Marine main engine-oriented state monitoring and diagnosing method and system

Abstract

The invention discloses a state monitoring and diagnosing method and system for a marine main engine, and the method comprises the steps: collecting the operation state parameters of the marine main engine, and carrying out the preprocessing; embedding a dynamic soft thresholding layer in each residual module of the residual shrinkage network, and executing channel-by-channel soft thresholding on the feature map; constructing a double-flow attention network, fusing the time-frequency domain characteristics of the vibration signals and the temperature field data of the space heat conduction model, and capturing a causal sequential relationship of multi-modal data through a gating mechanism; carrying out embedded optimization on the model by adopting depth separable convolution, adaptive load normalization and knowledge distillation; the model is carried on a shipborne industrial personal computer, and whether model updating needs to be triggered or not is judged by calculating the weight difference between the shipborne model and the global model of the cloud platform. According to the method, a whole-process closed loop from data acquisition to model deployment is realized, the anti-noise capability, causal association analysis, resource optimization and privacy protection are integrated, and the intelligent operation and maintenance requirements of the marine main engine are met.

Description

Technical Field

[0001] This invention relates to the field of ship main engine condition monitoring technology, and in particular to a method and system for condition monitoring and diagnosis of ship main engines. Background Technology

[0002] As the core component of a ship's propulsion system, the operating status of the main engine directly affects navigation safety and maintenance costs. Traditional main engine condition monitoring technology mainly relies on data from a single sensor, using threshold comparisons or simple statistical models to achieve fault early warning. However, with the increasing size and intelligence of ships, traditional methods face a contradiction between noise immunity and fault sensitivity under complex operating conditions. The operating conditions of ship main engines are highly variable, and traditional fixed-threshold noise suppression methods are difficult to adapt to dynamic changes in operating conditions. Under low-load conditions, weak fault characteristics are easily masked by noise; under high-load conditions, random vibration interference may lead to false alarms.

[0003] Existing technologies often employ data-level or decision-level fusion, but neglect the causal temporal relationships between different modalities. For example, the vibration frequency shift caused by bolt loosening can lag and lead to a local temperature increase. Traditional methods struggle to capture such cross-modal causal relationships, resulting in low efficiency in tracing the root cause of faults. General-purpose deep learning models perform well on public datasets, but they often fail to adapt to shipboard scenarios and cannot meet the requirements. Furthermore, model updates rely on manual annotation and offline training, with weekly updates being too frequent and impacting ship-shore collaboration efficiency. Summary of the Invention

[0004] In order to overcome the above-mentioned problems in the prior art, the present invention proposes a condition monitoring and diagnosis method and system for ship main engines.

[0005] The technical solution adopted by this invention to solve its technical problem is: a condition monitoring and diagnosis method for ship main engines, comprising the following steps: Step 1: Collect the ship's main engine operating status parameters and perform preprocessing; Step 2, construct the dynamic soft thresholding layer: embed the dynamic soft thresholding layer in each residual module of the residual shrinking network to perform channel-by-channel soft thresholding on the feature map; Step 3: Construct a dual-stream attention network, which integrates the time-frequency domain features of vibration signals with the temperature field data of the spatial heat conduction model, and captures the causal temporal relationships of multimodal data through a gating mechanism; Step 4: The model from Step 3 is embedded and optimized using depthwise separable convolution, adaptive load normalization, and knowledge distillation to obtain the state monitoring and diagnosis model. Step 5: The status monitoring and diagnostic model obtained in Step 4 is mounted on the shipborne industrial control computer. The weight difference between the shipborne model and the global model on the cloud platform is calculated to determine whether a model update needs to be triggered. After receiving the update package, the shipborne node fine-tunes the model with local data to adapt to individual differences of the host. In step 2, the dynamic soft thresholding layer drives the threshold through the running state parameters obtained in step 1.

[0006] The above-mentioned condition monitoring and diagnosis method for ship main engines includes, in step 1, data preprocessing specifically including: based on data alignment and synchronization technology, timing alignment of vibration signals, temperature field data and operating parameters is achieved through the PTP clock protocol; and the adaptive filter NLMS method is used to suppress high-frequency random noise and retain fault characteristic frequency bands.

[0007] In the aforementioned method for condition monitoring and diagnosis of ship main engines, the specific formula for step 2, soft thresholding, is as follows: in, Soft threshold, It is the original value of the input feature map at the c-th channel and spatial position (h,w); The sign function preserves the positive and negative information of the feature values; θc(t) is the dynamic soft threshold of the c-th channel; For soft thresholding, eigenvalues ​​with absolute values ​​less than θc(t) are set to zero, while eigenvalues ​​with absolute values ​​greater than θc(t) are retained and the threshold is subtracted.

[0008] The above-mentioned method for condition monitoring and diagnosis of ship main engines, wherein the specific calculation method for the dynamic soft threshold is as follows: Where θ(t) is the dynamic soft threshold, α is the load factor, Load(t) is the real-time load of the main unit, β is the speed factor, γ is the fuel quality factor, FuelQuality(t) is the real-time fuel quality index, δ is the fuel quality decay rate, and t is the main unit running time.

[0009] The above-mentioned condition monitoring and diagnostic method for ship main engines, specifically includes step 3 as follows: Step 3.1: Construct a vibration main branch one-dimensional convolutional neural network for extracting time-frequency domain features of vibration signals and a temperature auxiliary branch graph convolutional neural network for modeling the spatial heat conduction path of the temperature field. Step 3.2, through timing alignment Δ t Capturing the vibration-temperature cross-modal causal relationship, a gating mechanism dynamically adjusts the contributions of the two modes: G (t ) is the gate weight matrix, used to control the fusion ratio of vibration and temperature characteristics; s Use the Sigmoid activation function; Wg The weight matrix is ​​a learnable weight matrix; VibFeature ( t ) represents the main branch of vibration in time t The characteristic output; TempFeature ( t −Δ t ( ) is a temperature-dependent secondary branch in time t −Δ t The characteristic output; Δt is the heat conduction delay time; bg is the learnable bias term; Step 3.3, use classification loss. L_class Compared with loss L_contrast Cooperative joint loss function.

[0010] In the aforementioned method for condition monitoring and diagnosis of ship main engines, the joint loss function is specifically as follows: L_total For the total loss of the function, L_class Cross-entropy loss is used in classification loss. L_contrast Enhance contrast loss to improve cross-modal feature consistency; y true The one-hot encoded vector of the real label; y pred This is the predicted probability vector output by the model after Softmax activation; V vib Time-frequency features extracted from the main vibration branch; V temp Spatial heat conduction characteristics extracted by temperature auxiliary branch; V temp_neg Temperature characteristics of other samples within the batch; is the cosine similarity function; exp is the exponential function.

[0011] In the aforementioned method for condition monitoring and diagnosis of ship main engines, the specific calculation method for the weight difference between the shipborne model and the cloud platform global model in step 5 is as follows: Among them, the weight difference index of the D model, The current model weights of the shipborne edge nodes. Global model weights of the shore-based cloud platform.

[0012] In the aforementioned method for monitoring and diagnosing the condition of a ship's main engine, step 5 involves fine-tuning the cross-entropy loss in the classification loss using local data and updating the weights of the shipborne model. The specific weight update formula is as follows: in, The weights of the shipborne model after fine-tuning The shipboard model weights before fine-tuning or To control the step size of weight updates for the learning rate, ∇ L CE This is the gradient of the loss function with respect to the model weights.

[0013] A condition monitoring and diagnostic system for ship main engines, based on the aforementioned condition monitoring and diagnostic method for ship main engines, includes a multimodal data acquisition and processing layer, a dynamic soft thresholding layer, a cross-modal attention fusion network layer, a lightweight residual shrinking unit optimization layer, and a federated learning incremental update deployment layer. The multimodal data acquisition and processing layer includes a sensor network deployed on the host computer for acquiring host operating status data and preprocessing the acquired data; the dynamic soft thresholding layer drives the threshold through state parameters to achieve a dynamic balance between noise suppression and fault sensitivity; the cross-modal attention fusion network layer fuses the time-frequency domain features of vibration signals with the temperature field data of the spatial heat conduction model, captures the causal temporal relationship of multimodal data through a gating mechanism, designs a joint loss function, and emphasizes the causal relationship between vibration and temperature features in the hidden layer space while ensuring accurate identification of fault types; the lightweight residual shrinkage unit optimization layer optimizes the model through depthwise separable convolution, adaptive load normalization, and knowledge distillation; the federated learning incremental update deployment layer is used to deploy the model on the shipborne industrial control computer and update the shipborne model in real time.

[0014] The beneficial effects of this invention are as follows: It achieves multi-dimensional technological breakthroughs in ship main engine condition monitoring and diagnosis through a triple-customized improved deep residual shrinkage network. The dynamic soft thresholding layer dynamically adjusts the noise suppression threshold based on real-time operating parameters, enhancing the extraction capability of weak fault features under low-load conditions and improving anti-interference performance under high-load conditions. The cross-modal causal association attention fusion mechanism accurately captures the "vibration-temperature" cross-modal causal relationship by combining a two-stream network (vibration time-frequency features and a temperature spatial heat conduction model) with gated temporal alignment. The lightweight residual shrinkage unit employs deep separable convolution, adaptive load normalization, and knowledge distillation techniques. The federated learning incremental update mechanism, through a ship-shore collaborative training framework, combines differential privacy protection and model difference evaluation to achieve efficient and secure incremental learning while protecting data privacy. Attached Figure Description

[0015] Figure 1 This is a flowchart of the invention. Figure 2 This is a structural diagram of the condition monitoring and diagnostic system of the present invention; Figure 3 This is a comparison chart of loss deviations in the condition monitoring and diagnosis method of this invention. Detailed Implementation

[0016] To enable those skilled in the art to better understand the technical solution of the present invention, the present invention will be described in detail below with reference to the accompanying drawings and specific embodiments.

[0017] like Figure 1 As shown in the figure, this embodiment discloses a condition monitoring and diagnostic method for ship main engines, including the following steps: Step 1: Collect the operating status parameters of the ship's main engine and perform preprocessing.

[0018] 1.1 By deploying a sensor network, vibration acceleration sensors and temperature sensor arrays are installed in key parts of the host to synchronously collect operating parameters such as load, speed, and fuel quality.

[0019] 1.2 Based on data alignment and synchronization technology, the timing alignment of vibration signals with a sampling rate of 10kHz, temperature field data with a sampling rate of 1Hz, and operating parameters is achieved through the PTP clock protocol, with an error ≤1ms.

[0020] 1.3 The adaptive filter NLMS method is used to suppress high-frequency random noise and retain the 0-2kHz fault characteristic frequency band.

[0021] Step 2, construct the dynamic soft thresholding layer: embed the dynamic soft thresholding layer in each residual module of the residual shrinking network to perform channel-by-channel soft thresholding on the feature map.

[0022] 2.1 Threshold Parameter Calculation Host status parameters are passed through dynamic thresholds. i ( t This achieves real-time nonlinear coupling drive of operating parameters, as shown in the following formula: in, i ( t () is a dynamic soft threshold, which varies with time. t And changes in operating parameters are used to control the characteristic map. F The suppression strength of noise components in the middle. α The load factor controls the sensitivity to subtle fault characteristics under low load conditions. Load ( t This represents the real-time load of the host computer, ranging from 0-100%. For example, 10% of the rated load indicates a low load.α ⋅ Load ( t )leading i ( t Lower the threshold to retain low-amplitude fault signals. β This is the speed coefficient, used to balance vibration suppression under high-speed operating conditions. The unit of speed is... RPM (rpm), such as at 90% of the rated speed at high speed, β ⋅ RPM ( t )leading i ( t Increase the threshold to suppress high-frequency noise. c The fuel quality coefficient is used to simulate the long-term impact of fuel degradation on the threshold by combining an exponential decay term. Fuel quality is represented by a quantitative index. Poor quality fuel will reduce combustion efficiency and increase mechanical vibration and noise. FuelQuality ( t This is a real-time fuel quality indicator, collected by a fuel sensor and normalized to the 0-1 range. d For fuel quality degradation rate, control index term The rate of decay. t The host's uptime (in hours) is used as an example. This decay term assumes that the negative effects of inferior fuel gradually decrease over time (this assumption is dynamically corrected through federated learning). Federated learning updates parameters online. α , β , c , d By using a collaborative training framework to aggregate and optimize data from multiple ships, it can adapt to the operating characteristics of different mainframes.

[0023] 2.2 Soft Thresholding Operation A dynamic soft thresholding layer is embedded in each residual module of the residual shrinking network to process the feature map. Perform channel-by-channel soft thresholding: Soft threshold, The input feature map is at the th c Each channel, spatial location ( h , w The original value of ). The sign function preserves the positive or negative information of the eigenvalues. θc ( t ) is the first c The dynamic threshold of each channel is calculated by the formula in step 2.1 and allocated independently to adapt to the noise characteristics of different frequency bands. For soft thresholding, values ​​with absolute values ​​less than 1000 are considered soft thresholding. θc ( tThe eigenvalues ​​of features are set to zero, and eigenvalues ​​greater than the threshold are retained and subtracted from the threshold. (Feature map after thresholding) F soft Through residual structure and original input F in Superposition, i.e. F out = F soft + F in This avoids gradient vanishing and enhances feature propagation.

[0024] When under low load conditions Load ( t When ) < 30%, i ( t Reduce retention of low-amplitude fault characteristics such as gear meshing frequency deviation. High-load conditions. Load ( t >70%, i ( t Increased efficiency to suppress wideband vibrations and high-frequency noise caused by pump and valve impacts. This is especially important when using low-quality fuel. FuelQuality ( t When the threshold is less than 0.5, the threshold is temporarily increased by the exponential decay term to avoid noise overfitting; after long-term operation, the decay term gradually weakens the threshold to adapt to the improvement of fuel quality.

[0025] Traditional fixed threshold methods are prone to missed detections or misjudgments when operating conditions change, while dynamic soft thresholding layers drive the threshold through operating condition parameters, achieving a dynamic balance between noise suppression and fault sensitivity.

[0026] Table 1 Comparison of Dynamic Soft Thresholding Layer Processing Results The construction of a dynamic soft thresholding layer solves the problem that traditional fixed thresholds cannot adapt to dynamic changes in operating conditions, and achieves a dynamic balance between noise suppression and fault sensitivity. The ability to extract weak fault features under low load conditions (<30%) is improved by 25%, the anti-interference performance under high load conditions (>70%) is improved by 18%, and the computational efficiency is optimized by 30%. The processing results of the dynamic soft thresholding layer are shown in Table 1.

[0027] Step 3: Construct a dual-stream attention network, which integrates the time-frequency domain features of vibration signals with the temperature field data of the spatial heat conduction model, and captures the causal temporal relationships of multimodal data through a gating mechanism.

[0028] 3.1 Dual-stream network setup The vibration main branch uses a one-dimensional convolutional neural network (1D-CNN) with a kernel size of 3×1. The kernel length in the temporal dimension is 3, and the spatial dimension is 1. The convolution operation slides only along the time axis, and the step size for temporal feature extraction balances computational efficiency and feature resolution. The output dimension after convolution is [64, 128], with 64 channels and a temporal dimension length of 128 for each channel. This is used to extract time-frequency domain features of vibration signals, capturing high-frequency components from 0-2kHz, such as gear meshing and bearing failure.

[0029] A temperature-dependent graph convolutional neural network (Graph CNN) is constructed using an adjacency matrix based on the host thermal conductivity matrix. Nodes represent the locations of temperature sensors, and edge weights are determined by thermal conductivity. The output dimension after convolution is [32, 64], with 32 channels. The feature dimension after graph convolution aggregation has a spatial dimension length of 64 for each channel, used to model the spatial thermal conduction path of the temperature field and capture the thermal diffusion pattern caused by local overheating.

[0030] 3.2 Implementation of Gated Fusion Layer Through timing alignment Δ t By capturing the cross-modal causal relationship between vibration and temperature, a gating mechanism dynamically adjusts the contribution of the two modes. symbol G ( t ) is a gating weight matrix used to control the fusion ratio of vibration and temperature features. It has dimensions [1, C_out] and takes values ​​in the range (0,1) after Sigmoid activation. s The Sigmoid activation function is used to restrict the gate value to the (0,1) interval to achieve soft fusion. Wg The learnable weight matrix has dimensions [C_in, C_out] and is used for model initialization. No Normal distribution initialization avoids gradient vanishing. VibFeature ( t ) represents the main branch of vibration in time t The feature output has dimensions [B, C_vib, T_vib], where B is the batch size, C_vib = 64 channels, and T_vib = 128 time dimension. TempFeature ( t −Δ t ( ) is a temperature-dependent secondary branch in time t −Δ t The feature output has dimensions [B, C_temp, N_temp], where the number of channels C_temp=32, N_temp=64 is the number of temperature sensors corresponding to the number of nodes, Δt is the heat conduction delay time, and bg is the learnable bias term with dimensions [1, C_out].

[0031] 3.3 Design of Joint Loss Function Joint loss function design through classification loss L_class Compared with loss L_contrast Collaborative optimization ensures the model's accurate identification of fault types while forcing a causal relationship between vibration and temperature features in the hidden layer space. Contrastive loss addresses the problem of traditional methods neglecting cross-modal temporal dependencies by bringing positive samples closer to the multimodal features of the same fault event and pushing negative samples further away. This enables the model to trace the fault evolution path from vibration anomaly to temperature change, ultimately improving the reliability of diagnosis and the accuracy of root cause localization under complex operating conditions. L_total For the total loss of the function, L_class Cross-entropy loss is used in classification loss. L_contrast Enhance contrast loss to improve cross-modal feature consistency. l The loss weight coefficient is taken as its contribution to balancing the classification task and cross-modal consistency constraints. Where: y true This is the one-hot encoded vector of the true label, representing the true category of the sample. y pred is the predicted probability vector output by the model after Softmax activation, representing the probability that a sample belongs to each category. ∑ is the summation over all samples and all categories, used to measure the difference between the model's predictions and the true labels, driving the model to learn accurate fault classification capabilities. V vib The time-frequency features extracted by the vibration main branch have the following dimensions: [B, C_vib, T_vib]. ∑ is the sum of the similarities of all negative sample pairs within the batch. V temp Spatial heat conduction features extracted by the temperature auxiliary branch, with dimensions: [B, C_temp, N_temp]. V temp_neg Temperature characteristics of other samples within the batch. is the cosine similarity function, used to calculate the similarity between two feature vectors. exp is the exponential function, used to amplify differences in similarity.

[0032] Cross-modal causal correlation attention fusion solves the problem that traditional methods ignore the causal relationship of multimodal data, enabling accurate tracing of fault root causes, improving the timeliness of anomaly warning by 40%, reducing the false alarm rate by 35% under complex working conditions, and achieving a fault root cause location accuracy of 92%.

[0033] Table 2 Comparison of Cross-modal Causal Association Attention Fusion Results Step 4: The model from Step 3 is embedded and optimized using depthwise separable convolution, adaptive load normalization, and knowledge distillation to obtain the state monitoring and diagnostic model.

[0034] 4.1 Depthwise Separable Convolution Replacement By decoupling space and channels, the number of parameters is reduced, adapting to the resource constraints of shipborne industrial control computers.

[0035] Traditional convolution parameter count : Depthwise separable convolution parameter count : Parameter reduction percentage: The input feature map has 64 channels, representing the main branch of the vibration. K The convolution kernel spatial dimensions are DepthwiseConv = 3×3 and PointwiseConv = 1×1. The number of channels in the output feature map is reduced from 64 channels to 32 channels.

[0036] 4.2 Adaptive Channel Normalization pass Load ( t Weighted signal processing enhances weak signals under low load and suppresses noise under high load. m c For the first c The weighted average of each channel, N To reduce batch processing size, this addresses the failure issue of traditional BN layers under varying operating conditions and improves the model's environmental adaptability. No. c The first channel, the first i The original feature values ​​of each sample are dynamically adjusted to adjust the normalization intensity to adapt to changes in operating conditions. No. c The weighted variance of each channel, Load ( t ): Real-time load of the host (0-100%). The lower the load, the smaller the weight, and the mean / variance calculation focuses more on high-frequency characteristics; the higher the load, the larger the weight, and the lower the low-frequency interference is suppressed.

[0037] 4.3 Knowledge Distillation and Compression KL divergence loss: Total loss: The output of the teacher model (121 layers) for the i-th sample, Student model (34 layers) for the first i The output of each sample T The temperature coefficient and softening probability distribution reveal the hidden knowledge of the teacher model. α The weights for cross-entropy loss are used to balance the supervision of real labels. L CE Supervision of knowledge distillation L KD . L CE Cross-entropy loss measures the difference between the student model's predictions and the true labels. L KD KL divergence loss measures the difference in output distribution between the student model and the teacher model.

[0038] The lightweight residual shrinking unit optimization solves the problem of excessive resource consumption in general deep learning models, meets the deployment requirements of shipborne industrial control computers, reduces the number of model parameters from 29.3M to 11.5M, increases inference speed by 2.1 times, and reduces embedded platform resource consumption by 55%.

[0039] Step 5: The status monitoring and diagnostic model obtained in Step 4 is mounted on the shipborne industrial control computer. The weight difference between the shipborne model and the global model on the cloud platform is calculated to determine whether a model update needs to be triggered. After receiving the update package, the shipborne node fine-tunes the model with local data to adapt to individual differences of the host.

[0040] 5.1 Initialization of Shipborne Edge Nodes A lightweight model is deployed on the shipborne industrial control computer, and a local data caching mechanism is initialized to provide a basic environment for federated learning.

[0041] 5.2 Shore-based cloud platform aggregation The shore-based cloud platform aggregates the weights of multiple ship models using a federated averaging method and combines differential privacy technology to protect data privacy.

[0042] Federal average aggregation: Differential privacy noise addition: in, For the first t +1 round of global model weights,K The number of ships participating in the aggregation. No. k The ship in t The local model weights of the wheel. To add differential privacy noise weights, N (0, s 2 () has a mean of 0 and a variance of s 2 Gaussian noise. , Δ f Function sensitivity is used to determine the amount of noise that needs to be added in differential privacy. e For privacy budget ( e =1.0), controls noise intensity (the smaller the value, the stronger the privacy protection). d For failure probability ( d =1e-5), allowing for an extremely small probability of noise protection failure.

[0043] 5.3 Model Difference Assessment By calculating the weight difference between the shipborne model and the global model on the cloud platform, it is determined whether a model update needs to be triggered. The D-model weight difference index reflects the degree of deviation between the local model and the global model. When D > t Updates are triggered on time. t The threshold for triggering an update is set based on experience. The current model weights of the shipborne edge nodes. Global model weights of the shore-based cloud platform. Given the L2 Euclidean norm, calculate the square root of the sum of squares of the weight vectors.

[0044] 5.4 Incremental Learning Execution After receiving the update packet, the shipborne node fine-tunes the model with local data to adapt to individual differences in the host machine.

[0045] Fine-tuning the loss function: Weight update: L CE The cross-entropy loss function measures the difference between the predicted result and the true label. y true One-hot encoded vector of the real label, y pred The predicted probability vector output by the model after Softmax activation. The weights of the shipborne model after fine-tuning The shipboard model weights before fine-tuning or To control the step size of weight updates for the learning rate, ∇ L CE This is the gradient of the loss function with respect to the model weights.

[0046] The deployment of the federated learning incremental update mechanism addresses issues such as reliance on manual annotation for model updates, data privacy concerns, and insufficient cross-model generalization capabilities. The frequency of model updates has been reduced from once a week to once a month, while cross-model generalization capabilities have been improved by 30%.

[0047] Through the above steps, a closed loop is achieved from data collection to model deployment, which combines noise resistance, causal correlation analysis, resource optimization and privacy protection, thus meeting the needs of intelligent operation and maintenance of ship main engines.

[0048] Based on the above-described condition monitoring and diagnostic method for ship main engines, this embodiment also discloses a condition monitoring and diagnostic system for ship main engines, such as... Figure 2 As shown, it includes a multimodal data acquisition and processing layer, a dynamic soft thresholding layer, a cross-modal attention fusion network layer, a lightweight residual shrinking unit optimization layer, and a federated learning incremental update deployment layer.

[0049] The multimodal data acquisition and processing layer includes a sensor network deployed on the host computer for acquiring host operating status data and preprocessing the acquired data; the dynamic soft thresholding layer drives the threshold through state parameters to achieve a dynamic balance between noise suppression and fault sensitivity; the cross-modal attention fusion network layer fuses the time-frequency domain features of vibration signals with the temperature field data of the spatial heat conduction model, captures the causal temporal relationship of multimodal data through a gating mechanism, designs a joint loss function, and emphasizes the causal relationship between vibration and temperature features in the hidden layer space while ensuring accurate identification of fault types; the lightweight residual shrinkage unit optimization layer optimizes the model through depthwise separable convolution, adaptive load normalization, and knowledge distillation; the federated learning incremental update deployment layer is used to deploy the model on the shipborne industrial control computer and update the shipborne model in real time.

[0050] The aforementioned condition monitoring and diagnostic system and methods were applied to the six-cylinder main engine of a 210,000-ton bulk carrier for condition monitoring and diagnostics. The results were compared with existing diagnostic models and methods. Figure 3 As shown, the red curve before optimization uses a traditional residual shrinkage network without embedding a dynamic soft thresholding layer, and a fixed threshold ( i =0.15), unfused temperature field data, normalized using standard BN layer. The optimized blue curve uses dynamic soft thresholding + cross-modal attention fusion network, dynamic thresholding ( i ( tDriven by operating parameters, this method integrates vibration-temperature dual-modal data and employs adaptive load normalization. The optimized method outperforms traditional methods in prediction accuracy under most operating conditions, with only explainable local biases observed during extreme abrupt changes in operating conditions.

[0051] The above embodiments are merely exemplary embodiments of the present invention and are not intended to limit the present invention. Those skilled in the art can make various modifications or equivalent substitutions to the present invention within its scope and spirit, and such modifications or equivalent substitutions should also be considered to fall within the scope of protection of the present invention.

Claims

1. A condition monitoring and diagnostic method for ship main engines, characterized in that, Includes the following steps: Step 1: Collect the ship's main engine operating status parameters and perform preprocessing; Step 2, construct the dynamic soft thresholding layer: embed the dynamic soft thresholding layer in each residual module of the residual shrinking network to perform channel-by-channel soft thresholding on the feature map; Step 3: Construct a dual-stream attention network, which integrates the time-frequency domain features of vibration signals with the temperature field data of the spatial heat conduction model, and captures the causal temporal relationships of multimodal data through a gating mechanism; Step 4: The model from Step 3 is embedded and optimized using depthwise separable convolution, adaptive load normalization, and knowledge distillation to obtain the state monitoring and diagnosis model. Step 5: The status monitoring and diagnostic model obtained in Step 4 is mounted on the shipborne industrial control computer. The weight difference between the shipborne model and the global model on the cloud platform is calculated to determine whether a model update needs to be triggered. After receiving the update package, the shipborne node fine-tunes the model with local data to adapt to individual differences of the host. In step 2, the dynamic soft thresholding layer drives the threshold through the running state parameters obtained in step 1.

2. The condition monitoring and diagnostic method for ship main engines according to claim 1, characterized in that, The data preprocessing in step 1 specifically includes: based on data alignment and synchronization technology, timing alignment of vibration signals, temperature field data and operating parameters is achieved through the PTP clock protocol; and the adaptive filter NLMS method is used to suppress high-frequency random noise and retain fault characteristic frequency bands.

3. The condition monitoring and diagnostic method for ship main engines according to claim 1, characterized in that, The specific formula for soft thresholding in step 2 is as follows: in, Soft threshold, It is the original value of the input feature map at the c-th channel and spatial position (h,w); The sign function preserves the positive and negative information of the feature values; θc(t) is the dynamic soft threshold of the c-th channel; For soft thresholding, eigenvalues ​​with absolute values ​​less than θc(t) are set to zero, while eigenvalues ​​with absolute values ​​greater than θc(t) are retained and the threshold is subtracted.

4. The condition monitoring and diagnostic method for ship main engines according to claim 3, characterized in that, The specific calculation method for the dynamic soft threshold is as follows: Where θ(t) is the dynamic soft threshold, α is the load factor, Load(t) is the real-time load of the main unit, β is the speed factor, γ is the fuel quality factor, FuelQuality(t) is the real-time fuel quality index, δ is the fuel quality decay rate, and t is the main unit running time.

5. The condition monitoring and diagnostic method for ship main engines according to claim 1, characterized in that, Step 3 specifically includes: Step 3.1: Construct a vibration main branch one-dimensional convolutional neural network for extracting time-frequency domain features of vibration signals and a temperature auxiliary branch graph convolutional neural network for modeling the spatial heat conduction path of the temperature field. Step 3.2, through timing alignment Δ t Capturing the vibration-temperature cross-modal causal relationship, a gating mechanism dynamically adjusts the contributions of the two modes: G ( t ) is the gate weight matrix, used to control the fusion ratio of vibration and temperature characteristics; σ Use the Sigmoid activation function; Wg The weight matrix is ​​a learnable weight matrix; VibFeature ( t ) represents the main branch of vibration in time t The characteristic output; TempFeature ( t −Δ t ( ) is a temperature-dependent secondary branch in time t −Δ t The characteristic output; Δt is the heat conduction delay time; bg is the learnable bias term; Step 3.3, use classification loss. L_class Compared with loss L_contrast Cooperative joint loss function.

6. The condition monitoring and diagnostic method for ship main engines according to claim 5, characterized in that, The joint loss function is specifically as follows: L_total For the total loss of the function, L_class Cross-entropy loss is used in classification loss. L_contrast Enhance contrast loss to improve cross-modal feature consistency; y true The one-hot encoded vector of the real label; y pred This is the predicted probability vector output by the model after Softmax activation; V vib Time-frequency features extracted from the main vibration branch; V temp Spatial heat conduction characteristics extracted by temperature auxiliary branch; V temp_neg Temperature characteristics of other samples within the batch; is the cosine similarity function; exp is the exponential function.

7. The condition monitoring and diagnostic method for ship main engines according to claim 1, characterized in that, The specific calculation method for the weight difference between the shipborne model and the cloud platform global model in step 5 is as follows: Among them, the weight difference index of model D, The current model weights of the shipborne edge nodes. Global model weights of the shore-based cloud platform.

8. The condition monitoring and diagnostic method for ship main engines according to claim 1, characterized in that, In step 5, the cross-entropy loss in the classification loss is fine-tuned using local data, and the weights of the shipborne model are updated. The specific weight update formula is as follows: in, The weights of the shipborne model after fine-tuning The shipboard model weights before fine-tuning η To control the step size of weight updates for the learning rate, ∇ L CE This is the gradient of the loss function with respect to the model weights.

9. A condition monitoring and diagnostic system for ship main engines, characterized in that, A condition monitoring and diagnosis method for ship main engines based on any one of claims 1-8 includes a multimodal data acquisition and processing layer, a dynamic soft thresholding layer, a cross-modal attention fusion network layer, a lightweight residual shrinkage unit optimization layer, and a federated learning incremental update deployment layer. The multimodal data acquisition and processing layer includes a sensor network deployed on the host, used to acquire host operating status data and preprocess the acquired data; The dynamic soft thresholding layer drives the threshold through state parameters, achieving a dynamic balance between noise suppression and fault sensitivity. The cross-modal attention fusion network layer fuses the time-frequency domain features of vibration signals with the temperature field data of the spatial heat conduction model. It captures the causal temporal relationship of multimodal data through a gating mechanism, designs a joint loss function, and ensures accurate identification of fault types. It also emphasizes the causal relationship between vibration and temperature features in the hidden layer space. The lightweight residual shrinkage unit optimization layer optimizes the model through depthwise separable convolution, adaptive load normalization, and knowledge distillation; the federated learning incremental update deployment layer is used to deploy the model on the shipboard industrial control computer and update the shipboard model in real time.

Images