the reference distance threshold. This comparison process uses hardware-level floating-point comparison instructions to ensure completion within the ship's embedded system clock cycle, meeting the real-time response requirements of online monitoring. The comparison result is output as a Boolean logic value; a true value indicates that the current state meets the health judgment conditions, while a false value indicates an abnormal state requiring subsequent diagnostics.

[0051] It should be noted that the reference distance threshold has dynamic robustness in the case of ship operating conditions because it integrates the statistical confidence of the health status feature distribution through the kernel density probability model, which effectively overcomes the feature drift problem caused by fluctuations in ship main engine speed and load. The comparison logic design adopts an asymmetric fault-tolerant mechanism to smooth the Euclidean distance spikes caused by instantaneous disturbances through a sliding window, avoiding misjudgments caused by the complex vibration environment of the ship.

[0052] S5. When the Euclidean distance between the real-time temporal feature and the best matching unit is less than the reference distance threshold, the injector is determined to be in a healthy state.

[0053] In this embodiment, S5 further includes determining that the injector is currently in a fault state when the Euclidean distance between the real-time time domain feature and the best matching unit is greater than the reference distance threshold.

[0054] Specifically, the health assessment module determines that the injector is in a healthy state and updates the cumulative healthy runtime of the injector when the Euclidean distance between the real-time time-domain characteristics and the best-matching unit is less than a reference distance threshold. When the Euclidean distance between the real-time time-domain characteristics and the best-matching unit is greater than the reference distance threshold, the injector is determined to be in a fault state and an abnormal alarm is triggered. This judgment process incorporates a state transition lag interval design to avoid instantaneous state jumps caused by severe ship vibration or operating condition switching, ensuring the engineering reliability of the diagnostic conclusions. The final state is transmitted to the ship's engine room monitoring system via the ship's CAN bus, driving health management log recording and maintenance decisions.

[0055] Reference Figure 2 This is a schematic diagram of the system structure for health assessment of a common rail injector in a ship, as described in this application.

[0056] The system for health assessment of common rail injectors for ships includes a data acquisition module 1, a data processing module 2, a neural network training module 3, and a health assessment module 4.

[0057] The data acquisition module 1 consists of a pressure sensor 11 and a signal conditioning circuit 12, used to acquire the normal pressure signal of the common rail during normal injector operation, as well as the real-time pressure signal. The pressure sensor 11 is directly installed at the pressure measuring port of the injector's common rail. The sensor's operating temperature range covers the extreme thermal environment during the operation of the ship's main engine, and its dynamic response characteristics match the microsecond-level pressure fluctuations of fuel injection. The acquired normal pressure signal and real-time pressure signal are transmitted to the data processing module 2 via the ship's bus connection.

[0058] The data processing module 2 consists of an industrial-grade multi-core microprocessor 21 and an analog-to-digital converter 22. It is used to preprocess normal / real-time pressure signals to remove outliers and noise, extract time-domain features from the cleaned pressure signals, and normalize the normal / real-time time-domain features to a predetermined range. It is primarily implemented based on the industrial-grade multi-core microprocessor 21, which integrates the analog-to-digital converter 22 to receive the raw normal and real-time pressure signals acquired by the sensors. Time-domain feature extraction and normalization operations are performed on the industrial-grade multi-core microprocessor 21.

[0059] The neural network training module 3 is mainly composed of parallel computing units 31, which are used to train the self-organizing map neural network using normalized normal time-domain features, obtain the best matching unit and obtain the reference distance threshold, and output it to the health assessment module 4.

[0060] The health assessment module 4 mainly consists of a real-time computing processing unit (41 elements). It calculates the Euclidean distance between real-time temporal features and the best-matching unit in the output layer of the self-organizing map neural network, and compares this distance with a reference distance threshold. When the Euclidean distance between the real-time temporal features and the best-matching unit is less than the reference distance threshold, the injector is considered to be in a healthy state; when the distance is greater than the reference distance threshold, the injector is considered to be in a fault state.

[0061] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the above-described division of functional units and modules is used as an example. In practical applications, the above functions can be assigned to different functional units and modules as needed, that is, the internal structure of the device can be divided into different functional units or modules to complete all or part of the functions described above.

[0062] 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.

[0063] 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 health assessment of marine common rail injectors, characterized in that: This includes S1, which acquires the normal pressure signal of the common rail for normal fuel injector operation; S2. After preprocessing the normal pressure signal, extract the normal time domain features and normalize the normal time domain features. S3. Construct and train a self-organizing map neural network using the normalized normal temporal features to obtain the best matching unit and obtain the reference distance threshold; where the best matching unit is the node in the output layer of the self-organizing map neural network with the smallest Euclidean distance to the input normal temporal features. Specifically, S3 includes: constructing a self-organizing map neural network; inputting normalized normal temporal features into the self-organizing map neural network; calculating the Euclidean distance between the normalized normal temporal features and the weights of each network output layer node, and obtaining the output layer node with the smallest Euclidean distance as the best matching unit; updating the weights according to the learning rate based on the positions of the best matching unit and its neighboring nodes; repeating the iteration until the self-organizing map neural network converges, and automatically generating a reference distance threshold based on the kernel density probability model. The automatic generation of reference distance threshold based on kernel density probability model includes constructing a kernel density probability model of normal time domain features based on the output layer of a trained self-organizing map neural network; and automatically generating reference distance threshold according to a preset anomaly confidence level. Specifically, this includes collecting the mapping positions of all normalized normal temporal features in the output layer of a self-organizing map neural network; establishing a kernel density probability model of the normal temporal features using a kernel density estimation algorithm; locating the boundary positions corresponding to preset anomaly confidence levels in the kernel density probability model; using the arithmetic mean of the mapping positions of all normalized normal temporal features in the output layer of the self-organizing map neural network as the cluster center of the normalized normal temporal features; and calculating the Euclidean distance from the boundary positions to the cluster centers of the normalized normal temporal features as a reference distance threshold. S4. Process the real-time pressure signal acquired in real time according to step S2 and obtain the real-time time domain features. Calculate the Euclidean distance between the real-time time domain features and the best matching unit, and compare its magnitude with the reference distance threshold. S5. When the Euclidean distance between the real-time temporal feature and the best matching unit is less than the reference distance threshold, the injector is determined to be in a healthy state.

2. The method for health assessment of a ship common rail injector according to claim 1, characterized in that: S5 also includes determining that the injector is currently in a fault state when the Euclidean distance between the real-time time-domain feature and the best matching unit is greater than the reference distance threshold.

3. The method for health assessment of a ship common rail injector according to claim 1, characterized in that: S2 specifically includes cleaning the normal pressure signal to remove outliers and noise data; and extracting normal time-domain features from the cleaned normal pressure signal. The normal time-domain features are normalized to a predetermined interval.

4. The method for health assessment of a ship common rail injector according to claim 1, characterized in that: The step of updating weights according to the learning rate based on the position of the best matching unit and its neighboring nodes includes calculating the topological distance weights of the neighboring nodes based on the position coordinates of the best matching unit in the output layer; combining the learning rate and the topological distance weights to generate a dynamic decay update coefficient that decreases with the number of iterations of the self-organizing map neural network and decreases with the increase of the topological distance between the neighboring nodes and the best matching unit; and using the dynamic decay update coefficient to adjust the weight vector of the best matching unit and its neighboring nodes.

5. The method for health assessment of a ship common rail injector according to claim 4, characterized in that: The step of calculating the topological distance weights of neighboring nodes based on the position coordinates of the best matching unit in the output layer includes calculating the Euclidean distance between the position coordinates of the neighboring nodes and the best matching unit; and inputting the Euclidean distance into the radial basis function to generate the topological distance weights.

6. A system for health assessment of marine common rail injectors, characterized in that, It includes a data acquisition module, a data processing module, a neural network training module, and a health assessment module; The data acquisition module is used to acquire the normal pressure signal of the common rail when the injector is working properly, as well as the real-time pressure signal acquired in real time. The data processing module is used to preprocess normal / real-time pressure signals to remove outliers and noise data, extract time-domain features from the cleaned pressure signals, and normalize the normal / real-time time-domain features to a predetermined range. The neural network training module is used to train a self-organizing map neural network using normalized normal temporal features to obtain the best matching unit and obtain the reference distance threshold. The neural network training module is also used to collect the mapping positions of all normalized normal temporal features in the output layer of the self-organizing map neural network; apply the kernel density estimation algorithm to establish a kernel density probability model of the normal temporal features; locate the boundary position corresponding to the preset anomaly confidence level in the kernel density probability model; and use the arithmetic mean of the mapping positions of all normalized normal temporal features in the output layer of the self-organizing map neural network as the cluster center of the normalized normal temporal features. The Euclidean distance from the boundary location to the normalized normal temporal feature cluster center is calculated as a reference distance threshold. The health assessment module is used to calculate the Euclidean distance between the real-time temporal features and the best matching unit in the output layer of the self-organizing map neural network, and compare the relationship between the real-time temporal features and the best matching unit and the reference distance threshold. When the Euclidean distance between the real-time temporal features and the best matching unit is less than the reference distance threshold, the injector is judged to be in a healthy state. When the Euclidean distance between the real-time temporal features and the best matching unit is greater than the reference distance threshold, the injector is judged to be in a fault state.

Images