Fault monitoring system and method for ship power system

Abstract

The invention relates to the technical field of ships, in particular to a fault monitoring system and method for a ship power system, and the system comprises a multi-source data collection module which carries out the real-time collection of the operation parameters, environment parameters and equipment state parameters of the ship power system through a collection device. The fault early warning module is used for predicting the development trend of the fault after the fault diagnosis is completed; the early warning decision module generates early warning information of different levels according to the state evaluation result, the fault diagnosis result and the RUL prediction result, gives targeted operation and maintenance decision suggestions, and pushes the suggestions to a ship cockpit, a shore-based operation and maintenance center and an operation and maintenance personnel mobile terminal; 24-hour uninterrupted multi-parameter acquisition of the ship power system is realized through the multi-source sensor network, three types of parameters of operation, environment and state are covered, acquisition delay is reduced, and the problems of poor timeliness and incomplete parameter coverage of traditional manual inspection are solved.

Description

Technical Field

[0001] This application relates to the field of marine technology, and in particular to a fault monitoring system and method for a marine power system. Background Technology

[0002] The ship's propulsion system is the core power source for ship navigation, and its operating status directly determines the ship's navigation safety and operational efficiency. This system typically includes multiple subsystems such as the main engine, auxiliary engines, transmission, propulsion system, fuel system, lubrication system, and cooling system. These subsystems are closely interconnected, operate under complex conditions, and are subjected to harsh environments such as high load, high vibration, high humidity, and salt spray corrosion for extended periods. As a result, they are prone to potential failures such as component wear, performance degradation, pipeline leaks, and abnormal temperature increases.

[0003] Traditional methods of monitoring ship propulsion systems mainly rely on manual inspections and periodic maintenance, which have the following significant drawbacks: First, the monitoring timeliness is poor, as manual inspections cannot achieve 24-hour uninterrupted monitoring and cannot promptly capture signs of transient faults; second, the monitoring accuracy is low, as relying on human experience makes it susceptible to subjective factors and difficult to identify early, weak fault signals; third, the early warning lag is strong, often only being discovered after a fault has occurred or severely worsened, missing the optimal maintenance opportunity and potentially leading to serious consequences such as propulsion system paralysis and ship downtime; fourth, the data utilization rate is low, as traditional monitoring methods can only collect a small number of key parameters (such as speed, temperature, and pressure), and the data is not analyzed in depth to uncover the correlation between parameters and the evolution patterns of faults.

[0004] Existing technologies also include some monitoring systems based on single sensors or simple threshold judgments, such as monitoring the cylinder liner temperature of the main engine using a temperature sensor and issuing an alarm when the temperature exceeds a preset threshold. However, these systems have the following shortcomings: First, relying solely on single parameter monitoring cannot comprehensively reflect the overall operating status of the power system, and is prone to "false alarms" or "missed alarms" (for example, a brief exceedance of a threshold by a single parameter may be a normal fluctuation, while multiple parameters exhibiting abnormal behavior together is a precursor to a fault); second, they lack the ability to accurately diagnose fault types and predict fault development trends, and can only provide "fault alarms," ​​failing to inform maintenance personnel of the cause of the fault, the location of the fault, and the remaining service life, leading to significant uncertainty in maintenance decisions. Summary of the Invention

[0005] This application provides a fault monitoring system and method for a ship propulsion system to improve the following technical problems.

[0006] This application provides a fault monitoring system and method for a marine propulsion system, employing the following technical solution:

[0007] A fault monitoring system for a ship's propulsion system includes a multi-source data acquisition module, a data preprocessing module, a condition assessment module, a fault diagnosis module, a fault early warning module, an early warning decision module, and a data storage module.

[0008] The multi-source data acquisition module collects the operating parameters, environmental parameters, and equipment status parameters of the ship's power system in real time through the acquisition equipment;

[0009] The input end of the data preprocessing module is electrically connected to the output end of the multi-source data acquisition module, and the data preprocessing module is used to clean, reduce noise, normalize and extract features from the raw data acquired by the multi-source data acquisition module.

[0010] The status assessment module assesses the operating status of the ship's power system in real time based on the preprocessed feature data, and determines whether the ship's power system is in a normal, sub-healthy or abnormal state.

[0011] The fault diagnosis module is used to identify the fault type, fault location and fault severity when the condition assessment module determines that the ship's power system is in a sub-healthy or abnormal state.

[0012] The fault early warning module is used to predict the development trend of the fault after the fault diagnosis is completed, and the input terminal of the fault early warning module is electrically connected to the output terminal of the fault diagnosis module.

[0013] The early warning decision module generates early warning information of different levels based on the status assessment results, fault diagnosis results and RUL prediction results, and provides targeted operation and maintenance decision suggestions, which are pushed to the ship's cockpit, shore-based operation and maintenance center and operation and maintenance personnel's mobile terminals.

[0014] The data storage module is used to store data from various modules within the system, and it employs the time-series database InfluxDB and the relational database MySQL.

[0015] Furthermore, the operating parameters collected by the multi-source data acquisition module include: main engine speed, main engine power, fuel consumption, lubricating oil pressure, lubricating oil temperature, cooling water temperature, exhaust temperature, cylinder pressure, transmission speed, and propeller torque.

[0016] Environmental parameters include: seawater temperature, seawater salinity, atmospheric temperature, atmospheric humidity, and wind and wave levels in the sea area where the ship navigates;

[0017] Status parameters include: equipment vibration acceleration, equipment surface temperature, ultrasonic signals of pipeline leaks, motor current / voltage, and valve open / close status;

[0018] The acquisition device is a distributed sensor network based on the Internet of Things, and the multiple sensors include: vibration sensors, thermocouple temperature sensors, pressure sensors, ultrasonic leak sensors, current transformers, and GPS positioning modules.

[0019] Furthermore, in the data preprocessing module, a hybrid algorithm of the 3σ criterion and isolation forest is used for outlier rejection. The specific strategy is as follows: initially reject outliers beyond the reasonable range through the 3σ criterion. The 3σ criterion determines the normal data range as [μ - 3σ, μ + 3σ] based on the mean μ and standard deviation σ of the data. That is, when the data x satisfies |x - μ| > 3σ, it is determined as an obvious outlier and is rejected.

[0020] Then, use the isolation forest algorithm to deeply detect the remaining data. The isolation forest algorithm constructs multiple isolation trees, calculates the path length h(x) of each data point in the tree, and compares it with the expected path length E[h(x)]. When h(x) < E[h(x)], it is determined as a hidden outlier and is rejected.

[0021] When the data preprocessing module performs noise reduction on the original data, it uses an algorithm that combines variational mode decomposition and wavelet threshold denoising. The specific strategy is as follows:

[0022] First, set the number of modes in the variational mode decomposition algorithm (i.e., VMD algorithm) to 3 - 8, and the penalty parameter α = 2000. Decompose the original signal f(t) into multiple intrinsic mode functions uk(t). The objective function of the VMD algorithm is The constraint condition is ∑ k u k (t) = f(t), where w k is the central frequency corresponding to the k-th intrinsic mode function uk(t), δ(t) is the Dirac function, which is often used as an idealized impulse signal in signal processing to construct the operation form of the VMD algorithm; j' is the imaginary unit; is the partial derivative with respect to time t; solve this optimization problem through the alternating direction multiplier method to achieve the separation of the signal and noise;

[0023] Then, perform wavelet threshold denoising on the noisy IMF components. Select the db4 wavelet basis function and adopt the Birge - Massart strategy for the threshold rule. The wavelet threshold denoising first performs wavelet transform on the noisy signal to obtain the wavelet coefficients w j ,k, and then processes the wavelet coefficients according to the threshold λ. The processing rule is where, sign(w j ,k) is the sign function. When w j ,k > 0, sign(w j ,k) = 1; when w j ,k = 0, sign(wj ,k)=0;w j When k < 0, sign(w) j ,k)=-1; Finally, the processed wavelet coefficients are subjected to inverse wavelet transform to reconstruct the signal;

[0024] The data preprocessing module uses a max-min normalization algorithm to normalize the raw data, mapping parameters of different dimensions to the [0,1] interval. The calculation formula is as follows: Where, x norm The data is normalized, and x is the original data. max x min These are the maximum and minimum values ​​of the parameter in historical data, respectively.

[0025] Furthermore, the data preprocessing module extracts key features reflecting the device status from the preprocessed data. These key features include time-domain features, frequency-domain features, and time-frequency-domain features. The time-domain features include mean, variance, peak value, peak-to-peak value, kurtosis, and margin factor. The frequency-domain features are obtained by converting the time-domain signal x(t) into the frequency-domain signal X(f) using the Fast Fourier Transform (FFT), extracting the characteristic frequency amplitude, spectral centroid, and spectral variance. The specific strategy is as follows: The FFT calculation formula is: Extract the characteristic frequency amplitude |X(f0)| from the frequency domain signal, where f0 is the characteristic frequency of equipment fault; the formula for calculating the centroid of the spectrum is: The formula for calculating the spectral variance is: Among them, f cg Let f be the centroid of the spectrum, f be the frequency variable, and |X(f)| be the amplitude corresponding to frequency f. For spectral variance;

[0026] Time-frequency domain features are extracted using wavelet packet transform (WPT) to obtain wavelet packet energy entropy and wavelet packet singular entropy. The wavelet packet energy entropy E... e The calculation formula is: in, E i Let S be the energy of the i-th node, M be the number of nodes, and S be the singular entropy of the wavelet packet. e The calculation formula is: in, s i represents the singular values ​​of the wavelet packet coefficients, and L is the number of singular values;

[0027] After feature extraction, a feature matrix with dimension n×m is formed, where n is the number of sampling time points and m is the number of features.

[0028] Furthermore, the state assessment module adopts a model that combines a stacked autoencoder with fuzzy comprehensive evaluation;

[0029] The stacked autoencoder (SAE) training strategy is as follows:

[0030] Network structure design: The number of nodes in the input layer is consistent with the dimension m of the feature matrix. A 3-5 hidden layer structure is constructed. The number of nodes in each hidden layer decreases and then increases proportionally, set to m / 2, m / 4, m / 8, m / 4, and m / 2 respectively. The number of nodes in the output layer is set to m to realize the reconstruction output of the feature vector.

[0031] By introducing an L2 regularization term and a sparsity constraint term into the SAE loss function, the improved loss function formula is as follows: in, For the input feature vector, To output the reconstructed feature vector, λ' is the regularization coefficient, and the value of λ' ranges from 0.001 to 0.01; w lj The weight matrix represents the weights of the j-th neuron in the l-th layer; L is the total number of hidden layers; s l ρ is the number of neurons in layer l; β is the sparsity coefficient, and the value of β ranges from 0.1 to 1; ρ is the expected sparsity. And a lj (x) represents the input x from the j-th neuron in the l-th layer. i Activation value; Let KL divergence be a metric.

[0032] The SAE model is pre-trained and fine-tuned using historical data under normal operating conditions of the ship's power system, and the sample size of the historical data is no less than 10,000 sets.

[0033] The specific strategy for fuzzy comprehensive evaluation is as follows:

[0034] Establish an evaluation index system: This includes the reconstruction error E of the SAE model and the normalized values ​​x of each parameter. norm,1 ,x norm,2 ,…,x norm,k' As evaluation indicators, an indicator set U = {E, x} is formed. norm,1 ,x norm,2 ,…,x norm,k'}, where k' is the number of parameters;

[0035] The Analytic Hierarchy Process (AHP) is used to determine the weights of each evaluation indicator. Specifically, the judgment matrix A = (a ij ) (k+1)×(k+1) a ij To determine the importance of indicator i relative to indicator j; then, calculate the largest eigenvalue λ of the judgment matrix. maxAnd its corresponding eigenvector w; then perform a consistency check, and if the consistency ratio CR < 0.1, the judgment matrix is ​​considered to meet the consistency requirements, and the weight vector W = {w E ,w1,w2,…,w k}, where w E The weights for the reconstruction error;

[0036] Establish membership functions for three evaluation levels: "normal," "sub-healthy," and "abnormal." The membership functions use a triangular membership function. Specifically, let the evaluation level set V = {v1(normal), v2(sub-healthy), v3(abnormal)}. For an index u∈U, its membership function with respect to level v1 is: The membership function for level v2 is: The membership function for level v3 is: Where a1, b1, and b2 are thresholds determined based on historical data; the membership vector of the system state to each level is calculated through fuzzy synthesis operation. Where R is the membership matrix, R ij =μ j (u i ),and The final output health index is H = B1×1 + B2×0.65 + B3×0.25, where H ranges from 0 to 1. 1, 0.65, and 0.25 are the baseline values ​​for the health levels of "normal", "sub-healthy", and "abnormal", respectively. When H ≥ 0.8, it is considered normal; 0.5 ≤ H ≤ 0.8, it is considered sub-healthy; and H < 0.5, it is considered abnormal.

[0037] Furthermore, the fault diagnosis module employs a fusion diagnostic algorithm based on the attention mechanism of a convolutional neural network-long short-term memory network (CNN-LSTM-Attention model) and a K-nearest neighbor model (KNN). The CNN-LSTM-Attention model is used to extract spatial and temporal features from the feature data and strengthens fault-sensitive features through an attention mechanism. The specific strategy of the CNN-LSTM-Attention model is as follows:

[0038] The network structure is designed as follows:

[0039] Input layer: receives the feature matrix Where n represents the time step and m is the number of features;

[0040] CNN layers: consist of 2-3 convolutional layers, each using a 3×3 kernel with 32-64 kernels. The activation function is ReLU. The output of the t-th convolutional layer is: C t=ReLU(Conv(X) t-1 ,K t )+b t ), where K t Let b be the kernel of the t-th layer. t For bias terms;

[0041] Pooling layer: Employs max pooling operation with a pooling kernel size of 2×2. Achieve dimensionality reduction;

[0042] LSTM layer: contains 1-2 LSTM units, with the number of hidden layer nodes ranging from 64 to 128;

[0043] Attention layer: An additive attention mechanism is used to calculate the hidden state weights at each time step, specifically: calculating the score: s t =v T tanh(W h h t +b), This is the weight matrix. The bias is d, where d is the attention dimension and h is the number of hidden points in the LSTM layer; then the weights are normalized using the formula: The normalized weights are then weighted by features, using the following formula: Among them, h t For the hidden layer node output of the LSTM at time step t, s t Let s be the score at time t. k The score at time k;

[0044] Fully connected layer: The number of output layer nodes is the same as the number of fault types K. The Softmax activation function is used to distribute the probability p of the output fault type as follows: p = Softmax(W fc h att +b fc ),in, This is the weight matrix. It is the bias vector;

[0045] The model training strategy is as follows:

[0046] Dataset construction: Use historical data of marine power systems, including but not limited to 10-15 types of faults such as main engine cylinder liner wear, lubricating oil line leakage, cooling system blockage, and propeller bearing damage, with no less than 500 samples for each type of fault;

[0047] Data partitioning: The training set and the test set are partitioned in a 7:3 ratio;

[0048] Optimization method: Use cross-entropy loss function Among them, y ikp is the true label for fault type k corresponding to sample i. ik The model is trained using the Adam optimizer to predict probabilities.

[0049] The specific strategy of the KNN model is as follows:

[0050] Feature weights and distance weighting are introduced, and the feature weight vector w = {w1, w2, ..., w...} is determined based on the sensitivity of features to faults. m}, calculate the difference between the test sample x and the training sample x. i Weighted Euclidean distance Neighbor samples are weighted by the inverse distance, with the weights... k' is the nearest neighbor number.

[0051] Furthermore, the fault early warning module includes a trend prediction unit and a lifetime estimation unit. The trend prediction unit uses a Long Short-Term Memory (LSTM) network model based on historical fault evolution data and current fault characteristics to predict the changing trend of fault characteristic parameters over a future period. The specific strategy of the trend prediction unit is as follows:

[0052] The LSTM model is trained using complete time-series data covering the entire process from fault occurrence to deterioration, with each fault type having a time series length of at least 1000 time steps. The input to the LSTM model is the historical sequence X = {x} of the fault feature parameters. t-T'+1 ,x t-T'+2 ,...,x t}, where T' is the length of the input sequence, ranging from 30 to 50; the output is the predicted parameter values ​​for the next 100 time steps.

[0053] The number of hidden layer nodes is set to 64-128, and the root mean square error (RMSE) is used as the loss function, calculated as follows: Where M is the total number of prediction time steps, and the LSTM model is trained using the Adam optimizer;

[0054] The lifetime estimation unit uses the particle filter algorithm (PF) to predict the lifetime of the power system, and the lifetime prediction process is as follows:

[0055] The feature parameter sequence predicted by the LSTM model z, as an observation of the PF model t+1 ,z t+2 ,…,z t+T' The observation equation of the PF model is defined as: z t =W t +δ t , where δ t To conform to a normal distribution For the observed noise, the execution strategy is as follows:

[0056] Particle initialization: Randomly sample N initial particles from the prior distribution p(W0). And assign an initial weight to each particle.

[0057] State prediction: Calculate according to the state transition equation to generate the particle set at time t.

[0058] Weight update: Calculate the particle weights according to the observation equation. The weight calculation formula is where is the probability density function, and After the calculation, normalize the particle weights. The normalization formula is

[0059] Resampling: Calculate the number of effective particles Ne eff , when Ne eff < N / 2, perform the resampling operation to generate a new particle set And reset the weights of all particles to 1 / N;

[0060] State estimation: Calculate the estimated value of the wear amount at time t by weighted summation. The formula is

[0061] Loop and execute the above steps until the predicted wear amount reaches the pre-set critical failure threshold W th , define the difference between the time point t fail when the threshold is reached and the current time t now as the remaining useful life RUL = t fail - t now .

[0062] Furthermore, the different levels of warning information in the warning decision module are divided into:

[0063] The first-level warning is the normal state: The state health index H ≥ 0.8, indicating that the system is in a fault-free operation state, and push the system operation status report;

[0064] The second-level warning is the attention state: The state health index 0.5 ≤ H ≤ 0.8, indicating that the system has no obvious faults but there are potential risk hazards, and push attention suggestions;

[0065] The third-level warning is the warning state: The state health index H < 0.5, indicating that the system has identified a minor fault, and the remaining useful life RUL > 7 days, and push warning and maintenance preparation suggestions;

[0066] Level 4 warning indicates an emergency: This means the system has detected a moderate or severe fault, and the remaining service life is...

[0067] If the RUL is less than 7 days, an emergency warning and emergency response suggestions will be sent.

[0068] The strategy for generating operation and maintenance decision recommendations is as follows: based on the fault type and the real-time navigation status of the vessel, which includes key parameters such as geographical location (Lat / Lon), planned route (Route), and current cargo load (Load), a preset operation and maintenance decision knowledge base module is invoked. This knowledge base stores integrated fault handling standard procedures, spare parts reserve lists, and maintenance process specifications. The strategy uses case-based reasoning, i.e., the CBR fusion algorithm, to generate operation and maintenance decision recommendations.

[0069] Furthermore, the data storage module stores the following data:

[0070] Raw data: Stores the raw parameter values ​​collected by the sensor, indexed by timestamp;

[0071] Preprocessed data and feature matrix: InfluxDB stores the preprocessed parameter time series, MySQL stores the feature matrix, and the table structure of the feature matrix includes "feature ID, device number, timestamp, time domain feature, frequency domain feature, and time-frequency domain feature".

[0072] The results of condition assessment, fault diagnosis, fault prediction, and early warning decision-making are stored in MySQL, with the following table structures:

[0073] Status assessment results table: Assessment ID, Equipment number, Assessment time, Health index H, System status, Reconfiguration error E;

[0074] Fault Diagnosis Result Table: Diagnosis ID, Device Number, Diagnosis Time, Fault Type, Fault Location, Fault Severity, CNN-LSTM-Attention Confidence, KNN Confidence, Final Confidence;

[0075] Fault Prediction Results Table: Prediction ID, Equipment Number, Prediction Time, Fault Characteristic Parameter Prediction Sequence, RUL, Prediction Error Range;

[0076] Early warning decision results table: Early warning ID, device number, early warning time, early warning level, early warning information, operation and maintenance suggestions, and push recipients;

[0077] Historical fault data and maintenance records of marine power systems: MySQL stores historical fault data and maintenance records. The historical fault data includes: fault ID, equipment number, fault occurrence time, fault type, fault cause, handling process, and maintenance result. The maintenance records include: maintenance ID, equipment number, maintenance time, maintenance content, spare parts usage, and maintenance personnel.

[0078] Furthermore, a method for monitoring and early warning of a ship's power system includes the following steps:

[0079] S1. Data Acquisition: Start the multi-source data acquisition module to collect the operating parameters, environmental parameters and status parameters of the ship's power system in real time through a distributed sensor network. The acquisition frequency is set according to the parameter type.

[0080] S2. Data Preprocessing: The data preprocessing module reads raw data from the InfluxDB data storage module according to a fixed time window and performs cleaning, noise reduction, normalization, and feature extraction.

[0081] S3. State Assessment: The state assessment module reads the feature matrix of the current time window from MySQL, inputs it into the trained SAE model, calculates the reconstruction error, constructs the index set U and weight vector W of fuzzy comprehensive evaluation, calculates the membership degree of each index to "normal", "sub-healthy", and "abnormal", obtains the membership degree vector B through the weighted average method, calculates the health index, and judges the system state; the assessment results are stored in the state assessment result table of MySQL.

[0082] S4. Fault Diagnosis: If the status assessment result is "sub-healthy" or "abnormal", the fault diagnosis module reads the feature matrix from MySQL, inputs it into the CNN-LSTM-Attention model, obtains the fault type probability distribution p, and calculates the maximum confidence.

[0083] S5. Fault Prediction: After the fault diagnosis is completed, the fault early warning module reads the historical sequence of fault characteristic parameters of the device and the current diagnosis results from MySQL, inputs them into the LSTM model, predicts the trend of characteristic parameters for the next 7 days, and stores the prediction results in the fault prediction result table of MySQL.

[0084] S6. Early Warning Decision-Making: The early warning decision-making module reads the status assessment result (H), fault diagnosis result, and fault prediction result (RUL) from MySQL, and then calls the operation and maintenance decision knowledge base to store the early warning information and suggestions.

[0085] The MySQL early warning decision results table is pushed to the ship's cockpit display terminal, the shore-based operation and maintenance center management platform, and the operation and maintenance personnel's mobile APP via the MQTT protocol;

[0086] S7. Model Iteration and Optimization: An iteration cycle is conducted every 3 months. New historical fault data and maintenance records are read from MySQL to update the models of each module.

[0087] State Assessment (SAE) Model: Add 5000 sets of normal operation data, retrain the model, and optimize λ and β in the loss function;

[0088] Fault diagnosis CNN-LSTM-Attention model: Add new fault data with 200 sets of each fault, expand the training set, retrain the network, and adjust the number of convolutional kernels and the number of LSTM hidden layer nodes;

[0089] Fault prediction LSTM and PF models: Add fault evolution data with 300 time steps, retrain the LSTM model, and optimize the process noise and observation noise variance of the PF model; the iterated model replaces the original model.

[0090] In summary, this application includes at least one of the following beneficial technical effects:

[0091] By using a multi-source sensor network, the system can continuously collect multiple parameters of the ship's power system 24 hours a day, covering three major categories of parameters: operation, environment, and status. The acquisition latency is reduced, which solves the problems of poor timeliness and incomplete parameter coverage in traditional manual inspections. Data transmission adopts dual-mode communication to ensure data stability in ocean scenarios and provide a continuous and complete data foundation for subsequent analysis.

[0092] High accuracy of status assessment: The model adopts a fusion of SAE and fuzzy comprehensive evaluation. SAE improves the sensitivity to weak anomalies through regularization and sparse constraints. Fuzzy comprehensive evaluation combines multiple indicators and subjective weights to achieve quantitative assessment of status. The accuracy of status health index assessment is improved, avoiding the limitations of single threshold judgment. It can identify sub-health status in advance and buy time for fault prevention.

[0093] Good fault prediction foresight: Based on the fusion model of LSTM and PF, LSTM achieves accurate prediction of short-term parameter trends, while PF combines fault physics model and adaptive resampling to improve the accuracy of RUL estimation. It can predict the fault development trend and remaining service life 1-7 days in advance, allowing maintenance personnel sufficient time to prepare for maintenance and avoiding downtime losses caused by sudden failures.

[0094] Intelligent operation and maintenance decision-making: Through the four-level early warning classification and CBR fusion algorithm, not only can differentiated early warnings be generated based on the severity of the fault and RUL, but also targeted operation and maintenance suggestions (such as route adjustment, spare parts model, and maintenance time estimation) can be generated in combination with the ship's navigation status. This reduces the dependence on the experience of operation and maintenance personnel (new operation and maintenance personnel can also respond quickly) and improves operation and maintenance efficiency. Attached Figure Description

[0095] To more clearly illustrate the technical solutions in the embodiments of this application, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0096] Figure 1 This is a flowchart illustrating the fault monitoring system for the ship's power system in this application.

[0097] Figure 2 This is a flowchart illustrating the monitoring and early warning method for the ship's power system proposed in this application.

[0098] Explanation of reference numerals in the attached figures:

[0099] 1. Multi-source data acquisition module; 2. Data preprocessing module; 3. Status assessment module; 4. Fault diagnosis module; 5. Fault early warning module; 501. Trend prediction unit; 502. Lifetime estimation unit; 6. Early warning decision module; 7. Data storage module. Detailed Implementation

[0100] To make the technical problems, technical solutions, and beneficial effects to be solved by this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and are not intended to limit the scope of this application.

[0101] Please see Figure 1 A fault monitoring system for a ship propulsion system includes a multi-source data acquisition module 1, a data preprocessing module 2, a condition assessment module 3, a fault diagnosis module 4, a fault early warning module 5, an early warning decision module 6, and a data storage module 7.

[0102] The multi-source data acquisition module 1 collects the operating parameters, environmental parameters, and equipment status parameters of the ship's power system in real time through the acquisition equipment;

[0103] The operating parameters collected in the multi-source data acquisition module 1 include: main engine speed, main engine power, fuel consumption, lubricating oil pressure, lubricating oil temperature, cooling water temperature, exhaust temperature, cylinder pressure, transmission speed, and propeller torque;

[0104] Environmental parameters include: seawater temperature, seawater salinity, atmospheric temperature, atmospheric humidity, and wind and wave levels in the sea area where the ship navigates;

[0105] Status parameters include: equipment vibration acceleration, equipment surface temperature, ultrasonic signals of pipeline leaks, motor current / voltage, and valve open / close status;

[0106] The acquisition device is a distributed sensor network based on the Internet of Things, and the multiple sensors include: vibration sensors, thermocouple temperature sensors, pressure sensors, ultrasonic leakage sensors, current transformers, and GPS positioning modules;

[0107] The input end of the data preprocessing module 2 is electrically connected to the output end of the multi-source data acquisition module 1, and the data preprocessing module 2 is used to clean, denoise, normalize, and extract features from the raw data collected by the multi-source data acquisition module 1;

[0108] In the data preprocessing module 2, a hybrid algorithm of the 3σ criterion and isolation forest is used to remove outliers. The specific strategy is: initially remove outliers beyond the reasonable range through the 3σ criterion. The 3σ criterion determines the normal data range as [μ - 3σ, μ + 3σ] based on the mean μ and standard deviation σ of the data. That is, when the data x satisfies |x - μ| > 3σ, it is determined as an obvious outlier and removed;

[0109] Then, use the isolation forest algorithm to deeply detect the remaining data. The isolation forest algorithm constructs multiple isolation trees, calculates the path length h(x) of each data point in the tree, and compares it with the expected path length E[h(x)]. When h(x) < E[h(x)], it is determined as a hidden outlier and removed;

[0110] When the data preprocessing module 2 denoises the raw data, it uses an algorithm that combines variational mode decomposition and wavelet threshold denoising. The specific strategy is:

[0111] First, set the number of modes in the variational mode decomposition algorithm, i.e., the VMD algorithm, to 3 - 8, and the penalty parameter α = 2000. Decompose the original signal f(t) into multiple intrinsic mode functions uk(t). The objective function of the VMD algorithm is The constraint condition is ∑ k u k (t) = f(t), where w k is the central frequency corresponding to the k-th intrinsic mode function uk(t), δ(t) is the Dirac function, which is often used as an idealized pulse signal in signal processing to construct the operation form of the VMD algorithm; j' is the imaginary unit; is the partial derivative with respect to time t; solve this optimization problem through the alternating direction multiplier method to achieve the separation of the signal and noise;

[0112] Then, perform wavelet threshold denoising on the noisy IMF components. Select the db4 wavelet basis function and adopt the Birge - Massart strategy for the threshold rule. The wavelet threshold denoising first performs wavelet transform on the noisy signal to obtain the wavelet coefficients w j ,k, and then processes the wavelet coefficients according to the threshold λ. The processing rule is Where, sign(w j (k) is a sign function, when w j When k>0, sign(w) j ,k)=1; when w j When k = 0, sign(w) j ,k)=0;w j When k < 0, sign(w) j ,k)=-1; Finally, the processed wavelet coefficients are subjected to inverse wavelet transform to reconstruct the signal;

[0113] Data preprocessing module 2 uses a max-min normalization algorithm to normalize the raw data, mapping parameters of different dimensions to the [0,1] interval. The calculation formula is as follows: Where, x norm The data is normalized, and x is the original data. max x min These are the maximum and minimum values ​​of the parameter in historical data, respectively.

[0114] Data preprocessing module 2 extracts key features reflecting the equipment status from the preprocessed data. These key features include time-domain features, frequency-domain features, and time-frequency-domain features. Time-domain features include mean, variance, peak value, peak-to-peak value, kurtosis, and margin factor. Frequency-domain features are obtained by converting the time-domain signal x(t) into the frequency-domain signal X(f) using Fast Fourier Transform (FFT), extracting feature frequency amplitude, spectral centroid, and spectral variance. The specific strategy is as follows: The FFT calculation formula is: Extract the characteristic frequency amplitude |X(f0)| from the frequency domain signal, where f0 is the characteristic frequency of equipment fault; the formula for calculating the centroid of the spectrum is: The formula for calculating the spectral variance is: Among them, f cg Let f be the centroid of the spectrum, f be the frequency variable, and |X(f)| be the amplitude corresponding to frequency f. For spectral variance;

[0115] Time-frequency domain features are extracted using wavelet packet transform (WPT) to obtain wavelet packet energy entropy and wavelet packet singular entropy. The wavelet packet energy entropy E... e The calculation formula is: in, E i Let S be the energy of the i-th node, M be the number of nodes, and S be the singular entropy of the wavelet packet. e The calculation formula is: in, s i represents the singular values ​​of the wavelet packet coefficients, and L is the number of singular values;

[0116] After feature extraction, a feature matrix with dimension n×m is formed, where n is the number of sampling time points and m is the number of features.

[0117] The status assessment module 3 assesses the operating status of the ship's power system in real time based on the preprocessed feature data, and determines whether the ship's power system is in a normal, sub-healthy or abnormal state.

[0118] The state assessment module 3 adopts a model that combines a stacked autoencoder with fuzzy comprehensive evaluation;

[0119] The stacked autoencoder (SAE) training strategy is as follows:

[0120] Network structure design: The number of nodes in the input layer is consistent with the dimension m of the feature matrix. A 3-5 hidden layer structure is constructed. The number of nodes in each hidden layer decreases and then increases proportionally, set to m / 2, m / 4, m / 8, m / 4, and m / 2 respectively. The number of nodes in the output layer is set to m to realize the reconstruction output of the feature vector.

[0121] By introducing an L2 regularization term and a sparsity constraint term into the SAE loss function, the improved loss function formula is as follows: in, For the input feature vector, To output the reconstructed feature vector, λ' is the regularization coefficient, and the value of λ' ranges from 0.001 to 0.01; w lj The weight matrix represents the weights of the j-th neuron in the l-th layer; L is the total number of hidden layers; s l ρ is the number of neurons in layer l; β is the sparsity coefficient, and the value of β ranges from 0.1 to 1; ρ is the expected sparsity. And a lj (x) represents the input x from the j-th neuron in the l-th layer. i Activation value; Let KL divergence be a metric.

[0122] The SAE model is pre-trained and fine-tuned using historical data under normal operating conditions of the ship's power system, and the sample size of the historical data is no less than 10,000 sets.

[0123] The specific strategy for fuzzy comprehensive evaluation is as follows:

[0124] Establish an evaluation index system: This includes the reconstruction error E of the SAE model and the normalized values ​​x of each parameter. norm,1 ,x norm,2 ,…,x norm,k' As evaluation indicators, an indicator set U =

[0125] {E,x norm,1 ,x norm,2 ,…,x norm,k'}, where k' is the number of parameters;

[0126] The Analytic Hierarchy Process (AHP) is used to determine the weights of each evaluation indicator. Specifically, the judgment matrix A = (a ij ) (k+1)×(k+1) a ij To determine the importance of indicator i relative to indicator j; then, calculate the largest eigenvalue λ of the judgment matrix. max And its corresponding eigenvector w; then perform a consistency check, and if the consistency ratio CR < 0.1, the judgment matrix is ​​considered to meet the consistency requirements, and the weight vector W = {w E ,w1,w2,…,w k}, where w E The weights for the reconstruction error;

[0127] Establish membership functions for three evaluation levels: "normal," "sub-healthy," and "abnormal." The membership functions use a triangular membership function. Specifically, let the evaluation level set V = {v1(normal), v2(sub-healthy), v3(abnormal)}. For an index u∈U, its membership function with respect to level v1 is: The membership function for level v2 is: The membership function for level v3 is: Where a1, b1, and b2 are thresholds determined based on historical data; the membership vector of the system state to each level is calculated through fuzzy synthesis operation. Where R is the membership matrix, R ij =μ j (u i ),and The final output health index is H = B1×1 + B2×0.65 + B3×0.25, where H ranges from 0 to 1, and 1, 0.65, and 0.25 are the baseline values ​​for the health levels of "normal", "sub-healthy", and "abnormal", respectively. When H ≥ 0.8, it is considered normal; 0.5 ≤ H ≤ 0.8, it is considered sub-healthy; and H < 0.5, it is considered abnormal.

[0128] The fault diagnosis module 4 is used to initiate fault diagnosis to identify the fault type, fault location and fault severity when the status assessment module 3 determines that the ship's power system is in a sub-healthy or abnormal state.

[0129] The fault early warning module 5 is used to predict the development trend of the fault after the fault diagnosis is completed, and the input terminal of the fault early warning module 5 is electrically connected to the output terminal of the fault diagnosis module 4.

[0130] The fault early warning module 5 includes a trend prediction unit 501 and a lifespan estimation unit 502. The trend prediction unit 501 uses a Long Short-Term Memory (LSTM) network model based on historical fault evolution data and current fault characteristics to predict the changing trend of fault characteristic parameters over a future period. The specific strategy of the trend prediction unit 501 is as follows:

[0131] The LSTM model is trained using complete time-series data covering the entire process from fault occurrence to deterioration, with each fault type having a time series length of at least 1000 time steps. The input to the LSTM model is the historical sequence X = {x} of the fault feature parameters. t-T'+1 ,x t-T'+2 ,...,x t}, where T' is the length of the input sequence, ranging from 30 to 50; the output is the predicted parameter values ​​for the next 100 time steps.

[0132] The number of hidden layer nodes is set to 64-128, and the root mean square error (RMSE) is used as the loss function, calculated as follows: Where M is the total number of prediction time steps, and the LSTM model is trained using the Adam optimizer;

[0133] The life estimation unit 502 uses the particle filter algorithm (PF) to predict the life of the power system, and the life prediction process is as follows:

[0134] The feature parameter sequence predicted by the LSTM model z, as an observation of the PF model t+1 ,z t+2 ,…,z t+T' The observation equation of the PF model is defined as: z t =W t +δ t , where δ t To conform to a normal distribution To mitigate observation noise, the following strategy is implemented:

[0135] Particle initialization: N initial particles are generated by randomly sampling from the prior distribution p(W0). And assign initial weights to each particle.

[0136] State prediction: Based on the state transition anti-dust stacking calculation, generate the particle set at time t.

[0137]

[0138] Weight Update: The particle weights are calculated based on the observation equations. The weight calculation formula is as follows: in is a probability density function, and After the calculation, normalize the particle weights. The normalization formula is

[0139] Resampling: Calculate the number of effective particles N eff , when N eff < N / 2, perform the resampling operation to generate a new particle set and reset the weights of all particles to 1 / N;

[0140] State estimation: Calculate the estimated value of the wear amount at time t by weighted summation. The formula is

[0141] Loop and execute the above steps until the predicted wear amount reaches the preset fatal failure threshold W th , and define the difference between the time point t fail when the threshold is reached and the current time t now as the remaining useful life RUL = t fail - t now .

[0142] The fault diagnosis module 4 adopts a fusion diagnosis algorithm of a convolutional neural network-long short-term memory network based on the attention mechanism, namely the CNN-LSTM-Attention model and the K-nearest neighbor model, namely the KNN. The CNN-LSTM-Attention model is used to extract spatial and temporal features from the feature data, and strengthen the features sensitive to faults through the attention mechanism. The specific strategy of the CNN-LSTM-Attention model is as follows:

[0143] The network structure is designed as follows:

[0144] Input layer: Receive the feature matrix where n represents the time step and m is the number of features;

[0145] CNN layer: Consists of 2-3 convolutional layers. Each convolutional layer uses a 3×3 convolutional kernel, and the number is set in the range of 32-64. The activation function uses ReLU. The output of the t-th convolutional layer is: C t =

[0146] ReLU(Conv(X t-1 , K t ) + b t ), where K t is the t-th convolutional kernel and b t is the bias term;

[0147] Pooling layer: Employs max pooling operation with a pooling kernel size of 2×2. Achieve dimensionality reduction;

[0148] LSTM layer: contains 1-2 LSTM units, with the number of hidden layer nodes ranging from 64 to 128;

[0149] Attention layer: An additive attention mechanism is used to calculate the hidden state weights at each time step, specifically: calculating the score: s t =v T tanh(W h h t +b), This is the weight matrix. The bias is d, where d is the attention dimension and h is the number of hidden points in the LSTM layer; then the weights are normalized using the formula: The normalized weights are then weighted by features, using the following formula: Among them, h t For the hidden layer node output of the LSTM at time step t, s t Let s be the score at time t. k The score at time k;

[0150] Fully connected layer: The number of output layer nodes is the same as the number of fault types K. The Softmax activation function is used to distribute the probability p of the output fault type as follows: p = Softmax(W fc h att +b fc ),in, This is the weight matrix. It is the bias vector;

[0151] The model training strategy is as follows:

[0152] Dataset construction: Use historical data of marine power systems, including but not limited to 10-15 types of faults such as main engine cylinder liner wear, lubricating oil line leakage, cooling system blockage, and propeller bearing damage, with no less than 500 samples for each type of fault;

[0153] Data partitioning: The training set and the test set are partitioned in a 7:3 ratio;

[0154] Optimization method: Use cross-entropy loss function Among them, y ik p is the true label for fault type k corresponding to sample i. ik The model is trained using the Adam optimizer to predict probabilities.

[0155] The specific strategy of the KNN model is as follows:

[0156] Feature weights and distance weighting are introduced, and the feature weight vector w = {w1, w2, ..., w...} is determined based on the sensitivity of features to faults. m}, calculate the difference between the test sample x and the training sample x. i Weighted Euclidean distance Neighbor samples are weighted by the inverse distance, with the weights... k' is the nearest neighbor number.

[0157] The early warning decision module 6 generates early warning information of different levels based on the status assessment results, fault diagnosis results and RUL prediction results, and provides targeted operation and maintenance decision suggestions, which are pushed to the ship's cockpit, shore-based operation and maintenance center and mobile terminal of operation and maintenance personnel.

[0158] The different levels of early warning information in the early warning decision module 6 are divided into:

[0159] Level 1 warning indicates a normal state: the health index H ≥ 0.8, indicating that the system is operating without faults, and a system operation status report is pushed out;

[0160] Level 2 alert is a state of concern: the health index is 0.5≤H≤0.8, indicating that the system has no obvious faults, but there are potential risks and hidden dangers, and a suggestion to pay attention is pushed.

[0161] Level 3 warning is a warning state: the health index H < 0.5, indicating that the system has identified a minor fault, the remaining service life (RUL) is > 7 days, and a warning and maintenance preparation suggestions are pushed out;

[0162] Level 4 warning indicates an emergency: This means the system has detected a moderate or severe fault, and the remaining service life is...

[0163] If the RUL is less than 7 days, an emergency warning and emergency response suggestions will be sent.

[0164] The strategy for generating operation and maintenance decision recommendations is as follows: based on the fault type and the real-time navigation status of the vessel, which includes key parameters such as geographical location (Lat / Lon), planned route (Route), and current cargo load (Load), a preset operation and maintenance decision knowledge base module is invoked. This knowledge base stores integrated fault handling standard procedures, spare parts reserve lists, and maintenance process specifications. The strategy uses case-based reasoning, i.e., the CBR fusion algorithm, to generate operation and maintenance decision recommendations.

[0165] The data storage module 7 is used to store data from various modules within the system, and the data storage module 7 uses the time-series database InfluxDB and the relational database MySQL.

[0166] The following data is stored in data storage module 7:

[0167] Raw data: Stores the raw parameter values ​​collected by the sensor, indexed by timestamp;

[0168] Preprocessed data and feature matrix: InfluxDB stores the preprocessed parameter time series, MySQL stores the feature matrix, and the table structure of the feature matrix includes "feature ID, device number, timestamp, time domain feature, frequency domain feature, and time-frequency domain feature".

[0169] The results of condition assessment, fault diagnosis, fault prediction, and early warning decision-making are stored in MySQL, with the following table structures:

[0170] Status assessment results table: Assessment ID, Equipment number, Assessment time, Health index H, System status, Reconfiguration error E;

[0171] Fault Diagnosis Result Table: Diagnosis ID, Device Number, Diagnosis Time, Fault Type, Fault Location, Fault Severity, CNN-LSTM-Attention Confidence, KNN Confidence, Final Confidence;

[0172] Fault Prediction Results Table: Prediction ID, Equipment Number, Prediction Time, Fault Characteristic Parameter Prediction Sequence, RUL, Prediction Error Range;

[0173] Early warning decision results table: Early warning ID, device number, early warning time, early warning level, early warning information, operation and maintenance suggestions, and push recipients;

[0174] Historical fault data and maintenance records of marine power systems: MySQL stores historical fault data and maintenance records. The historical fault data includes: fault ID, equipment number, fault occurrence time, fault type, fault cause, handling process, and maintenance result. The maintenance records include: maintenance ID, equipment number, maintenance time, maintenance content, spare parts usage, and maintenance personnel.

[0175] Please see Figure 2 A method for monitoring and early warning of a ship's power system, the method comprising the following steps:

[0176] S1. Data Acquisition: Start the multi-source data acquisition module 1 to collect the operating parameters, environmental parameters and status parameters of the ship's power system in real time through a distributed sensor network. The acquisition frequency is set according to the parameter type.

[0177] S2. Data Preprocessing: The raw data is read from the InfluxDB data storage module 7 by the data preprocessing module 2 according to a fixed time window, and then cleaned, denoised, normalized, and feature extracted.

[0178] S3. State Assessment: The state assessment module 3 reads the feature matrix of the current time window from MySQL, inputs it into the trained SAE model, calculates the reconstruction error; constructs the index set U and weight vector W of fuzzy comprehensive evaluation, calculates the membership degree of each index to "normal", "sub-healthy" and "abnormal", obtains the membership degree vector B through the weighted average method, calculates the health index, and judges the system state; the assessment results are stored in the state assessment result table of MySQL.

[0179] S4. Fault Diagnosis: If the status assessment result is "sub-healthy" or "abnormal", the fault diagnosis module 4 reads the feature matrix from MySQL, inputs it into the CNN-LSTM-Attention model, obtains the fault type probability distribution p, and calculates the maximum confidence.

[0180] S5. Fault Prediction: After the fault diagnosis is completed, the fault early warning module 5 reads the historical sequence of fault characteristic parameters of the device and the current diagnosis results from MySQL, inputs them into the LSTM model, predicts the trend of characteristic parameters for the next 7 days, and stores the prediction results in the fault prediction result table of MySQL.

[0181] S6. Early Warning Decision-Making: The early warning decision-making module 6 reads the status assessment result H, fault diagnosis result, and fault prediction result (RUL) from MySQL, and then calls the operation and maintenance decision knowledge base to store the early warning information and suggestions.

[0182] The MySQL early warning decision results table is pushed to the ship's cockpit display terminal, the shore-based operation and maintenance center management platform, and the operation and maintenance personnel's mobile APP via the MQTT protocol;

[0183] S7. Model Iteration and Optimization: An iteration cycle is conducted every 3 months. New historical fault data and maintenance records are read from MySQL to update the models of each module.

[0184] State Assessment (SAE) Model: Add 5000 sets of normal operation data, retrain the model, and optimize λ and β in the loss function;

[0185] Fault diagnosis CNN-LSTM-Attention model: Add new fault data with 200 sets of each fault, expand the training set, retrain the network, and adjust the number of convolutional kernels and the number of LSTM hidden layer nodes;

[0186] Fault prediction LSTM and PF models: Add fault evolution data with 300 time steps, retrain the LSTM model, and optimize the process noise and observation noise variance of the PF model; the iterated model replaces the original model.

[0187] The above description is merely a preferred embodiment of this application and is not intended to limit this application. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of this application should be included within the protection scope of this application.

Claims

1. A fault monitoring system of a marine power system, comprising a multi-source data acquisition module (1), a data preprocessing module (2), a state evaluation module (3), a fault diagnosis module (4), a fault warning module (5), a warning decision module (6) and a data storage module (7), characterized in that: the multi-source data acquisition module (1) acquires the running parameters, environmental parameters and equipment state parameters of the marine power system in real time through acquisition equipment; the input end of the data preprocessing module (2) is electrically connected with the output end of the multi-source data acquisition module (1), and the data preprocessing module (2) is used for cleaning, denoising, normalizing and feature extraction of the original data collected by the multi-source data acquisition module (1); the state evaluation module (3) evaluates the running state of the marine power system in real time based on the preprocessed feature data, and judges whether the marine power system is in a normal, sub-health or abnormal state; the fault diagnosis module (4) is used to identify the fault type, fault location and fault severity when the state evaluation module (3) determines that the marine power system is in a sub-health or abnormal state; the fault warning module (5) is used to predict the development trend of the fault after the fault diagnosis is completed, and the input end of the fault warning module (5) is electrically connected with the output end of the fault diagnosis module (4); the warning decision module (6) generates different levels of warning information according to the state evaluation result, the fault diagnosis result and the RUL prediction result, and gives targeted operation and maintenance decision suggestions, which are pushed to the ship bridge, the shore operation and maintenance center and the operation and maintenance personnel mobile terminal; the data storage module (7) is used for storing the data of each module in the system, and the data storage module (7) adopts a time series database InfluxDB and a relational database MySQL.

2. A fault monitoring system of a marine power system according to claim 1, characterized in that, The running parameters collected in the multi-source data acquisition module (1) include: main engine speed, main engine power, fuel consumption, lubricating oil pressure, lubricating oil temperature, cooling water temperature, exhaust gas temperature, cylinder pressure, transmission device speed, and propeller torque; The environmental parameters include: sea water temperature, sea water salinity, atmospheric temperature, atmospheric humidity, and wind wave level of the sea area where the ship sails; The state parameters include: equipment vibration acceleration, equipment surface temperature, ultrasonic signal of pipeline leakage, motor current / voltage, and valve switch state; The acquisition equipment is a distributed sensor network based on Internet of Things, and the multiple sensors include: vibration sensor, thermocouple temperature sensor, pressure sensor, ultrasonic leakage sensor, current transformer, and GPS positioning module.

3. The fault monitoring system of a marine power system according to claim 1, characterized by, In the data preprocessing module (2), the mixed algorithm of 3σ criterion and isolation forest is used for outlier rejection, and the specific strategy is: the 3σ criterion is used to preliminarily reject the outliers outside the reasonable range, and the 3σ criterion determines the normal data range as [μ-3σ, μ+3σ] according to the mean μ and standard deviation σ, that is, when the data x satisfies |x-μ|>3σ, it is determined as an obvious outlier and is rejected. Then the residual data is detected by the Isolation Forest algorithm, which builds multiple Isolation Trees, calculates the path length h(x) of each data point in the tree, and compares it with the expected path length E[h(x)]. When h(x) < E[h(x)], it is determined to be a hidden outlier and is removed. The data preprocessing module (2) uses a combination of variational mode decomposition and wavelet threshold denoising algorithm when denoising the original data. The specific strategy is: First, set the number of modes in the variational mode decomposition algorithm, namely VMD algorithm, to be 3-8, the penalty parameter α=2000, and the original signal f(t) is decomposed into a plurality of intrinsic mode functions uk(t). The objective function of the VMD algorithm is The constraint condition is ∑ k u k (t)=f(t), where w k is the center frequency corresponding to the kth intrinsic mode function uk(t), δ(t) is the Dirac function, which is often used as an idealized pulse signal in signal processing, and is used to construct the operation form of the VMD algorithm; j' is the imaginary unit; is the partial derivative with respect to time t; the optimization problem is solved by the alternating direction multiplier method, and the signal and noise are separated. The noisy IMF component is denoised by using wavelet threshold, the wavelet base function is selected as db4, the threshold rule is selected as Birge-Massart strategy, the wavelet threshold denoising is performed on the noisy signal to obtain wavelet coefficients w j ,k, the wavelet coefficients are processed according to the threshold λ, and the processing rule is wherein, sign(w j ,k) is a sign function, when w j ,k>0, sign(w j ,k)=1; when w j ,k=0, sign(w j ,k)=0; when w j ,k<0, sign(w j ,k)=-1; finally, the processed wavelet coefficients are inverse wavelet transformed to reconstruct the signal; The data preprocessing module (2) adopts the maximum-minimum normalization algorithm when normalizing the original data, maps parameters of different dimensions to the interval [0, 1], and the calculation formula is: Wherein, x norm is the normalized data, x is the original data, x max and x min are the maximum and minimum values of the parameter in the historical data, respectively.

4. A fault monitoring system for a marine power system according to claim 3, characterised in that, The data preprocessing module (2) extracts key features reflecting the device state from the preprocessed data, including time domain features, frequency domain features, and time-frequency domain features, wherein the time domain features include mean, variance, peak value, peak-to-peak value, kurtosis, and margin factor; the frequency domain features are converted from the time domain signal x(t) to the frequency domain signal X(f) by fast Fourier transform (FFT), and the characteristic frequency amplitude, spectral centroid, and spectral variance are extracted, and the specific strategy is that the calculation formula of FFT is: The characteristic frequency amplitude |X(f0)| is extracted from the frequency domain signal, wherein f0 is the device fault characteristic frequency; the calculation formula of the spectral centroid is: The calculation formula of the spectral variance is: Wherein, f cg is the spectral centroid, f is the frequency variable, |X(f)| is the amplitude corresponding to the frequency f, is the spectral variance; The time-frequency domain features are extracted by wavelet packet transform (WPT) to obtain wavelet packet energy entropy and wavelet packet singular entropy e The calculation formula of the wavelet packet energy entropy E is as follows: E i is the energy of the i th node, and M is the number of nodes; the calculation formula of the wavelet packet singular entropy S e is as follows: is as follows: s i is a singular value of a wavelet packet coefficient, and L is the number of singular values. After feature extraction, a feature matrix with dimensions n x m is formed, where n is the number of sampling time points and m is the number of features.

5. The fault monitoring system of a marine power system according to claim 1, characterized by, The state evaluation module (3) uses a model combining stack autoencoder and fuzzy comprehensive evaluation; The stack autoencoder (SAE) training strategy is: Network structure design: The number of input layer nodes is consistent with the dimension m of the feature matrix. A 3-5 layer hidden layer structure is constructed, with the number of nodes in each hidden layer decreasing and then increasing by a certain proportion, set as m / 2, m / 4, m / 8, m / 4, m / 2 in turn. The number of output layer nodes is set to m to realize the reconstruction output of the feature vector. On the basis of SAE loss function, L2 regularization term and sparse constraint term are introduced, and the improved loss function formula is: Wherein, is the input feature vector, is the output reconstructed feature vector, λ' is the regularization coefficient, and λ' takes the value range of 0.001-0.01; w lj is the weight matrix, that is, the weight of the jth neuron in the lth layer; L is the total number of hidden layers; s l is the number of neurons in the lth layer; β is the sparse coefficient, and the value range of β is 0.1-1; ρ is the expected sparsity; And a lj (x) is the activation value of the jth neuron in the lth layer to the input x i ; is the KL divergence; The SAE model is pre-trained and fine-tuned using historical data of the ship power system under normal operating conditions, and the sample size of the historical data is not less than 10,000 groups. The specific strategy of fuzzy comprehensive evaluation is: Establishing evaluation index system: the reconstruction error E of SAE model, the normalized value x norm,1 ,x norm,2 ,…,x norm,k' as evaluation index, form index set U = {E,x norm,1 ,x norm,2 ,…,x norm,k'}, where k' is the number of parameters; The weight of each evaluation index is determined by using analytic hierarchy process (AHP), specifically: first, a judgment matrix A=(a ij ) (k+1)×(k+1) , a ij is the importance of index i relative to index j; then the maximum eigenvalue λ max of the judgment matrix and the corresponding eigenvector w are calculated; then consistency test is carried out, and the consistency ratio CR<0.1 is considered to meet the requirements of the consistency of the judgment matrix, and the weight vector W={w E ,w1,w2,…,w k} is obtained, wherein w E is the weight of reconstruction error; The membership functions of the three evaluation levels of "normal", "sub-health" and "abnormal" are established, and the membership functions are triangular membership functions. Specifically, assuming that the evaluation level set V = {v1 (normal), v2 (sub-health), v3 (abnormal)}, for an index u ∈ U, the membership function of the level v1 is: The membership function of the level v2 is: The membership function of the level v3 is: Wherein, a1, b1, b2 are threshold values determined according to historical data; the membership degree vector of the system state to each level B = W°R is calculated through fuzzy synthesis operation, wherein R is a membership matrix, R ij = μ j (u i ), and Finally, the state health index H = B1×1+B2×0.65+B3×0.25 is output, wherein the value range of H is 0-1, 1, 0.65 and 0.25 are respectively the health degree benchmark values corresponding to the levels of "normal", "sub-health" and "abnormal"; when H ≥ 0.8, it is normal, 0.5 ≤ H ≤ 0.8 is sub-health, and H < 0.5 is abnormal.

6. The fault monitoring system of a marine power system according to claim 1, characterized by, The fault diagnosis module (4) uses a fusion diagnosis algorithm based on the convolutional neural network-long short-term memory network-attention mechanism (CNN-LSTM-Attention) model and the K-nearest neighbor (KNN) model. The CNN-LSTM-Attention model is used to extract spatial and temporal features from the feature data and strengthen the features sensitive to faults through the attention mechanism. The specific strategy of the CNN-LSTM-Attention model is: Network structure design: Input layer: receives feature matrix where n represents time step, m is the number of features; CNN layer: composed of 2-3 convolutional layers, each convolutional layer uses a 3x3 size convolution kernel, the number is set to 32-64 intervals, the activation function uses ReLU, the output of the tth convolutional layer is: t = ReLU(Conv(X t-1 , K t )+b t ), where K t is the tth layer convolution kernel, and b t is the bias term; Pooling layer: the maximum pooling operation is adopted, and the pooling kernel size is 2x2, so that dimensionality reduction is realized; LSTM layer: contains 1-2 LSTM units, with the number of hidden layer nodes ranging from 64 to 128; Attention layer: the additive attention mechanism is used to calculate the weight of the time step hidden state, specifically: calculate the score: s t = v T tanh(W h h t + b), is the weight matrix, is the bias, d is the attention dimension, h is the number of LSTM hidden layer points; then normalize the weight, the formula is: a t = Feature weighting is performed on the normalized weight, the formula is: where, is the hidden layer node output of the LSTM at time step t, s t is the score at time t, s k is the score at time k; Fully connected layer: the number of output layer nodes is consistent with the number of fault types K, and the Softmax activation function is used to output the fault type probability p distribution: p = Softmax(W fc h att +b fc ), wherein, is a weight matrix, is a bias vector; Model training strategy: Data set construction: use historical data of the ship power system including but not limited to 10-15 types of faults such as main engine cylinder wear, lubricating oil pipeline leakage, cooling system blockage, and propeller bearing damage, with a sample size of not less than 500 groups for each fault; Data division: divide the training set and test set in a 7:3 ratio; Optimization method: cross-entropy loss function is adopted where y ik is the true label of sample i corresponding to fault type k, p ik is the predicted probability of the model, and the Adam optimizer is used for training; The specific strategy of the KNN model is: The feature weight and distance weight are introduced, the feature weight vector w = {w1, w2,..., wn} is determined according to the sensitivity of the feature to the fault, the weighted Euclidean distance between the test sample x and the training sample x m i is calculated The distance reciprocal weight is adopted for the adjacent sample, the weight k' is the number of neighbors.​ 7. The fault monitoring system of a marine power system according to claim 1, characterized by, The fault warning module (5) includes a trend prediction unit (501) and a life estimation unit (502). The trend prediction unit (501) uses a long short-term memory network (LSTM) model to predict the trend of fault feature parameters in the future based on historical fault evolution data and current fault features. The specific strategy of the trend prediction unit (501) is: The LSTM model is trained using complete time series data covering the whole process from the occurrence to the deterioration of the fault, and the time series length of each fault type is not less than 1000 time steps. The input of the LSTM model is the historical sequence X = {x t-T'+1 ,x t-T'+2 ,...,x t} of the fault characteristic parameters, wherein T' is the input sequence length, and the value range is 30-50; the output is the parameter prediction value of the future 100 time steps The number of hidden layer nodes is set to 64-128, the root mean square error (RMSE) is used as the loss function, and the calculation formula is: Wherein, M is the total number of prediction time steps, and the LSTM model is trained by the Adam optimizer. The life estimation unit (502) uses a particle filter (PF) algorithm to predict the life of the power system. The life prediction process is: The feature parameter sequence predicted by the LSTM model The observation value z of the PF model t+1 ,z t+2 ,…,z t+T' , The observation equation of the PF model is defined as: z t =W t +δ t , wherein δ t is observation noise conforming to a normal distribution , and the execution strategy is: Particle initialization: Generate N initial particles by random sampling from the prior distribution p(W0) and assign each particle an initial weight State prediction: generate the particle set at time t according to the state transition distribution Weight update: The particle weight is calculated according to the observation equation, and the weight calculation formula is wherein is a probability density function, and After the calculation, the particle weight is normalized, and the normalization formula is Resampling: Compute effective number of particles N eff When N eff Perform resampling operation to generate new set of particles and reset all particles' weights to 1 / N; State estimation: the estimation value of the wear amount at time t is calculated by weighted summation, and the formula is The above steps are executed in a loop until a predicted amount of wear reaches a pre-set fatal failure threshold W th The point in time t fail at which the threshold is reached is defined as the remaining useful life RUL = t now - t fail - t now .

8. The fault monitoring system of a marine power system according to claim 1, characterized by, The different levels of warning information in the warning decision module (6) are: Level 1 warning: normal state: state health index H >= 0.8, indicating that the system is in a fault-free operating state, and the system operating state report is pushed. The second level warning is a state of attention: the state health index 0.5<=H<=0.8, indicating that the system has no clear failure, but there are potential risks and hidden dangers, and the attention suggestion is pushed; The third level warning is a warning state: the state health index H<0.5, indicating that the system is identified as a mild failure, and the remaining useful life RUL>7 days, and the warning and maintenance preparation suggestions are pushed; The fourth level warning is an emergency state: indicating that the system detects moderate or high failure, and the remaining useful life RUL<7 days, and the emergency warning and emergency treatment suggestions are pushed; The operation and maintenance decision suggestion generation strategy is: based on the fault type and the real-time sailing state of the ship, and the real-time sailing state of the ship includes the key parameters of geographical position Lat / Lon, planned route Route and current load Load, the preset operation and maintenance decision knowledge base is called, and the operation and maintenance decision knowledge base stores: integrated fault handling standard process, spare parts reserve list and maintenance process specification, and the operation and maintenance decision suggestion is generated by using the case-based reasoning (CBR) fusion algorithm.

9. The fault monitoring system of a marine power system according to claim 1, characterized by, The following data is stored in the data storage module (7): Raw collected data: indexed by timestamp, storing the original parameter values collected by the sensor; Preprocessed data and feature matrix: InfluxDB stores the preprocessed parameter time series, and MySQL stores the feature matrix, and the table structure of the feature matrix includes "feature ID, device number, timestamp, time domain feature, frequency domain feature, time-frequency domain feature"; The result data of state evaluation, fault diagnosis, fault prediction and warning decision: MySQL stores the result data, and the table structure is respectively: State evaluation result table: evaluation ID, device number, evaluation time, health index H, system state, reconstruction error E; Fault diagnosis result table: diagnosis ID, device number, diagnosis time, fault type, fault location, fault severity, CNN-LSTM-Attention confidence, KNN confidence, final confidence; Fault prediction result table: prediction ID, device number, prediction time, fault feature parameter prediction sequence, RUL, prediction error range; Warning decision result table: warning ID, device number, warning time, warning level, warning information, operation and maintenance suggestion, push object; Ship power system historical fault data and operation and maintenance records: MySQL stores the historical fault data and operation and maintenance records, and the historical fault data includes: fault ID, device number, fault occurrence time, fault type, fault reason, processing process, repair result; the operation and maintenance records include: operation and maintenance ID, device number, operation and maintenance time, operation and maintenance content, spare parts usage, operation and maintenance personnel.

10. A method of monitoring and alerting for a marine power system, characterized by, The method comprises the following steps: S1. Data collection, start the multi-source data collection module (1), and collect the running parameters, environmental parameters and state parameters of the ship power system in real time through the distributed sensor network, and the collection frequency is set according to the parameter type; S2. Data preprocessing: the data preprocessing module (2) reads the original data of the InfluxDB of the data storage module (7) according to the fixed time window for cleaning, denoising, normalization and feature extraction S3. State assessment: The state assessment module (3) reads the feature matrix of the current time window from MySQL, inputs the trained SAE model, and calculates the reconstruction error; the index set U and weight vector W of fuzzy comprehensive evaluation are constructed, the membership degrees of each index to "normal", "sub-health" and "abnormal" are calculated, the membership degree vector B is obtained by weighted average method, the health index is calculated, and the system state is judged; the evaluation results are stored in the state assessment result table of MySQL; S4. Fault diagnosis: If the state assessment result is "sub-health" or "abnormal", the fault diagnosis module (4) reads the feature matrix from MySQL, inputs the CNN-LSTM-Attention model, obtains the fault type probability distribution p, and calculates the maximum confidence; S5. Fault prediction: After fault diagnosis, the fault warning module (5) reads the historical sequence of fault feature parameters and the current diagnosis result of the device from MySQL, inputs the LSTM model, predicts the trend of feature parameters in the next 7 days, and the prediction result is stored in the fault prediction result table of MySQL; S6. Early warning decision: The early warning decision module (6) reads the state assessment result H, fault diagnosis result and fault prediction result RUL from MySQL, then calls the operation and maintenance decision knowledge base, stores the early warning information and suggestions in the early warning decision result table of MySQL, and pushes it to the ship bridge display terminal, shore-based operation and maintenance center management platform and operation and maintenance personnel mobile APP through MQTT protocol; S7. Model iteration optimization: every 3 months is an iteration cycle, read the new historical fault data and operation records from MySQL, update the models of each module: State assessment SAE model: add 5000 groups of normal operation data, retrain the model, and optimize λ and β in the loss function; Fault diagnosis CNN-LSTM-Attention model: add 200 groups of fault data for each fault, expand the training set, retrain the network, and adjust the number of convolution kernels and the number of LSTM hidden layer nodes; Fault prediction LSTM and PF model: add 300 time steps of fault evolution data, retrain the LSTM model, and optimize the process noise and observation noise variance of the PF model; replace the original model with the iterated model.

Images