Cutter suction dredger steel pile trolley control method and system based on hydraulic excitation early warning

Abstract

The invention relates to the technical field of cutter suction dredger steel pile trolley control, and discloses a cutter suction dredger steel pile trolley control method and system based on hydraulic excitation early warning, and the method comprises the steps: collecting the influence degree of multiple parameters on the system stability, calculating a pressure fluctuation factor and the like to form a state feature set; constructing a dynamic pressure balance model, inputting data such as a pressure time sequence queue and a sea condition level, and outputting a dynamic pressure adjustment value; constructing a cooperative control correction model, inputting steel pile attitude data and outputting a cooperative correction coefficient; and calculating a final control output quantity by combining the adjustment value and the correction coefficient, and generating a hydraulic excitation control instruction set. The system comprises a multi-source sensor module, a state feature calculation module, a dynamic pressure balancing module and the like, and multi-dimensional data fusion and intelligent control are achieved. The stability, the control precision and the environmental adaptability of the steel pile trolley system are improved, and the method is suitable for the control scene of the cutter suction dredger steel pile trolley.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of steel pile trolley control of a cutter suction dredger, in particular to a steel pile trolley control method and system based on hydraulic excitation early warning for a cutter suction dredger. BACKGROUND

[0002] As the core equipment in the fields of water conservancy projects and channel dredging, the stability and control accuracy of the steel pile trolley system of a cutter suction dredger directly affect the construction efficiency and safety. The steel pile trolley system mainly realizes key operations such as positioning, plugging and moving of the steel pile through the coordinated operation of the hydraulic execution unit, the steel pile clamp and the walking mechanism. However, in actual operation, the system faces complex working condition challenges, resulting in significant limitations of existing control methods.

[0003] From the perspective of the characteristics of the hydraulic system, the traditional control method is insufficient in dynamic response to hydraulic pressure fluctuations. The cylinder pressure of the hydraulic execution unit is easily affected by factors such as load mutation and hydraulic oil temperature change in different operation stages (such as steel pile plugging and trolley walking), resulting in pressure over-limit or instability. For example, when the hydraulic oil temperature rises, the decrease in oil viscosity may cause thermal attenuation effect, causing the actual pressure to deviate from the preset threshold, thereby affecting the stability of the clamping force of the steel pile clamp and the displacement accuracy of the walking mechanism. At the same time, the problem of electromagnetic valve response delay cannot be ignored, and the traditional control model does not fully consider the influence of electromagnetic valve opening and closing delay time on system synchronization, which may cause action lag or impact when multiple execution mechanisms work together, reducing the operation efficiency.

[0004] In terms of mechanical motion control, the real-time monitoring and correction ability of the steel pile posture is another shortcoming of the traditional method. The steel pile is prone to tilt or deviation in complex sea conditions (such as wave and current action), and the existing control model cannot effectively integrate real-time posture data such as steel pile tilt angle and clamp clamping force distribution, so it cannot adjust the output of each control node in time to compensate for the posture deviation. In addition, the displacement deviation accumulation problem of the walking mechanism has not been effectively solved, and long-term operation may cause displacement accuracy to decrease due to mechanical resonance or environmental vibration, affecting the positioning accuracy of the dredger.

[0005] The complexity of environmental interference factors further increases the control difficulty. Changes in sea state level (such as wave height and period) will produce different degrees of vibration excitation to the trolley system, and the traditional control method lacks analysis of the spectral characteristics of environmental vibration, cannot identify the influence of frequency spectrum main peak frequency shift on system stability, and is difficult to maintain normal operation of the system in strong vibration working conditions. At the same time, the motion coupling relationship between multiple execution mechanisms (such as the linkage of hydraulic execution unit and walking mechanism) has not been fully modeled, resulting in the inability of the coordinated control strategy to accurately coordinate the actions of each mechanism, which may easily cause mechanical stress concentration or abnormal wear.

[0006] In the prior art, although some control systems introduce sensor monitoring and closed-loop control, they are mostly single-parameter feedback control, lacking the ability of fusion analysis and dynamic modeling of multi-source data (pressure, displacement, temperature, vibration, etc.). For example, only the pressure sensor feedback is used to adjust the hydraulic output, while the comprehensive influence of electromagnetic valve response delay, hydraulic oil temperature and environmental vibration is ignored, resulting in insufficient robustness of the system under complex working conditions. In addition, the traditional model is mostly based on fixed control parameters, which cannot adapt to the dynamic changes of different equipment models (such as oil cylinders with different rated pressures, clamps with different clamping surface structures) and sea conditions, limiting the universality and adaptive ability of the system. SUMMARY

[0007] The purpose of the present application is to provide a hydraulic excitation early warning based control method and system for the steel pile trolley of a cutter suction dredger, to solve the problems raised in the background art.

[0008] To achieve the above-mentioned purpose, the present application provides the following technical solution: a hydraulic excitation early warning based control method for the steel pile trolley of a cutter suction dredger, the method comprising: For any one control node of the hydraulic execution unit, steel pile clamp and walking mechanism in the trolley system, the influence degree of the oil cylinder pressure value, electromagnetic valve response time, displacement sensor data, hydraulic oil temperature and environmental vibration parameters on the system stability is collected respectively, and the pressure fluctuation factor, response delay factor, displacement deviation factor, thermal decay factor and mechanical resonance factor are calculated based on the obtained influence degrees to form a state feature set; A dynamic pressure equalization model is constructed, for any one control node, the actual pressure data of the previous M time units of the node in any control period is arranged in the order of operation to form a pressure time sequence queue, and the preset pressure threshold, target displacement, hydraulic oil temperature limit and equipment code information of the next operation stage are obtained, and the current sea condition level identifier is also obtained; the pressure time sequence queue, preset pressure threshold, target displacement, hydraulic oil temperature limit, equipment code information, sea condition level identifier and state feature set of the current control period are input into the dynamic pressure equalization model, and the dynamic pressure adjustment value of the node in the next operation stage is output; A cooperative control correction model is constructed, for any one control node, real-time steel pile attitude data is obtained in any control period; the inclination angle of the steel pile, clamp clamping force distribution and trolley displacement deviation of the current control period are input into the cooperative control correction model, and the cooperative correction coefficient of the node in the next operation stage is output; For any one control node, based on the dynamic pressure adjustment value and the cooperative correction coefficient obtained in the current control period, the final control output of the node in the next operation stage is calculated; The final control output of all control nodes is integrated to generate a set of hydraulic excitation control instructions for the trolley system.

[0009] Preferably, the influence degree of the collected cylinder pressure value, solenoid valve response time, displacement sensor data, hydraulic oil temperature and environmental vibration parameters on system stability specifically includes: For any control node, the actual pressure data of each operation stage in the historical operation cycle is obtained, and the solenoid valve opening and closing delay time, displacement sensor sampling value, hydraulic oil temperature peak value and environmental vibration spectrum data of each operation stage are collected; wherein the sea state level is divided into six categories according to wave height and period, and each category corresponds to a unique code; the equipment code information includes cylinder model identification and clamp type, and the cylinder model identification is divided into three levels according to the rated pressure, and the clamp type is divided into plane and curved surface according to the clamping surface structure; The state sequence information of each operation stage includes the pressure fluctuation amplitude of n continuous sampling points in the operation stage; the n sampling points of each operation stage are coded in time sequence as 1 to n, and the sampling point code with the highest frequency of pressure overrun in the historical cycle is counted; The influence degree of the cylinder pressure value, solenoid valve response time, displacement deviation, hydraulic oil temperature and environmental vibration on the stability of the current control node is calculated respectively, wherein the influence degree of the first four items is obtained by correlation analysis of historical data and system failure, and the influence degree of environmental vibration is obtained by calculating the offset of the main peak frequency of the spectrum.

[0010] Preferably, the calculation method of the state feature set is: based on the influence degree of the cylinder pressure value, solenoid valve response time, displacement deviation, hydraulic oil temperature and environmental vibration, the pressure fluctuation factor to the mechanical resonance factor is obtained by normalized weighted summation.

[0011] Preferably, the dynamic pressure balance model is constructed based on a multi-modal perception network, including a data fusion layer, a time sequence coding layer, a feature extraction stacking layer, a cross-modal attention layer, a decision generation layer and an output calibration layer; The data fusion layer is used to convert the pressure time sequence queue, the preset pressure threshold, the target displacement amount, the hydraulic oil temperature limit value, the equipment code information, the sea state level identification and the state feature set into a multi-dimensional feature tensor; The time sequence coding layer is used to generate operation sequence coding for each operation stage in the pressure time sequence queue to maintain the continuity of the operation process; The feature extraction stacking layer contains L parallel feature extraction modules, each module processing hydraulic parameter features, mechanical motion features and environmental interference features respectively; The cross-modal attention layer is used to perform cross-attention calculation on the multi-dimensional features of the state feature set and the feature extraction output to generate a dynamic weight distribution matrix; The decision generation layer is used to fuse the dynamic weight distribution matrix and the feature extraction result to generate a preliminary control strategy vector; The output calibration layer is used to map the preliminary control strategy vector to the dynamic pressure adjustment value.

[0012] Preferably, the training process of the dynamic pressure balancing model comprises: Collect historical control data sets of multiple operation cycles, each data sample containing pressure data of consecutive M+1 operation stages, influence factor data of the M+1 stage and a state feature set; divide the data set into a training set and a test set; Take the data of the first M stages and the influence factors as inputs, and take the pressure adjustment amount of the M+1 stage as the target output, and optimize the network parameters through error back propagation; Use the test set to evaluate the model control error, and if the error does not meet the convergence standard, adjust the number of feature extraction modules or the dimension of the attention head and retrain.

[0013] Preferably, the collaborative control correction model is constructed based on a graph neural network, including a topological relationship modeling branch and a collaborative optimization branch; The topological relationship modeling branch establishes the motion coupling relationship of each actuator of the trolley system through graph structure learning; The collaborative optimization branch fuses the motion coupling relationship and the steel pile posture data, and outputs a collaborative correction coefficient after multi-layer perceptron calculation.

[0014] Preferably, the calculation method of the final control output is to weight and fuse the dynamic pressure adjustment value and the collaborative correction coefficient according to a preset proportion, and combine the nonlinear compensation of the current oil cylinder load characteristics.

[0015] Preferably, the generation method of the hydraulic excitation control instruction set is to perform time domain synchronization calibration on the final control output of all control nodes, and generate an instruction sequence with timing constraints in combination with the device response characteristics.

[0016] Preferably, the model training process further comprises an adaptive optimization step: According to the distribution characteristics of the test set error, automatically adjust the connection mode of the feature extraction module or the activation function of the attention head; if the error presents periodic fluctuation, increase the residual connection structure and reduce the learning rate; if the error continuously diverges, reinitialize the parameters after reconstructing the network topology.

[0017] Preferably, the application further comprises a hydraulic excitation pre-warning based cutter suction dredger steel pile trolley control system for realizing the hydraulic excitation pre-warning based cutter suction dredger steel pile trolley control method as described above, the system comprising: A multi-source sensor module is arranged in the hydraulic execution unit, the steel pile clamp and the walking mechanism of the trolley system, and is used for collecting the cylinder pressure value, the electromagnetic valve response time, the displacement sensor data, the hydraulic oil temperature and the environmental vibration parameters of each control node in real time; A state feature calculation module is connected to the multi-source sensor module, and is used for calculating the pressure fluctuation factor, the response delay factor, the displacement deviation factor, the thermal attenuation factor and the mechanical resonance factor according to the collected data, and generating a state feature set; A time sequence data storage module is provided with a circular buffer unit, and is used for storing the actual pressure data of the previous M time units of each control node in the operation sequence, and forming a pressure time sequence queue; A sea condition recognition module integrates a ship attitude sensor and a wave monitoring unit, and is used for analyzing the current sea condition grade mark in real time; A dynamic pressure equalization module is constructed based on a multi-modal neural network, and the input end is connected to the state feature calculation module, the time sequence data storage module and the sea condition recognition module, and is used for receiving the pressure time sequence queue, the preset pressure threshold value, the target displacement amount, the hydraulic oil temperature limit value, the equipment code information and the sea condition grade mark, and generating a dynamic pressure adjustment value through a cross-modal attention mechanism calculation; A cooperative control correction module is provided with a steel pile attitude monitoring unit and a clamp stress analysis unit, and is used for obtaining the real-time steel pile inclination angle, the clamp clamping force distribution and the trolley displacement deviation, modeling the motion coupling relationship through a graph neural network, and outputting a cooperative correction coefficient; An instruction generation module is connected to the input end of the dynamic pressure equalization module and the cooperative control correction module, and is used for performing nonlinear fusion calculation on the dynamic pressure adjustment value and the cooperative correction coefficient, generating the final control output of each control node, and synchronously generating a hydraulic excitation control instruction set with time sequence constraints; A safety warning module is internally provided with a historical fault database and an abnormal pattern recognition unit, and is used for monitoring the abnormal fluctuation of the state feature set in real time, triggering a hierarchical alarm signal and feeding back to the instruction generation module for control amount compensation.

[0018] Compared with the prior art, the present application has the following advantages: The application collects multi-dimensional data such as oil cylinder pressure value, electromagnetic valve response time, displacement sensor data, hydraulic oil temperature and environmental vibration parameters in real time through multi-source sensors, and calculates the influence degree of each parameter on system stability based on historical data correlation analysis and spectrum analysis, to generate a state feature set containing pressure fluctuation factors, response delay factors and the like. This multi-parameter fusion analysis mechanism breaks through the limitations of traditional single parameter control, can comprehensively capture the comprehensive influence of hydraulic system dynamic characteristics, mechanical motion state and environmental interference, and provides rich feature input for precise control. For example, by analyzing the frequency offset of the main peak of the environmental vibration spectrum, the mechanical resonance risk can be identified in advance, and the displacement deviation accumulation or equipment damage caused by vibration can be avoided.

[0019] The dynamic pressure balancing model is constructed based on a multi-modal perception network, converts multi-modal data such as pressure time sequence, preset pressure threshold, sea state grade identifier and the like into a multi-dimensional feature tensor through a multi-layer structure such as a data fusion layer and a time sequence coding layer, and generates a dynamic weight distribution matrix using a cross-modal attention mechanism to realize dynamic weight adjustment of hydraulic parameters, mechanical motion characteristics and environmental interference. The model can predict the dynamic pressure adjustment value of the next stage according to the historical pressure time sequence data, and simultaneously combine the sea state grade and the equipment model (such as the rated pressure grade of the oil cylinder, the clamp type) for adaptive adjustment, effectively compensating for problems such as pressure fluctuation, thermal decay and electromagnetic valve response delay of the hydraulic system. For example, in high sea states, the model can automatically increase the pressure adjustment value to cope with the load fluctuation caused by wave impact, ensuring the stability of the steel pile clamp clamping force.

[0020] The cooperative control correction model is constructed based on a graph neural network, establishes the motion coupling relationship of each actuator of the trolley system through a topological relationship modeling branch, and combines real-time steel pile inclination angle, clamp clamping force distribution and other attitude data to output a cooperative correction coefficient through a cooperative optimization branch. This graph structure-based modeling method can accurately describe the linkage relationship between the hydraulic execution unit, the steel pile clamp and the walking mechanism, and real-time correct the control error caused by the inclination of the steel pile or the motion coupling of the mechanism. For example, when the inclination of the steel pile is detected, the model can quickly calculate the cooperative correction coefficient of each control node, adjust the oil cylinder pressure and the walking mechanism displacement, realize the dynamic calibration of the steel pile attitude, and avoid the positioning deviation or operation efficiency caused by inclination.

[0021] At the control output level, the dynamic pressure adjustment value and the collaborative correction coefficient are weighted and fused according to a preset ratio, and nonlinear compensation is performed in conjunction with the cylinder load characteristics to generate the final control output of each control node. This composite control strategy considers both the dynamic pressure balance of the hydraulic system and the collaborative correction of multiple mechanisms, effectively improving control accuracy and response speed. Simultaneously, the output of all control nodes is synchronously calibrated in the time domain to generate a hydraulic excitation control command set with time constraints, ensuring the synchronization and coordination of the actions of multiple actuators and reducing mechanical wear and energy loss caused by action lag or impact.

[0022] The safety early warning module incorporates a historical fault database and an anomaly pattern recognition unit. By monitoring abnormal fluctuations in the set of state characteristics in real time (such as excessive pressure fluctuation factors or sudden changes in mechanical resonance factors), it triggers tiered alarm signals and feeds them back to the command generation module for control quantity compensation. This early warning mechanism can identify potential system faults in advance (such as excessively high hydraulic oil temperature or abnormal solenoid valve response delay), and prevent major faults by dynamically adjusting control parameters or triggering protective measures, thus significantly improving the safety and reliability of the system.

[0023] The adaptive optimization mechanism during model training (such as automatically adjusting the connection method of the feature extraction module based on the error distribution of the test set and adding residual connection structures) enables the dynamic pressure equilibrium model to automatically optimize the network structure and parameters according to different working conditions, thereby improving the model's generalization ability and convergence speed. By combining sea state classification (six categories) and equipment coding information (cylinder model, fixture type), the system can automatically switch control strategies for different operating environments and equipment, significantly enhancing the system's versatility and adaptability to different working conditions. Attached Figure Description

[0024] Fig. 1 This is a schematic diagram illustrating the working principle of the cutter suction dredger steel pile trolley control method based on hydraulic excitation early warning as described in this invention. Fig. 2 This is a schematic diagram illustrating the principle of impact level acquisition and state feature set calculation. Fig. 3 This is a schematic diagram illustrating the working principle of the dynamic pressure equilibrium model. Detailed Implementation

[0025] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0026] Please see Figs. 1-3The present invention relates to a control method for a steel pile trolley of a cutter suction dredger based on hydraulic excitation early warning, the specific implementation steps of which are as follows: For any control node in the trolley system, including the hydraulic actuator, steel pile clamp, and traveling mechanism, the impact of cylinder pressure, solenoid valve response time, displacement sensor data, hydraulic oil temperature, and environmental vibration parameters on system stability is collected. In practice, a multi-source sensor module acquires raw data from each control node in real time. Cylinder pressure is collected by a pressure sensor; solenoid valve response time is calculated using the time difference between the solenoid valve's opening / closing signal and the system clock; displacement sensor data is directly read from the output value of a linear displacement sensor; hydraulic oil temperature is monitored in real time by a temperature sensor embedded in the hydraulic pipeline; and environmental vibration parameters are collected by a vibration acceleration sensor and subjected to spectrum analysis.

[0027] Based on the acquired levels of influence, pressure fluctuation factor, response delay factor, displacement deviation factor, thermal attenuation factor, and mechanical resonance factor are obtained through normalized weighted summation. These factors together constitute a set of state characteristics. The normalization process employs the Z-score standardization method, and the weighting coefficients are determined through correlation analysis between historical data and system faults. For example, the weight of the cylinder pressure value can be set to 40% based on its frequency of causing system faults, and the weights of other parameters are calculated accordingly.

[0028] A dynamic pressure equalization model based on a multimodal sensing network is constructed. For any control node, within any control cycle, the actual pressure data of the previous M time units of that node are arranged in the order of operation to form a pressure time sequence queue. This queue is stored and updated through the cyclic buffer unit of the time sequence data storage module. At the same time, the preset pressure threshold (set by system process parameters), target displacement (from trolley travel control command), hydraulic oil temperature limit (determined according to hydraulic oil type and working environment), and equipment coding information (including cylinder model identifier and clamp type, where cylinder model identifier is divided into three levels according to rated pressure, and clamp type is divided into planar and curved surfaces according to clamping surface structure) are obtained through the sea state identification module. The current sea state level identifier is obtained (divided into six categories according to wave height and period, each category corresponds to a unique code, such as sea state level 1 corresponding to code 01).

[0029] The pressure time sequence queue, the preset pressure threshold, the target displacement amount, the hydraulic oil temperature limit value, the equipment code information, the sea state grade identifier and the state feature set are input into a dynamic pressure balance model. Inside the model, multi-source data is converted into a multi-dimensional feature tensor through a data fusion layer, a time sequence coding layer generates operation sequence coding for the pressure time sequence queue, a feature extraction stack layer processes hydraulic parameters, mechanical movement and environmental interference features through L parallel modules respectively, a cross-modal attention layer calculates a dynamic weight distribution matrix, a decision generation layer generates a preliminary control strategy vector after fusion, and finally an output calibration layer maps it to the dynamic pressure adjustment value of the node in the next operation phase.

[0030] A collaborative control correction model based on a graph neural network is constructed. For any control node, in any control cycle, the real-time steel pile inclination angle is obtained through the steel pile posture monitoring unit, the clamp clamping force distribution data is obtained through the clamp stress analysis unit, and the trolley displacement deviation (the difference between the actual displacement and the target displacement) is read. The steel pile inclination angle, the clamp clamping force distribution and the trolley displacement deviation are input into the collaborative control correction model. The model establishes the motion coupling relationship (such as the correlation matrix of the steel pile inclination angle and the clamp clamping force) of each actuator of the trolley system through the topological relationship modeling branch, and the motion coupling relationship and the real-time posture data are fused through the collaborative optimization branch. After calculation by a multilayer perceptron, the collaborative correction coefficient of the node in the next operation phase is output.

[0031] For any control node, based on the dynamic pressure adjustment value and the collaborative correction coefficient obtained in the current control cycle, the weighted fusion is performed according to the preset proportion (for example, the dynamic pressure adjustment value weight accounts for 60%, and the collaborative correction coefficient weight accounts for 40%), and the nonlinear compensation is performed in combination with the load characteristics of the current oil cylinder (query through the oil cylinder load-pressure characteristic curve), and finally the final control output of the node in the next operation phase is obtained. The nonlinear compensation function adopts a piecewise linear interpolation method, and the compensation coefficient is automatically selected according to the load size.

[0032] The final control outputs of all control nodes are integrated. First, the outputs are time-domain synchronized (synchronize the node calculation period through the system clock), and then the instruction sequence with time sequence constraints is generated in combination with the device response characteristics (such as electromagnetic valve opening and closing delay time and oil cylinder action lag time), forming the trolley system hydraulic excitation control instruction set. The instruction sequence adopts an event-triggered mechanism, and the completion signal of the previous operation phase is used as the trigger condition of the next stage instruction.

[0033] The application will be further described below in combination with Examples 1 to 5: Example 1: The embodiment specifically limits the collection and calculation of the influence degree in the state feature set construction process. For any control node of the hydraulic execution unit, steel pile clamp and walking mechanism in the trolley system, the actual pressure data of each operation stage of the node in the historical operation cycle needs to be obtained, and the electromagnetic valve opening and closing delay time, displacement sensor sampling value, hydraulic oil temperature peak value and environmental vibration spectrum data of the corresponding stage are also collected. Among them, the length of the historical operation cycle is determined according to the equipment working condition, and is usually 10 to 50 complete cycles of continuous operation, each cycle contains clamp clamping, trolley positioning, steel pile lowering, excavation operation, trolley displacement and other operation stages, and each stage is identified by photoelectric switch or travel sensor.

[0034] The collection of electromagnetic valve opening and closing delay time is realized by connecting a high-precision time stamp module in the electromagnetic valve control loop. The module synchronously records the time of the driving signal sent by the controller and the time of the in-place signal feedback by the electromagnetic valve. The time difference between the two is the delay time, and the resolution can reach microseconds. The displacement sensor sampling value adopts a high-frequency sampling mode, with a sampling frequency of 20Hz, i.e. data is obtained every 50ms, which is stored after being filtered by a low-pass filter to remove high-frequency noise. The collection of hydraulic oil temperature peak value uses a temperature sensor with peak holding function. The sensor probe is embedded 10cm away from the valve body at the outlet pipeline of the hydraulic pump, which monitors the oil temperature change in real time and records the maximum value of each operation stage. The environmental vibration spectrum data is collected by a three-axis vibration acceleration sensor, which is fixed to the non-stressed plane of the steel pile base, with a collection frequency of 10kHz. The time domain signal is converted to frequency domain spectrum by fast Fourier transform (FFT), and the frequency spectrum distribution in the range of 0.1Hz to 200Hz is obtained.

[0035] The division of sea state level identification follows the International Maritime Organization (IMO) standard, which is divided into six categories according to wave height and period and assigned a unique code. Among them, 1st sea state corresponds to wave height ≤ 0.5m and period ≤ 4s, coded as "01"; 2nd sea state corresponds to 0.5m < wave height ≤ 1.25m and 4s < period ≤ 6s, coded as "02"; 3rd sea state corresponds to 1.25m < wave height ≤ 2.5m and 6s < period ≤ 8s, coded as "03"; 4th sea state corresponds to 2.5m < wave height ≤ 4.0m and 8s < period ≤ 10s, coded as "04"; 5th sea state corresponds to 4.0m < wave height ≤ 6.0m and 10s < period ≤ 12s, coded as "05"; 6th sea state corresponds to wave height > 6.0m or period > 12s, coded as "06". The sea state level is calculated in real time by the wave radar and attitude sensor equipped on the ship, and updated every 10 minutes.

[0036] The equipment code information includes the oil cylinder model identification and the clamp type. The oil cylinder model identification is divided into three levels according to the rated pressure: P1 level, the rated pressure ≤ 16 MPa, P2 level, the rated pressure is 16 MPa~25 MPa (including 25 MPa), P3 level, the rated pressure > 25 MPa; the clamp type is divided into plane (code F1) and curved surface (code F2) according to the clamping surface structure, wherein the plane clamp is suitable for the steel pile with smooth surface, and the curved surface clamp is suitable for the special-shaped steel pile with arc. The equipment code is input by manual input or RFID tag reading mode during system initialization, and is stored in the controller register, which can be called in real time with the control data.

[0037] The state sequence information of each operation stage includes the pressure fluctuation amplitude of the node at n continuous sampling points in the stage, and the value of n is determined according to the operation stage length, usually not less than 100 sampling points, to ensure the reliability of the statistical result. For example, if the operation stage length is 5 seconds, 100 sampling points can be obtained at a sampling frequency of 20 Hz, and each point is encoded in time sequence as 1 to n. By traversing the historical cycle data, the frequency of pressure overrun corresponding to each sampling point code is counted, and the overrun judgment standard is that the actual pressure exceeds the preset pressure threshold ± 10%. The frequency statistics adopts a sliding window algorithm, and the window size is 5 historical cycles. The proportion of the number of overruns of each sampling point in the window is calculated, and the sampling point code with the highest proportion is taken as the sensitive time characteristic parameter of the operation stage. For example, the 50th sampling point in a stage accumulates 15 overruns in 5 cycles, the total sampling number is 500, and the proportion is 3%. If the proportion is the highest in the stage, the code is recorded as the characteristic parameter.

[0038] The influence degree of the oil cylinder pressure value, the electromagnetic valve response time, the displacement deviation, the hydraulic oil temperature and the environmental vibration on the system stability is calculated as follows: The influence degree of the oil cylinder pressure value: through the correlation analysis of the historical data and the system failure, the odds ratio (OR value) is calculated by using the Logistic regression model. First, the historical data is divided into failure samples (marked as 1) and normal samples (marked as 0), the oil cylinder pressure value is taken as the independent variable, and the failure state is taken as the dependent variable, to construct a binary classification model. After the model training is completed, the OR value represents the change multiple of the failure probability when the oil cylinder pressure value changes by one unit, and the value is the influence degree coefficient. For example, if the OR value is 2.5, it means that the failure probability increases by 2.5 times when the pressure value increases by 1 MPa.

[0039] The influence degree of the electromagnetic valve response time: the same Logistic regression model is adopted, the independent variable is the response time, and the dependent variable is the failure state. The response time is quantified in milliseconds, and the OR value output by the model reflects the influence degree of the response time extension on the failure probability.

[0040] Impact of Displacement Deviation: Displacement deviation is defined as the absolute value of the difference between the actual displacement and the target displacement, in millimeters. The degree of impact is determined by calculating the Spearman rank correlation coefficient between this deviation value and the number of fault occurrences. The correlation coefficient ranges from [-1, 1], with a larger absolute value indicating a stronger correlation; here, the absolute value is used as the degree of impact coefficient.

[0041] Impact of Hydraulic Oil Temperature: The correlation analysis between peak hydraulic oil temperature and system failure uses the Cox proportional hazards model in survival analysis, with oil temperature as a covariate, and the hazard ratio (HR value) is calculated as the impact coefficient. The HR value represents the multiple by which the risk of system failure changes for every 1°C increase in oil temperature.

[0042] The degree of environmental vibration impact is determined by calculating the offset between the main peak frequency of the vibration spectrum and the natural frequency of the equipment. First, the natural frequency fnatural (in Hz) of the steel pile trolley system is obtained through hammer impact tests or finite element analysis. Then, the main peak frequency fmain is extracted from the vibration spectrum. The offset Δf = |fmain - fnatural|, and the impact coefficient = Δf / fnatural × 100%. When fmain coincides with fnatural, Δf = 0, and the impact coefficient is 0, indicating no resonance risk. When Δf exceeds fnatural × 20%, it is considered a high-risk state, with an impact coefficient ≥ 20%.

[0043] During data processing, all raw data undergo outlier removal and missing value imputation. Outlier detection uses the IQR (interquartile range) method to remove data points exceeding the range of Q1-1.5IQR or Q3+1.5IQR. Missing values ​​are imputed using linear interpolation or the mean of data from adjacent periods. After calculating the influence coefficient, normalization is performed to map the influence of each parameter to the [0,1] interval, facilitating the subsequent generation of the state feature set. The normalization formula is: Normalized value = (Parameter influence coefficient - Minimum value) / (Maximum value - Minimum value), where the minimum and maximum values ​​are taken from the statistical extreme values ​​in historical data.

[0044] Example 2: This embodiment specifies the calculation method for the state feature set and the architecture of the dynamic pressure equilibrium model. The state feature set consists of a pressure fluctuation factor, a response delay factor, a displacement deviation factor, a thermal attenuation factor, and a mechanical resonance factor. Each factor is obtained by normalizing and weighting the sum of the influence coefficients of the corresponding parameters. Specifically, the influence coefficients of cylinder pressure, solenoid valve response time, displacement deviation, hydraulic oil temperature, and environmental vibration are first obtained. Each coefficient is obtained through correlation analysis of historical data and system faults or spectrum analysis (as described in Example 1). Then, the coefficients are normalized using the Z-score normalization method, as shown in the formula:

[0045] wherein, is the mean of the corresponding coefficient in historical data, is the standard deviation. The normalized data is mapped to the standard normal distribution, which is convenient for subsequent weighted calculation. In the weighting process, the weight of each parameter is determined by the analytic hierarchy process (AHP), for example, the weight of the cylinder pressure value is set to 40% because it directly affects the stability of the hydraulic system; the weight of the solenoid valve response time is 20% because it affects the control accuracy; the weights of the displacement deviation, hydraulic oil temperature and environmental vibration are 15%, 15% and 10%, respectively. Finally, the factors are generated by weighted summation, and the formula is:

[0046] (Note: the calculation of each factor uses the sum of the product of the normalized value of the corresponding parameter and the weight. Here, the pressure fluctuation factor is taken as an example, and the rest are similar.) The dynamic pressure balancing model is based on a multi-modal perception network and includes a data fusion layer, a time sequence encoding layer, a feature extraction stack layer, a cross-modal attention layer, a decision generation layer, and an output calibration layer. The structure and function of each layer are as follows: ① Data fusion layer, this layer is responsible for converting multi-source heterogeneous input data into a unified dimension multi-dimensional feature tensor. The input data includes: Pressure time sequence queue: composed of actual pressure data of the previous M time units arranged in operation order, dimension , where M is the length of the historical window, usually 10 to 50, and is set according to the system response speed requirement.

[0047] Pre-set pressure threshold: the target pressure value of the next operation stage, determined by the system process parameters, dimension .

[0048] Target displacement: the expected displacement value of the trolley walking mechanism, from the upper control instruction, dimension .

[0049] Hydraulic oil temperature limit: the upper limit of the safety temperature set according to the type of hydraulic oil and the environmental temperature, dimension .

[0050] Device encoding information: including cylinder model identification (P1 / P2 / P3, one-hot encoding as a 3-dimensional vector) and clamp type (F1 / F2, one-hot encoding as a 2-dimensional vector), total dimension .

[0051] Sea state level identification: six levels (coded 01-06) divided according to wave height and period, one-hot encoding as vector.

[0052] State feature set: 5-dimensional vector containing five factors, dimension .

[0053] The data fusion layer first splices the above data into a raw feature vector with dimension , and then performs linear transformation through a fully connected neural network, with the formula:

[0054] where is the weight matrix, is the bias vector, optimized through model training. The ReLU activation function is used to introduce nonlinearity, with an output dimension of (D is the feature dimension set by the model, usually 128 or 256).

[0055] ② Time encoding layer, which generates operation sequence encoding for each operation stage in the pressure time queue to preserve the time dependency of the job flow. The encoding method uses sine-cosine position encoding, with the formula:

[0056] where is the time unit number (1 to M), is the feature dimension number. Position encoding generates a unique phase feature for each time unit through different frequencies of sine and cosine functions, avoiding the long-term dependency problem of recurrent neural network (RNN) models. The generated encoding is added element-wise to the feature vector of the pressure time queue to form an input representation containing time sequence information.

[0057] ③ Feature extraction stacking layer, which contains L parallel feature extraction modules (usually L=3) that process hydraulic parameter features, mechanical motion features, and environmental interference features respectively: Hydraulic parameter module: input is the pressure time queue, preset pressure threshold, hydraulic oil temperature limit, and pressure fluctuation factor and thermal decay factor in the state feature set, using 1D convolutional neural network (1D-CNN) for feature extraction. The convolution kernel size is 3, the step is 1, the padding is 1, the number of layers is 2, and each layer is followed by BatchNormalization and ReLU activation function, which is used to capture the time domain features of pressure fluctuation and the influence of oil temperature on the system.

[0058] Mechanical motion module: input is the target displacement, displacement sensor data, and displacement deviation factor in the state feature set, using a gated recurrent unit (GRU) network to process sequence data. The GRU unit contains a reset gate and an update gate, which can effectively capture the dynamic change trend of displacement deviation, with an output dimension consistent with the hydraulic parameter module.

[0059] Environmental interference module: the input is sea state level identification, environmental vibration spectrum characteristics (main peak frequency offset extracted by preprocessing) and mechanical resonance factor, which is processed by a fully connected neural network. This module combines discrete sea state level coding with continuous vibration characteristics, and outputs the comprehensive feature vector of environmental interference.

[0060] The feature vectors output by each module are spliced in the channel dimension to form a composite feature tensor containing multi-modal information, with a dimension of .

[0061] ④ Cross-modal attention layer, which uses a multi-head self-attention mechanism (Multi-Head Attention) to realize the interaction between the state feature set and the multi-modal composite feature. The specific steps are as follows: The composite feature tensor is linearly mapped to the query (Query), key (Key), and value (Value) matrix, with the formula:

[0062] where is a trainable weight matrix.

[0063] The dot product similarity of each query and all keys is calculated, divided by ( is the dimension of the key), and then the attention weight is generated by the Softmax function:

[0064] The multi-head mechanism divides the query, key, and value into h heads (usually h = 2), and parallelly calculates the attention and splices it, with the formula:

[0065] where is the output projection matrix.

[0066] Through cross-modal attention calculation, the model can dynamically allocate the weights of different modal features to generate a dynamic weight allocation matrix containing interaction information.

[0067] ⑤ Decision generation layer, which element-wise multiplies the dynamic weight allocation matrix output by the cross-modal attention with the feature extraction result to realize feature weighted fusion. The fused feature vector is reduced in dimension by a fully connected neural network to generate a preliminary control policy vector, with the formula:

[0068] where and are the parameters of the fully connected layer, and the activation function Tanh limits the output value to the [-1, 1] interval, representing the direction (positive or negative) and relative amplitude of pressure adjustment.

[0069] ⑥Output calibration layer, which maps the preliminary control policy vector to the dynamic pressure adjustment value (unit: MPa) of the actual physical quantity. The mapping process adopts linear transformation, and the formula is:

[0070] wherein, is a proportional factor, is an offset, determined by model training. For example, if the policy vector output is 0.5, , then the dynamic pressure adjustment value is .

[0071] In the model inference stage, the input data is processed according to the following process: first, the multi-source data is converted into a unified feature space representation by the data fusion layer, the time sequence encoding layer injects time sequence information, the feature extraction stack layer extracts deep features of different modalities in parallel, the cross-modal attention layer calculates the interaction weight between modalities, the decision generation layer generates a preliminary control policy, and finally the output calibration layer converts it into a specific pressure adjustment value. The calculation delay of the whole process is controlled within 50ms, meeting the real-time control requirements.

[0072] Embodiment 3: The training process and adaptive optimization steps of the dynamic pressure balancing model are specifically defined in this embodiment. The training of the dynamic pressure balancing model is based on a historical control data set, and the data collection covers multiple operation cycles, each cycle containing a complete trolley operation process, such as steel pile positioning, clamp clamping, excavation operation, trolley displacement, etc. Each data sample of the historical control data set is composed of a sequence of consecutive M+1 operation stages, of which the first M stages are input features and the M+1 stage is the target output label. The input features include the pressure time sequence queue of each stage, the preset pressure threshold, the target displacement, the hydraulic oil temperature limit, the equipment code information, the sea state grade identification and the state feature set; the target output is the pressure adjustment value of the M+1 stage, which is determined by artificial calibration or expert experience combined with actual working conditions, serving as the benchmark for model learning.

[0073] The construction of the data set follows the following steps: first, through the multi-source sensor module and the time series data storage module, the original data of each control node is collected and stored according to the control period (such as 100 ms), and the data storage time covers at least 200 operation periods; second, the original data is preprocessed, including outlier rejection (using the IQR method, rejecting data points exceeding Q1-1.5IQR or Q3+1.5IQR range), missing value filling (linear interpolation method) and stage marking (according to the photoelectric switch trigger signal to divide the operation stage); finally, the preprocessed data is divided into independent samples according to the operation stage sequence, each sample contains a continuous M+1 stage feature vector and the corresponding pressure adjustment amount label. The data set is divided into training set and test set according to the ratio of 7:3, wherein the training set is used for model parameter optimization, and the test set is used for evaluating the generalization ability.

[0074] The model training adopts a supervised learning mode, taking the feature data of the first M stages in the training set as input and the pressure adjustment amount of the M+1 stage as target output. The loss function uses mean square error (MSE), and the calculation formula is:

[0075] wherein, is the loss value, is the number of samples in the training batch, is the pressure adjustment amount predicted by the model, is the actual label value. The optimization algorithm uses the Adam algorithm, the initial learning rate is set to 0.001, the momentum parameters β1=0.9, β2=0.999, and ε=1e-8. During training, the batch size (batchsize) is set to 32, the number of iterations (epoch) is 100, the model parameters are saved every 5 epochs, and the loss value is evaluated on the validation set (20% of the training data is divided as the validation set). If the validation loss does not decrease for 10 consecutive epochs, the early stopping mechanism (EarlyStopping) is triggered to avoid overfitting.

[0076] In the test phase, the test set is used to evaluate the model control error, and the error indicators include root mean square error (RMSE) and mean absolute error (MAE). If the error does not meet the convergence standard (such as RMSE>0.8MPa), then enter the adaptive optimization step. Adaptive optimization is based on the distribution characteristics of the test set error, and automatically adjusts the model structure or parameters, which can be divided into two scenarios: Scenario one: the error presents periodic fluctuations, when the test set loss curve appears regular oscillation (such as fluctuating once every 5 epochs), it is judged that the network has gradient disappearance or mode confusion problem. At this time, the following measures are taken: Add residual connection structure: Add residual skip connection between parallel modules in the feature extraction stack layer, the formula is:

[0077] where, is the input of the l-th layer, is the nonlinear transformation output of the l-th layer, is the output after residual connection. Residual connection preserves the original features through identity mapping, alleviating the gradient vanishing problem of deep network. Reduce learning rate: Multiply the current learning rate by the decay factor 0.5, for example, from 0.001 to 0.0005, to slow down the parameter update step and avoid skipping the optimal solution in the optimization process.

[0078] Scenario two: error continues to diverge, when the test set loss curve is monotonically increasing or has no convergence trend, it is judged that the network topology structure is unreasonable or the parameter initialization is improper. At this time, the following measures are taken: Reconstruct network topology: Adjust the connection mode of the feature extraction module, such as changing the parallel structure to the series structure, or increasing / decreasing the number of modules (such as L from 3 to 4) to change the feature interaction path. For example, connect the hydraulic parameter module and the mechanical motion module in series, process the hydraulic features first and then input the motion features, and strengthen the causal relationship modeling.

[0079] Re-initialize parameters: Use Kaiming initialization method (He Initialization) to re-initialize the network weights, which automatically calculates the appropriate initialization variance according to the activation function type (such as ReLU), the formula is:

[0080] where, is the weight parameter, is the number of input neurons, ensuring that the initialized network activation value has a suitable distribution range, avoiding gradient explosion or disappearance.

[0081] In the adaptive optimization process, model structure adjustment and parameter update are automatically implemented through code without human intervention. After each adjustment, the training process is reinitialized until the test set error converges to the preset threshold (such as RMSE≤0.5MPa) or reaches the maximum iteration number (such as 200 times). The model parameters of the training completed are stored in the non-volatile memory (such as EEPROM) of the controller, supporting power saving and power loading, ensuring that the system can be put into use without retraining after restarting.

[0082] It should be noted that the pressure adjustment amount label in the historical control data set needs to be strictly verified by engineering, and the label value is determined by technical personnel with rich experience in dredger operation combined with equipment operating manual, to ensure the rationality and safety of the label value. During model training, it is prohibited to use unverified simulation data or hypothetical data, and all input features and output labels are based on real working scene collection to ensure the engineering practicability of the model. In addition, the training environment uses an embedded computing platform (such as NVIDIA Jetson AGX Orin) integrated with a GPU acceleration module to ensure the efficiency of large-scale data training, with a single round of training period (100 epochs) controlled within 2 hours to meet the time cost requirements of engineering applications.

[0083] Embodiment 4: The embodiment specifically defines the architecture of the cooperative control correction model and the calculation method of the final control output. The cooperative control correction model is based on a graph neural network, which aims to model the motion coupling relationship of each actuator of the trolley system, and generate a cooperative correction coefficient based on real-time steel pile attitude data to compensate for nonlinear errors when multiple mechanisms are linked. The model is divided into a topological relationship modeling branch and a cooperative optimization branch, and the implementation details of each branch are as follows: The topological relationship modeling branch is used to establish the motion coupling relationship of each actuator of the trolley system. The actuators of the trolley system include hydraulic cylinders, steel pile clamps, walking mechanisms, etc., and each mechanism is dynamically related through mechanical connection or hydraulic oil circuit. For example, the extension and retraction of the hydraulic cylinder will drive the steel pile to move up and down through the connecting rod mechanism, the change of the inclination angle of the steel pile will affect the clamping force distribution of the clamp, and the displacement deviation of the walking mechanism may cause uneven force on the steel pile. In the modeling process, first, each actuator is abstracted as a node in the graph structure, and the number of nodes is determined according to the complexity of the system, usually 5 to 10 (such as cylinder node, clamp left node, clamp right node, trolley chassis node, steel pile top node, etc.). Node features include physical parameters of the mechanism (such as cylinder cross-sectional area, clamp clamping force range) and real-time state parameters (such as current cylinder pressure, clamp opening degree).

[0084] The edges between nodes represent the motion coupling relationship between mechanisms, and the weight of the edge is determined by correlation analysis in historical data. For example, the Pearson correlation coefficient of the change of the cylinder pressure and the change of the inclination angle of the steel pile is calculated. If the coefficient is 0.7, the weight of the corresponding edge is set to 0.7, indicating that there is a strong positive correlation between the two; if the coefficient is -0.3, the weight is set to -0.3, indicating a weak negative correlation. The weight of the edge ranges from -1 to 1, and the greater the absolute value, the higher the coupling strength. The construction of the graph structure adopts an undirected graph model, that is, the weight matrix of the edge is a symmetric matrix, reflecting the bidirectional influence between mechanisms. The topological relationship modeling branch is trained through a graph convolution network (GCN), the input is the node feature matrix and the initial adjacency matrix, and the output is the updated adjacency matrix, which represents the dynamic coupling strength between mechanisms.

[0085] The collaborative optimization branch fuses the motion coupling relationship with real-time steel pile attitude data to generate a collaborative correction coefficient. Real-time steel pile attitude data includes steel pile inclination angle, clamp clamping force distribution, and trolley displacement deviation. The steel pile inclination angle is collected by a dual-axis inclination sensor installed at the top of the steel pile, with an accuracy of ±0.1°. The sensor uses magnetic attraction installation, which is convenient for calibration and maintenance. The clamp clamping force distribution is collected by a strain gauge sensor array inside the clamp. The array contains 4 to 8 measuring points, which are evenly distributed on the clamping surface. The force value of each measuring point is converted to a voltage signal through a Wheatstone bridge circuit, and then input to the control system after amplification and analog-digital conversion. The trolley displacement deviation is monitored in real time by a laser displacement sensor installed on the guide rail. The sensor has a resolution of 0.1 mm and a detection range of 0 to 5000 mm.

[0086] The processing flow of the collaborative optimization branch is as follows: First, the real-time collected steel pile inclination angle (unit: °), clamp clamping force distribution (unit: kN), and trolley displacement deviation (unit: mm) are normalized to map to the [0, 1] interval to eliminate the dimension effect; Then, the normalized attitude data is input into the graph neural network as the node features, and the adjacency matrix output by the topological relationship modeling branch is fused. The fusion process is realized through multi-layer graph convolution operation. The calculation formula of each layer of graph convolution is:

[0087] wherein, is the node feature matrix of the th layer, is the adjacency matrix with self-loop added, is the corresponding degree matrix, is the trainable weight matrix, is an activation function (e.g., ReLU). After multiple layers of convolution, the node features contain the coupling information between mechanisms and real-time pose information. Finally, a multi-layer perceptron (MLP) is used for nonlinear transformation to output the collaborative correction coefficient, which ranges from -0.2 to 0.2. The sign of the correction coefficient indicates the direction of the control adjustment: a positive value indicates increasing the output of the current control node (e.g., increasing the cylinder pressure), and a negative value indicates decreasing the output; the absolute value indicates the adjustment amplitude, for example, when the coefficient is 0.1, the output adjustment amplitude is 10% of the reference value.

[0088] The calculation of the final control output combines the dynamic pressure adjustment value and the collaborative correction coefficient, and is achieved through weighted fusion and nonlinear compensation. The dynamic pressure adjustment value is output by the dynamic pressure balancing model and represents the pressure adjustment reference value based on historical data and multi-modal features; the collaborative correction coefficient is output by the collaborative control correction model and represents the correction amount based on real-time pose and mechanism coupling. In weighted fusion, the preset proportion is determined according to the system response characteristics, for example, the dynamic pressure adjustment value weight is 60% and the collaborative correction coefficient weight is 40%, and the calculation formula is:

[0089] The nonlinear compensation link is used to correct the influence of the cylinder load characteristics on the control accuracy. The load characteristics of the cylinder are represented by the nonlinearity of the pressure-flow curve, for example, the pressure fluctuation is sensitive to the flow change under light load, and the system inertia increases under heavy load, resulting in response lag. The load characteristics are determined by the pre-tested cylinder load-pressure curve, and the curve acquisition method is as follows: different loads (e.g., 50kN, 100kN, 150kN) are applied to both ends of the cylinder, and the corresponding relationship between the input pressure and the piston movement speed is recorded to form a piecewise linear curve. In nonlinear compensation, the curve is queried according to the current cylinder load size to automatically adjust the compensation coefficient. For example, when the load is 200kN, the compensation coefficient is 1.1, and the preliminary output needs to be multiplied by the coefficient, and the formula is:

[0090] The compensation coefficient usually ranges from 0.9 to 1.2, and is determined according to the test data.

[0091] Taking the actual working condition of a steel pile trolley of a cutter suction dredger as an example: in the steel pile lowering operation stage, the pressure adjustment value output by the dynamic pressure balance model is +5 MPa (indicating that the pressure needs to be increased by 5 MPa), the collaborative control correction model detects that the steel pile tilts 0.8° to the left, the clamping force of the left clamp is 120 kN, the clamping force of the right clamp is 80 kN, and the trolley displacement deviation is +15 mm (offset to the right). Through graph neural network analysis, there is a strong coupling relationship between the tilting of the steel pile and the uneven clamping force of the left and right clamps (the corresponding weight of the adjacency matrix is 0.9), and the right deviation of the trolley aggravates the stress on the left. The model output collaborative correction coefficient is -0.15, indicating that the pressure adjustment amount needs to be reduced by 15%. The preliminary output after weighted fusion is At this time, the cylinder load is 180 kN, according to the load characteristic curve, the compensation coefficient is 1.05, and the final control output is That is, the actual pressure adjustment command input to the cylinder is to increase 2.52 MPa.

[0092] Example 5 The present embodiment specifically limits the generation method of the hydraulic excitation control instruction set and the hardware composition of the control system. The generation of the hydraulic excitation control instruction set needs to integrate the final control output of all control nodes, and combine the device response characteristics and timing constraints to form an executable instruction sequence. Taking the steel pile trolley system of a certain cutter suction dredger as an example, the system includes 4 control nodes: left front cylinder node, right front cylinder node, left rear clamp node, and right rear walking mechanism node, and the final control output of each node corresponds to cylinder pressure adjustment value, clamp clamping force adjustment value, walking motor speed instruction, etc.

[0093] The first step of instruction set generation is time domain synchronization calibration. The calculation period of each control node is 100 ms, but due to sensor sampling, data transmission and other delays, there may be slight time deviations. Synchronization calibration is achieved through a system global clock (accuracy ±1 μs), all nodes update data at the same time at the beginning of each calculation period (such as 0 ms, 100 ms, 200 ms, etc.), and complete control quantity calculation before the end of the period (such as 95 ms), to ensure that the instruction generation time is consistent. For example, the left front cylinder node calculates a pressure adjustment value of +3 MPa in the nth period, and the right rear walking mechanism node calculates a displacement increment of +200 mm, both of which are transmitted to the instruction generation module at 95 ms of the nth period, to ensure that they are executed at the same time at the beginning of the n+1 period.

[0094] The second step is to generate timing constraints in combination with the device response characteristics. The device response characteristics include electromagnetic valve opening and closing delay, oil cylinder action lag, motor starting inertia, etc. Taking the electromagnetic valve as an example, its opening and closing delay time is 20 ms, the lag time of the oil cylinder from receiving the pressure command to completing the action is 80 ms, and the walking motor needs 150 ms from the speed command to reach a stable speed. The command sequence needs to set the execution order and delay time according to these characteristics. For example, in the steel pile positioning stage, the following operations need to be performed in sequence: The left rear clamp node sends a clamping command, considering the electromagnetic valve delay of 20 ms, the command sending time is set to the beginning of the cycle (0 ms), and the actual clamping action completion time is 20 ms; The left front oil cylinder node sends a pressurization command, which needs to wait for the clamp clamping in place signal (20 ms feedback), so the command sending time is set to 25 ms, and the oil cylinder action completion time is 25 ms + 80 ms = 105 ms; The right rear walking mechanism node sends a displacement command, which needs to wait for the oil cylinder pressure to stabilize (set the stabilization time to 50 ms), so the command sending time is set to 105 ms + 50 ms = 155 ms, and the motor reaches the target speed time is 155 ms + 150 ms = 305 ms.

[0095] Through the above timing constraints, it is ensured that each mechanism action is executed in the required order, avoiding mechanical impact or control failure due to synchronization errors. The command sequence uses an event triggering mechanism, with the completion signal of the previous action (such as the clamp in place switch, pressure sensor stable signal) as the trigger condition for the next action command, rather than simply relying on time delay, to adapt to changes in device response speed under different working conditions.

[0096] The hardware components of the control system include a multi-source sensor module, a state feature calculation module, a timing data storage module, a sea state identification module, a dynamic pressure equalization module, a collaborative control correction module, an instruction generation module, and a safety warning module. The specific implementation of each module is as follows: The multi-source sensor module is deployed at key positions of the trolley system: The pressure sensor (precision ±0.5% FS) is installed on the inlet and outlet pipes of the hydraulic cylinder, with an armored stainless steel shell to adapt to the humid and vibrating environment, and outputs a 4-20 mA analog signal; The electromagnetic valve response time monitoring module is composed of a microsecond-level timer and a signal conditioning circuit, and is connected in parallel to the electromagnetic valve control loop, which monitors the time difference between the drive signal and the feedback signal in real time; The laser displacement sensor (precision ±1 mm) is installed on both sides of the trolley guide rail, which monitors the trolley displacement in real time through the triangulation method, and outputs an RS485 digital signal; The temperature sensor (accuracy ±1℃) is a Pt100 thermal resistor buried in the hydraulic oil tank 10 cm away from the oil pump inlet to monitor the oil temperature changes. The vibration sensor (frequency range 0.1-1000Hz) is a three-axis accelerometer fixed to the non-stressed surface of the steel pile base and installed through a magnetic seat for easy disassembly and calibration.

[0097] The state feature calculation module is realized based on an FPGA chip and includes an independent data preprocessing unit and a feature calculation unit. The preprocessing unit filters (50Hz power frequency notch), amplifies, and digitizes (16-bit ADC) the sensor signals, and the feature calculation unit calculates five factors such as pressure fluctuation factor and response delay factor in real time according to the method described in Embodiment 1, with a processing delay of ≤10ms, and the results are transmitted to the dynamic pressure equalization module through a high-speed bus.

[0098] The time series data storage module adopts a circular buffer (Circular Buffer) structure composed of an SRAM memory and an address counter, and can store pressure data of 200 operation stages (each stage contains 100 sampling points, each point occupies 4 bytes, and the total capacity is 200x100x4=80KB). The buffer supports first-in, first-out (FIFO) reading and writing, and automatically covers the earliest data when it is full, ensuring that the latest historical window data is always retained.

[0099] The sea condition recognition module integrates an inertial navigation system (INS) and a wave radar: The INS calculates the ship's roll, pitch, and heading angle by fusing gyroscopes, accelerometers, and magnetometers, with a sampling frequency of 100Hz; The wave radar is an X-band radar installed on the bow mast to monitor wave height, period, and direction, with an update frequency of 1Hz; The data processing unit fuses the ship's attitude and wave parameters through an extended Kalman filter algorithm and outputs a sea condition level identifier (01-06) to the control system through a CAN bus.

[0100] The dynamic pressure equalization module and the collaborative control correction module are realized based on an NVIDIA Jetson AGX Orin embedded platform, which integrates a 6-core ARM CPU and a 200TOPS power GPU and supports CUDA acceleration. The multi-modal perception network of the dynamic pressure equalization module and the graph neural network of the collaborative control correction module are optimized through the TensorRT framework, with inference delays of 40ms and 30ms, respectively, meeting the real-time requirements.

[0101] The instruction generation module adopts an STM32H7 series microcontroller, integrates a CANopen master controller and a D / A conversion module. The microcontroller receives the calculation results of each module, executes time domain synchronous calibration and timing constraint algorithms, and generates control instructions including: analog instructions, such as oil cylinder pressure adjustment values (output through 4-20 mA signals); digital instructions, such as electromagnetic valve opening and closing signals (24V level signals); pulse instructions, such as walking motor speed pulses (output through a quadrature encoder interface). The instruction output period is consistent with the control period (100 ms), and the reliability is ensured through a hardware watchdog timer.

[0102] The safety warning module has a built-in historical fault database and an abnormal pattern recognition unit: The historical fault database stores more than 5000 fault records, and each record contains a state feature set at the time of failure, an operation stage, and a fault type (such as pressure overrun, clamp loosening); The abnormal pattern recognition unit is based on the One-Class SVM algorithm, which monitors the fluctuation of the state feature set in real time. When any factor exceeds the warning threshold (such as a pressure fluctuation factor > 0.8), a three-level alarm is triggered: first-level alarm (warning): the operation interface displays a yellow warning, and records abnormal data; second-level alarm (warning): accompanied by sound and light alarms, automatically reduce the control amount adjustment amplitude by 20%; third-level alarm (emergency): trigger the safety shutdown program, cut off the hydraulic power source and lock the position of the trolley.

[0103] Taking the abnormal temperature of hydraulic oil during the steel pile lowering process as an example: when the temperature sensor detects that the oil temperature peak value reaches 65℃ (exceeding the limit value 60℃), the state feature calculation module generates a thermal decay factor of 0.9 (close to the threshold value 1.0), and the safety warning module identifies a second-level alarm. The instruction generation module automatically reduces the pressure adjustment value of each oil cylinder by 20%, and starts the hydraulic oil cooling pump until the oil temperature falls within the safe range.

[0104] It should be noted that in this text, relational terms such as first and second are used only to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any such actual relationship or order between the entities or operations. Moreover, the terms "include", "contain" or any other variants thereof are intended to cover non-exclusive inclusion, so that the process, method, article or equipment including a series of elements not only includes those elements, but also includes other elements not explicitly listed or inherent to such process, method, article or equipment.

[0105] While embodiments of the present application have been shown and described, it is to be understood that various modifications, substitutions, replacements and variations can be made to these embodiments without departing from the principles and spirit of the present application, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A control method for a steel pile trolley of a cutter suction dredger based on hydraulic excitation early warning, characterized in that, include: For any control node in the trolley system, including the hydraulic actuator, steel pile clamp, and traveling mechanism, the influence of cylinder pressure, solenoid valve response time, displacement sensor data, hydraulic oil temperature, and environmental vibration parameters on system stability is collected. Based on the obtained influence levels, pressure fluctuation factor, response delay factor, displacement deviation factor, thermal attenuation factor, and mechanical resonance factor are calculated to form a set of state characteristics. A dynamic pressure equalization model is constructed. For any control node, the actual pressure data of the previous M time units of the node are arranged into a pressure time sequence queue according to the operation sequence within any control cycle. The preset pressure threshold, target displacement, hydraulic oil temperature limit and equipment coding information of the next operation stage are obtained, and the current sea state level is obtained. The pressure time sequence queue, preset pressure threshold, target displacement, hydraulic oil temperature limit, equipment coding information, sea state level and state feature set of the current control cycle are input into the dynamic pressure equalization model, and the dynamic pressure adjustment value of the node in the next operation stage is output. Construct a collaborative control correction model. For any control node, acquire real-time steel pile attitude data within any control cycle. Input the steel pile tilt angle, clamping force distribution and trolley displacement deviation of the current control cycle into the collaborative control correction model, and output the collaborative correction coefficient of the node in the next operation stage. For any control node, based on the dynamic pressure adjustment value and collaborative correction coefficient obtained in the current control cycle, the final control output of the node in the next operation stage is calculated. Integrate the final control outputs of all control nodes to generate the hydraulic excitation control instruction set for the trolley system.

2. The control method for steel pile trolley of a cutter suction dredger based on hydraulic excitation early warning as described in claim 1, characterized in that, The impact of the collected cylinder pressure value, solenoid valve response time, displacement sensor data, hydraulic oil temperature, and environmental vibration parameters on system stability specifically includes: For any given control node, the actual pressure data for each operational stage of that node within the historical operating cycle is acquired, and the solenoid valve opening and closing delay time, displacement sensor sampling value, hydraulic oil temperature peak value, and environmental vibration spectrum data for each operational stage are collected. Among these, the sea state level is divided into six categories based on wave height and period, with each category corresponding to a unique code. The equipment coding information includes the cylinder model identifier and the clamp type. The cylinder model identifier is divided into three levels according to the rated pressure, and the clamp type is divided into planar and curved surfaces according to the clamping surface structure. The state sequence information for each operation phase includes the pressure fluctuation amplitude of n consecutive sampling points of that node within the operation phase; the n sampling points of each operation phase are encoded as 1 to n in chronological order, and the sampling point with the highest frequency of pressure exceeding the limit within the historical period is encoded. The impact of cylinder pressure, solenoid valve response time, displacement deviation, hydraulic oil temperature and environmental vibration on the stability of the current control node are calculated. The impact of the first four items is obtained through correlation analysis of historical data and system faults, and the impact of environmental vibration is obtained by calculating the offset of the main peak frequency of the spectrum.

3. The control method for steel pile trolley of a cutter suction dredger based on hydraulic excitation early warning as described in claim 2, characterized in that, The calculation method of the state feature set is as follows: based on the influence of cylinder pressure value, solenoid valve response time, displacement deviation, hydraulic oil temperature and environmental vibration, the pressure fluctuation factor to mechanical resonance factor is obtained by normalized weighted summation.

4. The control method for steel pile trolley of a cutter suction dredger based on hydraulic excitation early warning as described in claim 3, characterized in that, The dynamic pressure equalization model is built on a multimodal perception network and includes a data fusion layer, a temporal coding layer, a feature extraction stacking layer, a cross-modal attention layer, a decision generation layer, and an output calibration layer. The data fusion layer is used to convert pressure time series queues, preset pressure thresholds, target displacement, hydraulic oil temperature limits, equipment coding information, sea state level identifiers and state feature sets into multidimensional feature tensors; The timing coding layer is used to generate operation sequence codes for each operation stage in the pressure timing queue to maintain the continuity of the job flow; The feature extraction stack layer contains L parallel feature extraction modules, each of which processes hydraulic parameter features, mechanical motion features, and environmental disturbance features respectively. The cross-modal attention layer is used to perform cross-attention calculations on the state feature set and the multi-dimensional features of the feature extraction output to generate a dynamic weight allocation matrix. The decision generation layer is used to fuse the dynamic weight allocation matrix with the feature extraction results to generate a preliminary control strategy vector; The output calibration layer is used to map the initial control strategy vector to dynamic pressure adjustment values.

5. A control method for a steel pile trolley of a cutter suction dredger based on hydraulic excitation early warning, as described in claim 4, is characterized in that... The training process of the dynamic pressure equilibrium model includes: Collect historical control datasets from multiple work cycles. Each data sample contains pressure data for M+1 consecutive operation stages, influencing factor data for the M+1th stage, and a set of state features. Divide the dataset into a training set and a test set. Using the data and influencing factors of the first M stages in the training set as input, and the pressure adjustment amount of the (M+1)th stage as the target output, the network parameters are optimized through error backpropagation. The model control error is evaluated using a test set. If the error does not meet the convergence criteria, the number of feature extraction modules or the dimension of the attention head is adjusted and the model is retrained.

6. A control method for a cutter suction dredger steel pile trolley based on hydraulic excitation early warning, as described in claim 5, is characterized in that... The cooperative control correction model is constructed based on a graph neural network, including a topology modeling branch and a cooperative optimization branch; The topology modeling branch establishes the motion coupling relationships between the actuators of the trolley system through graph structure learning; The collaborative optimization branch integrates the motion coupling relationship with the steel pile attitude data, and outputs collaborative correction coefficients after calculation by a multilayer perceptron.

7. A control method for a cutter suction dredger steel pile trolley based on hydraulic excitation early warning, as described in claim 6, is characterized in that... The final control output is calculated by weighting and fusing the dynamic pressure adjustment value and the collaborative correction coefficient according to a preset ratio, and then performing nonlinear compensation in combination with the current cylinder load characteristics.

8. A control method for a cutter suction dredger steel pile trolley based on hydraulic excitation early warning, as described in claim 7, is characterized in that... The hydraulic excitation control instruction set is generated by performing time-domain synchronous calibration on the final control output of all control nodes and generating an instruction sequence with timing constraints based on the equipment response characteristics.

9. A control method for a cutter suction dredger steel pile trolley based on hydraulic excitation early warning, as described in claim 5, is characterized in that, The model training process further includes an adaptive optimization step: Based on the distribution characteristics of the test set error, the connection method of the feature extraction module or the activation function of the attention head is automatically adjusted; if the error shows periodic fluctuations, the residual connection structure is increased and the learning rate is reduced; if the error continues to diverge, the parameters are reinitialized after reconstructing the network topology.

10. A control system for a cutter suction dredger steel pile trolley based on hydraulic excitation early warning, used to implement the control method as described in any one of claims 1 to 9, characterized in that, include: Multi-source sensor modules are configured in the hydraulic actuator, steel pile clamp and walking mechanism of the trolley system to collect in real time the cylinder pressure value, solenoid valve response time, displacement sensor data, hydraulic oil temperature and environmental vibration parameters of each control node. The state feature calculation module, connected to the multi-source sensor module, is used to calculate the pressure fluctuation factor, response delay factor, displacement deviation factor, thermal attenuation factor, and mechanical resonance factor based on the collected data, and generate a set of state features. The time-series data storage module is equipped with a circular cache unit, which is used to store the actual pressure data of the first M time units of each control node in the order of operation, forming a pressure time-series queue. The sea state identification module integrates a ship attitude sensor and a wave monitoring unit to analyze the current sea state level in real time. The dynamic pressure equalization module, built on a multimodal neural network, is connected to the state feature calculation module, time-series data storage module, and sea state recognition module at its input end. It is used to receive pressure time-series queues, preset pressure thresholds, target displacement, hydraulic oil temperature limits, equipment coding information, and sea state level identifiers, and to calculate and generate dynamic pressure adjustment values ​​through a cross-modal attention mechanism. The collaborative control correction module is equipped with a steel pile attitude monitoring unit and a clamp stress analysis unit to obtain the real-time steel pile tilt angle, clamp clamping force distribution and trolley displacement deviation. It models the motion coupling relationship of the mechanism through a graph neural network and outputs the collaborative correction coefficient. The instruction generation module, with its input end connected to the dynamic pressure equalization module and the collaborative control correction module, is used to perform nonlinear fusion calculation of the dynamic pressure adjustment value and the collaborative correction coefficient to generate the final control output of each control node, and simultaneously generate a hydraulic excitation control instruction set with time constraints. The safety early warning module has a built-in historical fault database and an abnormal pattern recognition unit, which is used to monitor abnormal fluctuations in the set of status characteristics in real time, trigger hierarchical alarm signals, and feed them back to the instruction generation module for control quantity compensation.

Images