Leveling and stability evaluation method and system for self-adaptive terrain of vehicle-mounted crane outriggers

Abstract

The present invention provides a method and system for adaptive terrain leveling and stability assessment of vehicle-mounted crane outriggers, which relates to the field of automatic control technology. The method comprises the following steps: using data collected by sensors to establish a ground stiffness-deformation mapping relationship, constructing a nonlinear spring-damping contact force model, and obtaining outrigger load distribution data; obtaining a stability state matrix based on the outrigger load distribution data; inputting the stability state matrix into a model predictive controller, which uses a reinforcement learning method to construct a terrain adaptability evaluation function and generates outrigger leveling control instructions through an action value network; inputting the outrigger leveling control instructions into a distributed collaborative control unit, which sets the outriggers as network nodes and performs information exchange, and uses a feedforward-feedback composite controller with adaptive gain to control the outrigger displacement.

Description

Technical Field

[0001] The present invention relates to automatic control technology, and in particular to a method and system for evaluating the leveling and stability of a vehicle-mounted crane support leg that is adaptive to terrain. Background Art

[0002] Truck-mounted cranes are an important type of lifting equipment and are widely used in operations in various complex terrain environments. The leveling and stability of their outriggers are directly related to the safety and efficiency of lifting operations. The defects and shortcomings of existing technologies are mainly reflected in the following three aspects:

[0003] First, traditional outrigger leveling methods lack the ability to adapt to complex terrain. On uneven surfaces, existing methods struggle to precisely control the extension and retraction of the outriggers, leading to significant deviations in vehicle posture, affecting operational stability and even causing rollover accidents.

[0004] Secondly, existing outrigger-ground contact modeling is overly simplified, making it difficult to accurately reflect the interaction between the outrigger and the ground. Ignoring the nonlinear relationship between ground stiffness and deformation, as well as the unevenness of the outrigger contact surface, results in a deviation between the calculated outrigger load distribution and the actual situation, affecting the accuracy of stability assessments.

[0005] Finally, existing stability assessment methods are overly simplistic and lack comprehensive consideration of dynamic loads and stress distribution. Relying solely on simple geometric analysis and empirical formulas makes it difficult to effectively predict and warn of potential hazards, leading to higher operational risks. Summary of the Invention

[0006] The embodiments of the present invention provide a method and system for adaptive terrain leveling and stability assessment of vehicle-mounted crane outriggers, which can solve the problems in the prior art.

[0007] According to a first aspect of the embodiments of the present invention,

[0008] Provides a method for evaluating the adaptive terrain leveling and stability of the truck-mounted crane outriggers, including:

[0009] The terrain elevation data, outrigger pressure data, and vehicle body posture data collected by the sensors are filtered and weighted, and then input into the outrigger ground contact modeling unit. The outrigger ground contact modeling unit uses a historical evolution predictor based on a deep neural network to classify terrain features and establish a ground stiffness-deformation mapping relationship. The outrigger ground contact modeling unit uses an online learning contact parameter adaptor to construct a nonlinear spring-damping contact force model to obtain outrigger load distribution data.

[0010] The outrigger load distribution data is input into an outrigger stability assessment unit, the outrigger comprising a stability discriminator, a stress uniformity evaluator, and a dynamic load early warning device. The stability discriminator forms a support polygon based on the outrigger landing point and calculates the support stability margin in combination with the projection position of the vehicle body's center of gravity. The stress uniformity evaluator calculates a stress uniformity index based on the outrigger pressure variance. The dynamic load early warning device outputs an early warning signal based on the dynamic characteristics of the outrigger pressure. The outputs of the stability discriminator, the stress uniformity evaluator, and the dynamic load early warning device are weighted and fused using an adaptive weight coefficient that is adjusted in real time with the vehicle body posture to obtain a stability state matrix.

[0011] The stability state matrix is ​​input into a model predictive controller, which uses a reinforcement learning method to construct a terrain adaptability evaluation function and generates a leg leveling control instruction through an action value network; the leg leveling control instruction is input into a distributed collaborative control unit, which sets the legs as network nodes and performs information exchange, and uses a feedforward feedback composite controller with adaptive gain to control the leg displacement.

[0012] In an optional embodiment,

[0013] The outrigger ground contact modeling unit uses a historical evolution predictor based on a deep neural network to classify terrain features and establish a ground stiffness-deformation mapping relationship, and uses an online learning contact parameter adaptor to construct a nonlinear spring damping contact force model. The steps of obtaining outrigger load distribution data include:

[0014] A historical evolution predictor based on a two-branch densely connected convolutional neural network is used to classify terrain features. The two-branch densely connected convolutional neural network includes a high-frequency feature extraction branch for extracting local texture features and a low-frequency feature extraction branch for extracting regional morphological features. Each feature extraction branch is equipped with an independent spatial pyramid pooling module. The dual-branch features are adaptively fused through a series of attention modules to output the terrain classification result.

[0015] A hybrid ground mechanics model is established based on the terrain classification results. A three-layer back-propagation neural network is used to establish a mapping relationship between ground stiffness and deformation to obtain ground deformation characteristics. The ground deformation characteristics are combined with the outrigger pressure sequence, displacement sequence, and terrain classification results and input into a two-layer long short-term memory network for soil mechanics parameter identification. The first layer of the two-layer long short-term memory network outputs preliminary soil parameter estimates, and the second layer introduces historical data to perform time series correction on the preliminary soil parameter estimates.

[0016] A multi-physics field coupled nonlinear spring-damper contact force model is constructed based on soil mechanics parameters. The nonlinear spring-damper contact force model includes a stiffness function related to ground moisture content, a damping function related to temperature, and a friction coefficient function related to relative velocity and pressure. A parameter adaptor based on a motion evaluation framework is used to evaluate the contact state between the outrigger and the ground and to perform online optimization of the parameters of the stiffness, damping, and friction coefficient functions.

[0017] The dynamic coupling equation of the outrigger group is established based on the nonlinear spring-damping contact force model. Under the constraints of the dynamic coupling equation of the outrigger group, a load distributor with a prediction-correction structure is used to optimize the distribution. The outrigger force signal characteristics are extracted by wavelet decomposition. The adaptive notch filter and integral sliding mode controller are combined to compensate for high-frequency vibration and low-frequency drift, and the outrigger load distribution data is output.

[0018] In an optional embodiment,

[0019] The stress uniformity evaluator calculates a stress uniformity index based on the outrigger pressure variance, and the dynamic load warning device outputs a warning signal based on the dynamic characteristics of the outrigger pressure; and the outputs of the stability discriminator, the stress uniformity evaluator, and the dynamic load warning device are weightedly integrated using an adaptive weight coefficient that is adjusted in real time with the vehicle body posture to obtain a stability state matrix, including the following steps:

[0020] The outrigger pressure variance is calculated based on the outrigger load distribution data. The terrain correction coefficient related to the terrain slope and terrain relief is introduced to correct the outrigger pressure variance. The stress uniformity index is calculated by combining the corrected outrigger pressure variance with the maximum pressure.

[0021] Multi-scale dynamic feature extraction is performed on the outrigger pressure sequence. Dynamic features are extracted through orthogonal complementary decomposition and variational mode decomposition. A time-frequency joint feature index is constructed based on the link function. When the time-frequency joint feature index exceeds an adaptive warning threshold, a dynamic load warning signal is triggered. The adaptive warning threshold is dynamically adjusted according to changes in terrain slope and hoisting mechanism rotation speed.

[0022] A weight mapping function based on the vehicle body attitude angle, height and lifting mechanism rotation speed is constructed to perform initial weight allocation on the support stability margin, stress uniformity index and dynamic load warning signal. The state sensitivity matrix is ​​obtained by calculating the rate of change of the weight to the vehicle body attitude angle, height and lifting mechanism rotation speed according to the weight mapping function. The weight coefficient is dynamically updated based on the state sensitivity matrix, the vehicle body state parameter increment and the newly calculated weight value, and the weight optimization is performed under the constraint condition that the sum of the weight coefficients is one and the difference of the weight coefficients at adjacent moments does not exceed the preset threshold. The optimized weight is weightedly fused with the support stability margin, stress uniformity index and dynamic load warning signal to obtain the stability state matrix.

[0023] In an optional embodiment,

[0024] Multi-scale dynamic feature extraction is performed on the outrigger pressure sequence. Dynamic features are extracted through orthogonal complementary decomposition and variational mode decomposition. A time-frequency joint feature index is constructed based on the link function. When the time-frequency joint feature index exceeds an adaptive warning threshold, a dynamic load warning signal is triggered. The adaptive warning threshold is dynamically adjusted according to changes in terrain slope and hoisting mechanism rotation speed, including the following steps:

[0025] Acquiring a leg pressure signal, a leg deformation, and a ground reaction force signal, constructing an interactive dynamic characteristic matrix between the leg and the ground using the leg pressure signal, the leg deformation, and the ground reaction force signal, calculating a transient component based on the interactive dynamic characteristic matrix using an orthogonal complementary decomposition algorithm, and extracting a dynamic eigenvector based on the transient component;

[0026] Performing an improved variational modal decomposition on the transient component, wherein an adaptive window function dynamically adjusted according to local characteristics of the signal is introduced into the variational modal decomposition process to obtain multiple eigenmode components and their corresponding center frequencies;

[0027] Constructing a joint distribution of the dynamic eigenvector, the intrinsic mode component, and the center frequency based on a link function, dynamically selecting a generating function of the link function according to feature correlation, and obtaining a time-frequency joint feature index by integrating the joint distribution;

[0028] A multi-objective optimization function is constructed, which includes the time-frequency joint characteristic index, terrain slope and hoisting mechanism rotation speed. The time-frequency joint characteristic index is averaged and a cumulative sequence is constructed. A grey differential equation model is established based on the cumulative sequence. A background value sequence is constructed and a parameter estimation equation is established. The parameter estimation equation is solved by the least squares method to obtain a time response function. The predicted value of the time-frequency joint characteristic index is obtained by cumulative reduction, and the grey differential equation model is updated online through a sliding time window. A fuzzy rule base is constructed based on the terrain slope, the hoisting mechanism rotation speed and the predicted time-frequency joint characteristic index. The threshold adjustment amount is obtained through fuzzy reasoning, and the warning threshold is adaptively updated by dynamically adjusting the learning rate according to the prediction error. When the time-frequency joint characteristic index exceeds the warning threshold after adaptive update, a dynamic load warning signal of the outrigger is triggered.

[0029] In an optional embodiment,

[0030] The model predictive controller uses a reinforcement learning method to construct a terrain adaptability evaluation function, and the steps of generating leg leveling control instructions through an action value network include:

[0031] A terrain adaptability evaluation function was constructed that included outrigger height deviation, roll angle deviation, pitch angle deviation, ground contact force, and outrigger sliding speed. The terrain adaptability evaluation function adopted an adaptive weight allocation strategy that was dynamically adjusted based on terrain roughness, slope standard deviation, ground friction coefficient, and outrigger ground stress distribution standard deviation. A Sigmoid function with a periodic compensation term was used during the dynamic weight adjustment process, and the outrigger ground stress distribution was corrected using a Gaussian kernel function.

[0032] An action-value network structure is designed, comprising an evaluation network and a target network. The state spaces of the evaluation network and the target network include the outrigger height vector, vehicle body attitude angle, terrain feature vector, and outrigger stress distribution characteristics, and the action space is the outrigger velocity instruction. The evaluation network uses a multi-head self-attention mechanism and a residual convolution structure for feature extraction, and calculates the value estimate of the state-action pair based on the output of the terrain adaptability evaluation function. The evaluation network is trained and optimized in the model predictive controller, and the parameters of the evaluation network are soft-updated to the target network using an adaptive update rate based on the training error.

[0033] Establishing an optimization objective function including a short-term control cost term, a control amount adjustment term, and a terminal penalty term, wherein the short-term control cost term is provided by the evaluation network, the terminal penalty term is calculated based on the long-term value estimate provided by the target network, and the control amount adjustment term includes a square term of the control increments at adjacent moments;

[0034] The optimization objective function and the leg kinematic constraints, actuator limit constraints and control increment constraints constitute an optimization problem, and the optimization problem is solved by the alternating direction multiplier method based on Nesterov acceleration. The alternating direction multiplier method decomposes the original problem into a sequence of sub-problems through dual decomposition, and the sub-problems are solved by the conjugate gradient method; the solution of the optimization problem is output as the leg leveling control instruction.

[0035] In an optional embodiment,

[0036] The steps of training and optimizing the evaluation network in the model predictive controller and soft-updating the parameters of the evaluation network to the target network through an adaptive update rate according to the training error include:

[0037] Construct a distributed experience replay pool, which uses a hierarchical storage structure. The distributed experience replay pool calculates the environment complexity index based on terrain roughness and slope standard deviation, and the task difficulty index based on the leg height error vector, roll angle error, and pitch angle error. The state-action-reward value-next-state transition samples are stored hierarchically according to the weighted combination of the environment complexity index and the task difficulty index.

[0038] A hybrid priority sampling mechanism is designed based on the distributed experience replay pool. The hybrid priority sampling mechanism includes priority calculation based on Wasserstein distance and importance sampling weight calculation based on the ratio of control decision functions. The combined weight coefficient of the priority calculation and the importance sampling weight calculation is dynamically adjusted according to the training stage, and the sampling probability of samples of different difficulty levels is adjusted according to the combined weight coefficient.

[0039] Constructing a dual-gradient optimizer, which includes a control decision gradient optimizer and an evaluation network gradient optimizer. The control decision gradient optimizer uses a proximal policy optimization algorithm with trust region constraints to calculate the control decision update based on the ratio of the current control decision function to the historical control decision function. The evaluation network gradient optimizer uses an adaptive moment estimation algorithm with momentum term to construct a loss function based on the prediction error and gradient norm of the evaluation network.

[0040] Based on the optimization results of the dual-gradient optimizer, a multi-scale adaptive learning mechanism is designed. The base learning rate is adjusted according to the temporal difference error, the variance of the control decision update amount, and the loss value of the evaluation network. The exploration noise amplitude is adaptively adjusted based on the statistical distribution characteristics of the outrigger velocity command output.

[0041] A hierarchical network update strategy is adopted, and differentiated update frequencies are designed for the feature extraction layer and the control instruction output layer of the evaluation network. The feature extraction layer updates parameters according to a preset first update period, and the control instruction output layer updates parameters according to a preset second update period, where the first update period is greater than the second update period. The parameters of the evaluation network are weighted averaged according to the verification performance and then soft-updated to the target network.

[0042] In an optional embodiment,

[0043] The distributed collaborative control unit sets the legs as network nodes and performs information exchange, and adopts a feedforward feedback composite controller with adaptive gain to control the leg displacement, including the following steps:

[0044] A distributed collaborative control unit network is constructed, with each leg corresponding to a distributed collaborative control unit as a network node. The state vector of the network node includes leg height, roll angle, pitch angle, speed, and contact force information. The leg displacement error and error change rate are constructed based on the state vector of the network node. The network communication radius and bandwidth parameters as well as the state component weights are adaptively adjusted according to the leg displacement error and error change rate to obtain a dynamic connection weight and a weighted state vector.

[0045] Based on the dynamic connection weight and weighted state vector, an exponential function of the leg displacement error is calculated to generate an adaptive trigger function. When the leg displacement error is greater than the adaptive trigger function, information interaction is triggered. The leg displacement error is used to calculate the network node confidence. Based on the network node confidence, a time-varying information fusion weight is designed. A time-varying consistency protocol that considers communication delay is constructed, and an adaptive update law of the protocol gain matrix is ​​designed to output a consistency control quantity.

[0046] A feedforward-feedback composite controller is constructed based on the consistency control quantity, the feedforward-feedback composite controller including a disturbance observer and a neural network compensator, wherein the disturbance observer estimates the disturbance during the movement of the outrigger based on the feedforward control term and the outrigger displacement error, the neural network compensator obtains a motion compensation control value by online learning the weighted state vector, and the disturbance, the motion compensation control value and the consistency control quantity are combined to construct a composite control rate;

[0047] A predictive control and fault diagnosis mechanism is designed based on the composite control rate and weighted state vector, a long short-term memory network is used to predict the state of a network node, a prediction error compensation term is designed based on the error between the predicted state of the network node and the actual state, a fault index is calculated by the residual between the weighted state vector and the predicted state of the network node, a fault is determined to have occurred when the fault index is greater than a preset threshold, and when a fault occurs, the control rate is recalculated based on the state vector of the neighboring network node and the control amount of the faulty leg is compensated.

[0048] According to a second aspect of the embodiments of the present invention,

[0049] Provides a terrain-adaptive leveling and stability assessment system for the truck-mounted crane outriggers, including:

[0050] The first unit is used to filter and weight the terrain elevation data, outrigger pressure data, and vehicle body posture data collected by the sensor, and then input them into the outrigger ground contact modeling unit. The outrigger ground contact modeling unit uses a historical evolution predictor based on a deep neural network to classify terrain features and establish a ground stiffness-deformation mapping relationship. The outrigger ground contact modeling unit uses an online learning contact parameter adaptor to construct a nonlinear spring-damping contact force model to obtain outrigger load distribution data;

[0051] a second unit configured to input the outrigger load distribution data into an outrigger stability assessment unit, the stability assessment unit comprising a stability discriminator, a stress uniformity evaluator, and a dynamic load early warning device; the stability discriminator forming a support polygon based on the outrigger landing point and calculating the support stability margin in combination with the projection position of the vehicle body's center of gravity; the stress uniformity evaluator calculating a stress uniformity index based on the outrigger pressure variance; and the dynamic load early warning device outputting an early warning signal based on a dynamic characteristic analysis of the outrigger pressure; and the outputs of the stability discriminator, the stress uniformity evaluator, and the dynamic load early warning device being weightedly fused using an adaptive weight coefficient adjusted in real time with the vehicle body posture to obtain a stability state matrix;

[0052] The third unit is used to input the stability state matrix into a model predictive controller, the model predictive controller uses a reinforcement learning method to construct a terrain adaptability evaluation function, and generates a leg leveling control instruction through an action value network; the leg leveling control instruction is input into a distributed collaborative control unit, the distributed collaborative control unit sets the legs as network nodes and performs information exchange, and uses a feedforward feedback composite controller with adaptive gain to control the leg displacement.

[0053] According to a third aspect of the embodiments of the present invention,

[0054] An electronic device is provided, comprising:

[0055] processor;

[0056] a memory for storing processor-executable instructions;

[0057] The processor is configured to call the instructions stored in the memory to execute the aforementioned method.

[0058] According to a fourth aspect of the embodiments of the present invention,

[0059] A computer-readable storage medium is provided, on which computer program instructions are stored. When the computer program instructions are executed by a processor, the method described above is implemented.

[0060] The present invention can effectively avoid overturning or rollover accidents caused by uneven terrain and ensure operational safety by accurately modeling the terrain and combining it with outrigger load distribution and stability assessment.

[0061] The adaptive terrain leveling function of the present invention can quickly adapt to various terrains, reduce the time and energy of manual adjustment of the legs, and improve work efficiency.

[0062] The control strategy based on deep neural networks and reinforcement learning adopted in the present invention enables the truck crane to maintain stable and reliable operation under various complex and changing terrain conditions, enhancing its adaptability and reliability. BRIEF DESCRIPTION OF THE DRAWINGS

[0063] Figure 1 Schematic diagram of the flow of a method for adaptive terrain leveling and stability assessment of a vehicle-mounted crane outrigger according to an embodiment of the present invention;

[0064] Figure 2 The figure is a schematic structural diagram of a system for adaptive terrain leveling and stability assessment of vehicle-mounted crane legs according to an embodiment of the present invention. DETAILED DESCRIPTION

[0065] To make the objectives, technical solutions, and advantages of the embodiments of the present invention more clear, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts shall fall within the scope of protection of the present invention.

[0066] The following specific embodiments are used to describe the technical solution of the present invention in detail. The following specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described in detail in some embodiments.

[0067] Figure 1 FIG. 1 is a flow chart of a method for adaptively leveling and evaluating the stability of a vehicle-mounted crane support leg according to an embodiment of the present invention. Figure 1 As shown, the method includes:

[0068] S1. Terrain elevation data, outrigger pressure data, and vehicle body posture data collected by sensors are filtered and weighted, then input into an outrigger ground contact modeling unit. The outrigger ground contact modeling unit uses a deep neural network-based historical evolution predictor to classify terrain features and establish a ground stiffness-deformation mapping relationship. A nonlinear spring-damper contact force model is constructed using an online learning contact parameter adaptor to obtain outrigger load distribution data.

[0069] S2. Inputting the outrigger load distribution data into an outrigger stability assessment unit, the stability assessment unit comprising a stability discriminator, a stress uniformity evaluator, and a dynamic load warning device. The stability discriminator forms a support polygon based on the outrigger landing point and calculates the support stability margin in combination with the projected position of the vehicle body's center of gravity. The stress uniformity evaluator calculates a stress uniformity index based on the outrigger pressure variance. The dynamic load warning device outputs a warning signal based on the dynamic characteristics of the outrigger pressure. The outputs of the stability discriminator, stress uniformity evaluator, and dynamic load warning device are weighted and fused using adaptive weight coefficients adjusted in real time with the vehicle body posture to produce a stability state matrix.

[0070] S3. Input the stability state matrix into the model predictive controller, which uses the reinforcement learning method to construct a terrain adaptability evaluation function and generates leg leveling control instructions through the action value network; input the leg leveling control instructions into the distributed collaborative control unit, which sets the legs as network nodes and conducts information exchange, and uses a feedforward feedback composite controller with adaptive gain to control the leg displacement.

[0071] In an optional embodiment,

[0072] The outrigger ground contact modeling unit uses a historical evolution predictor based on a deep neural network to classify terrain features and establish a ground stiffness-deformation mapping relationship, and uses an online learning contact parameter adaptor to construct a nonlinear spring damping contact force model. The steps of obtaining outrigger load distribution data include:

[0073] A historical evolution predictor based on a two-branch densely connected convolutional neural network is used to classify terrain features. The two-branch densely connected convolutional neural network includes a high-frequency feature extraction branch for extracting local texture features and a low-frequency feature extraction branch for extracting regional morphological features. Each feature extraction branch is equipped with an independent spatial pyramid pooling module. The dual-branch features are adaptively fused through a series of attention modules to output the terrain classification result.

[0074] A hybrid ground mechanics model is established based on the terrain classification results. A three-layer back-propagation neural network is used to establish a mapping relationship between ground stiffness and deformation to obtain ground deformation characteristics. The ground deformation characteristics are combined with the outrigger pressure sequence, displacement sequence, and terrain classification results and input into a two-layer long short-term memory network for soil mechanics parameter identification. The first layer of the two-layer long short-term memory network outputs preliminary soil parameter estimates, and the second layer introduces historical data to perform time series correction on the preliminary soil parameter estimates.

[0075] A multi-physics field coupled nonlinear spring-damper contact force model is constructed based on soil mechanics parameters. The nonlinear spring-damper contact force model includes a stiffness function related to ground moisture content, a damping function related to temperature, and a friction coefficient function related to relative velocity and pressure. A parameter adaptor based on a motion evaluation framework is used to evaluate the contact state between the outrigger and the ground and to perform online optimization of the parameters of the stiffness, damping, and friction coefficient functions.

[0076] The dynamic coupling equation of the outrigger group is established based on the nonlinear spring-damping contact force model. Under the constraints of the dynamic coupling equation of the outrigger group, a load distributor with a prediction-correction structure is used to optimize the distribution. The outrigger force signal characteristics are extracted by wavelet decomposition. The adaptive notch filter and integral sliding mode controller are combined to compensate for high-frequency vibration and low-frequency drift, and the outrigger load distribution data is output.

[0077] For example, a two-branch densely connected convolutional neural network is first used to classify terrain features. The high-frequency feature extraction branch is responsible for extracting local terrain texture features, such as fine cracks and particle size; the low-frequency feature extraction branch is responsible for extracting regional morphological features, such as overall slope and undulation. Each branch is equipped with an independent spatial pyramid pooling module to integrate features at different scales. The features extracted by the two branches are adaptively fused through a series of attention modules, ultimately outputting the terrain classification results, such as different types of terrain: sand, clay, gravel, etc.

[0078] To illustrate this more clearly, let's assume we use a set of image data containing different terrain types, such as sand, clay, and gravel, for training. The training data includes the images themselves and the corresponding terrain type labels. Once the network learns the characteristics of different terrain types, it can accurately classify new terrain images. For example, if a new image is input to the network and the network determines that it is sand, it will output a sand type label.

[0079] Based on the terrain classification results, a hybrid ground mechanics model was established. A three-layer back-propagation neural network was used to map ground stiffness to deformation. This neural network uses terrain type as one of its input features and combines it with deformation to predict ground stiffness. This training data includes deformation and corresponding stiffness values ​​for different terrain types. Once the network learns the relationship between deformation and stiffness for different terrain types, it can predict stiffness based on the deformation. For example, given a sand type and a deformation value, the network can predict the corresponding stiffness value for the sand.

[0080] Ground deformation characteristics, outrigger pressure sequences, displacement sequences, and terrain classification results are input into a two-layer long short-term memory (LSTM) network for soil mechanical parameter identification. The first-layer LSTM network outputs preliminary estimates of soil parameters, such as Young's modulus and Poisson's ratio. The second-layer LSTM network incorporates historical data and performs time-series correction on the preliminary soil parameter estimates output by the first layer, improving the accuracy of parameter estimation.

[0081] Assume that a series of outrigger pressures, displacements, and terrain classification results have been obtained over a period of time. This data is fed into the first-layer LSTM network, which outputs preliminary estimates of soil parameters. These preliminary estimates, along with historical data, are then fed into the second-layer LSTM network, which outputs corrected soil parameter values. These corrected soil parameter values ​​more accurately reflect the mechanical properties of the soil.

[0082] Based on the identified soil mechanical parameters, a multi-physics coupled nonlinear spring-damper contact force model is constructed. This model includes a stiffness function related to ground moisture content, a damping function related to temperature, and a friction coefficient function related to relative velocity and pressure. A parameter adaptor based on a motion evaluation framework is used to evaluate the contact state between the outrigger and the ground and perform online optimization of the parameters of the stiffness, damping, and friction coefficient functions. For example, if outrigger slip is detected, the parameters of the friction coefficient function are adjusted to more accurately simulate the sliding process.

[0083] Assume that initial soil parameters are set in the contact force model. During the contact between the outrigger and the ground, the parameter adaptor adjusts model parameters such as stiffness, damping, and friction coefficient based on real-time monitoring of the contact state, such as the outrigger's displacement, velocity, and pressure. If a change in the contact state is detected, such as from static contact to sliding, the adaptor adjusts the parameters to make the model more accurately reflect the actual situation.

[0084] Based on a nonlinear spring-damper contact force model, a dynamic coupling equation for the outrigger group is established. Under the constraints of this equation, a predictor-corrector load distributor is employed to optimize load distribution. Wavelet decomposition is used to extract outrigger force signal characteristics. High-frequency vibration and low-frequency drift are compensated for using an adaptive notch filter and an integral sliding mode controller. Ultimately, outrigger load distribution data is output.

[0085] Assuming there are four outriggers, the load distributor calculates the load that each leg should bear based on the established dynamic coupling equations. Wavelet decomposition removes the effects of high-frequency vibration and low-frequency drift, improving the accuracy of the results. Finally, the load magnitude and direction for each leg are output.

[0086] Through deep learning models and parameter adaptive algorithms, the present invention can more accurately predict terrain features, establish a ground stiffness-deformation mapping relationship, and construct an accurate contact force model, ultimately improving the calculation accuracy of the leg load distribution; the online learning contact parameter adapter can adjust the model parameters according to the real-time contact state, enhance the adaptability and robustness of the model, and improve the applicability of the model under different terrains and working conditions; the use of an efficient deep learning algorithm and a load distributor with a prediction-correction structure can reduce the computational complexity while ensuring the calculation accuracy and improve the calculation efficiency.

[0087] In an optional embodiment,

[0088] The stress uniformity evaluator calculates a stress uniformity index based on the outrigger pressure variance, and the dynamic load warning device outputs a warning signal based on the dynamic characteristics of the outrigger pressure; and the outputs of the stability discriminator, the stress uniformity evaluator, and the dynamic load warning device are weightedly integrated using an adaptive weight coefficient that is adjusted in real time with the vehicle body posture to obtain a stability state matrix, including the following steps:

[0089] The outrigger pressure variance is calculated based on the outrigger load distribution data. The terrain correction coefficient related to the terrain slope and terrain relief is introduced to correct the outrigger pressure variance. The stress uniformity index is calculated by combining the corrected outrigger pressure variance with the maximum pressure.

[0090] Multi-scale dynamic feature extraction is performed on the outrigger pressure sequence. Dynamic features are extracted through orthogonal complementary decomposition and variational mode decomposition. A time-frequency joint feature index is constructed based on the link function. When the time-frequency joint feature index exceeds an adaptive warning threshold, a dynamic load warning signal is triggered. The adaptive warning threshold is dynamically adjusted according to changes in terrain slope and hoisting mechanism rotation speed.

[0091] A weight mapping function based on the vehicle body attitude angle, height and lifting mechanism rotation speed is constructed to perform initial weight allocation on the support stability margin, stress uniformity index and dynamic load warning signal. The state sensitivity matrix is ​​obtained by calculating the rate of change of the weight to the vehicle body attitude angle, height and lifting mechanism rotation speed according to the weight mapping function. The weight coefficient is dynamically updated based on the state sensitivity matrix, the vehicle body state parameter increment and the newly calculated weight value, and the weight optimization is performed under the constraint condition that the sum of the weight coefficients is one and the difference of the weight coefficients at adjacent moments does not exceed the preset threshold. The optimized weight is weightedly fused with the support stability margin, stress uniformity index and dynamic load warning signal to obtain the stability state matrix.

[0092] For example, first, the load data of the crane legs is collected, including the pressure value of each leg. Then, based on the structural parameters and load distribution of the crane, the stability margin of each leg is calculated. The stability margin refers to the difference between the load actually borne by the leg and its ultimate load-bearing capacity, which reflects the degree of stability of the leg. The stability margin calculation takes into account the influence of factors such as the inclination angle and height of the crane body and the extension length of the boom. For example, if the pressures of the four legs are 10 tons, 12 tons, 11 tons, and 13 tons respectively, and the ultimate load-bearing capacity of each leg is 20 tons, then the stability margin of each leg is 10 tons, 8 tons, 9 tons, and 7 tons respectively. These data can be stored in a database for use in subsequent steps.

[0093] Crane outrigger pressure data is collected and the outrigger pressure variance is calculated. The variance reflects the uniformity of the outrigger pressure distribution. The smaller the variance, the more uniform the outrigger pressure distribution. To more accurately reflect actual conditions, a terrain correction factor is introduced. This factor is related to the terrain slope and terrain relief. The greater the terrain slope and relief, the larger the terrain correction factor, and the greater the correction to the outrigger pressure variance. Assuming the calculated outrigger pressure variance is 2 and the terrain correction factor is 1.2, the corrected outrigger pressure variance is 2.4. Finally, the stress uniformity index is calculated by combining the corrected outrigger pressure variance and the maximum outrigger pressure. The closer the stress uniformity index is to 1, the more uniform the outrigger pressure distribution and the more stable the crane. For example, if the corrected outrigger pressure variance is 2.4 and the maximum pressure is 13 tons, the corresponding stress uniformity index, such as 0.8, can be calculated using a pre-set formula.

[0094] Multi-scale dynamic feature extraction is performed on the outrigger pressure sequence, using orthogonal complementary decomposition and variational mode decomposition methods to extract dynamic features such as frequency and amplitude. Based on these features, a joint time-frequency feature index is then constructed. When this index exceeds a preset adaptive warning threshold, a dynamic load warning signal is triggered. The adaptive warning threshold is dynamically adjusted based on the terrain slope and the lifting mechanism's rotation speed. For example, if the terrain slope is steep and the lifting mechanism's rotation speed is fast, the adaptive warning threshold will be increased accordingly. This processed data is also stored in the database.

[0095] A weight mapping function based on the vehicle body attitude angle, height, and hoisting mechanism rotation speed is constructed to initially assign weights to the support stability margin, stress uniformity index, and dynamic load warning signal. For example, on flat terrain, when operating at low speeds, the support stability margin weight can be set higher, while the weights of the other two indicators are relatively low. Then, the rate of change of the weight with respect to the vehicle body attitude angle, height, and hoisting mechanism rotation speed is calculated according to the weight mapping function to obtain the state sensitivity matrix. The weight coefficients are dynamically updated based on the state sensitivity matrix, the increment of the vehicle body state parameters, and the newly calculated weight values. During the weight update process, it is necessary to ensure that the sum of all weight coefficients is one, and the difference in weight coefficients at adjacent moments does not exceed the preset threshold to ensure the smoothness of the weight adjustment.

[0096] The optimized weights are then combined with the support stability margin, stress uniformity index, and dynamic load warning signals to create a final stability state matrix. This matrix reflects the crane's current stability status and can be used for real-time monitoring and early warning. For example, if the value in the stability state matrix falls below a preset threshold, the system issues an alarm, alerting the operator to safety.

[0097] By integrating multiple indicators, the present invention can more comprehensively evaluate crane stability, effectively reduce the probability of lifting accidents, monitor the crane status in real time, issue timely warnings, avoid operation interruptions due to stability problems, and improve work efficiency; dynamically adjust the weight coefficient to adapt to different working conditions and improve the reliability of stability assessment.

[0098] In an optional embodiment,

[0099] Multi-scale dynamic feature extraction is performed on the outrigger pressure sequence. Dynamic features are extracted through orthogonal complementary decomposition and variational mode decomposition. A time-frequency joint feature index is constructed based on the link function. When the time-frequency joint feature index exceeds an adaptive warning threshold, a dynamic load warning signal is triggered. The adaptive warning threshold is dynamically adjusted according to changes in terrain slope and hoisting mechanism rotation speed, including the following steps:

[0100] Acquiring a leg pressure signal, a leg deformation, and a ground reaction force signal, constructing an interactive dynamic characteristic matrix between the leg and the ground using the leg pressure signal, the leg deformation, and the ground reaction force signal, calculating a transient component based on the interactive dynamic characteristic matrix using an orthogonal complementary decomposition algorithm, and extracting a dynamic eigenvector based on the transient component;

[0101] Performing an improved variational modal decomposition on the transient component, wherein an adaptive window function dynamically adjusted according to local characteristics of the signal is introduced into the variational modal decomposition process to obtain multiple eigenmode components and their corresponding center frequencies;

[0102] Constructing a joint distribution of the dynamic eigenvector, the intrinsic mode component, and the center frequency based on a link function, dynamically selecting a generating function of the link function according to feature correlation, and obtaining a time-frequency joint feature index by integrating the joint distribution;

[0103] A multi-objective optimization function is constructed, which includes the time-frequency joint characteristic index, terrain slope and hoisting mechanism rotation speed. The time-frequency joint characteristic index is averaged and a cumulative sequence is constructed. A grey differential equation model is established based on the cumulative sequence. A background value sequence is constructed and a parameter estimation equation is established. The parameter estimation equation is solved by the least squares method to obtain a time response function. The predicted value of the time-frequency joint characteristic index is obtained by cumulative reduction, and the grey differential equation model is updated online through a sliding time window. A fuzzy rule base is constructed based on the terrain slope, the hoisting mechanism rotation speed and the predicted time-frequency joint characteristic index. The threshold adjustment amount is obtained through fuzzy reasoning, and the warning threshold is adaptively updated by dynamically adjusting the learning rate according to the prediction error. When the time-frequency joint characteristic index exceeds the warning threshold after adaptive update, a dynamic load warning signal of the outrigger is triggered.

[0104] For example, first, sensors are used to obtain outrigger pressure signals, outrigger deformation signals, and ground reaction force signals during crane operation. These signals are typically obtained through sensors installed on the outriggers and the ground, such as pressure sensors and displacement sensors. The sensor sampling frequency must be determined based on actual conditions to ensure that rapid changes in the dynamic load on the outriggers can be captured. For example, the sampling frequency can be set to 1000 Hz. The collected raw signals may contain noise and require preprocessing, such as filtering, to remove the effects of noise. Preprocessing methods can include wavelet filtering or mean filtering.

[0105] Next, the collected outrigger pressure signals, outrigger deformation signals, and ground reaction force signals are integrated. The values ​​of these signals at the same time point are combined into a vector, and these vectors are arranged in chronological order to construct the interaction dynamic characteristic matrix between the outrigger and the ground.

[0106] The interaction dynamics characteristic matrix is ​​decomposed into a trend component and a transient component using the orthogonal complementary decomposition algorithm. This algorithm is a nonlinear signal processing method that effectively separates the different components in the signal. The transient component represents the rapidly changing portion of the signal and contains important information about the dynamic loads on the outriggers. The transient component is of particular interest because it is directly related to the dynamic loads.

[0107] Next, a dynamic eigenvector is extracted from the extracted transient component. This dynamic eigenvector can include multiple features, such as mean, variance, peak value, kurtosis, and form factor, which can reflect the characteristics of the outrigger dynamic load. Feature extraction methods can be selected based on practical needs. For example, the energy or spectral entropy of the transient component can be calculated. Ultimately, a dynamic eigenvector containing multiple eigenvalues ​​is obtained, which represent certain characteristics of the outrigger dynamic load.

[0108] An improved variational modal decomposition (VMD) is performed on the resulting transient components. This improvement involves the introduction of an adaptive window function during the VMD process. The adaptive window function dynamically adjusts its size based on the local characteristics of the signal, better adapting to the non-stationary nature of the signal. For example, when the signal varies dramatically, a smaller window is used; when the signal varies gently, a larger window is used. This allows for more accurate signal decomposition, yielding multiple intrinsic mode components and their corresponding center frequencies.

[0109] The extracted dynamic eigenvectors are combined with the derived intrinsic mode components and their center frequencies using a copula function. The choice of copula function depends on the correlation between the features. If some features are highly correlated, a copula function that reflects this correlation can be selected. For example, an appropriate copula function can be selected based on the correlation coefficient between the features, such as a simple weighted average or a more complex nonlinear function. The joint distribution is then integrated to obtain the final time-frequency joint feature index.

[0110] A multi-objective optimization function was constructed, which included the time-frequency joint characteristic index, terrain slope, and hoisting mechanism slewing speed. The time-frequency joint characteristic index was averaged and a cumulative sequence was constructed. Based on this cumulative sequence, a grey differential equation model was established to predict the future time-frequency joint characteristic index. The parameter estimation equation was solved using the least squares method to obtain the time response function. The predicted value of the time-frequency joint characteristic index was obtained using cumulative reduction. The grey differential equation model was then updated online using a sliding time window.

[0111] A fuzzy rule base is constructed using the combined time-frequency characteristics of terrain slope, hoisting mechanism slewing speed, and predictions. Fuzzy inference is used to determine the threshold adjustment value, and the learning rate is dynamically adjusted based on the prediction error to adaptively update the warning threshold. For example, if the terrain slope is steep, the warning threshold should be increased accordingly; if the hoisting mechanism slews at a high speed, the warning threshold should also be increased accordingly.

[0112] When the calculated time-frequency joint characteristic index exceeds the adaptively updated warning threshold, the outrigger dynamic load warning signal is triggered.

[0113] The multi-scale dynamic feature extraction and adaptive warning threshold adjustment adopted by the present invention improve the accuracy of warnings and reduce the possibility of false alarms and missed alarms; the online updated grey differential equation model and adaptive warning threshold mechanism ensure the real-time nature of warnings and provide operators with timely and effective warning information; the system can adapt to different terrain slopes and lifting mechanism rotation speeds, enhance the robustness of the system, and improve the reliability and stability of the system.

[0114] In an optional embodiment,

[0115] The model predictive controller uses a reinforcement learning method to construct a terrain adaptability evaluation function, and the steps of generating leg leveling control instructions through an action value network include:

[0116] A terrain adaptability evaluation function was constructed that included outrigger height deviation, roll angle deviation, pitch angle deviation, ground contact force, and outrigger sliding speed. The terrain adaptability evaluation function adopted an adaptive weight allocation strategy that was dynamically adjusted based on terrain roughness, slope standard deviation, ground friction coefficient, and outrigger ground stress distribution standard deviation. A Sigmoid function with a periodic compensation term was used during the dynamic weight adjustment process, and the outrigger ground stress distribution was corrected using a Gaussian kernel function.

[0117] An action-value network structure is designed, comprising an evaluation network and a target network. The state spaces of the evaluation network and the target network include the outrigger height vector, vehicle body attitude angle, terrain feature vector, and outrigger stress distribution characteristics, and the action space is the outrigger velocity instruction. The evaluation network uses a multi-head self-attention mechanism and a residual convolution structure for feature extraction, and calculates the value estimate of the state-action pair based on the output of the terrain adaptability evaluation function. The evaluation network is trained and optimized in the model predictive controller, and the parameters of the evaluation network are soft-updated to the target network using an adaptive update rate based on the training error.

[0118] Establishing an optimization objective function including a short-term control cost term, a control amount adjustment term, and a terminal penalty term, wherein the short-term control cost term is provided by the evaluation network, the terminal penalty term is calculated based on the long-term value estimate provided by the target network, and the control amount adjustment term includes a square term of the control increments at adjacent moments;

[0119] The optimization objective function and the leg kinematic constraints, actuator limit constraints and control increment constraints constitute an optimization problem, and the optimization problem is solved by the alternating direction multiplier method based on Nesterov acceleration. The alternating direction multiplier method decomposes the original problem into a sequence of sub-problems through dual decomposition, and the sub-problems are solved by the conjugate gradient method; the solution of the optimization problem is output as the leg leveling control instruction.

[0120] For example, an evaluation function is constructed to assess the vehicle's terrain adaptability. This function considers several key factors: outrigger height deviation, roll angle deviation, pitch angle deviation, ground contact force, and outrigger slip velocity. To adapt the evaluation function to varying terrain conditions, an adaptive weighting strategy is employed. This strategy dynamically adjusts the weights of various factors based on real-time terrain information. Terrain information includes terrain roughness, slope standard deviation, ground friction coefficient, and outrigger ground stress distribution standard deviation.

[0121] Specifically, the system first acquires vehicle sensor data and calculates the aforementioned terrain information and vehicle state parameters. Next, a modified Sigmoid function (with a periodic compensation term added to account for periodic terrain variations) dynamically adjusts each weight. This Sigmoid function takes terrain information as input and outputs a weight value ranging from 0 to 1. The periodic compensation term is introduced to better handle periodic terrain variations, such as wavy roads. Finally, a Gaussian kernel function is used to smooth the ground stress distribution to reduce the impact of measurement noise.

[0122] For example, assuming the initial weights are: outrigger height deviation 0.5, roll angle deviation 0.2, pitch angle deviation 0.2, ground contact force 0.05, and outrigger slip velocity 0.05. On a relatively flat road (low terrain roughness, low slope standard deviation, high friction coefficient, and low standard deviation of the outrigger stress distribution), the outrigger height deviation weight may be slightly reduced, while the ground contact force weight may be slightly increased to ensure vehicle stability. Conversely, on a bumpy road (high terrain roughness, high slope standard deviation, low friction coefficient, and high standard deviation of the outrigger stress distribution), the outrigger height deviation weight may be increased to ensure the vehicle can adapt to uneven terrain.

[0123] Design an action-value network consisting of an evaluation network and a target network. The state space of the evaluation network and the target network includes: outrigger height vectors (the height of each outrigger), vehicle attitude angles (roll and pitch angles), terrain feature vectors (terrain roughness, slope standard deviation, ground friction coefficient, standard deviation of outrigger stress distribution), and outrigger stress distribution characteristics (e.g., mean and variance of outrigger stress distribution). The action space is the outrigger velocity command (the speed command for each outrigger).

[0124] The evaluation network uses a multi-head self-attention mechanism to capture the complex relationships between features in the state space and a residual convolutional structure to extract features. Finally, it combines the output of the terrain adaptability evaluation function to calculate the value estimate of the state-action pair. The target network has the same structure as the evaluation network, and its parameters are updated from the evaluation network via a soft update mechanism. This soft update mechanism uses an adaptive update rate, adjusting the update amplitude based on training error to improve training efficiency and stability.

[0125] Construct an optimization objective function consisting of a short-term control cost, a control amount adjustment term, and a terminal penalty term. The short-term control cost term, provided by the evaluation network, represents the cost of taking a specific action in the current state. The terminal penalty term, provided by the target network, represents a long-term value estimate and guides the controller to select actions that will yield long-term benefits. The control amount adjustment term penalizes changes in control increments between adjacent moments to ensure smooth control instructions and prevent severe control oscillation.

[0126] The optimization objective function is combined with outrigger kinematic constraints, such as outrigger travel limits, actuator position constraints, such as motor speed limits, and control increment constraints, such as control increment limits, to form a constrained optimization problem. This problem is solved using the alternating direction method of multipliers (ADMM) based on Nesterov acceleration. The ADMM method decomposes the original problem into multiple subproblems through dual decomposition, and then solves each subproblem using the conjugate gradient method. The final solution is the outrigger leveling control command. The resulting outrigger leveling control command is sent to the actuator to control the outrigger motion and achieve vehicle leveling.

[0127] The present invention can effectively cope with various terrain conditions through adaptive weight distribution strategy and multi-factor evaluation function; by optimizing objective function and constraint conditions, it can ensure that the vehicle is stable and steady during the leveling process, avoid dangerous situations such as violent shaking or overturning, and improve vehicle safety; based on reinforcement learning action value network and efficient optimization algorithm, it can quickly and accurately calculate the outrigger leveling control instructions, thereby realizing precise leveling control and improving control efficiency and accuracy.

[0128] In an optional embodiment,

[0129] The steps of training and optimizing the evaluation network in the model predictive controller and soft-updating the parameters of the evaluation network to the target network through an adaptive update rate according to the training error include:

[0130] Construct a distributed experience replay pool, which uses a hierarchical storage structure. The distributed experience replay pool calculates the environment complexity index based on terrain roughness and slope standard deviation, and the task difficulty index based on the leg height error vector, roll angle error, and pitch angle error. The state-action-reward value-next-state transition samples are stored hierarchically according to the weighted combination of the environment complexity index and the task difficulty index.

[0131] A hybrid priority sampling mechanism is designed based on the distributed experience replay pool. The hybrid priority sampling mechanism includes priority calculation based on Wasserstein distance and importance sampling weight calculation based on the ratio of control decision functions. The combined weight coefficient of the priority calculation and the importance sampling weight calculation is dynamically adjusted according to the training stage, and the sampling probability of samples of different difficulty levels is adjusted according to the combined weight coefficient.

[0132] Constructing a dual-gradient optimizer, which includes a control decision gradient optimizer and an evaluation network gradient optimizer. The control decision gradient optimizer uses a proximal policy optimization algorithm with trust region constraints to calculate the control decision update based on the ratio of the current control decision function to the historical control decision function. The evaluation network gradient optimizer uses an adaptive moment estimation algorithm with momentum term to construct a loss function based on the prediction error and gradient norm of the evaluation network.

[0133] Based on the optimization results of the dual-gradient optimizer, a multi-scale adaptive learning mechanism is designed. The base learning rate is adjusted according to the temporal difference error, the variance of the control decision update amount, and the loss value of the evaluation network. The exploration noise amplitude is adaptively adjusted based on the statistical distribution characteristics of the outrigger velocity command output.

[0134] A hierarchical network update strategy is adopted, and differentiated update frequencies are designed for the feature extraction layer and the control instruction output layer of the evaluation network. The feature extraction layer updates parameters according to a preset first update period, and the control instruction output layer updates parameters according to a preset second update period, where the first update period is greater than the second update period. The parameters of the evaluation network are weighted averaged according to the verification performance and then soft-updated to the target network.

[0135] Exemplarily, a hierarchically stored experience replay pool is constructed to store state-action-reward-next-state transition samples. First, define the environment complexity index and the task difficulty index. The environment complexity index is calculated based on terrain data, for example, a value is obtained by calculating the terrain roughness and slope standard deviation. The higher the value, the more complex the environment. The task difficulty index is calculated based on the robot control error. For example, a value is obtained by calculating the leg height error, roll angle error, and pitch angle error. The higher the value, the more difficult the task. Then, based on the weighted combination of the environment complexity index and the task difficulty index, the samples are stored in different levels of the replay pool. For example, the environment complexity index and the task difficulty index can be assigned weights of 0.6 and 0.4 respectively, and the weighted sum is calculated as the basis for grading. The weights can be adjusted according to actual conditions.

[0136] Based on the constructed distributed experience replay pool, a hybrid priority sampling mechanism is designed. This mechanism combines priority calculation based on the Wasserstein distance with importance sampling weight calculation based on the control decision function ratio. The Wasserstein distance measures the similarity between samples; smaller distances indicate higher priorities. The control decision function ratio reflects the difference between the current strategy and the historical strategy; larger ratios indicate higher importance sampling weights. In the early stages of training, sampling relies more heavily on the Wasserstein distance to avoid being dominated by a small number of high-reward samples too early. In the later stages of training, greater reliance is placed on importance sampling weights to improve sampling efficiency. The combined weight coefficients are dynamically adjusted over the training phase. For example, in the early stages, the Wasserstein distance weight is 0.8, and the importance sampling weight is 0.2; in the later stages, the reverse is true.

[0137] A dual-gradient optimizer is constructed, consisting of a control decision gradient optimizer and an evaluation network gradient optimizer. The control decision gradient optimizer uses a proximal policy optimization algorithm with trust region constraints to update the control decision by calculating the ratio of the current control decision function to the historical control decision function. The evaluation network gradient optimizer uses an adaptive moment estimation algorithm with momentum to construct a loss function based on the evaluation network's prediction error and gradient norm. Trust region constraints are used to limit the step size of each update to ensure the stability of the optimization process.

[0138] The base learning rate is dynamically adjusted based on the temporal difference error (TDE), the variance of the control decision updates, and the loss of the evaluation network. A larger TDE results in a smaller learning rate; a larger variance of the control decision updates results in a smaller learning rate; and a larger loss of the evaluation network results in a smaller learning rate. Furthermore, the exploration noise amplitude is adaptively adjusted based on the statistical distribution of the outrigger velocity command outputs. If the command outputs are too concentrated, the exploration noise amplitude is increased; otherwise, it is decreased.

[0139] For example, if the temporal difference error is large for multiple consecutive training steps, reduce the base learning rate to 50% of the original value; if the loss value of the evaluation network continues to decrease, consider increasing the learning rate.

[0140] The evaluation network's feature extraction layer and control output layer use different update frequencies. The feature extraction layer updates parameters at a lower frequency, for example, every 100 iterations, while the control output layer updates parameters at a higher frequency, for example, every iteration. Finally, the evaluation network's parameters are weighted averaged based on validation performance and then soft-updated to the target network. For example, parameters with better performance on the validation set are given a higher weight.

[0141] Through hierarchical storage of samples, mixed priority sampling, and dual-gradient optimizer, the controller of the present invention can better learn control strategies under different complexities and difficulties, thereby improving control accuracy and stability; the multi-scale adaptive learning mechanism and hierarchical network update strategy make the training process more efficient, fully utilize sample information, and avoid resource waste; the adaptive adjustment of exploration noise amplitude and learning rate enables the controller to better adapt to environmental changes and disturbances, thereby enhancing robustness.

[0142] In an optional embodiment,

[0143] The distributed collaborative control unit sets the legs as network nodes and performs information exchange, and adopts a feedforward feedback composite controller with adaptive gain to control the leg displacement, including the following steps:

[0144] A distributed collaborative control unit network is constructed, with each leg corresponding to a distributed collaborative control unit as a network node. The state vector of the network node includes leg height, roll angle, pitch angle, speed, and contact force information. The leg displacement error and error change rate are constructed based on the state vector of the network node. The network communication radius and bandwidth parameters as well as the state component weights are adaptively adjusted according to the leg displacement error and error change rate to obtain a dynamic connection weight and a weighted state vector.

[0145] Based on the dynamic connection weight and weighted state vector, an exponential function of the leg displacement error is calculated to generate an adaptive trigger function. When the leg displacement error is greater than the adaptive trigger function, information interaction is triggered. The leg displacement error is used to calculate the network node confidence. Based on the network node confidence, a time-varying information fusion weight is designed. A time-varying consistency protocol that considers communication delay is constructed, and an adaptive update law of the protocol gain matrix is ​​designed to output a consistency control quantity.

[0146] A feedforward-feedback composite controller is constructed based on the consistency control quantity, the feedforward-feedback composite controller including a disturbance observer and a neural network compensator, wherein the disturbance observer estimates the disturbance during the movement of the outrigger based on the feedforward control term and the outrigger displacement error, the neural network compensator obtains a motion compensation control value by online learning the weighted state vector, and the disturbance, the motion compensation control value and the consistency control quantity are combined to construct a composite control rate;

[0147] A predictive control and fault diagnosis mechanism is designed based on the composite control rate and weighted state vector, a long short-term memory network is used to predict the state of a network node, a prediction error compensation term is designed based on the error between the predicted state of the network node and the actual state, a fault index is calculated by the residual between the weighted state vector and the predicted state of the network node, a fault is determined to have occurred when the fault index is greater than a preset threshold, and when a fault occurs, the control rate is recalculated based on the state vector of the neighboring network node and the control amount of the faulty leg is compensated.

[0148] For example, first, a distributed collaborative control unit network is constructed. Each leg corresponds to a control unit, acting as a node in the network. Each node needs to collect and update its own state information in real time, including the leg's current height, roll angle, pitch angle, vertical velocity, and contact force with the ground. This data constitutes the state vector of each node.

[0149] Based on the state vector of each node, the leg displacement error is calculated. This error is the difference between the current displacement of the leg and the target displacement. At the same time, the rate of change of the error, that is, the speed at which the error changes over time, is calculated. Based on the calculated displacement error and error change rate, the system adaptively adjusts the connection strength between nodes in the network, that is, the communication radius and bandwidth, as well as the importance of each state component in control, that is, the state component weight. For example, if the error is large and the rate of change is fast, the connection strength and importance are increased; otherwise, they are reduced. These adjusted parameters constitute the dynamic connection weights, which are multiplied by the state vector to obtain the weighted state vector. For example, if the weight of the height component is set to 0.5, the weighted height information is 1.2×0.5=0.6.

[0150] An adaptive trigger function is calculated using the leg displacement error and dynamic connection weights. A node sends its status information to its neighbors only when the leg displacement error exceeds this trigger function, avoiding unnecessary communication. Furthermore, a confidence score is calculated for each node based on the leg displacement error. A higher confidence score indicates more reliable status information. For example, smaller errors indicate higher confidence scores.

[0151] Based on the node's confidence, a time-varying information fusion weight is designed to weighted average state information from neighboring nodes. This approach takes into account the impact of communication latency. Furthermore, a time-varying consensus protocol is designed to ensure that all nodes ultimately reach a consensus on the control objective. To optimize protocol performance, an adaptive update law is designed to adjust the protocol gain matrix. For example, if the control effect is poor, the gain is increased; otherwise, the gain is decreased. The final output is the consistent control value.

[0152] Based on the consistency control variable, a feedforward-feedback composite controller is constructed. This controller consists of a disturbance observer and a neural network compensator. The disturbance observer estimates various disturbances in the outrigger motion, such as ground unevenness or wind effects, based on the feedforward control term and the outrigger displacement error. The neural network compensator learns and compensates for the nonlinear characteristics of the outrigger motion through online learning of a weighted state vector, thereby obtaining a motion-compensated control value. Finally, the disturbance variable obtained by the disturbance observer, the motion-compensated control value obtained by the neural network compensator, and the consistency control variable are combined to construct the final composite control ratio.

[0153] A predictive control and fault diagnosis mechanism is designed based on a composite control rate and a weighted state vector. A long short-term memory (LSTM) network is used to predict the future state of each node. The predicted state is compared with the actual state, and a prediction error compensation term is calculated and added to the control rate. Furthermore, the residual between the weighted state vector and the predicted state is calculated as a fault indicator. When the fault indicator exceeds a preset threshold, a fault is identified. Once a fault occurs, the system recalculates the control rate based on the state vectors of neighboring nodes and compensates for the control loss of the faulty leg.

[0154] The present invention effectively suppresses the influence of disturbances and nonlinear factors through the combination of distributed collaborative control and feedforward-feedback composite control, significantly improving the accuracy and stability of outrigger displacement control; the introduction of adaptive adjustment of network connection, confidence weighting, and fault diagnosis and compensation mechanism enhances the robustness and fault tolerance of the system, ensuring the stable operation of the entire system even if some outriggers fail; and controls the frequency of information interaction through adaptive trigger functions, reducing the burden of network communication, lowering the computational complexity, and improving the real-time performance of the system.

[0155] Figure 2 FIG. 1 is a structural diagram of a system for adaptive terrain leveling and stability assessment of vehicle-mounted crane legs according to an embodiment of the present invention. Figure 2 As shown, the system includes:

[0156] The first unit is used to filter and weight the terrain elevation data, outrigger pressure data, and vehicle body posture data collected by the sensor, and then input them into the outrigger ground contact modeling unit. The outrigger ground contact modeling unit uses a historical evolution predictor based on a deep neural network to classify terrain features and establish a ground stiffness-deformation mapping relationship. The outrigger ground contact modeling unit uses an online learning contact parameter adaptor to construct a nonlinear spring-damping contact force model to obtain outrigger load distribution data;

[0157] a second unit configured to input the outrigger load distribution data into an outrigger stability assessment unit, the stability assessment unit comprising a stability discriminator, a stress uniformity evaluator, and a dynamic load early warning device; the stability discriminator forming a support polygon based on the outrigger landing point and calculating the support stability margin in combination with the projection position of the vehicle body's center of gravity; the stress uniformity evaluator calculating a stress uniformity index based on the outrigger pressure variance; and the dynamic load early warning device outputting an early warning signal based on a dynamic characteristic analysis of the outrigger pressure; and the outputs of the stability discriminator, the stress uniformity evaluator, and the dynamic load early warning device being weightedly fused using an adaptive weight coefficient adjusted in real time with the vehicle body posture to obtain a stability state matrix;

[0158] The third unit is used to input the stability state matrix into a model predictive controller, the model predictive controller uses a reinforcement learning method to construct a terrain adaptability evaluation function, and generates a leg leveling control instruction through an action value network; the leg leveling control instruction is input into a distributed collaborative control unit, the distributed collaborative control unit sets the legs as network nodes and performs information exchange, and uses a feedforward feedback composite controller with adaptive gain to control the leg displacement.

[0159] According to a third aspect of the embodiments of the present invention,

[0160] An electronic device is provided, comprising:

[0161] processor;

[0162] a memory for storing processor-executable instructions;

[0163] The processor is configured to call the instructions stored in the memory to execute the aforementioned method.

[0164] According to a fourth aspect of the embodiments of the present invention,

[0165] A computer-readable storage medium is provided, on which computer program instructions are stored. When the computer program instructions are executed by a processor, the method described above is implemented.

[0166] The present invention may be a method, an apparatus, a system and / or a computer program product. The computer program product may include a computer-readable storage medium carrying computer-readable program instructions for executing various aspects of the present invention.

[0167] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the above embodiments, or replace some or all of the technical features therein with equivalents. However, these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for evaluating the leveling and stability of a truck-mounted crane's outriggers based on adaptive terrain, characterized in that: include: The terrain elevation data, outrigger pressure data, and vehicle body posture data collected by the sensors are filtered and weighted, and then input into the outrigger ground contact modeling unit. The outrigger ground contact modeling unit uses a historical evolution predictor based on a deep neural network to classify terrain features and establish a ground stiffness-deformation mapping relationship. The outrigger ground contact modeling unit uses an online learning contact parameter adaptor to construct a nonlinear spring-damping contact force model to obtain outrigger load distribution data. The outrigger load distribution data is input into an outrigger stability assessment unit, the outrigger comprising a stability discriminator, a stress uniformity evaluator, and a dynamic load early warning device. The stability discriminator forms a support polygon based on the outrigger landing point and calculates the support stability margin in combination with the projection position of the vehicle body's center of gravity. The stress uniformity evaluator calculates a stress uniformity index based on the outrigger pressure variance. The dynamic load early warning device outputs an early warning signal based on the dynamic characteristics of the outrigger pressure. The outputs of the stability discriminator, the stress uniformity evaluator, and the dynamic load early warning device are weighted and fused using an adaptive weight coefficient that is adjusted in real time with the vehicle body posture to obtain a stability state matrix. The stability state matrix is ​​input into a model predictive controller, which uses a reinforcement learning method to construct a terrain adaptability evaluation function and generates a leg leveling control instruction through an action value network; the leg leveling control instruction is input into a distributed collaborative control unit, which sets the legs as network nodes and performs information exchange, and uses a feedforward feedback composite controller with adaptive gain to control the leg displacement.

2. The method according to claim 1, characterized in that The outrigger ground contact modeling unit uses a historical evolution predictor based on a deep neural network to classify terrain features and establish a ground stiffness-deformation mapping relationship, and uses an online learning contact parameter adaptor to construct a nonlinear spring damping contact force model. The steps of obtaining outrigger load distribution data include: A historical evolution predictor based on a two-branch densely connected convolutional neural network is used to classify terrain features. The two-branch densely connected convolutional neural network includes a high-frequency feature extraction branch for extracting local texture features and a low-frequency feature extraction branch for extracting regional morphological features. Each feature extraction branch is equipped with an independent spatial pyramid pooling module. The dual-branch features are adaptively fused through a series of attention modules to output the terrain classification result. A hybrid ground mechanics model is established based on the terrain classification results. A three-layer back-propagation neural network is used to establish a mapping relationship between ground stiffness and deformation to obtain ground deformation characteristics. The ground deformation characteristics are combined with the outrigger pressure sequence, displacement sequence, and terrain classification results and input into a two-layer long short-term memory network for soil mechanics parameter identification. The first layer of the two-layer long short-term memory network outputs preliminary soil parameter estimates, and the second layer introduces historical data to perform time series correction on the preliminary soil parameter estimates. A multi-physics field coupled nonlinear spring-damper contact force model is constructed based on soil mechanics parameters. The nonlinear spring-damper contact force model includes a stiffness function related to ground moisture content, a damping function related to temperature, and a friction coefficient function related to relative velocity and pressure. A parameter adaptor based on a motion evaluation framework is used to evaluate the contact state between the outrigger and the ground and to perform online optimization of the parameters of the stiffness, damping, and friction coefficient functions. The dynamic coupling equation of the outrigger group is established based on the nonlinear spring-damping contact force model. Under the constraints of the dynamic coupling equation of the outrigger group, a load distributor with a prediction-correction structure is used to optimize the distribution. The outrigger force signal characteristics are extracted by wavelet decomposition. The adaptive notch filter and integral sliding mode controller are combined to compensate for high-frequency vibration and low-frequency drift, and the outrigger load distribution data is output.

3. The method according to claim 1, characterized in that The stress uniformity evaluator calculates a stress uniformity index based on the outrigger pressure variance, and the dynamic load warning device outputs a warning signal based on the dynamic characteristics of the outrigger pressure; and the outputs of the stability discriminator, the stress uniformity evaluator, and the dynamic load warning device are weightedly integrated using an adaptive weight coefficient that is adjusted in real time with the vehicle body posture to obtain a stability state matrix, including the following steps: The outrigger pressure variance is calculated based on the outrigger load distribution data. The terrain correction coefficient related to the terrain slope and terrain relief is introduced to correct the outrigger pressure variance. The stress uniformity index is calculated by combining the corrected outrigger pressure variance with the maximum pressure. Multi-scale dynamic feature extraction is performed on the outrigger pressure sequence. Dynamic features are extracted through orthogonal complementary decomposition and variational mode decomposition. A time-frequency joint feature index is constructed based on the link function. When the time-frequency joint feature index exceeds an adaptive warning threshold, a dynamic load warning signal is triggered. The adaptive warning threshold is dynamically adjusted according to changes in terrain slope and hoisting mechanism rotation speed. A weight mapping function based on the vehicle body attitude angle, height and lifting mechanism rotation speed is constructed to perform initial weight allocation on the support stability margin, stress uniformity index and dynamic load warning signal. The state sensitivity matrix is ​​obtained by calculating the rate of change of the weight to the vehicle body attitude angle, height and lifting mechanism rotation speed according to the weight mapping function. The weight coefficient is dynamically updated based on the state sensitivity matrix, the vehicle body state parameter increment and the newly calculated weight value, and the weight optimization is performed under the constraint condition that the sum of the weight coefficients is one and the difference of the weight coefficients at adjacent moments does not exceed the preset threshold. The optimized weight is weightedly fused with the support stability margin, stress uniformity index and dynamic load warning signal to obtain the stability state matrix.

4. The method according to claim 3, characterized in that Multi-scale dynamic feature extraction is performed on the outrigger pressure sequence. Dynamic features are extracted through orthogonal complementary decomposition and variational mode decomposition. A time-frequency joint feature index is constructed based on the link function. When the time-frequency joint feature index exceeds an adaptive warning threshold, a dynamic load warning signal is triggered. The adaptive warning threshold is dynamically adjusted according to changes in terrain slope and hoisting mechanism rotation speed, including the following steps: Acquiring a leg pressure signal, a leg deformation, and a ground reaction force signal, constructing an interactive dynamic characteristic matrix between the leg and the ground using the leg pressure signal, the leg deformation, and the ground reaction force signal, calculating a transient component based on the interactive dynamic characteristic matrix using an orthogonal complementary decomposition algorithm, and extracting a dynamic eigenvector based on the transient component; Performing an improved variational modal decomposition on the transient component, wherein an adaptive window function dynamically adjusted according to local characteristics of the signal is introduced into the variational modal decomposition process to obtain multiple eigenmode components and their corresponding center frequencies; Constructing a joint distribution of the dynamic eigenvector, the intrinsic mode component, and the center frequency based on a link function, dynamically selecting a generating function of the link function according to feature correlation, and obtaining a time-frequency joint feature index by integrating the joint distribution; A multi-objective optimization function is constructed, which includes the time-frequency joint characteristic index, terrain slope and hoisting mechanism rotation speed. The time-frequency joint characteristic index is averaged and a cumulative sequence is constructed. A grey differential equation model is established based on the cumulative sequence. A background value sequence is constructed and a parameter estimation equation is established. The parameter estimation equation is solved by the least squares method to obtain a time response function. The predicted value of the time-frequency joint characteristic index is obtained by cumulative reduction, and the grey differential equation model is updated online through a sliding time window. A fuzzy rule base is constructed based on the terrain slope, the hoisting mechanism rotation speed and the predicted time-frequency joint characteristic index. The threshold adjustment amount is obtained through fuzzy reasoning, and the warning threshold is adaptively updated by dynamically adjusting the learning rate according to the prediction error. When the time-frequency joint characteristic index exceeds the warning threshold after adaptive update, a dynamic load warning signal of the outrigger is triggered.

5. The method according to claim 1, wherein The model predictive controller uses a reinforcement learning method to construct a terrain adaptability evaluation function, and the steps of generating leg leveling control instructions through an action value network include: A terrain adaptability evaluation function was constructed that included outrigger height deviation, roll angle deviation, pitch angle deviation, ground contact force, and outrigger sliding speed. The terrain adaptability evaluation function adopted an adaptive weight allocation strategy that was dynamically adjusted based on terrain roughness, slope standard deviation, ground friction coefficient, and outrigger ground stress distribution standard deviation. A Sigmoid function with a periodic compensation term was used during the dynamic weight adjustment process, and the outrigger ground stress distribution was corrected using a Gaussian kernel function. An action-value network structure is designed, comprising an evaluation network and a target network. The state spaces of the evaluation network and the target network include the outrigger height vector, vehicle body attitude angle, terrain feature vector, and outrigger stress distribution characteristics, and the action space is the outrigger velocity instruction. The evaluation network uses a multi-head self-attention mechanism and a residual convolution structure for feature extraction, and calculates the value estimate of the state-action pair based on the output of the terrain adaptability evaluation function. The evaluation network is trained and optimized in the model predictive controller, and the parameters of the evaluation network are soft-updated to the target network using an adaptive update rate based on the training error. Establishing an optimization objective function including a short-term control cost term, a control amount adjustment term, and a terminal penalty term, wherein the short-term control cost term is provided by the evaluation network, the terminal penalty term is calculated based on the long-term value estimate provided by the target network, and the control amount adjustment term includes a square term of the control increments at adjacent moments; The optimization objective function and the leg kinematic constraints, actuator limit constraints and control increment constraints constitute an optimization problem, and the optimization problem is solved by the alternating direction multiplier method based on Nesterov acceleration. The alternating direction multiplier method decomposes the original problem into a sequence of sub-problems through dual decomposition, and the sub-problems are solved by the conjugate gradient method; the solution of the optimization problem is output as the leg leveling control instruction.

6. The method according to claim 5, characterized in that The steps of training and optimizing the evaluation network in the model predictive controller and soft-updating the parameters of the evaluation network to the target network through an adaptive update rate according to the training error include: Construct a distributed experience replay pool, which uses a hierarchical storage structure. The distributed experience replay pool calculates the environment complexity index based on terrain roughness and slope standard deviation, and the task difficulty index based on the leg height error vector, roll angle error, and pitch angle error. The state-action-reward value-next-state transition samples are stored hierarchically according to the weighted combination of the environment complexity index and the task difficulty index. A hybrid priority sampling mechanism is designed based on the distributed experience replay pool. The hybrid priority sampling mechanism includes priority calculation based on Wasserstein distance and importance sampling weight calculation based on the ratio of control decision functions. The combined weight coefficient of the priority calculation and the importance sampling weight calculation is dynamically adjusted according to the training stage, and the sampling probability of samples of different difficulty levels is adjusted according to the combined weight coefficient. Constructing a dual-gradient optimizer, which includes a control decision gradient optimizer and an evaluation network gradient optimizer. The control decision gradient optimizer uses a proximal policy optimization algorithm with trust region constraints to calculate the control decision update based on the ratio of the current control decision function to the historical control decision function. The evaluation network gradient optimizer uses an adaptive moment estimation algorithm with momentum term to construct a loss function based on the prediction error and gradient norm of the evaluation network. Based on the optimization results of the dual-gradient optimizer, a multi-scale adaptive learning mechanism is designed. The base learning rate is adjusted according to the temporal difference error, the variance of the control decision update amount, and the loss value of the evaluation network. The exploration noise amplitude is adaptively adjusted based on the statistical distribution characteristics of the outrigger velocity command output. A hierarchical network update strategy is adopted, and differentiated update frequencies are designed for the feature extraction layer and the control instruction output layer of the evaluation network. The feature extraction layer updates parameters according to a preset first update period, and the control instruction output layer updates parameters according to a preset second update period, where the first update period is greater than the second update period. The parameters of the evaluation network are weighted averaged according to the verification performance and then soft-updated to the target network.

7. The method according to claim 1, characterized in that The distributed collaborative control unit sets the legs as network nodes and performs information exchange, and adopts a feedforward feedback composite controller with adaptive gain to control the leg displacement, including the following steps: A distributed collaborative control unit network is constructed, with each leg corresponding to a distributed collaborative control unit as a network node. The state vector of the network node includes leg height, roll angle, pitch angle, speed, and contact force information. The leg displacement error and error change rate are constructed based on the state vector of the network node. The network communication radius and bandwidth parameters as well as the state component weights are adaptively adjusted according to the leg displacement error and error change rate to obtain a dynamic connection weight and a weighted state vector. Based on the dynamic connection weight and weighted state vector, an exponential function of the leg displacement error is calculated to generate an adaptive trigger function. When the leg displacement error is greater than the adaptive trigger function, information interaction is triggered. The leg displacement error is used to calculate the network node confidence. Based on the network node confidence, a time-varying information fusion weight is designed. A time-varying consistency protocol that considers communication delay is constructed, and an adaptive update law of the protocol gain matrix is ​​designed to output a consistency control quantity. A feedforward-feedback composite controller is constructed based on the consistency control quantity, the feedforward-feedback composite controller including a disturbance observer and a neural network compensator, wherein the disturbance observer estimates the disturbance during the movement of the outrigger based on the feedforward control term and the outrigger displacement error, the neural network compensator obtains a motion compensation control value by online learning the weighted state vector, and the disturbance, the motion compensation control value and the consistency control quantity are combined to construct a composite control rate; A predictive control and fault diagnosis mechanism is designed based on the composite control rate and weighted state vector, a long short-term memory network is used to predict the state of a network node, a prediction error compensation term is designed based on the error between the predicted state of the network node and the actual state, a fault index is calculated by the residual between the weighted state vector and the predicted state of the network node, a fault is determined to have occurred when the fault index is greater than a preset threshold, and when a fault occurs, the control rate is recalculated based on the state vector of the neighboring network node and the control amount of the faulty leg is compensated.

8. A system for adaptive terrain leveling and stability assessment of a truck-mounted crane outrigger, used to implement the method according to any one of claims 1 to 7, characterized in that: include: The first unit is used to filter and weight the terrain elevation data, outrigger pressure data, and vehicle body posture data collected by the sensor, and then input them into the outrigger ground contact modeling unit. The outrigger ground contact modeling unit uses a historical evolution predictor based on a deep neural network to classify terrain features and establish a ground stiffness-deformation mapping relationship. The outrigger ground contact modeling unit uses an online learning contact parameter adaptor to construct a nonlinear spring-damping contact force model to obtain outrigger load distribution data; a second unit configured to input the outrigger load distribution data into an outrigger stability assessment unit, the stability assessment unit comprising a stability discriminator, a stress uniformity evaluator, and a dynamic load early warning device; the stability discriminator forming a support polygon based on the outrigger landing point and calculating the support stability margin in combination with the projection position of the vehicle body's center of gravity; the stress uniformity evaluator calculating a stress uniformity index based on the outrigger pressure variance; and the dynamic load early warning device outputting an early warning signal based on a dynamic characteristic analysis of the outrigger pressure; and the outputs of the stability discriminator, the stress uniformity evaluator, and the dynamic load early warning device being weightedly fused using an adaptive weight coefficient adjusted in real time with the vehicle body posture to obtain a stability state matrix; The third unit is used to input the stability state matrix into a model predictive controller, the model predictive controller uses a reinforcement learning method to construct a terrain adaptability evaluation function, and generates a leg leveling control instruction through an action value network; the leg leveling control instruction is input into a distributed collaborative control unit, the distributed collaborative control unit sets the legs as network nodes and performs information exchange, and uses a feedforward feedback composite controller with adaptive gain to control the leg displacement.

9. An electronic device, characterized in that: include: processor; a memory for storing processor-executable instructions; The processor is configured to call the instructions stored in the memory to execute the method according to any one of claims 1 to 7.

10. A computer-readable storage medium having computer program instructions stored thereon, characterized in that: When the computer program instructions are executed by a processor, the method according to any one of claims 1 to 7 is implemented.

Images