A control method for automatically adjusting the working depth of a post puller

Abstract

The application provides a control method for adaptive adjustment of working depth of a puller, and belongs to the technical field of control of adaptive adjustment of working depth of a puller. The application eliminates vibration interference through an adaptive notch filter, extracts a purified depth signal through empirical mode decomposition, identifies parameters of a modified Cambridge clay model in real time through an online recursive least square method combined with an extended Kalman filter, predicts multi-step hydraulic cylinder target speed through a soil quality mutation-resistance prediction fusion model, adaptively adjusts a deviation ratio driving control step length, and compensates for nonlinearity of a proportional valve through numerical inverse mapping of a Bouc-Wen hysteresis model combined with an artificial intelligence fusion model, so as to realize adaptive adjustment of working depth of the puller and solve the technical problem of low adjustment accuracy caused by mismatch between the adjustment control step length and the change in soil layer resistance of the puller under complex soil working conditions.

Description

Technical Field

[0001] This invention belongs to the field of control technology for adaptive adjustment of the working depth of a lever puller, and more specifically, relates to a control method for adaptive adjustment of the working depth of a lever puller. Background Technology

[0002] Pole extractors are widely used in engineering fields such as pile foundation construction, underground pipeline laying, and geological exploration. Their core function is to extract poles from the ground using a hydraulically driven vibratory hammer. In existing technologies, depth adjustment typically relies on a hydraulic proportional valve with a fixed step size, combined with a preset pressure threshold for manual or semi-automatic adjustment. Some advanced systems introduce closed-loop feedback based on a PID controller, using the hydraulic cylinder pressure signal as feedback to drive the proportional valve. These systems are suitable for homogeneous soil layers or conditions with slow soil changes and have been validated in engineering projects under standard geological conditions such as sand and soft clay.

[0003] However, existing fixed-step control methods have significant drawbacks in complex soil conditions such as gravel interlayers and alternating distributions of hard and soft clay. The broadband mechanical vibration generated by the vibratory hammer during operation is superimposed on the signals from the displacement and pressure sensors, distorting the feedback signals and causing the controller's estimation of soil resistance to deviate significantly from the true value. Furthermore, the fixed control step size cannot adapt to abrupt changes in soil resistance; when the soil conditions change rapidly, the controller response lags, leading to hydraulic cylinder overpressure or depth overshoot.

[0004] In existing technologies, the fixed step-size control law lacks the ability to perceive sudden changes in soil resistance in real time, and the sensor signals cannot accurately reflect the true depth state after being disturbed by vibration. This results in a serious mismatch between the control step size of the derrick for depth adjustment and the actual changes in soil resistance under complex soil conditions. In other words, existing technologies suffer from the technical problem of low adjustment accuracy due to the mismatch between the control step size of the derrick for depth adjustment and the changes in soil resistance under complex soil conditions. Summary of the Invention

[0005] In view of this, the present invention provides a control method for adaptive adjustment of the working depth of a bar puller, which can solve the technical problem in the prior art where the control step size for adjusting the working depth of the bar puller is mismatched with the changes in soil resistance under complex soil conditions, resulting in low adjustment accuracy.

[0006] This invention is implemented as follows: This invention provides a control method for adaptive adjustment of the working depth of a lever puller, comprising the following steps:

[0007] A static cone penetration tester was used to conduct penetration tests on the work area, and the cone tip resistance sequence and sidewall friction force sequence were collected to obtain static cone penetration soil profile data; a hydraulic pressure sensor was used to collect hydraulic cylinder pressure sequence, a displacement sensor was used to collect displacement sensor sequence, a six-axis torque sensor was used to collect six-axis torque sequence, and a vibration acceleration sensor was used to collect vibration acceleration sequence.

[0008] The vibration acceleration sequence is input into an adaptive notch filter to estimate the fundamental frequency of the vibratory hammer in real time, thereby eliminating vibration interference components in the hydraulic cylinder pressure sequence and displacement sensor sequence. The displacement sensor sequence after vibration interference elimination is then input into empirical mode decomposition processing to extract the sum of intrinsic mode functions and obtain the purification depth signal.

[0009] Using purification depth signal, hydraulic cylinder pressure sequence, six-axis torque sequence and static cone penetration soil profile data as input, the parameters of the modified Cambridge clay model are identified online by online recursive least squares method combined with extended Kalman filter, and the output cohesion, internal friction angle, overconsolidation ratio and critical stress ratio are updated by weighted confidence weight matrix.

[0010] The hydraulic cylinder pressure sequence, displacement sensor sequence, six-axis torque sequence, vibration acceleration sequence, and static cone penetration soil profile data are input into the soil abrupt change-resistance prediction fusion model, which outputs the hydraulic cylinder target velocity sequence for multiple future control steps. At the same time, the inference batch size and CUDA flow number of the soil abrupt change-resistance prediction fusion model are dynamically adjusted using a comprehensive adjustment quality index.

[0011] The current soil volumetric strain is calculated using the measured force of the hydraulic cylinder and the purification depth signal. The current effective confining pressure is inferred by modifying the Cambridge clay model flow law. The predicted resistance is calculated by combining the overconsolidation ratio and the critical stress ratio. The deviation ratio is obtained by comparing the predicted resistance with the measured force of the hydraulic cylinder. When the deviation ratio exceeds the step size halving threshold, the current control step size is halved and the soil constitutive parameters are re-identified. When the deviation ratio is lower than the step size doubling threshold, the current control step size is doubled. The target speed of the hydraulic cylinder for the current control step is output based on the adaptive step size and the target speed sequence of the hydraulic cylinder.

[0012] Using the numerical inverse mapping of the Bouc-Wen hysteresis model of the proportional valve, the desired input current is calculated for the desired valve core displacement corresponding to the target speed of the hydraulic cylinder. The desired input current is superimposed with the control current output by the adaptive back-propagation control law to drive the proportional valve to perform deep adjustment, completing one control cycle; then return to the aforementioned online identification step to execute the next control cycle.

[0013] Specifically, the adaptive notch filter uses a phase-locked loop to estimate the fundamental frequency of the vibrating hammer in real time from the vibration acceleration sequence, and places notch traps at the fundamental frequency of the vibrating hammer and its integer multiples of frequency. The center frequency of the notch traps is continuously updated by tracking the fundamental frequency of the vibrating hammer.

[0014] The elimination of vibration interference components in the hydraulic cylinder pressure sequence and displacement sensor sequence eliminates interference frequencies ranging from 10 to 500 Hz; the empirical mode decomposition process extracts intrinsic mode functions with frequencies lower than the excitation frequency band of the vibratory hammer.

[0015] Specifically, the online recursive least squares method combined with extended Kalman filtering is used to identify the parameters of the modified Cambridge clay model online. Specifically, the extended Kalman filtering is used to linearize the nonlinear observation equations of the modified Cambridge clay model using a first-order Taylor expansion, and the online recursive least squares method uses a forgetting factor to weight the historical residuals. The two methods are used alternately to update the parameter estimates and error covariance.

[0016] The elements of the confidence weight matrix range from 0.1 to 1.0 and are assigned values ​​after normalization of the signal-to-noise ratio (SNR) estimates of each sensor. The SNR estimates are calculated as the ratio of the signal variance to the noise variance within a continuous sampling period.

[0017] The backbone of the soil mutation-resistance prediction fusion model is a causal dilated convolutional network. Each dilated convolutional block is followed by a physical residual module. The physical residual module embeds the hydraulic power conservation equation and the soil Mohr-Coulomb failure criterion into the residual bypass in the form of differentiable tensor operations.

[0018] The soil mutation-resistance prediction fusion model includes a soil mutation detection subnet that outputs the probability of soil type change. When the probability of soil type change exceeds the mutation detection threshold, an internal state reset gate is triggered to clear the hidden state of the gated loop unit in the backbone network.

[0019] The output layer of the soil mutation-resistance prediction fusion model is configured with multiple parallel prediction heads. The parallel prediction heads are weighted and fused with confidence using Softmax probability weights to obtain the target velocity sequence of the hydraulic cylinder.

[0020] The calculation of the comprehensive adjustment quality index uses the current deviation ratio, the standard deviation of the purification depth signal, and the fundamental frequency of the vibratory hammer as input components. Each component is divided by its corresponding benchmark value and then weighted and summed. The sum of the three weights is 1. The comprehensive adjustment quality index determines the inference batch size, the number of CUDA streams, and the order of the intrinsic mode functions retained by empirical mode decomposition.

[0021] The Bouc-Wen hysteresis model uses an auxiliary hysteresis state variable to describe the nonlinear memory effect between the proportional valve spool displacement and the input current. The evolution equation of the auxiliary hysteresis state variable includes a linear recovery term and a nonlinear saturation term. The parameters of the Bouc-Wen hysteresis model are identified by driving the proportional valve with sinusoidal frequency sweep excitation and recording the spool displacement response. The mean square error between the model output and the measured spool displacement is minimized using a particle swarm optimization algorithm.

[0022] The adaptive back-pushing control law decomposes the hydraulic cylinder displacement tracking error system into multiple subsystems step by step, designs virtual control quantities for each subsystem, verifies the convergence of each subsystem using the Lyapunov stability theorem, and obtains the control current by combining the virtual control quantities at each level.

[0023] The step size reduction threshold is 15%, which is determined by conducting a soil abrupt change experiment under gravel interlayer conditions and statistically controlling the minimum deviation ratio when mismatch occurs; the step size doubling threshold is 3%, which is determined by conducting a steady-state experiment under homogeneous clay conditions and statistically controlling the maximum residual deviation ratio when the predicted resistance fully converges.

[0024] The training loss function of the soil mutation-resistance prediction fusion model consists of a weighted sum of three parts: the mean square error of hydraulic cylinder target speed prediction, the penalty term for physical residual violation, and the cross-entropy term for soil type classification. The weights of the three parts are 0.6, 0.3, and 0.1, respectively.

[0025] Specifically, when the overall adjustment quality index is greater than or equal to the upper quality threshold, the inference batch size is adjusted to 1, the number of CUDA streams is adjusted to 1, and the empirical mode decomposition retains only the first 3 intrinsic mode functions; when the overall adjustment quality index is between the lower and upper quality thresholds, the inference batch size is adjusted to 2, the number of CUDA streams is adjusted to 2, and the empirical mode decomposition retains the first 5 intrinsic mode functions; when the overall adjustment quality index is lower than the lower quality threshold, the inference batch size is adjusted to 4, the number of CUDA streams is adjusted to 4, and the empirical mode decomposition performs the full decomposition level.

[0026] Wherein, the upper quality threshold is 1.2, the lower quality threshold is 0.8, the mutation detection threshold is 0.7, and the weighting coefficient of the overall quality index is adjusted. The value range is 0.4 to 0.6, and the weighting coefficient is... The value range is 0.2 to 0.4, and the weighting coefficient is... The value range is 0.1 to 0.3, with typical values ​​being... , , .

[0027] This invention obtains the purification depth signal through adaptive notch filtering and empirical mode decomposition, uses online recursive least squares method combined with extended Kalman filtering to identify the parameters of the modified Cambridge clay model in real time, and uses an adaptive adjustment mechanism driven by the deviation ratio to control the step size. This solves the technical problem of low adjustment accuracy caused by the mismatch between the control step size and the change of soil resistance in the working depth adjustment of the auger under complex soil conditions.

[0028] The adaptive notch filter eliminates vibration interference from contaminating the sensor signal, ensuring that the depth signal accurately reflects the position of the member. Online soil constitutive parameter identification enables the controller to perceive the changing trend of soil resistance in real time. The deviation ratio mechanism converts the degree of deviation between the predicted resistance and the measured force into a step size adjustment command. When the deviation ratio exceeds the upper threshold, the step size is shortened to quickly respond to sudden changes in soil conditions. When the deviation ratio is below the lower threshold, the step size is increased to improve the operating efficiency under stable conditions, thus ensuring that the control step size always matches the current rate of change of soil resistance.

[0029] In summary, the present invention solves the technical problem mentioned in the background art of low adjustment accuracy caused by the mismatch between the control step size of the operating depth adjustment of the lever puller and the change of soil resistance under complex soil conditions. Attached Figure Description

[0030] Figure 1 This is a flowchart of the method of the present invention.

[0031] Figure 2 This is a depth-resistance distribution diagram of soil profile data from static cone penetration testing.

[0032] Figure 3 This is a comparison chart of displacement sensor sequences before and after purification.

[0033] Figure 4 This is a graph showing the change in soil type probability with depth.

[0034] Figure 5 The graph shows the variation of deviation ratio and control step size over operation time.

[0035] Figure 6 This is a graph showing the change in hydraulic cylinder depth adjustment error over operating time. Detailed Implementation

[0036] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below.

[0037] like Figure 1 The diagram shown is a flowchart of a control method for adaptive adjustment of the working depth of a lever puller provided by the present invention. This method includes the following steps:

[0038] S01. A static cone penetrator is used to conduct a penetration test on the work area, and the cone tip resistance sequence and sidewall friction force sequence are collected to obtain static cone penetration soil profile data; a hydraulic pressure sensor is used to collect the hydraulic cylinder pressure sequence, a displacement sensor is used to collect the displacement sensor sequence, a six-axis torque sensor is used to collect the six-axis torque sequence, and a vibration acceleration sensor is used to collect the vibration acceleration sequence.

[0039] S02. Input the vibration acceleration sequence into an adaptive notch filter to estimate the fundamental frequency of the vibratory hammer in real time. Set notch traps at integer multiples of the fundamental frequency of the vibratory hammer to eliminate vibration interference components in the hydraulic cylinder pressure sequence and displacement sensor sequence. Input the displacement sensor sequence after eliminating vibration interference into empirical mode decomposition processing to extract the sum of eigenmode functions with frequencies lower than the excitation frequency band of the vibratory hammer, and obtain the purification depth signal.

[0040] S03. Using the purification depth signal, hydraulic cylinder pressure sequence, six-axis torque sequence and static cone penetration soil profile data as input, the parameters of the modified Cambridge clay model are identified online by combining online recursive least squares method with extended Kalman filter. The outputs cohesion, internal friction angle, overconsolidation ratio and critical stress ratio are then updated with the above four soil constitutive parameters by weighted confidence weight matrix.

[0041] S04. Input the hydraulic cylinder pressure sequence, displacement sensor sequence, six-axis torque sequence, vibration acceleration sequence and static penetration soil profile data into the soil abrupt change-resistance prediction fusion model, output the hydraulic cylinder target velocity sequence for the next 3 control steps, and dynamically adjust the inference batch size and CUDA flow quantity of the soil abrupt change-resistance prediction fusion model with the comprehensive adjustment quality index.

[0042] S05. Calculate the current soil volumetric strain using the measured force of the hydraulic cylinder and the purification depth signal. Inversely deduce the current effective confining pressure by modifying the Cambridge clay model flow law. Calculate the predicted resistance by combining the overconsolidation ratio and the critical stress ratio. Obtain the deviation ratio by comparing the predicted resistance with the measured force of the hydraulic cylinder. When the deviation ratio exceeds 15%, halve the current control step size and trigger the re-identification of soil constitutive parameters. When the deviation ratio is below 3%, double the current control step size. Output the target hydraulic cylinder velocity for the current control step based on the adaptive step size and the target hydraulic cylinder velocity sequence.

[0043] S06. Using the numerical inverse mapping of the Bouc-Wen hysteresis model of the proportional valve, the desired input current is calculated for the desired valve core displacement corresponding to the target speed of the hydraulic cylinder. The desired input current is superimposed with the control current output by the adaptive back-propagation control law to drive the proportional valve to perform deep adjustment, completing one control cycle; return to S03 to execute the next control cycle.

[0044] The adaptive notch filter works as follows: a phase-locked loop estimates the fundamental frequency of the vibratory hammer in real time from the vibration acceleration sequence; notch traps are placed at the fundamental frequency and its integer multiples; the center frequency of the notch traps is continuously updated with the fundamental frequency of the vibratory hammer, thereby accurately eliminating known frequency interference components without introducing broadband phase delay; the interference frequency range eliminated by the adaptive notch filter is 10-500Hz, covering the wide frequency band of mechanical vibration generated during the operation of the auger; the boundary frequency of the vibratory hammer excitation frequency band is determined by statistical analysis of the main peak frequency distribution of the vibration acceleration spectrum after conducting no less than 15 operational experiments under three typical soil types.

[0045] The working principle of the empirical mode decomposition is as follows: the displacement sensor sequence after vibration interference is eliminated is adaptively decomposed, and the intrinsic mode functions arranged from high frequency to low frequency are extracted in sequence. Each intrinsic mode function satisfies the condition of local symmetry and zero mean. The intrinsic mode functions with frequencies lower than the vibration hammer excitation frequency band are summed to obtain the purification depth signal. The number of retained intrinsic mode functions is determined by the comprehensive adjustment quality index, and the specific rules are given in the description of the comprehensive adjustment quality index.

[0046] The working principle of the online recursive least squares method combined with extended Kalman filtering for online identification of parameters of the modified Cambridge clay model is as follows: The hydraulic cylinder pressure sequence, purification depth signal, and six-axis torque sequence are used as observation vectors, and cohesion, internal friction angle, overconsolidation ratio, and critical stress ratio are considered as random walk state variables. Extended Kalman filtering linearizes the nonlinear observation equations of the modified Cambridge clay model using a first-order Taylor expansion. The online recursive least squares method uses a forgetting factor to weight the historical residuals, and the two methods alternately update the parameter estimates and error covariance to achieve rapid tracking of soil abrupt changes. The elements of the confidence weight matrix range from 0.1 to 1.0, and are assigned values ​​after normalization of the signal-to-noise ratio estimates of each sensor. The signal-to-noise ratio estimates are calculated as the ratio of signal variance to noise variance over 100 consecutive sampling periods. The threshold values ​​of the confidence weight matrix elements are determined by conducting at least 20 penetration-extraction experiments on standard sand, clay, and gravel soils, and analyzing the inflection points of the identification error curves.

[0047] The specific structure of the soil mutation-resistance prediction fusion model is as follows: The input layer receives hydraulic cylinder pressure sequences, displacement sensor sequences, six-axis torque sequences, vibration acceleration sequences, and static cone penetration soil profile data, which are then spliced ​​into a multi-channel time-series tensor; the backbone is a causal dilated convolutional network with a kernel size of 3 and dilation steps of 1, 2, 4, 8, and 16, for a total of 5 dilation levels. Each level outputs a channel dimension of 64, and a causal mask ensures that only current and historical information is used at any given time; each dilated convolutional block is followed by a physical residual module, which embeds the hydraulic power conservation equation and the soil Mohr-Coulomb failure criterion into the residual bypass in the form of differentiable tensor operations, and superimposes the physical residuals onto the main feature using a learnable scaling factor; the network internally includes a soil mutation detection subnet, which is a 3-layer one-dimensional convolutional classifier, whose input is the current data of the hydraulic cylinder pressure sequence. The pressure gradient sequence within the front window outputs the soil layer type change probability. When the soil layer type change probability exceeds 0.7, an internal state reset gate is triggered, clearing the hidden state of the gated loop unit in the backbone network. The output layer sets three parallel prediction heads, each corresponding to the target speed of the hydraulic cylinder in the next three control steps. The three parallel prediction heads are weighted and fused with confidence using Softmax probability weights to obtain the hydraulic cylinder target speed sequence. The soil layer type change probability threshold of 0.7 is determined by conducting no less than 30 soil mutation experiments under gravel interlayer conditions and statistically analyzing the intersection of the false trigger rate and the missed detection rate. The resource scheduling mechanism is as follows: the inference batch size and the amount of video memory allocated are linearly proportional, and the number of CUDA streams is adjusted synchronously with the inference batch size. When the inference batch size is 1, a single CUDA stream is allocated, and when the inference batch size is 4, four parallel CUDA streams are allocated.

[0048] The technical principle of the soil mutation-resistance prediction fusion model is as follows: causal dilated convolution captures multi-scale temporal dependencies without increasing the number of parameters by exponentially growing the receptive field; physical residual embedding ensures that network features are continuously constrained by the hydraulic power conservation equation and the Mohr-Coulomb failure criterion during training, avoiding physically unreasonable prediction biases in sparse sample regions by the purely data-driven model; the collaborative mechanism between the soil mutation detection subnet and the internal state reset gate enables the model to actively remove interference from historical soil layer information when soil types change, improving the prediction response speed for new soil layer resistance; and the confidence-weighted fusion of multiple parallel prediction heads reduces the output variance of a single prediction branch due to insufficient local features, achieving an overall unity of physical consistency and data adaptability.

[0049] The specific steps for establishing the training dataset for the soil mutation-resistance prediction fusion model include: conducting no fewer than 50 complete pole-pulling operations in four soil types—standard sand, soft clay, hard clay, and gravel interlayer—and simultaneously recording hydraulic cylinder pressure sequences, displacement sensor sequences, six-axis torque sequences, vibration acceleration sequences, and static cone penetration test soil profile data at a sampling rate of 25kHz. The soil layer boundary positions and soil layer types for each operation are manually labeled. The raw data are processed by adaptive notch filtering and empirical mode decomposition to form purified data samples, which are then divided into training set, validation set, and test set in a 7:2:1 ratio.

[0050] The specific steps for training the soil mutation-resistance prediction fusion model include: using the AdamW optimizer with an initial learning rate of... The weight decay coefficient is The training batch size is 32, and the total number of training rounds is 200. The learning rate is multiplied by 0.5 when the loss on the validation set does not decrease for 10 consecutive rounds. The loss function consists of a weighted sum of three parts: the mean square error of the hydraulic cylinder target speed prediction, the penalty for physical residual violation, and the cross-entropy of soil type classification. The weights of the three parts are 0.6, 0.3, and 0.1, respectively. The weights are determined by a grid search experiment with a step size of 0.05 on the validation set.

[0051] The formula for calculating the comprehensive adjustment quality index is as follows: ;in This is the current deviation ratio. The baseline deviation ratio is calculated from the average deviation ratio of 15 standard operating condition tests. The standard deviation of the current purification depth signal. The standard deviation of the baseline purification depth signal is calculated from the average standard deviation of the purification depth signal from 15 standard operating condition experiments. This is the current fundamental frequency of the vibratory hammer. The rated excitation frequency is read from the equipment nameplate parameters. , , These are weighting coefficients; the sum of the three is 1, and their typical value range is... , , The value was determined by orthogonal experimental design with a step size of 0.05; when the comprehensive adjustment quality index... At that time, the inference batch size is adjusted to 1, the number of CUDA streams is adjusted to 1, and the empirical mode decomposition retains only the first 3 eigenmode functions; when the overall adjustment quality index is adjusted... At that time, the inference batch size was adjusted to 2, the number of CUDA streams was adjusted to 2, and the empirical mode decomposition retained the first 5 eigenmode functions; when the overall adjustment quality index was adjusted... At that time, the inference batch size was adjusted to 4, the number of CUDA streams was adjusted to 4, and the empirical mode decomposition was performed with a complete decomposition level; the boundary values ​​of the comprehensive adjustment quality index interval of 1.2 and 0.8 were determined by joint statistical analysis of control accuracy and computing resource utilization rate through no less than 20 operation experiments in each of the four soil types.

[0052] The Bouc-Wen hysteresis model works as follows: an auxiliary hysteresis state variable describes the nonlinear memory effect between the proportional valve core displacement and the input current. The evolution equation of the auxiliary hysteresis state variable includes a linear recovery term and a nonlinear saturation term. By identifying three types of parameters—linear recovery coefficient, saturation coefficient, and shape coefficient—offline, the hysteresis characteristics with asymmetric gain during forward and reverse motion are fully characterized. The identification method involves driving the proportional valve with a sinusoidal frequency sweep excitation of 1–50 Hz and recording the valve core displacement response. The particle swarm optimization algorithm is used to minimize the mean square error between the Bouc-Wen hysteresis model output and the measured valve core displacement. At least five steady-state cycles are collected at each frequency point, and at least three repeated experiments are conducted to confirm parameter consistency.

[0053] The working principle of the numerical inverse mapping of the Bouc-Wen hysteresis model is as follows: for a given desired valve core displacement, the desired input current that satisfies the evolution equation of the auxiliary hysteresis state variable is solved using the parameters of the Bouc-Wen hysteresis model. The desired input current is then superimposed with the control current output by the adaptive back-propagation control law to jointly drive the proportional valve, so that the error of the actual valve core displacement tracking the desired valve core displacement converges to the residual range of the Bouc-Wen hysteresis model, thus eliminating the influence of dead zone jumps and gain asymmetry on the depth adjustment accuracy.

[0054] The adaptive back-mapping control law works as follows: the hydraulic cylinder displacement tracking error system is decomposed into multiple subsystems step by step, a virtual control quantity is designed for each subsystem, the convergence of each subsystem is verified by Lyapunov stability theorem, and the control current is obtained by combining the virtual control quantities at each level; the adaptive law estimates the uncertain parameters of the system online and updates the control gain in real time, so that the closed-loop system remains asymptotically stable when the soil constitutive parameters change; the control current output by the adaptive back-mapping control law is superimposed with the expected input current output by the numerical inverse mapping of the Bouc-Wen hysteresis model as the final drive current of the proportional valve.

[0055] The threshold for triggering a halving of the step size when the deviation ratio exceeds 15% is determined by statistically analyzing the minimum deviation ratio at which mismatch occurs after conducting no less than 20 soil abrupt change experiments under gravel interlayer conditions; the threshold for triggering a doubling of the step size when the deviation ratio is below 3% is determined by statistically analyzing the maximum residual deviation ratio at which the predicted resistance fully converges after conducting no less than 20 steady-state experiments under homogeneous clay conditions.

[0056] The static cone penetration test (CPPT) soil profile data consists of the distribution data of the cone tip resistance sequence and sidewall friction force sequence with depth obtained by conducting penetration tests on the target area using a static cone penetration tester before operation. This data is used to initialize the overconsolidation ratio and critical stress ratio. The overconsolidation ratio is the ratio of the soil's historical maximum effective consolidation pressure to its current effective consolidation pressure. The critical stress ratio is the ratio of the soil's deviatoric stress to its effective average stress under critical conditions, which depends on the soil's internal friction angle. The modified Cambridge clay model is an elastoplastic constitutive model based on critical state soil mechanics, using the overconsolidation ratio and critical stress ratio as core parameters to describe the soil's deformation and strength characteristics under different stress paths.

[0057] Optionally, the present invention also provides a method for implementing a lever-pulling machine working depth adaptive adjustment system by means of a computer, wherein the computer is provided with a readable storage medium, the readable storage medium stores program instructions, and the program instructions are used to execute the above-described method when the computer is run.

[0058] The specific implementation of step S01 is as follows. Before the pole extraction operation begins, a static cone penetrometer is vertically inserted into the target area, simultaneously collecting data on the distribution of cone tip resistance and sidewall friction force sequences with depth. This data forms static cone penetrometer soil profile data, which is used to initialize and correct the overconsolidation ratio and critical stress ratio of the Cambridge clay model. The overconsolidation ratio is the ratio of the soil's historical maximum effective consolidation pressure to its current effective consolidation pressure, and the critical stress ratio depends on the soil's internal friction angle. During the operation, a hydraulic pressure sensor collects the hydraulic cylinder pressure sequence at a fixed sampling rate, a displacement sensor collects the displacement sequence, a six-axis torque sensor collects the six-axis torque sequence, and a vibration acceleration sensor collects the vibration acceleration sequence. The signals from these four sensors are simultaneously collected and stored in a real-time buffer for subsequent steps.

[0059] The specific implementation of step S02 is as follows. The adaptive notch filter uses a phase-locked loop (PLL) principle to estimate the fundamental frequency of the vibratory hammer in real time from the vibration acceleration sequence. The PLL continuously tracks the dynamic changes of the fundamental frequency through phase error feedback. Notch traps are placed at the fundamental frequency and its integer multiples. The center frequency of the notch traps is continuously updated with the fundamental frequency, eliminating vibration interference components in the 10–500 Hz range in the hydraulic cylinder pressure sequence and displacement sensor sequence, while not introducing broadband phase delay and retaining effective low-frequency information. The displacement sensor sequence after eliminating vibration interference enters empirical mode decomposition (EMD) processing. EMD adaptively decomposes the non-stationary signal into eigenmode functions arranged from high to low frequencies. Each eigenmode function satisfies the condition of local symmetry and zero mean. The eigenmode functions with frequencies lower than the vibratory hammer excitation frequency band are summed to obtain the purification depth signal. The purification depth signal truly reflects the actual depth of the member in the soil layer.

[0060] The specific implementation of step S03 is as follows. The modified Cambridge clay model is based on critical state soil mechanics, using cohesion, internal friction angle, overconsolidation ratio, and critical state stress ratio as core parameters to describe the deformation and strength characteristics of soil under different stress paths. During online identification, the above four parameters are treated as random walk state variables. The extended Kalman filter linearizes the nonlinear observation equations of the model using a first-order Taylor expansion, obtaining an approximate linear expression for the state transition matrix and observation matrix. The online recursive least squares method uses a forgetting factor to exponentially weight historical residuals, making the contribution of recent observation data to parameter estimation greater, thereby quickly responding to soil abrupt changes. The two algorithms alternately update the parameter estimates and error covariance. The elements of the confidence weight matrix are assigned values ​​after normalization of the signal-to-noise ratio estimates of each sensor within 100 consecutive sampling periods, with a value range of 0.1 to 1.0. The identification results of the four soil constitutive parameters are weighted and updated to suppress the identification error introduced by low signal-to-noise ratio sensors.

[0061] The specific implementation of step S04 is as follows. The input layer of the soil mutation-resistance prediction fusion model concatenates the hydraulic cylinder pressure sequence, displacement sensor sequence, six-axis torque sequence, vibration acceleration sequence, and static cone penetration soil profile data into a multi-channel temporal tensor. The expansion step size of the backbone causal dilated convolutional network is 1, 2, 4, 8, and 16, for a total of 5 expansion levels. The causal mask ensures that only current and historical information is used at any given time, achieving multi-scale temporal dependency capture. Each dilated convolutional block is followed by a physical residual module, which embeds the hydraulic power conservation equation and the Mohr-Coulomb failure criterion into the residual bypass in the form of differentiable tensor operations, and superimposes them onto the main feature with a learnable scaling factor to constrain the predictive physical rationality of the network in sparse sample regions. The soil mutation detection subnet consists of 3 layers of one-dimensional convolutional classifiers. The input is the pressure gradient sequence within the current window, and the output is the soil type transformation probability. When the probability exceeds the mutation detection threshold of 0.7, the hidden state of the gated recurrent unit is reset. The output layer uses Softmax probability weights to perform confidence-weighted fusion of multiple parallel prediction heads to obtain the future multi-step target velocity sequence of the hydraulic cylinder. The overall adjustment quality index is then calculated. ,in accordance with The system dynamically adjusts the inference batch size, the number of CUDA streams, and the order of the intrinsic mode functions retained by empirical mode decomposition to achieve a dynamic balance between computational resources and performance tuning.

[0062] The specific implementation of step S05 is as follows: The current soil volumetric strain is calculated using the measured force of the hydraulic cylinder and the purification depth signal. The current effective confining pressure is inferred by modifying the Cambridge clay model flow law. The predicted resistance is calculated by combining the currently identified overconsolidation ratio and critical stress ratio. The ratio of the predicted resistance to the measured force of the hydraulic cylinder is the deviation ratio, which reflects the degree of agreement between the soil resistance prediction model and the actual stress state. When the deviation ratio exceeds 15%, it is determined that the current soil resistance has abruptly changed. The control step size is halved, and the soil constitutive parameters are re-identified to quickly track the new soil state. When the deviation ratio is below 3%, it is determined that the soil resistance prediction has sufficiently converged. The control step size is doubled to improve operational efficiency. The current step size remains unchanged between the two threshold ranges. The target hydraulic cylinder speed for the current control step is selected from the target hydraulic cylinder speed sequence based on the adaptive step size.

[0063] The specific implementation of step S06 is as follows. The Bouc-Wen hysteresis model uses an auxiliary hysteresis state variable to describe the nonlinear memory effect between the proportional valve spool displacement and the input current. The evolution equation of the auxiliary hysteresis state variable includes a linear recovery term and a nonlinear saturation term. The hysteresis characteristics of gain asymmetry are characterized by offline identification of the linear recovery coefficient, saturation coefficient, and shape coefficient. The identification method is to minimize the mean square error between the model output and the measured spool displacement using a particle swarm optimization algorithm. Numerical inverse mapping is used to solve for the desired input current that satisfies the evolution equation for a given desired spool displacement, eliminating the influence of dead zone jumps on the depth adjustment accuracy. The adaptive back-propagation control law decomposes the hydraulic cylinder displacement tracking error system into multiple subsystems step by step. The convergence of each subsystem is guaranteed by the Lyapunov stability theorem. The control current is obtained by integrating the virtual control quantities at each level. After being superimposed with the desired input current, the proportional valve is driven. After completing one control cycle, the process returns to step S03 to execute the next control cycle.

[0064] It should be noted that the key technologies of this invention include: a two-stage signal purification technology of adaptive notch filtering and empirical mode decomposition, which solves the problem of frequency domain separation between broadband vibration interference and low-frequency depth signals by tracking the fundamental frequency of the vibratory hammer through a phase-locked loop and accurately setting notch traps in the frequency domain; empirical mode decomposition further eliminates non-stationary residual disturbances, ensuring the physical reliability of the purified depth signal; a modified Cambridge clay model online identification technology, in which the alternating iterative mechanism of extended Kalman filtering and online recursive least squares method enables soil constitutive parameters to track soil abrupt changes in real time, and the reliability weight matrix suppresses the contamination of identification results by low signal-to-noise ratio sensors; a soil abrupt change-resistance prediction fusion model technology, in which a causal dilated convolutional network captures multi-scale temporal dependencies, a physical residual module is embedded with conservation equation constraints, and a soil abrupt change detection subnet triggers hidden state reset, enabling the model to quickly reconstruct the prediction benchmark when soil changes; and a deviation ratio-driven adaptive step size mechanism, which converts the deviation between predicted resistance and measured force into step size adjustment commands, realizing real-time matching between the control step size and the rate of change of soil resistance. The four key technologies mentioned above work together: the purified signal provides accurate observation for parameter identification, the parameter identification provides a physical basis for the calculation of the deviation ratio, the fusion model provides multi-step prediction support for step size decision-making, and the step size mechanism uniformly drives the resource scheduling of signal acquisition, parameter update and model inference, forming a complete adaptive closed loop.

[0065] It should be noted that when the derrick continuously traverses multiple layers of alternating soil types, the depth adjustment system needs to complete the control switch from the old soil layer state to the new soil layer state in a very short time. When the derrick enters the gravel interlayer from the soft clay layer, the hydraulic cylinder resistance rises sharply within several control cycles. If the control system fails to detect this sudden change in time and maintain the original control gain and step size, it will cause the hydraulic system pressure to instantaneously exceed the limit or the depth to overshoot, which may lead to derrick bending or equipment damage in severe cases. The reason for the above technical problem is that the time scale of the soil resistance change is much smaller than the response time of the traditional fixed step size controller. Traditional methods rely on offline set soil parameters or manual intervention for gain switching, which cannot detect the location and type of unknown soil layer boundaries online. Moreover, vibration interference superimposed on the sensor signal further delays the controller's detection time of the resistance change, making the response lag problem more prominent. The usual solution to the above technical problem is to use model predictive control combined with an offline soil parameter database to pre-plan the control sequence under known geological profile conditions. However, the accuracy of offline databases is limited by exploration density, resulting in significant interpolation errors between exploration points. Furthermore, this method lacks any adaptive capability to real-time soil changes at the construction site; once the actual soil layer deviates from the database, the pre-planned sequence immediately fails, and the control system degenerates into an open-loop state. This invention effectively solves this technical problem. The soil abrupt change detection subnet of the soil abrupt change-resistance prediction fusion model takes the pressure gradient sequence as input. It can output a high-probability transition signal before the pressure gradient anomaly at the soil boundary has fully developed, triggering the resetting of the hidden state of the gated cyclic unit. This allows the model to switch to the new prediction benchmark in the first control cycle after entering a new soil layer, rather than waiting for the historical hidden state to be gradually covered by new data. Simultaneously, the deviation ratio mechanism immediately detects a significant deviation between the predicted resistance and the measured force within the control cycle after the soil abrupt change occurs, triggering a halving of the step size and parameter re-identification. This allows the constitutive parameters of the modified Cambridge clay model to converge to the true values ​​of the new soil layer within several control cycles. The purified depth signal eliminates the contamination of depth feedback by vibration interference, eliminating false resistance abrupt changes caused by vibration in traditional systems. This ensures that the trigger signals of the abrupt change detection subnet and the deviation ratio mechanism come from the actual soil boundary, not vibration noise. The synergistic effect of these three mechanisms ensures that the system's detection of soil abrupt changes, parameter updates, and step size adjustments are tightly linked on the time axis, fundamentally eliminating the response lag of traditional fixed-step-size methods in multi-layered complex soil conditions.

[0066] Specifically, the principle of this invention is:

[0067] The fundamental reason why this invention can solve the above-mentioned technical problems is that it establishes a clear processing path with a physical mechanism for the three mutually coupled links of signal contamination, unknown soil resistance and fixed step size, and integrates the three into an adaptive step size decision framework through the deviation ratio.

[0068] First, vibration interference is the root cause of feedback distortion in existing systems. This invention uses a phase-locked loop to track the fundamental frequency of the vibrating hammer in real time, and sets adaptive notch traps at the fundamental frequency and its integer multiples, thereby accurately eliminating known frequency interference without introducing broadband phase delay. Subsequently, empirical mode decomposition adaptively decomposes the denoised displacement sequence, extracting the sum of eigenmode functions with frequencies lower than the excitation frequency band as the purified depth signal, further eliminating residual non-stationary disturbances. These two stages of processing ensure the physical reliability of the depth feedback signal, providing an accurate observational basis for subsequent parameter identification.

[0069] Secondly, online sensing of soil resistance is a prerequisite for adaptive step size. This invention treats the cohesion, internal friction angle, overconsolidation ratio, and critical stress ratio of the modified Cambridge clay model as random walk state variables. An extended Kalman filter is used to linearize the nonlinear observation equations using a first-order Taylor expansion. Furthermore, the historical residuals are weighted by a forgetting factor in the online recursive least squares method, enabling parameter estimation to rapidly track abrupt changes in soil properties. A confidence weight matrix is ​​used to weight and fuse the identification results based on the signal-to-noise ratio of each sensor, avoiding contamination of parameter estimation by low-quality signals. Thus, the controller can obtain constitutive parameters reflecting the current physical state of the soil layer in each cycle.

[0070] Third, the soil abrupt change-resistance prediction fusion model uses a causal dilated convolutional network as its backbone. It captures multi-scale temporal dependencies through exponentially growing receptive fields. The physical residual module embeds the hydraulic power conservation equation and the Mohr-Coulomb failure criterion into the network using differentiable tensor operations, ensuring that the prediction results still satisfy physical constraints in sparse sample regions. The soil abrupt change detection subnet triggers a hidden state reset when the probability of soil type change exceeds a threshold, eliminating the interference of historical soil information on the prediction of new soil layers and ensuring the physical consistency of the future multi-step hydraulic cylinder target velocity sequence output by the model.

[0071] Fourth, the deviation ratio mechanism compares the predicted resistance derived from the Cambridge clay model with the measured force of the hydraulic cylinder. When the deviation ratio exceeds the upper threshold, the step size is halved and parameter re-identification is triggered; when the deviation ratio is below the lower threshold, the step size is doubled, allowing the control step size to dynamically contract or expand with the rate of change of soil resistance. This mechanism fundamentally solves the problem of mismatch between fixed step size and abrupt changes in soil resistance, achieving a dynamic balance between adjustment accuracy and operational efficiency.

[0072] Finally, the Bouc-Wen hysteresis model numerical inverse mapping compensates for the dead zone jump and gain asymmetry of the proportional valve. The adaptive back-propagation control law uses Lyapunov's stability theorem to ensure the asymptotic stability of the closed-loop system when parameters change. The superimposed output drive current enables the actual valve core displacement to accurately track the expected value, transforming the upper-level adaptive step size decision into a reliable physical execution action.

[0073] The following provides a specific embodiment 1 of the present invention, and the specific implementation of each step in this embodiment 1 is described in detail below.

[0074] The specific implementation of step S01 is as follows: Before operation, a static cone penetrator is used to conduct a vertical penetration test on the target area, and the cone tip resistance sequence is continuously collected along the depth direction. and sidewall friction sequence ,in For the first Each sampling depth point, measured in meters, together constitutes the static cone penetration test soil profile data, used for subsequent initialization of the overconsolidation ratio and critical stress ratio. Simultaneously, hydraulic cylinder pressure sequences are acquired using a hydraulic pressure sensor. The unit is Pa; displacement sequences are acquired using displacement sensors. The unit is meters; six-axis torque sequences are acquired using a six-axis torque sensor. ,in It is a 6-dimensional column vector. , , These represent the force components along the three coordinate axes, in N. , , These represent the torque components along the three coordinate axes, in N·m; vibration acceleration sequences were acquired using a vibration acceleration sensor. The unit is , For the first Each sampling time point has a sampling rate of 25kHz.

[0075] The specific implementation of step S02 is as follows: The vibration acceleration sequence... An adaptive notch filter is input, and the fundamental frequency of the vibratory hammer is estimated in real time using a phase-locked loop. The unit is Hz. Integer multiples of frequency points ( Set a notch trap. The highest harmonic order is obtained by rounding down the ratio of the 500Hz upper limit to the current fundamental frequency. ,in This indicates the floor function (rounding down). Notch transfer function. for:

[0076] ;

[0077] In the formula, Let Z be a complex variable, dimensionless. The sampling frequency is expressed in Hz. Let be the pole radius, dimensionless, with an empirical value of 0.95. The closer the bandwidth is to 1, the narrower the notch bandwidth. The harmonic order is a positive integer. All notch filters are cascaded and applied to... and By eliminating vibration interference components, the pressure sequence after vibration damping is obtained. The unit is Pa, and the displacement sequence after vibration damping. The unit is meters (m). Input empirical mode decomposition, adaptive decomposition to obtain intrinsic mode functions ( Each intrinsic modulus function satisfies the condition that the difference between the number of local extrema and the number of zero crossings does not exceed 1 and the local mean is zero, thus purifying the depth signal. for:

[0078] ;

[0079] In the formula, The lowest-order eigenmode function index is the frequency of the vibration hammer, which is lower than the excitation frequency band of the vibration hammer. To decompose the total series, For the first The intrinsic modulus function of order m, Units are in meters, retain series. The comprehensive adjustment quality index The decision and rules are detailed in step S04.

[0080] The specific implementation method of step S03 is: with , and As the observation vector, cohesion (Unit: Pa), Angle of Internal Friction (unit: rad) Overconsolidation ratio (Dimensionless) and critical state stress ratio (Dimensionless) Composition of state vector Treating them as random walk state variables, the state equation is:

[0081] ;

[0082] In the formula, Let be the process noise vector, which has a mean of zero and a covariance matrix of... Gaussian distribution, for The process noise covariance matrix is ​​empirically initialized as follows: Observation vector The nonlinear observation equation between the state vector and the state vector is:

[0083] ;

[0084] In the formula, To correct the nonlinear mapping function derived from the Cambridge clay model, the state vector is... Mapped to the observation space, The observed noise vector follows a pattern with a mean of zero and a covariance matrix of... Gaussian distribution, for Observe the noise covariance matrix. Extended Kalman filter pair. exist Perform a first-order Taylor expansion at this point:

[0085] ;

[0086] In the formula, For the first Prior parameter estimates at time step, with dimension 1 , The Jacobian matrix has dimensions of . Each element is Partial derivatives with respect to each state component. The extended Kalman filter prediction step is:

[0087] ,

[0088] ;

[0089] In the formula, The posterior parameter estimates from the previous time step are of dimension 1. , Let be the posterior error covariance matrix of the previous time step, with dimension . , Let be the prior error covariance matrix, with dimension . Using the credibility weight matrix The confidence weight matrix is ​​defined as follows: (The confidence weight matrix is ​​then applied to the observation residuals.)

[0090] ;

[0091] In the formula, ( ) is the first The confidence weight for each sensor channel is determined by the signal-to-noise ratio of the corresponding sensor over 100 consecutive sampling periods. After normalization, the value is assigned. For the first Channel signal variance For the first Channel noise variance, after normalization , It is a dimensionless diagonal matrix. Kalman gain. The parameters are updated as follows:

[0092] ,

[0093] ,

[0094] ;

[0095] In the formula, The Kalman gain matrix has a dimension of . , For the first Posterior parameter estimates at time step 1, with dimension 2. , Let be the posterior error covariance matrix, with dimension . , for Identity matrix. Online recursive least squares method with forgetting factor. Weighted historical residuals, their gain vector Sum of covariance matrix The update formula is:

[0096] ,

[0097] ;

[0098] In the formula, The forgetting factor has an empirical value of 0.98. The covariance matrix for parameter estimation using recursive least squares method has dimensions of . , The recursive least squares gain vector has dimension . The recursive least squares method and the extended Kalman filter are executed alternately for joint updates. .

[0099] The specific implementation of step S04 is as follows: ... , , , The static cone penetration test soil profile data were stitched together to form a multi-channel time series tensor. ,in For inference batch size, Input the number of channels. The time window length is denoted by , and all three are dimensionless. The backbone causal dilated convolutional network is... Level output characteristics for:

[0100] ;

[0101] In the formula, For the first Level convolution kernel, , Indicates the expansion step size is Causal dilated convolution operation, For the first Level bias, For the first Level output characteristics, , It is a linear rectified activation function. The output of the physical residual module is expressed as follows:

[0102] ;

[0103] In the formula, For the first Level-learnable scaling factor, dimensionless. For the residual terms of the hydraulic power conservation equation constraints, The residuals of the Mohr-Coulomb violation criterion constraints are both embedded in the form of differentiable tensors, with dimensions identical to those of the Mohr-Coulomb constraint. The corresponding features are consistent. The input to the soil abrupt change detection subnet is the pressure gradient sequence within the current window. Its definition is:

[0104] ;

[0105] In the formula, For the first The first-order difference of the hydraulic cylinder pressure after vibration damping, in Pa, is used to reflect the characteristics of pressure abrupt changes. The output soil layer type transformation probability is then processed by a three-layer one-dimensional convolution classifier. Dimensionless, when The internal state reset gate is triggered when the gate control loop unit is hidden. Reset to zero The hidden state vector is 64-dimensional, with the same dimension as the number of output channels (64) per stage. The output layer has three parallel prediction heads that output the target velocities of the hydraulic cylinders for the next three control steps. ( ), unit is Based on credibility weight Weighted fusion:

[0106] ,

[0107] ;

[0108] In the formula, The target speed of the hydraulic cylinder after fusion, in units of , For the first Normalized confidence weights for each predictor head, dimensionless. For the first The confidence score of each prediction head is obtained by a linear mapping from the output layer. get, For learnable weight vectors, For learnable bias scalars, This refers to the hidden feature vectors output from the final layer of the backbone network; they are dimensionless. The training loss function... The weighted summation formula is:

[0109] ;

[0110] In the formula, This is the mean square error term for the target speed prediction of the hydraulic cylinder, which is dimensionless (normalized by dividing the square of the speed error by the square of the reference speed). This is a dimensionless penalty term for breach of contract related to physical residuals. The cross-entropy term for soil type classification is dimensionless, and the sum of the weights of the three terms is 1. The weights are determined by a grid search experiment with a step size of 0.05 on the validation set. Dimensionless. Comprehensive adjustment quality index. The calculation formula is:

[0111] ;

[0112] In the formula, The current deviation ratio is dimensionless. This is the baseline deviation ratio, dimensionless, calculated from the average deviation ratio of 15 standard operating condition tests. The standard deviation of the current purification depth signal is expressed in meters. The standard deviation of the baseline cleanroom depth signal, in meters, is calculated from the average standard deviation of the cleanroom depth signal from 15 standard operating condition experiments. This is the current fundamental frequency of the vibratory hammer, in Hz. The rated excitation frequency, in Hz, is read from the equipment nameplate. , , These are weighting coefficients; the sum of the three is 1, and their typical value range is... , , The value was determined by orthogonal experimental design with a step size of 0.05. Dimensionless. When When the inference batch size is adjusted to 1, the empirical mode decomposition retains only the first 3 eigenmode functions; when When the inference batch size is adjusted to 2, the first 5 eigenmode functions are retained; when At that time, the inference batch size is adjusted to 4, and the complete decomposition level is executed.

[0113] The specific implementation of step S05 is as follows: The actual force measured by the hydraulic cylinder... and purification depth signal Calculate the current volumetric strain of the soil layer, and inversely deduce the current effective confining pressure by modifying the Cambridge clay model flow law. The unit is Pa, combined with the overconsolidation ratio. Stress ratio to critical state Calculate and predict resistance The unit is N, and the deviation ratio is... for:

[0114] ;

[0115] In the formula, For the first Time deviation ratio, dimensionless. and All units are in N. When At that time, the control step size is halved and soil constitutive parameters are re-identified; when When the control step size is doubled, the target hydraulic cylinder speed for the current control step is output based on the adaptive step size and the target hydraulic cylinder speed sequence. The unit is .

[0116] The specific implementation of step S06 is as follows: the valve core displacement is described using the Booker-Wynne hysteresis model of the proportional valve. (Unit: m) and input current Nonlinear hysteresis relationship between (unit: A), auxiliary hysteresis state variables The evolution equation (in meters) is:

[0117] ;

[0118] In the formula, The linear recovery coefficient is expressed in units of 1000 ppm. , This is the saturation coefficient, in units of... , This is the shape factor, in units of... , The nonlinear exponent, with an empirical value of 1 or 2, is used to drive the proportional valve via a 1–50 Hz sinusoidal sweep frequency excitation. The parameters are identified by minimizing the mean square error between the Booker-Wen model output and the measured valve core displacement using a particle swarm optimization algorithm. For a given desired valve core displacement... (Unit: meters, determined by the target speed of the hydraulic cylinder) (Originally derived from the hydraulic cylinder flow equation), the desired input current is obtained by numerically solving the above evolution equation. (Unit: A) Control current output by adaptive back-propagation control law (Unit: A) Superimposed, the final drive current of the proportional valve is obtained. (Unit: A):

[0119] ;

[0120] In the formula, The adaptive back-propagation control law is obtained by designing virtual control quantities step by step through Lyapunov's stability theorem and updating the control gain online, so that the closed-loop system remains asymptotically stable when the soil constitutive parameters change, thereby driving the proportional valve to complete the deep adjustment. After completing one control cycle, the system returns to step S03 to execute the next control cycle.

[0121] To better understand and implement this invention, the following is a specific application scenario of the invention, Example 2: To verify the technical effect of the invention, the technicians built a test environment for the depth adaptive adjustment of the rod pulling machine, and selected a stratum with alternating layers of standard sand and gravel as the test site. The rod diameter was 108mm, the rated working pressure of the hydraulic cylinder was 28MPa, the rated excitation frequency of the vibratory hammer was 35Hz, the sampling rate was 25kHz, and the control cycle was 20ms.

[0122] Before the test, technicians used a static cone penetrometer to conduct a penetration test on the test area, obtaining data on the distribution of the cone tip resistance sequence and the sidewall friction force sequence with depth, such as... Figure 2 As shown, the static cone penetration test (CPPT) soil profile data clearly reveals the standard sand layer within the 0–3 m range, the gravel interlayer within the 3–4.5 m range, and the boundary location of the soft clay layer below 4.5 m. The CPPT soil profile data was used to initialize the overconsolidation ratio to 1.8 and the critical stress ratio to 1.2, serving as the initial parameters for modifying the Cambridge clay model.

[0123] During operation, the vibration acceleration sequence collected by the vibration acceleration sensor is tracked in real time via a phase-locked loop, identifying the fundamental frequency of the vibratory hammer as 34.8Hz. An adaptive notch filter is used to sequentially set notch traps at integer multiples of frequencies such as 34.8Hz, 69.6Hz, and 104.4Hz to eliminate vibration interference in the hydraulic cylinder pressure sequence and displacement sensor sequence. Empirical mode decomposition is then applied to the denoised displacement sensor sequence, and the overall mass index is adjusted accordingly. Under the given conditions, the first 5 eigenmode functions are retained and summed to obtain the purification depth signal. The comparison between the purification depth signal and the original displacement sensor sequence is as follows: Figure 3 As shown, the signal before purification has obvious high-frequency vibration components superimposed on it, while the signal after purification smoothly reflects the actual position change trend of the rod.

[0124] In the process of online identification of the modified Cambridge clay model, the extended Kalman filter and the online recursive least squares method are alternately iterated. The signal-to-noise ratio estimates of each sensor and the corresponding confidence weight matrix elements are shown in Table 1.

[0125] Table 1. Estimated signal-to-noise ratio values ​​and confidence weight matrix elements for each sensor

[0126]

[0127] As shown in Table 1, the vibration acceleration sensor has the lowest signal-to-noise ratio under vibration conditions, and its confidence weight matrix elements are the smallest. Consequently, its contribution weight to the soil constitutive parameter identification results is reduced, effectively suppressing the interference of vibration noise on the identification accuracy.

[0128] When the member depth reaches 3.1m, the soil layer abrupt change probability output by the soil layer abrupt change detection subnet of the soil abrupt change-resistance prediction fusion model is 0.83, exceeding the abrupt change detection threshold of 0.7, triggering the hidden state reset of the gated cyclic unit. Figure 4 As shown, the soil layer type transformation probability curve exhibits a significant peak at a depth of 3.1m, which closely matches the initial depth of the gravel interlayer in the static cone penetration soil profile data. After resetting the hidden state, the model outputs a hydraulic cylinder target velocity sequence that conforms to the resistance characteristics of the new soil layer within the first control cycle of the gravel interlayer.

[0129] The dynamic change process of the deviation ratio is as follows: Figure 5 As shown in the table, upon entering the gravel interlayer, the deviation ratio rapidly increased from 2.1% to 18.7%, exceeding the step size reduction threshold of 15%, triggering a reduction of the control step size from 20ms to 10ms and initiating a re-identification of soil constitutive parameters. After four control cycles, the identified cohesion was updated from 12kPa to 35kPa, the internal friction angle from 28° to 38°, the overconsolidation ratio from 1.8 to 2.6, and the critical stress ratio from 1.2 to 1.45. The predicted resistance and the measured hydraulic cylinder force converged again, and the deviation ratio dropped back to 4.2%. The control step size was maintained at 10ms. After the rod entered the soft clay layer below 4.5m, the soil resistance tended to stabilize, and the deviation ratio remained below 3%. The control step size was gradually doubled from 10ms back to 20ms. The soil constitutive parameter identification results for each depth are summarized in Table 2.

[0130] Table 2. Constitutive parameter identification results for each soil layer

[0131]

[0132] The identification results of the Bouc-Wen hysteresis model show that the linear recovery coefficient is 1.24 and the saturation coefficient is 0.87 when the proportional valve moves forward, and 1.09 when it moves backward, with a gain asymmetry of approximately 14%. Numerical inverse mapping converts the desired valve spool displacement into the desired input current, which, when superimposed with the control current output by the adaptive backpropagation control law, drives the proportional valve. The error in the proportional valve spool displacement tracking the desired value converges to the residual range of the Bouc-Wen hysteresis model, effectively eliminating the influence of dead-zone jumps on depth adjustment accuracy. The overall hydraulic cylinder depth adjustment accuracy is as follows: Figure 6 As shown, the error between the purification depth signal and the target depth of the hydraulic cylinder experiences a brief fluctuation during the gravel interlayer crossing process and then converges rapidly, with the overall tracking error remaining within a small range.

[0133] Compared to traditional fixed-step hydraulic proportional valve control methods, the advancements of this invention are reflected in the following aspects: Traditional methods directly feed back sensor signals at a fixed sampling frequency. Vibration interference superimposed on the signal causes the controller to misjudge soil resistance. This invention, through two-stage signal purification, removes vibration interference from the physical mechanism, ensuring the accuracy of the feedback signal. Traditional methods rely on offline soil parameters and cannot detect sudden changes in soil conditions online. This invention, through alternating iterations of extended Kalman filtering and online recursive least squares method, continuously updates the parameters of the modified Cambridge clay model in each control cycle, achieving real-time perception of unknown soil boundary conditions. The fixed step size of traditional methods results in a lag response during sudden changes in soil conditions. The deviation ratio mechanism of this invention directly maps the deviation between predicted resistance and measured force into step size adjustment commands, ensuring that the control step size always matches the rate of change of soil resistance, thus eliminating the response lag problem from a fundamental mechanism perspective.

[0134] It should be noted that the variables involved in this invention are explained in detail in Tables 3 and 4.

[0135] Table 3. Variable Explanation Table (Part 1)

[0136]

[0137] Table 4. Variable Explanation Table (Part Two)

[0138]

[0139] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in the present invention should be included within the scope of protection of the present invention.

Claims

1. A control method for adaptively adjusting the working depth of a lever puller, characterized in that, Includes the following steps: A static cone penetration tester was used to conduct penetration tests on the work area, and the cone tip resistance sequence and sidewall friction force sequence were collected to obtain static cone penetration soil profile data; a hydraulic pressure sensor was used to collect hydraulic cylinder pressure sequence, a displacement sensor was used to collect displacement sensor sequence, a six-axis torque sensor was used to collect six-axis torque sequence, and a vibration acceleration sensor was used to collect vibration acceleration sequence. The vibration acceleration sequence is input into an adaptive notch filter to estimate the fundamental frequency of the vibratory hammer in real time, thereby eliminating vibration interference components in the hydraulic cylinder pressure sequence and displacement sensor sequence. The displacement sensor sequence after vibration interference elimination is then input into empirical mode decomposition processing to extract the sum of intrinsic mode functions and obtain the purification depth signal. Using purification depth signal, hydraulic cylinder pressure sequence, six-axis torque sequence and static cone penetration soil profile data as input, the parameters of the modified Cambridge clay model are identified online by online recursive least squares method combined with extended Kalman filter, and the output cohesion, internal friction angle, overconsolidation ratio and critical stress ratio are updated by weighted confidence weight matrix. The hydraulic cylinder pressure sequence, displacement sensor sequence, six-axis torque sequence, vibration acceleration sequence, and static cone penetration soil profile data are input into the soil abrupt change-resistance prediction fusion model, which outputs the hydraulic cylinder target velocity sequence for multiple future control steps. At the same time, the inference batch size and CUDA flow number of the soil abrupt change-resistance prediction fusion model are dynamically adjusted using a comprehensive adjustment quality index. The current soil volumetric strain is calculated using the measured force of the hydraulic cylinder and the purification depth signal. The current effective confining pressure is inferred by modifying the Cambridge clay model flow law. The predicted resistance is calculated by combining the overconsolidation ratio and the critical stress ratio. The deviation ratio is obtained by comparing the predicted resistance with the measured force of the hydraulic cylinder. When the deviation ratio exceeds the step size halving threshold, the current control step size is halved and the soil constitutive parameters are re-identified. When the deviation ratio is lower than the step size doubling threshold, the current control step size is doubled. The target speed of the hydraulic cylinder for the current control step is output based on the adaptive step size and the target speed sequence of the hydraulic cylinder. By numerical inverse mapping of the Bouc-Wen hysteresis model of the proportional valve, the desired input current is calculated for the desired valve core displacement corresponding to the target speed of the hydraulic cylinder. The desired input current is superimposed with the control current output by the adaptive back-propagation control law to drive the proportional valve to perform deep adjustment and complete one control cycle. Return to the aforementioned online identification steps to execute the next control cycle.

2. The control method according to claim 1, characterized in that, The adaptive notch filter specifically uses a phase-locked loop to estimate the fundamental frequency of the vibrating hammer in real time from the vibration acceleration sequence, and places notch traps at the fundamental frequency of the vibrating hammer and its integer multiples of frequency. The center frequency of the notch traps is continuously updated by tracking the fundamental frequency of the vibrating hammer.

3. The control method according to claim 2, characterized in that, The vibration interference components in the hydraulic cylinder pressure sequence and displacement sensor sequence are eliminated, and the frequency range of the eliminated interference is 10 to 500 Hz. The empirical mode decomposition process extracts the intrinsic mode functions with frequencies lower than the excitation frequency band of the vibratory hammer.

4. The control method according to claim 3, characterized in that, The online recursive least squares method combined with extended Kalman filtering is used to identify the parameters of the modified Cambridge clay model online. Specifically, the extended Kalman filtering is used to linearize the nonlinear observation equations of the modified Cambridge clay model using a first-order Taylor expansion, and the online recursive least squares method uses a forgetting factor to weight the historical residuals. The two methods are used alternately to update the parameter estimates and error covariance.

5. The control method according to claim 4, characterized in that, The elements of the confidence weight matrix range from 0.1 to 1.0 and are assigned values ​​after normalization of the signal-to-noise ratio estimates of each sensor. The signal-to-noise ratio estimates are calculated as the ratio of signal variance to noise variance within a continuous sampling period.

6. The control method according to claim 5, characterized in that, The backbone of the soil mutation-resistance prediction fusion model is a causal dilated convolutional network. Each dilated convolutional block is followed by a physical residual module. The physical residual module embeds the hydraulic power conservation equation and the soil Mohr-Coulomb failure criterion into the residual bypass in the form of differentiable tensor operations.

7. The control method according to claim 6, characterized in that, The soil mutation-resistance prediction fusion model is equipped with a soil mutation detection subnet, which outputs the soil type change probability. When the soil type change probability exceeds the mutation detection threshold, the internal state reset gate is triggered, and the hidden state of the gated loop unit in the backbone network is cleared.

8. The control method according to claim 7, characterized in that, The output layer of the soil mutation-resistance prediction fusion model is configured with multiple parallel prediction heads. The parallel prediction heads are then fused with confidence weights using Softmax probability weights to obtain the target velocity sequence of the hydraulic cylinder.

9. The control method according to claim 8, characterized in that, The calculation of the comprehensive adjustment quality index uses the current deviation ratio, the standard deviation of the purification depth signal, and the fundamental frequency of the vibratory hammer as input components. Each component is divided by its corresponding benchmark value and then weighted and summed. The sum of the three weights is 1. The comprehensive adjustment quality index determines the inference batch size, the number of CUDA streams, and the order of the intrinsic mode functions retained by empirical mode decomposition.

10. The control method according to claim 9, characterized in that, The Bouc-Wen hysteresis model describes the nonlinear memory effect between the proportional valve spool displacement and the input current using an auxiliary hysteresis state variable. The evolution equation of the auxiliary hysteresis state variable includes a linear recovery term and a nonlinear saturation term. The parameters of the Bouc-Wen hysteresis model are identified by driving the proportional valve with sinusoidal frequency sweep excitation and recording the spool displacement response. The mean square error between the model output and the measured spool displacement is minimized using a particle swarm optimization algorithm.

Images