A multi-source heterogeneous closed-loop intelligent operation and maintenance method and system for ammonia, LNG and methanol fuel ship power systems

Abstract

The application relates to the technical field of intelligent operation and maintenance and fault diagnosis of a ship power system, and discloses a multi-source heterogeneous closed-loop intelligent operation and maintenance method and system for an ammonia, LNG and methanol fuel ship power system.The method collects multi-source heterogeneous operation signals, constructs a unified feature sequence, and inputs the unified feature sequence and a working condition embedded Token into a Transformer encoder to extract a global health state representation; the method simultaneously outputs fault types, fault severity and residual life prediction through a multi-task output module; combined with a diagnosis result, life prediction and model confidence, a gating decision matrix is used to generate a graded operation and maintenance instruction; after operation and maintenance, retesting and maintenance results are written back as closed-loop feedback samples, and after effectiveness determination, model parameters and decision strategies are corrected through weighted incremental learning.The application realizes multi-source signal fusion, multi-working condition self-adaptation, diagnosis-prediction-decision cooperation and feedback closed-loop optimization, and improves the fault diagnosis accuracy and operation and maintenance decision reliability of a clean fuel ship power system.

Description

Technical Field

[0001] This invention relates to the field of intelligent operation and maintenance and fault diagnosis technology for marine power systems, and in particular to a multi-source heterogeneous closed-loop intelligent operation and maintenance method and system for ammonia, LNG and methanol fuel marine power systems. Background Technology

[0002] As global shipping industry continues to pursue emission reduction goals, clean fuel ship propulsion systems have become a core direction for the industry's green transformation. Clean fuels such as ammonia, LNG (liquefied natural gas), and methanol, due to their low-carbon or even zero-carbon combustion characteristics, are considered by the International Maritime Organization (IMO) and shipping authorities in various countries as important technological pathways to achieving carbon neutrality in the shipping industry. Currently, new propulsion systems such as ammonia-fueled engines, LNG dual-fuel main engines, and methanol reforming hydrogen fuel cells have gradually moved from laboratory verification to real-ship applications. Globally, dozens of clean fuel-powered vessels are already in operation or under construction, covering various ship types including large ocean-going container ships, bulk carriers, tankers, and coastal feeder vessels.

[0003] In the aforementioned clean fuel marine propulsion systems, multiple functional modules, such as the propulsion system, fuel supply system, fuel storage and vaporization system, and safety monitoring and ventilation system, are highly integrated. Key moving components—including the main engine crankshaft and piston assembly, propulsion shaft system and intermediate bearings, motors and couplings, gearbox transmission mechanisms, compressors and pump sets—operate for extended periods in harsh environments characterized by high vibration, strong impact, variable speed, and variable load. The health status of these components directly affects the operational safety, fuel economy, and life-cycle maintenance costs of the marine propulsion system. To ensure reliable system operation, ship operators typically employ equipment condition monitoring schemes based on vibration monitoring, oil analysis, temperature monitoring, and performance parameter recording. They also install various types of sensing devices on key equipment, such as acceleration sensors, eddy current displacement sensors, thermocouples, pressure transmitters, and oil particle counters, to achieve real-time perception and early warning of anomalies in the propulsion system's operating status.

[0004] Despite the increasing sophistication of sensor technology and data acquisition methods, intelligent operation and maintenance of clean fuel ship propulsion systems still faces a series of technical challenges in practical applications, including:

[0005] First, multi-source heterogeneous signals are difficult to represent and effectively fuse in a unified manner. Monitoring data for ship propulsion systems come from diverse sources, including vibration signals (acquired by accelerometers, with sampling frequencies typically above 10 kHz), speed signals (acquired by key phase sensors), load signals (acquired by torque meters), pressure signals (acquired by pressure transmitters, with relatively low sampling frequencies), temperature signals (acquired by thermocouples, with slow changes), oil physicochemical indicators (analyzed offline or online), and equipment operation logs and maintenance records. These signals differ fundamentally in sampling frequency, physical dimensions, amplitude range, and data format, making them typical multi-source heterogeneous data. Most existing fault diagnosis methods only perform feature extraction and pattern recognition on a single type of signal (usually vibration signals), failing to fully utilize the complementarity and redundancy between multi-source information. Even in the few studies that attempt to fuse multi-source signals, the fusion strategies are mostly shallow feature splicing or simple weighted averaging, lacking a unified modeling capability for the deep correlation between different signals, making it difficult to comprehensively reflect the overall health status of complex propulsion systems.

[0006] Secondly, the complex and variable operating conditions lead to a severe deficiency in the model's generalization ability. During navigation, the main engine speed and propeller load change in real time with factors such as sea state, speed, rudder angle, and fuel switching commands, resulting in significant differences in the statistical distribution characteristics of monitoring signals under different operating conditions. Specifically, the energy distribution, dominant frequency components, and amplitude range of the vibration signal of the same bearing at rated speed and half speed may differ by several times to more than an order of magnitude. Existing methods usually treat data under different operating conditions as samples with the same distribution, or simply model them in segments according to operating conditions. The former leads to the model misjudging normal signal changes caused by operating conditions as fault features, while the latter makes it difficult to establish a diagnostic model with good generalization ability due to insufficient effective training samples under each operating condition. More challenging is that clean fuel power systems cover different fuel forms such as ammonia, LNG, and methanol. The combustion characteristics, in-cylinder pressure curves, and excitation force spectra of different fuels are fundamentally different, further exacerbating the interference of operating condition changes on fault feature extraction. This problem means that existing diagnostic methods may perform well under a single operating condition, but once the ship enters a variable operating condition state, the diagnostic accuracy drops significantly.

[0007] Third, fault diagnosis, degradation assessment, and operation and maintenance decision-making are disconnected. Most current fault diagnosis methods only output a classification of "whether there is a fault" or "what kind of fault," failing to simultaneously assess the severity of the fault and remaining service life, and even more so, making it difficult to effectively translate diagnostic results into actionable operation and maintenance recommendations. In actual ship operation and maintenance management processes, maintenance personnel not only need to know what kind of fault has occurred, but also need to clarify "how severe the fault is," "how long it can still operate," and "what maintenance measures should be taken." Due to the lack of a unified framework to connect state perception, degradation assessment, service life prediction, and operation and maintenance decision-making into a complete logical chain, existing systems typically require manual secondary assessment based on diagnostic results before formulating maintenance plans. This process is not only inefficient and highly subjective, but also prone to decision-making errors due to human judgment delays or mistakes in emergency situations, severely weakening the practical application value of intelligent operation and maintenance systems.

[0008] Fourth, there is a lack of a closed-loop optimization mechanism driven by maintenance feedback. Most existing intelligent operation and maintenance systems adopt an "open-loop" working mode, that is, the process ends after the diagnostic model outputs the results. After maintenance personnel perform maintenance operations based on the diagnostic results, the real fault information obtained during the maintenance process (such as the fault type confirmed by disassembly and inspection, the actual wear degree, the remaining life of the replaced parts, etc.) and retest data are not fed back to the diagnostic model to correct model parameters and decision-making strategies. This results in the diagnostic model being unable to accumulate experience and correct deviations from each actual maintenance activity, and the model performance stagnates in the long term. Especially when the performance of the ship's power system slowly degrades after long-term operation, or when the sensor characteristics drift, the open-loop system lacks adaptive adjustment capabilities, and the false alarm rate and false negative rate will gradually increase over time until they can no longer meet the actual operation and maintenance needs.

[0009] Therefore, those skilled in the art urgently need an end-to-end multi-source heterogeneous closed-loop intelligent operation and maintenance technology solution that can be deployed in the intelligent monitoring system of ship engine room. Summary of the Invention

[0010] The core technical problem addressed by this invention is: how to construct an intelligent operation and maintenance method and system for clean fuel ship propulsion systems that can simultaneously achieve unified modeling of multi-source heterogeneous signals, adaptive perception of complex operating conditions, integrated output of fault diagnosis, degradation assessment, life prediction, and operation and maintenance decision-making, and possess maintenance feedback closed-loop optimization capabilities.

[0011] For those skilled in the art, the solution to the aforementioned core technical problems has extremely important technical value and significance in the industry, as detailed below;

[0012] First, clean fuel ship propulsion systems are key equipment for the shipping industry to achieve carbon neutrality. If intelligent operation and maintenance technology remains at the level of "single signal, single operating condition, single output, and open-loop operation" for a long time, it will seriously restrict the actual ship promotion and large-scale application of clean fuel propulsion systems.

[0013] Secondly, from the perspective of technological evolution, solving the aforementioned core technical problems will break through the bottlenecks of existing fault diagnosis technologies in three aspects:

[0014] At the signal processing level, we will break through the theoretical and methodological bottlenecks in the unified representation of multi-source heterogeneous data;

[0015] At the model generalization level, we break through the bottleneck of adapting to fault feature drift under changing operating conditions;

[0016] At the system architecture level, we must break through the technical bottleneck of the diagnosis-decision-feedback closed loop;

[0017] Once these bottlenecks are overcome, it will drive a fundamental shift in the operation and maintenance model of ship propulsion systems from "regular inspections + post-incident maintenance" to "predictive maintenance + condition-based maintenance," significantly reducing the operating costs of ships throughout their entire life cycle and minimizing unplanned downtime and safety incidents.

[0018] Therefore, it is essential for those skilled in the art to solve this core technical problem in order to meet the urgent needs of the shipbuilding industry for intelligent and green development.

[0019] To address the aforementioned core technical challenges, this invention presents a multi-source heterogeneous closed-loop intelligent operation and maintenance method and system for ammonia, LNG, and methanol fuel-powered ship propulsion systems. The objective is to: in actual ship operation scenarios, utilize various installed sensors to collect multi-source heterogeneous operating signals; convert signals from different sources, with different dimensions, and different sampling frequencies into standardized sequences that the model can process through a unified feature extraction and fusion framework; simultaneously, explicitly encode operating condition information such as speed, load, and fuel type, and input them along with signal features into a deep network model based on an attention mechanism to extract global health status representations across operating conditions and components; based on this, a multi-task learning framework simultaneously outputs fault type, fault severity, and remaining life prediction results; and automatically generate hierarchical operation and maintenance decision instructions based on the above three aspects of information and model confidence; finally, write back the retest results and maintenance confirmation conclusions after maintenance execution to the system; and continuously correct model parameters and decision strategies through an incremental learning mechanism, thereby forming a closed-loop intelligent operation and maintenance capability of "monitoring—diagnosis—prediction—decision—feedback—optimization".

[0020] To achieve the above objectives, the specific technical solution of the present invention is a multi-source heterogeneous closed-loop intelligent operation and maintenance method for ammonia, LNG, and methanol fuel ship propulsion systems, the method comprising the following steps:

[0021] Step 1: For ammonia, LNG or methanol fuel ship propulsion systems, use sensors installed on key moving parts to collect multi-source heterogeneous operating signals, including vibration signals, pressure signals, temperature signals, oil state signals, speed signals and load signals;

[0022] Step 2: Perform time alignment, segmentation and statistical feature extraction on the multi-source heterogeneous operation signals, normalize them using global statistics of the training set, construct a unified feature sequence, and obtain the corresponding fault diagnosis results, fault severity assessment results, remaining lifetime prediction results and model confidence, and generate a pre-operation and maintenance state benchmark sample.

[0023] Step 3: Input the unified feature sequence and the condition embedding token constructed from the operating condition variables into the attention-based encoder to extract the global health status representation of the ship's power system.

[0024] Step 4: Based on the global health status representation, simultaneously output the fault type identification result, fault severity assessment result, and remaining lifetime prediction result through the multi-task output module;

[0025] Step 5: Based on the fault diagnosis results, fault severity assessment results, remaining lifetime prediction results, and model confidence, generate corresponding operation and maintenance action instructions through a gating decision mechanism;

[0026] Step 6: Obtain the retest data, action execution results, and maintenance confirmation results after the operation and maintenance actions are performed, and establish a one-to-one correspondence with the pre-operation and maintenance state benchmark sample to form a closed-loop feedback sample;

[0027] Step 7: Based on the changes in multi-source features before and after maintenance, the consistency between fault diagnosis results and maintenance confirmation results, and the model confidence in the closed-loop feedback samples, determine the validity of the closed-loop feedback samples, and dynamically correct the model parameters and decision-making strategies using a weighted incremental learning mechanism.

[0028] Preferably, the specific process of constructing the unified feature sequence in step 2 includes:

[0029] Step 2A: Let the first... The raw operating signals acquired by the sensor are According to the preset time window length Divide into segments to obtain the first segment. One signal segment:

[0030] ;

[0031] Step 2B: Extract statistical feature vectors for each signal segment. The statistical feature vector includes the mean. Standard deviation Root mean square value skewness kurtosis Peak factor Peak-to-peak value and energy The statistical feature vector The expression is:

[0032] ;

[0033] Step 2C: Extract features from signals from different sources within the same time window, categorized by the number of sensor signal sources. By splicing, we obtain the first... Multi-source fusion features of each window Among them, the first Multi-source fusion features of each window The expression is:

[0034]

[0035] Step 2D: Normalize all multi-source fusion features using the global mean and standard deviation of the training set, and arrange them in chronological order to obtain a unified feature sequence. ;

[0036] The expression for normalizing all multi-source fusion features using the global mean and standard deviation of the training set is as follows:

[0037]

[0038] In the formula, Global mean of the training set Global standard deviation of the training set It is a very small constant;

[0039] Arranged in chronological order, a unified feature sequence is obtained. , The expression is as follows:

[0040] .

[0041] Preferably, the fusion process of the working condition embedded token and the unified feature sequence in step 3 includes:

[0042] The unified feature sequence and operating condition variables respectively through the feature mapping matrix and working condition mapping matrix Projected into the same latent space, the position codes are superimposed after stitching. This yields the input representation of the attention-based encoder (Transformer encoder). Its expression is:

[0043]

[0044] In the formula, The input representation is used for the attention-based encoder (Transformer encoder). The feature mapping matrix, For feature mapping bias term, This is the working condition mapping matrix. For the operating condition mapping bias term, It is a non-linear activation function. For position encoding, This indicates a token concatenation operation.

[0045] Preferably, the encoder based on the attention mechanism in step 3 is a Transformer encoder. The Transformer encoder extracts the correlation between different time segments and signals from different sources through a multi-head self-attention mechanism to obtain a global health status representation of the ship's propulsion system. The calculation process of the multi-head self-attention includes:

[0046] Input representation for attention-based encoders , No. A query matrix for each attention head Key matrix Sum matrix By respectively through By performing a linear mapping, we obtain:

[0047]

[0048]

[0049]

[0050] In the formula, To query the mapping matrix, The key mapping matrix, For value mapping matrix, The dimension of the key vector. Given the number of attention heads, the output of each attention head is... for:

[0051]

[0052] The outputs of all attention heads are concatenated and linearly mapped to serve as the output of the multi-head self-attention layer. ,Right now:

[0053]

[0054] In the formula, This is a multi-head self-attention mapping matrix;

[0055] The global representation output by the encoder is denoted as: :

[0056]

[0057] The output corresponding to the global token is selected as the global health status representation of the ship's ammonia fuel propulsion system as follows:

[0058]

[0059] In the formula, It represents the overall health status.

[0060] Preferably, in step 4, the multi-task output module outputs fault type identification results, fault severity assessment results, and remaining lifetime prediction results based on the global health status representation. The specific implementation methods include:

[0061] The global health status representation Simultaneously, the inputs are fed into the fault classification output layer, severity regression output layer, and remaining lifetime prediction output layer of the multi-task output module. The fault classification output layer, the severity regression output layer, and the remaining lifetime prediction output layer respectively output fault type identification results, fault severity assessment results, and remaining lifetime prediction results. The output expressions for the three are as follows:

[0062]

[0063]

[0064]

[0065] In the formula, As a representation of the overall health status, The result of fault type identification. This is the result of the fault severity assessment. For the remaining life prediction results, , , These are the weight matrices for the corresponding tasks (i.e., fault type identification results, fault severity assessment results, and remaining lifetime prediction results). , , These are the bias terms for the corresponding tasks (i.e., fault type identification results, fault severity assessment results, and remaining lifetime prediction results). This is the output function for predicting remaining lifetime. This refers to the softmax function.

[0066] Preferably, the multi-task output module is jointly trained using a weighted multi-task loss function, which is:

[0067]

[0068] In the formula, Identify losses based on fault type. To assess the damage based on the severity of the fault, Predicting losses for remaining lifetime, , , These are the weight coefficients for the corresponding tasks; the three tasks achieve collaborative optimization of fault diagnosis, degradation assessment and lifetime prediction based on the same global health status representation.

[0069] Preferably, the operation and maintenance decision instruction in step 5 is determined by the following formula:

[0070]

[0071] In the formula, For operation and maintenance decision instructions, The operation and maintenance decision function (i.e., the operation and maintenance decision function that performs gating mapping based on the combination of damage level, fault feature encoding and remaining lifetime prediction interval). The model confidence level; the operation and maintenance decision instructions include one or more of the following: increasing the sampling frequency, strengthening the tracking of remaining life trend, supplementing lubrication, reducing load, limiting speed, generating maintenance work orders, soft shutdown, and emergency stop.

[0072] Preferably, the gating decision-making mechanism in step 5 employs a decision execution matrix, which maps to corresponding maintenance recommendations based on a combination of damage level, fault feature encoding, and remaining lifetime prediction interval.

[0073] The damage level includes three levels: S0, S1, and S2, which correspond to the normal state, the abnormal warning state, and the severe fault state, respectively; the fault feature code includes R0 to R4, which correspond to different fault types and fault severity, respectively; the remaining life prediction interval includes a safe interval, a warning interval, and a dangerous interval.

[0074] When the decision execution matrix determines the damage level to be S0 and the remaining lifetime prediction is within the safe range, the operation and maintenance actions of increasing the sampling frequency and strengthening the tracking of the remaining lifetime trend are executed.

[0075] When the decision execution matrix determines the damage level to be S1, the fault feature is encoded as a specific interval, and the remaining life is predicted as a warning interval, the operation and maintenance action of reducing the load or limiting the speed to below 80% of the rated value is executed.

[0076] When the decision execution matrix determines the damage level to be S2, a soft stop or emergency stop command is executed.

[0077] Preferably, the closed-loop feedback sample in step 6 includes a pre-maintenance status baseline sample, recommended maintenance actions, actual executed actions, inspection confirmation results, and retest results, forming a closed-loop feedback sample structure of "pre-maintenance status - maintenance actions - post-maintenance status - inspection conclusion".

[0078] Preferably, the validity determination value of the closed-loop feedback sample in step 7 The calculation formula is:

[0079]

[0080] In the formula, The recovery gain value corresponds to the magnitude of changes in multi-source features before and after maintenance, and is used to quantify the actual degree of recovery of the equipment status after the maintenance action is performed; This is the consistency value between the fault diagnosis result and the maintenance confirmation result, used to evaluate the accuracy of the diagnostic output; This is the confidence gate value corresponding to the model confidence level, used to measure the reliability of the model output;

[0081] when > When the time is right, it is determined to be a valid feedback sample. A preset validity threshold is set; valid feedback samples are used to correct fault labels, remaining life labels, and maintenance action triggering conditions, and feedback samples with conflicting results are marked as samples to be verified.

[0082] Preferably, if the closed-loop feedback sample in step 7 fails any of the manual confirmation, retest verification, or security gating judgment, the closed-loop feedback sample is marked as a sample to be verified and does not participate in the fault diagnosis model parameter update.

[0083] Preferably, the model parameter update process for weighted incremental learning in step 7 is as follows:

[0084]

[0085] In the formula, These are the model parameters before the k-th update. For the updated model parameters, For learning rate, The original training data, To effectively provide feedback on the dataset, The feedback data loss weight coefficients are used to balance the contributions of the original training data and the feedback data to the model update;

[0086] Effective feedback dataset It is formed after maintaining event write-back, and its specific expression is:

[0087]

[0088] In the formula, Let i be the unified feature sequence corresponding to the i-th feedback sample. Let i be the operating condition variable corresponding to the i-th feedback sample. The correction label is written back based on the execution results, retest status, or maintenance confirmation conclusion. To provide feedback sample weights, This represents the number of feedback samples.

[0089] Through the aforementioned maintenance event write-back and incremental learning process, the model parameters and operation and maintenance decision-making strategies are dynamically corrected.

[0090] Preferably, the present invention also discloses a multi-source heterogeneous closed-loop intelligent operation and maintenance system for ammonia, LNG, and methanol fuel ship propulsion systems. The system operates based on the method described above and includes:

[0091] The data acquisition module is used to collect multi-source heterogeneous operating signals using sensors installed on key moving components of the ship's power system. The multi-source heterogeneous operating signals include vibration signals, pressure signals, temperature signals, oil state signals, speed signals, and load signals.

[0092] The feature processing module is communicatively connected to the data acquisition module. It is used to perform time alignment, segmentation, statistical feature extraction and global normalization of the training set on the multi-source heterogeneous operating signals, construct a unified feature sequence, and obtain the corresponding fault diagnosis results, fault severity assessment results, remaining life prediction results and model confidence, and generate a pre-operation and maintenance state benchmark sample.

[0093] The global representation extraction module is communicatively connected to the feature processing module. It is used to input the unified feature sequence and the condition embedding token constructed from the operating condition variables into the attention mechanism-based encoder to extract the global health status representation of the ship's power system.

[0094] The multi-task output module is communicatively connected to the global representation extraction module and is used to simultaneously output fault type identification results, fault severity assessment results, and remaining lifetime prediction results based on the global health status representation.

[0095] The gating decision module is communicatively connected to the multi-task output module and is used to generate corresponding operation and maintenance action instructions based on the fault diagnosis results, fault severity assessment results, remaining lifetime prediction results and model confidence through the gating decision mechanism.

[0096] The feedback sample construction module is connected to the gating decision module and is used to obtain the retest data, action execution results and maintenance confirmation results after the operation and maintenance action is executed. It establishes a one-to-one correspondence with the pre-operation and maintenance state benchmark sample to form a closed-loop feedback sample.

[0097] The incremental learning module, which is communicatively connected to the feedback sample construction module and the feature processing module, is used to determine the validity of the closed-loop feedback sample based on the change magnitude of multi-source features before and after maintenance, the consistency between fault diagnosis results and maintenance confirmation results, and the model confidence level. The valid feedback sample is then used to dynamically correct the model parameters and decision-making strategy through a weighted incremental learning mechanism.

[0098] Preferably, the statistical feature vector extracted by the feature processing module for each signal segment includes the mean. Standard deviation Root mean square value skewness kurtosis Peak factor Peak-to-peak value and energy Features extracted from signals from different sources within the same time window are spliced ​​together according to sensor type to form multi-source fusion features. These features are then normalized using global statistics from the training set and arranged in chronological order to form a unified feature sequence.

[0099] Preferably, in the global representation extraction module, the unified feature sequence and the operating condition variables are respectively processed through a feature mapping matrix. and working condition mapping matrix Projected into the same latent space, the position codes are superimposed after stitching. The input representation of the Transformer encoder is obtained. Its expression is:

[0100]

[0101] In the formula, The input representation is used for the attention-based encoder (Transformer encoder). The feature mapping matrix, For feature mapping bias term, This is the working condition mapping matrix. For the operating condition mapping bias term, It is a non-linear activation function. For position encoding, This indicates a token concatenation operation.

[0102] The operating condition variables include one or more of the following: speed, load, and fuel type.

[0103] Preferably, the attention-based encoder is a Transformer encoder, which extracts the correlation between different time segments and signals from different sources through a multi-head self-attention mechanism to obtain a global health status representation of the ship's power system.

[0104] The Transformer encoder comprises L encoder layers stacked sequentially. Each encoder layer includes a multi-head self-attention sublayer and a feedforward network sublayer. Each sublayer is followed by a residual connection and a layer normalization operation.

[0105] Preferably, the multi-task output module outputs fault type identification results, fault severity assessment results, and remaining lifetime prediction results based on the global health status representation through a fault classification output layer, a severity regression output layer, and a remaining lifetime prediction output layer, respectively; the multi-task output module employs a weighted multi-task loss function. Joint training is conducted to achieve synergistic optimization of fault diagnosis, degradation assessment, and life prediction.

[0106] Preferably, the operation and maintenance decision instructions in the gating decision module Determined by the following formula:

[0107]

[0108] In the formula, The result of fault type identification. This is the result of the fault severity assessment. For the remaining life prediction results, For model confidence, This is an operation and maintenance decision function that performs gating mapping based on a combination of damage level, fault feature encoding, and remaining lifetime prediction interval.

[0109] Preferably, the validity determination value of the closed-loop feedback samples in the incremental learning module The calculation formula is:

[0110]

[0111] In the formula, The recovery gain value corresponds to the magnitude of changes in multi-source features before and after maintenance, and is used to quantify the actual degree of recovery of the equipment status after the maintenance action is performed; This is the consistency value between the fault diagnosis result and the maintenance confirmation result, used to evaluate the accuracy of the diagnostic output; This is the confidence gate value corresponding to the model confidence level, used to measure the reliability of the model output;

[0112] when > The feedback samples are determined to be valid. Valid feedback samples are used to correct fault labels, remaining life labels and maintenance action trigger conditions. Feedback samples with conflicting results are marked as samples to be verified. Closed-loop feedback samples that fail manual confirmation, retest verification or safety gating judgment do not participate in model parameter updates.

[0113] Preferably, the present invention also discloses an intelligent operation and maintenance device for a clean fuel ship power system, which includes the aforementioned system. The intelligent operation and maintenance device is installed in the ship's engine room monitoring room and is communicatively connected to vibration sensors, pressure sensors, temperature sensors, and speed sensors installed on the main engine, propulsion shaft, motor, gearbox, and bearings. It is used to receive multi-source heterogeneous operating signals collected by the sensors and perform fault diagnosis, life prediction, and operation and maintenance decision command generation and feedback optimization.

[0114] Preferably, the present invention also discloses a computer-readable storage medium having a computer program stored thereon, the computer program being executed by a processor to implement the steps of the method.

[0115] Compared with the prior art, the technical solution disclosed in this application has the following non-obvious technical features:

[0116] Technical Feature 1: A unified characterization framework and global normalization strategy for multi-source heterogeneous signals; existing methods typically analyze only single signal types such as vibration, or simply splice different signals after applying their own independent preprocessing procedures; this application creatively establishes a unified feature extraction and characterization framework for multi-source heterogeneous signals such as vibration, pressure, temperature, oil, speed, and load: First, each type of signal is segmented according to a unified time window, and eight-dimensional statistical features such as mean, standard deviation, root mean square, skewness, kurtosis, crest factor, peak-to-peak value, and energy are extracted to form a feature vector. Then, feature vectors from different sources are spliced ​​along the feature dimensions to form a multi-source... The fusion features are then normalized using the global mean and standard deviation of the training set to form a unified serialized feature representation. The non-obviousness of this approach lies in the fact that traditional normalization methods are usually performed independently for each channel of a single signal type, while this application uses global statistics from the training set for normalization across signal sources. The underlying intention is to preserve the relative proportional relationship of physical amplitudes between different signal sources. This proportional relationship essentially reflects the difference in sensitivity of different monitoring parameters to the same fault event, and is an important physical clue for assessing the severity of the fault, rather than the "difference in dimensions" that needs to be eliminated in the traditional understanding.

[0117] Technical Feature Two: This technique explicitly encodes operating condition variables into learnable tokens and deeply integrates them with feature sequences. Existing methods for handling operating condition changes typically employ segmented modeling, hard-coding of operating condition labels, or simple concatenation with feature vectors. This application creatively projects continuous operating condition variables such as speed and load, as well as discrete operating condition variables such as fuel type, onto the same latent space as the feature sequence through independent learnable mapping matrices, forming operating condition embedding tokens. These tokens are then concatenated with multi-source feature tokens along the sequence dimension and input into the Transformer encoder. The profound innovation of this technique lies in the fact that operating condition tokens do not participate in model computation as static "additional information." Instead, through the Transformer's multi-head self-attention mechanism, they dynamically interact with other feature tokens across multiple attention heads, enabling the model to adaptively adjust the attention weights for various signal features under different operating conditions. This design transforms "operating condition" from a simple classification label into an adjustable variable that dynamically participates in feature reweighting during model inference, fundamentally changing the model's paradigm for handling operating condition changes.

[0118] Technical Feature 3: A multi-task collaborative output architecture that simultaneously outputs fault type, fault severity, and remaining lifetime prediction based on the same global health status representation; In existing technologies, fault classification, degradation assessment, and lifetime prediction are usually designed as three independent models or pipeline modules, each using an independent feature extraction branch; This application creatively adopts the global health status representation output by a single Transformer encoder as a shared foundation, and outputs fault type identification results, fault severity assessment results, and remaining lifetime prediction results through three independent task output heads, and uses a weighted multi-task loss function to achieve joint optimization of the three tasks during training; The non-obviousness of this architecture lies in the fact that the three tasks achieve mutual information gain through a shared representation layer—the fault classification task helps the encoder learn the feature boundaries that distinguish different fault modes, the severity assessment task forces the encoder to pay attention to the continuous change information in the feature amplitude, and the lifetime prediction task guides the encoder to capture the temporal degradation trend—the representation quality generated by the collaborative optimization of the three is significantly better than any single task-driven representation learning.

[0119] Technical Feature 4: A four-dimensional gated operation and maintenance decision-making mechanism combining fault type, severity, remaining lifetime, and model confidence. Existing methods typically only consider fault type or a single threshold criterion when generating operation and maintenance recommendations. This application establishes a gated decision matrix that includes four dimensions: damage level (S0 / S1 / S2), fault feature encoding (R0-R4), remaining lifetime prediction interval (safe / early warning / dangerous), and model confidence. Based on the combination of these four types of information, it automatically matches the corresponding operation and maintenance actions and outputs graded instructions ranging from "increase sampling frequency," "replenish lubrication," and "reduce load and limit speed" to "soft shutdown" and "emergency stop." The non-obviousness of this decision-making mechanism lies in the fact that model confidence, as one dimension of the decision gate, allows the system to automatically reduce the level of automatic operation and maintenance instructions or switch to manual review mode when the model lacks sufficient confidence in the current diagnostic results. This embeds safety redundancy logic while pursuing automated operation and maintenance, a safety design generally lacking in existing technologies.

[0120] Technical Feature 5: Closed-Loop Feedback and Weighted Incremental Learning Mechanism Based on Validity Judgment; In existing systems, maintenance feedback data is often simply recorded or completely ignored; This application creatively constructs a closed-loop feedback sample structure of "pre-maintenance state baseline sample - maintenance action - post-maintenance retest data - maintenance confirmation conclusion," and designs a validity judgment function containing three indicators: recovery gain value, diagnostic consistency, and confidence gate value. Only feedback samples that are judged to be valid are assigned non-zero weights to participate in incremental learning; samples with conflicting results are marked as pending verification and do not participate in model updates; At the same time, independent loss weight coefficients are applied to the feedback data during incremental learning, enabling the model to continuously absorb new knowledge from the feedback while maintaining its ability to fit the original training data; The non-obviousness of this mechanism lies in the fact that the three indicators of the validity judgment function evaluate the quality of feedback samples from three orthogonal dimensions: physical response, semantic consistency, and model reliability, effectively avoiding the damage of erroneous or noisy feedback to the model, and solving the engineering problem that existing incremental learning methods are difficult to implement effectively in actual ship maintenance scenarios due to the inconsistent quality of feedback data;

[0121] Technical Feature Six: The method proposed in this application has universality across fuel types, and its technical framework does not depend on specific fuel physicochemical properties. Existing fault diagnosis methods for ammonia, LNG, or methanol power systems are usually developed separately for specific fuel types, and their signal processing and feature extraction strategies often implicitly assume prior assumptions about the combustion excitation characteristics of specific fuels. The technical framework of this application is based entirely on the statistical characteristics of general physical quantities such as vibration, pressure, and temperature at the signal feature extraction level, and encodes the fuel type as a discrete operating condition variable at the operating condition token construction level, so that the entire method can be seamlessly transferred to ammonia, LNG, methanol, and even traditional diesel power systems. This universal design across fuel types is an extremely rare systematic breakthrough in existing technologies.

[0122] Compared with the prior art, the present invention has the following beneficial effects:

[0123] 1. Significantly improves the comprehensive utilization efficiency and fault information completeness of multi-source heterogeneous signals; This invention establishes a unified framework for feature extraction and fusion of multi-source heterogeneous signals, converting operating signals from different sources, with different sampling frequencies and different physical dimensions, such as vibration, pressure, temperature, oil, speed, and load, into a unified serialized feature representation, enabling the collaborative analysis of various monitoring data that were originally scattered and fragmented under a unified framework; Experimental verification shows that compared with a single signal diagnosis scheme that only uses vibration signals, the multi-source fusion scheme of this invention improves the accuracy of bearing fault diagnosis by about 12% to 18%, and the improvement is more significant in complex fault scenarios, fully demonstrating the enhancing effect of multi-source heterogeneous signal fusion on the completeness of fault information;

[0124] 2. Significantly improves the reliability and generalization ability of fault diagnosis under complex multi-operating conditions; This invention explicitly encodes operating condition variables into learnable operating condition embedding tokens and deeply integrates them with multi-source feature sequences in the Transformer encoder through a multi-head self-attention mechanism, enabling the model to adaptively adjust feature attention weights under different speeds, loads, and fuel types; Verified in cross-operating condition tests simulating different speeds (50%, 75%, 100% of rated speed) and different loads (25%, 50%, 75%, 100%), the diagnostic accuracy of this invention is controlled within 5% under cross-operating conditions, while the accuracy of traditional methods is as high as 20% to 35%, proving the superior generalization ability of this invention under complex operating conditions;

[0125] 3. Achieve end-to-end collaborative output of fault diagnosis, degradation assessment, and remaining life prediction, directly supporting operation and maintenance decisions; This invention, through a multi-task joint learning architecture, synchronously outputs three results—fault type, fault severity, and remaining life prediction—based on the same global health status representation. The information complementarity and collaborative optimization among the three outputs make the accuracy of each output superior to independent modeling schemes; In the remaining life prediction task, the prediction error (MAPE) of this invention is reduced by approximately 8% to 10% compared to a single-task model; More importantly, the synchronous output of the three results enables ship operation and management personnel to obtain a complete status assessment information chain from "what fault" to "how severe" to "how long can it still be used" on the same interface and at the same time, greatly shortening the time window from status awareness to operation and maintenance decision-making;

[0126] 4. Establish a closed-loop intelligent operation and maintenance system of "monitoring-diagnosis-prediction-decision-feedback-optimization" to achieve system self-evolution. This invention continuously feeds back the real fault information and retest results obtained from each maintenance and repair to the diagnostic model through maintenance event write-back and weighted incremental learning mechanisms, enabling the model to continuously accumulate experience and correct deviations in practical applications. Experimental results show that after 5 rounds of incremental learning updates, the average diagnostic confidence of the model on the target ship increased from the initial 0.78 to 0.91, and the false alarm rate decreased by about 45%, proving the significant improvement effect of the closed-loop feedback mechanism on the long-term operational reliability of the system. This invention provides a technically deployable basis for the transformation of the ship power system operation and maintenance mode from "periodic maintenance + post-event maintenance" to "predictive maintenance + condition-based maintenance".

[0127] 5. It possesses universality for power systems powered by various clean fuels such as ammonia, LNG, and methanol. At the technical architecture level, this invention does not rely on the combustion characteristics and physicochemical parameters of specific fuels. Its signal feature extraction is based on the statistical characteristics of general physical quantities, and its operating condition coding mechanism can accommodate any fuel type as discrete operating condition input. Therefore, it can be seamlessly adapted to various clean fuel marine power systems such as ammonia fuel engines, LNG dual-fuel main engines, and methanol fuel cell power systems. This universal design significantly reduces the secondary development costs of migrating and deploying technical solutions between different ship types and different fuel types, and has good prospects for industrialization and promotion. Attached Figure Description

[0128] Figure 1 This is a flowchart of the method described in this invention;

[0129] Figure 2 This is a flowchart of the multi-source heterogeneous feature extraction and working condition token construction described in this invention;

[0130] Figure 3 This is a schematic diagram of the structure of the Transformer multi-task fault diagnosis model described in this invention;

[0131] Figure 4 This is a flowchart of the closed-loop operation and maintenance action gating and maintenance event write-back process described in this invention;

[0132] Figure 5 This is a diagram showing the fault diagnosis results of the ship ammonia fuel power system under different fault conditions as described in this invention;

[0133] Figure 6 This is a comparison chart of the feedback effects before and after maintenance event write-back and incremental learning as described in this invention;

[0134] Figure 7 This is a comparison chart of the model confidence distribution curves before and after maintenance event write-back and incremental learning as described in this invention;

[0135] Figure 8 This is a schematic diagram of the system described in this invention;

[0136] Figure 9 It is the decision execution matrix pre-stored inside the gating decision module of the present invention. Detailed Implementation

[0137] The following is in conjunction with the appendix Figure 1 To be continued Figure 9 The main technical features and embodiments of the present invention are described in detail below;

[0138] In the technical solution of this application, the unified time alignment and segmentation processing mechanism for multi-source heterogeneous operating signals involves aligning and segmenting various types of signals collected by vibration sensors, pressure sensors, temperature sensors, oil sensors, speed sensors, and load sensors installed on different key components of the ship's power system according to a unified time reference and a preset time window length. This ensures that signals with different sampling frequencies and start times have aligned time segments within the same time window, thus providing a time-consistent basic data unit for subsequent cross-signal source feature fusion. This feature is the data entry foundation of the entire technical solution. If the signals are not precisely aligned in time, the subsequently extracted multi-source features will reflect the equipment status in different time intervals, and the global representation generated after fusion will be inconsistent in temporal semantics, which will seriously impair the accuracy of fault diagnosis. This feature is the primary prerequisite for ensuring the effective fusion of multi-source heterogeneous signals and is also a key basic design that distinguishes it from the existing method of "independent processing of each signal".

[0139] In the technical solution of this application, the unified signal characterization method based on multidimensional statistical features extracts eight statistical features—mean, standard deviation, root mean square value, skewness, kurtosis, crest factor, peak-to-peak value, and energy—from the data segment of each sensor signal within each time window, forming a feature vector of the signal source within that window. In this way, the original waveform data is converted into a standardized feature vector composed of eight statistical quantities, allowing for different physical dimensions (such as the unit of vibration acceleration m / s², the unit of pressure MPa, and the unit of temperature ℃) and different amplitude ranges (such as vibration amplitude typically ranging from 0.1 to 10). The original signals (on the order of m / s², with temperature amplitudes ranging from 20 to 600°C) all participate in subsequent fusion in the form of a unified eight-dimensional feature vector. This feature is the core means to solve the problem of difficulty in unified modeling of multi-source heterogeneous data with different dimensions, amplitudes, and distributions. By uniformly extracting statistical features instead of retaining the original waveforms, heterogeneous signals are abstracted to the same semantic level (statistical domain), laying the "homogenization" foundation for subsequent multi-source feature splicing. Its importance lies in the fact that if existing technologies are used to design feature extraction strategies independently for each signal, the feature dimensions and feature types of different signals will not be uniform, making effective cross-signal source fusion impossible.

[0140] In the technical solution of this application, the cross-signal source normalization strategy based on global statistics of the training set does not adopt the conventional approach of independent normalization for each channel after the multi-source features are spliced ​​together to form multi-source fusion features. Instead, it treats all fusion features of all training samples as a whole distribution, calculates the global mean and global standard deviation, and uses these two global statistics to normalize all fusion features. As a result, the amplitude ratio between different signal source features is completely preserved after normalization, which is reflected as the relative magnitude difference of the normalized values. This feature is one of the most core non-obvious features of this application. Conventional normalization methods perform independent normalization for each channel, which would erase the key physical information of the difference in sensitivity of different physical quantities to the same fault event. However, this application, through the global normalization strategy, preserves the proportional information reflecting the physical essence in the model input, which is that "the magnitude of change in vibration features is much greater than the magnitude of change in temperature features". This allows the fault severity assessment task to directly obtain strong relevant information from the relative amplitude of the input features. The importance of this feature is that it is a cognitive bridge connecting physical sensing information and machine learning models, and directly determines the upper limit of the accuracy of the fault severity assessment task.

[0141] In the technical solution of this application, the explicit learnable token encoding mechanism for operating condition variables projects operating condition information such as rotational speed (continuous variable), load (continuous variable), and fuel type (discrete variable, such as ammonia / LNG / methanol) onto the same latent space dimension as the multi-source features through an independent learnable mapping matrix, forming an operating condition embedding token. This token, as an independent element in the sequence, is parallel to all feature tokens in the sequence dimension, together constituting the complete input sequence of the encoder. This feature is the core technical means to achieve adaptive operation under complex operating conditions. Traditional methods treat operating conditions as static labels or simply concatenated auxiliary features, and the model cannot dynamically adjust the way it interprets signals using operating condition information. However, this application encodes operating condition variables into learnable tokens, enabling them to dynamically interact with all other feature tokens in the Transformer self-attention calculation, thereby adaptively changing the attention weights to different signal features under different operating conditions. This feature directly solves the cross-operating condition generalization problem of "the same equipment is normal under operating condition A but is misjudged as faulty under operating condition B".

[0142] In the technical solution of this application, the global correlation feature extraction architecture based on the Transformer multi-head self-attention mechanism uses the Transformer encoder as the core feature extractor. Through the multi-head self-attention mechanism, a direct dependency relationship between any two tokens in the input sequence is established in the same layer of computation. In the operation and maintenance scenario of marine power system, this means that the token from the first time window of the vibration signal and the token from the fifth time window of the temperature signal can directly interact with each other through attention, without the need for the stepwise propagation of RNN or the stacking of local receptive fields of CNN. This feature is the core architectural choice that distinguishes this application from the traditional CNN / LSTM architecture. The fault evolution of marine power system has the coupling characteristics of cross-component and cross-signal source. For example, the impact vibration caused by bearing failure may cause changes in shaft torsional vibration, which in turn leads to motor current fluctuations and abnormal temperature rises. These cross-signal and cross-time correlations are crucial for accurately judging complex faults. The Transformer's self-attention mechanism can capture these long-range cross-source dependencies with extremely low path length, while traditional RNN / CNN requires multiple layers to approximately achieve the same receptive field and is prone to gradient vanishing. The importance of this feature is that it is the architectural guarantee for achieving the completeness of global health state representation information.

[0143] In the technical solution of this application, the multi-task collaborative output architecture based on shared global representation utilizes the representation corresponding to the global token output by the Transformer encoder. As the sole shared health status representation, it is simultaneously input to three independent output heads—a fault classification head (softmax multi-classification), a severity regression head (linear regression), and a remaining life prediction head (linear regression with ReLU constraints). These three output heads share the same encoder parameters during training and are jointly optimized using a weighted multi-task loss function. This feature is a concrete implementation of the "diagnosis-assessment-prediction synergy" objective at the model architecture level. Conventional multi-task models typically design feature extraction branches independently for each task, leading to insufficient information sharing between tasks. This application forces the three tasks to share the same representation space, compelling the encoder to learn representations that simultaneously benefit classification, regression, and temporal prediction tasks—effectively achieving "representation regularization" and preventing the encoder from excessively favoring a single task while ignoring information needed by other tasks. The importance of this feature lies in its ability to create mutually beneficial positive feedback among the three outputs at the information level, rather than a piecemeal output, fundamentally solving the decision conflict problem of "contradictory diagnostic and prediction results" in traditional solutions.

[0144] In the technical solution of this application, the four-dimensional gating decision matrix and hierarchical operation and maintenance instruction generation mechanism establish a gating decision matrix with four input dimensions: damage level (S0 normal / S1 warning / S2 severe), fault feature encoding (R0-R4, corresponding to different fault type combinations), remaining life prediction interval (safe / warning / dangerous), and model confidence. Each set of four-dimensional input combinations uniquely maps to a kind of operation and maintenance suggestion, including but not limited to "increase sampling frequency," "replenish lubrication," "reduce load," "limit speed," "generate maintenance work order," "soft shutdown," and "emergency stop." The output level progresses step by step from information prompts to safety protection. This feature connects "model output" with "actual operation and maintenance actions." The key transition hub determines whether this system can truly leap from state awareness to automated operation and maintenance. In existing technologies, there is a "manual judgment" link between diagnostic output and maintenance actions, while this application automates this step through a predefined decision execution matrix. The importance of this feature is particularly reflected in the introduction of the confidence dimension—when the model confidence is lower than a preset threshold, the decision matrix will force the operation and maintenance recommendations to be downgraded to a "manual review" mode or only output information prompts, prohibiting the automatic execution of high-risk shutdown actions, thus embedding safety redundancy logic in the system design. This definition is based on a mechanism of decision output based on multi-dimensional combination rather than a single threshold, making the system interpretable, traceable, and engineering safe.

[0145] In the technical solution of this application, the structured construction and association identification mechanism of closed-loop feedback samples uses the execution time and action type of maintenance actions as association identifiers. It associates and binds five parts of information: the pre-maintenance state baseline sample (including multi-source features, diagnostic results, life prediction, and model confidence), the system-recommended maintenance actions, the actually executed maintenance actions, the retest data after maintenance execution, and the fault conclusion confirmed by inspection and disassembly. This forms a closed-loop feedback sample with a complete causal chain: "pre-maintenance state - maintenance action - post-maintenance state - maintenance conclusion." This feature is the information carrier foundation of the closed-loop intelligent maintenance system. Without structured feedback samples, incremental learning will lack supervised target values, and closed-loop optimization is impossible. The importance of this feature lies in ensuring that each piece of "posterior knowledge" (i.e., the fault type and actual damage degree confirmed by maintenance) in the feedback sample is accurately associated with its corresponding "prior output" (i.e., the model's diagnosis and prediction before maintenance). This provides high-quality supervision signals for incremental learning and avoids engineering risks such as "using the feedback of fault B to correct the diagnostic model of fault A" due to misalignment.

[0146] In the technical solution of this application, the independent control mechanism of feedback data loss weight in weighted incremental learning involves setting an independent feedback data loss weight coefficient in the model parameter update formula of incremental learning. This coefficient is independent of the learning rate of the original training data and is used to adjust the update magnitude of the model parameters by the feedback samples, preventing the gradient of the feedback samples (usually much smaller than the original training data) from causing excessive perturbation to the model parameters. This feature is a key hyperparameter design to ensure that the model maintains a "stability-plasticity balance" during continuous learning. In conventional incremental learning, new data and old data usually use the same loss weight, which can easily lead to "catastrophic forgetting"—the model overfits on new feedback samples and loses its generalization ability to the original wide range of working conditions. This application explicitly adjusts the gradient contribution of the feedback data through an independent coefficient, enabling the model to retain the memory of existing knowledge while absorbing new knowledge. The importance of this feature lies in the fact that it is the key to engineering from laboratory verification to long-term shipboard deployment, determining whether the system can maintain diagnostic reliability comparable to that at the time of initial deployment after several years of actual operation.

[0147] In the technical solution of this application, Transformer is a multi-head self-attention-based transformer, MLP is a multilayer perceptron, t-SNE is a t-distributed random neighborhood embedding, and Z-Score is zero-mean normalization.

[0148] refer to Figure 1 This document illustrates a flowchart of a multi-source heterogeneous closed-loop intelligent operation and maintenance method for ammonia, LNG, and methanol fuel-powered marine propulsion systems according to one or more embodiments of this application. This method is applied to a ship's engine room intelligent monitoring system, utilizing multiple types of sensors installed on key moving components of the clean fuel marine propulsion system to achieve automated closed-loop processing of equipment status perception, fault diagnosis, lifespan prediction, and operation and maintenance decision-making. It includes: Step 1, multi-source heterogeneous operating signal acquisition; Step 2, unified feature sequence construction and pre-operation and maintenance state benchmark sample generation; Step 3, global health status representation extraction; Step 4, multi-task collaborative output; Step 5, gating decision and operation and maintenance instruction generation; Step 6, closed-loop feedback sample construction; and Step 7, validity determination and weighted incremental learning.

[0149] In one or more embodiments, step 1 configures the ship's engine room data acquisition terminal to communicate with a sensor group installed on key moving components of the ship's power system via an industrial Ethernet or CAN bus, acquiring multi-source heterogeneous operating signals at a preset sampling frequency. Specifically, the key moving components include one or more of the following: main engine crankshaft and piston assembly, propulsion shaft and intermediate bearing, motor and coupling, gearbox transmission mechanism, and compressor and pump unit. The sensor group includes: a piezoelectric accelerometer installed on the bearing housing for acquiring vibration signals, with a sampling frequency set to 25.6 kHz; a key phase sensor installed at the end of the shaft system for acquiring speed signals; a torque meter installed at the coupling for acquiring load signals; a pressure transmitter installed in the fuel supply pipeline and cooling water system for acquiring pressure signals; a K-type thermocouple installed on the bearing pads and cylinder head for acquiring temperature signals; and an online oil particle counter and viscosity sensor installed on the return oil pipeline for acquiring oil state signals. The above-mentioned signals have different sampling frequencies, different physical dimensions, and different amplitude ranges, constituting a typical set of multi-source heterogeneous operating signals.

[0150] In one or more embodiments, step 2 configures the feature processing module to perform time alignment, segmentation processing, and statistical feature extraction on the acquired multi-source heterogeneous operating signals, normalize them using global statistics of the training set, construct a unified feature sequence, and obtain the corresponding fault diagnosis results, fault severity assessment results, remaining lifetime prediction results, and model confidence scores to generate a pre-maintenance state benchmark sample.

[0151] Specifically, step 2 includes the following sub-steps 2A to SD:

[0152] Step 2A – Signal Segmentation: Let the first segment be... The raw operating signals acquired by the sensor are According to the preset time window length Segmentation (as one implementation method) The value is set to 1024 sampling points, corresponding to a time length of approximately 40ms. This window length is sufficient to capture the characteristic frequency components of bearing faults and meet real-time processing requirements. One signal segment:

[0153] ;

[0154] Step S=2B — Multidimensional statistical feature extraction: Extract an eight-dimensional statistical feature vector for each signal segment. The statistical feature vector includes the mean. Standard deviation Root mean square value skewness kurtosis Peak factor Peak-to-peak value and energy Taking vibration signals as an example, when a bearing experiences an inner ring failure, the impact component in the vibration signal causes a significant increase in peak-to-peak value and kurtosis, while the response of the temperature signal is relatively smooth—this difference is fully preserved through the relative amplitude changes of the eight-dimensional features. The statistical feature vector... The expression is:

[0155] ;

[0156] Step 2C – Multi-source feature stitching: Features extracted from signals from different sources within the same time window are stitched together according to the number of sensor signal sources. By splicing, we obtain the first... Multi-source fusion features of each window :

[0157]

[0158] As a preferred embodiment of this step, The value is 6, which corresponds to six types of signal sources: vibration, rotational speed, load, pressure, temperature and oil. The multi-source fusion feature dimension after splicing is 48 dimensions (6 types of signals × 8-dimensional statistics for each type).

[0159] Step 2D – Global Normalization and Serialization: During the offline training phase, the global mean of the training set is pre-calculated and stored. and global standard deviation During the actual inference phase, the feature processing module calls the aforementioned pre-stored global statistics to normalize all multi-source fused features:

[0160]

[0161] In the formula, It is a very small constant (with a value of 10). -8 This global normalization strategy is used to prevent division by zero errors. Its deeper advantage lies in preserving the physical amplitude ratios between different signal source features. For example, the standard deviation of vibration signal features may be several orders of magnitude larger than that of temperature signal features. This difference, after being normalized using the same set of global statistics, is preserved as a relative numerical difference in different dimensions of the feature vector. This allows the model to directly perceive the differences in sensitivity of different physical quantities to the same fault event from the input features, providing crucial physical clues for subsequent fault severity assessment.

[0162] Subsequently, the feature processing module arranges the normalized fused features of all windows in chronological order to form a unified feature sequence. While forming a unified feature sequence, the feature processing module calls the current model to process the unified feature sequence. Inference is performed to obtain the corresponding fault diagnosis results, fault severity assessment results, remaining life prediction results, and model confidence scores. The above five pieces of information (unified feature sequence, fault diagnosis results, severity assessment results, remaining life prediction results, and model confidence) together constitute the pre-operation and maintenance state benchmark sample.

[0163] In one or more embodiments, step 3 configures the global representation extraction module (containing a Transformer encoder) to extract the unified feature sequence generated in step 2. Operating condition variables The constructed operating condition embedding tokens are used as common inputs to an encoder based on an attention mechanism to extract a global health status representation of the ship's power system.

[0164] Specifically, the operating condition variables will first be... Explicit encoding is performed. The operating condition variables include: main engine speed (a continuous variable, in r / min, for example, in actual ship conditions it may be 50%, 75%, or 100% of the rated speed), load factor (a continuous variable, ranging from 0 to 100%), and fuel type (a discrete variable, including "ammonia", "LNG", and "methanol"). The operating condition encoding module concatenates these variables into an operating condition vector. Subsequently, a unified feature sequence was established. Through the feature mapping matrix Projected into the latent space, the working condition vector Through the working condition mapping matrix Projected into the same latent space, the two are stitched together and their positions are superimposed with an encoding. The input representation of the Transformer encoder is obtained. :

[0165]

[0166] In the formula, The input representation is used for the attention-based encoder (Transformer encoder). The feature mapping matrix, For feature mapping bias term, This is the working condition mapping matrix. For the operating condition mapping bias term, It is a non-linear activation function. For position encoding, This indicates a token concatenation operation.

[0167] Through the above design, the working condition token is placed alongside all feature tokens as an independent element in the sequence, enabling the working condition information to dynamically interact with other feature tokens in the subsequent Transformer self-attention calculation.

[0168] Then, Input Transformer encoder, the Transformer encoder is composed of It consists of a series of stacked encoder layers (as one implementation). (Value is 6). Each encoder layer includes a multi-head self-attention sub-layer and a feedforward network sub-layer. Each sub-layer is followed by a residual connection and a layer normalization operation. The calculation process of multi-head self-attention is as follows:

[0169] For input representation , No. A query matrix for each attention head Key matrix Sum matrix By respectively through A linear mapping is performed to obtain (as one implementation method, the number of attention heads) The value is 8, representing the dimension of the key vector. (Value 64)

[0170] , ,

[0171] The output of a single attention head is:

[0172]

[0173] The outputs of all attention heads are concatenated and linearly mapped to serve as the output of the multi-head self-attention layer:

[0174]

[0175] The global representation of the encoder output is denoted as: .

[0176] The output corresponding to the global token is selected as the global health status representation of the ship's propulsion system. The dimension of this global representation is (As one implementation method, With a value of 512, it integrates information from multiple sources within different time windows, as well as operating condition information, and serves as the unified feature basis for the subsequent multi-task output of the entire system.

[0177] In one or more embodiments, step 4 configures the multi-task output module to be based on the global health status representation obtained in step S3. It also outputs fault type identification results, fault severity assessment results, and remaining life prediction results.

[0178] Specifically, the multi-task output module includes three independent output heads: a fault classification output layer, a severity regression output layer, and a remaining lifetime prediction output layer. Global health status characterization. The inputs are simultaneously fed into the three output layers mentioned above. The output expressions for the three layers are as follows:

[0179]

[0180]

[0181]

[0182] In the formula, The fault type identification results are output as a probability distribution vector containing categories such as "normal state", "bearing inner ring fault", "bearing outer ring fault", "rolling element fault" and "compound fault", with the category with the highest probability being used as the final diagnostic conclusion. The output is a continuous value between 0 and 1, representing the degree of damage of the current fault (0 represents no damage, and 1 represents complete failure). Output the remaining life prediction results in hours; The remaining lifetime prediction output function is (preferably, the ReLU activation function is used to ensure that the output is non-negative). , , These are the weight matrices for the corresponding tasks. , , These are the bias terms for the corresponding tasks.

[0183] During the offline training phase, the multi-task output module is jointly trained using a weighted multi-task loss function:

[0184]

[0185] In the formula, Loss is assigned to fault type identification (using the cross-entropy loss function). To assess the severity of the fault (using the mean squared error loss function). Predict the loss for remaining lifetime (using a smoothed L1 loss function). , , These are the weight coefficients for the corresponding tasks (in one implementation, the values ​​are 0.4, 0.3, and 0.3, respectively). This joint training mechanism allows the three tasks to share the same representation space, forcing the encoder to learn a global health state representation that is beneficial to all three tasks simultaneously: classification, regression, and temporal prediction. This essentially achieves the effect of "representation regularization," preventing the encoder from becoming overly biased towards a single task. For example, the fault classification task helps the encoder learn the feature boundaries that distinguish different fault modes, the severity assessment task forces the encoder to pay attention to continuous changes in feature amplitudes, and the lifetime prediction task guides the encoder to capture temporal degradation trends. The representation quality produced by the collaborative optimization of these three tasks is superior to any single task-driven representation learning.

[0186] In one or more embodiments, step 5 configures the gating decision module to respond to the fault diagnosis results output in step S4. Fault severity assessment results Remaining life prediction results and model confidence The corresponding operation and maintenance action instructions are generated through the gating decision matrix.

[0187] Operation and maintenance decision instruction 'a' is determined by the following formula:

[0188]

[0189] In the formula, This is an operation and maintenance decision function that performs gating mapping based on a combination of damage level, fault feature encoding, and remaining lifetime prediction interval.

[0190] Specifically, the gating decision module first assesses the results based on the severity of the fault. Determine the damage level: when When <0.2, it is judged as S0 level (normal state); when 0.2≤ When <0.6, it is judged as S1 level (abnormal warning state); when A value ≥0.6 is classified as S2 (severe fault state). Simultaneously, the gating decision module identifies the fault type based on the fault type recognition result. Fault feature codes are determined, with values ​​ranging from R0 to R4, where R0 represents no fault, R1 to R3 correspond to different fault types, and R4 represents a compound fault. Furthermore, the gating decision module uses the remaining lifetime prediction results... Divide the prediction interval: when >500 hours is considered a safe range, when 100 < ≤500 hours is defined as the warning range, when A time limit of 100 hours or less is considered a danger zone.

[0191] In one or more embodiments, the gating decision module pre-stores a decision execution matrix, such as Figure 9 As shown: The gating decision module matches the corresponding operation and maintenance suggestions in the decision execution matrix based on the above four-dimensional input. Among them, the model confidence level... As an independent dimension of security gating, it participates in decision-making—when When the value falls below a preset threshold (in one implementation, the threshold value is 0.7), regardless of the matrix mapping result, the decision module will downgrade the maintenance recommendation to "manual review" mode or only output an information prompt, prohibiting the automatic execution of high-risk shutdown actions. This embeds safety redundancy logic while pursuing automated maintenance. The final output maintenance action instructions include one or more of the following: increasing the sampling frequency, strengthening the tracking of remaining life trends, supplementing lubrication, reducing load, limiting speed, generating maintenance work orders, soft shutdown, and emergency stop. The instructions are sent to the corresponding execution unit (such as frequency converter, programmable logic controller, automatic oiler, etc.) via Modbus / TCP protocol or hard-wired I / O.

[0192] In one or more embodiments, step S6 configures the feedback sample construction module to use the execution time and action type of the operation and maintenance action as the association identifier, obtain the retest data, action execution results and maintenance confirmation results after the operation and maintenance action is executed, and establish a one-to-one correspondence with the pre-operation and maintenance state benchmark sample to form a closed-loop feedback sample.

[0193] Specifically, after the maintenance instructions generated in step 5 are executed (for example, the automatic lubrication machine completes the supplementary lubrication operation, or the maintenance personnel complete the bearing replacement and shaft dynamic balance calibration), the system re-executes the data acquisition and inference process of steps 1 to 4 after a preset time window (in one implementation, the retest time window is set to 30 minutes to 2 hours after the maintenance action is completed) to obtain the post-maintenance status data. At the same time, the system obtains the disassembly and inspection confirmation conclusions entered by the maintenance personnel through the human-machine interface—including information such as the actual fault type, the actual degree of damage, and the remaining physical life of the replaced parts. The feedback sample construction module associates and binds the above-mentioned pre-maintenance status benchmark sample (including unified feature sequence, diagnostic results, life prediction, and model confidence), the maintenance actions recommended by the system, the actual maintenance actions performed, the retest status after maintenance execution, and the fault conclusions confirmed by disassembly and inspection, forming a closed-loop feedback sample with a complete causal chain. Its structure is "pre-maintenance status - maintenance action - post-maintenance status - maintenance conclusion".

[0194] In one or more embodiments, step 7 configures the incremental learning module to determine the validity of the closed-loop feedback samples based on the change magnitude of multi-source features before and after maintenance, the consistency between fault diagnosis results and maintenance confirmation results, and the model confidence level, and to dynamically correct the model parameters and decision-making strategies using a weighted incremental learning mechanism for the valid feedback samples.

[0195] Specifically, for the first For each closed-loop feedback sample, the incremental learning module calculates the validity judgment value. :

[0196]

[0197] In the formula, The recovery gain value corresponds to the magnitude of changes in multi-source features before and after maintenance. It is calculated by comparing the unified feature sequence before maintenance. Feature sequences after maintenance and retesting The rate of change of the Euclidean distance between them; the greater the decrease in distance, the higher the recovery gain. This represents the consistency value between the fault diagnosis results and the maintenance confirmation results. When the fault type diagnosed by the model matches the fault type confirmed by the maintenance... ,otherwise =0; The confidence threshold value corresponding to the model confidence level represents the confidence level of the model when outputting pre-operational diagnostic results. ≥0.7 =1, otherwise =0.

[0198] when > Time (as one implementation method, A value of 0.5 is considered a valid feedback sample; when ≤ If a sample is deemed invalid or has conflicting feedback, it is marked as a sample to be verified and will not be included in the model parameter update.

[0199] After the valid feedback samples are relabeled, a valid feedback dataset is formed.

[0200]

[0201] In the formula, For the first The unified feature sequence corresponding to each feedback sample For the corresponding operating condition variables, The correction label (including the corrected fault type, severity, and remaining life) is written back based on the inspection confirmation conclusion. For feedback sample weights (as one implementation method, confidence level) The higher, (The larger the value) This represents the total number of feedback samples.

[0202] The incremental learning module is based on the original training data. and effective feedback dataset The model parameters are updated, and the parameter update process is as follows:

[0203]

[0204] In the formula, For the first Model parameters before the last update For the updated model parameters, The learning rate (as one implementation method) The value is 1×10 -5 ), As one implementation method, the feedback data loss weighting coefficient is used. The value ranges from 0.1 to 0.3 and is dynamically adjusted based on the quality of the feedback samples. This independent weighting coefficient is used to balance the contributions of the original training data and the feedback data to the model update. This prevents the gradient of the feedback samples (usually much smaller than the original training data) from causing excessive perturbation to the model parameters, effectively avoiding the "catastrophic forgetting" problem.

[0205] refer to Figure 8 This paper illustrates a schematic diagram of a multi-source heterogeneous closed-loop intelligent operation and maintenance system for ammonia, LNG, and methanol fueled marine propulsion systems according to one or more embodiments of this application. The system is deployed within a ship's engine room intelligent monitoring platform and works collaboratively with the physical equipment layer, sensor layer, and actuator layer. It includes the following modules: a data acquisition module, a feature processing module, a global representation extraction module, a multi-task output module, a gating decision module, a feedback sample construction module, and an incremental learning module. These modules are sequentially connected in communication to form a data-driven closed-loop intelligent operation and maintenance link.

[0206] In one or more embodiments, the data acquisition module is configured to acquire multi-source heterogeneous operating signals using a sensor array installed on key moving components of the ship's propulsion system. Specifically, the module includes: a first acquisition subunit communicatively connected to a piezoelectric accelerometer for acquiring vibration signals at a sampling frequency of 25.6 kHz; a second acquisition subunit communicatively connected to a key phase sensor for acquiring rotational speed signals; a third acquisition subunit communicatively connected to a torque meter for acquiring load signals; a fourth acquisition subunit communicatively connected to a pressure transmitter for acquiring pressure signals from the fuel supply line and cooling water system; a fifth acquisition subunit communicatively connected to a K-type thermocouple for acquiring temperature signals from bearing pads and cylinder heads; and a sixth acquisition subunit communicatively connected to an online oil particle counter and viscosity sensor for acquiring oil state signals. The signals output by the above six acquisition subunits are synchronized with a GPS clock as the reference in terms of timestamps to ensure the temporal consistency of subsequent feature extraction.

[0207] In one or more embodiments, the feature processing module is communicatively connected to the data acquisition module and configured to perform time alignment, segmentation processing, statistical feature extraction, and global normalization of the training set on the multi-source heterogeneous operating signals output by the data acquisition module, construct a unified feature sequence, and obtain the corresponding fault diagnosis results, fault severity assessment results, remaining lifetime prediction results, and model confidence, thereby generating a pre-maintenance state benchmark sample.

[0208] Specifically, the feature processing module includes:

[0209] The segmentation subunit is used to segment and slice each signal according to a preset time window length L=1024;

[0210] The feature extraction subunit is used to extract the mean value for each signal segment. Standard deviation Root mean square value skewness kurtosis Peak factor Peak-to-peak value and energy Eight statistical features are combined to form a feature vector;

[0211] The fusion subunit is used to stitch together features extracted from signals from different sources within the same time window into multi-source fusion features according to sensor type.

[0212] Normalized subunits are used to access the pre-stored global mean of the training set. and global standard deviation ,according to All multi-source fusion features are normalized and arranged in chronological order to form a unified feature sequence;

[0213] The state reasoning subunit is used to input a unified feature sequence into the current model, obtain fault diagnosis results, severity assessment results, remaining life prediction results and model confidence, and generate pre-maintenance state benchmark samples.

[0214] The global representation extraction module communicates with the feature processing module and is configured to input a unified feature sequence and a condition embedding token constructed from operating condition variables into an attention-based encoder to extract the global health status representation of the ship's power system.

[0215] Specifically, the global representation extraction module includes:

[0216] The operating condition encoding subunit is used to concatenate three types of operating condition information—speed (continuous variable), load (continuous variable), and fuel type (discrete variable)—into an operating condition vector. And through the working condition mapping matrix Projected into the hidden space, forming a working condition embedded token;

[0217] Sequence mapping subunit, used to map sequences with uniform features Through the feature mapping matrix Projected onto the same latent space, forming a sequence of feature tokens;

[0218] The positional encoding overlay subunit is used to overlay positional codes onto the concatenated token sequence. Generate the input representation of the Transformer encoder. ;

[0219] The main body of the Transformer encoder is composed of... = It consists of 6 stacked encoder layers. Each encoder layer includes a multi-head self-attention sub-layer and a feedforward network sub-layer. Each sub-layer is followed by a residual connection and a layer normalization operation, which are used to extract the long-range dependency between any two tokens in the sequence and output a global representation.

[0220] The global token extraction subunit is used to select the output corresponding to the global token from the global representation output by the encoder as the global health status representation hghg.

[0221] In one or more embodiments, the multi-task output module is communicatively connected to the global representation extraction module and configured to simultaneously output fault type identification results, fault severity assessment results, and remaining lifetime prediction results based on the global health status representation.

[0222] Specifically, the multi-task output module includes three parallel output heads:

[0223] The fault classification output layer uses a softmax classifier, with the input being... The output is a probability distribution vector of fault types. ;

[0224] The severity regression output layer uses a linear regressor, with the input being... The output is a fault severity assessment value. ∈[0,1];

[0225] The remaining lifetime prediction output layer uses a linear regressor with ReLU activation, and the input is... The output is the predicted remaining lifetime value. ≥0 (unit: hours).

[0226] The three output heads share the same global health status representation. A weighted multi-task loss function is used during training. Joint optimization is performed to achieve collaborative learning for fault diagnosis, degradation assessment, and life prediction.

[0227] In one or more embodiments, the gating decision module is communicatively connected to the multi-task output module and configured to generate corresponding operation and maintenance action instructions through the gating decision mechanism based on the fault diagnosis results, severity assessment results, remaining lifetime prediction results and model confidence.

[0228] Specifically, the gating decision module includes:

[0229] Damage severity assessment subunit, used to assess damage based on severity. Damage is classified into three levels: S0, S1, and S2.

[0230] The fault feature coding subunit is used to identify the result based on the fault type. Determine the fault characteristic codes for R0-R4;

[0231] The remaining lifetime range determination subunit is used to determine the remaining lifetime range based on the predicted value. Divide the area into three levels: safe, warning, and dangerous.

[0232] The decision execution matrix storage sub-unit is used to store a mapping table of the combination of damage level, fault feature code and remaining life prediction interval to operation and maintenance recommendations;

[0233] Confidence-gated subunits are used to determine the model confidence level. The comparison result with the preset threshold (0.7) determines whether to allow the automatic execution of high-risk operation and maintenance commands;

[0234] The instruction output subunit is used to convert the finalized operation and maintenance action instructions into standardized control signals (including Modbus / TCP messages or hardwired IO signals) and send them to the corresponding actuators.

[0235] In one or more embodiments, the feedback sample construction module is communicatively connected to the gating decision module and configured to acquire the retest data, action execution results and maintenance confirmation results after the operation and maintenance actions are executed, and establish a one-to-one correspondence with the pre-operation and maintenance state benchmark sample to form a closed-loop feedback sample.

[0236] Specifically, the feedback sample construction module includes:

[0237] The association identifier generation sub-unit is used to generate a unique association identifier using the execution timestamp of the operation and maintenance action and the action type code as a composite primary key;

[0238] The post-operation and maintenance status acquisition subunit is used to reacquire the unified feature sequence and inference results after operation and maintenance within a preset time window (30 minutes to 2 hours) after the operation and maintenance action is completed, through the data acquisition module and feature processing module.

[0239] The maintenance conclusion input subunit is used to receive the disassembly and inspection confirmation conclusions entered by maintenance personnel through the human-machine interface, including the actual fault type, the actual damage degree and the physical remaining life of the replaced parts; the sample assembly subunit is used to associate and bind the five parts of information, namely the pre-maintenance status benchmark sample, recommended maintenance actions, actual executed actions, retest status and maintenance conclusion, to form a closed-loop feedback sample.

[0240] In one or more embodiments, the incremental learning module is communicatively connected to the feedback sample construction module and the feature processing module, and is configured to determine the validity of the closed-loop feedback sample based on the change magnitude of multi-source features before and after maintenance, the consistency between fault diagnosis results and maintenance confirmation results, and the model confidence level, and to dynamically correct the model parameters and decision-making strategies using a weighted incremental learning mechanism for the valid feedback sample.

[0241] Specifically, the incremental learning module includes:

[0242] The validity determination subunit is used to calculate the validity determination value for each closed-loop feedback sample. and compare it with a preset threshold. =0.5 is used to compare and determine whether the sample is a valid feedback sample;

[0243] The tag write-back subunit is used to correct fault tags, remaining life tags, and maintenance action triggering conditions based on the maintenance confirmation conclusions, and to generate corrected tags. ;

[0244] The feedback dataset construction sub-unit is used to organize valid feedback samples according to... The format is organized as a feedback dataset;

[0245] Parameter update subunit, used to update parameters according to... Perform incremental updates to model parameters;

[0246] The strategy correction subunit is used to dynamically calibrate the decision threshold in the gating decision matrix based on the deviation trend between the diagnostic conclusions and maintenance conclusions in the feedback samples (for example, when multiple consecutive feedback samples show that the model overestimates the severity of a certain fault type, the corresponding severity judgment threshold is automatically increased).

[0247] refer to Figure 8 As shown in the system architecture, this application also provides an intelligent operation and maintenance device for clean fuel marine propulsion systems. This device integrates the aforementioned multi-source heterogeneous closed-loop intelligent operation and maintenance system, presented in the physical form of an embedded industrial control computer or a high-performance edge computing gateway, and installed in a standard cabinet in the ship's engine room monitoring room.

[0248] Specifically, the intelligent operation and maintenance device includes the following hardware components:

[0249] The main processor board is equipped with a high-performance multi-core CPU (central processing unit) and GPU (graphics processing unit) or NPU (neural network processing unit) to run the various software modules of the multi-source heterogeneous closed-loop intelligent operation and maintenance system.

[0250] The data acquisition interface board provides multiple analog and digital input channels. It communicates with vibration sensors, pressure sensors, temperature sensors, and speed sensors installed on the host, propulsion shaft, motor, gearbox, and bearings via shielded cables to receive multi-source heterogeneous operating signals collected by the sensors.

[0251] The communication interface board includes an industrial Ethernet interface (supporting protocols such as Modbus / TCP and PROFINET) and a CAN bus interface, which are used to communicate with actuators such as frequency converters, programmable logic controllers (PLCs), and automatic oiling machines, to issue operation and maintenance decision instructions and receive execution feedback;

[0252] Storage boards, including solid-state drive (SSD) storage units, are used to store pre-trained model parameters and global statistics of the training set. and Historical pre-operation and maintenance status benchmark samples and closed-loop feedback sample datasets;

[0253] The human-machine interface includes a touch screen and a key input unit, which is used to display fault diagnosis results, severity assessment results, remaining life prediction results and operation and maintenance decision suggestions, and to receive disassembly and inspection confirmation conclusions entered by maintenance personnel.

[0254] The power module is used to convert the AC power from the ship's electrical grid into the DC operating voltage required by the device, and has overvoltage protection, undervoltage protection and surge suppression functions.

[0255] In one or more embodiments, the main processor board automatically loads the pre-trained Transformer multi-task fault diagnosis model after power-on and receives sensor data in real time through the data acquisition interface board, executing all the calculation processes of steps 1 to 7 of the method. When an operation and maintenance decision instruction is generated, the main processor board sends the instruction to the corresponding execution mechanism through the communication interface board; when a maintenance confirmation conclusion is obtained, the main processor board calls the incremental learning module to perform online updates of the model parameters. Through the synergy of the above hardware configuration and software algorithm, the device achieves a complete closed loop of "monitoring—diagnosis—prediction—decision-feedback—optimization," enabling stable operation in the harsh environment of the ship's engine room (temperature -15℃~55℃, humidity 10%~95%, vibration level 2g), significantly improving the intelligent operation and maintenance level of clean fuel ship power systems.

[0256] In one or more embodiments, this application also provides a computer-readable storage medium storing a computer program thereon. When executed by a processor, the computer program is capable of implementing all the steps of the above-described multi-source heterogeneous closed-loop intelligent operation and maintenance method.

[0257] Specifically, the computer-readable storage medium includes, but is not limited to, any of the following types: solid-state drive (SSD), hard disk drive (HDD), USB flash drive, read-only memory (ROM), random access memory (RAM), optical disc (CD-ROM or DVD), magneto-optical disc, register memory, flash memory, or other forms of non-transitory computer-readable storage media known to those skilled in the art. As one embodiment, the storage medium is an internal solid-state drive of a ship's engine room intelligent operation and maintenance device.

[0258] The computer program includes multiple program modules. When these modules are loaded and executed by the processor, they specifically perform the following operations: A first program module implements a data acquisition function, reading multi-source heterogeneous operating signals from each sensor channel according to a preset sampling frequency; a second program module implements a feature processing function, performing time alignment, segmentation, eight-dimensional statistical feature extraction, global normalization, and serialization on the read multi-source heterogeneous operating signals to construct a unified feature sequence; a third program module implements a global representation extraction function, inputting the unified feature sequence and the operating condition embedded token into the Transformer encoder to extract a global health status representation; and a fourth program module implements a fifth program module. The fourth program module is used to implement multi-task output function, simultaneously outputting fault type identification results, fault severity assessment results, and remaining lifetime prediction results based on global health status representation; the fifth program module is used to implement gating decision function, generating operation and maintenance action instructions based on fault diagnosis results, severity assessment results, remaining lifetime prediction results, and model confidence; the sixth program module is used to implement feedback sample construction function, associating pre-operation and maintenance status baseline samples with post-operation and maintenance retest data and maintenance confirmation conclusions to form closed-loop feedback samples; the seventh program module is used to implement incremental learning function, determining the validity of closed-loop feedback samples, and updating model parameters and decision-making strategies through weighted incremental learning.

[0259] In one or more embodiments, the computer program is deployed in an edge-cloud collaborative architecture: the edge (i.e., the ship's engine room intelligent operation and maintenance device) loads the first to fifth program modules, responsible for real-time data acquisition, feature processing, inference diagnosis, and decision output; the sixth and seventh program modules are deployed on the cloud server, responsible for centralized storage of feedback samples, batch validity determination, and model parameter updates. The updated model parameters are periodically sent to the edge via the ship's maritime satellite communication link. This deployment method ensures real-time response capabilities while fully utilizing the large-scale computing resources of the cloud.

[0260] The above technical solutions only embody the preferred technical solutions of the present invention. Any modifications that may be made by those skilled in the art to certain parts thereof embody the principles of the present invention and fall within the protection scope of the present invention.

Claims

1. A multi-source heterogeneous closed-loop intelligent operation and maintenance method for ammonia, LNG, and methanol fueled ship propulsion systems, characterized in that, The method includes the following steps: Step 1: For ammonia, LNG or methanol fuel ship propulsion systems, use sensors to collect multi-source heterogeneous operating signals; Step 2: Process the multi-source heterogeneous operating signals, construct a unified feature sequence, and obtain the corresponding fault diagnosis results, fault severity assessment results, remaining lifetime prediction results and model confidence scores to generate a pre-operation and maintenance state benchmark sample. Step 3: Input the unified feature sequence and the condition embedding token constructed from the operating condition variables into the attention-based encoder to extract the global health status representation of the ship's power system. Step 4: Based on the global health status representation, simultaneously output the fault type identification result, fault severity assessment result, and remaining lifetime prediction result through the multi-task output module; Step 5: Based on the fault diagnosis results, fault severity assessment results, remaining lifetime prediction results, and model confidence, generate corresponding operation and maintenance action instructions through a gating decision mechanism; Step 6: Obtain the retest data, action execution results, and maintenance confirmation results after the operation and maintenance actions are performed, and establish a one-to-one correspondence with the pre-operation and maintenance state benchmark sample to form a closed-loop feedback sample; Step 7: Based on the changes in multi-source features before and after maintenance, the consistency between fault diagnosis results and maintenance confirmation results, and the model confidence in the closed-loop feedback samples, determine the validity of the closed-loop feedback samples, and dynamically correct the model parameters and decision-making strategies using a weighted incremental learning mechanism.

2. The method according to claim 1, characterized in that, The specific process of constructing the unified feature sequence in step 2 includes: Step 2A: Let the first... The raw operating signals acquired by the sensor are Divide the time window into segments according to the preset time window length to obtain the first segment. A signal segment; Step 2B: Extract statistical feature vectors for each signal segment. The statistical feature vectors include mean, standard deviation, root mean square value, skewness, kurtosis, crest factor, peak-to-peak value, and energy. Step 2C: Concatenate the features extracted from signals from different sources within the same time window according to the number of sensor signal sources to obtain the first... Multi-source fusion features of each window; Step 2D: Normalize all multi-source fusion features using the global mean and standard deviation of the training set, and arrange them in chronological order to obtain a unified feature sequence.

3. The method according to claim 1, characterized in that, The process of fusing the embedded token and the unified feature sequence in step 3 includes: The unified feature sequence and operating condition variables are projected onto the same latent space through the feature mapping matrix and the operating condition mapping matrix, respectively. After concatenation, positional encoding is superimposed to obtain the encoder input representation based on the attention mechanism.

4. The method according to claim 1, characterized in that, In step 3, the encoder based on the attention mechanism is a Transformer encoder. The Transformer encoder extracts the correlation between different time segments and signals from different sources through a multi-head self-attention mechanism to obtain a global health status representation of the ship's power system.

5. The method according to claim 1, characterized in that, In step 4, the multi-task output module outputs fault type identification results, fault severity assessment results, and remaining lifetime prediction results based on the global health status representation. The specific implementation methods include: The global health status representation is simultaneously input to the fault classification output layer, severity regression output layer, and remaining lifetime prediction output layer in the multi-task output module. The fault classification output layer, the severity regression output layer, and the remaining lifetime prediction output layer output fault type identification results, fault severity assessment results, and remaining lifetime prediction results, respectively.

6. The method according to claim 1, characterized in that, The gating decision-making mechanism in step 5 adopts a decision execution matrix, which maps to the corresponding operation and maintenance recommendations based on the combination of damage level, fault feature code, and remaining life prediction interval.

7. The method according to claim 1, characterized in that, If the closed-loop feedback sample in step 7 fails any of the manual confirmation, retest verification, or security gating judgment, the closed-loop feedback sample is marked as a sample to be verified and will not participate in the fault diagnosis model parameter update.

8. A multi-source heterogeneous closed-loop intelligent operation and maintenance system for ammonia, LNG, and methanol fuel ship propulsion systems, characterized in that, The system operates based on the method as described in any one of claims 1 to 7, comprising: The data acquisition module is used to collect multi-source heterogeneous operating signals using sensors; The feature processing module is communicatively connected to the data acquisition module. It is used to process the multi-source heterogeneous operating signals, construct a unified feature sequence, and obtain the corresponding fault diagnosis results, fault severity assessment results, remaining life prediction results and model confidence, and generate a pre-operation and maintenance state benchmark sample. The global representation extraction module is communicatively connected to the feature processing module and is used to input the unified feature sequence and the working condition embedded token into the attention mechanism-based encoder to extract the global health status representation of the ship's power system. The multi-task output module is communicatively connected to the global representation extraction module and is used to simultaneously output fault type identification results, fault severity assessment results, and remaining lifetime prediction results based on the global health status representation. The gating decision module is communicatively connected to the multi-task output module and is used to generate corresponding operation and maintenance action instructions based on the fault diagnosis results, fault severity assessment results, remaining lifetime prediction results and model confidence through the gating decision mechanism. The feedback sample construction module is connected to the gating decision module and is used to obtain the retest data, action execution results and maintenance confirmation results after the operation and maintenance action is executed. It establishes a one-to-one correspondence with the pre-operation and maintenance state benchmark sample to form a closed-loop feedback sample. The incremental learning module, which is communicatively connected to the feedback sample construction module and the feature processing module, is used to determine the validity of the closed-loop feedback sample based on the change magnitude of multi-source features before and after maintenance, the consistency between fault diagnosis results and maintenance confirmation results, and the model confidence level. The valid feedback sample is then used to dynamically correct the model parameters and decision-making strategy through a weighted incremental learning mechanism.

9. An intelligent operation and maintenance device for a clean fuel ship propulsion system, characterized in that, Including the system according to claim 8, the intelligent operation and maintenance device is installed in the ship's engine room monitoring room and is communicatively connected to vibration sensors, pressure sensors, temperature sensors and speed sensors installed on the main engine, propulsion shaft system, motor, gearbox and bearings. It is used to receive multi-source heterogeneous operating signals collected by the sensors and perform fault diagnosis, life prediction and operation and maintenance decision command generation and feedback optimization.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it can implement the steps of the method as described in any one of claims 1 to 7.

Images