Method for controlling a mechanical arm of a construction machine, concrete pumping device and storage medium

Abstract

The application discloses a mechanical arm control method of an engineering machine, a concrete pumping device and a storage medium, and belongs to the technical field of man-machine interaction control of engineering machines. The method comprises the following steps: acquiring first working condition data of a mechanical arm of the engineering machine at a current moment by detecting that an operator pulls a terminal hard rod of the mechanical arm; calling a physical model based on the first working condition data to generate a physical rough estimation of a terminal force on the terminal hard rod; acquiring observation conditions at the current moment; calling a conditional denoising diffusion probability model (DDPM) to generate a residual compensation amount by taking the observation conditions as prior information; and controlling the engineering machine based on the residual compensation amount and the physical rough estimation of the terminal force to drive the mechanical arm to move following the pulling intention of the operator. The application can improve the stability of the concrete pumping device when responding to the pulling intention of the operator and reduce costs.

Description

Technical Field

[0001] This application belongs to the field of human-machine interaction control technology for construction machinery, specifically relating to a robotic arm control method for construction machinery, concrete pumping equipment, and storage medium. Background Technology

[0002] With the continuous advancement of urbanization and the rapid development of large-scale infrastructure construction, concrete pumping equipment, as a core piece of equipment in modern engineering construction, has been widely used in the field of building construction.

[0003] During construction, the operator applies guiding force to the end effector of the robotic arm via a traction rod. The control system senses this force in real time and converts it into motion commands, enabling single-person operation. The core of this method is the accurate sensing of force at the end effector. Existing technologies typically involve directly installing multi-dimensional force sensors at the end effector, but this approach suffers from drawbacks such as high cost, susceptibility to damage, and vulnerability to signal interference. Summary of the Invention

[0004] The purpose of this application is to provide a method for controlling the robotic arm of engineering machinery, a concrete pumping device, and a storage medium, which can improve the stability of the concrete pumping device in response to the operator's traction intention and reduce costs.

[0005] To solve the above-mentioned technical problems, this application is implemented as follows: In a first aspect, embodiments of this application provide a method for controlling the robotic arm of engineering machinery, the method comprising: The system detects that the operator is pulling the end rigid bar of the robotic arm of the construction machinery, and acquires the first working condition data of the robotic arm at the current moment. Based on the first working condition data, the physical model is invoked to generate a physical coarse estimate of the end force on the end rigid bar; The observation conditions at the current moment are obtained; wherein the observation conditions are obtained by splicing together the multi-dimensional time-series features within the previous H control cycles at the current moment, and the multi-dimensional time-series features are obtained based on the second working condition data of the robotic arm; Using the observation conditions as prior information, the Conditional Denoising Diffusion Probability Model (DDPM) is invoked to generate residual compensation. Based on the residual compensation amount and the physical coarse estimate of the end force, the engineering machinery is controlled to drive the robotic arm to move in accordance with the operator's traction intention.

[0006] Secondly, embodiments of this application provide a concrete pumping device, which includes a processor, a memory, and a program or instructions stored in the memory and executable on the processor. When the program or instructions are executed by the processor, they implement the steps of the method described in the first aspect.

[0007] Thirdly, embodiments of this application provide a computer-readable storage medium on which a program or instructions are stored, which, when executed by a processor, implement the steps of the method described in the first aspect.

[0008] Fourthly, embodiments of this application also provide a computer program product, including a computer program that, when executed by a processor, implements the steps of the method described in the first aspect.

[0009] In this embodiment, by detecting the end effector of the robotic arm being pulled by the operator, the first working condition data of the robotic arm at the current moment is obtained. Based on the first working condition data, a physical model is invoked to generate a physical coarse estimate of the end force on the end effector. The observation conditions at the current moment are obtained; wherein, the observation conditions are obtained by splicing multidimensional time-series features within the previous H control cycles, and the multidimensional time-series features are obtained based on the second working condition data of the robotic arm. Using the observation conditions as prior information, the Conditional Denoising Diffusion Probability Model (DDPM) is invoked to generate a residual compensation amount. Based on the residual compensation amount and the physical coarse estimate of the end force, the robotic arm is controlled to drive the robotic arm to move in accordance with the operator's pulling intention. That is, without installing a multidimensional force sensor at the end, the guiding force applied by the operator to the end of the robotic arm by the traction rod can be obtained through the residual compensation amount and the physical coarse estimate of the end force, thereby improving the stability of the concrete pumping equipment in response to the operator's pulling intention and reducing costs. Attached Figure Description

[0010] Figure 1 This is one of the flowcharts illustrating a robotic arm control method for engineering machinery provided in some embodiments of this application; Figure 2 This is a schematic diagram of the overall architecture of the control system provided in some embodiments of this application; Figure 3 This is one of the flowcharts illustrating a robotic arm control method for engineering machinery provided in some embodiments of this application; Figure 4 These are internal structural diagrams of concrete pumping equipment provided in some embodiments of this application. Detailed Implementation

[0011] 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 some embodiments of this application, not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0012] The terms "first," "second," etc., used in the specification and claims of this application are used to distinguish similar objects and not to describe a specific order or sequence. It should be understood that such use of data can be interchanged where appropriate so that embodiments of this application can be implemented in orders other than those illustrated or described herein. Furthermore, in the specification and claims, "and / or" indicates at least one of the connected objects, and the character " / " generally indicates that the preceding and following objects are in an "or" relationship.

[0013] In one exemplary embodiment, this application proposes an applicable control system suitable for a three-section hydraulically driven concrete pumping equipment. Without installing an external force sensor at the end, the control system can achieve online accurate estimation of the external force at the end using only the existing pressure and angle sensors at the three hydraulically driven joints. The estimation results are then input into the admittance control frame to drive the robotic arm to follow the traction force applied by the construction worker through the end rigid rod, thereby realizing single-person traction construction operation.

[0014] The control system can be composed of the following three functional layers: The first layer is the signal preprocessing layer. For periodic pressure pulsations of the hydraulic pump and random disturbances such as wind load, an adaptive notch filter and a Kalman estimator are designed to output the purified pressure signal and the disturbance state estimate.

[0015] The second layer is the physical model layer, which uses a dynamic model with precise parameters of the rigid rod to convert the purification hydraulic data into a coarse estimate of the end force.

[0016] The third layer is the diffusion model residual compensation layer. With physical coarse estimation and multi-dimensional time series observation features as prior conditions, the residual compensation is modeled as a conditional generation problem using the Denoising Diffusion Probabilistic Model (DDPM). The probability distribution of the physical model residual is learned through the perturbation-aware temporal Transformer network and a high-precision compensation amount is generated. At the same time, the uncertainty quantification index is naturally output using the multiple parallel sampling mechanism. This index is superimposed with the coarse estimation to obtain the final accurate estimate of the end force.

[0017] In one exemplary embodiment, this application proposes a robotic arm control method for construction machinery. The robotic arm control method for construction machinery provided in this application will be described in detail below with reference to the accompanying drawings and specific embodiments and application scenarios.

[0018] Reference Figure 1 The method includes steps 102-110. Wherein: Step 102: Detect the operator pulling the end rigid rod of the robotic arm of the construction machinery, and obtain the first working condition data of the robotic arm at the current moment.

[0019] In some embodiments, the current moment can be the moment when the end rigid bar of the robotic arm of the construction machinery is detected being pulled by an operator.

[0020] In some embodiments, the operating condition data may include, but is not limited to: pressure signals, angle signals, and pressure pulsation frequency information.

[0021] The pressure signals include the rodless cavity pressure signals of each joint of the robotic arm. With rod chamber pressure signal The angle signal is the angle signal of each joint. .in, The value depends on the number of joints, for example, the value corresponding to a three-section lever robotic arm. .

[0022] Among them, the pressure pulsation frequency information is the harmonic frequency information of the hydraulic pump of each joint of the robotic arm. The harmonic frequency information carries a harmonic frequency that is an integer multiple of the pressure pulsation fundamental frequency. The calculation formula (1) is as follows: in, The harmonic frequency; The harmonic orders that a multi-order adaptive notch filter aims to align with and suppress. , The highest harmonic order that the multi-order adaptive notch filter aims to align with and suppress; This is the fundamental frequency of pressure pulsation.

[0023] In some embodiments, the pressure pulsation fundamental frequency can be determined by the rotational speed of the hydraulic pump. (Unit: r / min) and number of plungers It can be calculated using the following formula (2): in, The unit is Hz.

[0024] In some embodiments, when the speed sensor fails and cannot obtain data... In such cases, adaptive frequency estimation can be performed using the Least Mean Square (LMS) algorithm. Specifically, the fundamental frequency is extracted from the pressure signal using the LMS algorithm.

[0025] Step 104: Based on the first working condition data, call the physical model to generate a physical coarse estimate of the end force on the end rigid bar.

[0026] In some embodiments, since the original pressure signal is superimposed with periodic pressure pulsations of the hydraulic pump and random disturbances such as wind load, the random disturbances can be eliminated before calling the physical model, that is, the original pressure signal is preprocessed to obtain a purified pressure signal.

[0027] In some embodiments, preprocessing may involve first eliminating the disturbance force generated by the periodic pressure pulsation of the hydraulic pump in the original pressure signal, and then eliminating the disturbance force generated by the wind load.

[0028] In some embodiments, preprocessing may involve first eliminating the disturbance force generated by wind load in the original pressure signal, and then eliminating the disturbance force generated by the periodic pressure pulsation of the hydraulic pump.

[0029] Taking the elimination of disturbance forces generated by the periodic pressure pulsations of the hydraulic pump in the original pressure signal, followed by the elimination of disturbance forces generated by wind load, as an example, the elimination of disturbance forces generated by the periodic pressure pulsations of the hydraulic pump can be achieved by constructing a system targeting the fundamental frequency of the pulsation and its harmonics. A multi-order adaptive notch filter is constructed by cascading second-order infinite impulse response (IIR) notch elements. The first... The transfer function of the stage is shown in formula (3): in, For the first Second-order IIR notch filter element in The transfer function in the transform domain describes the frequency response of this stage of the notch filter element to the original pressure signal, i.e., at the center frequency. A deep trough is formed at the point to suppress the corresponding harmonics; Notch bandwidth factor The closer it is to 1, the narrower the notch bandwidth; For complex variables in the Z-transform, in the frequency domain ; The unit delay operator represents a delay of one sampling period in discrete time. Accordingly, Indicates a delay of 2 sampling periods in discrete time. ; For the first The center frequency of the notch filter at level 1, with Dynamically updated, thus adjusting the rotational speed. When the change occurs, the spectral line is continuously aligned with the pulsating line. The calculation formula (4) is as follows: in, The system (signal) sampling frequency, in Hz, is the sampling frequency used in digital signal processing and notch filter implementation; its function is to sample continuously fluctuating angular frequencies. Normalized to discrete angular frequency in the digital domain In this embodiment =500Hz (corresponding sampling period) =2 ms).

[0030] In some embodiments, =0.96.

[0031] In summary, the original pressure signal is transformed into a de-pulsating pressure signal after passing through the notch units of the multi-stage adaptive notch filter.

[0032] To eliminate the disturbance forces generated by wind loads, the disturbance forces generated by wind loads can be... The model is a second-order Markov (Autoregressive model of order 2, AR(2)) stochastic process. The slowly varying purification pressure P (i.e., the purified pressure signal after removing pump pulsation and wind load disturbance) and the disturbance force are used to form an augmented state vector. The state-space model is established as shown in the following formulas (5) and (6): in, This refers to the discrete sampling time sequence number. for The purification pressure value to be estimated at any time; for Wind-borne disturbance forces at any given moment; for Time (i.e.) The wind-borne disturbance force (at the moment before); for Time (i.e.) The augmented state vector at the next moment; for The pressure signal (measured value) at any given moment is de-pulsated. Indicates process noise. For process noise covariance, Indicates measurement noise. To measure the noise covariance; The measurement matrix is ​​as follows: in, This is the coupling coefficient between the disturbance force generated by wind load and the pressure channel.

[0033] in, The state transition matrix is ​​as follows: in, and The coefficient of the disturbance force generated by wind load can be obtained offline from the on-site wind load power spectrum.

[0034] Based on the above state-space model, the standard Kalman filter prediction-update iteration is performed, and the formulas (7) to (11) are as follows: in, For one-step prediction (prior) state estimation; for State estimation at time; The prior (prediction) error covariance matrix; for The posterior error covariance matrix at time t; for The posterior error covariance matrix at time t; for Kalman gain matrix at time step; It is the identity matrix; superscript This is the matrix transpose.

[0035] It should be noted that after the standard Kalman filter prediction-update iteration converges, the state estimation... Simultaneously extract the pressure estimate after purification. Estimation of disturbance forces generated by wind load The former is fed into the physical model, while the latter serves as one of the conditional features of the conditional denoising diffusion probability model.

[0036] The pressure signal after purification (after convergence) is as follows. ,Right now (and an explanation of the calculation process for estimating the disturbance force of random disturbances.)

[0037] The physical model is explained below: This physical model models the dynamic behavior of the boom system (i.e., the robotic arm), involving the calculation of the driving torque of each joint, the equivalent incorporation of the dynamic parameters of the end effector, multibody dynamics compensation, and a physical coarse estimate of the end effector force.

[0038] Specifically, the calculation of the driving torque for each joint is as follows: Utilizing the purified pressure signal Calculate the net driving thrust of each hydraulic cylinder and compensate for the sealing friction. Let the first... The effective working area of ​​the rodless cavity of each joint is The effective working area of ​​the rod cavity is Then we have the following formula (12).

[0039] in, For the first The driving force of each joint; For the first The rodless cavity pressure signal of each joint; For the first The pressure signal of the rod cavity in each joint. This can be obtained through the lever arm function. Converting to driving torque, the conversion formula (13) is as follows: in, For the first The driving torque of each joint; For the first The lever arm function of each joint is determined by the boom geometry and varies with the joint angle signal. change.

[0040] Specifically, regarding the equivalent incorporation of the dynamic parameters of the end-stiffened rod: Based on the accurate incorporation of the mass parameters of the end-mounted rigid rod (i.e., the traction rod) into the dynamic parameters of the end-arm, the parallel axis theorem is used to calculate the dynamic parameters of the end-arm. These dynamic parameters include the equivalent mass, equivalent center of mass position, and equivalent moment of inertia of the end-arm.

[0041] The equivalent mass is calculated using the following formula (14): in, For equivalent quality; The mass of the distal arm body; The mass of the rigid rod.

[0042] The position of the equivalent centroid is calculated using the following formula (15): in, The location of the equivalent center of mass; This is the position of the center of mass of the distal arm body; This is the location of the center of mass of the rigid rod.

[0043] The equivalent moment of inertia is calculated using the following formula (16): in, It is the equivalent moment of inertia; The moment of inertia of the end arm body; It is the square of the distance from the center of mass of the distal arm body to the equivalent center of mass; Let be the moment of inertia of the rigid rod; It is the square of the distance from the center of mass of the rigid bar to its equivalent center of mass.

[0044] For multibody dynamics compensation, specifically, the multibody dynamics equation (17) containing end forces (i.e., physical coarse estimates of end forces) is as follows: in, For the actual (driving) torque vector of all joints (the generalized torque vector of the joint), its first... Components ; The inertia matrix; Angular velocity, Angular acceleration; These are the Coriolis force and centrifugal force terms; This is the term related to gravity. This refers to the joint friction torque; This is the transpose of the terminal geometric Jacobian matrix; The external force at the end point to be solved; and Depend on It is obtained through differential and low-pass filtering.

[0045] In some embodiments, the joint friction torque is described using the Stribeck model. The parameters of the Stribeck model are obtained through offline identification, and its mathematical expression (18) is as follows: in, For the first Angular velocity of each joint; For the first The frictional torque of each joint; For the first Coulomb friction torque of each joint; For the first The static friction torque of each joint; For the first Stribeck characteristic velocities of each joint; For shape factor; For the first angular velocity of each joint The sign function takes values ​​of +1 / 0 / 1. Used to characterize the switching of Coulomb / static friction torque with the direction of joint movement; It is the coefficient of viscous friction.

[0046] For a rough physical estimate of the end force, specifically: By rearranging the terms of equation (17) above, we can obtain the measured torque residual after removing the terms of inertia, Coriolis, gravity, and friction. As shown in formula (19).

[0047] Based on the following formula (20), the generalized inverse of the Jacobi transpose can be used to... This is mapped to a physical coarse estimate of the end force.

[0048] in, A rough physical estimate of the end force; This is the generalized inverse of the transpose of Jacobi. .

[0049] Step 106: Obtain the observation conditions at the current moment; wherein, the observation conditions are obtained by splicing multi-dimensional time-series features within the previous H control cycles at the current moment, and the multi-dimensional time-series features are obtained based on the second working condition data of the robotic arm.

[0050] In some embodiments, H is a natural number, which is an empirical value, such as 50.

[0051] In some embodiments, multidimensional time series features can be obtained by concatenating the purified physical coarse estimates of pressure, angle, angular velocity, end force, disturbance force estimates, and pressure pulsation fundamental frequency estimates. The mathematical expression (21) of the multidimensional time series features is as follows: in, Multidimensional temporal features; For the current moment (No. Purification pressure (each control cycle); Before the current moment (No. From the perspective of (one control cycle); Before the current moment (No. angular velocity (in control cycles); For the current moment (No. A rough physical estimate of the end force (in one control cycle); For the current moment (No. Estimation of disturbance force (in control cycles); This is an estimate of the fundamental frequency of the pressure pulsation.

[0052] In some embodiments, when a speed sensor is available, multidimensional time-series characteristics can be obtained by splicing together physical coarse estimates of purification pressure, angle, angular velocity, end force, disturbance force estimates, and pressure pulsation fundamental frequency.

[0053] It is understandable that the observation conditions are... That is, the H control cycles prior to the current time. The multidimensional conditional vector obtained by sequentially splicing comprehensively describes the input and output states of the physical model and the level of environmental disturbance under the current working conditions. It can provide sufficient prior information for the conditional denoising diffusion probability model to accurately capture the distribution characteristics of the physical modeling residuals.

[0054] Step 108: Using the observation conditions as prior information, call the conditional denoising diffusion probability model to generate residual compensation.

[0055] In this embodiment, the residual compensation amount is first defined. The difference between the true value of the end force and the physical rough estimate of the end force mentioned above is given by formula (22).

[0056] in, This represents the true value of the end force. This is a rough physical estimate of the end force.

[0057] This embodiment models the residual compensation problem as a conditional probability generation problem, and uses DDPM to generate high-probability samples of the residual distribution from the observation conditions, thereby providing high-precision residual compensation while naturally outputting the estimated uncertainty.

[0058] It is important to emphasize that the technical motivation for choosing DDPM over other data-driven methods is as follows: (1) Compared with neural network regression models (such as multilayer perceptron (MLP) and long short-term memory (LSTM) networks) that only output deterministic point estimates, DDPM can model the multimodal characteristics and working condition dependence of hydraulic system residuals by learning the conditional probability distribution of residuals, and has better generalization ability in non-stationary working condition switching scenarios.

[0059] (2) Compared with other generative models such as Generative Adversarial Network (GAN), DDPM is more stable in the training phase, does not have the problem of mode collapse, and the iterative characteristics of the denoising process make it naturally robust to changes in conditional input.

[0060] (3) The randomness of multiple sampling in DDPM naturally provides the ability to measure uncertainty. Reliable confidence indexes can be obtained without the need to build an additional Bayesian inference framework or ensemble learning structure. This is a unique advantage that traditional regression methods and other generative models do not have.

[0061] (4) The denoising process of DDPM can be accelerated by sampling strategies such as Denoising Diffusion Implicit Models (DDIM) to significantly compress the number of inference steps, so as to meet the latency requirements of real-time control of engineering machinery.

[0062] The training phase of DDPM is a forward diffusion process. During forward diffusion, random Gaussian noise is added to the ground truth residuals to generate prediction noise. The loss is calculated by comparing the predicted noise with the actual added Gaussian noise, thereby adjusting the model parameters. Essentially, it learns the conditional distribution from the noisy ground truth residuals to the prediction noise. Specifically: The forward diffusion process can be defined as: The true value of the residual is evaluated at T time steps (e.g., T=100). The closed-form sampling expression of a Markov chain with progressively added Gaussian noise is given by the following formula (23): in, For diffusion time step Noisy samples at that time; ; The cumulative signal retention coefficient for the forward diffusion process is the coefficient up to the [number of]th ... Original residuals at step time The cumulative retention ratio, the larger the value, the more original signal is retained and the less noise, which is calculated by the following formula (24): in, For the first The noise scheduling coefficient (noise variance) of the first step, i.e., the noise variance of the first step in the forward diffusion process. The variance of Gaussian noise injected at each time step ∈(0,1).

[0063] The noise scheduling employs a cosine function strategy, resulting in slow noise growth in the early stages and a rapid approach to pure noise in the later stages. This is beneficial for preserving signal details during the inverse denoising process in the subsequent inference stage. Scheduling refers to the process that occurs with the diffusion time step... Pre-arranged noise variance sequence { The value of} is determined by the rules governing how much noise is injected at each step. This cosine function strategy is shown in formulas (25) and (26).

[0064] in, Auxiliary function exist The value at =0 is used as the normalization reference for cosine noise scheduling (making... =1); as follows: in, .

[0065] The inference phase of DDPM is a reverse denoising process, which outputs the prediction residual through conditional distribution and randomly generated Gaussian noise.

[0066] It should be noted that, in both the training and inference phases, the observation conditions are treated as prior information.

[0067] It should be noted that the inverse denoising process is performed by a parameterized denoising network. Driven by the current noisy sample, the diffusion time step embedding, and the observation condition c, the network predicts the noise component of the current step. This embodiment provides a perturbation-aware temporal Transformer network to improve upon DDPM. The core of the perturbation-aware temporal Transformer network lies in the collaboration of three sub-modules: (a) The observation feature encoder uses a one-dimensional causal convolution and multi-head self-attention layer cascade structure to extract features from the H-step observation conditions and output a fixed-dimensional conditional embedding vector (i.e., conditional features).

[0068] (b) The disturbance-aware gating fusion module estimates the disturbance force output by the Kalman filter. Encode as a perturbation feature separately Then, through learnable gating coefficients The weight of its influence on the denoising process is adaptively adjusted, and the formula (27) for the gated fusion output is as follows: in, The fusion characteristics of the gated fusion output; For conditional features; The disturbance characteristics are obtained by estimating and encoding the disturbance force. Element-wise multiplication; learnable gating coefficients It is calculated using the following formula (28): in, Use the Sigmoid activation function; and These are the learnable weight matrix and the bias term, respectively.

[0069] It is understandable that the disturbance-aware gating fusion module enables the improved DDPM to enhance its attention to disturbance-related residual components under strong disturbance conditions, and automatically reduce the contribution of the disturbance channel under weak disturbance conditions, thus avoiding interference from invalid information. This improves the robustness of the improved DDPM under harsh operating conditions.

[0070] (c) The denoising backbone network uses a multi-layer cross-attention Transformer block, where noisy samples and time step embeddings are used as queries. As key / value pairs, conditional information is used to finely guide the denoising process.

[0071] In some embodiments, the loss function of DDPM is further improved, specifically: The improvement lies in proposing a multi-objective composite loss. Specifically, during the training phase of DDPM, a multi-objective composite loss function is employed, which includes the following four components: diffusion principal loss... Physical consistency loss Disturbance robustness loss and amplitude constraint loss .

[0072] in, The mathematical expression (29) is as follows: in, For expectation operators; The square of the L2 norm; To predict noise.

[0073] in, The mathematical expression (30) is as follows: in, The mathematical expression (31) is as follows: in, For expectation operators; The predicted noise is the result of augmenting the training samples with simulated perturbations of different magnitudes.

[0074] in, The mathematical expression (32) is as follows: in, It is an activation function; This is the absolute value of the predicted residual; The residual threshold characterizes the physically reasonable range.

[0075] In this embodiment, Used to optimize the noise prediction accuracy of denoising networks; The residual compensation amount (i.e. the predicted residual) used to constrain the generation is superimposed with the physical coarse estimate of the end force to satisfy the joint torque balance relationship, ensuring that the output does not deviate from the physical law; The simulated perturbation augmentation (denoted by the superscript aug) is used to apply different amplitudes to the training samples, encouraging the improved DDPM to maintain stable compensation performance under multiple perturbation levels; Used to punish excess The residual output prevents abnormally large compensation amounts from occurring under distributed external inputs.

[0076] Based on the above four components, an uncertainty-weighted strategy can be used in multi-task learning to achieve learnable weights. The components are automatically balanced to obtain the total loss, which is the expression (33) of the multi-objective composite loss function, as follows: in, For the total loss, For the first Each component.

[0077] In some embodiments, during the offline pre-training phase, a high-precision reference force sensor can be temporarily installed at the end effector of the robotic arm to collect a dataset covering multiple working conditions, including static posture, dynamic motion, vibrations caused by different hydraulic pump speeds, multi-level wind load disturbances, and different rigid bar angle postures. The difference between the reference force sensor reading and the physical coarse estimate of the end effector force is used as the true residual label. Training is then conducted. After training is complete, the reference force sensors are removed during the deployment phase, and the model can operate independently using only the existing sensor data from the hydraulic joints, thus predicting the residual compensation amount.

[0078] Step 110: Based on the residual compensation amount and the physical coarse estimate of the end force, control the engineering machinery to drive the robotic arm to move in accordance with the operator's traction intention.

[0079] In some embodiments, during the inference phase, to meet the real-time requirements of the engineering machinery control system, a DDIM accelerated sampling strategy is adopted to reduce the number of denoising steps compared to the training phase. Specifically, the number of denoising steps is compressed from more during training to less during inference; for example, the T=100 denoising steps used in training are compressed to only 8 steps to complete high-quality sampling. The formula (34) for the DDIM single-step deterministic update (η=0) is as follows: In each control cycle, N (e.g., N=10) independent DDIM samplings are performed in parallel. The mean of each sampling result is used as the residual compensation output, and the standard deviation is used as the uncertainty quantification index. The formula for calculating the mean (35) is as follows: in, This is the mean of all sampling results; For the first The sampling results.

[0080] The formula for calculating the standard deviation (36) is as follows: in, As a quantification of uncertainty; This is the formula for calculating standard deviation.

[0081] It should be noted that this uncertainty quantification reflects the confidence level of DDPM in the residual estimation under the current operating conditions: when the boom system is in a normal operating condition with good training data coverage, the results of multiple samplings are highly consistent and the uncertainty is low; when the boom system encounters extreme operating conditions outside the training distribution, the sampling results are scattered and the uncertainty increases, thus providing a basis for safety decision-making for downstream control modules.

[0082] In some embodiments, to meet the high-frequency real-time requirements of the underlying control system of construction machinery, this embodiment adopts a hardware and software architecture that decouples the control and estimation frequencies during engineering deployment. Specifically, the joint PD controller and the physical model layer operate at high frequency in the on-board main controller to ensure the stability of the basic control; while the DDPM residual compensation layer is deployed in the edge AI computing unit, accelerated by parallel computing, outputting the residual compensation amount asynchronously at a relatively low frequency, and then fused with the high-frequency physical coarse estimate after smooth interpolation through a zero-order hold. This hardware and software architecture, combined with DDIM accelerated sampling, can solve the computational bottleneck problem of complex deep generative models in the real-time control of construction machinery.

[0083] In some embodiments, within each control cycle, the physical coarse estimate and the residual compensation amount are superimposed to obtain the final end force estimate, and the calculation formula (37) is as follows: in, For the final end force estimation; This is a rough physical estimate.

[0084] Subsequently, the following two types of physical constraint checks are performed on the final end force estimate: (1) Torque consistency verification requires that it be performed by Inverse calculation of joint torque The measured torque residual The deviation does not exceed the torque difference threshold. .

[0085] (2) Amplitude rationality verification, requiring || ||≤ ; This is the threshold value for the upper limit of the end force amplitude. Based on the structural strength of the robotic arm and end effector, the rated capacity of the hydraulic system, and the maximum traction force that the operator may apply during actual construction, combined with a safety factor, the value is determined offline by design parameters or on-site calibration (for example, approximately 1.2 to 1.5 times the rated / calibrated maximum traction force).

[0086] If any check fails, the estimation is deemed abnormal. rollback to alternative output or to Amplitude limiting is performed to ensure the safety of downstream control. The verification formula (38) is as follows: in, This is the output after verification.

[0087] exist In the case of, The input is fed into a human-machine motion intention recognition observer to smoothly weight and mix the preset admittance parameters for two working conditions. The two working conditions are a horizontal construction condition and a vertical construction condition. The admittance parameters include... and ,in, The formula for the weighted mixture (39) is as follows: in, The weighted mixed admittance-inertia parameters (desired mass / inertia matrix); Preset admittance inertia parameters (mass / inertia matrix) for horizontal construction conditions. Preset admittance inertia parameters (mass / inertia matrix) for vertical construction conditions. This is the weighting coefficient for mixed operating conditions. It is calculated using the following formula (40): in, The angle between the traction force and the horizontal direction; To switch the threshold angle; To switch the steepness coefficient. Among them, If it approaches 0, it indicates a bias towards a horizontal working condition; If it approaches 1, it indicates a bias towards vertical operating conditions.

[0088] in, The formula (41) for the weighted mixture is as follows: in, The weighted admittance damping parameters (i.e., the basic damping matrix) are the combined admittance damping parameters. The preset admittance and damping parameters (i.e., damping matrix) for horizontal construction conditions. The preset admittance and damping parameters (i.e., damping matrix) for vertical construction conditions.

[0089] Taking the target stiffness during traction servo operation (i.e., free guidance), the admittance control equation (42) at this time is: in, This is the acceleration of the rigid rod at the end; Uncertainty – Damping Gain Coefficient; The velocity of the end rigid rod; The fuzzy damping increment is determined by the fuzzy controller based on the end velocity. With traction Calculated in real time.

[0090] for The two inputs and output of the fuzzy controller are each divided into three fuzzy sets: {small, medium, and large}. Mamdani inference and the centroid method are used for defuzzification. The fuzzy rules are shown in Table 1 below: Table 1: Fuzzy rules.

[0091] The physical meaning of the above rules is as follows: When the force and speed are small, it is in the precise positioning stage, requiring greater damping to prevent the robotic arm from drifting; when the force and speed are large, the operator's traction intention is clear, and the damping should be reduced to improve follow-up sensitivity and ease of operation. Taking "medium damping" as an example when the traction force is "medium" and the end-effector speed is "small," the output is "medium damping." Its physical meaning is that the operator applies a medium traction force, but the robotic arm moves slowly, in an intermediate state transitioning from "precise positioning (large damping to prevent drift)" to "clear follow-up (small damping to improve sensitivity)," therefore, medium damping is applied to balance stability and follow-up sensitivity.

[0092] for This demonstrates the synchronous adjustment effect of uncertainty on damping. Specifically, when When the pressure increases, the damping automatically improves to suppress flutter that may be caused by a decrease in confidence, thereby achieving a dynamic balance between estimation accuracy and control safety. Furthermore, by adjusting the admittance and damping parameters in real time through a fuzzy controller, the human-machine motion intention recognition observer can automatically distinguish between horizontal and vertical construction conditions and switch seamlessly, thereby improving construction continuity and the user experience.

[0093] Admittance controller outputs desired end position increment The joint angle command increment is solved by the Jacobi pseudo-inverse method. Ultimately, the PD controller drives the hydraulic valves of each joint to achieve closed-loop position tracking.

[0094] It is understandable that the randomness of multiple samplings by the diffusion model naturally outputs an uncertainty quantification index, and the uncertainty is fed back to the damping adjustment loop of the admittance controller to achieve safe and compliant control driven by the estimated confidence level.

[0095] This embodiment innovatively introduces the diffusion generation model into the field of engineering machinery force estimation by combining a physical model with a conditional denoising diffusion probability model. It achieves high-precision online estimation of end force and quantification of uncertainty without relying on end force sensors. Combined with fuzzy adaptive damping admittance control and human-machine intention recognition, it provides an efficient, stable, and low-cost single-person traction construction control scheme for concrete pumping equipment, which has significant engineering application value.

[0096] It should be understood that although the steps in the flowcharts of the embodiments described above are shown sequentially according to the arrows, these steps are not necessarily executed in the order indicated by the arrows. Unless explicitly stated herein, there is no strict order restriction on the execution of these steps, and they can be executed in other orders. Moreover, at least some steps in the flowcharts of the embodiments described above may include multiple steps or multiple stages. These steps or stages are not necessarily completed at the same time, but can be executed at different times. The execution order of these steps or stages is not necessarily sequential, but can be performed alternately or in turn with other steps or at least some of the steps or stages of other steps.

[0097] For ease of understanding, such as Figure 2 and Figure 3 As shown, a specific embodiment is used for illustration: In this embodiment, the sampling period of the three hydraulic joints of the concrete pumping equipment is uniformly set to Ts, where Ts = 2ms, i.e., the sampling frequency fs = 500 Hz. The overall process includes the following steps: Step 1: Acquire raw pressure signals, including pressure signals from the rodless and rod-type cavities of each joint, as well as joint angle signals.

[0098] Step 2: Based on the original pressure signal, perform signal preprocessing to eliminate the periodic pulsation of the hydraulic pump and wind load, and estimate random disturbances to obtain the purified pressure signal and disturbance force estimate.

[0099] Step 3: Based on the purification pressure signal, disturbance force estimate and angle signal, call the physical model layer to calculate the coarse estimate of the end force.

[0100] Step 4: Using multidimensional time series observation characteristics as observation conditions, call the improved DDPM to generate residual compensation and uncertainty.

[0101] Step 5: Add the coarse estimate to the residual compensation amount to obtain the final end force estimate.

[0102] Step 6: Perform physical constraint verification on the final end force estimate. If the verification passes, input the uncertainty along with the result into the damping adaptive admittance controller.

[0103] Step 7: If the verification result is abnormal, the output is replaced with a physical coarse estimate and the process returns to step 3 to re-estimate.

[0104] Step 8: The damping adjustment loop of the admittance controller receives the uncertainty, combines it with the final estimated end force obtained from multiple samplings, and outputs the joint motion control signal.

[0105] The specific implementation methods are described in the above embodiments and will not be repeated here.

[0106] In one exemplary embodiment, a robotic arm control system for engineering machinery is provided, comprising: sensor components disposed at each hydraulic joint; an on-board main controller for performing signal preprocessing and physical model layer calculations; an edge AI calculation unit for generating a model of operating conditions and outputting residual compensation and uncertainty; a fusion and verification module for fusing physical coarse estimates and residual compensation and performing physical constraint verification; an admittance control module including a motion intention recognition observer and an adaptive admittance controller; and a hydraulic servo drive component for receiving joint motion commands.

[0107] In one exemplary embodiment, a concrete pumping device is provided, the internal structure of which can be shown in the following diagram. Figure 4As shown, the concrete pumping equipment includes a processor, memory, input / output (I / O) interfaces, and a communication interface. The processor, memory, and I / O interfaces are connected via a system bus, and the communication interface is also connected to the system bus via the I / O interfaces. The processor provides computational and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system and computer programs. The internal memory provides the environment for the operation of the operating system and computer programs stored in the non-volatile storage media. The I / O interfaces are used for exchanging information between the processor and external devices. The communication interface is used for communicating with external terminals via a network connection. When the computer program is executed by the processor, it implements a robotic arm control method for engineering machinery.

[0108] Those skilled in the art will understand that Figure 4 The structure shown is merely a block diagram of a portion of the structure related to the present application and does not constitute a limitation on the computer device to which the present application is applied. Specific computer devices may include more or fewer components than those shown in the figure, or combine certain components, or have different component arrangements.

[0109] In one embodiment, a computer-readable storage medium is provided, on which a program or instructions are stored, which, when executed by a processor, implement the steps in the above-described method embodiments.

[0110] In one embodiment, a computer program product is provided, including a computer program that, when executed by a processor, implements the steps in the above method embodiments.

[0111] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium, and when executed, it can include the processes of the embodiments of the above methods. Any references to memory, databases, or other media used in the embodiments provided in this application can include at least one of non-volatile memory and volatile memory. Non-volatile memory can include read-only memory (ROM), magnetic tape, floppy disk, flash memory, optical memory, high-density embedded non-volatile memory, resistive random access memory (ReRAM), magnetic random access memory (MRAM), ferroelectric random access memory (FRAM), phase change memory (PCM), graphene memory, etc. Volatile memory can include random access memory (RAM) or external cache memory, etc. By way of illustration and not limitation, RAM can take many forms, such as Static Random Access Memory (SRAM) or Dynamic Random Access Memory (DRAM). The databases involved in the embodiments provided in this application may include at least one type of relational database and non-relational database. Non-relational databases may include, but are not limited to, blockchain-based distributed databases. The processors involved in the embodiments provided in this application may be general-purpose processors, central processing units, graphics processing units, digital signal processors, programmable logic devices, quantum computing-based data processing logic devices, artificial intelligence (AI) processors, etc., and are not limited to these.

[0112] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this application.

[0113] The embodiments described above are merely illustrative of several implementation methods of this application, and while the descriptions are specific and detailed, they should not be construed as limiting the scope of this patent application. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of this application, and these all fall within the protection scope of this application. Therefore, the protection scope of this application should be determined by the appended claims.

Claims

1. A method for controlling the robotic arm of engineering machinery, characterized in that, The robotic arm control method of the engineering machinery includes: The system detects that the operator is pulling the end rigid bar of the robotic arm of the construction machinery, and acquires the first working condition data of the robotic arm at the current moment. Based on the first working condition data, the physical model is invoked to generate a physical coarse estimate of the end force on the end rigid bar; The observation conditions at the current moment are obtained; wherein the observation conditions are obtained by splicing together the multidimensional time series features within the previous H control cycles at the current moment, and the multidimensional time series features are obtained based on the second working condition data of the robotic arm; Using the observation conditions as prior information, the Conditional Denoising Diffusion Probability Model (DDPM) is invoked to generate residual compensation. Based on the residual compensation amount and the physical coarse estimate of the end force, the engineering machinery is controlled to drive the robotic arm to move in accordance with the operator's traction intention; The multidimensional temporal features are obtained by splicing together the purified pressure, angle, angular velocity, end force physical coarse estimation, disturbance force estimation, and pressure pulsation fundamental frequency; or, The multidimensional time series features are obtained by splicing together the purified pressure, angle, angular velocity, end force physical coarse estimation, disturbance force estimation and pressure pulsation fundamental frequency estimation; The engineering machinery further includes an admittance control module, which includes an adaptive admittance controller, and the adaptive admittance controller includes a fuzzy controller; the control of the engineering machinery based on the residual compensation amount and the physical coarse estimate of the end force includes: Within each control cycle, the residual compensation amount and the physical coarse estimate of the end force are calculated to obtain the final end force estimate; Based on the final end force estimate, the engineering machinery is controlled; The control of the engineering machinery based on the final end force estimate includes: The adaptive admittance controller controls the engineering machinery using the following formula: in, These are the admittance-inertia parameters after weighted mixing; This is the acceleration of the rigid rod at the end; These are the admittance and damping parameters after weighted mixing; The fuzzy damping increment is determined by the fuzzy controller based on... and Calculated in real time; Uncertainty—damping gain coefficient The uncertainty metric is the standard deviation of the sampling results of N independent denoised diffusion implicit model (DDIM) samplings. The velocity of the end rigid rod; This is the output after verification.

2. The robotic arm control method for engineering machinery according to claim 1, characterized in that, The first operating condition data includes pressure signal, angle signal, and pressure pulsation frequency information; The step of generating a coarse physical estimate of the end force on the end rigid bar based on the first working condition data and calling the physical model includes: Based on the pressure pulsation frequency information, the pressure signal is preprocessed to obtain the purified pressure signal and disturbance force estimation. Based on the purified pressure signal, the disturbance force estimation, and the angle signal, the physical model is invoked to generate a physical coarse estimate of the end force on the end rigid rod.

3. The robotic arm control method for engineering machinery according to claim 1, characterized in that, The method further includes: The DDPM is improved to obtain an improved DDPM; wherein, the improvement is: a perturbation-aware temporal Transformer network is provided, which includes an observation feature encoder, a perturbation-aware gating fusion module, and a denoising backbone network. The disturbance-aware gating fusion module adaptively fuses disturbance features and conditional features using the following formula: in, The fusion characteristics of the gated fusion output; For the conditional features; For learnable gating coefficients; For element-wise multiplication; The disturbance feature is obtained by encoding the disturbance force estimation. in, It is calculated using the following formula: in, Use the Sigmoid activation function; and These are the learnable weight matrix and the bias term, respectively.

4. The robotic arm control method for engineering machinery according to claim 3, characterized in that, The method further includes: The loss function of the DDPM is improved; wherein the improvement is that the loss function is a multi-objective composite loss function, which includes the following four components: diffusion principal loss, physical consistency loss, perturbation robustness loss and amplitude constraint loss; The expression for the multi-objective composite loss function is as follows: in, Total loss; For the first One component; These are learnable weights; in, The mathematical expression is as follows: in, For expectation operators; The square of the L2 norm; Gaussian noise was added to the image. To predict noise; in, The mathematical expression is as follows: in, This is the transpose of the terminal geometric Jacobian matrix; A rough physical estimate of the end force; To predict residuals; This represents the measured torque residual. in, The mathematical expression is as follows: in, For expectation operators; The predicted noise is augmented by simulated perturbations of different magnitudes applied to the training samples; in, The mathematical expression is as follows: in, It is an activation function; The absolute value of the predicted residual; This is the residual threshold.

5. The robotic arm control method for engineering machinery according to claim 1, characterized in that, During the inference phase of the DDPM, the DDIM accelerated sampling strategy is adopted to reduce the number of denoising steps compared to the training phase.

6. The robotic arm control method for engineering machinery according to claim 1, characterized in that, Before controlling the construction machinery based on the final end force estimate, the process includes: The final end force estimate is verified using the following formula: in, For the final end force estimation; This is the transpose of the terminal geometric Jacobian matrix; This represents the measured torque residual. The torque difference threshold; The threshold value is the upper limit of the end force amplitude; A rough physical estimate of the end force; exist In the case of [the situation], the step of controlling the engineering machinery based on the final end force estimate is performed.

7. A concrete pumping device, characterized in that, It includes a processor, a memory, and a program or instructions stored in the memory and executable on the processor, wherein when the program or instructions are executed by the processor, they implement the steps of the robotic arm control method for engineering machinery as described in any one of claims 1-6.

8. A readable storage medium, characterized in that, The readable storage medium stores a program or instructions that, when executed by a processor, implement the steps of the robotic arm control method for engineering machinery as described in any one of claims 1-6.

Images