Fault monitoring and adaptive control method and system for ship desulfurization system

Abstract

The application discloses a fault monitoring and self-adaptive regulation method and system for a ship desulfurization system, and the method comprises the following steps: collecting multi-source heterogeneous data in real time through a distributed sensor network, and performing pretreatment; extracting multi-dimensional statistical features, enhancing weak fault features by using an improved singular spectrum analysis algorithm, training a multi-channel attention mechanism LSTM network by using high-discrimination feature sequences, learning dynamic time sequence rules of normal modes and multiple early fault modes, and performing early probability prediction of faults; identifying complex fault sources caused by coupling of multiple potential factors; dynamically adjusting a detection threshold by using an EWMA model driven by reinforcement learning, and dynamically optimizing sensitivity and specificity of fault monitoring; calculating a severity index by using a fuzzy logic model, evaluating the severity of faults, generating graded early warnings, and iteratively optimizing self-adaptive regulation strategies, and the application improves the robustness and reliability of the system under strong noise and multivariate coupling.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of fault monitoring and technology, in particular to a fault monitoring and adaptive control method and system for a ship desulfurization system. BACKGROUND

[0002] The ship desulfurization tower is a system for removing sulfides in exhaust gas generated by fuel combustion devices such as main engines, auxiliary engines, and boilers of a ship. The system neutralizes sulfur oxides in the exhaust gas by using seawater or adding alkaline scrubbing liquid, so that the ship exhaust gas meets the emission standards. In order to ensure the safety and stability of the ship operation, the existing system is equipped with a fault monitoring and alarm function for the ship desulfurization tower system. The comprehensiveness and real-time performance of the ship desulfurization system fault diagnosis directly affect the reliability and maintenance efficiency of the system.

[0003] In the existing fault monitoring of the ship desulfurization system, there are still problems in fault positioning in realizing system comprehensive health state evaluation and early fault warning. The current ship desulfurization system fault monitoring method mainly relies on pre-set fixed thresholds for fault judgment, that is, the main realization is fault "after the fact" warning. For early faults such as slow degradation of control valve performance, slight drift of sensor, and implicit reduction of scrubbing tower efficiency, which have not caused parameter overrun, it is difficult to extract early weak feature signals representing performance degradation and potential faults from multi-parameter and strong noise background data, and it is difficult to make early prediction and warning for performance degradation or potential faults. At the same time, the existing ship desulfurization system fault monitoring system has low accuracy in fault positioning under multi-physical field coupling. That is, the diagnosis comprehensiveness and accurate positioning of the complex fault root caused by the coupling of multiple potential reasons such as exhaust gas quantity mutation, alkali injection fault, nozzle blockage, and pH sensor distortion are insufficient, and it is difficult to quickly and accurately trace back to the root fault source. SUMMARY

[0004] The purpose of the present application is to provide a fault monitoring and adaptive control method and system for a ship desulfurization system to solve the problems raised in the background.

[0005] The specific technical solutions provided by the present application are as follows: a fault monitoring and adaptive control method for a ship desulfurization system, comprising the following operation steps:

[0006] Step S1: Real-time acquisition of multi-source heterogeneous data by a distributed sensor network, pre-processing by using improved Z-score based outlier detection and adaptive wavelet threshold denoising.

[0007] Preferably, the preprocessing comprises: performing window dynamic adjustment based on an improved Z-score algorithm, detecting outliers using a dynamic sliding window, and self-adaptively adjusting the window size according to data volatility; calculating the Z-score value of data points in the window, marking the data points as outliers when the Z-score value of the data points in the window is greater than a set dynamic threshold, and removing the outliers, and triggering a window size adjustment mechanism when three consecutive points are marked as outliers; after removing the outliers, filling in missing values using a linear interpolation method based on time series; evaluating the quality of the multi-source data by calculating the data integrity rate and the proportion of outliers in real time as indexes, and triggering a sensor calibration instruction when the indexes exceed the standard; performing adaptive wavelet threshold denoising, setting wavelet transform parameters, using Morlet wavelet as a base function, estimating the noise standard deviation by using the median of wavelet coefficients, adjusting the set threshold based on Stein unbiased risk estimation, processing the wavelet coefficients by a soft threshold function, and retaining weak fault characteristics; and then performing signal reconstruction and verification, calculating the signal-to-noise ratio and mean square error after reconstructing the denoised signal, performing Z-score standardization on the denoised multi-source data, and combining a weighted fusion algorithm to merge the multi-sensor data into a comprehensive index.

[0008] Step S2: Extracting time domain, frequency domain and time-frequency domain multi-dimensional statistical features from the preprocessed multi-source data, performing weak fault feature enhancement by using an improved singular spectrum analysis algorithm, constructing a high-dimensional feature set to describe early signs of performance degradation, and forming high-discrimination input features.

[0009] Preferably, the weak fault feature enhancement comprises: performing feature enhancement by using an adaptive singular spectrum analysis algorithm, constructing a trajectory matrix from a one-dimensional feature sequence of a time domain variance sequence, dynamically determining an embedding dimension according to a data non-stationarity index, performing singular value decomposition on the trajectory matrix, reconstructing the trajectory matrix from the first Preferably, the weak fault feature enhancement comprises: performing feature enhancement by using an adaptive singular spectrum analysis algorithm, constructing a trajectory matrix from a one-dimensional feature sequence of a time domain variance sequence, dynamically determining an embedding dimension according to a data non-stationarity index, performing singular value decomposition on the trajectory matrix, reconstructing the trajectory matrix from the first

[0010] Step S3: Training a multi-channel attention mechanism LSTM network by using the output high-discrimination feature sequence, learning dynamic time series rules of normal modes and multiple early fault modes, and performing early probability prediction of faults.

[0011] Preferably, the early probability prediction of faults comprises:

[0012] Step S31: Construct a multi-channel attention mechanism LSTM model, organize the feature sequence of multiple sensors into a multi-channel time sequence by time step, each time step contains the feature values of all sensors, define a sliding window as the input of the model, and the fault probability of a future period of time as the output;

[0013] Step S32: Establish a multi-channel attention mechanism LSTM network architecture including a multi-channel input layer, a channel-specific LSTM layer, and a multi-head attention mechanism;

[0014] Step S33: Train the multi-channel attention mechanism LSTM model using historical data containing normal and multiple fault records, use weighted cross-entropy as the loss function to handle class imbalance, configure the Adam optimizer, and use early stopping to prevent overfitting;

[0015] Step S34: Real-time input of the latest window feature sequence into the model for online prediction, get the fault probability, generate the attention heat map, display the contribution of each sensor to the prediction at different time steps, and output the most important features for fault prediction by averaging the attention weights.

[0016] Step S4: Model the desulfurization system as a multi-physical field coupled graph structure, use a causally enhanced graph neural network to analyze the complex correlation between parameters, simulate the fault propagation path, and identify the complex fault root cause caused by the coupling of multiple potential factors.

[0017] Preferably, identifying the complex fault root cause caused by the coupling of multiple potential factors comprises:

[0018] Step S41: Construct a desulfurization system graph model, define the key physical parameters of the desulfurization system as graph nodes, each node contains its current value and attributes, and define edges representing the physical coupling relationship or causal relationship between nodes, calculate the correlation strength based on the mutual information of historical data;

[0019] Step S42: Construct a causally enhanced graph neural network, output each node as a probability of being a fault root cause;

[0020] Step S43: Simulate the fault propagation path between multi-physical field parameters, locate the root cause of complex faults based on the node fault probability distribution.

[0021] Step S5: Based on the real-time running state, historical performance and fault warning feedback of the system, use a reinforcement learning driven EWMA model to dynamically adjust the detection threshold, and dynamically optimize the sensitivity and specificity of fault monitoring.

[0022] Preferably, dynamically optimizing the sensitivity and specificity of fault monitoring comprises:

[0023] Step S51: Establish a multi-dimensional threshold baseline to compensate for performance degradation;

[0024] Step S52: Dynamically adjust the control limit coefficients and smoothing factor of the EWMA control chart through reinforcement learning;

[0025] Step S53: Optimize fault monitoring by combining fault probability, root fault location, and dynamic threshold;

[0026] Step S54: Continuously calculate the false positive rate and false negative rate, and trigger the reinforcement learning model update when performance degrades.

[0027] Step S6: Based on the predicted failure probability, impact range, and historical data, use a fuzzy logic model to calculate the severity index, assess the severity of the failure, generate a graded warning, and trigger different response processes.

[0028] Step S7: Optimize the control variables within the future time window based on the fault diagnosis and early warning results, adjust the operating parameters through the actuator, and iteratively optimize the adaptive control strategy.

[0029] Compared with existing technologies, the beneficial effects achieved by this invention are as follows: This invention, through weak signal enhancement and deep learning prediction, realizes the transformation from "post-event" alarm to "pre-event" early warning, improving the detection sensitivity of early faults in desulfurization systems. Simultaneously, by utilizing multiphysics coupling analysis and graph neural networks, it solves the problem of inaccurate root cause localization of complex faults, reducing the localization error; and by employing adaptive dynamic thresholds and control strategies, it overcomes the limitations of fixed thresholds. Based on the fusion of data-driven and physical model approaches, this invention improves the robustness and reliability of the system under strong noise and multivariate coupling, providing a comprehensive, real-time, and accurate intelligent monitoring solution for ship desulfurization systems. Attached Figure Description

[0030] Figure 1 This is a flowchart of the fault monitoring and adaptive control method for a ship desulfurization system provided in the embodiments of the present invention.

[0031] Figure 2 This is a schematic diagram of the functional modules of a fault monitoring and adaptive control system for a ship desulfurization system provided in an embodiment of the present invention. Detailed Implementation

[0032] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention are within the scope of protection of the present invention.

[0033] Example 1:

[0034] like Figure 1 As shown in the figure, the fault monitoring and adaptive control method for ship desulfurization systems described in this embodiment includes the following operating steps:

[0035] Step S1: Collect multi-source heterogeneous data in real time through a distributed sensor network, and perform data cleaning and denoising by using adaptive sliding window outlier detection based on improved Z-score and SURE optimized wavelet threshold denoising. Finally, generate high-quality input data through rolling Z-score standardization and weighted fusion to provide a reliable foundation for subsequent fault monitoring.

[0036] In this embodiment, high-precision sensors are deployed at key nodes of the desulfurization system, including pH sensors (measuring the acidity and alkalinity of the scrubbing liquid), electromagnetic flowmeters (monitoring exhaust gas flow and alkali flow), thermocouples (temperature monitoring), pressure transmitters (monitoring pressure within the tower), and vibration sensors (monitoring nozzle status). A distributed sensor network collects multi-source heterogeneous data from the ship's desulfurization system in real time, including exhaust gas flow, pH value, temperature, pressure, and alkali flow. The collected multi-source heterogeneous data is transmitted to edge computing nodes via CAN bus or industrial Ethernet protocol, with a sampling frequency set to 1Hz to ensure real-time data transmission. A timestamp alignment algorithm (such as the NTP protocol) is used to synchronize the collected multi-source data, with a tolerance error of less than 10ms. The data is encapsulated in a structured format (such as JSON), including sensor ID, timestamp, parameter values, and quality flags to ensure subsequent processing.

[0037] For example, in this embodiment, an improved Z-score algorithm is used to dynamically adjust the window, solving the problem of poor adaptability of fixed windows to non-stationary data. A dynamic sliding window is used to detect outliers, and the window size is adaptively adjusted according to data volatility. In this embodiment, the initial value is set to 60 seconds, the minimum window size is 30 seconds, the maximum is 120 seconds, and the sliding step size is 10 seconds. The Z-score value of the data points within the window is calculated using the following formula:

[0038]

[0039] in, This represents the original value of the single data point within the current sliding window that needs to be evaluated. This represents the median value of all data points within the current sliding window. The absolute median difference , Each data point within the window is marked as an outlier if its Z-value exceeds a set dynamic threshold. If three consecutive points are marked, a window resizing mechanism is triggered to reduce the window size and improve sensitivity. Unlike traditional standard deviation, MAD is also insensitive to outliers in the data, making the Z-value calculation more reliable and preventing outliers from "contaminating" the dispersion estimation itself. After removing outliers, a time-series-based linear interpolation method is used to fill missing values ​​to ensure data continuity. The interpolation formula is:

[0040]

[0041] in, For at a certain point in time The values ​​that need to be filled or repaired. For at a certain point in time The previous moment Normal data values, For at a certain point in time The next moment The normal data values ​​are used to assess and control data quality by calculating the data integrity rate (CR) and outlier ratio (OR) in real time. The target is CR>98% and OR<2%. When the indicators exceed the standard, the sensor calibration command is triggered.

[0042] In this embodiment, adaptive wavelet thresholding is employed for denoising. Wavelet transform parameters are set, and Morlet wavelet is used as the basis function due to its superiority in non-stationary signal processing. The number of wavelet decomposition levels is adaptively selected based on the signal length to ensure the preservation of key frequency components. The noise standard deviation is estimated using the median of the wavelet coefficients, and the set threshold is adjusted based on Stein's unbiased risk estimation (SURE). The adjustment formula is as follows: ,in To estimate the noise standard deviation using the median of wavelet coefficients, The adaptive factor is calculated using historical noise levels and applied through a soft thresholding function. The wavelet coefficients are processed to preserve weak fault characteristics. Signal reconstruction and verification are then performed; after reconstructing the denoised signal, the signal-to-noise ratio is calculated. and mean square error The target SNR improvement is set to >10dB and MSE <0.01. If the target is not met, the wavelet basis or decomposition level is automatically adjusted. Finally, Z-score standardization is performed on the denoised multi-source data. Combined with a weighted fusion algorithm, the multi-sensor data are merged into a comprehensive index (such as "washing efficiency index"). The weights are dynamically allocated according to the sensor accuracy and reliability, with higher-accuracy sensors receiving higher weights. Kalman filtering is used for updating, and the threshold is dynamically adjusted by Stein unbiased risk estimation to balance denoising effect and feature preservation, which is superior to traditional fixed threshold methods.

[0043] Step S2: Extract multi-dimensional statistical features in the time domain, frequency domain, and time-frequency domain from the preprocessed multi-source data. Use a weak fault feature enhancement algorithm based on adaptive window singular spectrum analysis to construct a high-discrimination feature set that highlights the weak signs of early performance degradation in the ship desulfurization system, providing high-discrimination input features for subsequent deep learning models.

[0044] In this embodiment, for the extraction of time-domain features, multiple statistical quantities such as mean, variance, kurtosis, skewness, peak factor, waveform factor, and impulse factor are calculated within a sliding window, and the slope of the data within the window is calculated using linear regression as a trend indicator. For the extraction of frequency-domain features, the time-domain signal is converted to the frequency domain by performing a Fast Fourier Transform on the window data, and the amplitude and phase of the main frequency components are extracted; simultaneously, the power spectral density features are calculated using the Welch method, and multiple frequency-domain indicators such as spectral centroid, spectral variance, and spectral kurtosis are extracted. For the extraction of time-frequency domain features, wavelet packet decomposition is used to decompose the signal into different frequency bands, resulting in multiple frequency band signals, and the relative energy and energy entropy are calculated for each frequency band.

[0045] For example, an adaptive singular spectrum analysis algorithm is used for feature enhancement, constructing a trajectory matrix from one-dimensional feature sequences such as the time-domain variance sequence of pH values. ,in It is a lag vector with an embedding dimension of length k. The embedding dimension k is adaptively selected based on the sequence length and dynamically determined according to the data non-stationarity index, which is superior to fixed-window SSA. For the trajectory matrix... Perform singular value decomposition and select the top performers based on the singular value contribution rate. The components are reconstructed, and the selected components are inversely transformed back to time series to obtain the enhanced feature sequence. The 2-norm ratio of the enhanced feature sequence to the original feature sequence is calculated. The calculated 2-norm ratio is compared with a set threshold to determine whether the feature has been effectively enhanced. When the calculated 2-norm ratio is greater than the set threshold, the feature is considered to have been effectively enhanced. The maximum correlation minimum redundancy (mRMR) algorithm is used to select the feature subset that is most relevant to the fault and has the least redundancy from the enhanced high-dimensional features. Principal component analysis (PCA) is performed for visualization to verify the separability of early fault features. This ensures that by combining time domain, frequency domain, time-frequency domain features and SSA enhancement, the significance of weak fault features is comprehensively improved.

[0046] Step S3: Using the high-discrimination feature sequence of the output, train a multi-channel attention mechanism LSTM network to learn the dynamic temporal patterns of normal operation and various early fault modes, including control valve degradation and sensor drift, to achieve early probability prediction of faults and break through the limitation of fixed thresholds.

[0047] In this embodiment, the steps for implementing event warning are as follows:

[0048] Step S31: Construct a multi-channel attention mechanism LSTM model, organize the feature sequences of multiple sensors into a multi-channel time series according to time steps, each time step contains the feature values ​​of all sensors, define a sliding window as the model input, and the failure probability in the future time period as the output.

[0049] Step S32: Establish a multi-channel attention mechanism LSTM network architecture including a multi-channel input layer, a channel-specific LSTM layer, and a multi-head attention mechanism.

[0050] For example, the feature dimensions are input through a multi-channel input layer, and the channels are divided according to the sensor type. In this embodiment, the input feature dimensions include: [batch size, time step 60, number of features 32], where the 32 features come from the feature selection in step S2. The channels divided according to the sensor type include exhaust gas, alkaline solution, temperature, and pressure, with 8 features per channel. Each sensor channel is independently passed through an LSTM layer to capture its own temporal dependence. The number of hidden units in each LSTM layer is set to 64. The output of each channel is weighted by an independent attention head, and the context vector of each channel is obtained by weighted summation. The context vectors of each channel are then concatenated and fused through a fully connected layer. The probability of each fault type is output using the sigmoid activation function (multi-task learning) or the overall fault probability is output using softmax. The established multi-channel attention LSTM architecture allows the model to independently learn the temporal patterns of different sensor types and can focus on key time points through the attention mechanism.

[0051] Step S33: Train a multi-channel attention mechanism LSTM model using historical data containing normal and multiple fault records, use weighted cross-entropy as the loss function to handle class imbalance, configure the Adam optimizer, and use early stopping to prevent overfitting.

[0052] Step S34: Input the feature sequence of the latest window into the model in real time for online prediction, obtain the fault probability, generate an attention heatmap, display the contribution of each sensor to the prediction at different time steps, sort based on the average attention weight, and output the features most important for fault prediction.

[0053] Step S4: Model the desulfurization system as a multi-physics coupled graph structure, use causal enhanced graph neural network to analyze the complex correlation between parameters, simulate the fault propagation path, and accurately identify the complex fault root causes caused by the coupling of multiple potential causes such as sudden changes in exhaust gas volume, alkali injection failure, nozzle blockage, and pH sensor distortion.

[0054] In this embodiment, identifying the root causes of complex failures resulting from potential causal coupling specifically includes:

[0055] Step S41: Construct a graph model of the desulfurization system, define the key physical parameters of the desulfurization system as graph nodes, each node contains its current value and attributes, define edges to represent the physical coupling or causal relationship between nodes, and calculate the correlation strength based on the mutual information of historical data.

[0056] For example, the key physical parameters in this embodiment include parameters selected from the following systems: Exhaust gas subsystem: exhaust gas flow rate, temperature, SOx concentration, pressure difference, etc.; Alkali solution subsystem: alkali solution flow rate, pressure, pH value, density, liquid level, etc.; Scrubber tower subsystem: tower temperature, pressure, liquid level, pressure drop, etc.; Nozzle subsystem: nozzle pressure difference, flow distribution, clogging index, etc.; Sensor subsystem: health status of each sensor, drift amount, etc. Edges representing the physical coupling or causal relationships between nodes are then defined, for example: exhaust gas flow rate. Scrubber pressure drop (physical relationship), alkali flow rate pH value (chemical reaction relationship), nozzle status Uniformity of alkali spray (geometric relationship); correlation strength calculated based on mutual information from historical data.

[0057] Step S42: Construct a causal augmented graph neural network and output the probability of each node being the root cause of the fault.

[0058] For example, information is aggregated from neighboring nodes for each defined node, and then updated via message passing. The update formula is: ,in Represents a node In the The feature representation or embedding vector of the layer (i.e., the updated layer), after message passing, is passed to the nodes in the graph neural network. The newly acquired features, which aggregate information about the feature itself and its neighboring nodes, can be used for subsequent fault classification or regression tasks. To represent a non-linear activation function, ReLU, Sigmoid, or Tanh functions can be used in this invention. Indicates the first The trainable weight matrix of the layer, It is an aggregate function. Representing neighboring nodes In the The feature representation of the layer (i.e., before the update). Represents a node and Normalized coefficients or edge weights between them Represents a node Belongs to node The set of neighboring nodes Furthermore, an attention mechanism is introduced, allowing nodes to dynamically adjust their attention to different neighbors. After multiple layers of graph convolution, each node obtains a high-level representation. The probability of each node being the root cause of a fault is output through fully connected layers and softmax. That is, each node aggregates neighbor information through a learnable and adaptive attention weight. This weight dynamically determines the relative importance of different neighbors. In the fault diagnosis of ship desulfurization systems, this means that the model can automatically learn which changes in other related parameters should be paid more attention to when judging whether a certain parameter has failed, thereby more accurately tracing the source of the fault.

[0059] Step S43: Simulate the fault propagation path between multiphysics parameters and locate the root cause of complex faults based on the node fault probability distribution.

[0060] In this embodiment, a random walk algorithm is used to simulate possible fault propagation paths in the graph. The random walk starts from the currently observed anomalous node and walks along the edges of the graph, probabilistically exploring possible fault sources. Each path consists of a series of consecutive nodes and edges, representing a possible path for the fault to propagate from the root node to the observed node. For each path, its fault contribution is calculated. This contribution is determined by the path's probability and the fault probability of the nodes along the path; a higher contribution indicates a more likely true fault propagation path. Finally, based on Bayesian inference, the probability of each candidate root fault is updated using the observed anomalous state of the current node, and the root fault with the highest posterior probability is selected as the diagnostic result output.

[0061] For example, the formula for calculating the probability of simulating a fault propagation path using the random walk algorithm is as follows:

[0062]

[0063] in, This represents the probability that a fault will propagate along a specific path. 、 express Iterate from 1 to Because fault propagation requires passing through every link in the path, the propagation probability of the entire path is the product of the transition probabilities of each edge segment (from one node to the next) on the path, ensuring the "chain-like property" of fault propagation. Therefore, when calculating the path probability, each connection segment on the path is considered sequentially. This indicates the total number of nodes contained in the path, for example, when the path nozzle is clogged. Abnormal alkali flow rate An abnormal pH value, i.e., containing 3 nodes, then ; Represents a node and nodes The edge weights between nodes, specifically referring to the causal strength coefficients in causal reinforcement GNNs, are calculated from physical models (such as mass conservation and chemical reaction kinetic equations). These weights quantify the causal strength of the nodes. For nodes The greater the weight of the physical impact, the more likely the fault originates from a node. propagation to nodes The stronger the tendency. Indicates a node Sum the weights of all outgoing edges. For nodes The set of neighboring nodes, that is, all nodes that are related to the node Directly connected nodes For traversing nodes Each of its neighboring nodes, Represents a node To one of his neighbors Edge weights.

[0064] The formula for calculating the fault contribution of each path is as follows:

[0065]

[0066] in, This represents the overall contribution of the path to the ultimately observed fault. It is a comprehensive indicator that considers not only the possibility of fault propagation but also the possibility of each node on the path itself failing. The higher the contribution of a path, the greater the likelihood that its root fault is the true root cause. This indicates performing a chain multiplication on all nodes contained in the path. Represents nodes in the path The failure probability is based on the early failure prediction output of the attention-based LSTM network. It represents the probability that the physical parameter represented by the node is in an abnormal state. For example, for the "alkali flow rate" node, It is the probability of "abnormal alkali flow rate" predicted by the LSTM model.

[0067] The formula for calculating the root failure probability is:

[0068]

[0069] in, Represents the posterior probability, given the evidence observed. Under the condition of candidate root failure The probability of the true root cause. It is the first One candidate root failure, This represents the likelihood probability, assuming the root fault is... At that time, evidence was observed The probability, in this invention, Contribution through the above paths Estimate, that is, summarize all from From Departure to Evidence The contribution of each abnormal node's path is calculated and normalized. Represents the prior probability, in the absence of any observational evidence, of the root fault. The probability of occurrence is typically set based on historical fault statistics, equipment maintenance records, or expert experience. For example, if historical data shows that "nozzle clogging" occurs much more frequently than "sensor distortion," then P(nozzle clogging) will be greater than P(sensor distortion). Let represent the normalization constant, also known as the marginal probability of the evidence. This summation term iterates through all possible candidate root faults, ensuring the posterior probability of all root faults. The sum is 1, which conforms to the axiom of probability.

[0070] In this embodiment, multi-level fault tracing achieves accurate inference from observed phenomena to root causes through three layers of calculation: path probability, path contribution, and Bayesian posterior probability. By combining the propagation characteristics of the graph structure with the real-time fault probability of nodes, and finally integrating prior knowledge using a Bayesian framework, it can effectively distinguish complex fault root causes caused by the coupling of multiple potential causes such as sudden changes in exhaust gas volume, alkaline injection failure, nozzle blockage, and pH sensor distortion in the complex environment of ship desulfurization systems with multiple coupled physical fields.

[0071] Step S5: By using a reinforcement learning-driven multi-condition adaptive EWMA threshold adjustment mechanism, combined with equipment health status perception and real-time performance feedback, the sensitivity and specificity of fault monitoring are dynamically optimized to achieve an intelligent balance between false alarm rate and false negative rate.

[0072] In this embodiment, adaptive dynamic threshold adjustment and optimized fault monitoring specifically include:

[0073] Step S51: Establish a multi-dimensional threshold baseline to compensate for performance degradation;

[0074] For example, a multi-dimensional threshold baseline is established, that is, multiple typical operating conditions are divided according to load rate, ambient temperature, seawater salinity, etc., and an independent parameter baseline is established for each operating condition. The impact of equipment performance degradation on the baseline is considered simultaneously, performance degradation is compensated, and the baseline is updated in real time to adapt to system changes.

[0075] Step S52: Dynamically adjust the control limit coefficients and smoothing factor of the EWMA control chart through reinforcement learning;

[0076] For example, calculate the EWMA statistic for each key parameter. , ,in, As a smoothing factor, For the current moment The parameter observations are derived from real-time sensor data. This is the EWMA statistic from the previous time step, used to maintain temporal continuity, and its initial value is defined as... Initialize based on the mean of historical normal data. Calculate control limits: , ,in for Upper control limit at any time, for Lower control limit at time, The target value for the process (usually the historical average). The control limit coefficient determines the width of the control limit. This is the standard deviation of the EWMA statistic, adjusted for reference based on current operating conditions (e.g., load, seawater temperature). The standard deviation of historical data is used to adjust the control limit coefficient and smoothing factor using reinforcement learning to optimize detection performance.

[0077] Step S53: Optimize fault monitoring by combining fault probability, root fault location, and dynamic threshold.

[0078] For example, optimizing fault monitoring includes:

[0079] a. Comprehensive decision-making: Combining the fault probability in step S3 and the root fault location result in step S4, an alarm is triggered when the fault probability exceeds the dynamic threshold and the root fault probability is clear.

[0080] b. Alarm Suppression: For transient anomalies, set a duration threshold to avoid false alarms caused by momentary interference.

[0081] c. Threshold backtracking update: When a fault is confirmed, the threshold is backtracked and adjusted to avoid similar missed detections; when a false alarm occurs, the threshold is adjusted to avoid similar false alarms.

[0082] Step S54: Continuously calculate the false positive rate (FAR) and false negative rate (MAR), and trigger the reinforcement learning model update when performance degrades.

[0083] Step S6: Based on the failure probability and the scope of impact, and using historical data, calculate the severity index using a fuzzy logic model to assess the severity of the failure, generate a graded warning, and trigger different response processes.

[0084] Step S7: Based on the fault diagnosis and early warning results, optimize the control variables within the future time window based on the model predictive control framework to maintain system performance, adjust parameters through actuators, monitor system response in real time, and iteratively optimize the control strategy.

[0085] Example 2:

[0086] like Figure 2 As shown in the figure, the fault monitoring and adaptive control system for ship desulfurization systems described in this embodiment includes: a data acquisition and preprocessing module, a feature extraction and enhancement module, a fault prediction and diagnosis module, an adaptive threshold adjustment and optimization module, and a fault early warning and adaptive control module.

[0087] For example, the data acquisition and preprocessing module is used to acquire multi-source heterogeneous sensor data from the ship's desulfurization system in real time and perform preprocessing operations. The data acquisition and preprocessing module also includes:

[0088] A distributed sensor network unit is used to integrate multiple sensors for multi-source data acquisition and transmission.

[0089] An adaptive data cleaning engine is used to clean, denoise, synchronize, and standardize the collected data.

[0090] The multi-source data fusion unit is used to standardize and align features of real-time data, and generates comprehensive operational indicators through weighted fusion.

[0091] For example, the feature extraction and enhancement module is used to extract multi-dimensional features from the preprocessed data and perform weak fault feature enhancement. The feature extraction and enhancement module also includes:

[0092] A multi-dimensional feature extraction engine is used to extract multi-dimensional features such as time domain, frequency domain, and time-frequency domain.

[0093] The weak fault feature enhancement unit is used to enhance weak fault features using algorithms such as singular spectrum analysis and construct a highly discriminative feature set.

[0094] The early fault prediction unit is used to identify early faults through an early fault prediction model.

[0095] For example, the fault prediction and diagnosis module is used for early fault prediction and fault tracing. The fault prediction and diagnosis module also includes:

[0096] The system graph model building unit is used to model the graph structure of key physical parameters;

[0097] The fault tracing engine is used for multi-physics field coupled fault tracing based on graph neural networks.

[0098] For example, the adaptive threshold adjustment and optimization module is used to dynamically adjust the fault detection threshold and make fault decisions. The adaptive threshold adjustment and optimization module also includes:

[0099] A dynamic threshold calculation engine is used to calculate and update real-time control limits;

[0100] The reinforcement learning optimization unit is used to dynamically adjust the fault detection threshold through a reinforcement learning-driven EWMA model.

[0101] The fault monitoring and optimization unit is used to optimize fault monitoring based on multi-fault fusion decision-making.

[0102] For example, a fault warning and adaptive control module is used for fault warning and adaptive control. The fault warning and adaptive control module also includes:

[0103] The intelligent early warning classification unit is used to classify early warning levels and push and confirm early warning information;

[0104] An adaptive control strategy engine is used to generate and execute control commands in real time through multi-variable collaborative optimization control.

[0105] The human-computer interaction and decision support unit is used to provide real-time feedback on the control effect through a visual monitoring interface.

[0106] It should be noted that, in this invention, relational terms such as "first" and "second" are used merely to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus.

[0107] Finally, it should be noted that the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for fault monitoring and adaptive control of ship desulfurization systems, characterized in that: The following steps are included: Step S1: Collect multi-source heterogeneous data in real time through a distributed sensor network, and preprocess the data using outlier detection based on improved Z-score and adaptive wavelet threshold denoising. Step S2: Extract multi-dimensional statistical features in the time domain, frequency domain, and time-frequency domain from the preprocessed multi-source data, use an improved singular spectrum analysis algorithm to enhance weak fault features, construct a high-dimensional feature set to describe early signs of performance degradation, and form highly discriminative input features. Step S3: Using the high-discrimination feature sequence output, train a multi-channel attention mechanism LSTM network to learn the dynamic temporal patterns of normal mode and various early fault modes, and perform early probability prediction of faults. Step S4: Model the desulfurization system as a graph structure with multi-physics coupling, use causal enhanced graph neural network to analyze the complex correlation between parameters, simulate the fault propagation path, and identify the complex fault root causes caused by the coupling of multiple potential factors. Step S5: Based on the system's real-time operating status, historical performance, and fault warning feedback, dynamically adjust the detection threshold using a reinforcement learning-driven EWMA model to dynamically optimize the sensitivity and specificity of fault monitoring; Step S6: Based on the predicted failure probability, impact range, and historical data, use a fuzzy logic model to calculate the severity index, assess the severity of the failure, generate a graded warning, and trigger different response processes. Step S7: Optimize the control variables within the future time window based on the fault diagnosis and early warning results, adjust the operating parameters through the actuator, and iteratively optimize the adaptive control strategy.

2. The fault monitoring and adaptive control method for a ship desulfurization system according to claim 1, characterized in that: The preprocessing in step S1 includes: The window is dynamically adjusted based on the improved Z-score algorithm. An outlier is detected using a dynamic sliding window, and the window size is adaptively adjusted according to the data volatility. Calculate the Z-score value of the data points within the window. When the Z-score value of a data point within the window is greater than the set dynamic threshold, the data point is marked as an outlier and removed. When three consecutive points are marked as outliers, the window size adjustment mechanism is triggered. After removing outliers, missing values ​​are filled using time-series-based linear interpolation. The quality of multi-source data is evaluated by calculating the data integrity rate and outlier ratio in real time. When the indicators exceed the limits, a sensor calibration command is triggered. Adaptive wavelet thresholding is used for noise reduction. Wavelet transform parameters are set, Morlet wavelet is used as the basis function, noise standard deviation is estimated by using the median of wavelet coefficients, the set threshold is adjusted based on Stein unbiased risk estimation, and wavelet coefficients are processed by a soft thresholding function to retain weak fault characteristics. Then, signal reconstruction and verification are performed. After reconstructing the denoised signal, the signal-to-noise ratio and mean square error are calculated. Z-score standardization is performed on the denoised multi-source data. Combined with a weighted fusion algorithm, the multi-sensor data are merged into a comprehensive index.

3. The fault monitoring and adaptive control method for a ship desulfurization system according to claim 2, characterized in that: Step S2 for enhancing weak fault characteristics includes: Feature enhancement is performed using an adaptive singular spectral analysis algorithm. A trajectory matrix is ​​constructed from the one-dimensional feature sequences of the time-domain variance sequence. The embedding dimension is dynamically determined based on the data non-stationarity index. Singular value decomposition is then performed on the trajectory matrix, and the top features are selected based on the singular value contribution rate. Reconstruct each component; The selected components are inversely transformed back to a time series to obtain the enhanced feature sequence. The 2-norm ratio of the enhanced feature sequence to the original feature sequence is calculated. The calculated 2-norm ratio is compared with the set threshold to determine whether the feature has been effectively enhanced. When the calculated 2-norm ratio is greater than the set threshold, the judgment feature is effectively enhanced. The maximum correlation minimum redundancy algorithm is used to select the feature subset that is most relevant to the fault and has the minimum redundancy from the enhanced high-dimensional features. Principal component analysis is then performed to visualize the feature set and verify the separability of the early fault features.

4. The fault monitoring and adaptive control method for a ship desulfurization system according to claim 3, characterized in that: Step S3 includes early probability prediction of the fault, which includes: Step S31: Construct a multi-channel attention mechanism LSTM model, organize the feature sequences of multiple sensors into a multi-channel time series according to time steps, each time step contains the feature values ​​of all sensors, define a sliding window as the model input, and the failure probability in the future time period as the output; Step S32: Establish a multi-channel attention mechanism LSTM network architecture including a multi-channel input layer, a channel-specific LSTM layer, and a multi-head attention mechanism; Step S33: Train a multi-channel attention mechanism LSTM model using historical data containing normal and multiple fault records, use weighted cross-entropy as the loss function to handle class imbalance, configure the Adam optimizer, and use early stopping to prevent overfitting; Step S34: Input the feature sequence of the latest window into the model in real time for online prediction to obtain the fault probability, generate an attention heatmap to show the contribution of each sensor to the prediction at different time steps, sort by average attention weight, and output the features most important for fault prediction.

5. The fault monitoring and adaptive control method for a ship desulfurization system according to claim 4, characterized in that: Step S4 identifies complex fault root causes resulting from the coupling of multiple potential factors, including: Step S41: Construct a graph model of the desulfurization system, define the key physical parameters of the desulfurization system as graph nodes, each node contains its current value and attributes, define edges to represent the physical coupling relationship or causal relationship between nodes, and calculate the correlation strength based on the mutual information of historical data. Step S42: Construct a causal augmentation graph neural network and output the probability of each node being the root cause of the fault; Step S43: Simulate the fault propagation path between multiphysics parameters and locate the root cause of complex faults based on the node fault probability distribution.

6. The fault monitoring and adaptive control method for a ship desulfurization system according to claim 5, characterized in that: Step S42 outputs the probability of each node being the root cause of the fault, which includes: aggregating information from neighboring nodes for each defined node, performing message passing updates, and using the following update formula: ,in Represents a node In the The feature representation or embedding vector of the layer, Represents a non-linear activation function. Indicates the first The trainable weight matrix of the layer, It is an aggregate function. Representing neighboring nodes In the Layer feature representation, Represents a node and Normalized coefficients or edge weights between them Represents a node Belongs to node The set of neighboring nodes Furthermore, an attention mechanism is introduced, allowing nodes to dynamically adjust their level of attention to different neighbors. After multiple layers of graph convolution, each node obtains a high-level representation, and the probability of each node being the root cause of the fault is output through a fully connected layer and softmax.

7. The fault monitoring and adaptive control method for a ship desulfurization system according to claim 6, characterized in that: Step S43 involves simulating the fault propagation path between multiphysics parameters, including: The possible propagation paths of a fault in a graph are simulated by using a random walk algorithm. The random walk starts from the currently observed anomalous node and walks along the edges of the graph to explore the source of the fault in a probabilistic manner. Each path consists of a series of consecutive nodes and edges, representing the path of the fault propagating from the root node to the observed node. For each path, calculate its failure contribution. This contribution is determined by the probability of the path and the failure probability of the nodes on the path. The higher the contribution, the more likely the path is to be a real failure propagation path. Finally, based on Bayesian inference, the probability of each candidate root fault is updated using the observed abnormal state of the current node, and the root fault with the highest posterior probability is selected as the diagnostic result output.

8. The fault monitoring and adaptive control method for a ship desulfurization system according to claim 7, characterized in that, Step S5 involves dynamically optimizing the sensitivity and specificity of fault monitoring, including: Step S51: Establish a multi-dimensional threshold baseline to compensate for performance degradation; Step S52: Dynamically adjust the control limit coefficients and smoothing factor of the EWMA control chart through reinforcement learning; Step S53: Optimize fault monitoring by combining fault probability, root fault location, and dynamic threshold; Step S54: Continuously calculate the false positive rate and false negative rate, and trigger the reinforcement learning model update when performance degrades.

9. The fault monitoring and adaptive control method for a ship desulfurization system according to claim 8, characterized in that: In step S52, the EWMA statistic is calculated for each key parameter. , ,in, As a smoothing factor, For the current moment The observed values ​​of the parameters, Calculate the control limits using the EWMA statistic from the previous time step: , ,in for Upper control limit at any time, for Lower control limit at time, The target value of the process, The control limit coefficient determines the width of the control limit. The standard deviation of the EWMA statistic is adjusted to the reference value based on the current operating conditions. The standard deviation of historical data is used to adjust the control limit coefficient and smoothing factor using reinforcement learning to optimize detection performance.

10. A fault monitoring and adaptive control system for ship desulfurization systems, characterized in that: The fault monitoring and adaptive control method for a ship desulfurization system according to any one of claims 1-9 includes the following functional modules: a data acquisition and preprocessing module, a feature extraction and enhancement module, a fault prediction and diagnosis module, an adaptive threshold adjustment and optimization module, and a fault early warning and adaptive control module; the data acquisition and preprocessing module is used to acquire multi-source heterogeneous sensor data of the ship desulfurization system in real time and perform preprocessing operations. The feature extraction and enhancement module is used to extract multi-dimensional features from the preprocessed data and enhance weak fault features. The fault prediction and diagnosis module is used for early fault prediction and fault tracing; the adaptive threshold adjustment and optimization module is used for dynamically adjusting the fault detection threshold and making fault decisions; the fault early warning and adaptive control module is used for fault early warning and adaptive control of operating parameters.

Images