Method for damping adjustment of a wave compensating device

Abstract

The application relates to the field of ocean engineering and discloses a damping adjustment method of a wave compensation device. The method comprises the following steps: collecting a heave acceleration sequence and a load sequence of a hydraulic cylinder of the wave compensation device within a preset time window; aligning and merging the heave acceleration sequence and the load sequence of the hydraulic cylinder according to timestamps to generate a joint state feature; inputting the joint state feature into a pre-trained time sequence feature extraction network to obtain a damping adjustment value for a current load state; and outputting a control instruction to a proportional overflow valve of the wave compensation device according to the damping adjustment value to adjust the damping force of the hydraulic cylinder. Through the technical scheme, the secondary impact risk of the compensation steel wire rope is controlled within a safe range.

Description

Technical Field

[0001] This application relates to the field of marine engineering, and in particular to a damping adjustment method for a wave compensation device. Background Technology

[0002] In the fields of marine engineering and offshore lifting operations, wave compensation devices are widely used to isolate the effects of ship heave on underwater loads.

[0003] In related technologies, wave compensation devices typically collect vertical displacement or acceleration data of the hull, calculate the amount of reverse displacement required for compensation, and use a PID algorithm to send adjustment commands to the proportional relief valve in the hydraulic circuit. By changing the damping force of the hydraulic cylinder, the relative motion caused by the waves is counteracted.

[0004] When faced with irregular and fluctuating sea conditions, the proportional relief valve itself has a certain phase lag in its mechanical response. This lag causes the damping force output by the hydraulic cylinder to be unable to be strictly synchronized with the instantaneous peak value of the wave. The compensating wire rope will experience a brief period of stress vacuum during the yielding phase, and will be subjected to a secondary transient impact force during subsequent tightening, which accelerates the metal fatigue wear of the load-bearing wire rope.

[0005] How to solve the problem of secondary impact on the wire rope caused by the above-mentioned damping force misalignment is a technical problem that needs to be solved by those skilled in the art. Summary of the Invention

[0006] This application provides a damping adjustment method for a wave compensation device to at least partially solve the above-mentioned technical problems.

[0007] To achieve the above objectives, this application provides a damping adjustment method for a wave compensation device, comprising: The heave acceleration sequence and hydraulic cylinder load sequence of the wave compensation device are collected within a preset time window; the heave acceleration sequence reflects the vertical motion state of the platform on which the wave compensation device is located; the hydraulic cylinder load sequence reflects the current load state of the wave compensation device. The heave acceleration sequence and the hydraulic cylinder load sequence are aligned and merged according to timestamps to generate joint state features; The joint state features are input into a pre-trained temporal feature extraction network to obtain the damping adjustment value for the current load state; the loss function of the temporal feature extraction network includes a penalty term for the rate of change of rope tension of the compensation wire rope of the wave compensation device; the penalty term increases the loss backpropagation weight when the rate of change of rope tension exceeds a preset tension fluctuation threshold. The control command is output to the proportional relief valve of the wave compensation device according to the damping adjustment value, thereby adjusting the damping force of the hydraulic cylinder.

[0008] In this embodiment of the application, the above technical solution enables the output damping adjustment value to suppress the instantaneous loss of tension of the wire rope during wave compensation, and controls the risk of secondary impact of the compensation wire rope within a safe range.

[0009] Other features and advantages of this application will be described in detail in the following detailed description section. Attached Figure Description

[0010] To more clearly illustrate the technical solutions in the embodiments of this application, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0011] Figure 1 This is a flowchart illustrating the steps of a damping adjustment method for a wave compensation device provided in an exemplary embodiment of this application. Detailed Implementation

[0012] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of this application, and not all of them. All other embodiments obtained by those skilled in the art based on the embodiments of this application without creative effort are within the protection scope of this application.

[0013] This application provides a damping adjustment method for a wave compensation device. Please refer to [link / reference]. Figure 1 The damping adjustment method of a wave compensation device provided in this application includes the following steps: Step 101: Collect the heave acceleration sequence and hydraulic cylinder load sequence of the wave compensation device within a preset time window. The heave acceleration sequence reflects the vertical motion state of the platform on which the wave compensation device is located; the hydraulic cylinder load sequence reflects the current load state of the wave compensation device. Specifically, the preset time window is backtracked from the current moment. Within this time window, the output data of the heave acceleration sensor and the hydraulic cylinder pressure transmitter are collected synchronously at a preset sampling frequency. The fluctuation pattern of the heave acceleration sequence reflects the intensity of the vertical motion of the ship under the current sea state, with the rising segment corresponding to the ship accelerating upward and the falling segment corresponding to the ship decelerating downward.

[0014] Step 102: Align and merge the heave acceleration sequence and the hydraulic cylinder load sequence according to their timestamps to generate a joint state feature. Specifically, since the heave acceleration sensor and pressure transmitter sample independently, there is a certain clock deviation between the two sensors, causing the heave acceleration value and load value at the same moment to be misaligned on the time axis. To solve this problem, using the timestamp of the heave acceleration sequence as a reference, the sampling points in the load sequence closest to each acceleration timestamp are resampled to the same time grid through linear interpolation to generate an aligned load sequence of the same length as the heave acceleration sequence. The aligned load sequence and the heave acceleration sequence are then concatenated along the channel dimension to generate a two-dimensional joint state feature with a time step and a feature channel dimension. The time step dimension corresponds to the number of sampling points in the acquisition window, and the feature channel dimension is 2, corresponding to the acceleration channel and the load channel, respectively.

[0015] Step 103: Input the joint state features into a pre-trained temporal feature extraction network to obtain the damping adjustment value for the current load state. Specifically, the temporal feature extraction network adopts a multi-scale one-dimensional convolutional architecture, performing layer-by-layer convolution operations on the input joint state features along the time dimension to extract the damping adjustment rules related to the current sea state and load state; the fully connected layer at the end of the temporal feature extraction network maps the convolutional features to the damping adjustment value; the damping adjustment value mentioned here refers to the proportional relief valve opening command value corresponding to the target damping force expected to be achieved by the hydraulic cylinder.

[0016] Step 104: Output control commands to the proportional relief valve of the wave compensation device according to the damping adjustment value to adjust the damping force of the hydraulic cylinder. Specifically, compare the damping adjustment value output by the timing feature extraction network with the current measured valve core opening of the proportional relief valve and calculate the absolute difference between the two. When the absolute difference exceeds the preset dead zone width, convert the damping adjustment value into a corresponding drive current pulse signal and send it to the proportional relief valve to drive the valve core to adjust the opening, thereby changing the damping force of the hydraulic cylinder. When the absolute difference falls within the dead zone width range, do not send control commands to avoid mechanical oscillation caused by excessive response of the proportional relief valve.

[0017] In one alternative, the dead zone width can be adjusted according to the short-term volatility of the hydraulic cylinder load sequence; the adaptive dead zone width is obtained by calculating the sliding variance of the load sequence within the historical time window and multiplying it by a preset scaling factor; the dead zone is reduced in calm sea conditions to improve response accuracy, and expanded in rough sea conditions to suppress noise interference.

[0018] In another implementation, when the instantaneous sag exceeds the abnormal wave warning threshold, the dead zone width is multiplied by a preset attenuation coefficient to obtain the emergency dead zone width, which is then used to replace the regular dead zone width in the control decision of the proportional relief valve, thereby releasing the control command for extreme waves and reducing the opening lag time of the proportional relief valve.

[0019] In this embodiment, by aligning the heave acceleration sequence and the hydraulic cylinder load sequence by timestamp and merging them into a joint state feature, which is then fed into a time-series feature extraction network, the calculation of the damping adjustment value takes into account the vertical motion state of the platform and the real-time load state of the hydraulic cylinder, thus avoiding the compensation lag caused by open-loop adjustment in the prior art. Since the time-series feature extraction network introduces the rope tension change rate into the penalty term of the loss function during training, the time-series feature extraction network includes the constraint condition of the wire rope force, thereby enabling the output damping adjustment value to suppress the instantaneous loss of tension of the wire rope during wave compensation and control the secondary impact risk of the compensation wire rope within a safe range.

[0020] In some embodiments, the heave acceleration sequence and the hydraulic cylinder load sequence are aligned and merged according to timestamps to generate a joint state feature, including: The system acquires the first device timestamp for each sampling point in the heave acceleration sequence and the second device timestamp for each sampling point in the hydraulic cylinder load sequence. Specifically, the heave acceleration sensor and the pressure transmitter each maintain an independent local clock, and their clock references have a certain frequency offset and phase shift, resulting in inconsistent timestamps for measurement data at the same moment in their respective sequences. The first device timestamp refers to the system time stamp assigned to each sampling point by the heave acceleration sensor data acquisition card; similarly, the second device timestamp is the timestamp recorded by the pressure transmitter data acquisition card.

[0021] Step 202: Using the timestamp of the first device as a reference, resample the load sequence of the hydraulic cylinder using a linear interpolation algorithm to obtain an aligned load sequence with the length of the heave acceleration sequence. Specifically, after determining the time axis of the heave acceleration sequence, linear interpolation estimation is performed on the load values ​​corresponding to the missing timestamps in the load sequence. The specific operation includes: traversing each timestamp of the heave acceleration sequence and calculating the interpolation result for two adjacent sampling points in the load sequence. It can be understood that linear interpolation assumes that the load signal changes linearly with time between two measured sampling points. This method maps two sets of data that were originally misaligned on the time axis onto a unified time grid.

[0022] Step 203: Combine the heave acceleration sequence and the aligned load sequence to generate a joint state feature with time step and feature channel dimension. Specifically, the heave acceleration sequence is used as the first feature channel and the aligned load sequence is used as the second feature channel. They are concatenated along the channel dimension to generate a two-dimensional array of shape (N,2), where N is the number of sampling points within the time window. This two-dimensional array serves as the input to the temporal feature extraction network. The first channel carries the vertical motion information of the heave acceleration, and the second channel carries the mechanical response information of the hydraulic cylinder load. The feature channel dimension refers to the number of columns of the joint state feature along the channel direction. The two channels correspond to different signal sources.

[0023] In one alternative approach, if the sampling frequencies of the heave acceleration sensor and the pressure transmitter are integer multiples of each other, time alignment can be achieved using upsampling or downsampling methods without the need for linear interpolation. When the sampling frequency of the load sensor is higher than that of the acceleration sensor, the load sequence is downsampled to align with the timestamp of the acceleration sequence. When the sampling frequency of the load sensor is lower than that of the acceleration sensor, the load sequence is upsampled to the sampling rate of the acceleration sensor using zero-order hold interpolation or linear interpolation.

[0024] In this embodiment, the clock deviation of multiple sensors is corrected by linear interpolation resampling of the hydraulic cylinder load sequence based on the heave acceleration timestamp, ensuring that the two input channels of the subsequent time-series feature extraction network are synchronized on the time axis. Time alignment is a prerequisite for multimodal data fusion. Excessive alignment error will cause distortion of the time correlation features extracted by the convolutional layer, thereby affecting the prediction accuracy of the damping adjustment value. After generating the joint state feature with a dual-channel structure, the time-series information of the two channels can be jointly modeled by the time-series feature extraction network within the same time window to capture the time-delay correlation between the change in heave acceleration and the hydraulic cylinder load response.

[0025] In some embodiments, inputting joint state features into a pre-trained temporal feature extraction network includes: The joint state features are input into a parallel multi-scale one-dimensional convolutional layer of a temporal feature extraction network. Specifically, the parallel multi-scale one-dimensional convolutional layer contains two branches: a first short-window convolutional kernel and a second long-window convolutional kernel. The two branches simultaneously perform one-dimensional convolution operations on the input joint state features in parallel. The first short-window convolutional kernel is used to capture the rapidly changing components in the heave acceleration and load sequence. It can be understood that the convolution result of the short-window convolutional kernel reflects the high-frequency rate of change of the signal within a local time window, and its response characteristics are similar to differential operations, making it sensitive to pulse-type wave loads.

[0026] The joint state features are extracted sequentially over time using a first short-window convolutional kernel and a second long-window convolutional kernel, generating transient and trend feature maps. Specifically, the window length of the second long-window convolutional kernel is set to a preset long-window size, covering a longer historical time range, and is used to extract the slow-varying trend components of heave acceleration and load sequences. The output of the long-window convolutional kernel reflects the statistical average level of the signal over a longer period, and its response characteristics are similar to low-pass filtering or integral operations, making it sensitive to continuous sea surface fluctuations. The outputs of the short-window and long-window convolutional kernels are denoted as transient feature maps and trend feature maps, respectively, both having the same time step dimension but different channel dimensions.

[0027] A fused feature map is generated by combining transient and trend feature maps through multi-channel features. Specifically, the transient and trend feature maps are concatenated along the channel dimension to form a multi-channel fused feature map containing both high-frequency detail information and low-frequency trend information. The number of channels in this fused feature map is equal to the sum of the number of output channels of the short-window convolutional kernel and the number of output channels of the long-window convolutional kernel. The multi-channel feature combination mentioned here refers to superimposing corresponding channels of the feature maps of the two branches at the same time position to form a composite temporal representation containing dual-scale information.

[0028] The inherent response delay time of the hydraulic cylinder is obtained; the inherent response delay time is multiplied by a preset sampling frequency, and the product is rounded to generate an integer pooling step size. Specifically, the inherent response delay time of the hydraulic cylinder refers to the time delay from when the controller issues a drive current pulse command to when the proportional relief valve spool actually moves to the target opening degree and the hydraulic cylinder output damping force changes; this delay time is determined by the mechanical inertia of the valve spool, the compressibility of the hydraulic oil, and the pipeline conduction time, and can be pre-calibrated experimentally; the integer pooling step size mentioned here is to convert the response delay time into the number of sampling points so as to align the time resolution of the timing feature extraction network output with the actual response speed of the hydraulic cylinder in subsequent pooling operations.

[0029] The fused feature map is input into the max pooling layer of the temporal feature extraction network. A sliding window extraction is performed along the time dimension of the fused feature map with an integer pooling step size to obtain salient peak features and output the dimensionality-reduced temporal feature vector. Specifically, the max pooling layer slides along the time dimension with an integer pooling step size as the window width, taking the maximum value of each channel within each window as the representative feature value of that window. Since the pooling step size corresponds to the hydraulic cylinder response delay, each element in the time-series feature vector after pooling dimensionality reduction corresponds to a response delay period, ensuring that the output of the temporal feature extraction network matches the action cycle of the hydraulic actuator. The salient peak features mentioned here refer to the local maxima of the fused feature map within each pooling window, reflecting the peak level of wave impact intensity within that period.

[0030] In one alternative, if the response delay of the hydraulic cylinder is nonlinear with load variation, a piecewise linear pooling strategy can be adopted: a smaller pooling step size is used in the light load range to retain more detailed features, and a larger pooling step size is used in the heavy load range to filter noise. In another implementation, max pooling can be replaced with weighted average pooling, with the weights determined according to the variance of the features within each window, and the window with the larger variance has a higher weight.

[0031] In this embodiment, the transient impact component and the continuous heave component in the joint state features are extracted simultaneously by parallel multi-scale one-dimensional convolutional layers. The complementary design of short and long windows enables the temporal feature extraction network to respond to both the transient peak and the overall trend of the wave. Based on the inherent response delay time of the hydraulic cylinder as the calculation benchmark for the pooling step size, the temporal resolution of the output of the temporal feature extraction network is matched with the response characteristics of the hydraulic actuator, avoiding the problem that the high-frequency jitter commands output by the temporal feature extraction network cannot be actually executed by the hydraulic system. The max pooling operation effectively compresses the dimension of the temporal feature vector and reduces the computational burden of the tail fully connected layer.

[0032] In some embodiments, obtaining the damping adjustment value for the current load state includes: The dimensionality-reduced temporal feature vector is split into a first sub-vector reflecting steady-state load characteristics and a second sub-vector reflecting transient impact characteristics. Specifically, the feature channel combination boundary on which the temporal feature vector splitting is based is formed in the multi-channel feature combination stage: this boundary defines the channel index boundary of the output features of the short-window convolution kernel and the long-window convolution kernel in the fused feature map; along the channel dimension, using this boundary as the dividing point, the part of the temporal feature vector belonging to the long-window convolution kernel channel is extracted as the first sub-vector, and the part belonging to the short-window convolution kernel channel is extracted as the second sub-vector; it can be understood that the first sub-vector carries the steady-state load change information caused by the continuous heave and sag of the ship, and its time change is relatively slow; the second sub-vector carries the sudden load jump information caused by the transient impact of waves, and its time change is drastic.

[0033] The first sub-vector is input into the first linear mapping branch of the fully connected layer at the end of the temporal feature extraction network to calculate the predicted value of the base damping opening. Specifically, the first linear mapping branch is composed of a single-layer fully connected network, with the activation function being a linear rectified function or an identity mapping, mapping the first sub-vector into a predicted value of the form; this predicted value reflects the baseline damping force level required to maintain compensation balance under the current steady-state load conditions.

[0034] The second sub-vector is input to the second nonlinear mapping branch of the fully connected layer at the tail end, and the mapping result is truncated using an activation function to output a damping compensation coefficient for the impact energy. Specifically, the second nonlinear mapping branch adds a nonlinear activation layer on top of the first linear mapping branch; this activation function has upper and lower thresholds to clamp the output within a preset mechanical saturation boundary; the damping compensation coefficient mentioned here refers to the impact correction coefficient superimposed on the predicted value of the basic damping opening. A value greater than 1 indicates that damping needs to be strengthened to resist the impact, while a value less than 1 indicates that damping can be appropriately relaxed to reduce energy loss; it can be understood that the truncation processing of the nonlinear activation layer prevents the time-series feature extraction network from outputting control commands that exceed the capabilities of the actuator under extreme sea conditions.

[0035] The predicted base damping opening is multiplied by the damping compensation coefficient, and the output reflecting the expected damping opening is used as the damping adjustment value. Specifically, the predicted base damping opening output from the first linear branch is multiplied by the damping compensation coefficient output from the second nonlinear branch to obtain the final damping adjustment value. When the damping compensation coefficient is greater than 1, the final threshold is greater than the base predicted value, triggering enhanced damping; when the damping compensation coefficient is less than 1, the final threshold is less than the base predicted value, allowing appropriate relaxation of damping. This product operation unifies the steady-state reference and transient correction in the same output, realizing stable and transient coordinated control of damping adjustment.

[0036] In one alternative, if the transient impact intensity of the second sub-vector exceeds the preset upper limit of impact intensity, the damping compensation coefficient corresponding to the upper limit can be output in a fixed manner, instead of relying on the nonlinear mapping result of the time-series feature extraction network to ensure that the damping compensation direction is always correct in extreme cases.

[0037] In some embodiments, the loss function of the temporal feature extraction network is constructed through the following process: During the backpropagation training phase of the temporal feature extraction network, the damping adjustment prediction sequence of the current batch is obtained, and the mean square error between the damping adjustment prediction sequence and the preset label is calculated as the basic mapping loss. Specifically, during training, each training batch contains several sets of input samples, and each set of samples corresponds to the joint state features of a time window and its corresponding expected damping adjustment label; the mean square error between the prediction sequence output by the forward propagation of the temporal feature extraction network and the label sequence is used as the basic term of the loss function; the preset label here refers to the expected damping opening value given by the historical best control record in the training dataset.

[0038] The measured tension sequence of the compensating wire rope, aligned with the damping adjustment prediction sequence, is obtained from the training set. Specifically, the measured tension sequence is acquired and recorded in real time by a tension sensor installed at the connection end between the compensating wire rope and the wave compensation device. The tension value at each sampling moment is aligned with the heave acceleration and hydraulic cylinder load at the same moment on the time axis.

[0039] Step 503: Perform differential calculations on adjacent sampling points of the measured tension sequence and obtain the absolute value of the differential results to generate the rope tension change rate representing the instantaneous tension change. Specifically, when wave impact causes insufficient damping force of the hydraulic cylinder, the compensating wire rope will suddenly change from a tensile state to a slack state and then suddenly tighten, at which point the tension sequence will show a sharp jump.

[0040] Step 504: Multiply the rate of change of rope tension by a pre-set penalty coefficient to construct a penalty term used to constrain the mechanical fatigue boundary. Specifically, the magnitude of the penalty coefficient determines the sensitivity of the loss function to tension jumps: the larger the penalty coefficient, the more the temporal feature extraction network tends to output a damping strategy that can suppress sudden changes in the wire rope during training; the mechanical fatigue boundary mentioned here refers to the critical condition for compensating for metal fatigue damage in the wire rope when it is repeatedly subjected to sudden loads. The design goal of constraining this boundary is to reduce the risk of fatigue fracture of the wire rope caused by repeated de-tensioning-tightening cycles.

[0041] Step 505: Add the basic mapping loss and the penalty term to generate the loss function. Specifically, in the total loss function, the basic mapping loss drives the temporal feature extraction network to learn the prediction accuracy of damping adjustment, and the penalty term drives the temporal feature extraction network to learn to suppress sudden changes in wire rope tension; after adding the two, the gradient is calculated through the backpropagation algorithm and the weight parameters of the temporal feature extraction network are updated.

[0042] In this embodiment, a penalty term for the rate of change of rope tension is introduced into the loss function, embedding the mechanical fatigue constraint of the compensating wire rope into the training objective of the temporal feature extraction network. The basic mapping loss ensures the accuracy of damping prediction, while the penalty term ensures the safety of the wire rope under stress, thus achieving a balance between accuracy and safety in the damping adjustment strategy output by the temporal feature extraction network. When the rate of change of rope tension is large, the value of the penalty term increases, forcing the temporal feature extraction network to adjust the damping parameters to suppress subsequent abrupt changes in the wire rope, thereby reducing the probability of fatigue damage to the compensating wire rope caused by repeated impacts.

[0043] In some embodiments, a penalty term for constraining mechanical fatigue boundaries is constructed by multiplying the rate of change of rope tension by a pre-defined penalty coefficient, including: The system determines whether the current tension value in the measured tension sequence is less than a preset relaxation threshold. Specifically, the relaxation threshold refers to the critical tension value at which the compensating wire rope is about to transition from a tensioned state to a relaxed state. This threshold can be determined through a calibration experiment: gradually reduce the preload on the compensating wire rope and record the tension reading when the wire rope begins to show visible relaxation. This reading is used as a reference value for the relaxation threshold. In actual control, when the measured tension drops below the relaxation threshold, it indicates that the traction force of the wave compensation device on the wire rope is insufficient to maintain its tension, and the wire rope is about to enter the untension stage.

[0044] When the current tension value is less than the relaxation threshold, the pre-set penalty coefficient is multiplied by a preset exponential amplification factor to generate an updated penalty coefficient. Specifically, this can be understood as the exponential amplification method causing the loss function to non-linearly increase the penalty weight for tension abrupt changes when the wire rope's stress deteriorates, forcing the temporal feature extraction network to more aggressively adjust the damping parameters during backpropagation training to restore the tensile tension of the wire rope.

[0045] Step 603: Multiply the rope tension change rate by the updated penalty coefficient to generate a penalty term. Specifically, under the same rope tension change rate, the penalty term value corresponding to the slack state is higher than the penalty term value corresponding to the normal tension state, thereby strengthening the constraint on the tension mutation during the slack phase of the wire rope in the loss function.

[0046] In one alternative, the exponential amplification factor can be replaced by a piecewise linear function: when the current tension is below the relaxation threshold, the penalty coefficient increases linearly rather than exponentially to reduce computational complexity; in another implementation, the duration of tension below the relaxation threshold can be further introduced as an additional amplification condition, with a higher penalty coefficient for longer duration.

[0047] In this embodiment, by exponentially amplifying the penalty coefficient when the current tension is below the relaxation threshold, the loss function becomes more sensitive to tension abrupt changes when the wire rope is in the relaxation stage. Because the weight of the penalty term in the loss function increases non-linearly, the temporal feature extraction network increases the gradient magnitude of the damping parameter adjustment during the backpropagation training phase, forcing the network to learn a predictive strategy of increasing damping in advance before the wire rope is about to lose tension. This scheme effectively reduces the response time window of damping compensation, lowers the peak tension of the compensated wire rope in the relaxation-tensioning cycle, and slows down the propagation rate of metal fatigue cracks.

[0048] In some embodiments, the proportional relief valve that outputs a control command to the wave compensation device according to the damping adjustment value includes: Step 701: Obtain the current actual valve spool opening of the proportional relief valve; calculate the absolute difference between the damping adjustment value and the actual valve spool opening. Specifically, the actual valve spool opening is measured in real time by the position feedback sensor of the proportional relief valve and transmitted back to the controller; the damping adjustment value is subtracted from the measured valve spool opening and the absolute value is taken to obtain the amplitude of the current control error; the dead zone width mentioned here refers to the minimum control error threshold that can be ignored. When the absolute difference falls within the dead zone width, no control command is sent to avoid mechanical oscillation of the proportional relief valve due to frequent small adjustments.

[0049] Step 702: Determine whether the absolute difference is greater than the preset dead zone width. Specifically, the dead zone width is determined through the following process: Obtain the hydraulic cylinder load sequence within a preset historical time window from the current time, and calculate the sliding variance of the load sequence; multiply the sliding variance by a preset scaling factor to generate the dead zone width; it can be understood that the sliding variance reflects the degree of fluctuation of the hydraulic cylinder load within the recent historical window. When the sea state is stable and the load fluctuation is small, the sliding variance is small, the corresponding dead zone width is narrow, and the control accuracy is high; when the waves are violent and the load jumps are large, the sliding variance is large, the corresponding dead zone width is wide, and the control strategy tends to be conservative to suppress noise interference; the dead zone width increases proportionally with the increase of the degree of fluctuation of the load sequence.

[0050] Step 703: When the absolute difference is determined to be greater than the dead zone width, the damping adjustment value is converted into a drive current pulse and output as a control command to the proportional relief valve. Specifically, there is a preset mapping relationship between the damping adjustment value and the drive current of the proportional relief valve. The damping adjustment value is converted into the corresponding drive current through this mapping relationship. The drive current is sent to the power drive circuit of the proportional relief valve in the form of a pulse signal through the controller output port, driving the valve core to move to the opening position corresponding to the target damping force.

[0051] The instantaneous judder value is obtained by taking the first derivative of the data within the current time window in the heave acceleration sequence. Specifically, the instantaneous judder refers to the first derivative of acceleration with respect to time, reflecting the rate of change of heave acceleration or the degree of abrupt changes in the higher order of vertical motion. It can be understood that when a ship encounters abnormal waves, the heave acceleration will undergo a drastic jump in a very short time, corresponding to the peak of the judder value, which can sensitively identify the arrival of extreme sea states.

[0052] A monitoring boundary condition is constructed for the instantaneous sag value to determine whether the instantaneous sag value exceeds a preset abnormal wave warning threshold. Specifically, the abnormal wave warning threshold is obtained by calibrating historical extreme sea state data: the sag peak value of the heave acceleration sequence is extracted from the recorded abnormal wave events, and a preset proportion of this peak value is used as the warning threshold; when the instantaneous sag first exceeds the abnormal wave warning threshold, it is determined that the current sea state has entered the abnormal wave stage, and the system immediately initiates the dead zone over-limit response.

[0053] When the instantaneous sag value exceeds the abnormal wave warning threshold, the dead zone over-limit response is activated. Specifically, the dead zone over-limit response multiplies the preset attenuation coefficient by the current dead zone width to obtain the emergency dead zone width. The emergency dead zone width is much smaller than the normal dead zone width. Its purpose is to significantly reduce the allowable negligible control error range, so that the proportional relief valve can respond to a certain adjustment demand that was originally filtered out by the dead zone.

[0054] The absolute difference determination of the proportional relief valve is achieved by replacing the normal dead zone width with the emergency dead zone width, thereby releasing control commands corresponding to transient wave characteristics and reducing the opening lag time of the proportional relief valve. Specifically, the absolute difference determination is performed again with the emergency dead zone width instead of the normal dead zone width: when the emergency dead zone width is extremely small, control commands that originally fell within the normal dead zone are released, and the proportional relief valve can receive control updates at a higher frequency during distorted waves, thus adjusting the damping force in time to cope with the transient impact of the waves; the reduction in opening lag time means that the time window from receiving the command to the actual movement of the valve core is compressed, so that the damping compensation response speed matches the rapid changes of the distorted wave.

[0055] In this embodiment, the control strategy of the proportional relief valve can be adjusted in real time according to sea conditions. When the wind and waves are calm, the dead zone is reduced to improve compensation accuracy, and when the waves are violently disturbed, the dead zone is expanded to filter sensor noise and load jitter, avoiding mechanical wear of the proportional relief valve due to over-response. It solves the contradiction that the dead zone anti-jitter strategy in the prior art causes compensation response lag under extreme sea conditions. When the jolt exceeds the abnormal wave warning threshold, the emergency dead zone wide-range forced compression of the dead zone range forces the proportional relief valve to release high-frequency control commands, effectively shortening the response delay of damping compensation under extreme wave conditions, and preventing the compensation wire rope from suffering a violent secondary breakage impact due to insufficient damping force adjustment under abnormal wave impact.

[0056] In some embodiments, the method further includes: The spectral energy distribution characteristics reflecting wave periodicity are obtained from the dimensionality-reduced temporal feature vector. Specifically, a discrete Fourier transform is performed along the time dimension on the dimensionality-reduced temporal feature vector to extract the energy amplitude of each frequency component, which constitutes the spectral energy distribution characteristics. The spectral energy distribution characteristics are a one-dimensional array indexed by frequency, where each element corresponds to the energy intensity of a specific frequency component. This can be understood as follows: if the energy is concentrated in the low-frequency band of the spectrum, it indicates that the current sea state is dominated by slow swells; if there are obvious peaks in the high-frequency band of the spectrum, it indicates the presence of more active wind and wave components.

[0057] The spectral energy distribution characteristics are input into the displacement transfer function model to predict the target displacement compensation amount of the wave compensation device in the next control cycle. Specifically, the displacement transfer function model takes the spectral energy distribution characteristics as input and predicts the displacement compensation amount that the hydraulic cylinder needs to output in the next control cycle through pre-trained linear or nonlinear mapping relationships. This model reflects the physical correspondence between wave spectral characteristics and hydraulic cylinder displacement response: when the spectral energy is concentrated in a certain frequency band, the corresponding wave period can be pre-calculated, and the displacement transfer function model predicts the amplitude and direction of the compensation displacement accordingly. The target displacement compensation amount mentioned here refers to the expected piston displacement amount that the hydraulic cylinder needs to output to offset the heave displacement in the current wave cycle.

[0058] The target displacement compensation amount is converted into a feedforward control current for the proportional relief valve. Specifically, based on the flow-pressure characteristic curve of the proportional relief valve, the target displacement compensation amount is mapped to the corresponding drive current value; this feedforward control current is input as a feedforward control signal before the proportional relief valve receives the feedback control command, so that the valve core has moved to a position close to the target before the feedback command arrives, thus shortening the overall response time.

[0059] When the current tension value is less than the relaxation threshold, the weight of the feedforward control current in the superposition calculation is increased to reduce the tension gap of the compensation wire rope. Specifically, when the measured tension of the compensation wire rope is detected to be lower than the relaxation threshold, the weight coefficient of the feedforward channel is increased in the superposition calculation of the feedforward control current and the feedback control current. The purpose of increasing the feedforward weight is to utilize the predictive nature of feedforward control to increase the opening of the proportional relief valve in advance, so that the hydraulic cylinder can output the compensation force more quickly to pull the compensation wire rope, thereby reducing the time gap between relaxation and re-tension of the wire rope.

[0060] When the rate of change of rope tension exceeds a preset transient impact threshold, a scaling factor is calculated based on the rate of change of rope tension. This scaling factor is then used to smooth and filter the feedforward control current, suppressing secondary impacts during the compensation process. Specifically, the transient impact threshold refers to the critical rate of change of tension required to determine if the compensation wire rope is subjected to a sudden external force impact. This threshold can be taken as the 75th percentile of the rope tension change rate in the training set. When the rate of change of rope tension exceeds this threshold, the filtered rate of change of the feedforward control current is suppressed, preventing secondary impacts from occurring in the proportional relief valve due to sudden changes in feedforward commands.

[0061] The weighted or filtered feedforward control current is added to the control command to generate a composite control signal, which is then output to the proportional relief valve. Specifically, a composite control signal is formed that simultaneously contains feedback correction information and feedforward prediction information; after this composite signal is sent to the proportional relief valve, the valve core responds to the predicted displacement demand from the feedforward and the real-time damping correction from the feedback.

[0062] In one alternative, the feedforward control current can also be output independently of the feedback control command: the feedforward component is only superimposed when tension relaxation or transient impact conditions are detected, and only the feedback control command is output during other periods to reduce the complexity of the control logic.

[0063] In this embodiment, the periodic characteristics of waves are extracted by the spectral energy distribution features and the target displacement compensation amount for the next control cycle is predicted by the displacement transfer function model, thus realizing the feedforward prediction of wave compensation displacement. When the compensation wire rope enters a slack state, the weight of the feedforward channel is increased, causing the proportional relief valve to respond in advance and reducing the tension gap. When a sudden change in rope tension is detected, the proportional scaling factor smooths and filters the feedforward current, avoiding the abrupt transmission of the feedforward command to the proportional relief valve and causing a secondary impact.

[0064] In some embodiments, before outputting control commands to the proportional relief valve of the wave compensation device, the following is also included: A training dataset containing real wave spectrum data and corresponding hydraulic cylinder displacement labels is obtained, and transient feature maps from the temporal feature extraction network are used as pre-input features for the displacement transfer function model to construct a bidirectional feature transfer path. Specifically, the training dataset contains N sets of samples, each set including real recorded wave spectrum data and corresponding hydraulic cylinder displacement labels; the displacement transfer function model uses transient feature maps as pre-input features, rather than only using spectral energy distribution features as input; the bidirectional feature transfer path refers to the following: the output error of the displacement transfer function model is backpropagated back through gradients to the one-dimensional convolutional layer of the temporal feature extraction network, and the transient feature map of the temporal feature extraction network is forward-propagated to the displacement transfer function model as input features. The two temporal feature extraction networks jointly update parameters during the training phase; this can be understood as the bidirectional transfer path transferring the error signal in the displacement prediction task to the convolutional layer, enabling the convolutional kernel to simultaneously learn damping adjustment features and displacement prediction features during training.

[0065] The mean square error between the predicted value of the target displacement compensation and the hydraulic cylinder displacement label is calculated as the initial displacement prediction error. Specifically, the output of the displacement transfer function model is compared with the hydraulic cylinder displacement labels labeled in the training dataset, and the initial displacement prediction error is calculated.

[0066] Obtain the maximum stroke boundary of the hydraulic cylinder and the maximum drive current boundary of the proportional relief valve. Specifically, the maximum stroke boundary of the hydraulic cylinder is determined by the structure of the hydraulic cylinder, referring to the maximum displacement limit of the piston rod within the cylinder barrel; the maximum drive current boundary of the proportional relief valve is determined by the rated output capacity of the power drive circuit, referring to the maximum drive current value corresponding to the valve core being fully open; both of these boundary parameters can be found in the equipment's factory parameter table or obtained through actual measurement and calibration.

[0067] The standard displacement loss is generated by dividing the initial displacement prediction error using the maximum travel boundary, and the dimensionless standard damping loss is generated by dividing the loss function using the maximum driving current boundary.

[0068] Obtain the current rate of change of rope tension in the training dataset, and when the current rate of change of rope tension increases, proportionally increase the first weight assigned to the standard damping loss, and correspondingly decrease the second weight assigned to the standard displacement loss, wherein the sum of the first weight and the second weight is one.

[0069] The standard displacement loss is multiplied by the second weight, and the standard damping loss is multiplied by the first weight. The sum of the two is calculated to generate a joint loss function. The joint loss function is then backpropagated to the one-dimensional convolutional layer of the temporal feature extraction network through a bidirectional feature propagation path for gradient calculation and weight update.

[0070] In some embodiments, the parameter optimization process of the displacement transfer function model includes: A training dataset containing real wave spectrum data and corresponding hydraulic cylinder displacement labels is obtained, and transient feature maps from the temporal feature extraction network are used as pre-input features for the displacement transfer function model to construct a bidirectional feature transfer path. The mean square error between the predicted value of the target displacement compensation and the hydraulic cylinder displacement label is used as the initial displacement prediction error. Obtain the maximum stroke boundary of the hydraulic cylinder and the maximum drive current boundary of the proportional relief valve; The standard displacement loss is generated by dividing the initial displacement prediction error using the maximum travel boundary, and the dimensionless standard damping loss is generated by dividing the loss function using the maximum drive current boundary. Obtain the current rate of change of rope tension in the training dataset, and when the current rate of change of rope tension increases, proportionally increase the first weight assigned to the standard damping loss, and correspondingly decrease the second weight assigned to the standard displacement loss, wherein the sum of the first weight and the second weight is one. Multiply the standard displacement loss by the second weight and the standard damping loss by the first weight, and calculate the sum of the two to generate a joint loss function; The joint loss function is backpropagated to the one-dimensional convolutional layer of the temporal feature extraction network through a bidirectional feature propagation path for gradient calculation and weight update.

[0071] In some embodiments, the training process of the time-series feature extraction network includes: acquiring the basic heave acceleration sequence and the basic hydraulic cylinder load sequence under historical normal working conditions; applying a random time offset that follows a uniform distribution to each sampling point timestamp of the basic hydraulic cylinder load sequence to generate a perturbed load sequence with phase noise; combining the basic heave acceleration sequence and the perturbed load sequence to construct a training set for iterative parameter updates of the time-series feature extraction network.

[0072] In some embodiments, acquiring the heave acceleration sequence of the wave compensation device and the load sequence of the hydraulic cylinder within a preset time window includes: acquiring the analog voltage signal output by the first acceleration sensor installed on the base of the wave compensation device; performing analog-to-digital conversion and low-pass filtering on the analog voltage signal to generate a discrete heave acceleration sequence; acquiring the real-time oil pressure value output by the pressure transmitter installed in the rodless chamber oil circuit of the hydraulic cylinder; calculating the instantaneous load force based on the real-time oil pressure value, and recording the instantaneous load force according to a preset sampling frequency to generate the load sequence of the hydraulic cylinder.

[0073] In some embodiments, the dead zone width is determined by the following process: obtaining the load sequence of the hydraulic cylinder within a preset historical time window from the current time; calculating the sliding variance of the load sequence of the hydraulic cylinder; multiplying the sliding variance by a preset scaling factor to generate the dead zone width; the dead zone width increases proportionally with the increase of the intensity of the load sequence fluctuation of the hydraulic cylinder.

[0074] In the description of this application, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of technical features indicated. Therefore, a feature defined as "first" or "second" may explicitly or implicitly include one or more features. In the description of this application, "multiple" means two or more, unless otherwise explicitly specified.

[0075] In the above embodiments, the descriptions of each embodiment have different focuses. For parts not described in detail in a certain embodiment, please refer to the relevant descriptions in other embodiments.

[0076] The embodiments, implementation methods, and related technical features of this application can be combined and substituted for each other without conflict.

[0077] The above are merely preferred embodiments of this application and are not intended to limit this application in any way. Any simple modifications, equivalent changes, and alterations made to the above embodiments based on the technical essence of this application without departing from the scope of the technical solution of this application shall still fall within the scope of the technical solution of this application.

Claims

1. A damping adjustment method for a wave compensation device, characterized in that, include: The heave acceleration sequence and hydraulic cylinder load sequence of the wave compensation device are collected within a preset time window; the heave acceleration sequence reflects the vertical motion state of the platform on which the wave compensation device is located; the hydraulic cylinder load sequence reflects the current load state of the wave compensation device. The heave acceleration sequence and the hydraulic cylinder load sequence are aligned and merged according to timestamps to generate joint state features; The joint state features are input into a pre-trained temporal feature extraction network to obtain the damping adjustment value for the current load state; The loss function of the temporal feature extraction network includes a penalty term for the rate of change of rope tension in the compensation wire rope of the wave compensation device; The penalty term increases the weight of loss feedback when the rate of change of rope tension exceeds a preset tension fluctuation threshold. The control command is output to the proportional relief valve of the wave compensation device according to the damping adjustment value, thereby adjusting the damping force of the hydraulic cylinder.

2. The method according to claim 1, characterized in that, The heave acceleration sequence and the hydraulic cylinder load sequence are aligned and merged according to timestamps to generate joint state features, including: Obtain the first device timestamp of each sampling point in the heave acceleration sequence; Obtain the second device timestamp for each sampling point in the load sequence of the hydraulic cylinder; Using the timestamp of the first device as a reference, the load sequence of the hydraulic cylinder is resampled using a linear interpolation algorithm to obtain an aligned load sequence with the same length as the heave acceleration sequence. The heave acceleration sequence is combined with the aligned load sequence to generate a joint state feature with time step and feature channel dimension.

3. The method according to claim 2, characterized in that, The joint state features are input into a pre-trained temporal feature extraction network, including: The joint state features are input into the parallel multi-scale one-dimensional convolutional layer of the temporal feature extraction network; the parallel multi-scale one-dimensional convolutional layer contains a first short-window convolutional kernel for transient wave impacts and a second long-window convolutional kernel for continuous ship heave. The joint state features are extracted by using the first short-window convolution kernel and the second long-window convolution kernel in time sequence to generate transient feature maps and trend feature maps. The transient feature map and the trend feature map are combined using multi-channel features to generate a fused feature map; Obtain the inherent response delay time of the hydraulic cylinder; the inherent response delay time reflects the time difference from receiving the control command to the hydraulic cylinder outputting the damping force. The inherent response delay time is multiplied by the preset sampling frequency and the product result is rounded to generate an integer pooling step size. The fused feature map is input into the max pooling layer of the temporal feature extraction network. A sliding window is used to extract the saliency peak features in the time dimension of the fused feature map with an integer pooling step size. The saliency peak features reflect the wave impact intensity under the current sea state.

4. The method according to claim 3, characterized in that, Obtain the damping adjustment value for the current load condition, including: The reduced time series feature vector is split into a first sub-vector reflecting steady-state load characteristics and a second sub-vector reflecting transient impact characteristics; The first sub-vector is input into the first linear mapping branch of the fully connected layer at the end of the temporal feature extraction network to calculate the predicted value of the basic damping opening. The second sub-vector is input to the second nonlinear mapping branch of the tail-end fully connected layer, and the mapping result is truncated using an activation function to output the damping compensation coefficient for the impact energy. Multiply the predicted basic damping opening value by the damping compensation coefficient, and output the damping adjustment value that reflects the expected damping opening value. Specifically, the dimensionality-reduced time-series feature vector is split into a first sub-vector reflecting steady-state load characteristics and a second sub-vector reflecting transient impact characteristics, including: Obtain the feature channel combination boundary formed when performing multi-channel feature combination; the feature channel combination boundary defines the data boundary between the output features of the first short-window convolution kernel and the second long-window convolution kernel; Based on the feature channel combination boundary, the reduced temporal feature vector is processed by vector slicing. Extract the slice data belonging to the channel dimension corresponding to the second long window convolution kernel into the first sub-vector; Extract the slice data belonging to the channel dimension corresponding to the first short-window convolution kernel into the second sub-vector.

5. The method according to claim 4, characterized in that, The loss function of the temporal feature extraction network is constructed through the following process: During the backpropagation training phase of the temporal feature extraction network, the damping adjustment prediction sequence of the current batch is obtained, and the mean square error between the damping adjustment prediction sequence and the preset label is calculated as the basic mapping loss. Obtain the measured tension sequence of the compensated wire rope aligned with the damping adjustment prediction sequence from the training set; The measured tension sequence records the actual stress state under the corresponding sea conditions; The difference calculation is performed on adjacent sampling points of the measured tension sequence and the absolute value of the difference result is obtained to generate the rope tension change rate representing the instantaneous tension change. The rate of change of rope tension is multiplied by a pre-set penalty coefficient to construct a penalty term used to constrain the mechanical fatigue boundary; The loss function is generated by adding the basic mapping loss to the penalty term.

6. The method according to claim 5, characterized in that, The penalty term used to constrain the mechanical fatigue boundary is constructed by multiplying the rate of change of rope tension by a pre-set penalty coefficient, including: Determine whether the current tension value in the measured tension sequence is less than the preset relaxation threshold; the relaxation threshold reflects the boundary stress value at which the compensating wire rope is about to lose its tensile state. When the current tension value is less than the relaxation threshold, the pre-set penalty coefficient is multiplied by the pre-set exponential amplification factor to generate an updated penalty coefficient; The penalty term is generated by multiplying the rate of change of rope tension by the updated penalty coefficient.

7. The method according to claim 6, characterized in that, The proportional relief valve of the wave compensation device outputs control commands according to the damping adjustment value, including: Obtain the current actual valve core opening of the proportional relief valve; Calculate the absolute difference between the damping adjustment value and the actual valve core opening; Determine whether the absolute difference is greater than the preset dead zone width; When the absolute difference is determined to be greater than the dead zone width, the damping adjustment value is converted into a drive current pulse and output as a control command to the proportional relief valve.

8. The method according to claim 7, characterized in that, The method also includes: The instantaneous jump value is obtained by performing the first derivative on the data within the current time window in the heave acceleration sequence; Construct monitoring boundary conditions for instantaneous sag values ​​to determine whether the instantaneous sag values ​​are greater than the preset abnormal wave warning threshold; When the instantaneous sag value exceeds the abnormal wave warning threshold, the dead zone over-limit response is initiated; Based on the dead zone over-limit response, the dead zone width is forcibly reduced by multiplying it by a preset attenuation coefficient to generate an emergency dead zone width; The emergency dead zone width is replaced by the absolute difference value of the proportional relief valve to release control commands corresponding to transient wave characteristics and reduce the opening lag time of the proportional relief valve.

9. The method according to claim 8, characterized in that, Before the proportional relief valve of the wave compensation device outputs the control command, it also includes: Extract the spectral energy distribution characteristics that reflect the wave cycle from the dimensionality-reduced temporal feature vector; The spectral energy distribution characteristics are input into the displacement transfer function model to predict the target displacement compensation amount of the wave compensation device in the next control cycle. Convert the target displacement compensation amount into a feedforward control current for the proportional relief valve; When the current tension value is less than the relaxation threshold, the weight of the feedforward control current in the superposition calculation is increased to reduce the gap of the wire rope to compensate for tension loss. When the rate of change of rope tension is greater than the preset transient impact threshold, the proportional scaling factor is calculated based on the rate of change of rope tension, and the feedforward control current is smoothed and filtered by the proportional scaling factor to suppress the secondary impact during the compensation process. The weighted or filtered feedforward control current is added to the control command to generate a composite control signal, which is then output to the proportional relief valve.

10. The method according to claim 9, characterized in that, The parameter optimization process of the displacement transfer function model includes: A training dataset containing real wave spectrum data and corresponding hydraulic cylinder displacement labels is obtained, and transient feature maps from the temporal feature extraction network are used as pre-input features for the displacement transfer function model to construct a bidirectional feature transfer path. The mean square error between the predicted value of the target displacement compensation and the hydraulic cylinder displacement label is used as the initial displacement prediction error. Obtain the maximum stroke boundary of the hydraulic cylinder and the maximum drive current boundary of the proportional relief valve; The standard displacement loss is generated by dividing the initial displacement prediction error using the maximum travel boundary, and the dimensionless standard damping loss is generated by dividing the loss function using the maximum drive current boundary. Obtain the current rate of change of rope tension in the training dataset, and when the current rate of change of rope tension increases, proportionally increase the first weight assigned to the standard damping loss, and correspondingly decrease the second weight assigned to the standard displacement loss, wherein the sum of the first weight and the second weight is one. Multiply the standard displacement loss by the second weight and the standard damping loss by the first weight, and calculate the sum of the two to generate a joint loss function; The joint loss function is backpropagated to the one-dimensional convolutional layer of the temporal feature extraction network through a bidirectional feature propagation path for gradient calculation and weight update.

Images