Intelligent control system for dense cabinet movement

Abstract

This invention discloses an intelligent control system for moving mobile shelving units, belonging to the field of intelligent control. It includes a sensing and detection module, a modal fusion module, an identification and prediction module, a parameter correction module, a predictive control module, a speed adjustment module, a monitoring and suppression module, a drive control module, a braking recovery module, a management optimization module, a log auditing module, and a monitoring and visualization module. This invention significantly reduces overshoot and swaying, improves stopping accuracy, enables early warning and graded braking, reduces external energy consumption, improves overall energy efficiency and system endurance, and achieves fully closed-loop intelligent control from prediction and execution to correction.

Description

Technical Field

[0001] This invention relates to the field of intelligent control, and more particularly to an intelligent control system for the movement of mobile shelving units. Background Technology

[0002] With the rapid development of informatization and intelligent manufacturing, the requirements for space utilization, security, and automation levels in fields such as archives management, library storage, intelligent warehousing, and equipment management are constantly increasing. As an important device for achieving high-density storage and optimized space utilization, mobile shelving systems have evolved from traditional mechanical structures towards electrification, networking, and intelligence. However, existing mobile shelving motion control systems still have many shortcomings in terms of technical implementation and operational performance. First, in terms of motion control accuracy, traditional electric mobile shelving mainly relies on position encoders and simple PID control algorithms to achieve start-stop speed regulation. This type of control cannot fully consider nonlinear factors such as changes in cabinet mass, differences in track friction, and inertial offset, leading to stop point errors, overshoot, and swaying under high-speed operation or heavy load conditions, affecting cabinet positioning accuracy and user safety. Second, in terms of environmental perception and safety protection, most systems only rely on infrared or proximity sensors to detect obstacles, resulting in delayed response and a high false alarm rate to dynamic obstacles (such as people accidentally entering or falling objects). The lack of multimodal fusion and dynamic recognition capabilities makes mobile shelving potentially susceptible to collisions and trapping risks in complex or personnel-interacting environments. Against this backdrop, the research on intelligent mobile shelving control systems has become an inevitable trend in the industry.

[0003] Existing intelligent control systems for moving mobile shelving units often exhibit overshoot and swaying phenomena, reducing stopping accuracy, failing to provide early warning and graded braking, increasing external energy consumption, and reducing overall energy efficiency and system endurance. Therefore, we propose an intelligent control system for moving mobile shelving units. Summary of the Invention

[0004] The purpose of this invention is to address the shortcomings of existing technologies by proposing an intelligent control system for the movement of mobile shelving units.

[0005] To achieve the above objectives, the present invention adopts the following technical solution: An intelligent control system for moving mobile shelving units includes a sensing and detection module, a modal fusion module, an identification and prediction module, a parameter correction module, a predictive control module, a speed adjustment module, a monitoring and suppression module, a drive control module, a braking and recovery module, a management optimization module, a log auditing module, and a monitoring and visualization module. The sensing and detection module is used to collect information on the operating environment and equipment status of the mobile shelving, and to monitor the cabinet position, load weight, friction changes and surrounding environmental conditions in real time. The modal fusion module constructs a real-time three-dimensional environment map by fusing millimeter-wave radar and 3D depth camera data; The identification and prediction module identifies obstacle types and analyzes the movement trends of dynamic targets based on a real-time 3D environment map. The parameter correction module calculates the current cabinet load mass through pressure sensor and motor current feedback, and corrects the motion parameters accordingly. The predictive control module predicts the stopping position of the cabinet in real time based on historical motion data and dynamically adjusts the brake trigger point and deceleration curve. The speed adjustment module adjusts the braking intensity and motor braking current in real time based on the predictive control results and the current load status. The monitoring and suppression module is used to analyze vibration signals during operation and compensate for them through motor output torque. The drive control module is used to receive control commands and control the cabinet to start, accelerate, maintain constant speed, decelerate and brake. The regenerative braking module is used to dynamically allocate and recover electrical energy, and to compensate for peak power consumption during motor startup or system standby power supply. The management optimization module is used to monitor energy status in real time and optimize charging and discharging strategies. The log auditing module is used to record all operation events and synchronize them to the cloud auditing platform; The monitoring and visualization module is used to display the cabinet's operating status, energy flow, obstacle distribution, and control logs in real time.

[0006] As a further aspect of the present invention, the specific steps for the modality fusion module to construct a real-time three-dimensional environment map are as follows: S1.1: Receive the raw radar points and camera point clouds collected in real time by millimeter-wave radar and 3D depth camera, remove noise from each radar point and camera point cloud by moving window filtering, remove outliers in each radar point and camera point cloud by box line method, and then resample each radar point and camera point cloud to the target synchronization time by time interpolation. S1.2: Use offline or online calibration to obtain the rotation matrix and translation vector from the camera to the aircraft, as well as the rotation and translation from the radar to the aircraft. Then, back-project the camera point cloud according to the camera intrinsic parameters to obtain 3D points in the camera coordinate system. Then, map it to the aircraft coordinate system through rigid body transformation. Then, directly use rigid body transformation to map the radar points to the aircraft coordinate system. After that, perform unit unification and scale correction on the transformed point cloud to generate radar points and camera points in the aircraft coordinate system. S1.3: Perform coarse registration between radar points and camera points and divide the space by voxels. Statistically count the observation sets of each voxel from various sensors. Then, perform weighted merging based on observation confidence and measurement variance to obtain the occupancy probability of each voxel. After that, perform Euclidean clustering on the merged voxels to divide the spatial points into candidate target clusters and calculate the centroid, velocity and shape characteristics of each cluster. S1.4: Calculate the Mahalanobis distance between each existing historical trajectory and each currently detected cluster. If there is a Mahalanobis distance lower than a preset threshold, the corresponding historical trajectory and cluster are used as candidate matches. If a cluster has multiple candidate matches, the redundant candidate matches are filtered out using the Hungarian algorithm. Then, the motion state of each matched target is updated in real time using an acceleration model to generate a dynamically updated real-time 3D environment map.

[0007] As a further aspect of the present invention, the specific steps of the identification and prediction module in identifying obstacle types and analyzing the movement trends of dynamic targets are as follows: S2.1: Initialize the log odds value of each voxel in the 3D environment map, then update the occupied probability of the voxels related to the observation according to the observation model, and then convert the log odds value into occupied, idle or unknown labels, while retaining the class probability of each voxel. Iterate through all voxels marked as "occupied" in the latest 3D environment map, divide the occupied voxels into multiple connected components, and treat each connected component as a candidate instance. S2.2: Construct the corresponding minimum bounding box for each connected component, and extract the point cloud subset within the bounding box as the original dataset of the instance. Extract the geometric boundary and timestamp information of each candidate instance, store each set of extracted information into the local instance table according to the unique instance ID, and then calculate the basic geometric quantities of the original dataset of each candidate instance. S2.3: Extract the feature values ​​of each candidate instance and normalize them to generate corresponding shape descriptors. Calculate the mean and variance of the point reflection intensity and the occurrence frequency and position variance of each candidate instance within the preset time frame. Then, input all the data into the trained classifier model. The classifier model outputs the non-normalized score of each category based on the forward propagation algorithm, and then converts it into a category probability distribution through softmax. S2.4: Based on the output category probability distribution, construct the category transition matrix, and then use Bayes-Markov update to combine the current frame's instantaneous classification probability with the transition prior probability to generate the corresponding temporal consistency probability. Read the current tracked state vector and velocity uncertainty index from the local instance table. S2.5: Progress the state step by step according to the preset time step and record the predicted position at each step. At the same time, calculate the position uncertainty growth at each predicted time step, collect the position-velocity sequence of each candidate instance at multiple times, and input each position-velocity sequence into the trained sequence prediction network. At the same time, output the trajectory samples or parameterized distribution of multiple future times through the sequence prediction network. S2.6: Based on the time consistency probability and the multimodal position distribution of each future step, calculate the occupancy probability of each candidate instance in the influence area of ​​the device's motion trajectory. Then, perform weighted accumulation based on the corresponding category probability distribution to generate a corresponding risk score. If the risk score exceeds the preset alarm threshold, an alarm will be issued at the control terminal to remind the staff.

[0008] As a further aspect of the present invention, the specific steps of the predictive control module in predicting the stopping position of the cabinet in real time and dynamically adjusting the brake trigger point and deceleration curve are as follows: S3.1: Collect data segments from multiple time steps in the real-time data stream as input sequences, and perform scale normalization and detrending processing on each input sequence. At the same time, establish the corresponding input matrix according to the collection time order of each input sequence. S3.2: The established input matrix is ​​transmitted to the LSTM network. The LSTM network captures the medium-to-long-term time dependencies of various types of data in the input matrix based on the gating mechanism and outputs the hidden state and candidate information at each time step. Based on the output of the LSTM sub-network of each modality data, a corresponding set of representation vectors is established. S3.3: Regress the set of representation vectors through a fully connected layer, and output the mean of the predicted stopping position and the prediction uncertainty. Based on the prediction uncertainty, the current encoder position and the current speed, calculate the remaining travel distance from the corresponding cabinet to the target stopping position. S3.4: Calculate the safe trigger distance based on the current speed, the maximum braking acceleration that the system can provide, and the system delay. If the remaining travel is less than or equal to the safe trigger distance and the system is not braking, then issue a braking command. If the system is already braking, then gradually adjust the target braking force according to the planned deceleration curve. S3.5: Based on the prediction uncertainty and the maximum braking acceleration, dynamically calculate the shape of the deceleration curve and generate the corresponding time series acceleration trajectory. Then, send the obtained deceleration curve to the driver and execute the corresponding deceleration strategy.

[0009] As a further aspect of the present invention, the data segment in S3.1 specifically includes: encoder position reading, timestamp, load indication of motor current estimation, current braking command status, and the time interval between the most recent braking start.

[0010] As a further aspect of the present invention, the specific steps of the speed adjustment module in real-time adjusting the braking intensity and the motor braking current are as follows: S4.1: Collect the stop position prediction and its uncertainty output by the predictive control module, as well as the current instantaneous motion measurement, and read the load estimate. Then, map the stop position prediction to the motion axis and calculate the remaining displacement error. If the difference between the stop position prediction and the current motion state exceeds the set tolerance, it is marked as a high uncertainty situation, and the conversion factor of the next time step is adjusted. S4.2: Map the uncertainty corresponding to the stop position prediction to a conservative factor, and calculate the corresponding nominal average deceleration through the remaining displacement error and the current velocity. At the same time, amplify the result through the conservative factor and output the target average deceleration. Then, convert the target average deceleration into the braking force required for the drive shaft. S4.3: Map the braking force required by the drive shaft to the torque required by the motor, and calculate the target torque on the corresponding motor side or brake side based on the actual effective lever arm and transmission ratio, as well as the constraints of mechanical efficiency and friction loss of the braking device. S4.4: If braking requires energy recovery and the power supply can accept it, then all current is allowed to be in the regenerative direction; if energy recovery is saturated or voltage limitation exists, then the braking current is limited and switched to consumable braking or mechanical braking. Based on the judgment result, the target torque is converted into a current command at the motor end. S4.5: Based on the maximum and minimum allowable current values, adjust the current command and send the adjusted current command to the driver. The driver executes the current command and reads the actual current, actual motor torque estimate or load displacement response in real time during execution, and calculates the execution error. If the actual effect is lower than the preset expectation, the current command is feedforward corrected, and the local gain is updated or the bias is corrected based on the observed long-term system deviation.

[0011] As a further aspect of the present invention, the specific steps for the log auditing module to record all operation events are as follows: S5.1: Map all operation events generated by the device into a standardized set of transactions. Each transaction contains a unique transaction ID, a timestamp string, an event type code, an event payload, and optional local node metadata. Compact the transaction payload and calculate a summary of each transaction. S5.2: Place multiple sets of transactions into the list to be packaged in the order of collection, and generate a transaction summary list for the packaged set. Construct the corresponding binary tree based on the transaction summary list. The bottom layer of the binary tree is the transaction summary. From bottom to top, perform binary concatenation on the binary tree, extract the hash value of the concatenation result, obtain the summary of the corresponding parent node, and recursively go to the root node to generate the root summary of the block. S5.3: Record the proof path corresponding to each transaction, and merge the root digest and block header metadata to generate the corresponding block. Use the digest reference of the previous block header, the root digest of the current block, the current edge node identifier, the timestamp of this block, the sequence number of this block, and the status metadata as the current block header fields. S5.4: After arranging the block header fields in a fixed order, calculate the block header digest, then concatenate the current block header digest with the previous block header digest to form a hash chain reference, and then use the private key of the current edge node to digitally sign the block header digest; S5.5: The block header generated and signed on the chain by the edge node is periodically synchronized with the block digest and signature to the cloud or multiple trusted audit nodes. At the same time, the block header received by the edge node from other nodes is verified to verify the validity of the block header signature and its consistency with the hash reference of the previous block. S5.6: When the auditor or local query requests to verify whether a transaction is recorded on the chain, the edge node provides the coded content of the transaction, the transaction summary, and the corresponding proof path. The verifier recalculates the transaction summary and iteratively calculates the hash value of the parent node according to the proof path to finally obtain the calculated root node. At the same time, it verifies whether the root node in the provided block header is consistent with the calculated root, and then verifies whether the signature of the block header hash is valid to obtain the proof of the existence and integrity of the transaction.

[0012] Compared with the prior art, the beneficial effects of the present invention are as follows: This invention establishes a voxel occupancy model in a 3D environment map, integrates sensor observations to update the probability of each voxel, extracts occupancy areas to form candidate instances, normalizes the instance features, inputs them into a classification network, outputs class probabilities, and performs temporal consistency updates using a Bayesian-Markov method. Subsequently, combining the motion sequence information of the instances, a sequence prediction network is used to predict their future trajectory and position distribution, calculates the risk probability of dynamic targets within the equipment path, and generates risk scores. If a threshold is exceeded, an alarm is triggered. Simultaneously, multi-source motion data is collected in parallel, normalized and detrended, and then input into an LSTM network to output the cabinet's stopping position and uncertainty. The system calculates the safe trigger distance and deceleration curve based on the current speed and braking characteristics. According to the load status and prediction error, it dynamically corrects the target deceleration and braking torque, and generates the final current command according to the energy recovery and current limit conditions. Finally, it encapsulates all operation events into blockchain transactions, forms an immutable record structure through hash tree and chain encryption, and ensures data integrity and traceability with digital signature and multi-node verification mechanism. This significantly reduces overshoot and swaying, improves stopping accuracy, realizes early warning and graded braking, reduces external energy consumption, improves overall energy efficiency and system endurance, and realizes a closed-loop intelligent control from prediction, execution to correction. Attached Figure Description

[0013] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used together with the embodiments of the invention to explain the invention and do not constitute a limitation thereof.

[0014] Figure 1 This is a system block diagram of an intelligent control system for moving mobile shelving units proposed in this invention. Detailed Implementation

[0015] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Example

[0016] Reference Figure 1 An intelligent control system for moving mobile shelving units includes a sensing and detection module, a modal fusion module, an identification and prediction module, a parameter correction module, a predictive control module, a speed adjustment module, a monitoring and suppression module, a drive control module, a braking and recovery module, a management optimization module, a log auditing module, and a monitoring and visualization module.

[0017] The perception and detection module is used to collect information on the operating environment and equipment status of the mobile shelving, and to monitor the cabinet position, load weight, friction changes and surrounding environmental conditions in real time; the modal fusion module constructs a real-time three-dimensional environmental map by fusing millimeter-wave radar and 3D depth camera data.

[0018] Specifically, the system receives raw radar points and camera point clouds acquired in real time by millimeter-wave radar and 3D depth cameras. Noise in each radar point and camera point cloud is removed using a moving window filter. Outliers in each radar point and camera point cloud are then removed using box-line scattering. Next, each radar point and camera point cloud is resampled to the target synchronization time using time interpolation. The rotation matrix and translation vector from the camera to the aircraft, as well as the rotation and translation of the radar to the aircraft, are obtained using offline or online calibration. The camera point cloud is then back-projected according to the camera's intrinsic parameters to obtain 3D points in the camera coordinate system. These points are then mapped to the aircraft coordinate system using rigid body transformation. Rigid body transformation is then directly used to map the radar points to the aircraft coordinate system. Finally, the transformed point cloud undergoes unit unification and scale correction to generate radar points and phase vectors in the aircraft coordinate system. The system performs coarse registration of radar and camera points and divides the space by voxels. It statistically analyzes the observation sets of each voxel from various sensors, and then performs weighted merging based on observation confidence and measurement variance to obtain the occupancy probability of each voxel. Then, it performs Euclidean clustering on the merged voxels to divide the spatial points into candidate target clusters, and calculates the centroid, velocity, and shape characteristics of each cluster. It calculates the Mahalanobis distance between each existing historical trajectory and each currently detected cluster. If there is a Mahalanobis distance below a preset threshold, the corresponding historical trajectory and cluster are used as candidate matches. If a cluster has multiple candidate matches, the redundant candidate matches are filtered out using the Hungarian algorithm. Finally, it uses an acceleration model to update the motion state of each matched target in real time to generate a dynamically updated real-time 3D environment map.

[0019] The recognition and prediction module identifies obstacle types and analyzes the movement trends of dynamic targets based on a real-time 3D environment map.

[0020] Specifically, the log-odds value of each voxel in the 3D environment map is initialized. Then, the voxels related to the observation are updated with occupancy probabilities according to the observation model. The log-odds values ​​are then converted into occupied, idle, or unknown labels, while retaining the class probability of each voxel. All voxels marked as "occupied" in the latest 3D environment map are traversed, and occupied voxels are divided into multiple connected components. Each connected component is considered a candidate instance. A minimum bounding box is constructed for each connected component, and a subset of the point cloud within the bounding box is extracted as the original dataset for the instance. The geometric boundaries and timestamp information of each candidate instance are extracted, and each set of extracted information is stored in a local instance table with a unique instance ID. Then, the basic geometric quantities of the original dataset for each candidate instance are calculated, the feature values ​​of each candidate instance are extracted, and they are normalized to generate corresponding shape descriptors. The mean and variance of point reflection intensity are calculated, and the frequency and location variance of each candidate instance within a preset time frame are calculated. Finally, all data are input into a trained classifier model, which outputs each category based on the forward propagation algorithm. The non-normalized score is then converted into a class probability distribution using softmax. Based on the output class probability distribution, a class transition matrix is ​​constructed. A Bayesian-Markov update is then used to combine the instantaneous classification probability of the current frame with the transition prior probability to generate the corresponding temporal consistency probability. The current tracked state vector and velocity uncertainty index are read from the local instance table. The state is progressively advanced according to a preset time step, and the predicted position is recorded at each step. At the same time, the position uncertainty growth at each predicted time step is calculated. The position-velocity sequences of each candidate instance at multiple time steps are collected, and each position-velocity sequence is input into a trained sequence prediction network. The sequence prediction network outputs trajectory samples or parameterized distributions for multiple future time steps. Based on the temporal consistency probability and the multimodal position distribution at each future step, the occupancy probability of each candidate instance within the influence area of ​​the device's motion trajectory is calculated. Then, the corresponding risk score is generated by weighted accumulation according to the corresponding class probability distribution. If the risk score exceeds the preset alarm threshold, an alarm is issued at the control terminal, and staff are alerted.

[0021] The parameter correction module calculates the current cabinet load mass through pressure sensor and motor current feedback, and corrects the motion parameters; the predictive control module predicts the cabinet's stopping position in real time based on historical motion data, and dynamically adjusts the brake trigger point and deceleration curve.

[0022] Specifically, multiple time-step data segments are collected from the real-time data stream as input sequences. Each input sequence undergoes scale normalization and detrending processing. Simultaneously, a corresponding input matrix is ​​established based on the acquisition time order of each input sequence. This input matrix is ​​then transmitted to an LSTM network. The LSTM network uses a gating mechanism to capture the medium- to long-term dependencies of various data types in the input matrix and outputs the hidden states and candidate information for each time step. Based on the output of the LSTM sub-network for each modality, a corresponding set of representation vectors is established. Regression is performed on this set of representation vectors through a fully connected layer, outputting the mean of the predicted stopping positions and the predicted... Uncertainty is assessed based on the predicted uncertainty, the current encoder position, and the current speed. The remaining travel distance from the corresponding cabinet to the target stopping position is calculated. The safe trigger distance is calculated based on the current speed, the maximum braking acceleration that the system can provide, and the system delay. If the remaining travel distance is less than or equal to the safe trigger distance and the system is not braking, a braking command is issued. If the system is already braking, the target braking force is gradually adjusted according to the planned deceleration curve. The shape of the deceleration curve is dynamically calculated based on the predicted uncertainty and the maximum braking acceleration, and the corresponding time-series acceleration trajectory is generated. The acquired deceleration curve is then sent to the driver, and the corresponding deceleration strategy is executed.

[0023] It should be further noted that the data segment specifically includes: encoder position reading, timestamp, load indication of motor current estimation, current braking command status, and the time interval between the most recent braking start. Example

[0024] Reference Figure 1 An intelligent control system for moving mobile shelving units includes a sensing and detection module, a modal fusion module, an identification and prediction module, a parameter correction module, a predictive control module, a speed adjustment module, a monitoring and suppression module, a drive control module, a braking and recovery module, a management optimization module, a log auditing module, and a monitoring and visualization module.

[0025] The speed adjustment module adjusts the braking intensity and motor braking current in real time based on the predictive control results and the current load status.

[0026] Specifically, the system acquires the stop position prediction and its uncertainty from the predictive control module, along with the current instantaneous motion measurement. It also reads the load estimate. The stop position prediction is then mapped to the motion axis, and the remaining displacement error is calculated. If the difference between the stop position prediction and the current motion state exceeds a set tolerance, it is marked as a high uncertainty situation. Simultaneously, the conversion factor for the next time step is adjusted, mapping the uncertainty corresponding to the stop position prediction to a conservative factor. The nominal average deceleration is calculated using the remaining displacement error and the current velocity, and amplified using the conservative factor to output the target average deceleration. This target average deceleration is then converted into the braking force required by the drive shaft. The braking force required by the drive shaft is mapped to the required torque value of the electric motor, based on the actual effective lever arm, transmission ratio, and braking device. Given the constraints of mechanical efficiency and friction loss, the target torque is calculated for the corresponding motor side or brake side. If braking requires energy feedback and the power supply can accept it, all current is allowed to be in the regenerative direction. If energy recovery is saturated or voltage limitations exist, the braking current is limited and replaced with consumable braking or mechanical braking. Based on the judgment result, the target torque is converted into a current command at the motor end. Based on the maximum and minimum allowable current values, the current command is adjusted and sent to the driver. The driver executes the current command and simultaneously reads the actual current, actual motor torque estimate, or load displacement response during execution in real time, and calculates the execution error. If the actual effect is lower than the preset expectation, the current command is feedforward corrected, and the local gain is updated or the bias is corrected based on the observed long-term system deviation.

[0027] The monitoring and suppression module is used to analyze vibration signals during operation and compensate for them through motor output torque; the drive control module is used to receive control commands and control the cabinet to start, accelerate, maintain constant speed, decelerate and brake.

[0028] The regenerative braking module is used to dynamically allocate and recover electrical energy and compensate for peak power consumption during motor startup or system standby power supply; the management optimization module is used to monitor energy status in real time and optimize charging and discharging strategies; the log auditing module is used to record all operation events and synchronize them to the cloud auditing platform.

[0029] Specifically, all operational events generated by the device are uniformly mapped to a standardized set of transactions. Each transaction contains a unique transaction ID, a timestamp string, an event type code, an event payload, and optional local node metadata. The transaction payload is compacted, and a digest of each transaction is calculated. Multiple sets of transactions are placed into a package list in the order of collection, and a transaction digest list for the package set is generated. A corresponding binary tree is constructed based on the transaction digest list, where the bottom layer of the binary tree is the transaction digest. From bottom to top, the binary tree is concatenated using binary methods, and the hash value of the concatenation result is extracted to obtain the digest of the corresponding parent node. This process is recursively repeated until the root node is reached to generate the root digest of the block. The proof path corresponding to each transaction is recorded, and the root digest and block header metadata are merged to generate the corresponding block. The digest reference of the previous block header, the root digest of the current block, the current edge node identifier, the timestamp of this block, the sequence number of this block, and the status metadata are used as the current block header. The process involves arranging the block header fields in a fixed order, calculating the block header digest, and then concatenating the current block header digest with previous block header digests to form a hash chain reference. The current edge node's private key is then used to digitally sign the block header digest. The edge node periodically synchronizes the block digest and signature to the cloud or multiple trusted audit nodes. Simultaneously, when the edge node receives block headers from other nodes, it verifies the validity of the block header signature and its consistency with the previous block hash reference. When an auditing party or local query requests verification of whether a transaction is recorded on the chain, the edge node provides the coded content of the transaction, the transaction digest, and the corresponding proof path. The verifier recalculates the transaction digest and iteratively calculates the parent node hash value according to the proof path, ultimately obtaining the calculated root node. It also verifies whether the root node in the provided block header matches the calculated root node and then verifies the validity of the block header hash signature to obtain proof of the transaction's existence and integrity.

[0030] The monitoring and visualization module is used to display the cabinet's operating status, energy flow, obstacle distribution, and control logs in real time.

Claims

1. An intelligent control system for moving mobile shelving units, characterized in that, It includes a perception and detection module, a modal fusion module, a recognition and prediction module, a parameter correction module, a predictive control module, a speed adjustment module, a monitoring and suppression module, a drive control module, a braking recovery module, a management and optimization module, a log auditing module, and a monitoring and visualization module; The sensing and detection module is used to collect information on the operating environment and equipment status of the mobile shelving, and to monitor the cabinet position, load weight, friction changes and surrounding environmental conditions in real time. The modal fusion module constructs a real-time three-dimensional environment map by fusing millimeter-wave radar and 3D depth camera data; The identification and prediction module identifies obstacle types and analyzes the movement trends of dynamic targets based on a real-time 3D environment map. The parameter correction module calculates the current cabinet load mass through pressure sensor and motor current feedback, and corrects the motion parameters accordingly. The predictive control module predicts the stopping position of the cabinet in real time based on historical motion data and dynamically adjusts the brake trigger point and deceleration curve. The speed adjustment module adjusts the braking intensity and motor braking current in real time based on the predictive control results and the current load status. The monitoring and suppression module is used to analyze vibration signals during operation and compensate for them through motor output torque. The drive control module is used to receive control commands and control the cabinet to start, accelerate, maintain constant speed, decelerate and brake. The regenerative braking module is used to dynamically allocate and recover electrical energy, and to compensate for peak power consumption during motor startup or system standby power supply. The management optimization module is used to monitor energy status in real time and optimize charging and discharging strategies. The log auditing module is used to record all operation events and synchronize them to the cloud auditing platform; The monitoring and visualization module is used to display the cabinet's operating status, energy flow, obstacle distribution, and control logs in real time.

2. The intelligent control system for moving mobile shelving units according to claim 1, characterized in that, The specific steps for the modal fusion module to construct a real-time 3D environment map are as follows: S1.1: Receive the raw radar points and camera point clouds collected in real time by millimeter-wave radar and 3D depth camera, remove noise from each radar point and camera point cloud by moving window filtering, remove outliers in each radar point and camera point cloud by box line method, and then resample each radar point and camera point cloud to the target synchronization time by time interpolation. S1.2: Use offline or online calibration to obtain the rotation matrix and translation vector from the camera to the aircraft, as well as the rotation and translation from the radar to the aircraft. Then, back-project the camera point cloud according to the camera intrinsic parameters to obtain 3D points in the camera coordinate system. Then, map it to the aircraft coordinate system through rigid body transformation. Then, directly use rigid body transformation to map the radar points to the aircraft coordinate system. After that, perform unit unification and scale correction on the transformed point cloud to generate radar points and camera points in the aircraft coordinate system. S1.3: Perform coarse registration between radar points and camera points and divide the space by voxels. Statistically count the observation sets of each voxel from various sensors. Then, perform weighted merging based on observation confidence and measurement variance to obtain the occupancy probability of each voxel. After that, perform Euclidean clustering on the merged voxels to divide the spatial points into candidate target clusters and calculate the centroid, velocity and shape characteristics of each cluster. S1.4: Calculate the Mahalanobis distance between each existing historical trajectory and each currently detected cluster. If there is a Mahalanobis distance lower than a preset threshold, the corresponding historical trajectory and cluster are used as candidate matches. If a cluster has multiple candidate matches, the redundant candidate matches are filtered out using the Hungarian algorithm. Then, the motion state of each matched target is updated in real time using an acceleration model to generate a dynamically updated real-time 3D environment map.

3. The intelligent control system for moving mobile shelving units according to claim 2, characterized in that, The specific steps of the identification and prediction module in identifying obstacle types and analyzing the movement trends of dynamic targets are as follows: S2.1: Initialize the log odds value of each voxel in the 3D environment map, then update the occupied probability of the voxels related to the observation according to the observation model, and then convert the log odds value into occupied, idle or unknown labels, while retaining the class probability of each voxel. Iterate through all voxels marked as "occupied" in the latest 3D environment map, divide the occupied voxels into multiple connected components, and treat each connected component as a candidate instance. S2.2: Construct the corresponding minimum bounding box for each connected component, and extract the point cloud subset within the bounding box as the original dataset of the instance. Extract the geometric boundary and timestamp information of each candidate instance, store each set of extracted information into the local instance table according to the unique instance ID, and then calculate the basic geometric quantities of the original dataset of each candidate instance. S2.3: Extract the feature values ​​of each candidate instance and normalize them to generate corresponding shape descriptors. Calculate the mean and variance of the point reflection intensity and the occurrence frequency and position variance of each candidate instance within the preset time frame. Then, input all the data into the trained classifier model. The classifier model outputs the non-normalized score of each category based on the forward propagation algorithm, and then converts it into a category probability distribution through softmax. S2.4: Based on the output category probability distribution, construct the category transition matrix, and then use Bayes-Markov update to combine the current frame's instantaneous classification probability with the transition prior probability to generate the corresponding temporal consistency probability. Read the current tracked state vector and velocity uncertainty index from the local instance table. S2.5: Progress the state step by step according to the preset time step and record the predicted position at each step. At the same time, calculate the position uncertainty growth at each predicted time step, collect the position-velocity sequence of each candidate instance at multiple times, and input each position-velocity sequence into the trained sequence prediction network. At the same time, output the trajectory samples or parameterized distribution of multiple future times through the sequence prediction network. S2.6: Based on the time consistency probability and the multimodal position distribution of each future step, calculate the occupancy probability of each candidate instance in the influence area of ​​the device's motion trajectory. Then, perform weighted accumulation based on the corresponding category probability distribution to generate a corresponding risk score. If the risk score exceeds the preset alarm threshold, an alarm will be issued at the control terminal to remind the staff.

4. The intelligent control system for moving mobile shelving units according to claim 3, characterized in that, The specific steps of the predictive control module in predicting the cabinet's stopping position in real time and dynamically adjusting the brake trigger point and deceleration curve are as follows: S3.1: Collect data segments from multiple time steps in the real-time data stream as input sequences, and perform scale normalization and detrending processing on each input sequence. At the same time, establish the corresponding input matrix according to the collection time order of each input sequence. S3.2: The established input matrix is ​​transmitted to the LSTM network. The LSTM network captures the medium-to-long-term time dependencies of various types of data in the input matrix based on the gating mechanism and outputs the hidden state and candidate information at each time step. Based on the output of the LSTM sub-network of each modality data, a corresponding set of representation vectors is established. S3.3: Regress the set of representation vectors through a fully connected layer, and output the mean of the predicted stopping position and the prediction uncertainty. Based on the prediction uncertainty, the current encoder position and the current speed, calculate the remaining travel distance from the corresponding cabinet to the target stopping position. S3.4: Calculate the safe trigger distance based on the current speed, the maximum braking acceleration that the system can provide, and the system delay. If the remaining travel is less than or equal to the safe trigger distance and the system is not braking, then issue a braking command. If the system is already braking, then gradually adjust the target braking force according to the planned deceleration curve. S3.5: Based on the prediction uncertainty and the maximum braking acceleration, dynamically calculate the shape of the deceleration curve and generate the corresponding time series acceleration trajectory. Then, send the obtained deceleration curve to the driver and execute the corresponding deceleration strategy.

5. The intelligent control system for moving mobile shelving units according to claim 4, characterized in that, The specific steps by which the speed adjustment module adjusts the braking intensity and motor braking current in real time are as follows: S4.1: Collect the stop position prediction and its uncertainty output by the predictive control module, as well as the current instantaneous motion measurement, and read the load estimate. Then, map the stop position prediction to the motion axis and calculate the remaining displacement error. If the difference between the stop position prediction and the current motion state exceeds the set tolerance, it is marked as a high uncertainty situation, and the conversion factor of the next time step is adjusted. S4.2: Map the uncertainty corresponding to the stop position prediction to a conservative factor, and calculate the corresponding nominal average deceleration through the remaining displacement error and the current velocity. At the same time, amplify the result through the conservative factor and output the target average deceleration. Then, convert the target average deceleration into the braking force required for the drive shaft. S4.3: Map the braking force required by the drive shaft to the torque required by the motor, and calculate the target torque on the corresponding motor side or brake side based on the actual effective lever arm and transmission ratio, as well as the constraints of mechanical efficiency and friction loss of the braking device. S4.4: If braking requires energy recovery and the power supply can accept it, then all current is allowed to be in the regenerative direction; if energy recovery is saturated or voltage limitation exists, then the braking current is limited and switched to consumable braking or mechanical braking. Based on the judgment result, the target torque is converted into a current command at the motor end. S4.5: Based on the maximum and minimum allowable current values, adjust the current command and send the adjusted current command to the driver. The driver executes the current command and reads the actual current, actual motor torque estimate or load displacement response in real time during execution, and calculates the execution error. If the actual effect is lower than the preset expectation, the current command is feedforward corrected, and the local gain is updated or the bias is corrected based on the observed long-term system deviation.

6. The intelligent control system for moving mobile shelving units according to claim 1, characterized in that, The specific steps for the log auditing module to record all operation events are as follows: S5.1: Map all operation events generated by the device into a standardized set of transactions. Each transaction contains a unique transaction ID, a timestamp string, an event type code, an event payload, and optional local node metadata. Compact the transaction payload and calculate a summary of each transaction. S5.2: Place multiple sets of transactions into the list to be packaged in the order of collection, and generate a transaction summary list for the packaged set. Construct the corresponding binary tree based on the transaction summary list. The bottom layer of the binary tree is the transaction summary. From bottom to top, perform binary concatenation on the binary tree, extract the hash value of the concatenation result, obtain the summary of the corresponding parent node, and recursively go to the root node to generate the root summary of the block. S5.3: Record the proof path corresponding to each transaction, and merge the root digest and block header metadata to generate the corresponding block. Use the digest reference of the previous block header, the root digest of the current block, the current edge node identifier, the timestamp of this block, the sequence number of this block, and the status metadata as the current block header fields. S5.4: After arranging the block header fields in a fixed order, calculate the block header digest, then concatenate the current block header digest with the previous block header digest to form a hash chain reference, and then use the private key of the current edge node to digitally sign the block header digest; S5.5: The block header generated and signed on the chain by the edge node is periodically synchronized with the block digest and signature to the cloud or multiple trusted audit nodes. At the same time, the block header received by the edge node from other nodes is verified to verify the validity of the block header signature and its consistency with the hash reference of the previous block. S5.6: When the auditor or local query requests to verify whether a transaction is recorded on the chain, the edge node provides the coded content of the transaction, the transaction summary, and the corresponding proof path. The verifier recalculates the transaction summary and iteratively calculates the hash value of the parent node according to the proof path to finally obtain the calculated root node. At the same time, it verifies whether the root node in the provided block header is consistent with the calculated root, and then verifies whether the signature of the block header hash is valid to obtain the proof of the existence and integrity of the transaction.

Images