A method for multi-station precise parking control of a vertical elevator

Abstract

The application provides a kind of vertical hoist multi-station precision parking control method, belongs to hoist control technical field, the application is by being arranged sensor on drive unit and guide rail, real-time extraction elastic deformation characteristics is adopted by bidirectional long short term memory unit, utilize state entropy increase constraint asynchronous multi-flow network to carry out low-dimensional alignment and three-way parallel fusion processing to heterogeneous sensor data, then by artificial bee colony algorithm in low-dimensional latent space collaborative search optimal control voltage, simultaneously drive Gaussian process hidden variable deep expansion model to flexible body elastic dynamics prior probability correction each station target position, finally through space self-organizing evolution control law dynamic reconstruction braking smooth curve and real-time restrain buffeting high frequency term, drive vertical hoist to complete multi-station precision parking, solved the technical problem that vertical hoist is in multi-station high frequency switching working condition due to steel wire rope elastic deformation and guide rail thermal expansion coupling causes parking precision dynamic degradation.

Description

Technical Field

[0001] This invention belongs to the field of hoist control technology, and more specifically, relates to a method for precise multi-station docking control of a vertical hoist. Background Technology

[0002] Vertical hoists are widely used in intelligent warehousing, industrial production lines, and material conveying in high-rise buildings. Precise multi-station docking control is a core element in ensuring system efficiency. Existing technologies typically employ a position-closed-loop proportional-integral-derivative (PID) controller with encoder feedback for docking positioning. Some solutions introduce sliding mode control or model predictive control to improve robustness and use temperature compensation coefficients for static correction of guide rail thermal expansion. Under high-frequency multi-station switching conditions, the wire rope undergoes nonlinear elastic creep during frequent starts and stops. This creep dynamically changes with load mass, operating speed, and the current temperature field distribution. Simultaneously, the guide rail thermal expansion exhibits a spatially non-uniform distribution, causing the target position at each station to drift continuously over time. Existing control methods rely on fixed compensation coefficients or offline calibrated models, failing to capture the dynamic coupling characteristics of elastic and thermal deformation in real time. This leads to overconfidence in the control output during sudden changes in system entropy caused by high-frequency switching, resulting in dynamic degradation of docking accuracy. In other words, existing technologies present a technical problem where the dynamic degradation of docking accuracy in vertical hoists under high-frequency multi-station switching conditions is caused by the coupling between wire rope elastic deformation and guide rail thermal expansion. Summary of the Invention

[0003] In view of this, the present invention provides a method for precise docking control of a vertical hoist at multiple workstations, which can solve the technical problem in the prior art where the docking accuracy of a vertical hoist dynamically deteriorates due to the coupling of elastic deformation of the wire rope and thermal expansion of the guide rail under high-frequency switching conditions at multiple workstations.

[0004] This invention is implemented as follows: This invention provides a method for precise multi-station docking control of a vertical hoist, comprising the following steps:

[0005] Displacement sensors and temperature sensors are installed on the drive unit and guide rail of the vertical hoist to collect the elastic deformation signal of the wire rope and the temperature field distribution data of the guide rail. The signals are then connected to a bidirectional circular queue, and the elastic deformation features are extracted in real time by a high-order fusion algorithm of streaming time series.

[0006] The elastic deformation characteristics and temperature field distribution data are input into the state entropy increase constraint asynchronous multi-channel network, projected to a low-dimensional orthogonal space by the shared latent space low-dimensional alignment operator, and processed in parallel by three channels: dynamic feature channel, time domain statistical channel and frequency domain transformation channel, and output control feature vector.

[0007] Using the control feature vector as the nectar source in the artificial bee colony, the guiding bee algorithm searches for the initial control voltage that satisfies the dynamic constraints in the low-dimensional potential space, the following bee algorithm performs weighted collaborative iterative optimization of the three flow channel outputs based on the real-time error confidence, and the reconnaissance bee algorithm adaptively reconstructs the local activation operator when the system entropy increases abruptly.

[0008] The control feature vector is synchronously input into the Gaussian process latent variable deep expansion model. After deep hierarchical latent variable spatial probability mapping, the prior mean function driven by the elastic dynamic boundary equation of the flexible body corrects the target position of each station and outputs the docking correction amount.

[0009] The spatial self-organizing evolution control law is adopted to integrate the docking correction amount with the real-time position error. The braking smooth curve is dynamically reconstructed through a local activation operator. In the streaming calculation, the high-frequency term of chattering is suppressed in real time, and the vertical hoist is driven to complete the multi-station precise docking at the corrected target position.

[0010] The bidirectional circular queue is a first-in-first-out ring buffer structure. The queue depth is determined experimentally based on the product of the duration of the high-frequency pulse burst and the sampling frequency. When the rate of new sampled data arrival exceeds the processing rate, the head of the queue automatically overwrites the oldest data.

[0011] The high-order fusion algorithm for streaming time series is based on bidirectional long short-term memory units. It performs forward and reverse time series modeling on the time series of multiple sensors in a bidirectional circular queue, extracts the elastic deformation feature vector containing elastic creep hysteresis correction, captures the phase delay through reverse time series modeling, and then compresses the forward and reverse hidden states through a linear projection layer to eliminate time domain hysteresis error.

[0012] The shared latent space low-dimensional alignment operator is a two-layer fully connected network that projects heterogeneous input data from different sensors onto a low-dimensional orthogonal feature space. The inter-layer activation function is a linear rectified function with leakage.

[0013] In the state entropy increase constraint asynchronous multi-channel network, the dynamic feature channel is composed of a convolutional neural network, the time-domain statistical channel is composed of stacked local asymmetric gated recurrent units, the frequency domain transformation channel applies a learnable discrete cosine transform basis to the input time series and then extracts the frequency domain statistical features by a fully connected network, and the outputs of the three channels are spliced ​​in the convergence layer and then fused into a control feature vector by attention weighting.

[0014] The ratio of the forward propagation gating time constant to the feedback correction gating time constant of the local asymmetric gated loop unit satisfies the asymmetric setting value, with fast forward propagation and slow feedback correction to ensure the real-time performance of streaming inference.

[0015] The state entropy increase constraint asynchronous multi-channel network is equipped with multiple confidence classifiers based on the state entropy increase threshold. When the posterior entropy is lower than the preset threshold, a dynamic hierarchical early exit mechanism is triggered, and a docking control command is directly output. The weight matrix of each layer in the network adopts sparse mask constraint, and the sparsity rate is determined through experimental iteration based on the activation frequency statistics of neurons in each layer.

[0016] In the artificial bee colony algorithm, the leader bee, follower bee, and scout bee run in parallel in a multi-threaded manner during the inference phase. When the entropy of the system state exceeds the entropy increase threshold, the scout bee algorithm abandons the current nectar source and re-randomly initializes the local activation operator parameters.

[0017] The Gaussian process latent variable deep expansion model sets up multiple latent variable layers. Each latent variable space is dynamically defined by an independent Gaussian process kernel function. The kernel function adopts a weighted combination of radial basis functions and linear kernels. The kernel hyperparameters are automatically optimized during the training phase through variational inference algorithm.

[0018] The Gaussian process latent variable deep expansion model is equipped with an adaptive parameter dynamic weight update algorithm. It dynamically adjusts the update step size of the kernel hyperparameters of each layer according to the real-time estimated value of the variance of each latent variable. When the output variance of the low-level latent variable is lower than the variance threshold, the docking correction amount is directly output.

[0019] The boundary equations of the flexible body's elastic dynamics are a set of partial differential equations describing the elastic creep and vibration response of the wire rope. The boundary conditions change dynamically with the target position, load mass and thermal expansion of the guide rail at each workstation. After embedding the prior mean function, the stopping correction amount satisfies the elastic dynamic constraints at the physical mechanism level.

[0020] The calculation method for the state entropy increment is as follows: the probability density of the system state vector in the most recent control cycles is estimated, and the Kullback-Leibler divergence between the estimated distribution and the reference steady-state distribution is calculated as a normalization measure. The reference steady-state distribution is determined by fitting historical data of the hoist operating under standard conditions.

[0021] In this process, after the multi-station precise docking is completed, the elastic deformation characteristics, temperature field distribution data, docking correction amount and docking accuracy error of this docking are written into the training dataset, triggering the online adaptive update of the state entropy increase constraint asynchronous multi-channel network and the Gaussian process latent variable deep expansion model.

[0022] Among them, the comprehensive evaluation index of docking accuracy Based on the root mean square position error of the most recent docking Temperature field drift of the guide rail With the increase of state entropy We obtain the result by weighted linear summation, i.e. Weighting coefficient , , This was determined through field experiments and data regression analysis.

[0023] The queue depth range of the bidirectional circular queue is: ~ The sampling points are [number]; the dimension of the elastic deformation feature vector ranges from 32 to 128; the dimension of the low-dimensional orthogonal feature space ranges from 16 to 64; the sparsity ranges from 30% to 70%; and the variance threshold search range is [number]. ~ mm Weighting coefficient The initial value range is 0.4 to 0.6. The initial value range is 0.2 to 0.4. The initial value range is 0.1 to 0.3; the learning rate satisfies... , proportion range The range is 5 to 20; the boundary penalty coefficient for increasing state entropy is within the range of... ~ .

[0024] This invention extracts elastic deformation features in real time by constructing a high-order fusion algorithm for streaming time series data. It utilizes an asynchronous multi-channel network constrained by state entropy increase to asynchronously and parallelly fuse elastic deformation features and temperature field distribution data. Furthermore, it employs an artificial bee colony algorithm to collaboratively optimize the control voltage in a low-dimensional latent space. Simultaneously, it introduces a Gaussian process latent variable deep extension model to probabilistically correct the target position of each station using prior elastic dynamics of the flexible body. Finally, it uses a spatial self-organizing evolution control law to suppress high-frequency chattering terms in real time, achieving precise multi-station docking. This invention integrates the dynamic features of elastic deformation and the non-uniform distribution information of the temperature field into a probabilistic control framework. This allows the system to adaptively tighten the control output boundary when the entropy increase is abruptly caused by high-frequency switching, avoiding the extrapolation failure of the fixed compensation model under dynamic coupling conditions. This overcomes the deficiency of existing control methods that cannot perceive the dynamic coupling of elastic and thermal deformation in real time, leading to decreased docking accuracy. In summary, this invention solves the technical problem mentioned in the background art: the dynamic deterioration of docking accuracy of vertical hoists under multi-station high-frequency switching conditions caused by the coupling of wire rope elastic deformation and guide rail thermal expansion. Attached Figure Description

[0025] Figure 1 This is a flowchart of the method of the present invention.

[0026] Figure 2 A schematic diagram of the data flow direction of three parallel channels in an asynchronous multi-channel network constrained by increasing state entropy.

[0027] Figure 3 This is a distribution diagram of docking correction amounts under representative loads at each workstation.

[0028] Figure 4 Comprehensive evaluation index for docking accuracy Curve showing the change with the number of stops.

[0029] Figure 5 This is a statistical chart showing the error in stopping positions for 40 round trips.

[0030] Figure 6 A schematic diagram of the attention-weighted fusion structure of the convergence layer. Detailed Implementation

[0031] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below.

[0032] like Figure 1 The diagram shown is a flowchart of a method for precise multi-station docking control of a vertical hoist provided by the present invention. This method includes the following steps:

[0033] S01. Displacement sensors and temperature sensors are installed on the drive unit and guide rail of the vertical hoist to collect the elastic deformation signal of the wire rope and the temperature field distribution data of the guide rail. The signals are then connected to a bidirectional circular queue, and the elastic deformation features are extracted in real time by a high-order fusion algorithm of streaming time series.

[0034] S02. Input the elastic deformation characteristics and temperature field distribution data into the state entropy increase constraint asynchronous multi-channel network, project it to the low-dimensional orthogonal space through the shared potential space low-dimensional alignment operator, and process it in parallel by three channels: dynamic feature channel, time domain statistical channel and frequency domain transformation channel, and output the control feature vector.

[0035] S03. Using the control feature vector as the nectar source in the artificial bee colony, the guiding bee algorithm quickly searches for the initial control voltage that satisfies the dynamic constraints in the low-dimensional potential space. The following bee algorithm performs weighted collaborative iterative optimization of the three flow channel outputs based on the real-time error confidence. The reconnaissance bee algorithm adaptively reconstructs the local activation operator when the system entropy increases abruptly.

[0036] S04. The control feature vector is synchronously input into the Gaussian process latent variable deep expansion model. After deep hierarchical latent variable spatial probability mapping, the prior mean function driven by the elastic dynamic boundary equation of the flexible body corrects the target position of each station and outputs the docking correction amount.

[0037] S05. Using a spatial self-organizing evolution control law, the docking correction amount is fused with the real-time position error. The braking smooth curve is dynamically reconstructed through a local activation operator. In the streaming calculation, the high-frequency term of the chattering is suppressed in real time, and the vertical hoist is driven to complete the multi-station precise docking at the corrected target position.

[0038] S06, Optionally, also includes: after docking is completed, the elastic deformation characteristics, temperature field distribution data, docking correction amount and docking accuracy error of this docking are written into the training dataset, triggering the online adaptive update of the state entropy increase constraint asynchronous multi-channel network and the Gaussian process latent variable deep extension model, and iteratively optimizing the docking accuracy.

[0039] The displacement sensor is a micrometer-level resolution displacement sensor. The sampling frequency is determined through multiple docking experiments and spectrum analysis based on the guide rail length and the number of workstations, with a recommended range of 1000–5000 Hz. The temperature sensors are evenly distributed along the guide rail, and the spacing is determined through experimental calibration based on the guide rail thermal expansion coefficient and guide rail length, with a recommended spacing range of 0.5–2 m.

[0040] The bidirectional circular queue is a first-in, first-out (FIFO) ring buffer structure. The queue depth is determined experimentally based on the product of the high-frequency pulse burst duration and the sampling frequency; a recommended depth range is [insert range here]. ~ Each sampling point; when the rate of new sampled data arrival exceeds the processing rate, the head of the queue automatically overwrites the oldest data to prevent memory overflow.

[0041] The high-order fusion algorithm for streaming time series is based on bidirectional long short-term memory units. It performs forward and reverse time series modeling on the time series of multiple sensors in a bidirectional circular queue, and extracts elastic deformation feature vectors containing elastic creep hysteresis correction. The dimension of the elastic deformation feature vector is determined according to the number of sensors and the modeling order, and the suggested range is 32 to 128 dimensions. The hysteresis correction principle of the algorithm is as follows: the phase delay introduced by high-frequency pulses in the sensor acquisition link is captured by reverse time series modeling, and the forward and reverse hidden states are concatenated and compressed by a linear projection layer to eliminate the time domain hysteresis error of traditional displacement sensors under high-frequency pulse conditions.

[0042] The specific structure of the state entropy increase constraint asynchronous multi-channel network is as follows: the front end consists of a low-dimensional alignment operator with a shared latent space. This operator is a two-layer fully connected network that forcibly projects heterogeneous input data from different sensors onto a low-dimensional orthogonal feature space of 16–64 dimensions. The connection weights between neurons in each layer are updated independently using gradient descent, and the inter-layer activation function uses a linear rectified function with leakage. The network internally constructs three fully parallel heterogeneous computing pipelines. There is no synchronization or waiting dependency between the three pipelines, and each maintains an independent memory allocation region. The memory is divided according to the channel, and each channel is allocated memory. The ratio is dynamically determined by the computational load of each channel, based on the average computational delay of each channel in the most recent batches of inference. A monitoring window of 16–64 batches is recommended. The dynamic feature channel consists of 3–5 layers of convolutional neural networks to capture the spatial local structural features of elastic deformation. The temporal statistical channel consists of 2–4 layers of locally asymmetric gated loop units stacked together. The ratio of the forward propagation gate time constant to the feedback correction gate time constant of these locally asymmetric gated loop units is set to 3:1–8:1. This asymmetric information flow, characterized by fast forward propagation and slow feedback correction, ensures the real-time performance of streaming inference. The domain transformation channel first applies a learnable discrete cosine transform basis to the input time series, concentrating the signal energy into low-frequency coefficients. Then, a 2-3 layer fully connected network extracts frequency domain statistical features. The output features of the three channels are concatenated at the convergence layer and then fused into a control feature vector through attention weighting. The attention weights are dynamically allocated based on the current confidence level of each channel. Multiple confidence classifiers based on state entropy increment thresholds are set within the network. These classifiers calculate the posterior entropy of the current convergence feature. When the posterior entropy falls below a preset threshold, a dynamic hierarchical early exit mechanism is triggered, directly outputting docking control commands without further propagation. The subsequent deep network; the weight matrix between neurons in each layer of the network adopts sparse mask constraints, and the sparsity rate is determined through experimental iteration based on the activation frequency statistics of neurons in each layer, with a suggested range of 30% to 70%, in order to compress the memory-level feature storage space and reduce inference latency; the batch size in the data loop is dynamically adjusted by the real-time load monitoring module according to the current memory occupancy rate, with a suggested batch size range of 1 to 16; the network is based on deep learning and is trained using the backpropagation algorithm, and the state entropy increase dynamic constraint is embedded in the loss function in the form of a boundary penalty term, with the penalty coefficient determined through multiple ablation experiments, and a suggested range of [missing information]. ~ .

[0043] The steps for establishing the training dataset for the state entropy increase constraint asynchronous multi-channel network specifically include: collecting the original time series of displacement sensors and temperature sensors in the actual operating environment of the vertical hoist, covering all workstations, the full load range, and the full temperature range; labeling each time series with the corresponding docking accuracy error, elastic deformation amplitude, and guide rail thermal expansion; calculating the state entropy increase of the dataset, mixing high-entropy increase samples with low-entropy increase samples in a certain proportion, with a suggested proportion of 30% to 60% for high-entropy increase samples; and dividing the samples into training set, validation set, and test set after normalization, with a suggested ratio of 7:2:1.

[0044] The specific steps of training the asynchronous multi-channel network with state entropy increase constraint include: using the sum of mean squared error loss and state entropy increase boundary penalty term as the total loss function, and using an adaptive moment estimation optimizer to update parameters; the initial learning rate is determined by the following docking accuracy adaptive adjustment function; after each training round, the confidence threshold of the confidence classifier is evaluated on the validation set, and the trigger threshold of the dynamic layer early exit mechanism is adjusted according to the evaluation results; the above process is repeated until the validation set loss does not decrease for several consecutive rounds, and it is recommended that the early stop rounds be 5 to 15 rounds.

[0045] The design method of the docking accuracy adaptive adjustment function is as follows: Define the docking accuracy comprehensive evaluation index. , Based on the root mean square position error of the most recent docking (unit ), guide rail temperature field drift (Unit: °C) and the increase in state entropy We obtain the result by weighted linear summation of (dimensionless normalized quantities), i.e. The weighting coefficient , , Based on multiple field experiments and data regression analysis, the recommended initial value ranges are as follows: , , ;when At that time, a smaller learning rate was used. And disable the dynamic hierarchical early exit mechanism to maintain model stability, where The docking accuracy that meets the design specifications after model convergence in the experiment. The statistical mean of the values ​​has been determined, and the suggested range is [range missing]. Equivalent; when At that time, a medium learning rate was used. A dynamic hierarchical early exit mechanism is enabled to balance speed and accuracy. The corresponding time when the docking accuracy degrades significantly in the experiment The statistical mean of the values ​​has been determined, and the suggested range is [range missing]. Equivalent; when At that time, a larger learning rate was used. Simultaneously, it triggers the reconnaissance bee algorithm to reconstruct local activation operators to quickly adapt to sudden environmental changes, where the learning rate satisfies... The recommended ratio range is The and The specific values ​​were obtained by conducting no fewer than 30 docking tests covering all operating conditions on the target hoist, and collecting data from each docking. , and Then, k-fold cross-validation was used. Regression analysis was used to determine this.

[0046] The principle of the state entropy increase dynamic constraint is as follows: the rate of increase of the thermodynamic state entropy of the hoist guide rail and wire rope transmission system is used as a measure of system uncertainty and applied in the loss function as a boundary penalty term. This forces the network to output a more conservative docking correction amount under high entropy increase conditions, thereby avoiding over-adjustment caused by abrupt changes in the nonlinear thermal deformation of the guide rail. The technical effect of the constraint on the scheme is that the network can adaptively tighten the control output boundary in scenarios of abrupt changes in temperature field or entropy increase caused by high-frequency switching of multiple workstations, avoiding docking errors caused by overconfident output of the model, and improving the robustness of multi-workstation docking control under extreme conditions.

[0047] The technical advantages of the asynchronous parallel pipeline multi-channel design are as follows: the three heterogeneous computing pipelines do not wait or block each other synchronously; the dynamic characteristic channel, time-domain statistical channel, and frequency-domain transformation channel can advance synchronously on different hardware computing units; when a channel experiences a sudden jitter in the input data, causing a computational delay, the output of the other two channels can still support the confidence classifier to exit early and output control commands, fundamentally eliminating the accumulation of control delay during high-frequency switching of multiple workstations.

[0048] The principle of the artificial bee colony algorithm is as follows: the leading bee algorithm quickly explores candidate control voltages that satisfy dynamic constraints in a low-dimensional potential space with random perturbation, essentially completing a coarse-grained search in the low-dimensional projection of the high-dimensional control space; the following bee algorithm performs probability-weighted sampling based on the real-time error confidence of each candidate control voltage, and performs collaborative iterative refinement of the output of the three channels, converging to a control voltage with higher accuracy; the scout bee algorithm abandons the current nectar source and re-randomly initializes the local activation operator parameters when the system state entropy increases beyond a threshold, achieving a rapid response to sudden changes in operating conditions; the three types of bee colony algorithms run in parallel in a multi-threaded manner during the inference phase, eliminating the blocking delay of multi-threaded synchronous waiting; the technical effects of the algorithm on the solution are as follows: traditional gradient descent post-processing optimization is limited in accuracy due to insufficient iterations when switching between multiple workstations at high frequency, while the artificial bee colony algorithm, through distributed parallel collaborative search, completes the evaluation and screening of multiple candidate solutions within a single inference cycle, enabling the control voltage output to take into account both real-time performance and global optimality, significantly improving the streaming high-concurrency control capability for precise multi-workstation docking.

[0049] The specific structure of the Gaussian process latent variable deep expansion model is as follows: The model is based on a combination of deep learning and Gaussian processes, exhibiting a deep hierarchical topology with 3-5 latent variable layers. The dimension of the latent variable space decreases progressively with each layer, and the top layer has 2-8 dimensions. Each latent variable space is dynamically defined by an independent Gaussian process kernel function. The kernel function uses a weighted combination of radial basis functions and linear kernels. The kernel hyperparameters are automatically optimized during training using a variational inference algorithm. There is no sharing of kernel hyperparameters between layers to avoid the memory overhead of generating explicit fully connected correlation matrices. The conditional dependencies between latent variables in each layer are parameterized through a variational posterior distribution. The mean and variance of each layer's output are passed to the next layer as inputs to the prior mean function. The elastic dynamic boundary equations of the flexible body are embedded in the probability distribution of the top-level Gaussian process in the form of a prior mean function. The parameters of the prior mean function are determined by regression fitting of the elastic modulus of the wire rope, the thermal expansion coefficient of the guide rail, and the load mass data collected in multiple full-condition docking experiments. It is recommended that the number of regression experiments be no less than 50. An adaptive parameter dynamic weight update algorithm is set within the model. This algorithm dynamically adjusts the update step size of the kernel hyperparameters of each layer based on the real-time estimated value of the variance of the latent variables in each layer. Layers with larger variances are assigned larger update step sizes to accelerate convergence, while layers with smaller variances are assigned smaller update step sizes to prevent overfitting. A dynamic layer early exit mechanism is constructed within the model using a confidence classifier. When the output variance of a lower-layer latent variable is lower than the variance threshold corresponding to millimeter-level docking accuracy, the docking correction amount is directly output without continuing to derive higher-order latent variable layers. The variance threshold is determined by searching on the validation set for the maximum variance value that allows the docking accuracy to meet the design specifications. A search range of [missing information] is recommended. ~ The model exhibits an asymmetric flow of fast forward probability mapping and slow feedback kernel update. During the forward inference phase, the mean values ​​of latent variables in each layer are rapidly propagated in the hierarchical order. During the feedback kernel update phase, only layers with variance exceeding the threshold are subjected to kernel hyperparameter gradient backpropagation, thereby compressing memory storage space and reducing inference latency. In terms of hierarchical allocation, the lower latent variable layers are allocated more computational resources for fast forward mapping, while the top latent variable layers are allocated more memory resources for storing prior mean function parameters and historical states of kernel hyperparameters.

[0050] The steps for establishing the training dataset for the Gaussian process latent variable deep extension model specifically include: collecting historical docking data of the vertical hoist under all positions, all loads, and all temperature ranges; each sample includes the input control feature vector, guide rail temperature field distribution data, measured docking accuracy error, and target docking correction; fitting the flexible body elastic dynamic boundary equation to the prior docking correction data for each position to obtain the initial parameters of the prior mean function; it is recommended that the number of samples be no less than 500, and supplemented by data augmentation in small sample scenarios, with an augmentation factor of 3 to 10 times.

[0051] The specific steps for training the Gaussian process latent variable deep extension model include: using the variational evidence lower bound as the objective function, and simultaneously optimizing the kernel hyperparameters and latent variable posterior distribution parameters of each layer of the Gaussian process through a variational inference algorithm; the initial learning rate is adaptively adjusted by the docking accuracy function. The value is determined; after each training round, the variance of the latent variables of each layer is evaluated on the validation set, and the variance threshold of the dynamic layer early exit mechanism is adjusted; repeat the above process until the lower bound of the variational evidence on the validation set does not improve for several consecutive rounds. It is recommended that the early stop rounds be 5 to 15.

[0052] The technical benefits of the Gaussian process latent variable deep extension model are as follows: Compared with parametric deep networks, the nonparametric probability mapping of Gaussian processes naturally outputs the uncertainty estimate of each docking correction, enabling the control system to quantify the confidence interval of the correction in small-sample cross-station migration scenarios. This allows the system to proactively trigger conservative control strategies when the uncertainty is large, avoiding docking errors caused by insufficient samples. The deep hierarchical topology allows the model to refine the abstract representation from the original elastic deformation features to the high-level dynamic constraints layer by layer, taking into account both feature representation ability and interpretability of probabilistic reasoning.

[0053] The principle of the local activation operator is as follows: independent nonlinear mapping functions are applied to each dimension of the control feature vector. The shape parameters of the mapping functions are adaptively reconstructed by the reconnaissance bee algorithm when the system entropy increases abruptly, so that the local shape of the braking smooth curve can follow the dynamic boundary changes caused by the nonlinear deformation of the guide rail and the airflow disturbance in real time, thereby suppressing the high-frequency chattering term generated by traditional sliding mode control.

[0054] The principle of the spatial self-organizing evolution control law is as follows: the braking smooth curve is regarded as a manifold that self-organizes and evolves in the position-velocity state space. The local curvature of the manifold is dynamically adjusted by the local activation operator, so that the control trajectory converges to the target stopping position along the direction of smooth curvature change, avoiding the high-frequency chattering caused by the rigidity of the switching surface in traditional sliding mode control. In the streaming calculation, the manifold curvature of the current state point is estimated in real time and the local activation operator parameters are updated in each control cycle to achieve online suppression of the high-frequency chattering term.

[0055] The technical effect of the shared latent space low-dimensional alignment operator is as follows: heterogeneous data from displacement sensors, temperature sensors and other sensors with different physical dimensions are uniformly projected into the same low-dimensional orthogonal space before entering the three parallel flow channels. This eliminates the feature coupling interference caused by the heterogeneity of different sensor data in terms of dimensions, amplitude and sampling rate, and provides a standardized feature basis for the subsequent independent parallel processing of the three flow channels.

[0056] The boundary equations of the flexible body elastic dynamics are defined as a set of partial differential equations describing the elastic creep and vibration response of the wire rope during frequent start-stop processes at multiple workstations. The boundary conditions of the equations change dynamically with the target position, load mass, and thermal expansion of the guide rail at each workstation. By embedding the equations into the prior mean function of the Gaussian process latent variable deep extension model, the docking correction output by the model satisfies the elastic dynamic constraints at the physical mechanism level, thus avoiding the extrapolation failure of the pure data-driven model near the physical boundary.

[0057] Wherein, the state entropy increases The calculation method is as follows: The probability density of the system state vector (including position error, velocity error, and temperature field drift) over the most recent control cycles is estimated, and the Kullback-Leibler divergence between the estimated distribution and the reference steady-state distribution is calculated as... The normalized metric, with reference steady-state distribution, is determined by fitting historical data of the hoist operating under standard conditions; the window length of the several control cycles is determined experimentally based on the system entropy increase response time, with a suggested range of 10 to 50 cycles.

[0058] Furthermore, the present invention also provides a computer-based multi-station precision docking control system for vertical hoists. The computer is equipped with a readable storage medium, which stores program instructions. When the program instructions are run in the computer, they execute all the steps in the above method. The computer is communicatively connected to the displacement sensor, temperature sensor, and drive unit of the vertical hoist, thereby realizing fully automatic control of the multi-station precision docking process.

[0059] The specific implementation of step S01 is as follows: Micrometer-resolution displacement sensors are installed along the entire length of the guide rail and the fixed end of the drive unit of the vertical hoist. The sampling frequency is determined through multiple docking experiments and spectrum analysis, with a recommended range of 1000–5000 Hz. Temperature sensors are evenly distributed along the guide rail, with the spacing determined through experimental calibration, and a recommended range of 0.5–2 m. The collected elastic deformation signals of the wire rope are input into a bidirectional circular queue. This queue is a first-in, first-out (FIFO) ring buffer structure. The queue depth is determined by the product of the high-frequency pulse burst duration and the sampling frequency, with a recommended range of [missing information]. ~ Each sampling point; when the rate of new data arrival exceeds the processing rate, the oldest data is automatically overwritten at the head of the queue to prevent memory overflow. The high-order fusion algorithm for streaming time series is based on bidirectional long short-term memory units. It performs forward and backward time series modeling on the time series of multiple sensors in the queue. Forward modeling extracts the elastic creep trend along the positive time direction, while backward modeling captures the phase delay introduced by high-frequency pulses in the sensor acquisition link along the negative time direction. After concatenating the forward hidden state and the backward hidden state, it is compressed into a 32-128 dimension elastic deformation feature vector through a linear projection layer. This eliminates the time domain hysteresis error of traditional displacement sensors under high-frequency pulse conditions and ensures the time domain accuracy of the elastic deformation feature vector on which subsequent control depends.

[0060] The specific implementation of step S02 is as follows. The input end consists of a low-dimensional alignment operator with a shared latent space. This operator is a two-layer fully connected network that projects heterogeneous data from displacement sensors and temperature sensors into a low-dimensional orthogonal feature space with dimensions of 16 to 64. The inter-layer activation function adopts a linear rectified function with leakage, and the weights of each layer are updated independently by the gradient descent method to eliminate feature coupling interference caused by different physical dimensions and sampling rates. Each of the three parallel channels maintains an independent memory allocation region, with the memory ratio dynamically determined based on the average computational latency of each channel in the most recent 16-64 batches of inference. The dynamic feature channel consists of 3-5 layers of convolutional neural networks, extracting the spatial local structural features of elastic deformation. The temporal statistics channel consists of 2-4 layers of locally asymmetric gated recurrent units, with the ratio of forward propagation gating time constant to feedback correction gating time constant set to 3:1-8:1 to ensure the real-time performance of streaming inference. The frequency domain transformation channel first applies a learnable discrete cosine transform basis to the input sequence to concentrate the signal energy to low-frequency coefficients, and then extracts frequency domain statistical features using a 2-3 layer fully connected network. The outputs of the three channels are concatenated at the convergence layer and then fused into a control feature vector through attention weighting, with the attention weights dynamically allocated by the current confidence of each channel. Multiple confidence classifiers within the network calculate the posterior entropy of the convergence features. When the posterior entropy falls below a preset threshold, a dynamic hierarchical early exit mechanism is triggered, directly outputting a docking control command. The weight matrices of each layer are constrained by sparse masks with a sparsity range of 30% to 70% to compress memory and reduce inference latency.

[0061] The specific implementation of step S03 is as follows. Using the control feature vector as the nectar source, the three types of bees in the artificial bee colony algorithm run in parallel with multiple threads. The lead bee quickly searches for candidate control voltages that satisfy the dynamic constraints in the low-dimensional potential space using random perturbation, completing a coarse-grained global exploration; the follower bees perform probability-weighted sampling based on the real-time error confidence level corresponding to each candidate control voltage, and perform collaborative iterative refinement of the three flow channel outputs, converging to a control voltage with higher accuracy; the scout bee abandons the current nectar source when the system state entropy increases beyond the entropy increase threshold, and re-randomly initializes the local activation operator parameters to achieve a rapid response to sudden changes in operating conditions. The parallel multi-threading of the three types of bee colonies eliminates the blocking delay of synchronous waiting, enabling the control voltage to complete the evaluation and screening of multiple candidate solutions within a single inference cycle, balancing real-time performance and global optimality. State entropy increase... The calculation is based on the probability density estimation of the system state vector over the most recent 10 to 50 control cycles. The Kullback-Leibler divergence is used to measure the deviation between the current distribution and the reference steady-state distribution, which is determined by fitting historical data of the hoist under standard operating conditions.

[0062] The specific implementation of step S04 is as follows. The Gaussian process latent variable deep expansion model has a deep hierarchical topology, with 3 to 5 latent variable layers. The dimension of the top-level latent variable space is 2 to 8 dimensions, and the dimensions of each layer decrease progressively. Each latent variable space is dynamically defined by an independent Gaussian process kernel function. The kernel function adopts a weighted combination of radial basis functions and linear kernels. The kernel hyperparameters are automatically optimized during the training phase through a variational inference algorithm, and there is no sharing between the kernel hyperparameters of each layer. The boundary equation of the flexible body elastic dynamics is embedded in the top-level Gaussian process in the form of a prior mean function. The parameters of the prior mean function are determined by regression fitting of the elastic modulus of the wire rope, the thermal expansion coefficient of the guide rail, and the load mass data collected in no less than 50 full-condition docking experiments. The adaptive parameter dynamic weight update algorithm dynamically adjusts the update step size of the kernel hyperparameters of each layer according to the real-time estimated value of the variance of each latent variable. The layer with larger variance is assigned a larger step size to accelerate convergence, and the layer with smaller variance is assigned a smaller step size to prevent overfitting. When the output variance of the lower-level latent variables is lower than the variance threshold (search range), ~ mm When the dynamic hierarchical early exit mechanism is triggered, the docking correction amount is directly output, and the higher-order hidden variable layer is not further derived, so as to compress the inference delay.

[0063] The specific implementation of step S05 is as follows. The spatial self-organizing evolution control law treats the braking smooth curve as a manifold that self-organizes and evolves in the position-velocity state space. In each control cycle, the manifold curvature of the current state point is estimated in real time, and the local curvature of the manifold is dynamically adjusted through a local activation operator. The local activation operator applies an independent nonlinear mapping function to each dimension of the control feature vector. The shape parameters of the mapping function are adaptively reconstructed by the reconnaissance bee algorithm in step S03 when entropy increases abruptly. After fusing the docking correction with the real-time position error, the braking smooth curve is reconstructed by the local activation operator. In the streaming calculation, the high-frequency term of chattering is suppressed cycle by cycle, so that the control trajectory converges to the target docking position along the direction of smooth curvature change. This avoids the high-frequency chattering caused by the rigidity of the switching surface in traditional sliding mode control, and drives the vertical hoist to complete precise docking at multiple positions.

[0064] The specific implementation of step S06 is as follows: After each docking is completed, the elastic deformation characteristics, temperature field distribution data, docking correction amount, and docking accuracy error of this docking are written into the training dataset. Docking accuracy comprehensive evaluation index. From root mean square position error Temperature field drift of the guide rail With the increase of state entropy The weighted linear summation was used to calculate the weighting coefficients, which were determined through field experiments and data regression analysis. Below the lower threshold Use a smaller learning rate And disable the dynamic hierarchical early exit mechanism; when In to Use a moderate learning rate between And enable a dynamic hierarchical early exit mechanism; when Not less than Use a larger learning rate And triggers the reconnaissance bee algorithm to reconstruct the local activation operator, with three learning rates satisfying , proportion range The range is 5 to 20. The state entropy increase-constrained asynchronous multi-channel network uses the sum of mean squared error loss and state entropy increase boundary penalty term as the total loss function, and uses an adaptive moment estimation optimizer for parameter updates; the Gaussian process latent variable deep expansion model uses variational evidence lower bound as the objective function, and optimizes kernel hyperparameters and latent variable posterior distribution parameters simultaneously through variational inference algorithm. Both models are equipped with an early stopping mechanism, which stops training when the validation set index does not improve for 5 to 15 consecutive rounds, and iteratively optimizes the docking accuracy.

[0065] It should be noted that the key technologies of this invention include: First, the state entropy increase-constrained asynchronous multi-channel network uses a parallel architecture with three heterogeneous computing pipelines that are asynchronous and waiting for each other to block. Combined with the state entropy increase boundary penalty term, it forces the network to output a conservative correction amount in the high entropy increase state, which solves the problem of overconfidence in the control output leading to docking deviation when high-frequency switching causes entropy increase mutations. Second, the Gaussian process latent variable deep extension model embeds the elastic dynamic boundary equation of the flexible body into the prior mean function, and outputs the docking correction amount and its uncertainty estimate with nonparametric probability mapping. This enables the system to quantify the confidence interval of the correction amount in small sample cross-station migration scenarios, avoiding the extrapolation failure of the pure data-driven model near the physical boundary. Third, the artificial bee colony algorithm and the local activation operator work together to guide the bees and follower bees to search for the optimal control voltage in parallel multi-threaded searches. The scout bees reconstruct the local activation operator when entropy increase mutations occur. The spatial self-organizing evolution control law then dynamically adjusts the curvature of the braking stability curve with the reconstructed operator, organically combining global search capability and local stability control capability. With the synergistic effect of the three key technologies, feature perception, probability correction and smooth control form a complete closed loop, which together ensures the dynamic stability of docking accuracy under high-frequency switching conditions of multiple workstations, and overcomes the systematic defects of single compensation or single control methods in existing technologies under the coupling conditions of elastic deformation and thermal deformation.

[0066] It should be noted that in situations where the vertical hoist needs to traverse numerous workstations and experience frequent and significant changes in load mass, the system needs to complete multiple braking operations continuously within a very short time. Each braking operation involves significant differences in the elastic deformation of the wire rope, the thermal expansion at different locations on the guide rail, and the current speed curve compared to the previous stop. The dynamic drift of the target stopping position cannot be pre-calibrated using a fixed compensation coefficient. The reason for this technical problem is that the wire rope undergoes nonlinear elastic creep during frequent starts and stops at multiple workstations. The creep continuously drifts with the coupled changes in load mass, operating speed, and temperature field. The thermal expansion at different locations on the guide rail exhibits a spatially non-uniform distribution due to varying heat dissipation conditions across different sections. The combined effect of these two factors results in different actual offsets at each workstation's target position at each stop, and these offsets exhibit highly non-stationary characteristics as the system's operating state evolves. Any method relying on offline calibration models or fixed compensation coefficients cannot track this coupled drift in real time. The usual solution to the aforementioned technical problems is to establish a parameterized physical model that includes the elastic modulus and coefficient of thermal expansion, calibrate the model parameters offline through numerous experiments before the system goes live, and calculate the correction amount using the calibrated parameters during runtime. However, offline calibration cannot cover all load and temperature combinations, and the parameterized model generates systematic correction errors under boundary conditions due to model mismatch. Furthermore, the calibration parameters cannot be automatically updated with system aging and environmental changes, leading to a continuous decline in correction accuracy after long-term operation. This invention effectively solves this technical problem. The Gaussian process latent variable deep extension model embeds the elastic dynamic boundary equations of the flexible body in the form of a priori mean functions, which not only preserves the physical mechanism constraints to prevent boundary extrapolation failure, but also performs data-driven correction of the prior through nonparametric probability mapping. The deep hierarchical latent variable space progressively refines the abstract representation from the original elastic deformation characteristics to the high-level dynamic constraints, and the model can output reliable uncertainty estimates even in small-sample cross-station migration scenarios. The online adaptive update mechanism in step S06 allows the model parameters to continuously iterate with each docking data, and the docking accuracy comprehensive evaluation index The learning rate is adaptively switched to ensure that the model maintains its correction accuracy under system aging and environmental drift conditions, thus overcoming the systematic failure problem of offline calibration methods under dynamic coupling conditions from the root.

[0067] Specifically, the principle of this invention is as follows: This invention can solve the above-mentioned core technical problems, and its principle is as follows. First, the real-time extraction of elastic deformation features relies on the forward and reverse synchronous modeling of the time series of high-frequency multi-channel sensors in the bidirectional circular queue by the bidirectional long short-term memory unit. The reverse time series modeling can capture the phase delay introduced by high-frequency pulses in the sensor acquisition link. After splicing the forward and reverse hidden states, the linear projection compression is performed to eliminate the time domain lag error of traditional displacement sensors under high-frequency pulse conditions, thereby ensuring the time domain accuracy of the elastic deformation feature vector. Secondly, the state entropy increase constraint asynchronous multi-channel network projects heterogeneous sensor data into a low-dimensional orthogonal space through a shared latent space low-dimensional alignment operator, eliminating feature coupling interference between different physical dimensions. The three heterogeneous computing pipelines—dynamic feature channel, time-domain statistical channel, and frequency-domain transformation channel—are asynchronous and do not wait for blockages, but advance in parallel on different hardware units. When a channel is delayed due to sudden data jitter, the other channels can still support the confidence classifier to exit early and output control commands, thereby eliminating the accumulation of control delays during high-frequency switching of multiple stations. The state entropy increase dynamic constraint inside the network is embedded in the loss function in the form of a boundary penalty term, forcing the network to output a more conservative docking correction amount under high entropy increase state, avoiding over-adjustment caused by sudden changes in nonlinear thermal deformation of the guide rail. Furthermore, in the artificial bee colony algorithm, the lead bees conduct a coarse-grained search for candidate control voltages in a low-dimensional potential space, the follower bees perform weighted collaborative iterative refinement of the three flow channel outputs based on real-time error confidence, and the scout bees adaptively reconstruct the local activation operator parameters when entropy increases beyond the threshold. These three types of bee colonies operate in parallel with multiple threads, ensuring both real-time performance and global optimality of the control voltage. In addition, the Gaussian process latent variable deep extension model embeds the elastic dynamic boundary equations of the flexible body into the prior mean function. Through deep hierarchical latent variable spatial probability mapping, it outputs the docking correction amount and its uncertainty estimate for each workstation. This allows the system to quantify the confidence interval of the correction amount in small-sample cross-workstation migration scenarios and actively triggers a conservative control strategy when uncertainty is large, avoiding the extrapolation failure of the pure data-driven model near the physical boundary. Finally, the spatial self-organizing evolution control law treats the braking smooth curve as a self-organizing manifold in the position-velocity state space. It dynamically adjusts the local curvature of the manifold through local activation operators, suppressing high-frequency chattering terms in real-time during each control cycle, causing the control trajectory to converge to the target docking position along the smooth curvature change direction. The above-mentioned links logically form a complete closed loop from feature perception, asynchronous fusion, probability optimization to smooth control. Each link supports the others and jointly ensures the docking accuracy under high-frequency switching conditions at multiple workstations.

[0068] The following provides a specific embodiment 1 of the present invention, and the specific implementation of each step in this embodiment 1 is described in detail below.

[0069] The specific implementation of step S01 is as follows: Micrometer-resolution displacement sensors and temperature sensors are respectively installed on the drive unit and guide rail of the vertical hoist. The sampling frequency of the displacement sensors is... The recommended Hz range is 1000–5000 Hz, determined through multiple docking experiments and spectrum analysis based on the guide rail length and number of workstations. Temperature sensors are evenly distributed along the guide rail with appropriate spacing. The spacing is determined experimentally based on the thermal expansion coefficient and length of the guide rail, with a recommended range of 0.5–2 m. The acquired displacement and temperature signals are input into a bidirectional circular queue, with a queue depth of… satisfy:

[0070] ;

[0071] In the formula, The duration of the high-frequency pulse burst, measured in seconds. The unit is Hz. The unit is a sampling point, it is recommended Each sampling point. The high-order fusion algorithm for streaming time series data is based on bidirectional long short-term memory units, processing data from the queue. Road sensor time series Simultaneous forward and backward time-series modeling, forward hidden state With reverse hidden state After being stitched and compressed by a linear projection layer, the elastic deformation feature vector is obtained. The formula is expressed as follows:

[0072] ;

[0073] In the formula, Represents the forward hidden state vector With the reverse hidden state vector The splicing result, The linear projected weight matrix is ​​obtained through gradient descent training. Let be the projection bias vector. The recommended dimension range is 32 to 128 dimensions. Inverse time-series modeling eliminates the time-domain hysteresis error of traditional displacement sensors under high-frequency pulse conditions by capturing the phase delay introduced by high-frequency pulses.

[0074] The specific implementation of step S02 is as follows: The elastic deformation feature vector... With temperature field distribution data The concatenated input shares a low-dimensional alignment operator in the latent space. This operator is a two-layer fully connected network that projects the heterogeneous input to a dimension of [dimensional value missing]. The projection formula for the low-dimensional orthogonal feature space is as follows:

[0075] ;

[0076] In the formula, The low-dimensional aligned feature vector has a dimension of . , for and The concatenated vector, , The weight matrix consists of two fully connected layers, obtained through gradient descent training. , For the corresponding bias vector It is a linear rectified function with leakage, whose input is a real vector and output is a real vector of the same dimension. For each component... implement The empirical value for the leakage coefficient is 0.01. The data is then fed into three parallel flow channels: the dynamic feature flow channel is processed by a 3-5 layer convolutional neural network, which outputs a dynamic feature vector. The time-domain statistical flow path is processed by stacking 2 to 4 layers of locally asymmetric gated loop units, outputting a time-domain statistical feature vector. Its forward propagation gating time constant With feedback correction of gating time constant ratio Set to 3:1 to 8:1; the frequency domain transformation channel first... Apply a learnable discrete cosine transform basis The frequency domain coefficient vector is obtained after transformation. Then, a 2-3 layer fully connected network extracts the frequency domain statistical features and outputs a frequency domain statistical feature vector. ,in The learnable parameter matrix is ​​optimized using gradient descent during the training phase. The three outputs are then weighted and fused at the convergence layer to form a control feature vector. The formula is expressed as follows:

[0077] ;

[0078] In the formula, For the attention weights of each flow channel, satisfy the following conditions: The confidence level of each flow channel is dynamically allocated. Represents the output feature vector of the three flow channels , , The first in road.

[0079] The specific implementation method of step S03 is: with As a nectar source, it guides bee algorithms in low-dimensional potential spaces. The system explores a set of candidate control voltages that satisfy dynamic constraints using random perturbation. The formula is expressed as follows:

[0080] ;

[0081] In the formula, For the first Updated values ​​of candidate control voltages, in V. For the first The current candidate control voltages, in volts (V). The control voltage is a randomly selected neighboring candidate voltage from existing nectar sources, expressed in volts (V). For interval Uniformly random numbers, dimensionless. The number of nectar sources. The follower bee algorithm is based on the real-time error confidence level corresponding to each candidate control voltage. Probability-weighted sampling is performed, and the outputs of the three flow channels are refined through collaborative iteration. The confidence-weighted formula is expressed as follows:

[0082] ;

[0083] In the formula, For the first The probability that a nectar source will be chosen by a following bee. For the first The real-time error confidence level corresponding to each candidate control voltage is dimensionless. For the first The real-time error confidence level corresponding to each candidate control voltage. The reconnaissance bee algorithm addresses the increase in system state entropy. Exceeding the threshold When the current nectar source is abandoned and the local activation operator parameters are randomly re-initialized, a rapid response to sudden changes in the working conditions is achieved. The three types of bee colony algorithms run in parallel in a multi-threaded manner during the inference phase.

[0084] The specific implementation of step S04 is as follows: ... Input a Gaussian process latent variable deep extension model. The model has a total of [number] settings. There are several layers of latent variables, with the dimension of the latent variable space decreasing progressively in each layer, and the top layer having a dimension of 2 to 8. Layer Gaussian process kernel function The weighted combination of radial basis functions and linear kernels is expressed by the following formula:

[0085] ;

[0086] In the formula, and For the first Two input points in the layer's implicit variable space, For the first Layer signal variance hyperparameter For the first Layer radial basis function length-scale hyperparameter For the first Layer linear kernel weights, For the first The kernels are linearly biased across layers, and their hyperparameters are not shared between layers; all hyperparameters are automatically optimized during the training phase using a variational inference algorithm. This is the inner product of two latent variable vectors. The boundary equations of the elastic dynamics of a flexible body are based on the prior mean function. The form embeds a top-level Gaussian process to describe the combined effect of wire rope elastic creep and guide rail thermal expansion on the docking correction, expressed by the following formula:

[0087] ;

[0088] In the formula, This is a priori estimate of the docking correction output from the top-level prior mean function, in units of m. This is the dimension conversion factor, with units of . This transforms the elastic dynamics term from the dimension of force to the dimension of displacement. This refers to the elastic modulus of the steel wire rope, expressed in Pa. This represents the cross-sectional area of ​​the wire rope, in units of... , The displacement field of the wire rope is expressed in meters (m). The coordinates are along the rope direction, in meters. Time, in seconds. This refers to the linear density of the steel wire rope, in units of... , This is the temperature field drift correction factor, in units of... , This represents the temperature field drift of the guide rail, expressed in °C. This is the load quality correction factor, in units of , Load mass, in kg. The parameters, expressed in meters (m), are determined by regression fitting of at least 50 full-condition stop experiments. The model ultimately outputs the stop correction amount for each workstation. and its uncertainty variance , The unit is meters. Units are The conditional dependencies between latent variables at each level are parameterized using variational posterior distributions. Mean of layer output With variance Passed to the first The layer serves as the input to the prior mean function, where Unit and No. Hidden variables are consistent across layers. This represents the variance of the corresponding uncertainty.

[0089] The specific implementation of step S05 is as follows: The spatial self-organizing evolution control law is adopted to adjust the docking correction amount. With real-time position error The braking smooth curve is then dynamically reconstructed using a local activation operator. The local activation operator independently applies a nonlinear mapping to each dimension of the control feature vector, as expressed in the following formula:

[0090] ;

[0091] In the formula, For the first The output after mapping by the local activation operator, To control the feature vector The dimensional components, For the first The shape amplitude parameters are adaptively reconstructed by the reconnaissance bee algorithm during abrupt entropy increases. For the first The shape slope parameter is also adaptively reconstructed using the reconnaissance bee algorithm. The function is the hyperbolic tangent. The spatial self-organizing evolution control law treats the braking stationary curve as a self-organized manifold in the position-velocity state space, estimating the local curvature of the manifold in real time. The local activation operator parameters are updated to ensure that the control trajectory converges to the corrected target position along the direction of curvature smoothing. The formula is expressed as follows:

[0092] ;

[0093] In the formula, For the first The nominal target location of the workstation is in meters. The corrected target position is shown in meters. The unit for the corresponding workstation docking correction is meters (m). High-frequency chattering terms are suppressed in real-time during streaming calculations to drive the vertical hoist to achieve precise docking at multiple workstations.

[0094] The specific implementation of step S06 is as follows: After docking is completed, the elastic deformation characteristics of this docking are... Temperature field distribution data Dock correction amount docking accuracy error Write the training dataset to trigger online adaptive updates of the asynchronous multi-channel network with state entropy increase constraints and the Gaussian process latent variable deep expansion model. Total loss function. Defined as the sum of the mean squared error loss and the boundary penalty term for the increase in state entropy, the formula is as follows:

[0095] ;

[0096] In the formula, This represents the batch sample size. For the model to the first The stop correction amount predicted for each sample, in meters. For the first The annotation docking correction for each sample, in meters. This is the entropy increase boundary penalty coefficient, dimensionless, with a suggested range of [missing value]. ~ , The entropy increase threshold is dimensionless. The function takes the larger of the value within the parentheses and 0, ensuring the penalty term is non-negative. This function allows the network to only apply the penalty term to zero. Those subject to restrictions and penalties. Comprehensive evaluation index of docking accuracy. The definition formula is expressed as follows:

[0097] ;

[0098] In the formula, , , The weighting coefficients were determined through multiple field experiments and data regression analysis. This represents the root mean square position error of the most recent stops, in mm. This represents the temperature field drift of the guide rail, expressed in °C. The unit is , The unit is mm, so that The overall dimension is unified to mm equivalent. Learning rate And disable the dynamic hierarchical early exit mechanism; when Learning rate And enable a dynamic hierarchical early exit mechanism; when Learning rate Simultaneously, it triggers the reconnaissance bee algorithm to reconstruct the local activation operator, and the three learning rates satisfy... and , equivalent, The equivalent was determined by at least 30 docking tests covering all operating conditions, combined with 5-10 fold cross-validation regression analysis. State entropy increment. The Koolbek-Leibler divergence measure between the probability density estimate of the system state vector and the reference steady-state distribution is used, and the formula is expressed as follows:

[0099] ;

[0100] In the formula, The system state vector over the most recent 10 to 50 control cycles The estimated probability density, This is a reference steady-state probability density fitted to historical data under standard operating conditions. This is the system state vector that includes position error, velocity error, and temperature field drift. To normalize dimensionless quantities, and Same dimensions Dimensionless, after integration Dimensionless.

[0101] To better understand and implement this invention, the following is a specific application scenario of this invention, Example 2:

[0102] This embodiment focuses on a multi-station precision docking control test of a 12-station vertical lifting machine. The total length of the guide rail is 36m, with 12 docking stations and a 3m spacing between adjacent stations. The load range is 200-1200kg, the ambient temperature range is 10-45℃, and the coefficient of thermal expansion of the guide rail is [missing information]. ℃ A total of 14 displacement sensors are deployed, with a sampling frequency set to 2000Hz and a bidirectional circular queue depth set to [missing value]. There are 37 sampling points; temperature sensors are evenly distributed along the guide rail at 1m intervals. The elastic deformation feature vector dimension is set to 64 dimensions, and the low-dimensional orthogonal feature space dimension is set to 32 dimensions.

[0103] During the system initialization phase, 620 historical docking data points of the vertical hoist were first collected across all 12 workstations, the entire load range, and the entire temperature range, with high-entropy increase samples accounting for 42%. Each sample was labeled with docking accuracy error, elastic deformation amplitude, and guide rail thermal expansion. The prior docking correction data were then fitted with the flexible body elastic dynamics boundary equations, with 58 fitting experiments performed. The training, validation, and test sets were divided in a 7:2:1 ratio. Figure 2 As shown, the data flow of the three parallel channels of the state entropy increasing constraint asynchronous multi-channel network on the hardware computing unit is as follows: the dynamic feature channel receives low-dimensional aligned features of dimension 32, and outputs 128-dimensional local structural features through a 4-layer convolutional neural network; the temporal statistics channel outputs 64-dimensional temporal statistical features through a 3-layer local asymmetric gated recurrent unit, with the ratio of the forward propagation gate time constant to the feedback correction gate time constant set to 5:1; the frequency domain transformation channel outputs 64-dimensional frequency domain statistical features through a learnable discrete cosine transform basis and a 2-layer fully connected network; the three features are concatenated into a 256-dimensional vector at the convergence layer and then fused into a 128-dimensional control feature vector through attention weighting.

[0104] The Gaussian process latent variable deep extension model has four latent variable layers with dimensions of 32, 16, 8, and 4 respectively, and a top-level layer with 4 dimensions. The hyperparameters of the kernel functions in each layer are optimized using a variational inference algorithm on 620 training samples. Early stopping rounds are set to 10 rounds; training stops when the lower bound of the variational evidence on the validation set fails to improve for 10 consecutive rounds. The variance threshold is set to... mm When the value falls below this threshold, a dynamic hierarchical early exit mechanism is triggered.

[0105] like Figure 3 As shown in Table 1, the distribution of docking correction amounts for each workstation under different load conditions in this embodiment is as follows:

[0106] Table 1. Statistics of Dock Correction Quantities under Representative Loads at Each Workstation

[0107]

[0108] As shown in Table 1, with the increase of the workstation number, both the elastic deformation and the thermal expansion correction increase accordingly, and the overall docking correction shows a monotonically increasing trend. This trend stems from the physical mechanism that the increase in guide rail length leads to the accumulation of elongation in the wire rope suspension section, coupled with the greater accumulation of thermal expansion in the lower section of the guide rail. This indicates that the fixed compensation coefficient cannot cover the correction changes across the entire workstation range, while the Gaussian process latent variable deep extension model of this invention achieves adaptive modeling of this trend through workstation-by-workstation probability mapping.

[0109] During the formal operation phase, a high-frequency round-trip switching scenario between workstations 6 and 9 was selected for verification. The switching cycle was 45 seconds, and a total of 40 round-trip stops were performed. In the artificial bee colony algorithm, the size of the leader bee colony was set to 20, the size of the follower bee colony was set to 20, and the size of the scout bee colony was set to 5. The entropy increase threshold was determined by fitting a reference steady-state distribution based on historical data under standard operating conditions. Figure 4 As shown, the comprehensive evaluation index of docking accuracy The curve showing the change with the number of stops exhibits the following trend: During the first 10 stops, because the model has not yet accumulated enough online update data, The value fluctuated between 1.2 and 1.8 mm equivalent, and the learning rate was in the medium range. During the 10th to 25th stops, as the amount of data updated online adaptively increased, As the value gradually decreases to 0.6–0.9 mm equivalent, the learning rate is switched to a lower setting. The dynamic hierarchical early exit mechanism is disabled; during the 26th to 40th docking, under the condition that the load suddenly changes from 800kg to 1200kg, the system state entropy increases. If the entropy increase threshold is briefly exceeded, the reconnaissance bee algorithm is triggered. After the local activation operator parameters are reconstructed, the curvature of the braking stationary curve is adjusted accordingly. The value recovered to below 0.8 mm equivalent within 3 dockings. (The above...) The decreasing trend of the value with the number of stops reflects, from a physical perspective, that the online adaptive update mechanism gradually approximates the true coupling law of elastic deformation and thermal expansion under the current working condition, reducing the systematic deviation of the stop correction amount. Meanwhile, the reconnaissance bee algorithm quickly reconstructs the local activation operator when entropy increases abruptly, avoiding chattering errors caused by curvature mismatch in the braking smooth curve under sudden load changes, demonstrating the robustness of the invention under dynamic coupling conditions.

[0110] like Figure 5 As shown in Table 2, the position error distribution for 40 round trip stops is as follows:

[0111] Table 2. Statistical Table of Stopping Position Errors for 40 Round Trips

[0112]

[0113] As shown in Table 2, the root mean square position error decreases with the number of stops, decreasing from 1.54 mm in the first stage to 0.82 mm in the second stage, reflecting the continuous correction effect of the online adaptive update mechanism on the model parameters. In the third stage, the error slightly rebounded due to a sudden entropy increase, but it recovered quickly after the intervention of the scout bee algorithm. The maximum position error decreased from 2.87 mm in the initial stage to 1.43 mm in the second stage, indicating that the uncertainty estimation of the Gaussian process latent variable deep extension model effectively guided the triggering of the conservative control strategy and avoided excessive output of correction.

[0114] In this embodiment, the main structure of the state entropy increase constraint asynchronous multi-channel network is as follows: the shared latent space low-dimensional alignment operator adopts a two-layer fully connected network, with an input of 94 dimensions (a 64-dimensional elastic deformation feature vector concatenated with a 30-dimensional temperature field feature), and an output of 32-dimensional low-dimensional aligned features; the dynamic feature channel adopts a 4-layer convolutional neural network with a kernel size of 1×3, and the number of channels in each layer is 32, 64, 128, and 128 respectively; the time-domain statistical channel adopts a 3-layer locally asymmetric gated recurrent unit, with a hidden state dimension of 64; the frequency domain transform channel adopts a learnable discrete cosine transform basis (32 basis functions) plus a 2-layer fully connected network, with a hidden layer dimension of 64. Figure 6 As shown, the outputs of the three channels are spliced ​​in the convergence layer and then fused into a 128-dimensional control feature vector by two-head attention weighting. The attention weights are dynamically allocated by the current confidence of each channel.

[0115] As an equivalent replacement, the locally asymmetric gated recurrent unit in the time-domain statistical channel can be replaced with a temporal convolutional network. This network consists of multiple layers of causal dilated convolutions with exponentially increasing dilation coefficients. The input is a 32-dimensional low-dimensional aligned feature sequence, with the number of channels in each layer being 64, 64, and 128 respectively. It outputs 64-dimensional time-domain statistical features through residual connections. The data flow is consistent with that of the locally asymmetric gated recurrent unit, both using low-dimensional aligned feature sequences as input and time-domain statistical features as output. The learnable discrete cosine transform basis in the frequency-domain transform channel can be replaced with a learnable Fourier transform basis. After transforming the input sequence to the frequency domain, the real and imaginary parts are concatenated, and then 64-dimensional frequency-domain statistical features are extracted through a two-layer fully connected network. The feature extraction dimension and the convergence target of the loss function are highly consistent with the main scheme. This is an equivalent replacement that can be implemented by those skilled in the art without creative effort.

[0116] It should be noted that the variables involved in this invention are explained in detail in Tables 3 and 4.

[0117] Table 3. Variable Explanation Table (Part 1)

[0118]

[0119] Table 4. Variable Explanation Table (Part Two)

[0120]

[0121] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for precise multi-station docking control of a vertical lifting machine, characterized in that, Includes the following steps: Displacement sensors and temperature sensors are installed on the drive unit and guide rail of the vertical hoist to collect the elastic deformation signal of the wire rope and the temperature field distribution data of the guide rail. The signals are then connected to a bidirectional circular queue, and the elastic deformation features are extracted in real time by a high-order fusion algorithm of streaming time series. The elastic deformation characteristics and temperature field distribution data are input into the state entropy increase constraint asynchronous multi-channel network, projected to a low-dimensional orthogonal space by the shared latent space low-dimensional alignment operator, and processed in parallel by three channels: dynamic feature channel, time domain statistical channel and frequency domain transformation channel, and output control feature vector. Using the control feature vector as the nectar source in the artificial bee colony, the guiding bee algorithm searches for the initial control voltage that satisfies the dynamic constraints in the low-dimensional potential space, the following bee algorithm performs weighted collaborative iterative optimization of the three flow channel outputs based on the real-time error confidence, and the reconnaissance bee algorithm adaptively reconstructs the local activation operator when the system entropy increases abruptly. The control feature vector is synchronously input into the Gaussian process latent variable deep expansion model. After deep hierarchical latent variable spatial probability mapping, the prior mean function driven by the elastic dynamic boundary equation of the flexible body corrects the target position of each station and outputs the docking correction amount. The spatial self-organizing evolution control law is adopted to integrate the docking correction amount with the real-time position error. The braking smooth curve is dynamically reconstructed through a local activation operator. In the streaming calculation, the high-frequency term of chattering is suppressed in real time, and the vertical hoist is driven to complete the multi-station precise docking at the corrected target position.

2. The method for precise multi-station docking control of a vertical lifting machine according to claim 1, characterized in that, The bidirectional circular queue is a first-in-first-out ring buffer structure. The queue depth is determined experimentally based on the product of the duration of the high-frequency pulse burst and the sampling frequency. When the rate of new sampled data arrival exceeds the processing rate, the head of the queue automatically overwrites the oldest data.

3. The method for precise multi-station docking control of a vertical lifting machine according to claim 2, characterized in that, The high-order fusion algorithm for streaming time series is based on bidirectional long short-term memory units. It performs forward and reverse time series modeling on the time series of multiple sensors in a bidirectional circular queue, extracts the elastic deformation feature vector containing elastic creep hysteresis correction, captures the phase delay through reverse time series modeling, and then compresses the forward and reverse hidden states through a linear projection layer to eliminate time domain hysteresis error.

4. The method for precise multi-station docking control of a vertical lifting machine according to claim 3, characterized in that, The shared latent space low-dimensional alignment operator is a two-layer fully connected network that projects heterogeneous input data from different sensors onto a low-dimensional orthogonal feature space. The interlayer activation function uses a linear rectified function with leakage.

5. The method for precise multi-station docking control of a vertical lifting machine according to claim 4, characterized in that, In the state entropy increase constraint asynchronous multi-channel network, the dynamic feature channel is composed of a convolutional neural network, the time-domain statistical channel is composed of stacked local asymmetric gated recurrent units, the frequency domain transformation channel applies a learnable discrete cosine transform basis to the input time series and then extracts the frequency domain statistical features by a fully connected network, and the outputs of the three channels are spliced ​​in the convergence layer and then fused into a control feature vector by attention weighting.

6. The method for precise multi-station docking control of a vertical lifting machine according to claim 5, characterized in that, The ratio of the forward propagation gating time constant to the feedback correction gating time constant of the locally asymmetric gated loop unit satisfies the asymmetric setting value, with fast forward propagation and slow feedback correction to ensure the real-time performance of streaming inference.

7. The method for precise multi-station docking control of a vertical lifting machine according to claim 6, characterized in that, The state entropy increase constraint asynchronous multichannel network is internally configured with multiple confidence classifiers based on the state entropy increase threshold. When the posterior entropy is lower than the preset threshold, a dynamic hierarchical early exit mechanism is triggered, and docking control commands are directly output. The weight matrix of each layer in the network adopts sparse mask constraint, and the sparsity rate is determined through experimental iteration based on the activation frequency statistics of neurons in each layer.

8. The method for precise multi-station docking control of a vertical lifting machine according to claim 7, characterized in that, In the artificial bee colony algorithm, the leader bee, follower bee, and scout bee run in parallel in a multi-threaded manner during the inference phase. When the entropy of the system state exceeds the entropy increase threshold, the scout bee algorithm abandons the current nectar source and re-randomly initializes the local activation operator parameters.

9. The method for precise multi-station docking control of a vertical lifting machine according to claim 8, characterized in that, The Gaussian process latent variable deep extension model sets up multiple latent variable layers. The latent variable space of each layer is dynamically defined by an independent Gaussian process kernel function. The kernel function adopts a weighted combination of radial basis functions and linear kernels. The kernel hyperparameters are automatically optimized during the training phase through variational inference algorithm.

10. The method for precise multi-station docking control of a vertical lifting machine according to claim 9, characterized in that, The Gaussian process latent variable deep expansion model is equipped with an adaptive parameter dynamic weight update algorithm. It dynamically adjusts the update step size of the kernel hyperparameters of each layer according to the real-time estimated value of the variance of each latent variable. When the output variance of the lower-level latent variable is lower than the variance threshold, the docking correction amount is directly output.

Images