A method for overlapping control of stacker crane pick-and-place axis movements based on load detection

Abstract

This invention discloses a stacker crane's overlapping action control method for picking and placing goods based on load detection, belonging to the field of warehouse automation control technology. The method includes: 1. Establishing picking and placing task parameters and multi-axis initial state information; 2. Collecting high-frequency loads on the forks and predicting material release times; 3. Identifying material types and dynamically matching action control strategies. This invention can reduce waiting time in traditional serial actions, improve the cycle time efficiency of a single picking operation, and enable the stacker crane to automatically adjust motion parameters for different materials. This improves operational stability and control accuracy under different working conditions, increases the degree of multi-axis action overlap, reduces idle time, improves the overall throughput efficiency of the stacker crane, enhances action coordination and operational smoothness, and can gradually form an optimal action control strategy adapted to different storage locations and material characteristics. This improves long-term operating efficiency and adaptability, enhances the intelligence level of stacker crane picking and placing, and reduces reliance on manual parameter adjustment.

Description

Technical Field

[0001] This invention relates to the field of warehouse automation control technology, specifically to a stacker crane's overlapping control method for picking and placing goods based on load detection. Background Technology

[0002] With the rapid development of automated storage and retrieval systems (AS / RS), intelligent logistics systems, and flexible storage equipment, stacker cranes have been widely used in pallet storage, bin handling, and high-density warehousing operations. Existing stacker cranes typically include a Z-axis lifting mechanism, a Y-axis fork extension and retraction mechanism, and a traveling mechanism, completing material handling operations through the coordinated movement of multiple axes. The operating efficiency of a stacker crane largely depends on the efficiency of the motion connection between its axes and the degree of overlap in their movements. Most existing stacker crane control methods employ a fixed-sequence control strategy, meaning that the next axis only starts moving after the previous one has completely finished its movement. For example, during picking, it is usually necessary to wait for the forks to fully lift the material and the Z-axis to be fully stable before the Y-axis is allowed to retract; similarly, during unloading, it is necessary to wait for the Z-axis to completely stop before extending or retracting the forks. While this sequential action mode has simple control logic, it leads to long waiting times between multiple axes, limiting the overall picking and unloading cycle time and making it difficult to meet the efficiency requirements of high-throughput warehousing systems. Furthermore, existing control methods often lack real-time sensing capabilities of the actual load status. Due to differences in weight, center of gravity distribution, structural shape, and friction characteristics among various materials, stacker cranes exhibit different dynamic response characteristics during actual loading and unloading operations. However, traditional control systems mostly employ fixed acceleration / deceleration curves and fixed triggering times, failing to dynamically adjust motion parameters based on the actual material conditions. When faced with heavy loads, off-center loads, or irregularly shaped materials, problems such as cargo swaying, increased impact, increased positioning errors, and unstable cycle times easily arise. Furthermore, traditional stacker crane control systems typically lack self-learning capabilities based on historical operating data. Existing systems struggle to effectively utilize the vast amounts of historical motion data generated under different storage locations, materials, and operating conditions, resulting in control strategies remaining fixed and unable to continuously optimize motion overlap and parameter configurations with accumulated operational experience. Consequently, the system struggles to continuously improve loading and unloading efficiency and operational stability over long-term operation. Therefore, a load-detection-based method for controlling the overlap of loading and unloading axis actions in stacker cranes is urgently needed.

[0003] Existing stacker crane pick-and-place axis motion overlap control methods suffer from long waiting times in sequential actions, reducing the efficiency of single pick-and-place cycles. They also cannot automatically adjust motion parameters for different materials, have poor multi-axis motion overlap, reducing the overall throughput efficiency of the stacker crane, and cannot form an optimal motion control strategy that adapts to different storage locations and material characteristics, relying too much on manual parameter tuning. To address this, we propose a stacker crane pick-and-place axis motion overlap control method based on load detection. Summary of the Invention

[0004] The purpose of this invention is to overcome the deficiencies in the existing technology and provide a stacker crane's overlapping control method for picking and placing goods based on load detection.

[0005] This invention proposes a method for controlling the overlapping motion of the stacker crane's picking and placing shafts based on load detection. The technical solution adopted to solve the technical problem is as follows:

[0006] Ⅰ. Establish picking and placing task parameters and multi-axis initial status information: After receiving the picking or placing task issued by the upper-level scheduling system, the stacker crane reads the target storage location coordinate information and obtains the current stacker crane operating status parameters. At the same time, it establishes the picking and placing action path based on the target storage location and generates the corresponding basic action sequence.

[0007] II. Collect high-frequency loads from the forks and predict the material removal time: The stacker crane controls the movement of the forks along the Z and Y axes based on the basic motion sequence. It also collects load data in real time during the contact phase between the forks and the target storage location using load sensors at a preset sampling frequency. Subsequently, it preprocesses the continuously collected load signals and predicts the time when the material completely leaves the storage location, thereby controlling the Z-axis of the forks to rise.

[0008] Ⅲ. Identify material type and dynamically match motion control strategy: During the process of lifting the material, based on real-time collected load data, record the complete load change curve from the time the forks contact the material to the time the material is completely removed from the storage position, and establish a corresponding load spectrum to identify the current material type. Then, based on the identified material type, automatically match the corresponding shaft motion speed curve parameters.

[0009] IV. Collaborately plan the stacker crane fork movements and dynamically adjust its speed curve: When the forks enter the lifting stage, dynamically adjust the Y-axis start timing based on various data during the Z-axis and Y-axis control process, and monitor the remaining Z-axis stroke and real-time operating status in real time during the Y-axis retraction process, while dynamically adjusting the Y-axis speed parameters until the current picking action ends.

[0010] V. Record historical action data and perform self-learning optimization: After each picking and placing task is completed, the complete action process data of the forks is automatically recorded. Then, a historical action record table is established for the corresponding storage location. Based on the historical operation results, the action strategy is optimized and learned to form an action control strategy that adapts to different storage locations and different material characteristics.

[0011] As a further aspect of the present invention, the stacker crane operating status parameters in step I specifically include the current Z-axis height position of the forks, the current Y-axis extension and retraction position of the forks, the current operating speed of each axis, the maximum allowable acceleration and deceleration of each axis, the current position of the forks, the current load status, and the historical action records of the corresponding storage location.

[0012] As a further aspect of the present invention, the specific steps for preprocessing the continuously acquired load signals in step II are as follows:

[0013] S1.1: After installing load sensors at the fork root or loading platform, the data acquisition task is started when the fork is detected to enter the picking contact area. After starting, each load sensor acquires the stacker crane fork load at a sampling frequency of ≥1kHz, and records the current load data and timestamp at each sampling to form the original load data arranged in chronological order. Then, the load data within the preset sliding time window is filtered by median.

[0014] S1.2: After median filtering, calculate the local mean and local standard deviation of the load data in each sliding time window, and obtain the anomaly threshold of the corresponding sliding time window based on the preset anomaly discrimination coefficient and the local standard deviation of the sliding time window at different times. Then, take the absolute value of the difference between each load data and the corresponding local mean as its deviation degree, and compare it with the corresponding anomaly threshold.

[0015] S1.3: If the deviation is less than or equal to the abnormal threshold, the current load data is considered to be within the normal fluctuation range; otherwise, it is marked as a local outlier and replaced with the local mean of the corresponding sliding time window. After the outlier replacement is completed, the load data at each time point after correction in each sliding time window is filtered twice using exponential smoothing. The smoothed load data generated after the second filtering is then organized into a stable load sequence in chronological order. After that, the load data in the stable load sequence are unified to the same dimension through Max-Min normalization.

[0016] As a further aspect of the present invention, the specific steps for predicting the time when the material is completely removed from the storage location and controlling the fork lifting in step II are as follows:

[0017] S2.1: Count the number of load data and sampling interval in each group of stable load sequences, and obtain the load change rate at the corresponding time by comparing the load data at each time with the load data at the previous time. Then, perform local averaging on the load change rate of each stable load sequence to obtain the smooth load change rate at different time windows.

[0018] S2.2: Set the rate of change threshold. If the smooth load rate of change is greater than or equal to the rate of change threshold, the judgment flag of the current time window is set to 1, indicating that the current rate of change has reached the condition for entering the lifting stage; otherwise, it is set to 0, indicating that the condition has not yet been reached. At the same time, when the judgment flag is 1, the total number of times the judgment flag is 1 is counted backward. If the total number of times is equal to the preset continuous judgment window length, the material is judged to have entered the lifting stage.

[0019] S2.3: Once the lifting phase is determined, the local growth slope within the preset time window is obtained in real time. Then, the historical lifting model is called, and the remaining removal time of the material at the target location is predicted based on the current stable load value, stable load data, smooth load change rate, and local growth slope.

[0020] S2.4: Collect historical data on the transmission and execution preparation time of the stacker crane fork Z-axis control command, the time required for the motor to respond and start speeding up, and the compensation time required to reach the target high-speed lifting state. Sum the response data as the comprehensive advance of the fork Z-axis. Then calculate the difference between the predicted remaining disengagement time and the comprehensive advance to obtain the advance triggering time of the fork Z-axis acceleration command.

[0021] S2.5: If the system time corresponding to the current control cycle is greater than or equal to the advance trigger time, an acceleration command is immediately sent to the Z-axis of the fork. After the acceleration command is issued, the subsequent load change rate and the running status of the Z-axis of the fork are continuously monitored to confirm whether the acceleration has covered the predicted disengagement time. If the judgment flag is 1 and the system time is greater than or equal to the advance trigger time, it means that the advance acceleration trigger has been completed within the lifting phase.

[0022] As a further aspect of the present invention, the specific construction steps of the historical uplift model described in S2.3 are as follows:

[0023] P1.1: Filter out the complete historical lifting process from the existing picking records, that is, the entire action data from when the forks start to lift the material until the material is completely removed from the target position. Then, record each complete lifting process as a historical sample number, and save the load curve, equipment status and action end time of each historical sample at the same time. Then, calculate the difference between the start time of the current historical sample and the time when the material in the sample is completely removed from the storage position, and use it as the remaining removal time of each historical sample in the corresponding state to establish a complete historical lifting sample.

[0024] P1.2: Time alignment of each historical uplift sample is performed by resampling. Then, each historical uplift sample is synchronously truncated according to a time window of preset length. After truncating, the local average load value, load increase amplitude, fluctuation intensity and average change amplitude of each historical uplift sample within the time window are calculated. The four sets of data are then integrated into the load characteristics of the corresponding historical uplift sample.

[0025] P1.3: Integrate the initial position state quantity, running state quantity and structural working condition quantity corresponding to the lifting of the fork in each historical lifting sample into the equipment state vector. Then, concatenate the load characteristics with the state vector to establish the final input sample of each historical lifting sample, and normalize each input sample.

[0026] P1.4: A historical rise model is constructed using a weighted linear regression approach. Then, each input sample is sequentially input into the historical rise model. Based on the initial model weights and model bias terms, the remaining escape time for each input sample is predicted. Then, based on the predicted remaining escape time and the actual remaining escape time for each historical rise sample, the training loss value of the historical rise model in the current training period is obtained.

[0027] P1.5: Based on the training loss value, the gradient descent method is used to adjust the weights and bias terms of the historical rise model. The iteration is repeated until the training loss value converges to the preset threshold interval. Then, the input samples of the historical rise samples participating in the training are used as the validation set and input into the trained historical rise model. The validation error of the historical rise model on the validation set is obtained.

[0028] P1.6: If the verification error meets the preset requirements, the current corresponding weights and bias terms are solidified as the final parameters of the historical rise model; otherwise, the historical rise samples and the extracted input samples are reorganized for model training. Then, the mapping relationship between the remaining escape time and the current state is established through the trained historical rise model.

[0029] As a further aspect of the present invention, the specific steps for establishing the corresponding load map and identifying the current material type in step III are as follows:

[0030] S3.1: When the fork contacts the material at the target location in the historical sample, select the minimum sampling time where the load data at the previous moment is less than the preset contact judgment threshold and the load data at the current moment is greater than or equal to the preset contact judgment threshold. Use this as the starting sampling index of the picking contact segment. At the same time, retrieve its ending sampling index and retain all load data from the starting sampling index to the ending sampling index in chronological order to form a complete historical lifting curve.

[0031] S3.2: Resample each historical uplift curve and then normalize the amplitude to unify the load spectrum values ​​of each historical uplift curve into the same interval. Then, map the time coordinates of each point in each historical uplift curve to the interval [0, 1] to establish the corresponding load spectrum segment. Then, extract multiple sets of time domain feature parameters from each load spectrum segment and integrate them into the input sample.

[0032] S3.3: Associate each input sample with its corresponding material category label, count the number of input samples for each material category, then split the input samples into training set and validation set according to the preset category ratio, then integrate each input sample into a root node, and at the same time count the proportion of each material category in the root node to obtain the Gini impurity of the current root node, then select any candidate feature and candidate splitting threshold to perform trial splitting on the root node.

[0033] S3.4: After the trial split, the splitting benefit of this split is calculated. Then, the candidate feature with the highest splitting benefit and the candidate splitting threshold are selected as the best splitting combination. Based on the best splitting combination, the root node is recursively split into left and right child nodes, and the same splitting process is repeated for each child node until the node tree depth is greater than or equal to the preset maximum allowable depth, or the total number of samples in the node is less than or equal to the preset minimum number of samples required for the node to continue splitting, or the Gini impurity of the node is less than or equal to the preset purity stopping threshold.

[0034] S3.5: When a node stops splitting, the corresponding child node that finally stops splitting is set as a leaf node, and its category output is set to the category with the most samples in that leaf node to complete the construction of the decision tree model. The current decision tree model is then pruned to generate multiple candidate subtrees. After that, all leaf nodes in each candidate subtree are counted, and the training set is input into each candidate subtree. The training misclassification rate on the training set is counted. Based on the obtained training misclassification rate, the total cost of each candidate subtree is calculated. Then, the candidate subtree with the smallest total cost is selected as the optimal decision tree model.

[0035] S3.6: After pruning, input the validation set into the decision tree model, and obtain the validation accuracy of the decision tree model based on the predicted category and the true category of each sample in the validation set. If the validation accuracy does not meet the preset requirements, return to the splitting or pruning stage and readjust the threshold set, tree depth limit or sample splitting method. Otherwise, use it as the final lightweight decision tree model, and then repeat to generate multiple sets of lightweight decision tree models.

[0036] S3.7: In the new picking process, extract the time-domain feature parameters of each group of the current load map and input them into the pre-trained decision tree models. The decision tree models judge the splitting conditions of each node layer by layer until they reach any leaf node, and read the category probability distribution in the leaf node. At the same time, the category with the highest probability is output as the category label of the material. Then, the category labels output by each decision tree model are weighted and summed to obtain the corresponding voting score. The category label with the highest voting score is output as the final category label of the material, and the corresponding category probability is used as the recognition confidence of its category label.

[0037] As a further aspect of the present invention, the time-domain characteristic parameters mentioned in S3.2 specifically include: peak load, load rise slope, fluctuation frequency, stabilization time, load jitter amplitude, curve slope change characteristics, and local impact characteristics;

[0038] The material category labels mentioned in S3.3 include: light-duty, medium-duty, heavy-duty, and irregularly shaped materials;

[0039] The candidate features mentioned in S3.3 are the temporal feature parameters in each training sample; the candidate splitting threshold is specifically the median value of each temporal feature parameter at different adjacent sampling times, and the candidate splitting thresholds are arranged in ascending order of the median value.

[0040] As a further aspect of the present invention, the specific steps for automatically matching the corresponding shaft motion speed curve parameters according to the identified material type in step III are as follows:

[0041] P2.1: Based on the current material category label, select the pre-stored basic axis motion speed curve parameters of the material from the strategy library, then randomly select a conservative adjustment coefficient from the [0,1] interval, and obtain the corresponding parameter correction coefficient based on the identification confidence of the current material category label and the conservative adjustment coefficient.

[0042] P2.2: If the difference between the parameter correction coefficient and 1 is lower than the preset threshold, the motion speed curve parameters of the basic shaft are used directly; otherwise, the motion speed curve parameters of each basic shaft are corrected based on the parameter correction coefficient.

[0043] P2.3: After the correction is completed, the corrected axis motion speed curve parameters are repackaged into the final executable axis speed curve parameter group and written into the current task buffer of the stacker crane axis controller. When the axis controller executes the picking and placing task, it reads the axis speed curve parameter group in the current task buffer to generate the actual speed trajectory of the Z axis and Y axis.

[0044] As a further aspect of the present invention, the basic axis motion speed curve parameters mentioned in P2.1 include: Z-axis target acceleration, Z-axis maximum running speed, Z-axis deceleration trigger position, Y-axis start-up timing, Y-axis retraction acceleration, and multi-axis motion overlap ratio.

[0045] As a further aspect of the present invention, the specific steps in step IV of dynamically adjusting the Y-axis start-up timing based on various data of the Z-axis and Y-axis of the fork, and monitoring the remaining travel and real-time operating status of the Z-axis during the Y-axis retraction process, while dynamically adjusting the Y-axis speed parameters, are as follows:

[0046] S4.1: When the stacker crane fork Z-axis enters the high-speed lifting stage, read the current speed, allowable deceleration and current completed lifting position of the Z-axis in the high-speed section. Then, based on the current speed and allowable deceleration of the fork Z-axis, obtain the remaining deceleration time and remaining deceleration distance required for the Z-axis to decelerate from the current speed to zero.

[0047] S4.2: Collect the remaining stroke of the fork Y-axis to be retracted, the maximum allowable speed, acceleration, and control response delay, and use a symmetrical acceleration and deceleration curve to obtain the total execution time for the fork Y-axis to complete retraction. Then, based on the single-cycle update time of the stacker crane axis controller and the average network delay in the command transmission or execution link, obtain the start-up safety margin.

[0048] S4.3: The sum of the time when the Z-axis enters the high-speed lifting stage and the start safety margin is used as the lower limit of the Y-axis startable window. The difference between the sum of the remaining deceleration time of the Z-axis and the start safety margin and the total execution time of the Y-axis is used as the upper limit of the Y-axis startable window. A complete Y-axis startable window is established based on the obtained lower limit and upper limit.

[0049] S4.4: Establish a synchronization cost function based on the Y-axis startable window, and select the start time with the minimum synchronization cost as the optimal start time. Compare the current control time with the optimal start time. If the current control time is greater than or equal to the optimal start time, the axis controller immediately sends a recovery command to the Y-axis.

[0050] S4.5: After Y-axis recovery is started, continuously monitor the synchronization deviation between the remaining stop time of Z-axis and the remaining recovery time of Y-axis. If the synchronization deviation is higher than the preset upper limit threshold or lower than the preset lower limit threshold, establish the Y-axis speed correction coefficient based on the synchronization deviation and issue a command speed to the Y-axis for correction.

[0051] As a further aspect of the present invention, the specific form of the synchronization cost function described in S4.4 is as follows:

[0052]

[0053]

[0054] In the formula, Representative candidate launch time The corresponding synchronization cost; Represents the remaining deceleration time along the Z-axis; This represents the total execution time for the Y-axis to complete the recycling process; This indicates the moment when the Z-axis enters the high-speed lifting phase; Candidate values ​​representing the start time of the Y-axis; Represents the synchronization error weight; This represents the late-start penalty weight; This represents the optimal Y-axis start time; This indicates that the window can be launched along the Y-axis.

[0055] As a further aspect of the present invention, the specific steps of establishing a historical action record table for the corresponding storage location in step V, and optimizing the action strategy based on the historical operation results to form an action control strategy adapted to different storage locations and different material characteristics are as follows:

[0056] S5.1: After each pickup and delivery task is completed, the action process data generated during the entire task is organized into a single historical sample, and a corresponding task number and end mark are added to each historical sample to construct a complete historical record sample. The historical record samples are then categorized according to the storage location to establish a historical record set for each storage location.

[0057] S5.2: Encode the historical records in the historical records set of each storage location into a state vector containing storage location structure, material category, equipment operating health and residual error of the action. Then, standardize each state vector and obtain the corresponding comprehensive reward value based on the total execution time, total energy consumption, action impact and end swing of the task in each historical record sample.

[0058] S5.3: Count the number of historical records of any action combination used under each state vector, and record the comprehensive reward value of each historical record sample under different action combinations into the action value table, and adjust the action value of each action combination under each state through iterative updates.

[0059] S5.4: Select the action combination with the highest action value from the action value table as the optimal action combination in the current state, and bind it to the corresponding storage location to form a storage location strategy record. After each subsequent pick-up and drop-off task, update the storage location strategy record until the difference between the action values ​​of two adjacent storage location strategy records is lower than the preset convergence threshold. This indicates that the action combination of the corresponding storage location under the current material distribution has formed a stable optimal strategy.

[0060] As a further aspect of the present invention, the action process data in S5.1 specifically includes: the action sequence of each axis, speed change curve, start and stop time, sensor response delay, load change curve, cargo location information, material type, action completion time, and operation stability index.

[0061] The beneficial effects of this invention are:

[0062] This invention constructs a stable load sequence through median filtering, outlier replacement, and exponential smoothing. Then, based on the load change rate and continuity judgment mechanism, it identifies when materials enter the lifting phase and predicts the remaining material release time using a historical lifting model. Based on the Z-axis comprehensive lead time, it triggers the Z-axis acceleration command in advance, achieving high-speed lifting at the moment of material release. Subsequently, it resamples, normalizes, and extracts features from the historical lifting curves to construct a load spectrum segment. Multiple lightweight decision tree models are then constructed to output material type identification and identification confidence. After the Z-axis enters the high-speed lifting phase, a time-velocity coordination model between the Z and Y axes is established in real time to dynamically solve for the optimal start time on the Y-axis. A synchronization deviation correction mechanism is used to achieve synchronous overlapping control of the Y-axis recovery process and the Z-axis deceleration process. After the task is completed, the entire process is analyzed. Action data is archived historically to build a historical data set for each storage location. Based on a comprehensive reward value and action value iterative update mechanism, the system performs self-learning optimization on action combinations under different storage locations and material conditions. This gradually forms a stable and optimal storage location action control strategy, which can reduce waiting time in traditional serial actions, improve the cycle time efficiency of single picking, and enable the stacker crane to automatically adjust motion parameters for different materials. This improves the operational stability and control accuracy under different working conditions, increases the degree of overlap of multi-axis actions, reduces idle time, improves the overall throughput efficiency of the stacker crane, and enhances action coordination and operational stability. It can gradually form an optimal action control strategy that adapts to different storage locations and material characteristics, improves long-term operating efficiency and adaptability, enhances the intelligence level of stacker crane picking and placing, and reduces reliance on manual parameter adjustment. Attached Figure Description

[0063] The present invention will now be further described with reference to the accompanying drawings.

[0064] Figure 1 This is a framework diagram of a stacker crane's overlapping control method for picking and placing goods based on load detection. Detailed Implementation

[0065] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0066] Example 1: This embodiment of the invention provides a method for controlling the overlapping motion of the stacker crane's picking and placing axes based on load detection. See also... Figure 1 , Figure 1 This is a framework diagram of a stacker crane's overlapping movement control method based on load detection, provided as an embodiment of the present invention. The method includes the following steps:

[0067] Establish pick-up and drop-off task parameters and multi-axis initial status information: After receiving the pick-up or drop-off task issued by the upper-level scheduling system, the stacker crane reads the target storage location coordinate information and obtains the current stacker crane operating status parameters. At the same time, it establishes the pick-up and drop-off action path based on the target storage location and generates the corresponding basic action sequence.

[0068] The high-frequency load of the forks is collected, and the material detachment time is predicted. The stacker crane controls the movement of the forks along the Z and Y axes based on the basic action sequence. The load sensor collects the load data in real time during the contact stage between the forks and the target storage location at a preset sampling frequency. Then, the continuously collected load signals are preprocessed, and the time when the material completely detaches from the storage location is predicted to control the Z-axis of the forks to rise.

[0069] Specifically, after installing load sensors at the fork root or loading platform, a data acquisition task is initiated when the forks are detected entering the picking contact zone. After initiation, each load sensor acquires the stacker crane fork load at a sampling frequency of ≥1kHz, recording the current load data and timestamp at each sampling, forming raw load data arranged in chronological order. Then, the load data within a preset sliding time window undergoes median filtering. After median filtering, the local mean and local standard deviation of the load data in each sliding time window are calculated. Based on a preset anomaly detection coefficient and the local standard deviation of the sliding time window at different times, the anomaly threshold for the corresponding sliding time window is obtained. The absolute value of the difference between each load data point and the corresponding local mean is used as its deviation degree. This value is then compared with the corresponding anomaly threshold. If the deviation degree is less than or equal to the anomaly threshold, the current load data is considered to be within the normal fluctuation range. Otherwise, it is marked as a local anomaly and replaced with the local mean of the corresponding sliding time window. After the anomaly replacement is completed, the load data at each time point after correction in each sliding time window is filtered twice using exponential smoothing. The smoothed load data generated after the second filtering is then organized into a stable load sequence in chronological order. Finally, Max-Min normalization is used to unify the load data in the stable load sequence to the same dimension.

[0070] Specifically, the number of load data points and sampling intervals in each stable load sequence are counted. The load change rate at each time point is obtained by comparing the load data at each time point with the load data at the previous time point. Then, the load change rate of each stable load sequence is locally averaged to obtain the smoothed load change rate for different time windows. A change rate threshold is set. If the smoothed load change rate is greater than or equal to the change rate threshold, the judgment flag for the current time window is set to 1, indicating that the current change rate has met the conditions for entering the lifting stage; otherwise, it is set to 0, indicating that the condition has not yet been met. Simultaneously, when the judgment flag is 1, the total number of times the judgment flag is 1 is counted backwards. If the total number of times equals the preset continuous judgment window length, the material is determined to have entered the lifting stage. Once the lifting stage is determined, the local growth slope within the preset time window is obtained in real time. Then, the historical lifting model is called, and the load is stabilized based on the current stable load value. Data, smoothed load change rate, and local growth slope are used to predict the remaining detachment time of materials at the target location. Historical stacker crane fork Z-axis control command transmission and execution preparation time, motor response and start-up time, and compensation time required to reach the target high-speed lifting state are collected. The sum of all response data is used as the comprehensive advance of the fork Z-axis. Then, the difference between the predicted remaining detachment time and the comprehensive advance is calculated to obtain the advance trigger time of the fork Z-axis acceleration command. If the system time corresponding to the current control cycle is ≥ the advance trigger time, an acceleration command is immediately sent to the fork Z-axis. After the acceleration command is issued, the subsequent load change rate and fork Z-axis operating status are continuously monitored to confirm whether the acceleration has covered the predicted detachment time. If the judgment flag is 1 and the system time is ≥ the advance trigger time, it means that the advance acceleration trigger has been completed within the lifting stage.

[0071] Identify material type and dynamically match motion control strategy: During the process of lifting the material, based on real-time collected load data, record the complete load change curve from the time the forks contact the material to the time the material is completely removed from the storage position, establish the corresponding load spectrum, identify the current material type, and then automatically match the corresponding shaft motion speed curve parameters according to the identified material type.

[0072] Specifically, when the forks contact the material at the target location in the historical samples, the smallest sampling time where the previous load data is less than the preset contact threshold and the current load data is greater than or equal to the preset contact threshold is selected and used as the starting sampling index of the picking contact segment. Simultaneously, the ending sampling index is retrieved, and all load data from the starting to the ending sampling index are retained in chronological order to form a complete historical lifting curve. Each historical lifting curve is resampled and then normalized to unify the load spectrum values ​​within the same interval. Then, the time coordinates of each point in each historical lifting curve are uniformly mapped to the [0, 1] interval to establish corresponding load spectrum segments. Finally, the load spectrum values ​​are analyzed from each load spectrum segment. Multiple sets of temporal feature parameters are extracted from the spectral band and integrated into input samples. Each input sample is associated with its corresponding material category label, and the number of input samples for each material category is counted. Then, the input samples are split into training and validation sets according to a preset category ratio. All input samples are then integrated into a root node, and the proportion of each material category in the root node is calculated to obtain the Gini impurity of the current root node. Next, any candidate feature and candidate splitting threshold are selected to perform a trial split on the root node. After the trial split, the splitting gain is calculated, and the candidate feature and candidate splitting threshold with the highest splitting gain are selected as the optimal splitting combination. Based on the optimal splitting combination, the root node is recursively split into left and right child nodes, and each child node is re-evaluated. The same partitioning process is repeated until the following conditions are met: the node tree depth is greater than or equal to the preset maximum allowed depth; the total number of samples in the node is less than or equal to the preset minimum number of samples required for the node to continue splitting; or the Gini impurity of the node is less than or equal to the preset purity stopping threshold. When a node stops splitting, the corresponding child node that finally stops splitting is set as a leaf node, and its category output is set to the category with the most samples in that leaf node, thus completing the construction of the decision tree model. The current decision tree model is then pruned to generate multiple sets of candidate subtrees. After that, all leaf nodes in each candidate subtree are counted, and the training set is input into each candidate subtree to calculate its training misclassification rate on the training set. Based on the obtained training misclassification rate, the total cost of each candidate subtree is calculated, and then the candidate subtree with the minimum total cost is selected as the optimal decision tree. After pruning, the validation set is input into the decision tree model. Based on this data, the model calculates the predicted and true classes of each sample in the validation set to obtain the validation accuracy. If the validation accuracy does not meet the preset requirements, the process returns to the splitting or pruning stage, readjusting the threshold set, tree depth limit, or sample partitioning method. Otherwise, it is used as the final lightweight decision tree model. Multiple lightweight decision tree models are then repeatedly generated. During a new loading process, the temporal feature parameters of each set of the current load map are extracted and input into the pre-trained decision tree models. The decision tree model proceeds layer by layer along the splitting conditions of each node until it reaches any leaf node, at which point the class probability distribution is read.Simultaneously, the category with the highest probability is output as the category label for the material. Then, the category labels output by each decision tree model are weighted and summed to obtain the corresponding voting score. The category label with the highest voting score is output as the final category label for the material, and the corresponding category probability is used as the recognition confidence level for its category label.

[0073] Example 2: This embodiment of the invention provides a method for controlling the overlapping movements of the stacker crane's picking and placing axes based on load detection. See also... Figure 1 , Figure 1 This is a framework diagram of a stacker crane's overlapping movement control method based on load detection, provided as an embodiment of the present invention. The method includes the following steps:

[0074] Collaboratively plan the stacker crane fork movements and dynamically adjust their speed curves: When the forks enter the lifting phase, dynamically adjust the Y-axis start timing based on various data during the Z-axis and Y-axis control process, and monitor the remaining Z-axis stroke and real-time operating status in real time during the Y-axis retraction process, while dynamically adjusting the Y-axis speed parameters until the current picking action ends.

[0075] Specifically, when the stacker crane fork Z-axis enters the high-speed lifting phase, the real-time motion status of the current speed, allowable deceleration, and current completed lifting position of the Z-axis in the high-speed segment is read. Subsequently, based on the current speed and allowable deceleration of the fork Z-axis, the remaining deceleration time and remaining deceleration distance required for the Z-axis to decrease from the current speed to zero are obtained. The remaining stroke, maximum allowable speed, acceleration, and control response delay of the fork Y-axis to be retracted are collected, and a symmetrical acceleration and deceleration curve is used to obtain the total execution time for the fork Y-axis to complete retraction. Then, based on the single-cycle update time of the stacker crane axis controller and the average network delay in the command transmission or execution link, the starting safety margin is obtained. The sum of the time when the Z-axis enters the high-speed lifting phase and the starting safety margin is used as the starting time of the Y-axis. The lower limit of the window is used, and the difference between the sum of the remaining deceleration time of the Z-axis and the starting safety margin and the total execution time of the Y-axis is used as the upper limit of the Y-axis startable window. A complete Y-axis startable window is established based on the obtained lower and upper limits. A synchronization cost function is established based on the Y-axis startable window, and the starting time with the minimum synchronization cost is selected as the optimal starting time. The current control time is compared with the optimal starting time. If the current control time is greater than or equal to the optimal starting time, the axis controller immediately sends a recovery command to the Y-axis. After the Y-axis recovery starts, the synchronization deviation between the remaining stop time of the Z-axis and the remaining recovery time of the Y-axis is continuously monitored. If the synchronization deviation is higher than the preset upper limit threshold or lower than the preset lower limit threshold, a Y-axis speed correction coefficient is established based on the synchronization deviation, and the command speed sent to the Y-axis is corrected.

[0076] Record historical action data and perform self-learning optimization: After each picking and placing task is completed, the complete action process data of the forks is automatically recorded. Then, a historical action record table is established for the corresponding storage location. Based on the historical operation results, the action strategy is optimized and learned to form an action control strategy that adapts to different storage locations and different material characteristics.

[0077] Specifically, after each pickup and delivery task is completed, the motion process data generated throughout the entire task is compiled into a single historical sample. Each historical sample is then assigned a corresponding task number and an end marker to construct a complete historical record sample. These historical record samples are categorized according to storage location, creating a historical record set for each location. The historical record samples in each location's historical record set are encoded as state vectors containing the storage location structure, material category, equipment operational health, and residual motion errors. These state vectors are then standardized. Finally, based on the total execution time, total energy consumption, motion impact, and end effector oscillation of each historical record sample, a corresponding comprehensive reward value is obtained, and the state vectors are statistically analyzed. The system records the number of historical samples of any action combination used, and records the comprehensive reward value of each historical sample under different action combinations into the action value table. The action value of each action combination under each state is adjusted through iterative updates. The action combination with the highest action value is selected from the action value table as the optimal action combination for the current state, and it is bound to the corresponding storage location to form a storage location strategy record. After each subsequent pick-up and drop-off task, the storage location strategy record is updated until the difference between the action values ​​of two adjacent storage location strategy records is lower than the preset convergence threshold. This indicates that the action combination of the corresponding storage location under the current material distribution has formed a stable optimal strategy.

[0078] The foregoing has provided a detailed description of one embodiment of the present invention, but this description is merely a preferred embodiment and should not be construed as limiting the scope of the invention. All equivalent variations and modifications made within the scope of the claims of this invention should still fall within the patent coverage of this invention.

Claims

1. A stacker crane's overlapping control method for picking and placing goods based on load detection, characterized in that, Includes the following steps: Ⅰ. Establish picking and placing task parameters and multi-axis initial status information: After receiving the picking or placing task issued by the upper-level scheduling system, the stacker crane reads the target storage location coordinate information and obtains the current stacker crane operating status parameters. At the same time, it establishes the picking and placing action path based on the target storage location and generates the corresponding basic action sequence. II. Collect high-frequency loads from the forks and predict the material removal time: The stacker crane controls the movement of the forks based on the basic action sequence, and collects load data in real time during the contact stage between the forks and the material at the target location through load sensors at a preset sampling frequency. Then, the continuously collected load signals are preprocessed and the time when the material completely leaves the location is predicted, and the Z-axis of the forks is controlled to rise. Ⅲ. Identify material type and dynamically match motion control strategy: During the process of lifting the material, based on real-time collected load data, record the complete load change curve from the time the forks contact the material to the time the material is completely removed from the storage position, and establish a corresponding load spectrum to identify the current material type. Then, based on the identified material type, automatically match the corresponding shaft motion speed curve parameters. IV. Collaborately plan the stacker crane fork movements and dynamically adjust their speed curves: When the forks enter the lifting phase, dynamically adjust the Y-axis start timing based on the data of the Z-axis and Y-axis of the forks, and monitor the remaining travel and real-time running status of the Z-axis in real time during the Y-axis retraction process, while dynamically adjusting the Y-axis speed parameters until the current picking action ends. V. Record historical action data and perform self-learning optimization: After each picking and placing task is completed, the complete action process data of the forks is automatically recorded. Then, a historical action record table is established for the corresponding storage location, and the action strategy is optimized and learned based on the historical operation results to form an action control strategy that adapts to different storage locations and different material characteristics.

2. The stacker crane's overlapping action control method for picking and placing goods based on load detection according to claim 1, characterized in that, The specific steps for preprocessing the continuously acquired load signals in step II are as follows: S1.1: After installing load sensors at the fork root or loading platform, the data acquisition task is started when the fork is detected to enter the picking contact area. After starting, each load sensor acquires the stacker crane fork load at a sampling frequency of ≥1kHz, and records the current load data and timestamp at each sampling to form the original load data arranged in chronological order. Then, the load data within the preset sliding time window is filtered by median. S1.2: After median filtering, calculate the local mean and local standard deviation of the load data in each sliding time window, and obtain the anomaly threshold of the corresponding sliding time window based on the preset anomaly discrimination coefficient and the local standard deviation of the sliding time window at different times. Then, take the absolute value of the difference between each load data and the corresponding local mean as its deviation degree, and compare it with the corresponding anomaly threshold. S1.3: If the deviation is less than or equal to the abnormal threshold, the current load data is considered to be within the normal fluctuation range; otherwise, it is marked as a local outlier and replaced with the local mean of the corresponding sliding time window. After the outlier replacement is completed, the load data at each time point after correction in each sliding time window is filtered twice using exponential smoothing. The smoothed load data generated after the second filtering is then organized into a stable load sequence in chronological order. After that, the load data in the stable load sequence are unified to the same dimension through Max-Min normalization.

3. The stacker crane's overlapping action control method for picking and placing goods based on load detection according to claim 2, characterized in that, The specific steps for predicting the time it takes for the material to completely leave the storage location and controlling the fork lifting, as described in Step II, are as follows: S2.1: Count the number of load data and sampling interval in each group of stable load sequences, and obtain the load change rate at the corresponding time by comparing the load data at each time with the load data at the previous time. Then, perform local averaging on the load change rate of each stable load sequence to obtain the smooth load change rate at different time windows. S2.2: Set the rate of change threshold. If the rate of change of the smooth load is greater than or equal to the rate of change threshold, then set the judgment flag of the current time window to 1, indicating that the current rate of change has reached the condition for entering the lifting stage. Conversely, it is set to 0, indicating that the condition has not yet been met; at the same time, when the judgment flag is 1, the total number of times the judgment flag is 1 is counted backward. If the total number of times equals the preset continuous judgment window length, the material is judged to have entered the lifting stage. S2.3: Once the lifting phase is determined, the local growth slope within the preset time window is obtained in real time. Then, the historical lifting model is called, and the remaining removal time of the material at the target location is predicted based on the current stable load value, stable load data, smooth load change rate, and local growth slope. S2.4: Collect historical data on the transmission and execution preparation time of the stacker crane fork Z-axis control command, the time required for the motor to respond and start speeding up, and the compensation time required to reach the target high-speed lifting state. Sum the response data as the comprehensive advance of the fork Z-axis. Then calculate the difference between the predicted remaining disengagement time and the comprehensive advance to obtain the advance triggering time of the fork Z-axis acceleration command. S2.5: If the system time corresponding to the current control cycle is greater than or equal to the advance trigger time, an acceleration command is immediately sent to the Z-axis of the fork. After the acceleration command is issued, the subsequent load change rate and the running status of the Z-axis of the fork are continuously monitored to confirm whether the acceleration has covered the predicted disengagement time. If the judgment flag is 1 and the system time is greater than or equal to the advance trigger time, it means that the advance acceleration trigger has been completed within the lifting phase.

4. The stacker crane's overlapping action control method for picking and placing goods based on load detection according to claim 1, characterized in that, The specific steps for establishing the corresponding load map and identifying the current material type in step III are as follows: S3.1: When the fork contacts the material at the target location in the historical sample, select the minimum sampling time where the load data at the previous moment is less than the preset contact judgment threshold and the load data at the current moment is greater than or equal to the preset contact judgment threshold. Use this as the starting sampling index of the picking contact segment. At the same time, retrieve its ending sampling index and retain all load data from the starting sampling index to the ending sampling index in chronological order to form a complete historical lifting curve. S3.2: Resample each historical uplift curve and then normalize the amplitude to unify the load spectrum values ​​of each historical uplift curve into the same interval. Then, map the time coordinates of each point in each historical uplift curve to the interval [0, 1] to establish the corresponding load spectrum segment. Then, extract multiple sets of time domain feature parameters from each load spectrum segment and integrate them into the input sample. S3.3: Associate each input sample with its corresponding material category label, count the number of input samples for each material category, then split the input samples into training set and validation set according to the preset category ratio, then integrate each input sample into a root node, and at the same time count the proportion of each material category in the root node to obtain the Gini impurity of the current root node, then select any candidate feature and candidate splitting threshold to perform trial splitting on the root node. S3.4: After the trial split, the splitting benefit of this split is calculated. Then, the candidate feature with the highest splitting benefit and the candidate splitting threshold are selected as the best splitting combination. Based on the best splitting combination, the root node is recursively split into left and right child nodes, and the same splitting process is repeated for each child node until the node tree depth is greater than or equal to the preset maximum allowable depth, or the total number of samples in the node is less than or equal to the preset minimum number of samples required for the node to continue splitting, or the Gini impurity of the node is less than or equal to the preset purity stopping threshold. S3.5: When a node stops splitting, the corresponding child node that finally stops splitting is set as a leaf node, and its category output is set to the category with the most samples in that leaf node to complete the construction of the decision tree model. The current decision tree model is then pruned to generate multiple candidate subtrees. After that, all leaf nodes in each candidate subtree are counted, and the training set is input into each candidate subtree. The training misclassification rate on the training set is counted. Based on the obtained training misclassification rate, the total cost of each candidate subtree is calculated. Then, the candidate subtree with the smallest total cost is selected as the optimal decision tree model. S3.6: After pruning, input the validation set into the decision tree model, and obtain the validation accuracy of the decision tree model based on the predicted category and the true category of each sample in the validation set. If the validation accuracy does not meet the preset requirements, return to the splitting or pruning stage and readjust the threshold set, tree depth limit or sample splitting method. Otherwise, use it as the final lightweight decision tree model, and then repeat to generate multiple sets of lightweight decision tree models. S3.7: In the new picking process, extract the time-domain feature parameters of each group of the current load map and input them into the pre-trained decision tree models. The decision tree models judge the splitting conditions of each node layer by layer until they reach any leaf node, and read the category probability distribution in the leaf node. At the same time, the category with the highest probability is output as the category label of the material. Then, the category labels output by each decision tree model are weighted and summed to obtain the corresponding voting score, and the category label with the highest voting score is output as the final category label of the material.

5. The stacker crane's overlapping action control method for picking and placing goods based on load detection according to claim 1, characterized in that, The specific steps in step IV, which involve dynamically adjusting the Y-axis start-up timing based on various data from the Z-axis and Y-axis of the fork, and monitoring the remaining travel and real-time operating status of the Z-axis during the Y-axis retraction process while dynamically adjusting the Y-axis speed parameters, are as follows: S4.1: When the stacker crane fork Z-axis enters the high-speed lifting stage, read the current speed, allowable deceleration and current completed lifting position of the Z-axis in the high-speed section. Then, based on the current speed and allowable deceleration of the fork Z-axis, obtain the remaining deceleration time and remaining deceleration distance required for the Z-axis to decelerate from the current speed to zero. S4.2: Collect the remaining stroke of the fork Y-axis to be retracted, the maximum allowable speed, acceleration, and control response delay, and use a symmetrical acceleration and deceleration curve to obtain the total execution time for the fork Y-axis to complete retraction. Then, based on the single-cycle update time of the stacker crane axis controller and the average network delay in the command transmission or execution link, obtain the start-up safety margin. S4.3: The sum of the time when the Z-axis enters the high-speed lifting stage and the start safety margin is used as the lower limit of the Y-axis startable window. The difference between the sum of the remaining deceleration time of the Z-axis and the start safety margin and the total execution time of the Y-axis is used as the upper limit of the Y-axis startable window. A complete Y-axis startable window is established based on the obtained lower limit and upper limit. S4.4: Establish a synchronization cost function based on the Y-axis startable window, and select the start time with the minimum synchronization cost as the optimal start time. Compare the current control time with the optimal start time. If the current control time is greater than or equal to the optimal start time, the axis controller immediately sends a recovery command to the Y-axis. S4.5: After Y-axis recovery is started, continuously monitor the synchronization deviation between the remaining stop time of Z-axis and the remaining recovery time of Y-axis. If the synchronization deviation is higher than the preset upper limit threshold or lower than the preset lower limit threshold, establish the Y-axis speed correction coefficient based on the synchronization deviation and issue a command speed to the Y-axis for correction.

6. The stacker crane's overlapping action control method for picking and placing goods based on load detection according to claim 1, characterized in that, The specific steps in step V, which involve establishing a historical action record table for the corresponding storage location and optimizing the action strategy based on historical results to form an action control strategy that adapts to different storage locations and material characteristics, are as follows: S5.1: After each pickup and delivery task is completed, the action process data generated during the entire task is organized into a single historical sample, and a corresponding task number and end mark are added to each historical sample to construct a complete historical record sample. The historical record samples are then categorized according to the storage location to establish a historical record set for each storage location. S5.2: Encode the historical records in the historical records set of each storage location into a state vector containing storage location structure, material category, equipment operating health and residual error of the action. Then, standardize each state vector and obtain the corresponding comprehensive reward value based on the total execution time, total energy consumption, action impact and end swing of the task in each historical record sample. S5.3: Count the number of historical records of any action combination used under each state vector, and record the comprehensive reward value of each historical record sample under different action combinations into the action value table, and adjust the action value of each action combination under each state through iterative updates. S5.4: Select the action combination with the highest action value from the action value table as the optimal action combination in the current state, and bind it to the corresponding storage location to form a storage location strategy record. After each subsequent pick-up and drop-off task, update the storage location strategy record until the difference between the action values ​​of two adjacent storage location strategy records is lower than the preset convergence threshold. This indicates that the action combination of the corresponding storage location under the current material distribution has formed a stable optimal strategy.

Images