Intelligent speed regulation method and system for conveyor belt-like brain robot and storage medium

Abstract

The application discloses a kind of intelligent speed regulation method, system and storage medium of conveying belt brain-like robot, intelligent speed regulation method, system and storage medium of conveying belt brain-like robot, comprising obtaining initial perception instruction output by heterogeneous brain-like unit, extract the decision cycle of initial perception instruction, obtain the period characteristic sequence containing timestamp information, judge whether period characteristic sequence exists time sequence conflict, if there is time sequence conflict, then initial perception instruction is timestamp alignment processing, obtain the alignment instruction set of time dimension unification;The intelligent speed regulation method, system and storage medium of the conveying belt brain-like robot, can reduce the interference of overall adjustment to the speed regulation relationship of other stations, reduce speed fluctuation and response delay, improve the speed stability, response accuracy and collaborative control ability of conveying belt brain-like robot under multi-station, dynamic load working condition.

Description

Technical Field

[0001] This invention relates to the field of intelligent manufacturing and intelligent control technology for conveying equipment, specifically to an intelligent speed regulation method, system, and storage medium for a brain-like robot for conveyor belts. Background Technology

[0002] In intelligent manufacturing systems, conveyor belts are used to connect multiple production stations and complete the continuous transport of materials. Their operating speed directly affects the material flow cycle, station collaboration efficiency, and overall stability of the production line. With the application of neuromorphic robot technology in intelligent manufacturing scenarios, different stations can be configured with heterogeneous neuromorphic units with different perception methods and decision-making capabilities. Various neuromorphic units can output corresponding speed adjustment commands according to local working conditions, thereby improving the autonomous perception and adjustment capabilities of the conveying system.

[0003] However, the perception methods, processing procedures, and decision-making cycles of different heterogeneous brain-like units differ, which may lead to inconsistencies in the generation time, arrival time, and update frequency of multiple speed control commands. When multiple commands participate in conveyor belt speed control simultaneously, timing conflicts are likely to occur. If the commands are still processed according to their arrival order or simply superimposed, some control commands may not match the current working conditions, resulting in speed fluctuations or response delays. At the same time, there is a continuous material flow relationship between multiple stations on the conveyor belt. Adjusting the speed of one station may further change the material arrival rhythm, load status, and operating requirements of subsequent stations. Therefore, the speed control commands output by different stations not only have a temporal sequence but may also form cross-station correlations and continuous influences. Existing speed control methods typically focus on comparing or integrating multiple commands at a single moment, making it difficult to effectively identify the spatial relationships, propagation relationships, and sequential dependencies between commands from different workstations. Furthermore, when the speed control demands of multiple workstations influence each other, the degree of impact of different speed control commands on the overall operating state of the conveyor belt may vary. Using fixed priorities or fixed weights for command integration makes it difficult to dynamically distinguish the importance of each command based on actual working conditions, which can easily lead to the weakening of critical speed control demands or the overemphasis of local speed control demands. Moreover, after a unified speed control command is formed, the actual operating speed of the conveyor belt may still deviate from the expected operating speed due to load changes and multi-workstation linkages, and existing overall correction methods are unable to effectively adjust the main commands that cause speed deviations. Summary of the Invention

[0004] The purpose of this invention is to provide an intelligent speed control method, system, and storage medium for a conveyor belt-type brain-like robot, thereby solving the problems existing in the prior art.

[0005] To achieve the above objectives, the present invention provides the following technical solution: an intelligent speed regulation method for a conveyor belt-type neuromorphic robot, comprising: S1. Obtain the initial perception command output by the heterogeneous brain-like unit, extract the decision cycle of the initial perception command, and obtain the cycle feature sequence. S2. If there is a time sequence conflict in the periodic feature sequence, the initial sensing command is timestamped to obtain an alignment command set. S3. Extract the spatial distribution features of the alignment instruction set when crossing workstations to obtain a spatial feature vector representing the topological relationship of the workstations. S4. Construct a spatiotemporal dimension matrix by using spatial feature vectors and periodic feature sequences, track the propagation path of alignment instructions across workstations, and extract the causal order in the propagation path to obtain spatiotemporal correlation features. S5. Based on the spatiotemporal correlation characteristics, if the alignment instruction set that crosses workstations produces a chain reaction, then calculate the instruction priority that crosses workstations and determine the target priority sequence. S6. The alignment instruction set is numerically weighted according to the target priority sequence to obtain a weighted instruction matrix that eliminates control errors. S7. Calculate the expected running speed of the conveyor belt based on the weighted instruction matrix to obtain the initial belt speed adjustment instruction for the conveyor belt brain-like robot. S8. If the actual running speed corresponding to the initial belt speed adjustment command is greater than the expected running speed, then the weighted command corresponding to the target priority sequence is located to determine the final belt speed adjustment command of the conveyor belt brain-like robot.

[0006] Preferably, step S1 includes acquiring the initial perception instruction output by the heterogeneous brain-like unit, encoding the initial perception instruction to obtain an instruction parsing matrix; extracting the temporal dependency of the instruction parsing matrix through a long short-term memory network to generate a hidden layer state vector; determining the decision cycle based on the hidden layer state vector, extracting the timestamp information of the decision cycle to generate a timestamp aligned sequence; and fusing the hidden layer state vector with the timestamp aligned sequence to obtain a periodic feature sequence containing timestamp information.

[0007] Preferably, step S2 includes acquiring a periodic feature sequence and determining whether there is a timing conflict in the periodic feature sequence; if there is a timing conflict, extracting the clock drift feature of the periodic feature sequence; constructing a synchronization reference sequence based on the clock drift feature; using the synchronization reference sequence to perform timestamp alignment processing on the initial sensing instruction to obtain a first sensing instruction; filtering the first sensing instruction to obtain a second sensing instruction; and performing aggregation processing according to the execution cycle of the second sensing instruction to obtain a time-uniform alignment instruction set.

[0008] Preferably, step S3 includes obtaining an alignment instruction set; constructing a workstation node matrix based on the instruction flow path of the alignment instruction set; calculating adjacency weights based on the workstation node matrix and updating the workstation node matrix to obtain a topology connection graph; processing the topology connection graph using a graph convolutional network to extract the spatial distribution features of the topology connection graph; performing dimensionality reduction processing on the spatial distribution features to obtain a feature aggregation vector; constructing a spatial mapping matrix through the feature aggregation vector and performing normalization processing to obtain a spatial feature vector representing the topology relationship of the workstations.

[0009] Preferably, step S4 includes acquiring spatial feature vectors and periodic feature sequences to construct a spatiotemporal dimension matrix; extracting unidirectional workstation topology relationships based on the spatiotemporal dimension matrix; determining instruction propagation paths according to unidirectional workstation topology relationships and timestamp order; determining whether there is a closed loop in the initial causal order of the aligned instruction propagation paths; if there is a closed loop, disconnecting the connection edges to obtain an updated causal order; and obtaining spatiotemporal correlation features with sequential dependencies through the updated causal order.

[0010] Preferably, step S5 includes extracting conflict nodes and delayed transmission paths based on spatiotemporal correlation features; if the conflict nodes and delayed transmission paths form a closed loop, it is determined that the alignment instruction set generates a chain reaction, resulting in a subset of associated instructions; calculating instruction priority values ​​based on the subset of associated instructions; and sorting the instruction priority values ​​to determine the target priority sequence.

[0011] Preferably, step S6 includes constructing a conflict detection matrix based on the execution cycle and resource status of the alignment instruction set; mapping the target priority sequence through the conflict detection matrix to obtain a priority mapping table; assigning weights to the alignment instruction set according to the priority mapping table to obtain initial weighted instructions; if the initial weighted instructions trigger a deviation threshold, generating an error compensation vector for the initial weighted instructions; and correcting the initial weighted instructions through the error compensation vector to obtain a weighted instruction matrix that eliminates control errors.

[0012] Preferably, step S7 includes extracting load tension distribution features based on the weighted instruction matrix and target priority sequence, calculating the slip compensation value corresponding to the load tension distribution features in combination with the friction resistance matrix; determining the speed adjustment amount based on the slip compensation value, and sequentially fusing the speed adjustment amounts corresponding to each workstation according to the target priority sequence to obtain the global dynamic traction force; if the global dynamic traction force exceeds the traction force threshold, then the global dynamic traction force is limited to obtain the target dynamic traction force; the expected operating speed is calculated through the target dynamic traction force, and the initial belt speed adjustment command for the global collaborative conveyor belt neuromorphic robot is obtained based on the expected operating speed.

[0013] Preferably, step S8 includes obtaining the actual running speed and the expected running speed corresponding to the initial belt speed adjustment command; if the actual running speed is greater than the expected running speed, calculating the speed deviation value; matching the target priority sequence and locating the weighted command based on the speed deviation value; using the weighted command to perform directional numerical decrement processing on the initial belt speed adjustment command to obtain the speed adjustment execution sequence; and determining the final belt speed adjustment command of the conveyor belt brain-like robot based on the speed adjustment execution sequence.

[0014] An intelligent speed control system for a conveyor belt-type neuromorphic robot, comprising the steps of the intelligent speed control method for the conveyor belt-type neuromorphic robot, the system including: The periodic module acquires the initial perception instructions output by the heterogeneous brain-like unit, extracts the decision cycle of the initial perception instructions through the long short-term memory network, and obtains a periodic feature sequence containing timestamp information. The alignment module determines whether there is a temporal conflict in the periodic feature sequence. If there is a temporal conflict, the initial sensing instructions are timestamped to obtain a unified set of alignment instructions in the time dimension. The spatial module uses a graph convolutional network to process the alignment instruction set, extracts the spatial distribution features of the alignment instruction set when crossing workstations, and obtains a spatial feature vector representing the topological relationship of the workstations. The association module constructs a spatiotemporal dimension matrix through spatial feature vectors and periodic feature sequences. It tracks the propagation path of alignment instructions across workstations according to the topological relationship of workstations and the order of timestamps, and extracts the causal order in the propagation path to obtain spatiotemporal association features with sequential dependencies. The sorting module determines whether the alignment instruction set that crosses workstations will produce a chain reaction based on the spatiotemporal correlation characteristics. If a chain reaction occurs, the priority of the instructions that cross workstations is calculated to determine the target priority sequence. The weighting module assigns numerical weights to the alignment instruction set through the target priority sequence to obtain a weighted instruction matrix that eliminates control errors. The speed control module, based on the weighted instruction matrix, sequentially integrates the speed adjustment amounts corresponding to each workstation according to the target priority sequence, calculates the expected running speed of the conveyor belt, and obtains the initial belt speed control instruction for the global collaborative conveyor belt neuromorphic robot. The correction module obtains the actual running speed corresponding to the initial belt speed adjustment command. If the actual running speed is greater than the expected running speed, it locates the corresponding weighted command according to the target priority sequence and performs directional numerical decrement processing on the initial belt speed adjustment command to determine the final belt speed adjustment command of the conveyor belt brain-like robot.

[0015] A computer-readable storage medium storing computer instructions thereon, which, when executed, perform the intelligent speed control method of the conveyor belt brain-like robot.

[0016] As can be seen from the above technical solution, the present invention has the following beneficial effects: This intelligent speed regulation method, system, and storage medium for a conveyor belt neuromorphic robot acquires initial perception commands from multiple heterogeneous neuromorphic units. It analyzes and unifies the decision-making cycle and timestamp information of different commands, reducing timing conflicts caused by inconsistent decision-making cycles and command arrival times among different neuromorphic units. Furthermore, by combining the topological relationships between workstations and the propagation paths of commands across workstations, it identifies the sequential dependencies and continuous effects between different speed regulation commands, enabling the conveyor belt speed regulation process to consider the local states of each workstation and their correlation with subsequent workstations. Based on this, it determines the optimal command based on the influence relationships of cross-workstation commands. Prioritizing and weighting the commands helps avoid the problem of weakening critical speed control needs or over-amplifying local speed control needs caused by fixed priorities or simple superposition, thereby improving the rationality of multi-source speed control command fusion and global coordination. At the same time, the expected operating speed of the conveyor belt is determined based on the weighted command results, and when the actual operating speed exceeds the expected operating speed, the corresponding weighted commands are directionally corrected in combination with the target priority sequence. This can reduce the interference of overall adjustment on the speed control relationship of other workstations, reduce speed fluctuations and response lag, and improve the speed control stability, response accuracy and collaborative control capability of the conveyor belt neuromorphic robot under multi-workstation and dynamic load conditions. Attached Figure Description

[0017] Figure 1 This is a schematic diagram of the overall structure of the present invention; Figure 2 This is a bar chart comparing the steady-state speed regulation accuracy of three methods under constant load in Embodiment 2 of the present invention; Figure 3 This is a comparison curve of the speed response of the three methods after a sudden load change in Embodiment 2 of the present invention; Figure 4 This is a bar chart comparing the weight allocation of instructions at each workstation under the load change condition in Embodiment 2 of the present invention. Figure 5 This is a line graph showing the instruction fusion deviation of the three methods when the number of instruction sources increases in Embodiment 2 of the present invention. Detailed Implementation

[0018] 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.

[0019] Example 1:

[0020] like Figure 1 As shown, the present invention provides a technical solution: an intelligent speed regulation method for a conveyor belt-type neuromorphic robot, comprising: S1. Obtain the initial perception command output by the heterogeneous brain-like unit, extract the decision cycle of the initial perception command through the long short-term memory network, and obtain the periodic feature sequence containing timestamp information. S2. Determine whether there is a time sequence conflict in the periodic feature sequence. If there is a time sequence conflict, perform timestamp alignment processing on the initial sensing instructions to obtain a unified set of alignment instructions in the time dimension. S3. Use graph convolutional networks to process the alignment instruction set, extract the spatial distribution features of the alignment instruction set when crossing workstations, and obtain spatial feature vectors that characterize the topological relationship of workstations. S4. Construct a spatiotemporal dimension matrix by using spatial feature vectors and periodic feature sequences. Track the propagation path of alignment instructions across workstations according to the topological relationship of workstations and the order of timestamps. Extract the causal order in the propagation path to obtain spatiotemporal correlation features with sequential dependencies. S5. Based on the spatiotemporal correlation characteristics, determine whether the alignment instruction set that crosses workstations produces a chain reaction. If a chain reaction occurs, calculate the instruction priority that crosses workstations and determine the target priority sequence. S6. The alignment instruction set is numerically weighted according to the target priority sequence to obtain a weighted instruction matrix that eliminates control errors. S7. Based on the weighted instruction matrix, the speed adjustment amount corresponding to each workstation is fused sequentially according to the target priority sequence to calculate the expected running speed of the conveyor belt and obtain the global collaborative initial belt speed adjustment instruction of the conveyor belt brain-like robot. S8. Obtain the actual running speed corresponding to the initial belt speed adjustment command. If the actual running speed is greater than the expected running speed, locate the corresponding weighted command according to the target priority sequence, and perform directional numerical decrement processing on the initial belt speed adjustment command to determine the final belt speed adjustment command of the conveyor belt brain-like robot.

[0021] In this embodiment, several workstations are set along the conveyor belt, and each workstation is equipped with a heterogeneous brain-like unit.

[0022] The heterogeneous neuromorphic unit is used to collect the material status, equipment operation status, load status and workstation processing status of the corresponding workstation, and outputs initial sensing instructions based on the collection results; each initial sensing instruction includes at least a workstation identifier, instruction timestamp, speed adjustment amount and instruction status identifier; the control system receives the initial sensing instructions output by each heterogeneous neuromorphic unit in the order of the timestamp from front to back, and inputs the continuously acquired initial sensing instructions into the long short-term memory network.

[0023] Furthermore, the Long Short-Term Memory (LSTM) network processes consecutive initial sensing instructions in ascending order of timestamps, and extracts the time interval between adjacent initial sensing instructions, the duration of consecutive instructions at the same workstation, and the state change relationship between consecutive instructions.

[0024] By continuously memorizing and updating the aforementioned time information and state change information, the decision cycle corresponding to each heterogeneous brain-like unit is determined. For each workstation, the control system establishes a correspondence between the decision cycle corresponding to the workstation and the timestamp of the output instruction of the workstation, and then arranges the periodic features corresponding to all workstations in ascending order of timestamp, thereby obtaining a periodic feature sequence containing timestamp information.

[0025] In addition, the control system determines whether there is a timing conflict based on the periodic characteristic sequence.

[0026] Specifically, for two workstations with a direct topological connection, the timestamps of the initial sensing instructions corresponding to the two workstations are read, and the time interval between the two timestamps is calculated. When the time interval exceeds a preset time tolerance threshold, it is determined that the corresponding instruction has a timing conflict. Alternatively, when the timestamp of the instruction corresponding to the upstream workstation is later than the timestamp of the instruction corresponding to the downstream workstation, and the material conveying direction between the two workstations is from the upstream workstation to the downstream workstation, it is also determined that the corresponding instruction has a timing conflict. When neither of the above two situations occurs, it is determined that the corresponding instruction does not have a timing conflict.

[0027] Furthermore, when a timing conflict is detected, the control system establishes a unified time axis, which consists of continuously set control moments, with a preset control cycle between two adjacent control moments.

[0028] For each initial sensing command, calculate the absolute value of the time difference between the command's timestamp and each control moment in the unified time axis, and map the initial sensing command to the control moment with the smallest absolute value of the time difference. If the absolute value of the time difference between an initial sensing command and two control moments before and after it is the same, then map the command to the control moment with the smaller timestamp value. After completing the mapping of all initial sensing commands, a set of aligned commands with a unified time dimension is obtained.

[0029] The control system treats each station on the conveyor belt as a node in the graph structure and establishes directed connections between nodes according to the material conveying direction.

[0030] Specifically, when materials can be directly transported from the first workstation to the second workstation, a directed connection is established from the first workstation to the second workstation; when materials are diverted from one workstation to two target workstations, directed connections are established from the diversion starting workstation to the two target workstations respectively; when two transport branches converge into the same workstation, directed connections are established from the workstations corresponding to the two branches to the converging workstation respectively; a workstation topology is formed through all workstation nodes and the directed connection relationships between nodes.

[0031] Furthermore, the control system will input the instruction parameters corresponding to each station in the alignment instruction set into the graph convolutional network.

[0032] For any target workstation, the graph convolutional network reads the instruction parameters of the target workstation itself and the instruction parameters of the workstations that have a directed connection with the target workstation, and performs feature aggregation according to the pre-established workstation topology. After feature aggregation is completed, the spatial features corresponding to the target workstation are formed. After performing the same processing on all workstations in sequence, all spatial features are arranged in order of workstation number to obtain the spatial feature vector representing the workstation topology.

[0033] Spatial feature vectors are used to record the instruction status of each workstation and the instruction status of other workstations that have a topological connection with that workstation.

[0034] When the speed adjustment command of a workstation changes, the change information is transmitted through a graph convolutional network to the spatial features corresponding to the workstations that have a topological connection with it. Therefore, the spatial feature vector can reflect the spatial distribution state when the command propagates along the workstation connection relationship.

[0035] The control system combines spatial feature vectors and periodic feature sequences according to a unified timestamp and constructs a spatiotemporal dimension matrix.

[0036] The spatiotemporal dimension matrix is ​​arranged according to two dimensions: workstation and control time. One dimension is used to represent different workstations, and the other dimension is used to represent continuous control times in a unified time axis. Each matrix position records the spatial characteristics, periodic characteristics, and alignment instruction parameters of the corresponding workstation at the corresponding control time, thereby simultaneously representing the spatial relationship between different workstations and the temporal change relationship of the same workstation at continuous control times.

[0037] The control system uses a spatiotemporal dimension matrix to track the propagation path of alignment instructions between different workstations.

[0038] First, select the source workstation where the instruction changes and record the timestamp of the corresponding instruction. Then, search for the next workstation according to the directed connection direction in the workstation topology diagram, and filter the instructions corresponding to the next workstation whose timestamps are later than the timestamps of the source workstation instructions. Calculate the time interval between the two, and if the time interval does not exceed the preset propagation time threshold, add the next workstation to the current propagation path. Then, continue to search for subsequent workstations along the workstation topology direction in the same way until there are no subsequent workstations that simultaneously meet the topology connection condition, timestamp increment condition, and propagation time threshold condition, thus obtaining the corresponding propagation path.

[0039] In this embodiment, the causal order in the propagation path is determined by both the workstation topology direction and the timestamp order.

[0040] For two adjacent workstations in the propagation path, if the preceding workstation points to the following workstation through a directed connection, and the timestamp of the instruction corresponding to the preceding workstation is less than the timestamp of the instruction corresponding to the following workstation, then the instruction corresponding to the preceding workstation is determined to be the preceding instruction and the instruction corresponding to the following workstation is determined to be the following instruction. According to the relationship between the preceding and following instructions among all adjacent workstations, a spatiotemporal correlation feature with sequential dependency is formed.

[0041] The control system determines whether the alignment instruction set that crosses workstations will produce a chain reaction based on the spatiotemporal correlation characteristics.

[0042] When a propagation path contains at least two workstations, and any two adjacent workstations in the propagation path satisfy the following conditions: the preceding workstation points to the following workstation, the timestamp of the instruction corresponding to the preceding workstation is less than the timestamp of the instruction corresponding to the following workstation, and the time interval between the two instructions does not exceed a preset propagation time threshold, it is determined that the alignment instruction set corresponding to the propagation path generates a chain reaction; if any adjacent workstation does not meet the above conditions, it is determined that the corresponding propagation path does not generate a chain reaction.

[0043] Furthermore, once a chain reaction is determined, the control system prioritizes all instructions in the propagation path.

[0044] First, the instructions are arranged in ascending order of their timestamps, with the instruction with the smallest timestamp value placed first. If two instructions have the same timestamp, the number of workstations covered by the propagation path of each instruction is counted, and the instruction covering more workstations is placed first. If two instructions have the same timestamp and the same number of workstations covered, the workstation order is determined according to the material conveying direction, and the instruction corresponding to the workstation at the front of the conveying direction is placed first. After sorting all the instructions according to the above rules, a target priority sequence is formed.

[0045] The control system assigns numerical weights to each alignment instruction according to the target priority sequence.

[0046] In this process, the instruction that appears earlier in the sequence has a weight that is no less than the instruction that appears later in the sequence. After all instructions have been assigned weights, each weight is normalized so that the sum of all weights equals a complete weight unit. Then, the speed adjustment amount corresponding to each instruction is multiplied by the weight corresponding to that instruction to obtain the corresponding weighted speed adjustment amount. All weighted speed adjustment amounts are arranged according to the positional relationship of each workstation and each control moment to obtain the weighted instruction matrix.

[0047] The control system calculates the expected operating speed of the conveyor belt based on the weighted instruction matrix.

[0048] First, read the current running speed of the conveyor belt. Then, read the weighted speed adjustment amount corresponding to each station in order from front to back according to the target priority sequence. For the weighted speed adjustment amount that represents acceleration, add the corresponding value to the current running speed. For the weighted speed adjustment amount that represents deceleration, reduce the corresponding value to the current running speed. After superimposing all the weighted speed adjustment amounts according to the target priority sequence, the calculated running speed is obtained.

[0049] Furthermore, the calculated running speed is compared with the preset upper limit and lower limit of the allowed running speed.

[0050] Specifically, when the calculation speed is greater than the preset allowed upper limit, the preset allowed upper limit is determined as the expected speed; when the calculation speed is less than the preset allowed lower limit, the preset allowed lower limit is determined as the expected speed; when the calculation speed is between the preset allowed lower limit and the preset allowed upper limit, the calculation speed is directly determined as the expected speed.

[0051] The control system generates an initial belt speed adjustment command for the global collaborative operation of the conveyor belt neuromorphic robot based on the expected operating speed, and sends the initial belt speed adjustment command to the speed execution unit of the conveyor belt.

[0052] Specifically, the speed execution unit adjusts the output of the drive motor according to the initial belt speed adjustment command, so that the actual running speed of the conveyor belt changes to the expected running speed. Since the initial belt speed adjustment command is formed by all valid weighted commands in the target priority sequence, the speed adjustment command also includes the time sequence of each workstation, the topological relationship of the workstations, and the cross-workstation propagation order.

[0053] Furthermore, after the initial belt speed adjustment command is executed, the control system acquires the actual operating speed of the conveyor belt according to the preset sampling period and compares the actual operating speed with the expected operating speed.

[0054] When the actual operating speed is less than or equal to the expected operating speed, the directional numerical reduction process is not initiated; when the actual operating speed is greater than the expected operating speed, the difference between the actual operating speed and the expected operating speed is taken as the speed deviation to be eliminated and the directional numerical reduction process is initiated.

[0055] In addition, when performing directional numerical reduction processing, the control system checks the speed adjustment direction corresponding to each weighted command sequentially, starting from the last position in the target priority sequence.

[0056] When the current weighted instruction corresponds to an acceleration direction, the weighted instruction is determined as an instruction to be decremented; when the current weighted instruction corresponds to a deceleration direction or the speed adjustment amount is zero, the instruction is not decremented, and the weighted instruction corresponding to the previous sequence position is checked.

[0057] For a identified instruction to be reduced, the current speed deviation that has not yet been eliminated is compared with the weighted speed adjustment amount corresponding to the instruction to be reduced.

[0058] Specifically, when the weighted speed adjustment is greater than or equal to the current unresolved speed deviation, the current unresolved speed deviation is directly deducted from the weighted speed adjustment, and the current reduction process ends; when the weighted speed adjustment is less than the current unresolved speed deviation, the weighted speed adjustment is reduced to zero, and the current unresolved speed deviation is subtracted from the reduced speed adjustment to obtain the new unresolved speed deviation, and the previous acceleration direction weighted instruction in the target priority sequence is processed.

[0059] When the remaining speed deviation is reduced to zero, the subsequent reduction operation is stopped immediately; when all acceleration direction weighted commands in the target priority sequence have been processed, the current directional value reduction process is also ended; the control system re-executes the speed adjustment fusion based on all weighted commands after the reduction process, and generates the final belt speed control command for the conveyor belt neuromorphic robot based on the re-obtained speed control result.

[0060] In this embodiment, the control system receives initial sensing commands and conveyor belt running status data output by the heterogeneous neuromorphic unit, and performs time feature extraction, timestamp alignment, spatial feature extraction, spatiotemporal correlation analysis, command priority calculation, weight allocation, speed fusion, and feedback correction processing. The control system includes a processor, a memory, a data input interface, and a data output interface. The processor calls the computer program stored in the memory to process the received data and sends the final belt speed adjustment command obtained from the processing to the conveyor belt drive control unit through the data output interface. The drive control unit adjusts the output of the drive motor according to the final belt speed adjustment command so that the conveyor belt runs at a determined target speed. The control system can be implemented using an industrial control computer, a programmable controller, an edge computing controller, or an embedded controller.

[0061] Example 2:

[0062] like Figure 2 and Figure 5 As shown, experiments were designed to verify the actual effect of this application in a multi-station collaborative speed regulation scenario.

[0063] The experiment was conducted on a self-built four-station conveyor belt speed control test bench, with the following main configuration: Conveyor belt line: 4 drive stations are arranged in series, each station is driven by a 0.75kW variable frequency motor, and the rated running speed of the belt is 1.5m / s; Brain-like unit: Each workstation is equipped with one heterogeneous brain-like computing unit (each using a different coding system such as spiking neural network and rate coding network), which outputs the perception speed adjustment command of its workstation. The decision cycle is different between 20-50ms, which is used to simulate the clock difference of real heterogeneous brain-like units. Loading device: Station 2 is equipped with a weight-type loading tray, which can instantly add load during operation to simulate sudden load changes. Measurement system: Each workstation is equipped with a rotary encoder for speed measurement (sampling frequency 100Hz), and the host computer records the speed curves and command logs of each workstation.

[0064] To highlight the role of each step in this method, two common engineering comparison methods are set up: Comparison Method 1 (Fixed Priority): Instructions from each workstation are sorted and merged according to a pre-set fixed priority, without timestamp alignment or consideration of cross-workstation propagation relationships; Comparison Method 2 (Simple Weighted Summation): Instructions from each workstation are summed using fixed weighting coefficients without dynamic priority adjustment or targeted correction. This method, following the process described in this application, sequentially completes timestamp alignment, graph convolutional spatial feature extraction, spatiotemporal causal sequence tracking, chain reaction determination and priority ranking, weight allocation, velocity fusion, and overspeed directional reduction correction.

[0065] Experiment 1: Comparison of Steady-State Speed ​​Regulation Accuracy under Constant Load Experimental procedure: The belt was started under no-load, with the target speed set at 1.5 m / s. After the system entered steady state (starting 60 seconds after startup), the actual running speed of the belt was continuously recorded for 30 seconds. The steady-state speed deviation (the absolute value of the difference between the 30-second average speed and the target speed) and the speed fluctuation rate (the percentage of the difference between the maximum and minimum speeds during the recording period relative to the average speed) were calculated. The three methods were repeated three times each, and the average was taken.

[0066] Experimental results: Table 1 Comparison of Steady-State Speed ​​Regulation Accuracy under Constant Load

[0067] From Table 1 and Figure 2 As can be seen, the steady-state velocity deviation of this method is 0.033 m / s, which is only 27.5% of that of the fixed-priority method and 37.5% of that of the simple weighted superposition method; the velocity fluctuation rate is 2.74%, which is 68.2% lower than that of the fixed-priority method. The reason is that the fixed-priority method does not handle the differences in decision-making cycles of heterogeneous brain-like units, resulting in inconsistent instruction arrival times and residual temporal conflicts in the fusion results, which manifest as continuous velocity deviation and fluctuations; although the simple weighted superposition method alleviates the problem of single instruction dominance, it still does not eliminate inconsistencies in the temporal dimension. This method extracts decision-making cycle features through long short-term memory networks and performs timestamp alignment and filtering aggregation on instructions, so that the instructions entering the fusion stage are unified in the temporal dimension, thereby significantly improving the steady-state accuracy.

[0068] Experiment 2: Comparison of dynamic responses under sudden load changes at a single workstation Experimental procedure: The belt was running steadily at a target speed of 1.5 m / s. At 30 seconds, 50% of the rated load was instantaneously added at station 2 to simulate a sudden load change. The belt speed curve was recorded within 20 seconds after the sudden change. Three indicators were statistically analyzed: maximum drop (the percentage drop of the lowest speed point after the sudden change relative to the target speed), maximum overshoot (the percentage exceedance of the highest speed point relative to the target speed during the recovery process), and settling time (the time required for the speed to enter and remain within the target speed ±2% bandwidth).

[0069] Experimental results: Table 2 Comparison of dynamic response under load change conditions

[0070] like Figure 3As shown in Table 2, after a load mutation, the maximum drop in this method was 3.33%, which is 38.4% for the fixed priority method and 55.5% for the simple weighted summation method; the maximum overshoot was 4.48%, and the settling time was 1.18s, which is 74.9% shorter than that of the fixed priority method. Figure 3 The fixed priority method exhibits significant deep drops and high overshoot oscillations after abrupt changes, followed by the simple weighted summation method. In contrast, the velocity curve of this method shows shallow drops, fast recovery, and small oscillations.

[0071] like Figure 4 As shown, the weight allocation of instructions at each workstation under abrupt changes is presented: the weight of the fixed priority method is unrelated to the actual impact on the workstation (workstation 1 has the highest weight, 0.35); this method assigns the highest weight of 0.46 to workstation 2, which is directly affected by the load, and gives downstream workstations 3 and 4 compensation weights of 0.20 and 0.18 respectively according to the causal order of the propagation path; this shows that this method identifies the propagation path of workstation 2 being affected by the load and downstream workstations being affected through spatiotemporal correlation features, and dynamically sorts priorities and allocates weights accordingly, so that the critical speed adjustment requirements are not weakened; at the same time, when the actual speed is detected to exceed the expected speed during the recovery phase, only the weighted instructions of the corresponding priority are reduced in a targeted manner to avoid overall downsizing that would interfere with other workstations, thus resulting in small overshoot and short adjustment time.

[0072] Experiment 3: Comparison of instruction fusion deviation as the number of brain-like units (instruction sources) increases Experimental procedure: 2, 3, 4, and 5 heterogeneous neuromorphic units were activated sequentially (one spare neuromorphic unit was added to each of stations 1 and 3 when 5 units were activated) and speed adjustment commands were output simultaneously. The remaining operating conditions maintained a constant load and a target speed of 1.5 m / s. A 60-second command fusion log was recorded for each configuration. The speed adjustment amount to be executed was calculated based on the actual speed of the encoder at each station. The relative deviation between the fused command and the benchmark was calculated, i.e., the command fusion deviation. The average value of 3 repetitions was taken.

[0073] Experimental results: Table 3 Comparison of instruction fusion deviations under different numbers of instruction sources (%)

[0074] like Figure 5As shown in Table 3, as the number of instruction sources increases from 2 to 5, the fusion bias of the fixed priority method rapidly increases from 8.6% to 25.3%, and the simple weighted superposition method increases from 7.9% to 18.6%, both showing an almost linear deterioration trend. In contrast, the fusion bias of our method only increases slowly from 2.6% to 4.7%, and with 5 instruction sources, it is only 18.6% and 25.3% of the two comparative methods, respectively. The reason is that the more instruction sources there are, the more complex the differences in decision-making cycles and cross-workstation propagation relationships become. Fixed priority and equal weighted superposition cannot distinguish the sequential dependencies between instructions, and conflicting instructions cancel each other out or amplify each other. Our method uses a spatiotemporal dimension matrix to track the propagation path, determine chain reactions, and dynamically allocate weights, so that newly added instruction sources are incorporated into the fusion process in an orderly manner, and the growth of fusion bias is gradual.

[0075] Example 3:

[0076] S1 includes acquiring the initial perception instructions output by the heterogeneous brain-like unit, encoding the initial perception instructions to obtain the instruction parsing matrix; extracting the temporal dependencies of the instruction parsing matrix through a long short-term memory network to generate a hidden layer state vector; determining the decision cycle based on the hidden layer state vector, extracting the timestamp information of the decision cycle, and generating a timestamp-aligned sequence; and fusing the hidden layer state vector with the timestamp-aligned sequence to obtain a periodic feature sequence containing timestamp information.

[0077] In this embodiment, the initial perception instructions output by each heterogeneous brain-like unit are first obtained, and the initial perception instructions are encoded.

[0078] Specifically, the initial sensing instruction includes at least the workstation number, instruction generation time, speed adjustment value, and instruction status. During encoding, corresponding data is extracted according to a fixed field order of workstation number, instruction generation time, speed adjustment value, and instruction status. The workstation number is converted into a consecutive number according to the conveying direction; the instruction generation time is uniformly converted into a time value in the same time unit; the speed adjustment value retains its numerical value and positive / negative direction; and the instruction status is converted into a corresponding status code according to a pre-defined status category. Initial sensing instructions generated consecutively from the same workstation are arranged in ascending order of timestamp, and the encoding result of each instruction is treated as a row of data, with the encoded values ​​of each field forming different columns in that row, thus creating an instruction parsing matrix. This processing ensures that instructions generated by different heterogeneous neuromorphic units have a consistent data dimension and arrangement order, allowing for direct input into the Long Short-Term Memory (LSTM) network.

[0079] Furthermore, the instruction parsing matrix is ​​input into the Long Short-Term Memory (LSTM) network in chronological order. When processing the instruction encoding result at the current moment, the LSM network simultaneously calls the state information stored at the previous moment and updates its internal state according to the change relationship between the current instruction and the instruction at the previous moment.

[0080] Specifically, the workstation number, speed adjustment value, and instruction status are compared for two consecutive time points. When the workstation number is the same and the instruction status has not changed, the status information corresponding to the previous time point is retained, and the speed adjustment change at the current time point is written into the current status. When the instruction status changes, the changed instruction status and its corresponding time position are written into the current status. After processing all rows of data in the instruction parsing matrix in the same way, a hidden layer status is output at each processing time point. The hidden layer statuses are arranged in the order of the original timestamps to form a hidden layer status vector. This hidden layer status vector records the continuous instruction changes at the same workstation and the time dependency between adjacent instructions.

[0081] In addition, when determining the decision cycle based on the hidden layer state vector, the change results of adjacent hidden layer states at the same workstation are compared sequentially, and the instruction change nodes that recur in the hidden layer states are located.

[0082] Here, the instruction change node refers to the location where the speed adjustment direction or instruction state changes. For two adjacent instruction change nodes of the same type, their corresponding timestamps are read. The timestamp of the previous change node is subtracted from the timestamp of the subsequent change node to obtain one cycle candidate value. The same calculation is then performed on subsequent instruction change nodes of the same type in chronological order to obtain the set of cycle candidate values ​​corresponding to that workstation. The values ​​in the cycle candidate value set are compared one by one, and the cycle candidate value that appears most frequently is determined as the decision cycle of this heterogeneous brain-like unit. When two cycle candidate values ​​appear the same number of times, the cycle candidate value that appears earlier in the chronological order is determined as the decision cycle. Through this process, the decision cycle is determined by the actual instruction change interval, rather than directly using the sampling period.

[0083] Furthermore, after determining the decision-making cycle, the timestamp information that makes up each decision-making cycle is extracted.

[0084] Specifically, the timestamp corresponding to the first instruction change node within a decision cycle is used as the cycle start timestamp, and the timestamp corresponding to the last instruction change node within the same decision cycle is used as the cycle end timestamp. The original timestamps corresponding to all initial perception instructions within the cycle are retained. Then, the cycle start timestamp with the smallest value among all heterogeneous brain-like units is used as the alignment starting point, and continuous control moments are generated sequentially according to the preset control cycle in Example 1. For each original timestamp, the absolute value of the time difference between the original timestamp and its adjacent control moments is calculated, and the original timestamp is mapped to the control moment with the smallest absolute time difference. When the absolute value of the time difference between the original timestamp and two adjacent control moments is the same, the original timestamp is mapped to the control moment with the smaller time value. All mapping results are arranged in order of control moments from front to back to obtain the timestamp alignment sequence.

[0085] When fusing the hidden layer state vector using timestamp alignment sequence, the original timestamp corresponding to each hidden layer state is first read, then the control time corresponding to the original timestamp is determined according to the timestamp alignment sequence, and the hidden layer state is configured to the corresponding control time.

[0086] If one control moment corresponds to only one hidden layer state, then that hidden layer state is directly retained; if one control moment corresponds to two or more hidden layer states, then the original timestamps corresponding to each hidden layer state are arranged in ascending order, and then the features are concatenated according to the arrangement order so that all instruction change information under the same control moment is retained; if a certain control moment does not have a corresponding hidden layer state, then the hidden layer state corresponding to the most recent control moment before that control moment is called to fill the gap, so as to ensure that consecutive control moments have corresponding state features.

[0087] After completing the above processing, the timestamp information of each control moment and the corresponding hidden layer state are written into the same feature record and arranged in ascending order of control moment to form a periodic feature sequence containing timestamp information.

[0088] Example 4:

[0089] S2 includes acquiring a periodic feature sequence and determining whether there is a timing conflict in the periodic feature sequence; if there is a timing conflict, extracting the clock drift feature of the periodic feature sequence; constructing a synchronization reference sequence based on the clock drift feature; using the synchronization reference sequence to perform timestamp alignment processing on the initial sensing command to obtain the first sensing command; filtering the first sensing command to obtain the second sensing command; and performing aggregation processing according to the execution cycle of the second sensing command to obtain a time-uniform aligned command set.

[0090] In this embodiment, after obtaining the periodic feature sequence formed in Embodiment 3, the timestamp, decision cycle and hidden layer state corresponding to each workstation are extracted according to the workstation number, and the time series comparison is performed on the workstations with direct connection relationships according to the workstation topology.

[0091] For two adjacent workstations along the conveying direction, first read the timestamp corresponding to the current cycle feature of the preceding workstation, then read the timestamp corresponding to the current cycle feature of the succeeding workstation. Subtract the timestamp of the preceding workstation from the timestamp of the succeeding workstation to obtain the actual time interval between the two workstations. Then compare this actual time interval with a pre-set propagation time threshold. If the actual time interval is greater than the propagation time threshold, it is determined that there is a timing conflict in the corresponding cycle feature. Similarly, if the timestamp of the succeeding workstation is earlier than the timestamp of the preceding workstation, it is also determined that there is a timing conflict. Only when the timestamp order conforms to the conveying direction and the actual time interval is not greater than the propagation time threshold, it is determined that there is no timing conflict between the two cycle features.

[0092] Furthermore, after detecting timing conflicts, clock drift features are extracted from the periodic features of the conflict.

[0093] In specific processing, the unified control time formed in Example 3 is used as the reference time. The original timestamps from each cycle characteristic of the same workstation are read, and the difference between each original timestamp and its corresponding reference time is calculated. This difference characterizes the offset of the local clock relative to the unified control time at the current control time. Subsequently, two consecutive control times are selected in chronological order, and the offset value corresponding to the later control time is subtracted from the offset value corresponding to the earlier control time to obtain the drift increment within that time interval. All consecutive control times of the workstation are processed in the same way, and the drift increments are arranged in chronological order to obtain the clock drift characteristics corresponding to that workstation. A positive drift increment indicates that the local time has shifted backward relative to the unified control time; a negative drift increment indicates that the local time has shifted forward relative to the unified control time.

[0094] When constructing the synchronization reference sequence based on the clock drift characteristics, the unified alignment starting point determined in Example 3 is used, and continuous reference control times are generated sequentially according to the preset control period determined in Example 1.

[0095] For each workstation, starting from a unified alignment point, the drift increments corresponding to each consecutive control moment are sequentially accumulated to obtain the cumulative drift value for each control moment. Then, the current reference control moment is subtracted from the corresponding cumulative drift value to obtain the corrected time position of the current control moment for that workstation. All corrected time positions are saved in chronological order according to the reference control moments, thus forming the synchronization reference sequence corresponding to that workstation. Using this method, the time positions of different heterogeneous neuromorphic units can be converted to the unified time reference used in Example 1.

[0096] In addition, when using the synchronization reference sequence to align the timestamps of the initial sensing instructions, the workstation number and original timestamp of each initial sensing instruction are read one by one, and the synchronization reference sequence of the corresponding workstation is called according to the workstation number.

[0097] Specifically, find two corrected time positions before and after the original timestamp, calculate the time difference between the original timestamp and the two corrected time positions, and take the absolute value of the two time differences; map the original timestamp to the corrected time position with the smaller absolute value of the time difference; when the two absolute values ​​of the time difference are equal, map to the corrected time position with the smaller time value; after completing the mapping, replace the original timestamp of the initial sensing instruction with the corresponding corrected time position, while keeping the workstation number, speed adjustment value, and instruction status unchanged. The processed instruction is the first sensing instruction.

[0098] When filtering the first sensing instructions, they are first grouped according to the workstation number, and then the first sensing instructions within the same workstation are arranged in ascending order of the correction timestamp.

[0099] Specifically, a filtering window is formed by three consecutive first sensing commands. The speed adjustment values ​​from each of the three commands are read and arranged in ascending order. The second-ranked speed adjustment value is taken as the filtering result of the central command of the filtering window. The station number, correction timestamp, and command status of the central command remain unchanged; only the original speed adjustment value of the central command is replaced by the filtering result. After processing the current window, the filtering window is moved forward one command along the time direction, and the same process continues until all first sensing commands for that station are filtered. After processing each station in the same way, the second sensing command is obtained.

[0100] For the first sensing instruction located at the beginning and end of the instruction sequence at the same workstation, which cannot form a three-instruction filtering window, its speed adjustment value is directly retained, and median replacement is not performed.

[0101] This ensures that the filtering process will not delete instructions or change the already aligned timestamp order; the filtered second sensing instruction still maintains a one-to-one correspondence with the original first sensing instruction, and only the speed adjustment value of abnormal jumps is corrected.

[0102] When performing aggregation processing according to the execution cycle of the second sensing instruction, the decision cycle of each workstation determined in Example 3 is used as the execution cycle of the corresponding second sensing instruction; the unified alignment starting point is used as the starting point of the first time interval, and the execution cycle is used as the length of the time interval to divide the continuous time intervals in sequence; for each second sensing instruction, the time interval to which the instruction belongs is determined according to its correction timestamp, and second sensing instructions whose timestamps fall in the same time interval and belong to the same workstation are grouped into the same aggregation group.

[0103] When processing each aggregation group, the workstation number corresponding to the aggregation group is retained, and the start control time of the time interval is used as the timestamp after aggregation.

[0104] Specifically, all speed adjustment values ​​within a group are added one by one in order of timestamp to obtain the total speed adjustment value. Then, the total speed adjustment value is divided by the number of second sensing instructions within the group, and the result is used as the speed adjustment value of the aggregation group. For the instruction status, the second sensing instruction with the largest timestamp value in the aggregation group is read, and the instruction status corresponding to the instruction is used as the aggregated instruction status. After processing, only one aggregated instruction corresponding to the workstation is retained in each execution cycle.

[0105] Furthermore, after each workstation completes its execution cycle aggregation, all aggregation instructions are arranged from front to back according to the unified control time; when there are more than two workstation aggregation instructions at the same control time, they are arranged according to the workstation number order.

[0106] If a workstation does not have a new aggregate instruction at the current control time, the aggregate instruction corresponding to the previous control time of that workstation is called, and only the timestamp is updated to the current control time, while the speed adjustment value and instruction status remain unchanged. Through the above processing, the instructions of all workstations correspond to the same control time, forming a unified set of aligned instructions with a unified time dimension.

[0107] Example 5:

[0108] S3 includes obtaining an alignment instruction set, constructing a workstation node matrix based on the instruction flow path of the alignment instruction set, calculating adjacency weights based on the workstation node matrix and updating the workstation node matrix to obtain a topology connection graph, processing the topology connection graph using a graph convolutional network to extract the spatial distribution features of the topology connection graph, performing dimensionality reduction on the spatial distribution features to obtain a feature aggregation vector, constructing a spatial mapping matrix through the feature aggregation vector and performing normalization processing to obtain a spatial feature vector representing the topology relationship of the workstations.

[0109] In this embodiment, after obtaining the alignment instruction set formed by the processing in Embodiment 4, the instruction flow path is first determined based on the workstation number, timestamp, and instruction transmission direction of each alignment instruction.

[0110] Specifically, taking the alignment instruction corresponding to any workstation as the starting instruction, the alignment instruction corresponding to the next workstation is searched according to the material conveying relationship between workstations, and the timestamps of the two instructions are compared. If there is a direct conveying relationship between the current workstation and the next workstation, and the timestamp of the next instruction is greater than the timestamp of the previous instruction, the workstations corresponding to the two instructions are recorded as continuous flow nodes. The same method is used to continue searching until there are no subsequent instructions that satisfy the conveying relationship and time sequence relationship, thereby determining one instruction flow path.

[0111] Furthermore, after forming the instruction flow path, a workstation node matrix is ​​constructed using workstations as nodes.

[0112] Specifically, each workstation corresponds to one row in the matrix, arranged in the order of workstation number; each row must contain at least the speed adjustment value, instruction status, execution cycle, and the position number of the workstation in the current instruction flow path; for workstations that do not generate new instructions at the same control moment, the aggregated instruction corresponding to the previous control moment of the workstation in Example 4 is used to fill the row, so as to ensure that each workstation has corresponding node data at the same control moment; thus, a workstation node matrix reflecting the instruction status of each workstation at the current moment is obtained.

[0113] When calculating the adjacency weights based on the workstation node matrix, for any two workstations, it is first determined whether they are located on the same instruction flow path.

[0114] If there is no sequential flow relationship between two workstations, the weight of the corresponding adjacency relationship is set to 0; if there is a sequential flow relationship between two workstations, the timestamp of the corresponding instruction of the two workstations and the position number of the two workstations in the flow path are extracted respectively.

[0115] For two workstations with a sequential flow relationship, first, subtract the timestamp of the previous workstation's instruction from the timestamp of the instruction at the subsequent workstation to obtain the actual time interval between the two instructions. Then, subtract the position number of the previous workstation from the position number of the subsequent workstation in the flow path to obtain the node interval number. According to a pre-defined time association table, the actual time interval is mapped to a time association weight. When the actual time interval equals the corresponding execution cycle, the base time weight is used. For each execution cycle increase in the actual time interval, the time weight is reduced by one level according to a preset decreasing step size. Next, the path association weight is determined based on the node interval number. When the node interval number is 1, the first-level path association weight is used. For each additional node interval number, the path association weight is reduced by one level according to a preset decreasing step size. Multiplying the time association weight by the path association weight yields the adjacency relationship weight between the two workstations.

[0116] Furthermore, after calculating the adjacency weights between all workstation nodes, the weights of each adjacency are written into the connection relationships of the corresponding nodes.

[0117] Specifically, two nodes with an adjacency weight of 0 are not connected; two nodes with an adjacency weight greater than 0 are connected in a directed manner from the preceding workstation to the following workstation according to the direction of instruction flow, and the calculated adjacency weight is used as the edge weight of the connection; the topology connection graph is formed by workstation nodes, node features, directed connections and edge weights.

[0118] When using graph convolutional networks to process the topology connection graph, for each target workstation, the node features of that workstation itself and the features of all preceding nodes pointing to that target workstation are read.

[0119] For each preceding node, the feature values ​​corresponding to the preceding node are multiplied by the adjacency weight between the preceding node and the target workstation to obtain the weighted feature of the preceding node to the target workstation. Then, the weighted features of all preceding nodes are added one by one according to the same feature position, and added to the data of the corresponding position of the node feature of the target workstation itself to obtain the aggregated node feature of the target workstation.

[0120] Furthermore, after inputting the aggregated node features into the feature transformation layer of the graph convolutional network, the feature values ​​in the aggregated node features are weighted and accumulated according to the pre-trained convolution parameters. The weighted accumulation results are then written into the corresponding output feature positions to obtain the updated node features of the target workstation. After performing the same processing on all workstations, the updated node features of each workstation are arranged in order of workstation number to form the spatial distribution features corresponding to the topological connection graph.

[0121] Since the updated node features contain both the current workstation's own information and the information of the preceding workstations passed along the instruction flow path, they can reflect the spatial distribution of each workstation under the topological connection relationship.

[0122] In particular, when performing dimensionality reduction processing on spatial distribution features, a linear mapping layer with a pre-determined output dimension is used.

[0123] Specifically, for each workstation, the feature values ​​are read in the original column order of the spatial distribution characteristics, and each feature value is multiplied by its corresponding mapping coefficient. The product results belonging to the same output position are accumulated to obtain the dimensionality-reduced feature value of that output position. After calculating all output positions in the same way, the dimensionality-reduced feature of that workstation is obtained. The dimensionality reduction process is completed for all workstations in sequence, and then the dimensionality-reduced features of each workstation are connected in sequence according to the workstation number to form a feature aggregation vector.

[0124] In addition, when constructing the spatial mapping matrix using the feature aggregation vector, the feature aggregation vector is divided into corresponding feature segments according to the number of workstations, with each feature segment corresponding to one workstation; each feature segment is written into the corresponding row of the matrix according to the workstation number, so that each row in the matrix corresponds to one workstation and each column corresponds to one dimension-reduced spatial feature, thus obtaining the spatial mapping matrix.

[0125] When normalizing the spatial mapping matrix, each column is treated as an independent processing object.

[0126] Specifically, first read all the values ​​in the current column and determine the maximum and minimum values ​​of the column; for any 1 value in the current column, first subtract the minimum value of the current column from the value, and then divide the resulting difference by the difference between the maximum and minimum values ​​of the current column to obtain the normalized feature value; if the maximum and minimum values ​​of the current column are equal, then set all normalization results of the column to 0; after processing all columns in the same way, each feature column is transformed to the same numerical range.

[0127] Finally, the data in each row of the normalized spatial mapping matrix is ​​read sequentially according to the workstation number, and the normalized features in each row are connected sequentially to obtain a spatial feature vector representing the topological relationship between workstations. This spatial feature vector preserves the instruction flow relationship between workstations, the node connection strength, and the spatial association result after graph convolution processing.

[0128] Example 6:

[0129] S4 includes acquiring spatial feature vectors and periodic feature sequences to construct a spatiotemporal dimension matrix; extracting unidirectional workstation topology relationships based on the spatiotemporal dimension matrix; determining the instruction propagation path according to the unidirectional workstation topology relationships and timestamp order; determining whether there is a closed loop in the initial causal order of the aligned instruction propagation path; if there is a closed loop, disconnecting the connection edge to obtain the updated causal order; and obtaining spatiotemporal correlation features with sequential dependencies through the updated causal order.

[0130] In this embodiment, the spatial feature vector obtained in Embodiment 5 and the periodic feature sequence obtained in Embodiment 3 are obtained, and a spatiotemporal dimension matrix is ​​established using the workstation number and unified control time as corresponding indices.

[0131] Specifically, the spatial feature vector is first divided into spatial feature segments corresponding to each workstation according to the workstation number. Then, the periodic features corresponding to each workstation in the periodic feature sequence are read according to the unified control time. For the same workstation and the same control time, the spatial feature segment and the corresponding periodic feature of the workstation are written into the same matrix position, so that the rows of the matrix correspond to the workstation and the columns correspond to the unified control time. Each matrix position contains the workstation topology features, instruction timestamps and timing status. Thus, a spatiotemporal dimension matrix that reflects both the spatial relationship of the workstation and the temporal relationship of the instruction is obtained.

[0132] When extracting the topological relationship of a unidirectional workstation based on the spatiotemporal dimension matrix, the connection direction and adjacency weight between each workstation are read one by one, and only connections that are consistent with the material conveying direction and have an adjacency weight greater than 0 are retained.

[0133] If the material at station 1 can be directly transported to station 2, then the connection from station 1 to station 2 is retained; if there is no direct transport relationship, or the connection direction is opposite to the material transport direction, then no corresponding one-way connection is established; after processing all stations in sequence, a one-way station topology relationship arranged along the material transport direction is obtained.

[0134] When determining the instruction propagation path based on the unidirectional workstation topology and timestamp sequence, start with the alignment instruction with the smallest timestamp value in the current analysis period and take its workstation as the propagation starting point.

[0135] Furthermore, the next workstation is searched along the unidirectional topology connection of the current workstation, and the alignment instruction in the next workstation is read. If there are more than two candidate instructions in the next workstation, only the candidate instruction with the smallest timestamp value and a timestamp greater than the current instruction's timestamp is retained. The timestamp of the current instruction is then subtracted from the timestamp of the candidate instruction to obtain the propagation time interval between the two instructions. When the propagation time interval is not greater than the preset propagation time threshold, the candidate instruction is added to the current propagation path. When the propagation time interval is greater than the preset propagation time threshold, the propagation tracking in the current direction is terminated. The same process is continued to search backwards until there are no subsequent instructions that meet the topology connection conditions and time conditions, thus forming a complete alignment instruction propagation path.

[0136] When establishing the initial causal order based on each propagation path, each alignment instruction in the propagation path is used as a causal node, and connecting edges are established between adjacent causal nodes according to the propagation direction.

[0137] For two adjacent nodes, a causal connection is established only if the workstation to which the previous node belongs is located at the front end of the conveyor of the workstation to which the next node belongs, and the timestamp of the previous node is less than the timestamp of the next node; after all propagation paths have been processed, the initial causal order is obtained.

[0138] When determining whether a closed loop exists in the initial causal sequence, start from any causal node and visit subsequent nodes level by level along the connection direction, and write the visited node numbers into the current path record in the order of visit.

[0139] Specifically, for each new node visited, it is determined whether the node number already exists in the current path record. If it already exists, it means that the path returns to the node that has appeared in the current path along the connection direction, and a closed loop is formed. If a node without subsequent connection edges is visited, and there are no duplicate nodes during the visit, then the visit path does not have a closed loop. The above checks are performed on all causal nodes in sequence to complete the closed loop identification.

[0140] When a closed loop is detected, the time interval is calculated for each connecting edge that constitutes the closed loop.

[0141] Specifically, the time interval of each connecting edge is obtained by subtracting the timestamp of the preceding node from the timestamp of the subsequent node; the time intervals of all connecting edges within the closed loop are compared, and the connecting edge with the largest time interval value is selected as the first edge to be broken; if there are two or more connecting edges with the same time interval value, the adjacency weights of the corresponding connecting edges are compared, and the connecting edge with the smallest adjacency weight value is selected as the edge to be broken; if both the time interval and the adjacency weight are the same, the first connecting edge is selected to be broken in ascending order of the timestamp of the endpoint node.

[0142] Furthermore, after completing one edge break, the path is accessed again from the starting node involved in the closed loop along the remaining connecting edges, and it is determined whether it is still possible to return to the already visited node.

[0143] If a closed loop still exists, the next connection edge to be disconnected will be determined according to the aforementioned time interval, adjacency weight, and endpoint timestamp, until all access paths can no longer return to the visited nodes in the current path, thus obtaining the updated causal order. The edge disconnection process only deletes the connection edges that cause circular dependencies, without modifying the original workstation number, timestamp, speed adjustment value, and instruction status of each alignment instruction.

[0144] In addition, when obtaining spatiotemporal correlation features through the updated causal order, the preceding and following nodes are read one by one according to the direction of the connecting edges.

[0145] Specifically, for each group of nodes with direct causal connections, the workstation to which the preceding node belongs, the workstation to which the following node belongs, the timestamp of the preceding node, the timestamp of the following node, the time interval between the two nodes, the corresponding adjacency weight, and the spatial and periodic characteristics of the two nodes are recorded. For continuous propagation paths containing more than three nodes, each level of association record is generated sequentially according to the connection order from the preceding node to the following node, so that the following node in the previous level association record becomes the preceding node in the next level association record.

[0146] After completing the association records for all propagation paths, first arrange them in ascending order of the timestamp of the starting node of each propagation path; when the timestamps of the starting nodes are the same, then arrange them in ascending order of the starting workstation number; the sorted association records together constitute the spatiotemporal association features.

[0147] Example 7:

[0148] S5 includes extracting conflict nodes and delayed transmission paths based on spatiotemporal correlation features; if the conflict nodes and delayed transmission paths form a closed loop, it is determined that the alignment instruction set produces a chain reaction, resulting in a subset of associated instructions; calculating instruction priority values ​​based on the subset of associated instructions; and sorting the instruction priority values ​​to determine the target priority sequence.

[0149] In this embodiment, after obtaining the spatiotemporal correlation features formed in Embodiment 6, the conflict relationship and delay transmission relationship between instructions are extracted according to the preceding workstation, following workstation, preceding instruction timestamp, following instruction timestamp, speed adjustment direction, time interval, and adjacency relationship weight recorded in each spatiotemporal correlation feature.

[0150] When extracting conflict nodes, two instructions that have a direct sequential dependency relationship in the same propagation path are grouped together for comparison.

[0151] Specifically, the speed adjustment directions corresponding to the two instructions are first read. When the preceding instruction requires increasing the conveying speed and the following instruction requires decreasing the conveying speed, the nodes corresponding to the two instructions are marked as conflict nodes. Similarly, when the preceding instruction requires decreasing the conveying speed and the following instruction requires increasing the conveying speed, they are also marked as conflict nodes. When the speed adjustment directions of the two instructions are consistent, no conflict marking is performed. For cases where one instruction conflicts with two or more subsequent instructions in the same direction, only one node identifier is retained, and the connection relationship between this node and each subsequent conflict node is recorded, thus forming a set of conflict nodes.

[0152] As a preferred embodiment, the delayed transmission path is determined based on the actual transmission time between instructions.

[0153] Specifically, for each preceding and following connection in the spatiotemporal correlation features, the timestamp of the preceding instruction is subtracted from the timestamp of the following instruction to obtain the actual transmission time corresponding to the connection. The actual transmission time is then compared with the preset propagation time threshold corresponding to the connection. If the actual transmission time is greater than the preset propagation time threshold, the preceding and following connection is determined to be a delayed connection. If the actual transmission time is not greater than the preset propagation time threshold, it is not considered a delayed connection. Subsequently, starting from each conflict node, subsequent delayed connections are searched sequentially along the updated causal order formed in Embodiment Six. The consecutively existing delayed connections are connected in the preceding and following order to form a delayed transmission path corresponding to the conflict node. If there is no subsequent delayed connection at any node, the delayed transmission path ends at that node.

[0154] When determining whether a conflict node and a delayed propagation path form a closed loop, the associated path is established by combining the conflict relationship and the delayed propagation relationship.

[0155] Specifically, starting from a conflict node, subsequent nodes are visited sequentially along the delayed propagation path corresponding to that conflict node. Upon reaching the end node of the delayed propagation path, it is determined whether there is a speed adjustment direction conflict relationship between the end node and the starting conflict node. If a conflict relationship exists, the conflict relationship is used as a return relationship to connect the end node with the starting conflict node, forming a closed loop consisting of the starting conflict node, continuous delayed connections, and the return conflict relationship. If no conflict relationship exists, no closed loop is formed. Thus, the closed loop determined in this step is used to characterize the instruction after delayed propagation having a reverse control effect on the original conflict node again.

[0156] When the aforementioned closed loop exists, it is determined that the set of alignment instructions corresponding to the closed loop will generate a chain reaction.

[0157] Subsequently, starting from the initial conflict node of the closed loop, all instruction nodes traversed by the delayed propagation path are extracted sequentially, and the terminal nodes that form a return conflict relationship with the initial conflict node are extracted. The extracted instructions are arranged in ascending order of their original timestamps. If two instructions have the same timestamp, they are arranged in ascending order of their workstation numbers. If the same instruction is extracted repeatedly because it participates in two delayed propagation paths, it is retained only once, and its workstation number, timestamp, speed adjustment value, instruction status, and propagation path position are retained, thus obtaining the associated instruction subset.

[0158] Specifically, when calculating the instruction priority value based on the subset of associated instructions, the time sequence value, propagation coverage value, and path position value are determined for each associated instruction.

[0159] Specifically, the time sequence value is determined according to the timestamp order in the associated instruction subset. The instruction with the smallest timestamp value is recorded as the first sequence position, and the sequence position increases by 1 for each subsequent instruction. Then, the current instruction's sequence position is subtracted from the total number of instructions in the associated instruction subset, and then increased by 1. The resulting value is used as the time sequence value of that instruction. Thus, the instruction with the earlier timestamp gets a larger time sequence value.

[0160] In this embodiment, the propagation coverage value is obtained by performing node statistics along the causal order updated in Embodiment Six.

[0161] Starting from the current associated instruction, count the number of workstation nodes that can be reached along its subsequent connection direction. The workstation to which the current instruction belongs is included in the count. If the same workstation appears repeatedly in the same propagation path, it is counted only once. The total number of workstations obtained is used as the propagation coverage value of the associated instruction.

[0162] In this embodiment, the path location value is determined based on the position of the instruction in the delayed transmission path corresponding to the closed loop.

[0163] Specifically, first count the total number of instruction nodes in the current delayed delivery path, then determine the current instruction's position starting from the initial conflict node; subtract the current instruction's position from the total number of nodes in the delayed delivery path, and then add 1 to obtain the instruction's path position value; therefore, the path position value obtained by the instruction at the beginning of the delayed delivery path is greater than that of the instruction at the end of the path.

[0164] Furthermore, after completing the above three numerical calculations, the time sequence value, propagation coverage value, and path position value corresponding to the same associated instruction are added together in sequence, and the sum is used as the instruction priority value of the associated instruction.

[0165] By using the same calculation rules to process all instructions in the associated instruction subset, the priority of each instruction is determined by the specific time sequence, the number of propagation stations, and the path location, without using subjective levels assigned manually.

[0166] Furthermore, after obtaining all instruction priority values, they are sorted from largest to smallest, with the instruction with the largest value at position 1 of the target priority sequence, and the remaining instructions arranged in descending order of value. If two related instructions have the same priority value, their timestamps are compared, with the instruction with the smaller timestamp value placed first. If both the priority value and timestamp are the same, the corresponding workstation positions are compared according to the conveying direction, with the instruction at the workstation at the front of the conveying direction placed first. After the above sorting, a uniquely determined target priority sequence is obtained.

[0167] Example 8:

[0168] S6 includes constructing a conflict detection matrix based on the execution cycle and resource status of the alignment instruction set; mapping the target priority sequence through the conflict detection matrix to obtain a priority mapping table; assigning weights to the alignment instruction set according to the priority mapping table to obtain initial weighted instructions; if the initial weighted instructions trigger the deviation threshold, generating an error compensation vector for the initial weighted instructions; and correcting the initial weighted instructions through the error compensation vector to obtain a weighted instruction matrix that eliminates control errors.

[0169] In this embodiment, the execution cycle, workstation number, speed adjustment value and resource usage information corresponding to each instruction in the alignment instruction set are first obtained, and the resource status of each workstation in the current execution cycle is read.

[0170] Specifically, the resource status includes the currently allocable drive capacity, the currently occupied drive capacity, and the current remaining speed adjustment margin; for each alignment instruction, its execution cycle, the workstation to which it belongs, the speed adjustment direction, the required drive capacity, and the current resource status are written into the same record, and then arranged in order according to the execution cycle and workstation number, thereby constructing a conflict detection matrix.

[0171] In constructing the conflict detection matrix, instructions within the same execution cycle are compared pairwise.

[0172] Specifically, if two instructions act on the same workstation, and one instruction corresponds to the acceleration direction while the other corresponds to the deceleration direction, then the matrix positions corresponding to the two instructions are recorded as conflict states. If two instructions act on different workstations but call the same restricted drive resource, then the required drive capabilities of the two instructions are added together and compared with the allocable drive capabilities of the corresponding resource within the execution cycle. When the sum of the required drive capabilities is greater than the allocable drive capabilities, the corresponding matrix position is recorded as a conflict state; otherwise, it is recorded as a non-conflict state. After comparing each item, a conflict detection matrix that reflects the conflict of instruction direction and resource occupation within the current execution cycle is obtained.

[0173] When mapping the target priority sequence obtained in Example 7 to the conflict detection matrix, instructions are read one by one from front to back according to the target priority sequence, and the conflict status of the instruction in the conflict detection matrix is ​​searched. Instructions without conflict status retain their original positions. For two instructions with conflict status, their positions in the target priority sequence are compared. The instruction with the earlier position in the sequence is recorded as the instruction to be executed first, and the instruction with the later position in the sequence is recorded as the instruction to be executed later. After all instructions are mapped in sequence, a priority mapping table is formed. The priority mapping table records at least the instruction number, workstation number, execution cycle, target priority position, conflict status, and execution order.

[0174] In addition, when assigning weights according to the priority mapping table, the number of valid instructions that participate in the processing within the current execution cycle is first counted, and basic weight values ​​are assigned according to the execution order.

[0175] Specifically, the first valid instruction is assigned a base weight value equal to the number of valid instructions, the second instruction is assigned a base weight value one less than the first, and the remaining instructions are assigned a base weight value decreasing by one in turn until the last instruction has a base weight value of 1. For instructions marked as to be executed later and that have resource conflicts, one is deducted from their base weight value. If the result after deduction is 0, it is retained as 1 to ensure that the instruction still participates in subsequent weighting processing.

[0176] Furthermore, after determining the basic weight values, the basic weight values ​​of all valid instructions in the current execution cycle are added together to obtain the total weight value. Then, the basic weight value corresponding to each instruction is divided by the total weight value to obtain the assigned weight of that instruction.

[0177] Subsequently, the speed adjustment value of each alignment instruction is multiplied by the corresponding assigned weight to obtain the initial weighted speed adjustment value, and the workstation, execution cycle and instruction status are retained to form the initial weighted instruction.

[0178] When determining whether the initial weighted instruction triggers the deviation threshold, all initial weighted speed adjustment values ​​in the current execution cycle are first merged according to direction; the initial weighted speed adjustment values ​​in the acceleration direction are added sequentially, and the initial weighted speed adjustment values ​​in the deceleration direction are subtracted sequentially according to absolute value to obtain the net speed adjustment value corresponding to the current execution cycle; then the current running speed of the conveyor belt is merged with the net speed adjustment value to obtain the expected execution speed of the current execution cycle, and the absolute value of the difference between the expected execution speed and the expected running speed in Example 1 is calculated.

[0179] The deviation threshold is determined based on the speed fluctuation range of the conveyor belt under stable operating conditions.

[0180] Specifically, when the conveyor belt is running at a constant target speed, at least 30 actual operating speed values ​​are continuously collected. The absolute value of the difference between each actual operating speed and the target speed is calculated. Then, all the absolute values ​​of the difference are added together and divided by the number of samples to obtain the average speed deviation. Subsequently, the maximum value among all the absolute values ​​of the difference is compared with the average speed deviation, and the larger of the two values ​​is taken as the deviation threshold. The deviation threshold determined in this way corresponds to the upper limit of speed fluctuation that the current conveyor belt can have under stable operating conditions.

[0181] When the absolute value of the difference between the expected execution speed and the expected running speed is greater than the deviation threshold, the initial weighted instruction is determined to trigger the deviation threshold, and error compensation processing begins; when the absolute value of the difference is less than or equal to the deviation threshold, the current initial weighted instruction is retained, and no error compensation vector is generated.

[0182] Furthermore, after triggering the deviation threshold, the direction of the deviation between the expected execution speed and the expected running speed is first determined.

[0183] If the expected execution speed is greater than the expected operating speed, the expected execution speed is subtracted from the expected operating speed, and the result is the total amount of compensation to be reduced; if the expected execution speed is less than the expected operating speed, the expected operating speed is subtracted from the expected execution speed, and the result is the total amount of compensation to be increased.

[0184] When it is necessary to reduce the total compensation, the initial weighting instruction for the acceleration direction is searched from back to front according to the priority mapping table, and the reduction is executed starting from the last position of the sequence.

[0185] If the initial weighted speed adjustment value of the current acceleration command is greater than or equal to the unallocated total compensation amount, the unallocated total compensation amount is directly deducted from the weighted speed adjustment value, and the remaining compensation amount is recorded as 0; if the current weighted speed adjustment value is less than the unallocated total compensation amount, the weighted speed adjustment value is reduced to 0, and the deducted value is subtracted from the unallocated total compensation amount, and the previous acceleration direction command is processed again.

[0186] When it is necessary to increase the total compensation, the initial weighting instruction for the deceleration direction is first searched from back to front according to the priority mapping table, and the corresponding deceleration amount is reduced in turn.

[0187] If the current deceleration amount can completely offset the unallocated total compensation amount, the corresponding deceleration amount is reduced directly, and the compensation ends; if the current deceleration amount is insufficient to offset the unallocated total compensation amount, the deceleration amount is adjusted to 0, and the previous deceleration command is processed again; if there is still a remaining compensation amount after all deceleration commands have been processed, the weighted speed adjustment value of the acceleration direction command is increased from front to back according to the priority mapping table until the remaining compensation amount is 0.

[0188] The increase or decrease amount corresponding to each initial weighted instruction is arranged in order of workstation number and execution cycle to form an error compensation vector. Each compensation value in the error compensation vector corresponds one-to-one with an initial weighted instruction, where the decrease processing is recorded as negative compensation and the increase processing is recorded as positive compensation.

[0189] When correcting the initial weighted instruction using the error compensation vector, the compensation value corresponding to the same workstation and the same execution cycle is merged with the initial weighted speed adjustment value; negative compensation is deducted from the original initial weighted speed adjustment value, and positive compensation is added to the original initial weighted speed adjustment value. During the correction process, the workstation number, execution cycle, and instruction status remain unchanged, and only the corresponding weighted speed adjustment value is updated.

[0190] After all initial weighted instructions have been corrected, the corrected weighted speed adjustment value is written to the corresponding position according to the workstation number as the row index and the execution cycle as the column index; when there is no valid instruction for a certain workstation in a certain execution cycle, 0 is written to the corresponding position; thus forming a weighted instruction matrix.

[0191] Example 9:

[0192] S7 includes extracting load tension distribution features based on the weighted instruction matrix and target priority sequence, calculating the slip compensation value corresponding to the load tension distribution features in combination with the friction resistance matrix; determining the speed adjustment amount based on the slip compensation value, and sequentially fusing the speed adjustment amounts corresponding to each workstation according to the target priority sequence to obtain the global dynamic traction force; if the global dynamic traction force exceeds the traction force threshold, then the global dynamic traction force is limited to obtain the target dynamic traction force; the expected operating speed is calculated through the target dynamic traction force, and the initial belt speed adjustment command for the global collaborative conveyor belt neuromorphic robot is obtained based on the expected operating speed.

[0193] In this embodiment, the weighted instruction matrix formed in Embodiment 8 and the target priority sequence formed in Embodiment 7 are obtained, and the weighted speed adjustment value of each workstation in the current execution cycle is read one by one according to the target priority sequence.

[0194] Specifically, the inlet side tension, outlet side tension, current material mass, and belt mass of the corresponding conveyor section are simultaneously acquired for each workstation. For each workstation, the inlet side tension is first subtracted from the outlet side tension to obtain the tension difference for that workstation. Then, the outlet side tension and the inlet side tension are added together to obtain the tension reference value. When the tension reference value is not 0, the tension difference is divided by the tension reference value to obtain the tension distribution value for that workstation. When the tension reference value is 0, the corresponding tension distribution value is recorded as 0. The tension difference, tension distribution value, and material mass of each workstation are arranged in order of workstation number to form the load tension distribution characteristics.

[0195] The friction resistance matrix is ​​established based on the load-bearing state and resistance parameters of each conveying section.

[0196] Specifically, for each conveying section corresponding to a workstation, the mass of the material carried in the section is added to the mass of the conveyor belt in the section to obtain the total mass of the section; then the total mass of the section is multiplied by the gravitational acceleration to obtain the normal load acting on the bearing surface of the section; subsequently, the normal load is multiplied by the pre-calibrated running resistance coefficient of the section to obtain the frictional resistance of the corresponding workstation; the frictional resistance of each workstation is written into the main diagonal position of the matrix; for two adjacent workstations with a direct conveying relationship, the frictional resistances of the two workstations are added and divided by 2, and the resulting value is written into the corresponding connection position; for workstations without a direct conveying relationship, 0 is written between them, thus obtaining the frictional resistance matrix.

[0197] The operating resistance coefficient is determined during the equipment calibration phase.

[0198] Specifically, when the conveyor belt is running unloaded and at a constant speed, the stable traction force at the drive end is read, and the normal load generated by the belt mass of the conveying section is obtained. The stable traction force is divided by the corresponding normal load, and the resulting value is used as the running resistance coefficient of the section. For conveying sections with different resistance characteristics under load and unload conditions, the same calibration process can be performed again under rated load conditions, and the running resistance coefficient obtained under rated load conditions can be used in the calculation of the current conveying operation.

[0199] When calculating the slip compensation value based on the load tension distribution characteristics and friction resistance matrix, the tension difference and corresponding friction resistance are read for each station.

[0200] Specifically, the absolute value of the tension difference is first taken, and then the frictional resistance of the corresponding station is subtracted to obtain the effective tension difference that can be used to overcome slip at the current station. If the result is greater than 0, the effective tension difference is divided by the traction stiffness corresponding to the station to obtain the slip compensation value. If the result is less than or equal to 0, the frictional resistance is subtracted from the absolute value of the tension difference, and the difference is divided by the traction stiffness to obtain the slip compensation value that needs to be increased. The compensation direction is determined according to the relationship between the tension on the inlet side and the tension on the outlet side. When the tension on the outlet side is greater than the tension on the inlet side, it is used to reduce the corresponding speed adjustment amount. When the tension on the outlet side is less than the tension on the inlet side, it is used to increase the corresponding speed adjustment amount.

[0201] The traction stiffness is obtained through equipment calibration.

[0202] Specifically, under the same load conditions, the conveyor belt is run in two stable traction states, and the traction force and corresponding stable linear speed are recorded twice. The traction force difference is obtained by subtracting the first traction force from the second traction force. The speed difference is obtained by subtracting the first stable linear speed from the second stable linear speed. The speed difference is obtained by dividing the speed difference by the traction force difference. The result is used as the traction stiffness corresponding to the station. Calibration data with a speed difference of 0 are not included in the traction stiffness calculation, and the calibration is re-performed.

[0203] When determining the speed adjustment amount based on the slip compensation value, the weighted speed adjustment value of the corresponding workstation in the weighted instruction matrix is ​​read.

[0204] If the slip compensation direction is to increase speed, then the slip compensation value is added to the weighted speed adjustment value; if the slip compensation direction is to decrease speed, then the slip compensation value is deducted from the weighted speed adjustment value; if the deduction result crosses 0, then the corresponding speed adjustment amount is limited to 0, and its adjustment direction is not changed; thus, the speed adjustment amount after slip compensation for each station is obtained.

[0205] In addition, when integrating the speed adjustment amounts of each workstation according to the target priority sequence, the dynamic traction force corresponding to the speed adjustment amount of each workstation is calculated first.

[0206] Specifically, for the current workstation, the mass of the material carried at the workstation is added to the mass of the corresponding conveyor belt section to obtain the equivalent motion mass of the workstation; the speed adjustment amount of the workstation is divided by the duration of the current execution cycle to obtain the corresponding speed change rate; the equivalent motion mass is multiplied by the speed change rate to obtain the inertial traction force used to change the running speed; the inertial traction force is then added to the frictional resistance corresponding to the workstation to obtain the dynamic traction force of the workstation; when the speed adjustment amount corresponds to the deceleration direction, the inertial traction force is included according to the direction of decreasing traction.

[0207] Furthermore, after completing the dynamic traction force calculation for each workstation, the dynamic traction force of each workstation is read sequentially from the first position according to the target priority sequence.

[0208] First, obtain the basic traction force required by the conveyor belt under the current load and current running speed. Then, increase the dynamic traction force corresponding to the acceleration direction in the target priority sequence to the basic traction force in turn, and deduct the dynamic traction force corresponding to the deceleration direction from the current traction force in turn. After completing the processing of all workstations, the obtained traction force is used as the global dynamic traction force.

[0209] In this embodiment, the basic traction force is obtained through the stable operating state under the current load conditions.

[0210] Specifically, with the current material load remaining unchanged, the conveyor belt is kept at its current operating speed without any speed adjustment. When the actual operating speed remains within the preset stable speed error range for three consecutive execution cycles, the traction force at the drive end is read, and the traction forces corresponding to the three execution cycles are added together and divided by 3. The resulting value is used as the base traction force. The stable speed error range uses the allowable measurement error of the equipment speed detection unit as the boundary, without introducing any additional empirical proportions.

[0211] The traction threshold is determined based on the allowable output traction force of the drive unit and the anti-slip traction capability between the drive roller and the conveyor belt.

[0212] First, read the maximum permissible output traction force set on the drive unit nameplate or controller; then read the static friction coefficient between the drive roller and the conveyor belt and the current normal pressure. Multiply the static friction coefficient and the normal pressure to obtain the upper limit of the anti-slip traction force; compare the maximum permissible output traction force with the upper limit of the anti-slip traction force, and determine the smaller value as the traction force threshold.

[0213] The static friction coefficient is obtained through the calibration process before the equipment is put into operation.

[0214] Specifically, while keeping the drive roller pressed in a constant state, the drive output is gradually increased starting from a low traction force, while simultaneously monitoring the roller circumferential speed and the conveyor belt linear speed. When the two first become unequal, the traction force corresponding to the previous stage when no relative slippage occurred is recorded, and this traction force is divided by the normal pressure at that time. The resulting value is used as the static friction coefficient. The normal pressure is determined based on the tension force of the tensioning device and the force relationship corresponding to the wrap angle of the drive roller, and is used for subsequent traction force threshold calculations under the same tensioning state.

[0215] The global dynamic traction force is compared with the traction force threshold.

[0216] When the global dynamic traction force is less than or equal to the traction force threshold, the global dynamic traction force is directly used as the target dynamic traction force; when the global dynamic traction force is greater than the traction force threshold, the global dynamic traction force is subtracted from the traction force threshold to obtain the traction force value that needs to be reduced, and the traction force threshold is used as the amplitude limiting target.

[0217] In addition, when performing amplitude limiting processing, the workstation that generates positive traction force increment is searched from back to front according to the target priority sequence.

[0218] If the positive traction force increment corresponding to the current workstation is greater than or equal to the traction force value that has not yet been reduced, then all the values ​​that have not yet been reduced are deducted from the dynamic traction force of that workstation, and the amplitude limiting process ends; if the positive traction force increment of the current workstation is less than the traction force value that has not yet been reduced, then the positive traction force increment of that workstation is reduced to 0, and the amount of reduction is subtracted from the traction force value that has not yet been reduced, and then the processing continues for the first workstation in the target priority sequence; after processing, the target dynamic traction force is obtained.

[0219] When calculating the expected operating speed using the target dynamic traction force, the frictional resistance of each station in the current execution cycle is first added together to obtain the total frictional resistance. Then, the total frictional resistance is subtracted from the target dynamic traction force to obtain the net traction force used to change the operating state of the conveyor belt.

[0220] Subsequently, the mass of the conveyor belt's moving portion is added to the mass of the material currently being carried to obtain the total moving mass; the net traction force is divided by the total moving mass to obtain the speed change rate of the current execution cycle; and the speed change rate is multiplied by the execution cycle duration to obtain the speed change amount within the current execution cycle.

[0221] When the speed change is positive, the speed change is added to the current actual operating speed; when the speed change is negative, its absolute value is subtracted from the current actual operating speed to obtain the calculated operating speed. The calculated operating speed is then compared with the upper and lower limits of the allowable operating speed of the conveyor equipment. If the operating speed exceeds the upper limit, the upper limit is used; if it is below the lower limit, the lower limit is used; otherwise, the calculated operating speed is used to determine the expected operating speed. Based on the expected operating speed, the initial belt speed adjustment command for the global collaborative conveyor belt neuromorphic robot is generated.

[0222] Example 10:

[0223] S8 includes obtaining the actual running speed and expected running speed corresponding to the initial belt speed adjustment command; if the actual running speed is greater than the expected running speed, calculating the speed deviation value; matching the target priority sequence and locating the weighted command based on the speed deviation value; using the weighted command to perform directional numerical decrement processing on the initial belt speed adjustment command to obtain the speed adjustment execution sequence; and determining the final belt speed adjustment command of the conveyor belt brain-like robot based on the speed adjustment execution sequence.

[0224] In this embodiment, after executing the initial belt speed adjustment command obtained in Embodiment Nine, the expected running speed recorded in the command is read, and the actual running speed of the conveyor belt is collected within the same execution cycle.

[0225] The actual operating speed is obtained by continuous sampling by the speed detection device. When there are more than two sampling values ​​in one execution cycle, all sampling values ​​in the execution cycle are added together and then divided by the number of samples. The resulting value is taken as the actual operating speed corresponding to that execution cycle. This ensures that the actual operating speed and the expected operating speed have the same time reference.

[0226] In addition, the actual operating speed is compared with the expected operating speed.

[0227] When the actual operating speed is less than or equal to the expected operating speed, the initial belt speed adjustment command remains unchanged; when the actual operating speed is greater than the expected operating speed, the expected operating speed is subtracted from the actual operating speed, and the positive value obtained is used as the speed deviation value; this speed deviation value represents the overspeed amount that needs to be eliminated from the original acceleration adjustment amount within the current execution cycle.

[0228] When matching the target priority sequence based on the speed deviation value, the target priority sequence determined in Example 7 is invoked, and the weighted instruction matrix formed in Example 8 is read.

[0229] Read the corresponding weighted instructions one by one from back to front according to the target priority sequence, and retain only the instructions with the speed adjustment direction of acceleration as candidate correction instructions; instructions with the speed adjustment direction of deceleration or speed adjustment value of 0 are not included in this deceleration process; thus, a set of candidate weighted instructions consistent with the current overspeed direction is obtained.

[0230] In addition, when locating candidate weighted instructions, the weighted speed adjustment value corresponding to the last candidate instruction in the target priority sequence is first read and compared with the current speed deviation value.

[0231] If the weighted speed adjustment value is greater than or equal to the speed deviation value, the instruction is identified as a partial decrease instruction, and the forward positioning is stopped. If the weighted speed adjustment value is less than the speed deviation value, the instruction is identified as a complete decrease instruction, and the weighted speed adjustment value is subtracted from the speed deviation value to obtain the remaining speed deviation. Then, the previous candidate weighted instruction is read and compared in the same way. The above process is repeated until the remaining speed deviation is eliminated or all candidate weighted instructions have completed positioning.

[0232] When using the weighted command after positioning to directionally decrease the initial belt speed adjustment command, for a completely decreasing command, its weighted speed adjustment value is directly reduced to 0; for a partially decreasing command, its original weighted speed adjustment value is subtracted from the current remaining speed deviation, and the resulting value is used as the corrected weighted speed adjustment value; acceleration commands that are not positioned and all deceleration commands remain unchanged.

[0233] This decreasing method ensures that the acceleration instructions at the back of the target priority sequence take priority in handling speed deviation corrections, while avoiding weakening the control role of the instructions at the front of the sequence.

[0234] After the reduction is completed, the modified weighted instructions are rearranged according to the original target priority sequence.

[0235] Specifically, for instructions that decrease to 0, their workstation number, execution cycle, and sequence position are retained, and only the speed adjustment value is recorded as 0; for instructions that decrease in execution, the original workstation number and execution cycle are retained, and the original value is replaced with the decreased speed adjustment value; the remaining instructions remain unchanged; thus forming the speed adjustment execution sequence.

[0236] When determining the final belt speed adjustment command based on the speed adjustment execution sequence, the actual running speed at the beginning of the current execution cycle is used as the reference speed. The corrected speed adjustment values ​​corresponding to each workstation are merged sequentially from front to back according to the speed adjustment execution sequence. The speed adjustment value in the acceleration direction is accumulated to the reference speed, and the speed adjustment value in the deceleration direction is deducted from the reference speed to obtain the corrected target execution speed.

[0237] In this embodiment, the corrected target execution speed is compared with the upper limit and lower limit of the allowable operating speed of the conveyor belt.

[0238] When the corrected target execution speed is greater than the upper limit of the allowed running speed, the upper limit of the allowed running speed shall be used; when the corrected target execution speed is less than the lower limit of the allowed running speed, the lower limit of the allowed running speed shall be used; when it is between the two limits, the corrected target execution speed shall be used directly.

[0239] Finally, the determined target execution speed, current execution cycle, and workstation execution order in the speed adjustment execution sequence are written into the final belt speed adjustment command and sent to the conveyor belt drive control unit. Thus, when the actual running speed of the initial belt speed adjustment command is higher than the expected running speed, a weighted command with positive speed increment can be generated according to the predetermined target priority sequence, and the corresponding speed deviation can be eliminated by the decrementing process from back to front to obtain the final belt speed adjustment command.

[0240] Example 11: An intelligent speed control system for a conveyor belt-type neuromorphic robot, comprising the steps of the intelligent speed control method for the conveyor belt-type neuromorphic robot, the system including: The periodic module acquires the initial perception instructions output by the heterogeneous brain-like unit, extracts the decision cycle of the initial perception instructions through the long short-term memory network, and obtains a periodic feature sequence containing timestamp information. The alignment module determines whether there is a temporal conflict in the periodic feature sequence. If there is a temporal conflict, the initial sensing instructions are timestamped to obtain a unified set of alignment instructions in the time dimension. The spatial module uses a graph convolutional network to process the alignment instruction set, extracts the spatial distribution features of the alignment instruction set when crossing workstations, and obtains a spatial feature vector representing the topological relationship of the workstations. The association module constructs a spatiotemporal dimension matrix through spatial feature vectors and periodic feature sequences. It tracks the propagation path of alignment instructions across workstations according to the topological relationship of workstations and the order of timestamps, and extracts the causal order in the propagation path to obtain spatiotemporal association features with sequential dependencies. The sorting module determines whether the alignment instruction set that crosses workstations will produce a chain reaction based on the spatiotemporal correlation characteristics. If a chain reaction occurs, the priority of the instructions that cross workstations is calculated to determine the target priority sequence. The weighting module assigns numerical weights to the alignment instruction set through the target priority sequence to obtain a weighted instruction matrix that eliminates control errors. The speed control module, based on the weighted instruction matrix, sequentially integrates the speed adjustment amounts corresponding to each workstation according to the target priority sequence, calculates the expected running speed of the conveyor belt, and obtains the initial belt speed control instruction for the global collaborative conveyor belt neuromorphic robot. The correction module obtains the actual running speed corresponding to the initial belt speed adjustment command. If the actual running speed is greater than the expected running speed, it locates the corresponding weighted command according to the target priority sequence and performs directional numerical decrement processing on the initial belt speed adjustment command to determine the final belt speed adjustment command of the conveyor belt brain-like robot.

[0241] In this embodiment, the intelligent speed regulation system is used to execute the method described in any one of S1 to S8. Each module establishes a data transmission relationship in the order of periodic processing, time alignment, spatial feature extraction, spatiotemporal correlation analysis, priority sorting, weight allocation, speed calculation, and feedback correction. Each module can be implemented by different program units in the same controller, or it can be deployed in the corresponding processing unit and the intermediate processing results can be transmitted through the internal data bus.

[0242] The periodic module receives the initial perception commands output by each heterogeneous brain-like unit.

[0243] Specifically, each initial perception instruction includes at least a workstation number, an instruction timestamp, a speed adjustment value, and an instruction status. The cycle module categorizes instructions according to workstation numbers and arranges consecutive instructions from the same workstation in ascending order of timestamp. It then encodes each instruction field to form an instruction parsing matrix. This instruction parsing matrix is ​​input into a Long Short-Term Memory (LSTM) network, which updates the hidden layer states chronologically and determines the decision cycle based on recurring instruction change nodes in consecutive hidden layer states. The cycle module further extracts the timestamps corresponding to each decision cycle and arranges them according to a unified time base, ultimately generating a cycle feature sequence containing timestamp information. This cycle feature sequence is then sent to the alignment and association modules.

[0244] The alignment module receives the periodic feature sequence and the initial sensing instruction, and determines whether there is a timing conflict based on the timestamp order and propagation time between adjacent workstations.

[0245] For two workstations with a direct transmission relationship, if the timestamp of the instruction corresponding to the subsequent workstation is earlier than the timestamp of the instruction corresponding to the preceding workstation, or if the time interval between the two timestamps exceeds the corresponding propagation time threshold, a timing conflict is determined. After detecting a timing conflict, the alignment module calculates the clock drift of the local time of each workstation relative to the unified control time based on the periodic feature sequence, and generates a synchronization reference sequence accordingly. Then, the original timestamp of the initial sensing instruction is mapped to the corresponding control time in the synchronization reference sequence, and the corrected speed adjustment value is filtered and aggregated according to the execution cycle of each workstation, thereby forming a time-dimensional unified alignment instruction set. This alignment instruction set is output to the spatial module, the weighting module, and the subsequent speed adjustment processing link.

[0246] Furthermore, after receiving the alignment instruction set, the space module constructs a workstation node matrix based on the workstation corresponding to each instruction and the direction of instruction flow.

[0247] Specifically, each workstation is treated as a node, and the node data includes at least the speed adjustment value, instruction status, execution cycle, and the position of the workstation in the current instruction flow path. The spatial module further calculates the adjacency weights between nodes with preceding and following flow relationships, and establishes directed connections based on the adjacency weights, thereby forming a topological connection graph. After inputting the topological connection graph into the graph convolutional network, the module reads the node features of each target workstation and the node features of the preceding workstation, and performs weighted aggregation based on the adjacency weights to obtain the spatial distribution features corresponding to each workstation. The spatial distribution features are then processed by dimensionality reduction and normalization to form a spatial feature vector, which is then output to the association module.

[0248] The association module receives spatial feature vectors and periodic feature sequences, and establishes a two-dimensional correspondence based on the workstation number and unified control time.

[0249] Specifically, each matrix position simultaneously writes the spatial features of the corresponding workstation and the periodic features at that control moment, thereby constructing a spatiotemporal dimension matrix. The association module extracts the unidirectional workstation topology relationship formed along the material conveying direction based on the spatiotemporal dimension matrix, and determines the propagation path of the alignment instruction in the order from front to back based on the timestamp. For each propagation path, a causal connection is established in the direction from the preceding workstation to the following workstation. If a closed loop that can return the visited node is detected in the causal connection, the connection edge to be broken is determined according to the time interval and adjacency weight of the connection edge, and the closed loop detection and edge breaking process are repeated until a causal order without closed loops is formed. The association module extracts the preceding node, following node, time interval, adjacency weight, and corresponding spatial and periodic features according to the updated causal order, forming a spatiotemporal association feature with sequential dependency.

[0250] Furthermore, after receiving the spatiotemporal correlation features, the sorting module extracts conflicting nodes with opposite speed adjustment directions and calculates the actual transmission time based on the timestamps of adjacent instructions.

[0251] When the actual transmission time exceeds the corresponding propagation time threshold, the corresponding connection is identified as a delayed connection, and consecutive delayed connections are arranged into a delayed transmission path along the causal order. If there is a directional conflict between the end node of the delayed transmission path and the starting conflict node, the corresponding alignment instruction is determined to form a chain reaction, and the instruction nodes constituting the chain relationship are extracted as a subset of associated instructions. The sorting module calculates the instruction priority value based on the time sequence position of each associated instruction, the number of workstations covered by the propagation, and the position in the delayed transmission path, and then arranges them in descending order of priority value. When the values ​​are the same, the order is determined according to the timestamp and the workstation delivery position, thereby obtaining the target priority sequence.

[0252] The weighting module receives the target priority sequence and alignment instruction set, and reads the execution cycle and resource status of each workstation.

[0253] A conflict detection matrix is ​​constructed based on the execution cycle, speed adjustment direction, drive resource occupancy status, and available drive capabilities. The target priority sequence is then mapped to the conflict detection matrix to form a priority mapping table. The weighting module assigns basic weights to each alignment instruction according to the mapped execution order. Then, the weights of subsequent instructions with resource conflicts are corrected, and all weights are normalized. The speed adjustment value of each alignment instruction is multiplied by its corresponding weight to form the initial weighted instruction.

[0254] The weighting module further calculates the expected execution speed of all initially weighted instructions in the current execution cycle and compares it with the expected running speed.

[0255] When the absolute value of the difference between the two exceeds the deviation threshold, the total compensation amount is determined according to the direction of the difference and allocated to the corresponding initial weighted instruction according to the priority mapping table; the compensation amount is arranged according to the workstation number and execution cycle to form an error compensation vector, and then the error compensation vector is used to correct each initial weighted speed adjustment value; the corrected weighted speed adjustment value is written into the corresponding matrix position according to the workstation number and execution cycle to form a weighted instruction matrix and sent to the speed control module.

[0256] Furthermore, after receiving the weighted instruction matrix and target priority sequence, the speed control module reads the inlet-side tension, outlet-side tension, mass of the material being carried, and resistance parameters of the corresponding conveying section for each station. It then calculates the tension difference and tension distribution value for each station, forming a load tension distribution characteristic. Based on the mass of each conveying section, normal load, and running resistance coefficient, it calculates the frictional resistance of each station, forming a frictional resistance matrix.

[0257] The speed control module determines the traction gap of each station based on the load tension distribution characteristics and friction resistance matrix, and calculates the slip compensation value in combination with the corresponding traction stiffness.

[0258] Furthermore, after correcting the speed adjustment value in the weighted instruction matrix based on the slip compensation value, the speed control module calculates the dynamic traction force corresponding to each workstation in sequence according to the target priority sequence, and merges it with the current basic traction force to obtain the global dynamic traction force.

[0259] Subsequently, the global dynamic traction force is compared with the traction force threshold. The traction force threshold is determined by the smaller of the traction force allowed to be output by the drive device and the anti-slip allowable traction force. When the global dynamic traction force exceeds the traction force threshold, the traction amount corresponding to the workstation that generates the positive traction force increment is reduced from back to front according to the target priority sequence until the global dynamic traction force is reduced to within the traction force threshold range, thereby obtaining the target dynamic traction force.

[0260] The speed control module calculates the net traction force based on the target dynamic traction force, and then calculates the speed change based on the total motion mass formed by the conveyor belt and the currently carried material. The speed change is then combined with the current operating speed to obtain the expected operating speed. The expected operating speed is further compared with the upper and lower limits of the allowable operating speed, and the result within the limit range is taken as the final expected operating speed. Based on the expected operating speed, a global coordinated initial belt speed control command is generated and sent to the correction module and the conveyor belt drive control unit.

[0261] In addition, after the initial belt speed control command is executed, the correction module obtains the actual running speed within the corresponding execution cycle and reads the expected running speed determined by the speed control module.

[0262] If the actual operating speed is less than or equal to the expected operating speed, the initial belt speed control command is maintained; if the actual operating speed is greater than the expected operating speed, the expected operating speed is subtracted from the actual operating speed to obtain the speed deviation value.

[0263] Furthermore, based on the speed deviation value, the correction module calls the target priority sequence and the weighted instruction matrix, and searches for the weighted instruction in the acceleration direction from back to front according to the target priority sequence; when the current weighted speed adjustment value is less than the remaining speed deviation, the weighted speed adjustment value is reduced to 0, and the reduction amount is subtracted from the remaining speed deviation, and the previous acceleration instruction is processed; when the current weighted speed adjustment value is greater than or equal to the remaining speed deviation, the remaining speed deviation is subtracted from the weighted speed adjustment value, and the current reduction ends; thus, the corrected weighted instructions are obtained, and a speed adjustment execution sequence is formed according to the original target priority sequence.

[0264] The correction module then merges the corrected speed adjustment values ​​sequentially according to the speed adjustment execution sequence to obtain the corrected target execution speed, and compares it with the upper limit and lower limit of the allowed running speed.

[0265] The final target execution speed, current execution cycle, and workstation execution sequence are written together into the final belt speed control command and sent to the conveyor belt drive control unit for execution, thereby completing the closed-loop speed control process from the input of the heterogeneous brain-like unit sensing command to the output of the final belt speed control command.

[0266] Example 12: A computer-readable storage medium storing computer instructions thereon, which, when executed, perform the intelligent speed control method of the conveyor belt brain-like robot.

[0267] In this embodiment, a computer-readable storage medium is used to store computer instructions for implementing the intelligent speed regulation method described in any one of S1 to S8.

[0268] The computer-readable storage medium is a non-volatile storage medium capable of storing digital instructions and being read by the processor. The computer instructions are stored in the execution order of data acquisition, periodic feature extraction, timestamp alignment, spatial feature extraction, spatiotemporal correlation processing, priority determination, instruction weighting, speed calculation and feedback correction, and are respectively configured to receive and output the data corresponding to each processing step.

[0269] When the computer instructions are executed, the processor first reads the initial perception instruction from the data interface corresponding to the heterogeneous brain-like unit, and writes the workstation number, timestamp, speed adjustment value and instruction status into the running storage area; the processor calls the periodic processing instruction in the storage medium to encode the initial perception instruction, and inputs the encoding result into the pre-loaded long short-term memory network, extracts the decision cycle and timestamp information according to the processing process described in Example 3, and forms a periodic feature sequence.

[0270] Furthermore, the processor continues to execute time alignment instructions, and determines whether there are timing conflicts between instructions at different workstations based on the cycle feature sequence. If there are timing conflicts, clock drift features are extracted and a synchronization reference sequence is established. The timestamps of the initial sensing instructions are corrected to the unified control time, and then filtering and execution cycle aggregation are completed to obtain the alignment instruction set.

[0271] The aligned instruction data is stored in the runtime memory area and used as input for subsequent space processing instructions.

[0272] When executing spatial processing instructions, the processor establishes a workstation node matrix and the adjacency relationships between nodes based on the alignment instruction set, and calls the stored graph convolutional network parameters to complete node feature aggregation, dimensionality reduction, and normalization processing to obtain spatial feature vectors. Subsequently, the spatial feature vectors and periodic feature sequences are correlated according to the workstation number and control time to construct a spatiotemporal dimension matrix, and the instruction propagation path is tracked according to the workstation topology direction and timestamp order. The processor performs causal order extraction and loop closure detection on the propagation path. When causal loops exist, the loops are eliminated according to the edge-breaking method described in Example 6 to obtain spatiotemporal correlation features.

[0273] The processor continues to execute priority processing instructions based on the spatiotemporal correlation characteristics, extracts conflict nodes and delay propagation paths, and determines whether the two form a correlation closed loop used to characterize the chain reaction.

[0274] When a chain reaction occurs, the corresponding subset of related instructions is extracted, and the priority value of each related instruction is calculated according to the time sequence value, propagation coverage value and path position value determined in Example 7. The order of arrangement is determined according to the priority value, timestamp and workstation position to form a target priority sequence.

[0275] Furthermore, after obtaining the target priority sequence, the processor reads the execution cycle and resource status of each workstation, constructs a conflict detection matrix, and maps the target priority sequence to the conflict detection matrix; according to the mapping result, weights are assigned to the alignment instructions, and the speed adjustment value of each instruction and the corresponding weight are numerically processed to obtain the initial weighted value; when the absolute value of the difference between the expected execution speed and the expected running speed exceeds the deviation threshold determined in Example 8, the processor generates an error compensation vector according to the speed difference, and uses the error compensation vector to correct the corresponding initial weighted instruction to form a weighted instruction matrix.

[0276] When the processor executes the speed calculation instruction, it reads the load tension data corresponding to each station according to the weighted instruction matrix, and calculates the slip compensation value in combination with the friction resistance data. Then, it corrects the speed adjustment amount of each station with the slip compensation value. The speed adjustment amount of each station is converted and merged into the global dynamic traction force according to the target priority sequence. When the global dynamic traction force exceeds the traction force threshold determined in Example 9, the excess part is subjected to amplitude limiting processing to obtain the target dynamic traction force. The speed change is calculated based on the target dynamic traction force, friction resistance, motion mass and execution cycle. The speed change is then merged with the current running speed to obtain the expected running speed, and the initial belt speed adjustment instruction is generated accordingly.

[0277] Furthermore, after the initial belt speed adjustment instruction is executed, the processor reads the actual running speed of the corresponding execution cycle through the speed detection interface and compares the actual running speed with the expected running speed.

[0278] Specifically, when the actual operating speed is greater than the expected operating speed, the speed deviation value is obtained by subtracting the expected operating speed from the actual operating speed. Then, the weighted instructions for the acceleration direction are located from back to front according to the target priority sequence. The processor allocates the speed deviation value to the located weighted instructions in the decreasing order described in Example 10, performs a full or partial decrease on the corresponding speed adjustment value, forms a speed adjustment execution sequence, and obtains the final belt speed adjustment instruction based on the speed adjustment execution sequence.

[0279] The final belt speed control command is sent by the processor to the conveyor belt drive control unit through the control interface, which includes at least the target execution speed, the corresponding execution cycle, and the workstation execution sequence.

[0280] Therefore, after the computer instructions stored in the computer-readable storage medium are loaded and executed by the processor, the processing from the initial sensing instruction acquisition to the final belt speed adjustment instruction output can be completed in accordance with the data processing order defined by S1 to S8.

[0281] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. An intelligent speed regulation method for a brain-like robot operating on a conveyor belt, characterized in that, include: S1. Obtain the initial perception command output by the heterogeneous brain-like unit, extract the decision cycle of the initial perception command, and obtain the cycle feature sequence. S2. If there is a time sequence conflict in the periodic feature sequence, the initial sensing command is timestamped to obtain an alignment command set. S3. Extract the spatial distribution features of the alignment instruction set when crossing workstations to obtain a spatial feature vector representing the topological relationship of the workstations. S4. Construct a spatiotemporal dimension matrix by using spatial feature vectors and periodic feature sequences, track the propagation path of alignment instructions across workstations, and extract the causal order in the propagation path to obtain spatiotemporal correlation features. S5. Based on the spatiotemporal correlation characteristics, if the alignment instruction set that crosses workstations produces a chain reaction, then calculate the instruction priority that crosses workstations and determine the target priority sequence. S6. The alignment instruction set is numerically weighted according to the target priority sequence to obtain a weighted instruction matrix that eliminates control errors. S7. Calculate the expected running speed of the conveyor belt based on the weighted instruction matrix to obtain the initial belt speed adjustment instruction for the conveyor belt brain-like robot. S8. If the actual running speed corresponding to the initial belt speed adjustment command is greater than the expected running speed, then the weighted command corresponding to the target priority sequence is located to determine the final belt speed adjustment command of the conveyor belt brain-like robot.

2. The intelligent speed regulation method for a brain-like robot using a conveyor belt according to claim 1, characterized in that: S1 includes: The initial perception commands output by the heterogeneous brain-like units are obtained, and the initial perception commands are encoded to obtain the command parsing matrix; The temporal dependencies of the instruction parsing matrix are extracted using a long short-term memory network to generate hidden layer state vectors; The decision cycle is determined based on the hidden layer state vector, the timestamp information of the decision cycle is extracted, and a timestamp aligned sequence is generated. The hidden layer state vector is fused using timestamp-aligned sequences to obtain a periodic feature sequence containing timestamp information.

3. The intelligent speed regulation method for a brain-like robot using a conveyor belt according to claim 1, characterized in that: S2 includes: Obtain the periodic feature sequence and determine whether there is a temporal conflict in the periodic feature sequence; If timing conflicts exist, extract the clock drift features of the periodic feature sequence; A synchronization reference sequence is constructed based on clock drift characteristics. The initial sensing command is timestamped using the synchronization reference sequence to obtain the first sensing command. The first sensing command is filtered to obtain the second sensing command; Aggregation processing is performed according to the execution cycle of the second perception instruction to obtain a unified set of alignment instructions in the time dimension.

4. The intelligent speed regulation method for a brain-like robot using a conveyor belt according to claim 1, characterized in that: S3 includes: Obtain the alignment instruction set and construct the workstation node matrix based on the instruction flow path of the alignment instruction set; Calculate the adjacency weights based on the workstation node matrix and update the workstation node matrix to obtain the topology connection graph; A graph convolutional network is used to process the topology connectivity graph and extract its spatial distribution features. The spatial distribution features are reduced in dimensionality to obtain feature aggregation vectors. A spatial mapping matrix is ​​constructed using the feature aggregation vectors and then normalized to obtain spatial feature vectors representing the topological relationships of workstations.

5. The intelligent speed regulation method for a brain-like robot on a conveyor belt according to claim 1, characterized in that: S4 includes: Construct a spatiotemporal dimension matrix by obtaining spatial feature vectors and periodic feature sequences; Extract the unidirectional workstation topology relationship based on the spatiotemporal dimension matrix, and determine the instruction propagation path according to the unidirectional workstation topology relationship and the timestamp sequence; Determine if a closed loop exists in the initial causal order of the alignment instruction propagation path; If a closed loop exists, disconnect the connecting edge to obtain the updated causal order; Spatiotemporal correlation features with sequential dependencies are obtained by updating the causal order.

6. The intelligent speed regulation method for a brain-like robot on a conveyor belt according to claim 1, characterized in that: S5 includes: Extract conflict nodes and delayed propagation paths based on spatiotemporal correlation features; If the conflicting node and the delayed propagation path form a closed loop, it is determined that the alignment instruction set produces a chain reaction, resulting in a subset of associated instructions; Calculate instruction priority values ​​based on the associated instruction subset; The target priority sequence is determined by sorting the instructions according to their priority values.

7. The intelligent speed regulation method for a brain-like robot using a conveyor belt according to claim 1, characterized in that: S6 includes: Construct a conflict detection matrix based on the execution cycle and resource status of the alignment instruction set; A priority mapping table is obtained by mapping the target priority sequence through the collision detection matrix. Weights are then assigned to the alignment instruction set according to the priority mapping table to obtain the initial weighted instructions. If the initial weighted instruction triggers the deviation threshold, then the error compensation vector of the initial weighted instruction is generated; By correcting the initial weighted instructions using the error compensation vector, a weighted instruction matrix that eliminates control errors is obtained.

8. The intelligent speed regulation method for a brain-like robot on a conveyor belt according to claim 1, characterized in that: S7 includes: The load tension distribution characteristics are extracted based on the weighted instruction matrix and the target priority sequence, and the slip compensation value corresponding to the load tension distribution characteristics is calculated by combining the friction resistance matrix. The speed adjustment amount is determined based on the slip compensation value, and the speed adjustment amounts corresponding to each workstation are sequentially merged according to the target priority sequence to obtain the global dynamic traction force. If the global dynamic traction force exceeds the traction force threshold, the global dynamic traction force is limited to obtain the target dynamic traction force. The expected operating speed is calculated by the target dynamic traction force, and the initial belt speed adjustment command for the global collaborative brain-like robot is obtained based on the expected operating speed.

9. The intelligent speed regulation method for a brain-like robot using a conveyor belt according to claim 1, characterized in that: S8 includes: Obtain the actual and expected operating speeds corresponding to the initial belt speed control command; If the actual operating speed is greater than the expected operating speed, the speed deviation value is calculated. Match the target priority sequence based on the speed deviation value and locate the weighted command; The initial belt speed control command is processed by weighted instructions to perform directional numerical decrementing to obtain the speed control execution sequence; The final belt speed control command for the conveyor belt neuromorphic robot is determined based on the speed control execution sequence.

10. An intelligent speed control system for a conveyor belt-type neuromorphic robot, used to implement the steps of the intelligent speed control method for the conveyor belt-type neuromorphic robot according to any one of claims 1-9, characterized in that, The system includes: The periodic module acquires the initial perception instructions output by the heterogeneous brain-like unit, extracts the decision cycle of the initial perception instructions through the long short-term memory network, and obtains a periodic feature sequence containing timestamp information. The alignment module determines whether there is a temporal conflict in the periodic feature sequence. If there is a temporal conflict, the initial sensing instructions are timestamped to obtain a unified set of alignment instructions in the time dimension. The spatial module uses a graph convolutional network to process the alignment instruction set, extracts the spatial distribution features of the alignment instruction set when crossing workstations, and obtains a spatial feature vector representing the topological relationship of the workstations. The association module constructs a spatiotemporal dimension matrix through spatial feature vectors and periodic feature sequences. It tracks the propagation path of alignment instructions across workstations according to the topological relationship of workstations and the order of timestamps, and extracts the causal order in the propagation path to obtain spatiotemporal association features with sequential dependencies. The sorting module determines whether the alignment instruction set that crosses workstations will produce a chain reaction based on the spatiotemporal correlation characteristics. If a chain reaction occurs, the priority of the instructions that cross workstations is calculated to determine the target priority sequence. The weighting module assigns numerical weights to the alignment instruction set through the target priority sequence to obtain a weighted instruction matrix that eliminates control errors. The speed control module, based on the weighted instruction matrix, sequentially integrates the speed adjustment amounts corresponding to each workstation according to the target priority sequence, calculates the expected running speed of the conveyor belt, and obtains the initial belt speed control instruction for the global collaborative conveyor belt neuromorphic robot. The correction module obtains the actual running speed corresponding to the initial belt speed adjustment command. If the actual running speed is greater than the expected running speed, it locates the corresponding weighted command according to the target priority sequence and performs directional numerical decrement processing on the initial belt speed adjustment command to determine the final belt speed adjustment command of the conveyor belt brain-like robot.

11. A computer-readable storage medium, characterized in that, The computer instructions are stored thereon, and when the computer instructions are executed, the intelligent speed regulation method of the conveyor belt brain-like robot according to any one of claims 1-9 is performed.

Images