Intelligent monitoring linkage control integrated treatment method and system for methanol-containing bilge water

Abstract

The invention provides an intelligent monitoring linkage control integrated processing method and system for methanol-containing bilge water, and the method comprises the steps: deploying a collection terminal integrated with various types of sensors at a ship position, collecting the navigation state, sea condition, bilge water quality and equipment operation parameters, and employing a high-frequency sampling mode to achieve the intelligent monitoring linkage control of the bilge water; sorting historical related data, classifying and marking according to a navigation state, a sea condition grade and a bilge water load interval, extracting methanol concentration characteristic parameters, and performing data cleaning to form a working condition and concentration associated characteristic library; the method comprises the following steps: training an LSTM (Long Short Term Memory) model according to working conditions and a concentration associated feature library, inputting multi-source real-time parameters, outputting a methanol concentration change trend and a pre-judgment result, automatically starting retraining when the model pre-judgment precision does not reach the standard, generating a dynamic control threshold value based on the model pre-judgment result and the real-time working conditions, and adjusting a threshold value interval by adapting to different working conditions. And the association processing equipment carries out real-time load matching processing intensity.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of knowledge graph, in particular to a methanol-containing bilge water intelligent monitoring and linkage control integrated processing method and system. BACKGROUND

[0002] The application of methanol fuel in the field of ship navigation is becoming more and more widespread, and the amount of methanol-containing bilge water is increasing. The accurate control of the methanol concentration is directly related to the compliance of maritime emissions and the protection of the marine ecosystem. The current maritime regulations continue to tighten the limit requirements for methanol emissions from ship bilge water, and higher standards are required for the monitoring and control accuracy of the processing system. However, the existing monitoring process of methanol-containing bilge water treatment has obvious shortcomings. The integration of sensors that adapt to the complex environment of high vibration, salt spray, and explosion-proof of ships is low. Most of them use single-point static sampling mode, and the sampling frequency is insufficient, which cannot capture the instantaneous fluctuations of methanol concentration. Moreover, the data collection does not achieve multi-dimensional integration of sailing state, sea conditions, and equipment operating parameters, making it difficult to explore the correlation between working conditions and methanol concentration, and prone to monitoring blind spots and data distortion problems.

[0003] The traditional control method of methanol-containing bilge water mostly uses fixed threshold control, which cannot adapt to the changes of dynamic working conditions such as uniform speed, turning, and different sea conditions of ships. The methanol concentration prediction lacks effective time series analysis model, and there are problems of prediction lag and insufficient accuracy. Moreover, the matching degree of the running load of the processing equipment and the processing intensity is low, which is prone to equipment overload or insufficient processing. At the same time, the monitoring, processing, and alarm subsystems are independent of each other, the communication protocol is not unified, the collaborative linkage is poor, the composite fault recognition highly depends on manual inspection, and it is difficult to achieve automatic and rapid fault determination and interlock prevention and control. The system as a whole lacks self-optimization and redundancy protection capabilities, and the response timeliness and control stability under complex working conditions cannot meet the intelligent control needs of actual navigation. SUMMARY

[0004] In view of the above-mentioned problems, in combination with the first aspect of the present application, the methanol-containing bilge water intelligent monitoring and linkage control integrated processing method is provided, which comprises: Deploying a collection terminal integrated with multiple types of sensors at the position of the ship, collecting sailing state, sea conditions, bilge water quality, and equipment operating parameters, using high-frequency sampling mode, sorting historical related data and classifying and labeling according to sailing state, sea condition level, and bilge water load interval, extracting methanol concentration characteristic parameters, and forming a working condition and concentration correlation feature library after data cleaning; Training an LSTM model based on the working condition and concentration correlation feature library, inputting multi-source real-time parameters, outputting methanol concentration change trend and prediction result, automatically starting retraining when the model prediction accuracy is not up to standard, generating dynamic control threshold based on the model prediction result and real-time working condition, adjusting the threshold interval for different working conditions, and matching the processing intensity with the real-time load of the processing equipment. The edge computing technology is introduced to complete data preprocessing near the collection terminal, data transmission is carried out by using industrial real-time Ethernet, each subsystem of monitoring, alarm, processing and discharge is connected to a unified distributed collaborative control platform, each subsystem is given a unique communication identifier and data interaction interface, and linkage trigger conditions and execution priorities are clearly defined; According to the linkage level divided according to the methanol concentration prediction result and the real-time detection value, different actions of pretreatment, main treatment, emergency treatment and emergency disposal are respectively corresponding to each level, so that the linkage process is quickly responded; A composite fault feature library is established, real-time multi-source data is analyzed by using a decision tree algorithm, automatic identification of composite faults is realized, corresponding intelligent interlocking logic is designed for different types of composite faults, a matching processing flow is started, and key valve locking prevention measures are taken; A response time monitoring unit is arranged on the collaborative control platform, the response states of each link of data processing, instruction transmission and equipment execution are monitored in real time, when the response states of each link do not meet the requirements, a standby channel or redundant equipment is automatically started, and related time information is recorded for optimizing the communication link and the control program.

[0005] In another aspect, the embodiment of the present application also provides a methanol-containing bilge water intelligent monitoring and linkage control integrated processing system, characterized by comprising: A processor; a machine-readable storage medium for storing machine-executable instructions of the processor; wherein the processor is configured to execute the machine-executable instructions to perform the above-mentioned methanol-containing bilge water intelligent monitoring and linkage control integrated processing method.

[0006] In another aspect, the embodiment of the present application also provides a computer program product, which comprises machine-executable instructions stored in a computer-readable storage medium, a processor of a computer device reads the machine-executable instructions from the computer-readable storage medium, and the processor executes the machine-executable instructions to make the computer device execute the above-mentioned methanol-containing bilge water intelligent monitoring and linkage control integrated processing method.

[0007] Based on the above aspects, through multi-dimensional data acquisition and working condition-concentration correlation feature library construction, combined with an LSTM network model adapted to the dynamic working condition of the ship, dual-precision output of methanol concentration quantitative prediction and trend classification is realized, and the problems of prediction lag and poor adaptability of traditional methods are effectively solved. The innovative working condition-trend dual-dimension dynamic threshold adjustment mechanism, combined with the precise matching of equipment load and processing intensity and safety constraints, not only ensures that the methanol concentration strictly meets the maritime emission standard, but also avoids overloading operation of the equipment, significantly improves the compliance of the bilge water treatment and the service life of the equipment.

[0008] Leveraging edge computing and a distributed collaborative control platform, the system achieves efficient collaboration in data preprocessing, command transmission, and device linkage, ensuring rapid response under complex operating conditions. The combination of a composite fault feature library and a decision tree algorithm enables automatic identification and differentiated interlocking control of composite faults, significantly reducing manual intervention costs and the risk of fault propagation. Simultaneously, through incremental model training and iterative rule optimization, the system possesses the self-optimization capability to continuously adapt to new operating conditions and regulations, exhibiting significantly superior long-term operational stability and reliability compared to traditional control systems. Attached Figure Description

[0009] Figure 1 This is a schematic diagram of the execution flow of the integrated intelligent monitoring and linkage control method for methanol-containing bilge water provided in this embodiment of the invention.

[0010] Figure 2 This is a schematic diagram of exemplary hardware and software components of an unconventional water source problem tracing system utilizing knowledge graphs provided in an embodiment of the present invention. Detailed Implementation

[0011] The present invention will now be described in detail with reference to the accompanying drawings. Figure 1 This is a flowchart illustrating an embodiment of the intelligent monitoring and linkage control integrated processing method for methanol-containing bilge water provided by the present invention. The following is a detailed description of the intelligent monitoring and linkage control integrated processing method for methanol-containing bilge water.

[0012] Step S110: Deploy a data acquisition terminal integrating multiple types of sensors at the ship's location to collect navigation status, sea state, bilge water quality and equipment operating parameters. Use a high-frequency sampling mode to sort out relevant historical data and classify and label them according to navigation status, sea state level and bilge water load range. Extract methanol concentration characteristic parameters and form a working condition and concentration correlation feature library after data cleaning. MS-6000 integrated multi-parameter acquisition terminals are deployed at key monitoring points in the ship's main control area of ​​the engine room, at the inlet and outlet flanges of the storage tanks, in the bilge water treatment equipment room, and along the pipelines. These terminals integrate a GPS positioning module, a three-axis gyroscope, an ultrasonic wave height sensor, an electrochemical methanol concentration sensor, an electromagnetic flow sensor, and a motor power sensor. They simultaneously collect data on navigation status (speed, angular velocity, acceleration), sea state (wave height, wave frequency), bilge water quality (methanol concentration, pH value), and equipment operating parameters (motor power, valve opening, processing flow rate). The sampling frequency for the high-frequency sampling mode is set to 300Hz to ensure the capture of instantaneous parameter fluctuations. Data from the past three years' worth of ship AIS navigation logs, equipment PLC operation records, daily water quality reports, and maritime sea state monitoring data are also collected. According to the IMO MSC.1 / Circ.1376 standard, the navigation status is classified as follows: constant speed if the rate of change of speed is ≤ ±0.3 m / s², acceleration / deceleration if the rate of change of speed is ≤ 1.8 m / s² if the speed is ≤ 0.3 m / s², and turning if the angular velocity of heading is ≥ 2.5° / s. The sea state is classified according to the GB / T12763.7-2021 standard. The bilge water load is classified according to the processing flow threshold. The 15-second sliding window method is used to extract characteristic parameters such as the mean, variance, peak value, and slope of methanol concentration. A three-level cleaning rule is constructed: data exceeding the range is directly removed, missing data in a single frame is filled with the mean of the next 5 frames, and abnormal fluctuation data is filtered according to the 3σ principle. Finally, an associated feature library of operating condition labels, feature parameter sets, and methanol concentration time series curves is formed, which supports fast retrieval by operating condition type.

[0013] Step S111: Select an integrated data acquisition terminal suitable for the high vibration, salt spray, and explosion-proof environment of a ship. The terminal includes four types of anti-interference wide-temperature sensors for navigation status, sea state, bilge water quality, and equipment operation status. It is deployed in the engine room power area, storage tank inlet and outlet, processing equipment inlet and outlet, and pipeline nodes, with no monitoring blind spots and redundant installation interfaces reserved at the points. The selected integrated data acquisition terminal (model: MarineSense-800) meets IP69K protection standards, has an operating temperature range of -45℃ to 90℃, vibration resistance up to IEC60068-2-6:2007, salt spray protection up to ASTMB117-2021, and an explosion-proof rating of ExiaIIBT6Ga. The terminal incorporates four types of anti-interference, wide-temperature sensors: a navigation status sensor (GPS + BeiDou dual-mode positioning, three-axis accelerometer with a range of ±16g and an accuracy of ±0.01g); a sea state sensor (ultrasonic wave height sensor with a measurement range of 0~15m and an accuracy of ±0.05m; and a wave frequency sensor with a measurement range of 0.05~15Hz); and a bilge water quality sensor (methanol concentration sensor). The sensors have an accuracy of ±0.05 mg / L; the pH sensor has a range of 0~14 and an accuracy of ±0.02; the turbidity sensor has a range of 0~100 NTU and an accuracy of ±1 NTU. Equipment operation status sensors include Hall current sensors (range 0~800 A, accuracy ±0.2% FS; pressure sensors (range 0~1.6 MPa, accuracy ±0.1% FS)). The terminals use flange fixing and magnetic assisted installation. Deployment locations and quantities are as follows: 4 units in the engine room power area (covering the main pump, standby pump, and chemical pump); 3 units each at the storage tank inlet and outlet; 2 units each at the treatment equipment inlet and outlet; and 1 unit every 5 meters at key pipeline nodes, totaling 28 units, achieving zero monitoring blind spots. The terminals have 4 reserved M12 redundant installation interfaces to support future sensor expansion.

[0014] Step S112: The unified sensor interface is ModbusRTU or CANopen protocol, the terminal integrates a data synchronization control module, and the local cache module is configured to store raw data; All sensor interfaces are standardized to the Modbus RTU protocol (9600bps baud rate, 8 data bits, even parity, 1 stop bit). Some high-speed sensors use the CANopen protocol (500kbps bit rate, conforming to the DS301 application layer specification). The terminal has a built-in protocol conversion module (Model: Modbus-CAN-200) to achieve data interoperability between the two protocols. The acquisition terminal integrates an STM32H750VBT6 data synchronization control module, which ensures the consistency of timestamps of multi-sensor data through a precise clock synchronization algorithm (time synchronization error ≤1ms). A 32GB industrial-grade SSD local cache module is configured with a circular storage mechanism, and the original data storage time is ≥72 hours. It also supports cache locking when an anomaly is triggered (retaining data for 10 minutes before and after the fault). The terminal has a built-in data encryption module that uses AES-256 encryption to encrypt cached data to prevent data tampering. A cached data backup mechanism is established, and the cached data is synchronized to the ship's central server every hour through the ship's internal LAN. If synchronization fails, it automatically switches to local encrypted storage and retransmits the data after the network is restored to ensure that the original data is not lost.

[0015] Step S113: Set the sampling trigger mode to timed and event-linked, and the data format includes millisecond-level timestamps, sensor type, parameter values, and sensor working status identifiers. Enable the terminal preprocessing function to complete data format conversion, verification, and removal of invalid data exceeding the range. The sampling trigger mode is set to a dual mode: timed trigger and event-linked trigger. The timed trigger cycle is set to 200ms / time to ensure data continuity under normal operating conditions. The event-linked trigger conditions include: methanol concentration exceeding the warning threshold by ±10%, equipment current sudden change ≥15%, and wave height sudden change ≥0.5m. After triggering, the sampling frequency is automatically increased to 500ms / time, and after continuous sampling for 5 minutes, it returns to the normal frequency. The data format adopts JSON structure, including millisecond-level timestamps (format: YYYYMMDDHHMMSSFFF), sensor unique identifier (e.g., CH3OH-001), parameter values ​​(including units), and sensor working status identifiers (0=normal, 1=fault, 2=low battery). The terminal preprocessing function is enabled to first complete the data format conversion (converting the sensor's original binary data to standardized JSON format), then verify the continuity of timestamps and the rationality of parameter values ​​through a verification algorithm (e.g., methanol concentration exceeding the range of 0~50mg / L is judged as invalid), and finally, invalid data from sensors that are out of range or have no response are removed. A checksum (CRC16) is added to the valid data after preprocessing to ensure data integrity during transmission.

[0016] Step S114: Collect historical data, navigation logs, equipment operation records and inspection records through the ship's local area network or satellite communication, establish standardized rules to unify parameter units and formats, convert them into a unified structure dataset and sort them by time; Real-time data is collected via the ship's EtherNet / IP industrial LAN, and supplemented by FleetXpress maritime satellite data transmission during ocean voyages. The collected data includes: historical terminal data from the past three years, ship AIS navigation logs, equipment PLC operation records, and crew inspection paper records (scanned and digitized). A standardized rule base is established: unified parameter units (speed in km / h, flow rate in m³ / h, concentration in mg / L), unified data formats (timestamp format unified to UTC, parameter values ​​retained to two decimal places), and unified anomaly indicators (using "-999" to indicate missing data). Python PaaS is used for this purpose. The ndas library transforms heterogeneous data into a unified structured dataset with 18 core fields, including UTC timestamp, ship MMSI code, navigation status, sea state level, bilge water load, methanol concentration, pH value, motor power, and processing flow rate. The dataset is sorted in ascending order by UTC timestamp, and a time-series index is built. A data aggregation delay threshold of ≤1s is set, and the aggregation status is monitored in real-time. If a data delay exceeds 5s, a local cached data retransmission mechanism is automatically triggered, and an alarm message is pushed to the central control system. The final standardized dataset is stored in a PostgreSQL database on the ship's server, supporting fast queries by time range and parameter type.

[0017] Step S115: Classify navigation status as constant speed / acceleration / deceleration / turning, sea state as calm / mild swell / moderate swell / severe swell, bilge water load as low / medium / high, and form a labeled classification dataset according to the standard annotation dataset; According to the IMO standard for ship maneuvering performance (MSC.1376), navigation states are classified as: constant speed (speed change rate ≤ ±0.3 m / s², duration ≥ 10 s), acceleration (speed change rate > 0.3 m / s² and ≤ 1.8 m / s²), deceleration (speed change rate < -0.3 m / s² and ≥ -1.8 m / s²), and turning (heading angular velocity ≥ 2.5° / s, duration ≥ 3 s); according to GB / T12763.7-2021 "Marine Survey Specifications Part 7: Marine Meteorological Observation", sea state grades are classified as: calm (wave height H ≤ 0.5 m), slight swell (0.5 m < H ≤ 1.25 m), moderate swell (1.25 m < H ≤ 2.5 m), and severe swell (H > 2.5 m); and load zones are classified according to the rated flow rate of the bilge water treatment equipment (20 m³ / h). The dataset was categorized into three load ranges: low load (processing flow Q ≤ 6 m³ / h), medium load (6 m³ / h < Q ≤ 14 m³ / h), and high load (Q > 14 m³ / h). The PythonLabelStudio tool was used for automatic labeling of the unified structure dataset. The labeling logic was as follows: the algorithm identified the navigation status, sea state level, and load range corresponding to each data point in real time, and automatically added labels (Navigation status: 0 = constant speed, 1 = acceleration, 2 = deceleration, 3 = turning; Sea state level: 0 = calm, 1 = slight, 2 = moderate, 3 = severe; Load range: 0 = low, 1 = medium, 2 = high). After labeling, 10% of the dataset was randomly selected for manual review to ensure an accuracy rate ≥ 98%. Errors found during the review were corrected through iterative algorithm optimization, ultimately resulting in a categorized dataset with three-dimensional labels.

[0018] Step S116: Group the datasets according to the labeled categories, analyze the methanol concentration variation pattern, extract the average concentration and fluctuation amplitude feature parameters, and form a structured feature parameter set through statistical quantification; The dataset was grouped into 48 groups based on a three-dimensional label combination of navigation status, sea state, and bilge water load. The methanol concentration time-series data in each group was analyzed using the MATLAB signal processing toolbox to extract time-domain characteristic parameters: concentration mean (arithmetic mean of all data within the window), fluctuation amplitude (difference between the maximum and minimum values ​​within the window), standard deviation (reflecting the degree of concentration dispersion), peak factor (ratio of peak value to effective value), and slope (calculated by linear regression fitting to determine the rate of concentration change per unit time). The analysis window size was set to 30 seconds, and the sliding step size to 10 seconds to ensure the continuity of the characteristic parameters. The extracted feature parameters are statistically quantified: the distribution interval, median, and quartiles of each feature parameter in each data set are calculated, and outlier features (deviations from the group mean by 3 times the standard deviation) are removed; the quantified feature parameters are bound to the corresponding working condition labels to form a structured feature parameter set, with fields including: working condition label combination, concentration mean, fluctuation range, standard deviation, peak factor, slope of change, and statistical window duration; the Min-Max normalization method is used to map the feature parameters to the [0,1] interval to eliminate dimensional differences; the final structured feature parameter set is stored in Parquet format, supporting rapid data interaction with subsequent model training modules.

[0019] Step S117: Construct cleaning rules for numerical rationality, logical consistency, and trend continuity; process abnormal data in a hierarchical manner; integrate data according to operating condition category, characteristic parameters, and methanol concentration change data to construct a feature library; and use distributed storage.

[0020] Three-level data cleaning rules are established: ① Numerical rationality rule: methanol concentration 0~50mg / L, pH value 0~14, motor power 0~120% of rated power; data outside these ranges are considered invalid. ② Logical consistency rule: pump power and processing flow rate are positively correlated (correlation coefficient ≥0.7), and there is no contradiction between sailing speed and heading angular velocity (heading angular velocity ≤2° / s under uniform speed conditions); violations of logic are marked as abnormal. ③ Trend continuity rule: methanol concentration change between adjacent data points ≤1mg / L, and the concentration change trend of three consecutive data points is consistent (all increasing / all decreasing / all stable); a sudden change in trend triggers a secondary verification. Abnormal data is processed in stages: Level 1 abnormalities (over-range, serious logical contradictions) are directly removed; Level 2 abnormalities (sudden trend changes) are further removed. Linear interpolation is used to fill in missing data (variable or single-frame missing data); data for level 3 anomalies (slight numerical deviations) are retained and anomaly markers are added; cleaned operating condition classification data, structured feature parameters, and methanol concentration change time series data are integrated to construct an operating condition-concentration correlation feature library, using a Hadoop distributed storage architecture (3 data nodes + 1 name node), with 3 data replicas to ensure storage reliability; the feature library adopts a two-level index structure of operating condition type and time range, supporting millisecond-level data retrieval; a feature library update mechanism is established: newly added data from the previous day is automatically collected at 2:00 AM every day, cleaned, and feature extracted before being stored in the library; the integrity of the feature library is verified at the end of each month, missing data is supplemented, and feature dimensions are optimized to ensure the timeliness and accuracy of the feature library.

[0021] Step S120: Train an LSTM model based on the feature library of working conditions and concentration correlation, input multi-source real-time parameters, output methanol concentration change trend and prediction results, automatically start retraining when the model prediction accuracy is not up to standard, generate dynamic control thresholds based on model prediction results and real-time working conditions, adjust threshold ranges to adapt to different working conditions, and correlate processing equipment real-time load matching processing intensity. A methanol concentration prediction cycle of 5 minutes is set. An LSTM model is trained based on a feature library of operating conditions and concentration correlation, with 12-dimensional multi-source parameters such as navigation status and sea state as input. The model outputs the concentration time series value for the next 5 minutes and the rising / falling / stable trend, requiring a quantitative prediction MAE ≤ 3% and a trend accuracy ≥ 92%. If the accuracy does not meet the standard, a retraining process is automatically triggered. Based on maritime standards, a basic threshold of 4 mg / L (early warning) and 4.75 mg / L (emergency) is set for nearshore areas. The threshold range is dynamically adjusted according to the complexity of the operating conditions and the concentration trend. The load rate is calculated by collecting equipment power in real time, and four levels of processing intensity are matched according to the load rate ≤ 70%, 70%-85%, and > 85%, to achieve coordinated adaptation of operating conditions, thresholds, and load.

[0022] Step S121: Extract multi-source parameters such as navigation status, sea state, bilge water quality, and equipment operation status from the operating condition and concentration correlation feature library. Use the Min-Max normalization method to eliminate the dimensional differences between different parameters. Design a time series sliding slicing strategy for the dynamic operating conditions of ships. Extract multi-source parameter input sequences and methanol concentration output sequence sample pairs according to time window and sliding step size mode, so that the time dimension of the sample pairs adapts to the methanol concentration prediction cycle requirements. Divide the training set and validation set as needed. A 12-dimensional multi-source parameter was extracted from the feature library and mapped to the [0,1] interval using Min-Max normalization. For a 5-minute prediction period, a time series sliding slicing strategy was designed: the time window was set to 300s (consistent with the prediction period), and the sliding step size was 60s (adapting to changes in operating conditions). Sample pairs of "multi-source parameter input sequence - concentration output sequence" were extracted. The training and validation sets were divided in an 8:2 ratio. The training set contained 48 types of operating condition combinations, and the validation set covered all operating condition types, ensuring that the sample distribution was consistent with the actual operating scenario and supporting the training of the model's generalization ability.

[0023] Step S1211: Define the methanol concentration prediction period, set the basic duration of the time window with a reasonable multiple that matches the prediction period, set the basic value of the sliding step size with the minimum change period of the ship's operating conditions, establish a database of operating conditions and parameter change characteristics, statistically analyze the fluctuation amplitude and change rate of multi-source parameters under different operating conditions, and mark the effective duration of parameter characteristics corresponding to each operating condition. The methanol concentration prediction cycle is defined as 3 minutes, and the base duration of the time window is set at twice this cycle (6 minutes). The minimum change cycle of ship operating conditions is statistically determined to be 30 seconds, and the base value of the sliding step is set to 30 seconds. A "operating condition type - parameter characteristics" library is constructed. Through historical data, the parameter fluctuation amplitude (e.g., wave height fluctuation ±0.3m) and change rate (e.g., speed change ≤0.5m / s²) of 48 operating conditions such as uniform speed / wave height are statistically analyzed. The effective duration of parameter characteristics for each operating condition is marked (80 seconds for high complexity conditions and 120 seconds for low complexity conditions), providing a basis for dynamic window adjustment.

[0024] Step S1212: Extract multi-source parameters and methanol concentration data of navigation status, sea state, bilge water quality, equipment operation status, which have been normalized to Min-Max in the working condition and concentration correlation feature library. Using the unified high-precision timestamp of ship GPS time as the reference axis, correct the sensor sampling delay, remove abnormal data points, and complete the sampling interval to generate a time series dataset with a continuous time axis and consistent sampling frequency. Calculate the change rate of multi-source parameters in real time to classify the working condition complexity level, and adaptively adjust the window duration and sliding step size according to the level. Extract sample pairs of multi-source parameter input sequence and methanol concentration output sequence, and retain valid samples by filtering through information entropy. Normalized multi-source parameters and concentration data were extracted, and a unified timestamp was created using GPS timing (synchronization error ≤1ms). Linear interpolation was used to complete the sampling interval, and outliers were removed using the 3σ principle to generate a 300Hz continuous time-series dataset. The parameter change rate was calculated in real time to classify the operating conditions into three levels of complexity: high, medium, and low. The window for high complexity was extended to 7 minutes with a step size of 20 seconds, while the window for low complexity was shortened to 5 minutes with a step size of 40 seconds. After extracting sample pairs, the information entropy was calculated, and valid samples with an entropy value ≤1.2 were retained, while redundant and noisy data were removed.

[0025] Step S1213: Classify the effective samples according to navigation status, sea state level, and bilge water load range. For each combination of operating conditions, extract training and validation sets in a stratified manner according to a preset ratio. For rare combinations of operating conditions with insufficient sample size, generate virtual samples to supplement the training set by slightly perturbing the parameter values ​​to ensure that the sample size of each combination of operating conditions is balanced. The valid samples were divided into 48 operating condition combinations based on "4 types of navigation conditions × 4 types of sea states × 3 types of loads". For each combination, training and validation sets were extracted stratified at an 8:2 ratio. For rare operating conditions (such as turning + severe surge + high load), virtual samples were generated using parameter values ​​with a small amplitude of ±5%, ensuring that the proportion of samples in each operating condition was ≥1% and the total number of training samples was ≥100,000. This avoids model prediction bias for rare operating conditions and ensures adaptability to all operating conditions.

[0026] Step S1214: After accumulating a preset number of new samples, calculate the predictive contribution of sample pairs under different working condition combinations. If the predictive accuracy of a certain working condition does not meet the preset standard, recalibrate the window / step parameters corresponding to that working condition, and update the working condition type, parameter change feature library and incorporate new working condition data features.

[0027] A threshold of 10,000 new samples was set, and cosine similarity was used to calculate the predictive contribution of sample pairs with different working conditions. If the quantitative MAE of a certain working condition is greater than 3% or the trend accuracy is less than 92%, its time window (adjustment range ±1 minute) and sliding step size (±10s) are recalibrated; the working condition-parameter feature library is updated synchronously to include the parameter fluctuation and rate of change features of new working conditions, and a full feature library optimization is completed every quarter to ensure that the samples are adapted to the model iteration requirements.

[0028] Step S122: Build an LSTM network for predicting methanol concentration time series to adapt to complex dynamic operating conditions of ships. The input layer dimension of the network is matched with the feature dimension of the multi-source parameters. A dropout layer is added to the hidden layer to suppress model overfitting. A dual-output layer design is adopted to simultaneously output the specific predicted value of methanol concentration and the classification result of concentration change trend, which respectively meet the needs of quantitative control and qualitative decision-making. Multiple types of loss functions are integrated to optimize the training objective. The input layer has 12 neurons (matching 12-dimensional core parameters), with two hidden layers (60 and 50 neurons respectively), each followed by a dynamic dropout layer (rate 0.1-0.4). The dual output layers output concentration time-series values ​​via linear activation and trend classification results via softmax activation, respectively. MAE loss (quantitative) and cross-entropy loss (qualitative) are integrated, dynamically weighted according to operating conditions (0.7 for nearshore quantitative analysis and 0.4 for offshore analysis), and L2 regularization (λ=1e-5) is added to suppress overfitting. After training, deployment is initiated once the full-condition validation is successful.

[0029] Step S1221: Integrate four types of parameters: ship navigation status, sea state, bilge water quality, and equipment operation status, and count the effective feature dimensions of each parameter. Use the Pearson correlation coefficient method to calculate the correlation between each parameter and methanol concentration changes, remove low-correlation redundant parameters, and retain feature dimensions. The system integrates 18 features across four categories, including navigation status and sea state. The Pearson correlation coefficient method is used to calculate the correlation between each parameter and methanol concentration. A threshold of |r|≥0.3 is set, and 6 low-correlation redundant parameters, such as bilge water temperature, are removed. The system retains 12 core feature dimensions, such as speed, wave height, and initial methanol concentration, to ensure that the input parameters are both correlated and concise, thereby reducing the computational complexity of the model and improving the real-time performance of predictions.

[0030] Step S1222: The LSTM network architecture is designed around the characteristics of ship operating conditions. The input layer incorporates ship operating condition label encoding features, which are input in conjunction with the time series of multi-source parameters. The hidden layer adopts a multi-level structure and is configured with a dynamic dropout mechanism that adapts to the complexity of the operating conditions. The output layer constructs a dual output mode of quantitative and qualitative analysis. At the same time, by adding an operating condition attention layer, parameters are given higher weights based on historical data, which strengthens the influence of key parameters on the prediction results. Overall, a deep adaptation between ship operating conditions and methanol concentration time series prediction is achieved. The LSTM network incorporates a ship operating condition adaptation design: the input layer concatenates a 12-dimensional parameter time series with an 8-dimensional operating condition label embedding vector; the hidden layer has three layers and is configured with a dynamic dropout mechanism (high complexity 0.4, medium complexity 0.2, low complexity 0.1); an additional operating condition attention layer is added, and the parameter weight matrix for 48 operating conditions is trained based on historical data. The dual output layers respectively meet the needs of quantitative control (concentration value) and qualitative decision-making (trend), achieving deep adaptation between operating conditions and concentration prediction through hierarchical collaboration, with a prediction accuracy of ≥92% for all operating conditions.

[0031] Step S12221: Simplify the features and determine the feature dimensions of the multi-source parameter time series sequences of ship navigation status, sea state, bilge water quality, and equipment operation status. Set the number of neurons in the input layer to match the feature dimensions. Use one-hot encoding to convert the discrete labels of ship operating conditions into high-dimensional sparse vectors. Connect them to the embedding layer to map them into low-dimensional dense vectors. Concatenate the low-dimensional operating condition feature vectors with the multi-source parameter time series sequences according to their dimensions to form an input tensor, which is then input into the LSTM layer. A 12-dimensional core feature dimension was determined, and the number of neurons in the input layer was set to 12. Forty-eight types of work condition labels, such as uniform velocity / wave height, were converted into 48-dimensional sparse vectors using one-hot encoding. These vectors were then mapped into 8-dimensional dense vectors in the embedding layer to eliminate dimensional redundancy. The 8-dimensional work condition vectors were concatenated with the 12-dimensional parameter time series according to their dimensions to form an input tensor (time step × 20), which was then input into the LSTM layer. This allows the model to perceive the work condition background from the initial stage, improving its ability to adapt to dynamic work conditions.

[0032] Step S12222: Set up 2-3 LSTM hidden layers. The number of neurons in the first hidden layer is 4-6 times the feature dimension. The number of neurons in the second / third layer decreases by 10%-20% layer by layer. Connect a dropout layer to the output of each LSTM hidden layer. Divide the ship's operating condition complexity level based on historical data and configure the corresponding dropout rate for different levels of operating conditions. Identify the current ship operating condition complexity in real time and match the corresponding dropout rate. Three LSTM hidden layers are set up. The number of neurons in the first layer is 5 times that of the core feature dimension (12) (60 neurons). The number of neurons in the second layer is reduced by 17% to 50 neurons, and the number of neurons in the third layer is reduced by 20% to 40 neurons. Each layer's output is connected to a dropout layer. Based on historical data, three types of working conditions are divided into high (turning + severe surge), medium (acceleration + moderate surge), and low (uniform speed + calm) complexity, corresponding to dropout rates of 0.4, 0.2, and 0.1, respectively. The current working condition is identified in real time, and the dropout rate is automatically matched to balance feature preservation and overfitting suppression.

[0033] Step S12223: After mapping the dimension of the output tensor of the last LSTM hidden layer into the fully connected layer, the output layer is constructed in two ways. The quantitative output branch is connected to the fully connected layer and a linear activation function is used. The output dimension is matched with the methanol concentration prediction time step and the specific predicted value of methanol concentration is output. The qualitative output branch is connected to the fully connected layer and a Softmax activation function is used. The output dimension is set to 3-dimensional to output the classification result of the concentration change trend. The final LSTM layer outputs a 40-dimensional tensor, which is then mapped to a 30-dimensional feature vector by a fully connected layer (corresponding to 5 minutes × 6 seconds / point). The output layer is constructed in two ways: the quantitative output branch uses a fully connected layer + linear activation function to output 30 concentration time-series values ​​(unit: mg / L); the qualitative output branch uses a fully connected layer + Softmax activation function to output a 3-dimensional probability vector (corresponding to rising / falling / stationary), and the maximum probability is taken as the trend result, simultaneously meeting the needs of precise control and rapid decision-making.

[0034] Step S12224: Add a working condition attention layer between the hidden layer and the dual output layer. Construct a training set based on the ship's historical working conditions and methanol concentration correlation data. Train the working condition type and parameter weight matrix that match the output dimension and parameter dimension of the hidden layer through supervised learning. When the network runs, identify the current ship working condition type, retrieve the corresponding parameter weight vector and multiply it element by element with the output tensor of the hidden layer to strengthen the parameter feature signal. A working condition attention layer is added between the hidden layer and the dual-output layer. A "working condition-parameter weight" training set is constructed based on three years of historical data. A 40×12-dimensional weight matrix is ​​trained through supervised learning (matching the 40-dimensional output of the hidden layer with the 12-dimensional parameters). During network operation, the current working condition type is identified in real time, and the corresponding 12-dimensional parameter weight vector is retrieved and multiplied element-wise with the 40-dimensional output tensor of the hidden layer to amplify the feature signals of core parameters (such as wave height and pump power) and strengthen their influence on the prediction results.

[0035] Step S12225: Connect the input layer, LSTM hidden layer, dropout layer, working condition attention layer, and dual output layer to build the overall network architecture. Input the full working condition samples of the ship into the network for training, monitor the prediction effect under different working conditions in real time, and adjust the dropout rate and attention weight of the corresponding working conditions accordingly. After training, select test samples of typical working conditions of the ship for full coverage verification.

[0036] The input layer, three LSTM hidden layers, a dynamic dropout layer, a working condition attention layer, and a dual output layer are connected in series according to the data flow to form a complete network architecture. Training is performed using 120,000 samples from 48 different working conditions, with real-time monitoring of the prediction accuracy for each condition: if the MAE (Modified Effectiveness) for a certain condition is greater than 3%, the corresponding dropout rate is increased by 0.05, and the attention weights are adjusted. After training, 2000 typical working condition test samples are selected for full coverage validation to ensure that the quantitative MAE for all working conditions is ≤3% and the trend accuracy is ≥92%.

[0037] Step S1223: Select the mean absolute error loss function for the quantitative output layer and the cross-entropy loss function for the qualitative output layer. Construct a mapping rule for operating conditions and loss weights. Dynamically adjust the weights of the two types of losses according to the ship operation scenario. Sum the two types of losses according to their weights to obtain the total loss function and add a regularization term to suppress overfitting. For the quantitative output layer, the MAE loss function (strong robustness against outliers) is selected, while for the qualitative output layer, the cross-entropy loss function (adapted to classification tasks) is selected. A "operating condition-loss weight" mapping rule is constructed: for near-shore navigation (strict emission regulations), the quantitative weight is 0.7, and the qualitative weight is 0.3; for ocean navigation (large fluctuations in operating conditions), the quantitative weight is 0.4, and the qualitative weight is 0.6; when operating conditions change abruptly, the qualitative weight is temporarily increased to 0.7. The total loss function = quantitative loss × weight + qualitative loss × (1 - weight) + L2 regularization term (λ = 1e-5) to suppress overfitting.

[0038] Step S1224: Initialize the weights of the LSTM layer and the fully connected layer using a preset initialization method, and initialize the working condition label embedding layer using random distribution. Select an adaptive optimizer and configure a learning rate with adaptive decay for working conditions. At the same time, add a learning rate restart mechanism, dynamically adjust the batch size according to the working condition sample size, and train the network using a phased training method based on working conditions. During the training process, monitor the validation set loss under different working conditions. If the prediction accuracy does not meet the standard, adjust the network parameters and retrain. After training, verify through full coverage of working conditions to ensure that the prediction requirements for all working conditions of the ship are met.

[0039] The weights of the LSTM and fully connected layers are initialized using Xavier, and the work condition label embedding layers are initialized using a random normal distribution (mean 0, variance 0.01). The Adam optimizer is used with an initial learning rate of 0.001. For high-complexity work conditions, the learning rate decays by 10% every 10 rounds, and for low-complexity work conditions, it decays by 10% every 20 rounds. If the validation loss does not decrease for 5 consecutive rounds, the process restarts at 50% of the current rate. The batch size is dynamically adjusted according to the number of work condition samples (16 for rare work conditions and 32 for regular work conditions). Training is conducted in stages using "low, medium, and high" work conditions, and deployment is performed after validation is successful across all work conditions.

[0040] Step S123: Input the preprocessed samples into the LSTM model for training, introduce an early stopping method to avoid the risk of overfitting, and construct a verification system with methanol concentration prediction accuracy as the core. If the model prediction accuracy does not meet the control requirements for methanol concentration in ship bilge water, the model hyperparameters will be automatically adjusted and the model will be retrained until the accuracy meets the standard. The preprocessed training set (100,000 records) is input into the LSTM model, and an early stopping mechanism (patience=10, stopping if the validation loss does not decrease for 10 consecutive rounds) is introduced to avoid overfitting. A validation system is constructed: the quantitative metric is MAE (≤3%), and the qualitative metric is trend accuracy (≥92%), covering 48 operating conditions. If the accuracy does not meet the standard, the hyperparameters are automatically adjusted (learning rate ±30%, dropout rate ±0.05), and retraining is performed until the requirements are met. The training process log is automatically stored on the ship's server.

[0041] Step S124: Collect multi-source parameters of ship navigation, sea state, bilge water quality, and equipment operation in real time. After preprocessing by normalization method, input them into the trained LSTM model, output methanol concentration prediction results, and construct dynamic accuracy verification rules. When the prediction accuracy is not up to standard, automatically retrieve the newly added actual ship operating condition data from the operating condition and concentration correlation feature library to supplement the training set, re-execute the training process, and update the system model parameters. Multi-source parameters are collected at a sampling frequency of 500ms. After outlier removal with 3σ and Min-Max normalization, the data are input into an LSTM model, which outputs 5-minute concentration predictions and trends. A "stratified operating condition + dual-dimensional" verification rule is constructed: high operating condition MAE ≤ 5%, accuracy ≥ 85%; medium / low operating condition MAE ≤ 3%, accuracy ≥ 90%; if 20% of samples fail to meet the standard for 10 consecutive times or within 30 minutes, the training set is automatically supplemented with newly added data from the past 7 days, and incremental training is used to update the model parameters to ensure real-time adaptation to changes in operating conditions.

[0042] Step S1241: Deploy the ship-side multi-dimensional data acquisition module, collect ship navigation status, sea state, bilge water quality, and equipment operation status parameters at the same sampling frequency as the model training, use the 3σ principle to remove outliers of the original parameters, and map each parameter to the [0,1] interval through Min-Max normalization based on the historical statistical extreme values ​​of the parameters in the working condition and concentration correlation feature library to eliminate dimensional differences; Deploy MarineSense-800 multi-dimensional acquisition modules to collect parameters such as navigation status and sea state at a frequency of 500ms; use the 3σ principle to remove outliers in parameters such as current and concentration; based on the historical extreme values ​​of parameters in the operating condition-concentration feature library (such as methanol concentration 0-50mg / L), use the Min-Max normalization formula (normalized value = (original value - minimum value) / (maximum value - minimum value)) to map each parameter to the [0,1] interval, eliminate dimensional differences, and ensure the consistency of input data.

[0043] Step S1242: The normalized multi-source parameters are concatenated into a time series according to the time window length during model training, and integrated into the encoding vector of the current ship operating condition label to form an input tensor. This tensor is then input into the trained LSTM model. The model outputs a quantitative prediction value of methanol concentration in the future period and reverses it to restore the actual concentration value. At the same time, it outputs a qualitative classification result of the concentration change trend. After integration, the result is output to the ship's bilge water control system. The normalized 12-dimensional parameters are concatenated into a time series within a 300-second time window and integrated into the current operating condition's 8-dimensional embedding vector to form an input tensor of (300, 20). This tensor is then input into the trained LSTM model, which outputs 30 normalized concentration values. These values ​​are then converted back to mg / L quantitative results using the inverse normalization formula (actual value = normalized value × (50 - 0) + 0), along with three trend classification results. The integrated results are then output to the ship's bilge water control system via the Modbus RTU protocol, with a latency ≤ 500ms.

[0044] Step S1243: Construct a hierarchical, two-dimensional accuracy verification rule for working conditions. Divide the working conditions of ships into three levels: high, medium, and low. Configure different quantitative dimension average absolute error thresholds and qualitative dimension trend matching accuracy thresholds for different levels of working conditions. Set continuous failure judgment conditions. If the prediction accuracy exceeds the threshold for a single working condition for a set number of consecutive preset times, or if the prediction accuracy of a preset proportion of the prediction samples within a preset time range in the entire working condition range is not up to standard, it is judged as continuous failure of prediction accuracy. Construct a "layered working condition + dual-dimensional" accuracy verification rule: Divide the working condition into three levels: high, medium, and low. For the quantitative dimension, the MAE for high working conditions is ≤5%, and for medium / low conditions it is ≤3%. For the qualitative dimension, the accuracy for high working conditions is ≥85%, and for medium / low conditions it is ≥90%. Set continuous failure criteria: If a single working condition exceeds the threshold for 10 consecutive predictions, or if ≥20% of the samples fail to meet the standard within 30 minutes across all working conditions, it is judged as continuous failure, triggering the model retraining process.

[0045] Step S1244: When the accuracy of the judgment continues to be substandard, the system automatically filters the newly added actual ship operating condition data within the preset time period from the operating condition and concentration correlation feature library. Priority is given to filtering samples that are consistent with the type of operating condition that is substandard and contain complete multi-source parameter time series, actual methanol concentration value and operating condition label. The newly added data after filtering is added to the training set and validation set according to the preset ratio. When adding, the proportion of each type of operating condition in the training set is kept balanced. If there are not enough new samples for a certain type of operating condition, virtual samples are generated to supplement the data by slightly perturbing the parameter values. If the accuracy of the judgment continues to fail to meet the standard, the system automatically filters newly added data from the feature library within the past 7 days, prioritizing samples that match the type of non-compliant operating condition and contain complete 12-dimensional parameters, concentration values, and operating condition labels. These samples are then added to the training and validation sets at an 8:2 ratio, ensuring a balanced distribution of the 48 operating condition categories and that rare operating condition samples account for ≥1%. If there are insufficient new samples for a certain type of operating condition, virtual samples are generated by slightly perturbing the parameter values ​​by ±5% to avoid imbalance in the training set distribution.

[0046] Step S1245: Adopt an incremental training strategy, initialize the LSTM model based on the original model parameters, retrain the LSTM model using the supplemented training set, configure hyperparameters to adapt to the working conditions, monitor the accuracy of the validation set during training, stop training when the accuracy of the validation set meets the standard and there is no decrease for a consecutive preset number of rounds, package the retrained model parameters, automatically replace the original parameters in the ship end system, switch to the new parameter model to perform real-time prediction and continuously monitor the accuracy.

[0047] An incremental training strategy was adopted, initializing the model parameters based on the original model and training for 50 epochs using only the supplemented training set. Hyperparameters were configured to adapt to the characteristics of the operating conditions: the learning rate decay rate was increased by 15% for high-complexity operating conditions, and the batch size was set to 16. During training, the accuracy of the validation set was monitored, and training was stopped when the MAE ≤ 3% and showed no decrease for 5 consecutive epochs. The new model parameters were packaged and automatically replaced the original parameters through the ship's local area network. After the switch, the accuracy was continuously monitored to ensure adaptation to the current operating conditions.

[0048] Step S125: Determine the basic control threshold for methanol concentration based on the maritime emission standards, and innovatively construct a two-dimensional threshold adjustment coefficient mapping rule based on operating condition type and concentration prediction trend. Combine the real-time operating conditions of the ship with the concentration prediction results output by the LSTM model to dynamically calculate the control threshold and refresh it frequently. Based on IMO methanol emission standards, a compliance limit of 5 mg / L is set for nearshore areas (4 mg / L for basic warning and 4.75 mg / L for emergency response), and a compliance limit of 10 mg / L is set for offshore areas (8 mg / L for basic warning and 9.5 mg / L for emergency response). A two-dimensional adjustment coefficient matrix of "3 operating conditions × 3 trends" is constructed (e.g., high operating condition + upward trend coefficient 0.945). Operating conditions and model prediction results are collected in real time and calculated according to the formula "dynamic threshold = basic threshold × operating condition coefficient × trend coefficient", with a high frequency refresh of 1 second. The threshold constraint is within 70%-100% of the compliance limit.

[0049] Step S1251: Extract the methanol concentration compliance limits under different scenarios, construct a two-layer basic control threshold of early warning threshold and emergency control threshold based on the compliance limits, mark the navigation area boundaries applicable to each basic control threshold, and enter the basic control threshold and area rules into the ship threshold management library; We analyzed IMO and nearshore / ocean-going / environmentally protected zone methanol emission regulations, extracting compliance limits of ≤5 mg / L for nearshore, ≤10 mg / L for ocean-going, and ≤3 mg / L for special environmental protection zones. Based on these limits, we constructed a two-tiered basic control threshold: early warning threshold = compliance limit × 80%, and emergency control threshold = compliance limit × 95%. We marked the applicable area boundaries for each threshold (12 nautical miles from the coastline for nearshore / ocean-going). The basic thresholds and area rules were entered into the ship threshold management database as a dynamic adjustment benchmark.

[0050] Step S1252: By quantifying the dual-level classification of ship operating conditions and methanol concentration evolution trends, a synergistic and adaptable coefficient combination system is constructed. After being verified and solidified by historical data, the threshold adjustment can accurately respond to the dynamic operating conditions and concentration change patterns of the ship. Based on the complexity of operating conditions, three levels—high, medium, and low—adjustment coefficients were set at 1.05-1.1, 1.0, and 0.95-0.98, respectively. Based on concentration trends, three categories were established—rising, stable, and declining—with trend adjustment coefficients set at 0.9-0.95, 1.0, and 1.02-1.05, respectively. Nine sets of "operating condition × trend" coefficient combination matrices were constructed. Simulation using one year of historical data was conducted to ensure that the adjusted thresholds were not lower than 70% of the compliance limit while reducing false triggering due to operating condition fluctuations. Once verified, the results were solidified into the threshold calculation system.

[0051] Step S1253: Deploy the operating condition identification module to collect ship navigation status, sea state, and equipment load data in real time and automatically determine the complexity level of the current operating condition. Simultaneously receive the quantitative prediction value and qualitative trend result of methanol concentration output by the LSTM model, analyze the trend type and related features, and perform real-time verification on the collected operating condition data and prediction results to remove abnormal data and ensure data continuity. A working condition identification module is deployed to collect navigation status, sea state, and equipment load data at a frequency of 500ms, automatically determining the complexity level of the working condition (delay ≤ 1s); it simultaneously receives the 5-minute concentration prediction values ​​and trend results output by the LSTM model, and analyzes the trend type and growth characteristics. The collected data is verified in real time: if the working condition identification is ambiguous or the trend is contradictory, it is judged as abnormal, and the valid data from the previous period is temporarily used to ensure the continuity of threshold calculation.

[0052] Step S1254: Based on the preset formula, calculate the dynamic early warning threshold and dynamic emergency control threshold through the basic control threshold, the operating condition adjustment coefficient and the trend adjustment coefficient. Set a high frequency refresh frequency. Read the latest operating condition and prediction results and recalculate the threshold each time it is refreshed. At the same time, set hard constraints on the threshold to ensure that the dynamic threshold does not exceed the preset ratio range between the compliance limit and the basic threshold. Push the calculated dynamic threshold to the ship's bilge water control system in real time as the control basis. The dynamic early warning threshold is calculated based on the preset formulas "Dynamic Early Warning Threshold = Basic Early Warning Threshold × Operating Condition Coefficient × Trend Coefficient" and "Dynamic Emergency Threshold = Basic Emergency Threshold × Operating Condition Coefficient × Trend Coefficient". A high-frequency refresh rate of 1 second is set, and the latest operating conditions and prediction results are read each time for recalculation. Hard constraints are set: the dynamic threshold must not be higher than the compliance limit and must not be lower than 70% of the basic threshold; if it exceeds these limits, it is corrected to the boundary value. After calculation, the data is pushed to the control system via the industrial bus as the basis for control.

[0053] Step S1255: Real-time statistical analysis of the application effect of dynamic thresholds, including threshold trigger accuracy and false trigger rate; periodically iterative optimization of coefficient combination matrix based on statistical results; and synchronously update basic control thresholds in conjunction with updates to maritime emission standards.

[0054] Real-time statistics are collected on the accuracy rate of dynamic threshold triggering (whether the concentration is close to the compliance limit after the emergency threshold is triggered) and the false triggering rate. The coefficient matrix is ​​iteratively optimized weekly based on the statistical results (such as adjusting the high-operating-condition coefficient ±0.02). The basic control threshold is updated quarterly in accordance with the updates to the marine emission standards (such as adding an environmental protection zone limit of 3 mg / L) to ensure that the dynamic threshold always meets the latest regulatory requirements and operating condition adaptation needs.

[0055] Step S126: Collect the operating load parameters of the bilge water treatment equipment and calculate the real-time load rate, construct dynamic threshold range and equipment treatment intensity association rules, match the corresponding treatment intensity based on the generated dynamic threshold and send control instructions to the equipment controller, add an equipment load safety constraint mechanism, and automatically reduce the treatment intensity when the equipment load is close to the rated value.

[0056] Hall current sensors and electromagnetic flow sensors are deployed to collect equipment operating parameters. Real-time load rate is calculated using the power method (load rate = actual power / rated power × 100%). A correlation rule is constructed between four dynamic threshold ranges and four levels of processing intensity (pump frequency 50%-90%), with a condition correction factor added. Based on the dynamic threshold matching processing intensity, Modbus RTU protocol control commands are issued. Two load safety thresholds are set: 85% (early warning) and 90% (mandatory). Upon triggering, the processing intensity is automatically reduced to balance concentration control and equipment safety.

[0057] Step S1261: Deploy the data acquisition module to collect the operating parameters of the bilge water treatment equipment, such as motor operating current / power, medium processing flow rate, inlet and outlet pressure, and reagent dosing amount, through the equipment PLC, current transformer, and flow sensor. Preprocess the collected raw parameters, filter out instantaneous fluctuation abnormal values, and use historical period averages to fill in missing parameters. Calculate the real-time load rate based on the preprocessed parameters using the power method. The system deploys ACS800 current transformers and DFM-50 electromagnetic flow sensors, and collects operating parameters such as motor current / power and medium flow rate via the equipment's PLC. A 5-second moving average filter is used to eliminate instantaneous fluctuations and outliers; if a single parameter is missing, the average of the previous five cycles is used to complete the calculation. Based on the pre-processed parameters, the real-time load factor is calculated using the power method (formula: load factor = (actual operating power / rated power) × 100%), with an accuracy ensured to be ±0.5%FS.

[0058] Step S1262: Connect to the output results of the dynamic control threshold of methanol concentration, divide the dynamic threshold intervals, each interval corresponds to a clear concentration control target, decompose the equipment processing intensity into quantifiable and controllable dimensions, define multiple levels of processing intensity, construct a one-to-one correspondence rule between the dynamic threshold interval and the processing intensity level, add the operating condition correction factor, and verify the rules through simulation using historical equipment operation data to ensure the concentration control effect and equipment load safety, and solidify the verified rules into the control system; The system integrates with dynamic threshold outputs, dividing the data into four intervals: low (<80% warning), medium (80%-100% warning), high (warning-emergency), and emergency (>emergency). The treatment intensity is broken down into three dimensions: pump frequency, valve opening, and reagent dispensing rate, defining four intensity levels (pump frequency 50%, 65%, 80%, 90%). A one-to-one correspondence rule between intervals and intensities is established, incorporating a working condition correction factor (increasing the level by 0.5 for high working conditions). Simulation verification using three months of historical data ensures that concentration control meets standards and the load does not exceed safe limits. Once verification is successful, the rules are solidified.

[0059] Step S1263: Receive the dynamic threshold calculation results of methanol concentration in real time, identify the current threshold range, match the initial treatment intensity level according to the association rules, fine-tune the initial treatment intensity in combination with the real-time load rate, convert the matched treatment intensity level into the control instructions of the device controller communication protocol, and clarify the specific control requirements of each control dimension. The system receives dynamic threshold calculation results in real time, identifies the corresponding interval, and matches the initial processing intensity level. It then fine-tunes the intensity level based on real-time load rate: lowering the level by 0.5 when the load rate is ≥70% and raising it by 0.5 when it is ≤50%. The adjusted intensity level is converted into control commands adapted to the Modbus RTU protocol, specifying concrete values ​​for pump frequency, valve opening, and chemical dosing rate (e.g., "pump frequency 75%, valve opening 85%)". The command generation delay is ≤500ms to ensure timely response.

[0060] Step S1264: Based on the rated load rate of the equipment, set two levels of safety constraint thresholds and corresponding processing intensity reduction gradients, monitor the real-time load rate of the equipment in real time, and trigger corresponding constraint operations according to the threshold range of the load rate, including reducing the processing intensity, pushing early warning prompts, pausing non-core control dimensions, locking the processing intensity increase operation, and designing a callback mechanism to gradually restore the processing intensity to the target level when the load rate falls back to the preset range. Based on the equipment's rated load rate, 85% is set as the warning threshold and 90% as the mandatory reduction threshold, corresponding to a reduction in processing intensity of 1 or 2 levels (down to level 1). Real-time monitoring of the load rate: when ≥85% and <90%, the intensity is reduced by 1 level and a warning is pushed; when ≥90%, the intensity is reduced by 2 levels, reagent dispensing is suspended, and upward adjustment operations are locked. If the load rate is ≤75% for 10 consecutive seconds, the processing intensity is gradually restored according to the original threshold range (0.5 levels / 5 seconds) to avoid a sudden increase in load.

[0061] Step S1265: Send control commands to the equipment controller via the industrial communication bus, collect equipment execution feedback in real time to confirm that the commands have been executed, monitor the changes in methanol concentration in the bilge water and the equipment load rate simultaneously, dynamically adjust the treatment intensity based on the monitoring results, and periodically calculate the command execution accuracy, the number of load overruns, and the concentration control compliance rate. Iteratively optimize the association rules and safety constraints based on the statistical data to lower the gradient.

[0062] Control commands are sent to the equipment controller via an industrial communication bus, and real-time feedback such as valve position and pump frequency is collected to confirm that the commands are executed correctly. Concentration changes and load rates are monitored at a frequency of 1 second: if the concentration does not meet the standard and the load does not exceed the safety threshold within 5 minutes, the level is increased by 0.5; if the concentration meets the standard but the load is close to the warning threshold, the level is decreased by 0.5. Daily statistics are compiled on command execution accuracy, number of load exceedances, and concentration compliance rate. The correlation rules and adjustment gradients are iteratively optimized monthly to balance control effectiveness and equipment safety.

[0063] Step S130: Introduce edge computing technology to complete data preprocessing at the acquisition terminal nearby, use industrial-grade real-time Ethernet for data transmission, connect the monitoring, alarm, processing and emission subsystems to a unified distributed collaborative control platform, assign each subsystem a unique communication identifier and data interaction interface, and clarify the linkage triggering conditions and execution priorities; An STM32H750 edge computing module is integrated into the acquisition terminal to perform data filtering, normalization, and outlier removal preprocessing locally, with a latency of ≤500ms. Profinet industrial real-time Ethernet (communication rate 1Gbps) is used to transmit data, with a packet loss rate of ≤0.01%. A distributed collaborative control platform (based on Docker container architecture) is built to connect six subsystems, including monitoring and alarm systems. Each subsystem is assigned a unique UUID communication identifier, and the unified data interaction interface is the Modbus RTU protocol. The linkage triggering conditions (such as concentration exceeding warning, equipment failure) and execution priorities are clearly defined: emergency response > emergency handling > main processing > preprocessing, to ensure cross-system collaborative response.

[0064] Step S140: Based on the methanol concentration prediction results and real-time detection values, the linkage levels are divided. Each level corresponds to different actions such as pretreatment, main treatment, emergency treatment, and emergency response to ensure rapid response of the linkage process. Based on the methanol concentration prediction results and real-time detection values, four linkage levels are defined: Level 1 (predicted value < 80% of the warning threshold) corresponds to pretreatment, fine-tuning the reagent dosing rate (±5%), with a response time ≤ 1s; Level 2 (warning threshold 80%-100%) corresponds to main treatment, increasing the pump frequency of the treatment equipment to 65%-80%, with a response time ≤ 500ms; Level 3 (warning threshold - emergency threshold) corresponds to emergency treatment, activating backup treatment equipment, with a response time ≤ 300ms; Level 4 (> emergency threshold) corresponds to emergency response, locking key valves and initiating emergency venting, with a response time ≤ 200ms. The actions of each level switch automatically without manual intervention.

[0065] Step S150: Establish a composite fault feature library, use decision tree algorithm to analyze real-time multi-source data, realize automatic identification of composite faults, design corresponding intelligent interlocking logic for different types of composite faults, start the matching processing flow and take control measures to lock key valves. Twenty types of composite faults in ship bilge water systems (such as "pump jamming + methanol concentration exceeding emergency threshold") were identified, and a composite fault feature library containing fault IDs and 30 feature combinations was constructed. The CART decision tree algorithm (accuracy ≥92%) was used to analyze real-time multi-source data and automatically identify fault types and levels. Differentiated intelligent interlocking logic was designed according to four fault levels, and locking commands for key valves (feed valves, isolation valves, etc., 10 types) were defined. Commands were issued through the Modbus TCP protocol, and if the command execution failed, the manual control backup path was activated to ensure that fault prevention and control were comprehensive.

[0066] Step S151: Identify the types of composite faults in the ship's bilge water system. Based on historical fault cases, equipment manuals, and maintenance records, extract the fault symptoms, associated multi-source parameters, fault triggering sequence relationships, and fault levels for various composite faults. Standardize the extracted features, unify parameter dimensions and judgment thresholds, clarify the logical relationships between features, and store the processed feature data according to fault type. Construct a composite fault feature library containing fault ID, feature combination, associated parameters, fault level, and historical processing records. Regularly incorporate new fault cases to update feature dimensions and association rules. Based on historical fault cases, equipment manuals, and maintenance records from the past 5 years, 20 types of composite faults, such as "equipment fault + parameter anomaly" and "sudden change in operating conditions + pipeline fault," were identified. Fault symptoms (such as current fluctuations), 12 types of associated multi-source parameters, triggering timing relationships, and 4 fault levels were extracted. Min-Max normalization was used to unify the parameter dimensions (mapped to [0,1]) and the logical relationships between features were clarified (such as "power drop + concentration increase" corresponding to pump faults). A composite fault feature library in PostgreSQL format was constructed, containing 8 fields such as fault ID and feature combination. New fault cases were added monthly to update the feature dimensions and association rules.

[0067] Step S152: Select the decision tree algorithm, use historical data in the composite fault feature library as the training set, take the composite fault type as the target variable, and the fault symptoms and related parameters as the input features. Divide the training set and test set and train the model. Optimize the model's generalization ability through pruning operations. After the test set verifies that it meets the standard, deploy it. Collect multi-source data in real time and extract real-time feature vectors. After preprocessing, input the pre-trained decision tree model. The model infers and outputs the composite fault type, confidence level and fault level in real time. Filter the effective fault identification results according to the preset judgment conditions. The CART decision tree algorithm was selected, and 100,000 historical data points from a composite fault feature library were used as the training set (divided into training and test sets in an 8:2 ratio). Composite fault type was used as the target variable, and fault symptoms and associated parameters were used as input features. The model was optimized using a combination of pre-pruning (tree depth limited to ≤15) and post-pruning. The test set validation required a fault identification accuracy of ≥92% and a confidence level of ≥85%. After deployment, multi-source data was collected at a frequency of 500ms, and real-time feature vectors (after preprocessing) were extracted and input into the model. The valid judgment condition was set as "three consecutive collection cycles + confidence level ≥85%". The fault type, confidence level, and grade were output.

[0068] Step S153: Based on the compound fault type and fault level, formulate differentiated intelligent interlocking logic, clarify the interlocking target and execution priority, and set interlocking action sequences including equipment control, pipeline regulation and alarm notification for different levels of compound faults. All interlocking logics are embedded with working condition adaptation factors, and the execution sequence is dynamically adjusted according to the real-time working condition complexity. Differentiated intelligent interlocking logic is formulated based on four levels of composite faults (minor / moderate / severe / urgent), clearly defining the interlocking objectives (fault isolation / risk prevention, etc.) and execution priorities (urgent > severe > moderate > minor). Emergency fault interlocking actions include: cutting off the power supply to the faulty equipment, locking the feed / discharge valves, shutting off the chemical dosing system, initiating emergency venting, and sending audible and visual alarms plus SMS messages. Severe fault actions include: reducing the system load to 30%, locking the upstream and downstream isolation valves of jammed valves, and opening backup pipelines. All logic incorporates operating condition adaptation factors, prioritizing valve locking under high sea conditions, and then adjusting the load to avoid secondary faults.

[0069] Step S154: Identify the list of key valves corresponding to various composite faults and the valve action instructions under different faults, bind the valve action logic with the interlock logic, send locking instructions to the valve controller through the industrial communication protocol, collect valve feedback signals in real time, start the backup interlock path in case of instruction execution failure, start the matching fault handling process simultaneously, push maintenance notifications and provide handling suggestions according to the fault level, monitor system parameters in real time, and automatically unlock key valves or upgrade interlock measures according to the parameter recovery status; Clearly define the list of key valves corresponding to various complex faults (including 10 categories such as feed valves and isolation valves), and label the valve action commands (close / lock / hold); issue locking commands via ModbusTCP / Profinet protocol, and automatically retry twice if no valve position feedback is received within 3 seconds. If it still fails, push a valve fault alarm and activate the manual control backup path; simultaneously push maintenance notifications (emergency fault SMS + audible and visual alarms) according to the fault level, providing fault characteristics and handling suggestions; monitor system parameters in real time, and automatically unlock the valve if the parameters recover within 30 minutes. If the situation does not improve, upgrade the interlocking measures.

[0070] Step S155: After the fault handling is completed, the fault identification, interlock execution, valve status changes and handling results are automatically recorded to form a fault closed-loop log. Based on the closed-loop log, the fault identification accuracy, interlock logic execution effectiveness and valve locking response speed are analyzed. The composite fault feature library and decision tree model are backtracked and optimized. The frequency of composite fault occurrence and handling effect are statistically analyzed regularly to optimize the interlock logic and key valve locking list.

[0071] After the fault is handled, 15 fields, including fault identification time, interlock action details, and valve status changes, are automatically recorded to form a fault closed-loop log. Based on the log analysis, the fault identification accuracy, interlock execution effectiveness, and valve locking response speed (required to be ≤3 seconds) are evaluated. If the false alarm rate is >5%, the feature library (adding new symptoms) and decision tree model (adjusting feature weights) are backtracked and optimized. The frequency of compound faults is statistically analyzed monthly, and the interlock logic (adjusting action sequence) and key valve locking list are optimized quarterly to improve the reliability of prevention and control.

[0072] Step S160: Set up a response time monitoring unit on the collaborative control platform to monitor the response status of each link in data processing, command transmission and equipment execution in real time. When the response status of each link does not meet the requirements, automatically start the backup channel or redundant equipment and record relevant time information to optimize the communication link and control program.

[0073] A response time monitoring unit is set up in the distributed collaborative control platform to monitor three key links in real time: data processing (threshold ≤ 500ms), command transmission (threshold ≤ 200ms), and equipment execution (threshold ≤ 300ms). When any link fails to meet the timeliness requirements, the system automatically switches to the backup communication channel (e.g., Profinet switches to EtherNet / IP) or activates redundant equipment (backup pumps / valves). Timeliness data (including timeout count and switching duration) is recorded in real time, and an optimization report is generated monthly to adjust the communication link bandwidth allocation and control program execution logic to ensure system response stability.

[0074] Based on the same inventive concept, please refer to Figure 2 The diagram shows a schematic block diagram of an unconventional water source problem tracing system 100 using a knowledge graph, which is provided in an embodiment of this application for performing the above-mentioned integrated processing method for intelligent monitoring and linkage control of methanol-containing bilge water. The unconventional water source problem tracing system 100 using a knowledge graph may include a communication unit 110, a machine-readable storage medium 120, and a processor 130.

[0075] In this embodiment, both the machine-readable storage medium 120 and the processor 130 are located within the unconventional water source problem tracing system 100 utilizing a knowledge graph and are separately configured. However, it should be understood that the machine-readable storage medium 120 may also be independent of the unconventional water source problem tracing system 100 utilizing a knowledge graph and may be accessed by the processor 130 via a bus interface. Alternatively, the machine-readable storage medium 120 may also be integrated into the processor 130 and may communicate with external systems via the communication unit 110.

[0076] The processor 130 is the control center of the knowledge graph-based unconventional water source problem tracing system 100. It connects to various parts of the system via various interfaces and lines. By running or executing software programs and / or modules stored in the machine-readable storage medium 120, and by calling data stored in the machine-readable storage medium 120, it performs various functions and processes data of the knowledge graph-based unconventional water source problem tracing system 100, thereby providing overall monitoring of the system. Optionally, the processor 130 may include one or more processing cores; for example, the processor 130 may integrate an application processor and a modem processor, wherein the application processor mainly handles the operating system, user interface, and applications, and the modem processor mainly handles wireless communication. It is understood that the modem processor may also not be integrated into the processor. The machine-readable storage medium 120 is used to store machine-executable instructions for executing the scheme of this application, and the processor 130 is used to execute the machine-executable instructions stored in the machine-readable storage medium 120 to realize the integrated processing method for intelligent monitoring and linkage control of methanol-containing bilge water provided in the aforementioned method embodiments.

[0077] It should be noted that, in order to simplify the description of the present invention and thus help to understand one or more embodiments of the invention, multiple features may sometimes be grouped into one embodiment, drawing or description thereof in the foregoing description of the embodiments of the present invention.

Claims

1. An integrated intelligent monitoring and control method for treating methanol-containing bilge water, characterized in that: Includes the following steps: A data acquisition terminal integrating multiple types of sensors is deployed at the ship's location to collect navigation status, sea state, bilge water quality, and equipment operating parameters. A high-frequency sampling mode is adopted, and historical relevant data is sorted out and classified and labeled according to navigation status, sea state level, and bilge water load range. Methanol concentration characteristic parameters are extracted, and a database of operating conditions and concentration correlation features is formed after data cleaning. The LSTM model is trained based on the feature library of working conditions and concentration correlation. Multiple real-time parameters are input, and the methanol concentration change trend and prediction results are output. When the model prediction accuracy is not up to standard, retraining is automatically started. Dynamic control thresholds are generated based on the model prediction results and real-time working conditions. The threshold range is adjusted to adapt to different working conditions, and the real-time load of the associated processing equipment is matched with the processing intensity. Edge computing technology is introduced to complete data preprocessing at the acquisition terminal nearby, and industrial-grade real-time Ethernet is used for data transmission. The monitoring, alarm, treatment and emission subsystems are connected to a unified distributed collaborative control platform, and each subsystem is given a unique communication identifier and data interaction interface, and the linkage triggering conditions and execution priorities are clearly defined. Based on the methanol concentration prediction results and real-time detection values, the linkage levels are divided, and each level corresponds to different actions such as pretreatment, main treatment, emergency treatment, and emergency response to ensure rapid response of the linkage process. Establish a composite fault feature library, use decision tree algorithm to analyze real-time multi-source data to achieve automatic identification of composite faults, design corresponding intelligent interlocking logic for different types of composite faults, start matching processing flow and take control measures to lock key valves; A response time monitoring unit is set up in the collaborative control platform to monitor the response status of each link in data processing, command transmission and equipment execution in real time. When the response status of each link does not meet the requirements, the backup channel or redundant equipment is automatically activated, and relevant time information is recorded for optimization of communication links and control programs.

2. The integrated intelligent monitoring and control method for methanol-containing bilge water according to claim 1, characterized in that: The aforementioned data acquisition terminal, deployed at the ship's location and integrating multiple types of sensors, collects navigation status, sea state, bilge water quality, and equipment operating parameters. It employs a high-frequency sampling mode, organizes historical data, and categorizes and labels it according to navigation status, sea state level, and bilge water load range. Methanol concentration characteristic parameters are extracted, and after data cleaning, a database of operating conditions and concentration correlation features is formed, including: We selected integrated data acquisition terminals suitable for high vibration, salt spray, and explosion-proof environments on ships. The terminals include four types of anti-interference wide-temperature sensors for navigation status, sea state, bilge water quality, and equipment operation status. They are deployed in the engine room power area, storage tank inlet and outlet, processing equipment inlet and outlet, and pipeline nodes, with no monitoring blind spots and redundant installation interfaces reserved at the points. The unified sensor interface is ModbusRTU or CANopen protocol, and the terminal integrates a data synchronization control module and is configured with a local cache module for raw data storage. Set the sampling trigger mode to timed and event-linked, and the data format includes millisecond-level timestamps, sensor type, parameter values, and sensor working status identifiers. Enable the terminal preprocessing function to complete data format conversion, verification, and removal of invalid data exceeding the range. Historical data, navigation logs, equipment operation records, and inspection records are collected through ship local area networks or satellite communications. Standardized rules are established to unify parameter units and formats, which are then converted into a unified structured dataset and sorted by time. The navigation status is divided into constant speed / acceleration / deceleration / turning; the sea state is divided into calm / slight swell / moderate swell / severe swell; the bilge water load is divided into low / medium / high; and the datasets are labeled according to the standard to form a categorized dataset. Data sets were grouped by label category, and the variation pattern of methanol concentration was analyzed. The average concentration and fluctuation range feature parameters were extracted and statistically quantified to form a structured feature parameter set. We construct cleaning rules based on numerical rationality, logical consistency, and trend continuity; process abnormal data in a hierarchical manner; integrate data according to operating condition, characteristic parameters, and methanol concentration changes to build a feature library; and use distributed storage.

3. The integrated intelligent monitoring and control method for methanol-containing bilge water according to claim 1, characterized in that: The LSTM model is trained based on a feature library related to operating conditions and concentration. It takes multi-source real-time parameters as input and outputs the methanol concentration change trend and prediction results. If the model's prediction accuracy is insufficient, it automatically restarts retraining. Based on the model's prediction results and real-time operating conditions, it generates dynamic control thresholds, adjusts the threshold range to suit different operating conditions, and correlates the real-time load of the processing equipment with the processing intensity, including: Multi-source parameters such as navigation status, sea state, bilge water quality, and equipment operation status are extracted from the operating condition and concentration correlation feature library. The Min-Max normalization method is used to eliminate the dimensional differences between different parameters. For the dynamic operating conditions of ships, a time series sliding slicing strategy is designed. Sample pairs of multi-source parameter input sequences and methanol concentration output sequences are extracted according to the time window and sliding step size mode, so that the time dimension of the sample pairs is adapted to the methanol concentration prediction cycle requirements. The training set and validation set are divided as needed. A methanol concentration time series prediction LSTM network adapted to complex dynamic operating conditions of ships was constructed. The input layer dimension of the network is matched with the feature dimension of the multi-source parameters. A dropout layer is added to the hidden layer to suppress model overfitting. A dual-output layer design is adopted to simultaneously output the specific predicted value of methanol concentration and the classification result of concentration change trend, which respectively meet the needs of quantitative control and qualitative decision-making. Multiple types of loss functions are integrated to optimize the training objective. The preprocessed samples are input into the LSTM model for training. An early stopping rule is introduced to avoid the risk of overfitting. A verification system with methanol concentration prediction accuracy as the core is constructed. If the model prediction accuracy does not meet the control requirements for methanol concentration in ship bilge water, the model hyperparameters are automatically adjusted and retrained until the accuracy meets the standard. The system collects multi-source parameters such as ship navigation, sea state, bilge water quality, and equipment operation in real time. After preprocessing by normalization, the parameters are input into the trained LSTM model, which outputs methanol concentration prediction results. Dynamic accuracy verification rules are constructed. When the prediction accuracy fails to meet the standard, the system automatically retrieves new ship actual operating condition data from the operating condition and concentration correlation feature library to supplement the training set, re-executes the training process, and updates the system model parameters. Based on maritime emission standards, a basic control threshold for methanol concentration is determined. An innovative two-dimensional threshold adjustment coefficient mapping rule is constructed, which combines the ship's real-time operating conditions with the concentration prediction results output by the LSTM model. The control threshold is dynamically calculated and refreshed at high frequency. The system collects the operating load parameters of the bilge water treatment equipment and calculates the real-time load rate. It constructs dynamic threshold ranges and equipment treatment intensity association rules, matches the corresponding treatment intensity based on the generated dynamic thresholds, and sends control commands to the equipment controller. It also adds an equipment load safety constraint mechanism, which automatically reduces the treatment intensity when the equipment load approaches the rated value.

4. The integrated intelligent monitoring and control method for methanol-containing bilge water according to claim 3, characterized in that: To address the dynamic operating characteristics of ships, a time-series sliding slicing strategy is designed. This strategy extracts sample pairs of multi-source parameter input sequences and methanol concentration output sequences according to a time window and sliding step size pattern. This ensures that the time dimension of the sample pairs adapts to the methanol concentration prediction cycle requirements. Training and validation sets are then divided as needed, including: Define the methanol concentration prediction period, set the base duration of the time window with a reasonable multiple that matches the prediction period, set the base value of the sliding step size with the minimum change period of the ship's operating conditions, establish a database of operating conditions and parameter change characteristics, statistically analyze the fluctuation amplitude and change rate of multi-source parameters under different operating conditions, and mark the effective duration of parameter characteristics corresponding to each operating condition. Extract multi-source parameters and methanol concentration data from the operating condition and concentration correlation feature library that have been normalized to Min-Max, including navigation status, sea state, bilge water quality, equipment operating status, and methanol concentration data. Using the unified high-precision timestamp of ship GPS time as the reference axis, correct sensor sampling delay, remove abnormal data points, and complete the sampling interval to generate a time-series dataset with a continuous time axis and consistent sampling frequency. Calculate the change rate of multi-source parameters in real time to classify the operating condition complexity level, and adaptively adjust the window duration and sliding step size according to the level. Extract sample pairs of multi-source parameter input sequences and methanol concentration output sequences, and retain valid samples by filtering through information entropy. The effective samples are classified according to navigation status, sea state level, and bilge water load range. For each combination of operating conditions, training and validation sets are extracted in a stratified manner according to a preset ratio. For rare combinations of operating conditions with insufficient sample size, virtual samples are generated to supplement the training set by slightly perturbing the parameter values, ensuring that the sample size of each combination of operating conditions is balanced. After accumulating a preset number of new samples, the predictive contribution of sample pairs under different working condition combinations is calculated. If the predictive accuracy of a certain working condition does not meet the preset standard, the window / step parameters corresponding to that working condition are recalibrated, and the working condition type, parameter change feature library is updated and new working condition data features are included.

5. The integrated intelligent monitoring and control method for methanol-containing bilge water according to claim 3, characterized in that: The proposed LSTM network for predicting methanol concentration in a time series, adapted to the complex dynamic conditions of ships, matches the input layer dimension with the feature dimension of the multi-source parameters. A dropout layer is added to the hidden layer to suppress overfitting. A dual-output layer design is employed to simultaneously output the specific predicted value of methanol concentration and the classification result of concentration change trend, respectively meeting the needs of quantitative control and qualitative decision-making. Multiple types of loss functions are integrated to optimize the training objective, including: The system integrates four types of parameters: ship navigation status, sea state, bilge water quality, and equipment operation status, and statistically analyzes the effective feature dimensions of each parameter. The Pearson correlation coefficient method is used to calculate the correlation between each parameter and methanol concentration changes. Low-correlation redundant parameters are removed, and feature dimensions are retained. The LSTM network architecture is designed around the characteristics of ship operating conditions. The input layer incorporates ship operating condition label encoding features, which are input in conjunction with the time series of multi-source parameters. The hidden layer adopts a multi-level structure and is configured with a dynamic dropout mechanism that adapts to the complexity of the operating conditions. The output layer constructs a dual output mode of quantitative and qualitative analysis. At the same time, by adding an operating condition attention layer, parameters are given higher weights based on historical data, which strengthens the influence of key parameters on the prediction results. Overall, it achieves a deep adaptation between ship operating conditions and methanol concentration time series prediction. For the quantitative output layer, the mean absolute error loss function is selected, and for the qualitative output layer, the cross-entropy loss function is selected. The working condition and loss weight mapping rules are constructed, and the two types of loss weights are dynamically adjusted according to the ship operation scenario. The two types of losses are weighted and summed to obtain the total loss function, and a regularization term is added to suppress overfitting. The weights of the LSTM layer and fully connected layer are initialized using a preset initialization method, and the working condition label embedding layer is initialized using a random distribution. An adaptive optimizer is selected and a learning rate with adaptive decay for working conditions is configured. At the same time, a learning rate restart mechanism is added, and the batch size is dynamically adjusted according to the working condition sample size. The network is trained using a phased training method based on working conditions. During the training process, the loss of the validation set under different working conditions is monitored. If the prediction accuracy does not meet the standard, the network parameters are adjusted and the network is retrained. After the training is completed, the network is verified through full coverage of working conditions to ensure that it meets the requirements for prediction of all working conditions of the ship.

6. The integrated intelligent monitoring and control method for methanol-containing bilge water according to claim 5, characterized in that: The LSTM network architecture is designed around the characteristics of ship operating conditions. It incorporates ship operating condition label encoding features into the input layer, which are then input in conjunction with multi-source parameter time series sequences. The hidden layer adopts a multi-level structure and is configured with a dynamic dropout mechanism adapted to the complexity of the operating conditions. The output layer constructs a dual output mode of quantitative and qualitative analysis. Furthermore, by adding an operating condition attention layer, parameters are given higher weights based on historical data, strengthening the influence of key parameters on the prediction results. Overall, this achieves a deep adaptation between ship operating conditions and methanol concentration time series prediction, including: The feature simplification and feature dimension determination are performed on the multi-source parameter time series of ship navigation status, sea state, bilge water quality and equipment operation status. The number of neurons in the input layer is set to match the feature dimension. One-hot encoding is used to convert the discrete labels of ship operating condition types into high-dimensional sparse vectors. The vectors are then connected to the embedding layer and mapped to low-dimensional dense vectors. The low-dimensional operating condition feature vectors are concatenated with the multi-source parameter time series according to the dimension to form an input tensor, which is then input into the LSTM layer. Set up 2-3 LSTM hidden layers. The number of neurons in the first hidden layer is 4-6 times the feature dimension. The number of neurons in the second / third layer decreases by 10%-20% layer by layer. Connect a dropout layer to the output of each LSTM hidden layer. Based on historical data, classify the ship's operating condition complexity level and configure the corresponding dropout rate for different levels of operating conditions. In real time, identify the current ship operating condition complexity and match the corresponding dropout rate. After the output tensor of the last LSTM hidden layer is connected to the fully connected layer for dimension mapping, the output layer is constructed in two ways. The quantitative output branch is connected to the fully connected layer and uses a linear activation function. The output dimension is matched with the methanol concentration prediction time step and outputs the specific predicted value of methanol concentration. The qualitative output branch is connected to the fully connected layer and uses a Softmax activation function. The output dimension is set to 3-dimensional to classify the results of the output concentration change trend. A working condition attention layer is added between the hidden layer and the dual output layer. A training set is constructed based on the ship's historical working conditions and methanol concentration correlation data. The working condition type and parameter weight matrix that match the output dimension and parameter dimension of the hidden layer are trained through supervised learning. When the network runs, it identifies the current ship working condition type and retrieves the corresponding parameter weight vector and multiplies it element-wise with the output tensor of the hidden layer to enhance the parameter feature signal. The overall network architecture is constructed by connecting the input layer, LSTM hidden layer, dropout layer, working condition attention layer, and dual output layer. Samples of all ship working conditions are input into the network for training. The prediction effect under different working conditions is monitored in real time, and the dropout rate and attention weights of the corresponding working conditions are adjusted accordingly. After training, test samples of typical ship working conditions are selected for full coverage verification.

7. The integrated intelligent monitoring and control method for methanol-containing bilge water according to claim 3, characterized in that: The system collects multi-source parameters in real time, including ship navigation, sea state, bilge water quality, and equipment operation. After preprocessing using a normalization method, these parameters are input into the trained LSTM model, which outputs methanol concentration prediction results. A dynamic accuracy verification rule is constructed: when the prediction accuracy consistently fails to meet the standard, newly added actual ship operating condition data from the operating condition and concentration correlation feature library are automatically retrieved to supplement the training set. The training process is then re-executed, and the system model parameters are updated, including: Deploy a multi-dimensional data acquisition module on the ship end, and collect parameters of ship navigation status, sea state, bilge water quality and equipment operation status at the same sampling frequency as the model training. Use the 3σ principle to remove outliers of the original parameters. Based on the historical statistical extreme values ​​of parameters in the operating condition and concentration correlation feature library, map each parameter to the [0,1] interval through Min-Max normalization to eliminate dimensional differences. The normalized multi-source parameters are concatenated into a time series according to the time window length during model training. This sequence is then integrated into the encoding vector of the current ship operating condition label to form an input tensor. This tensor is then input into the trained LSTM model. The model outputs a quantitative prediction of methanol concentration in the future period and reverses it to restore the actual concentration value. At the same time, it outputs a qualitative classification result of the concentration change trend. After integration, the result is output to the ship's bilge water control system. Construct a hierarchical, two-dimensional accuracy verification rule for working conditions. Divide the ship's working conditions into three levels: high, medium, and low. Configure differentiated quantitative dimension average absolute error thresholds and qualitative dimension trend matching accuracy thresholds for different levels of working conditions. Set continuous failure criteria. If the prediction accuracy exceeds the threshold for a single working condition for a set number of consecutive preset times, or if the prediction accuracy of a preset proportion of the prediction samples within a preset time range in the entire working condition range is not up to standard, it is determined that the prediction accuracy is continuously unsatisfactory. When the accuracy of the judgment is consistently substandard, the system automatically filters newly added actual ship operating condition data within the preset time period from the operating condition and concentration correlation feature library. Priority is given to samples that are consistent with the type of operating condition that does not meet the accuracy standard and contain complete multi-source parameter time series, actual methanol concentration value and operating condition label. The newly added data after screening is added to the training set and validation set according to a preset ratio. When adding data, the proportion of each type of operating condition in the training set is kept balanced. If there are not enough new samples for a certain type of operating condition, virtual samples are generated to supplement the data by slightly perturbing the parameter values. An incremental training strategy is adopted. The LSTM model is retrained based on the original model parameters and the supplemented training set. Hyperparameters are configured to adapt to the working conditions. The accuracy of the validation set is monitored during the training process. When the accuracy of the validation set meets the standard and there is no decrease for a consecutive preset number of rounds, the training is stopped. The retrained model parameters are packaged and automatically replaced with the original parameters in the ship end system. The model with the new parameters is switched to perform real-time prediction and the accuracy is continuously monitored.

8. The integrated intelligent monitoring and control method for methanol-containing bilge water according to claim 3, characterized in that: The method uses maritime emission standards as a benchmark to determine the basic control threshold for methanol concentration, and innovatively constructs a two-dimensional threshold adjustment coefficient mapping rule based on operating condition type and concentration prediction trend. Combining real-time ship operating conditions with the concentration prediction results output by the LSTM model, the control threshold is dynamically calculated and refreshed frequently, including: Extract the methanol concentration compliance limits under different scenarios, construct a two-layer basic control threshold based on the compliance limits, and mark the navigation area boundaries applicable to each basic control threshold. Enter the basic control thresholds and area rules into the ship threshold management library. By quantifying both ship operating conditions and methanol concentration evolution trends in a dual-level manner, a synergistic and adaptable coefficient combination system is constructed. After being verified and solidified by historical data, the threshold adjustment can accurately respond to the dynamic operating conditions and concentration change patterns of ships. Deploy a working condition identification module to collect real-time data on ship navigation status, sea state, and equipment load, and automatically determine the complexity level of the current working condition. Simultaneously receive the quantitative prediction value and qualitative trend result of methanol concentration output by the LSTM model, analyze the trend type and related features, and perform real-time verification on the collected working condition data and prediction results to remove abnormal data and ensure data continuity. Based on the preset formula, the dynamic early warning threshold and dynamic emergency control threshold are calculated by using the basic control threshold, the operating condition adjustment coefficient and the trend adjustment coefficient. A high-frequency refresh frequency is set, and the latest operating conditions and prediction results are read and the threshold is recalculated each time it is refreshed. At the same time, hard constraints on the threshold are set to ensure that the dynamic threshold does not exceed the preset ratio range between the compliance limit and the basic threshold. The calculated dynamic threshold is pushed to the ship's bilge water control system in real time as the basis for control. The application effect of dynamic thresholds is statistically analyzed in real time, including threshold triggering accuracy and false triggering rate. The coefficient combination matrix is ​​periodically optimized based on the statistical results, and the basic control thresholds are updated synchronously in conjunction with the updates to maritime emission standards.

9. The integrated intelligent monitoring and control method for methanol-containing bilge water according to claim 3, characterized in that: The system collects the operating load parameters of the bilge water treatment equipment and calculates the real-time load rate. It then constructs dynamic threshold ranges and equipment treatment intensity association rules. Based on the generated dynamic thresholds, it matches the corresponding treatment intensity and sends control commands to the equipment controller. An equipment load safety constraint mechanism is added, automatically reducing the treatment intensity when the equipment load approaches the rated value. This includes: A data acquisition module is deployed to collect operating parameters of the bilge water treatment equipment, such as motor operating current / power, medium processing flow rate, inlet and outlet pressure, and reagent dosing amount, through the equipment PLC, current transformer, and flow sensor. The collected raw parameters are preprocessed, and instantaneous fluctuation abnormal values ​​are filtered out. Missing parameters are filled in using historical period averages. The real-time load rate is calculated based on the preprocessed parameters using the power method. The dynamic threshold output of methanol concentration is connected to the dynamic threshold range, and each range corresponds to a clear concentration control target. The equipment processing intensity is decomposed into a quantifiable and adjustable dimension, and multiple processing intensity levels are defined. A one-to-one correlation rule between the dynamic threshold range and the processing intensity level is constructed, and a working condition correction factor is added. The rules are verified by simulating the historical data of equipment operation to ensure the concentration control effect and equipment load safety. The verified rules are then solidified into the control system. The system receives the dynamic threshold calculation results of methanol concentration in real time, identifies the current threshold range, matches the initial treatment intensity level according to the association rules, fine-tunes the initial treatment intensity based on the real-time load rate, converts the matched treatment intensity level into control instructions that are compatible with the communication protocol of the device controller, and clarifies the specific control requirements of each control dimension. Based on the rated load rate of the equipment, two levels of safety constraint thresholds and corresponding processing intensity reduction gradients are set. The real-time load rate of the equipment is monitored in real time. According to the threshold range of the load rate, corresponding constraint operations are triggered, including reducing the processing intensity, pushing early warning prompts, pausing non-core control dimensions, and locking the processing intensity increase operation. A callback mechanism is designed so that when the load rate falls back to the preset range, the processing intensity is gradually restored to the target level. Control commands are sent to the equipment controller via an industrial communication bus. Real-time feedback from the equipment is collected to confirm that the commands have been executed. Changes in methanol concentration in the bilge water and equipment load rate are monitored simultaneously. The treatment intensity is dynamically adjusted based on the monitoring results. The accuracy of command execution, number of load overruns, and concentration control compliance rate are statistically analyzed regularly. Based on the statistical data, the association rules and safety constraints are iteratively optimized to lower the gradient.

10. An integrated intelligent monitoring and control system for methanol-containing bilge water, characterized in that: include: processor; A machine-readable storage medium for storing machine-executable instructions of the processor; The processor is configured to execute the integrated processing method for intelligent monitoring and linkage control of methanol-containing bilge water as described in any one of claims 1 to 9 by executing the machine-executable instructions.

Images