WO2021195970A1 - 工业系统的预测模型学习方法、装置和系统 - Google Patents

工业系统的预测模型学习方法、装置和系统 Download PDF

Info

Publication number
WO2021195970A1
WO2021195970A1 PCT/CN2020/082459 CN2020082459W WO2021195970A1 WO 2021195970 A1 WO2021195970 A1 WO 2021195970A1 CN 2020082459 W CN2020082459 W CN 2020082459W WO 2021195970 A1 WO2021195970 A1 WO 2021195970A1
Authority
WO
WIPO (PCT)
Prior art keywords
simulation
parameter
industrial system
parameter group
value
Prior art date
Application number
PCT/CN2020/082459
Other languages
English (en)
French (fr)
Inventor
李婧
徐蔚峰
李明
Original Assignee
西门子股份公司
西门子（中国）有限公司
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by 西门子股份公司, 西门子（中国）有限公司 filed Critical 西门子股份公司
Priority to EP20928458.7A priority Critical patent/EP4122654A4/en
Priority to PCT/CN2020/082459 priority patent/WO2021195970A1/zh
Priority to CN202080096949.XA priority patent/CN115135463A/zh
Priority to US17/916,234 priority patent/US20230153640A1/en
Publication of WO2021195970A1 publication Critical patent/WO2021195970A1/zh

Links

Images

Classifications

    • G—PHYSICS
    • G05—CONTROLLING; REGULATING
    • G05B—CONTROL OR REGULATING SYSTEMS IN GENERAL; FUNCTIONAL ELEMENTS OF SUCH SYSTEMS; MONITORING OR TESTING ARRANGEMENTS FOR SUCH SYSTEMS OR ELEMENTS
    • G05B19/00—Programme-control systems
    • G05B19/02—Programme-control systems electric
    • G05B19/418—Total factory control, i.e. centrally controlling a plurality of machines, e.g. direct or distributed numerical control [DNC], flexible manufacturing systems [FMS], integrated manufacturing systems [IMS] or computer integrated manufacturing [CIM]
    • G05B19/41885—Total factory control, i.e. centrally controlling a plurality of machines, e.g. direct or distributed numerical control [DNC], flexible manufacturing systems [FMS], integrated manufacturing systems [IMS] or computer integrated manufacturing [CIM] characterised by modeling, simulation of the manufacturing system
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N5/00—Computing arrangements using knowledge-based models
    • G06N5/02—Knowledge representation; Symbolic representation
    • G06N5/022—Knowledge engineering; Knowledge acquisition
    • B—PERFORMING OPERATIONS; TRANSPORTING
    • B25—HAND TOOLS; PORTABLE POWER-DRIVEN TOOLS; MANIPULATORS
    • B25J—MANIPULATORS; CHAMBERS PROVIDED WITH MANIPULATION DEVICES
    • B25J9/00—Programme-controlled manipulators
    • B25J9/16—Programme controls
    • B25J9/1656—Programme controls characterised by programming, planning systems for manipulators
    • B25J9/1671—Programme controls characterised by programming, planning systems for manipulators characterised by simulation, either to verify existing program or to create and verify new program, CAD/CAM oriented, graphic oriented programming systems
    • G—PHYSICS
    • G05—CONTROLLING; REGULATING
    • G05B—CONTROL OR REGULATING SYSTEMS IN GENERAL; FUNCTIONAL ELEMENTS OF SUCH SYSTEMS; MONITORING OR TESTING ARRANGEMENTS FOR SUCH SYSTEMS OR ELEMENTS
    • G05B2219/00—Program-control systems
    • G05B2219/30—Nc systems
    • G05B2219/31—From computer integrated manufacturing till monitoring
    • G05B2219/31342—Design of process control system

Definitions

  • the invention relates to the field of industrial automation, and in particular to a method, device and system for learning a predictive model of an industrial system.
  • predicting the performance of complex systems is not easy.
  • One method is to evaluate system performance through simulation. Generate multiple scenarios and perform tests in simulation to generate a large number of simulation results for performing evaluation. However, it takes a lot of simulation execution time to explore the design space of a target system (such as parameter tuning), and additional man-hours are required to establish and execute simulations in a scenario. Therefore, a typical simulation platform is easy to test and evaluate complex systems, but it is not suitable for online prediction of system performance.
  • one of the solutions in the prior art is to use the learning method thereof to establish training data for refreshing machinery, select a prediction model template from the database, and optimize the prediction model based on the refreshed data debugging data.
  • this solution relies on collecting real data without data analysis, not through a simulation mechanism.
  • it needs a library of predictive templates, and then refines the templates through additional training.
  • the forecast goal of this scheme is business-related, especially sales.
  • the prior art also proposes a scheme, which is a control parameter debugging mechanism, which is based on simulation control parameter optimization. Model errors are resolved as a problem, and control parameter adjustments with higher accuracy can be identified.
  • this solution is not aimed at complex systems, nor is it based on simulation.
  • this scheme is limited to servo motors and parameter debugging controllers.
  • Another solution in the prior art is a neural network robot trained in a simulation environment, which collects comprehensive demo statistics (bin) from an algorithm supervisor in a pybullet simulator, and then it learns neural network through the comprehensive data obtained from the simulator
  • the network policy also trains the neural network robot controller based on the trajectory data obtained from the simulator.
  • this solution is useful for intelligent robot systems and collected simulation data, but its focus is on training robot controllers, not for complex systems and simulators to enhance the simulator to obtain a predictive model generator.
  • the first aspect of the present invention provides a predictive model learning method of an industrial system, wherein the industrial system executes simulation according to simulation tasks on a simulation platform, and the predictive model learning method of the industrial system includes the following steps: S1, using statistical indicators Quantify the simulation task of the industrial system to extract simulation task features; S2, extract a parameter group from the system module of the industrial system that executes the simulation and adjust the value of the parameter group, triggering the simulation platform to respond to the industrial system Perform simulation based on multiple parameter groups with different values; S3, record performance indicators according to the values of the parameter groups after adjusting the parameters; S4, based on the simulation task characteristics and corresponding parameter groups and performance indicators of different values through machine learning algorithms Train the data and generate a predictive model.
  • step S2 includes the following steps: S21, extracting a parameter group from the system module of the industrial system that performs the simulation, and marking the adjustable parameter group; S22, adjusting according to the adjustment range corresponding to each adjustable parameter group The value of the adjustable parameter group triggers the simulation platform to perform simulation on the multiple parameter groups of the industrial system based on different values; S23, synchronize and record the parameter group combination and value after adjusting the parameters.
  • step S3 includes the following steps: S31, recording the performance index according to the value of the parameter group after adjusting the parameter; S32, recording the value of the performance index based on the corresponding recording performance index and the simulation result.
  • the performance index is a key performance index or determined based on customer needs.
  • step S4 based on the input simulation task characteristics and the numerical value of the parameter group, query the prediction model to obtain the numerical value of the performance index.
  • the industrial system is a context-aware robot, wherein the simulation task feature includes multiple indicators of the depth color image, and the indicators include any one or more of the following: image resolution; target ratio in the image; The average number of objects.
  • the context-aware robot is a grasping robot, and its system modules include: a vision module, which performs image processing on the input depth color image; a path planning module, which calculates the motion of the robot arm through the grasping points obtained by image recognition The path to the grab point; the action control module, which controls the movement of the robotic arm according to the path planned by the path planning module.
  • the second aspect of the present invention provides a predictive model learning device of an industrial system, wherein the industrial system executes simulation according to simulation tasks on a simulation platform, and the predictive model learning device of the industrial system: a task feature management device, which uses statistics The index quantifies the simulation task of the industrial system to extract simulation task characteristics; a parameter adjustment management device that extracts a parameter group from the system module of the industrial system that performs the simulation and adjusts the value of the parameter group to trigger the simulation platform Perform simulation on multiple parameter groups of the industrial system based on different values; performance recording device, which records performance indicators according to the values of the parameter groups after adjusting the parameters; training data management device, which is based on simulation task characteristics and corresponding differences Numerical parameter groups and their performance indicators train data through machine learning algorithms; learning management devices, which generate predictive models.
  • a task feature management device which uses statistics The index quantifies the simulation task of the industrial system to extract simulation task characteristics
  • a parameter adjustment management device that extracts a parameter group from the system module of the industrial system that performs the simulation and adjusts the
  • the parameter adjustment management device further includes: a parameter feature group recording device, which extracts parameter groups from the system modules of the industrial system that executes the simulation, and marks the adjustable parameter groups; and a parameter adjustment device, which is based on each The adjustment range corresponding to the adjustable parameter group adjusts the value of the adjustable parameter group, triggering the simulation platform to perform simulation on multiple parameter groups of the industrial system based on different values; a parameter recording device that synchronizes and records the adjusted parameters The combination and value of the parameter group.
  • a parameter feature group recording device which extracts parameter groups from the system modules of the industrial system that executes the simulation, and marks the adjustable parameter groups
  • a parameter adjustment device which is based on each The adjustment range corresponding to the adjustable parameter group adjusts the value of the adjustable parameter group, triggering the simulation platform to perform simulation on multiple parameter groups of the industrial system based on different values
  • a parameter recording device that synchronizes and records the adjusted parameters The combination and value of the parameter group.
  • the performance recording device further includes: a performance registration device that records performance indicators according to the values of the parameter group after adjusting the parameters; a performance recording sub-device that records the performance indicators and simulation results corresponding to each other The value of the performance index.
  • the performance index is a key performance index or determined based on customer needs.
  • the learning management device is also used to query the prediction model to obtain the value of the performance index based on the input simulation task feature and the value of the parameter group.
  • the third aspect of the present invention provides a predictive model learning system for an industrial system, which includes: a processor; and a memory coupled to the processor, the memory having instructions stored therein, and the instructions are executed by the processor.
  • the action includes: S1, using statistical indicators to quantify the simulation task of the industrial system to extract simulation task features; S2, extracting a parameter set from the system module of the industrial system that performs the simulation And adjust the value of the parameter group, trigger the simulation platform to perform simulation on the multiple parameter groups of the industrial system based on different values; S3, record performance indicators according to the value of the parameter group after adjusting the parameters; S4, based on simulation
  • S1 using statistical indicators to quantify the simulation task of the industrial system to extract simulation task features
  • S2 extracting a parameter set from the system module of the industrial system that performs the simulation And adjust the value of the parameter group, trigger the simulation platform to perform simulation on the multiple parameter groups of the industrial system based on different values
  • S4 based
  • the action S2 further includes: S21, extracting a parameter group from the system module of the industrial system that executes the simulation, and marking the adjustable parameter group; S22, adjusting the station according to the adjustment range corresponding to each adjustable parameter group.
  • the value of the adjustable parameter group triggers the simulation platform to perform simulation on the multiple parameter groups of the industrial system based on different values; S23, synchronize and record the parameter group combination and value after adjusting the parameters.
  • the action S3 further includes: S31, recording the performance index according to the value of the parameter group after adjusting the parameter; S32, recording the value of the performance index based on the corresponding recording performance index and the simulation result.
  • the method further includes: querying the prediction model to obtain the value of the performance index based on the input simulation task feature and the value of the parameter group.
  • the fourth aspect of the present invention provides a computer program product, which is tangibly stored on a computer-readable medium and includes computer-executable instructions that, when executed, cause at least one processor to execute The method described in the first aspect of the present invention.
  • the fifth aspect of the present invention provides a computer-readable medium on which computer-executable instructions are stored, and when executed, the computer-executable instructions cause at least one processor to perform the method according to the first aspect of the present invention.
  • the present invention considers the data in the simulation task to train and construct the prediction model.
  • the step of adjusting parameters of the present invention adopts a systematic and automatic way, allowing simulation to generate a large number of and various parameters for data training, and this part of the prior art is often done manually. Therefore, these system parameters can be explored and tested in a structured manner to avoid incomplete consideration of manual methods.
  • simulation data can be converted into online prediction models, which can guide users to dynamically and more efficiently select parameter combinations and values for a complex system.
  • the present invention can simultaneously help customers build an online prediction system.
  • the present invention allows industrial customers to predict their system performance online and dynamically adjust parameters.
  • Fig. 1 is a schematic structural diagram of a predictive model learning system of an industrial system according to a specific embodiment of the present invention
  • Figure 2 is a schematic diagram of a scene where the grabbing robot grabs parts from the box;
  • FIG. 3 is a schematic structural diagram of a parameter adjustment management device of a predictive model learning system of an industrial system according to a specific embodiment of the present invention
  • FIG. 4 is a logic diagram of parameter adjustment of a predictive model learning mechanism of an industrial system according to a specific embodiment of the present invention.
  • FIG. 5 is a schematic structural diagram of a performance recording device of a predictive model learning system of an industrial system according to a specific embodiment of the present invention
  • Fig. 6 is a schematic diagram of a decision tree C5.0 of a predictive model learning mechanism of an industrial system according to a specific embodiment of the present invention.
  • the present invention is an extension of the existing simulation platform, which can automatically retrieve the parameter group of the complex system that needs to be simulated and record the main simulation data to learn a latent prediction model for the target industrial system.
  • the present invention records the data of interest for each system module, constructs a feature group of scenes and adjusted parameters, and finally uses device sampling to generate a predictive model for the simulated system.
  • this prediction model users can quickly predict the performance of a complex system based on specific scenarios and parameters, and then optimize the system by automatically selecting system parameters, where the automatic selection uses the predictive model to meet performance requirements.
  • the present invention can simultaneously help customers build an online prediction system for their complex system.
  • the target system includes an intelligent robot system, a manufacturing unit or a production line.
  • the present invention can allow customers to simultaneously predict their system performance online and automatically adjust parameters.
  • the present invention can be applied to many neighbors, especially complex industrial systems, including industrial context-aware manufacturing systems, such as autonomous equipment, workstations, and workshops.
  • the present invention is suitable for context sensing devices, such as autonomous driving, intelligent robots, and unmanned aerial vehicles.
  • Fig. 1 is a schematic structural diagram of a predictive model learning system of an industrial system according to a specific embodiment of the present invention.
  • the predictive model learning system 100 of the industrial system includes a task feature management device 110, a parameter adjustment management device 120, a performance recording device 130, a training data management device 140, and a learning management device 150.
  • the predictive model learning system 100 of the industrial system provided by the present invention is a function extension of the existing simulation platform 200.
  • the input of the simulation platform 200 is the simulation task, and the output is the simulation result.
  • the first aspect of the present invention provides a predictive model learning method of an industrial system, which includes the following steps:
  • step S1 is executed, and the task feature management device 110 uses statistical indicators to quantify the simulation tasks of the industrial system to extract simulation task features.
  • the task feature management device 110 is used to manage and load the features of the simulation task, that is, use statistical metrics to quantify the features of the input simulation task, and provide the simulation task feature data to the training data after the simulation is executed. ⁇ 140 ⁇ Management device 140.
  • the industrial system is a process in which a grabbing robot grabs multiple components from the box B'.
  • the simulation platform 200 executes the simulation of the aforementioned process of the grasping robot B, wherein the simulation platform 200 includes a plurality of system modules of the grasping robot B, including a vision module 210, a path planning module 220, and actions. Control module 230.
  • the input of the simulation platform 200 is an image, particularly a deep color image (RGBD image).
  • the vision module 210 performs image processing based on the input depth color image
  • the path planning module 220 calculates the path from the robot arm to the grab point through the grab points obtained by image recognition
  • the action control module 230 controls the path according to the path planned by the path planning module 220 Movement of the robotic arm.
  • the input of the task feature management device 110 is image data
  • the simulation task calculated by it is the statistical index group of the image data.
  • These indicators and features can be pre-defined in the task feature management device 110 and extracted features that need to be simulated for the input simulation task.
  • the grasping robot B is a context-aware robot
  • the simulation task feature is a series of multiple indicators of the depth color image, one of which is the feature vector I:
  • I [image resolution, target proportion in the image, average number of objects]
  • the above-mentioned characteristic values can be calculated from task data based on each specific simulation execution and specified by the simulation platform.
  • the task feature data value is synchronized with the index by the parameter adjustment management device 120 and stored as data, and later as a part of the training quantity set.
  • step S2 is executed.
  • the parameter adjustment management device 120 extracts a parameter group from the system module of the industrial system that executes the simulation and adjusts the value of the parameter group, triggering the simulation platform 200 to perform the multiplication of the industrial system based on different values.
  • the simulation is performed for each parameter group.
  • the step S2 includes three sub-steps S21, S22 and S23.
  • the parameter adjustment management device 120 is composed of three secondary devices, which are a parameter feature group recording device 121, a parameter adjustment device 122, and a parameter recording device 123, respectively.
  • the parameter adjustment management device 120 is also used to adjust to automatically record parameters for multiple system modules, which can systematically generate simulation data for the learning management device 150.
  • the feature group recording device 121 extracts the parameter group from the system module of the industrial system that executes the simulation, and marks the adjustable parameter group.
  • the feature group recording device 121 sends the adjustable parameter group to the parameter adjustment device 122.
  • the parameter feature group recording device 121 collects the adjustable parameter group from each system module, that is, collects the adjustable parameter group from the vision module 210, the path planning module 220, and the action control module 230.
  • the above-mentioned parameters can be Marked as "adjustable", these adjustable parameters and their adjustable ranges are then notified to the parameter adjustment device 122 in the parameter adjustment management device 120 for subsequent analysis.
  • the output of the parameter characteristic group recording device 121 is the parameter characteristic value.
  • the parameter feature group recording device 121 marking the adjustable parameter group from the input parameter group includes:
  • perception is an adjustable parameter group extracted from the vision module 210
  • Path Planner is an adjustable parameter group extracted from the path planning module 220
  • Motion Controller is an adjustable parameter group extracted from the motion control module 230.
  • the adjustable parameter group perception extracted from the vision module 210 includes x 1 , x 2 , x 3
  • the adjustable parameter groups x 1 , x 2 , and x 3 also have a value range defined as x 1min to x 1max , x 2min to x 2max , x 3min to x 3max .
  • the parameter feature group recording device 121 uses the LINEMOD algorithm
  • x 1 is the similarity of templates
  • x 2 is non-maximum suppression
  • x 3 is contrast ratio.
  • the adjustable parameter group Path Planner extracted from the path planning module 220 includes y 1 and y 2 , and y 1 and y 2 have discrete values defined as y 11 , y 12 , y 13 and y 21 , y 22 , and y 23, respectively .
  • y 1 and y 2 are routing parameters, including starting point and path point.
  • the adjustable parameter group Motion Controller extracted from the motion control module 230 includes z 1 and z 2 , which respectively represent the endpoint velocity and acceleration of the grasping robot B, and z 1 and z 2 can be continuous Variable or discrete value.
  • the parameter adjustment device 122 adjusts the value of the parameter group according to the adjustment range corresponding to each adjustable parameter group, and triggers the simulation platform 200 to perform simulation on multiple parameter groups of the industrial system based on different values. .
  • the parameter adjustment device 122 sends the value of the parameter group to the parameter recording device 123.
  • the parameter adjustment device 122 is used to systematically and automatically select parameter characteristic values from the parameter characteristic group, then adjust the parameters and push the adjusted parameters to the simulation platform 200, and start the simulation of the target industrial system to select these characteristics Perform the corresponding test in the simulation.
  • Many strategies are applied to perform parameter adjustments based on application requirements and customer needs, one of which is to use all possible combinations of values in the feature group, and the other is to select variable values based on a specific distribution, such as a normal distribution.
  • Fig. 4 is a logic diagram of parameter adjustment of a predictive model learning mechanism of an industrial system according to a specific embodiment of the present invention.
  • the parameter adjustment logic is basically to adjust the parameter x first, then adjust the parameter y, and finally adjust the parameter z.
  • the initial state of the above-mentioned parameters includes:
  • x 1 x 1min
  • x 2 x 2min
  • x 3 x 3min
  • the adjusted parameters x, y, and z are sent to the vision module 210, the path planning module 220, and the action control module 230, respectively.
  • the parameter recording device 123 synchronizes and records the parameter group combination and value after adjusting the parameter.
  • step S3 is executed, and the performance recording device 130 records the performance index according to the value of the parameter group after adjusting the parameter, where the performance index is a key performance index (KPI).
  • KPI key performance index
  • the performance recording device 130 and the parameter adjustment management device 120 cooperate to record the key performance index (KPI) of the industrial system, which is consistent with each value of the parameter feature group.
  • the performance recording device 130 further includes two sub-modules: a performance registration device 131 and a performance recording sub-device 132.
  • the step S3 includes sub-step S31 and sub-step S32.
  • the performance recording device 130 further includes two sub-modules: a performance registration device 131 and a performance recording sub-device 132.
  • the performance registration device 131 cooperates to record the performance index according to the value of the parameter group after adjusting the parameter.
  • the performance registration device 131 cooperates with the parameter adjustment management device 120 to record the key performance index KPI through each value of the performance characteristic group.
  • the performance registration device 131 collects performance metrics (performance metrics) of the industrial system that is the simulation target, and the performance metrics can be marked in the simulation platform 200.
  • performance indicators are related to KPIs that are of interest to customers. Multiple metrics can be recorded in the performance recording device 130.
  • the performance metric of the grabbing robot B can optionally be "whether to successfully complete the removal of a part (for example, the first part p1, the second part p2, and the third part p3) from the box
  • Another exemplary performance index can optionally be time t, which represents the time to complete the entire grasping process of the grasping robot B.
  • the performance recording sub-device 132 records the value of the performance index based on the recording performance index and the simulation result.
  • the performance index value is related to the index of the parameter adjustment management device 120, and it needs to complete each round of data recording together with the parameter adjustment management device 120.
  • the combination and value of each parameter group and its simulation results correspond one-to-one, and each simulation result corresponds to each performance indicator.
  • the performance index is a key performance index or determined based on customer needs.
  • step S4 is executed, and the data is trained through the machine learning algorithm based on the simulation task characteristics and the corresponding parameter groups of different values and their performance indicators, and a prediction model is generated.
  • the training data management device 140 obtains simulation task feature data from the task feature management device 110, the parameter group combination and numerical data from the parameter adjustment management device 120, and the performance indicators and numerical data from the performance recording device 130, respectively.
  • the learning management device 150 prepares training data.
  • the training data management device 140 is also used for data cleaning and data preprocessing tasks.
  • the input of the training data management device 140 is:
  • [Simulation task characteristics, parameter group values, performance index values] [I, x 1 , x 2 , x 3 , y 1 , y 2 , z 1 , z 2 , r, t]
  • I is the simulation task feature of the grabbing robot B acquired from the task feature management device 110, that is, one of the feature vectors of the image data of the grabbing robot B.
  • the numerical combination of the parameter group includes x 1 , x 2 , x 3 , y 1 , y 2 , z 1 , z 2 .
  • R is a performance index matrix for different parameter group combinations and values, and r and t are the weight vectors of the performance index R.
  • R is exemplarily determined by two weight vectors r and t, as follows:
  • w 1 and w 2 are the weights of each performance indicator.
  • the performance indicators are added together as a indicator to execute the machine algorithm. Therefore, a training data includes:
  • the training data management device 140 also performs data cleaning functions, such as cleaning up data noise including missing data and erroneous data, and it also applies specific standard algorithms to preprocess these data to obtain appropriate training data.
  • the learning management device 150 applies different machine learning algorithms to obtain prediction models, that is, latent rules. Based on the generation rules of the predictive model, users can predict system performance and choose to optimize the parameters of the entire industrial system more efficiently, which can be retrieved in the simulation as much as possible.
  • the learning management device 150 can embed existing data mining and its learning algorithms, such as decision tree C5.0 (decision tree C5.0) or artificial neural network (ANN, Artificial Neural Network) to learn from simulation data Patterns and hints of knowledge.
  • the simulation data is the marked training data
  • the learning goal is a rule, which can obtain the performance index of the given task feature and parameter group value.
  • rules users can quickly predict system performance given any scenario and parameters, and can also adjust system parameters online to use predictive models to meet performance requirements.
  • a decision tree as a learning algorithm for acquiring knowledge from simulation results.
  • the decision tree training data set we can finally obtain the rule knowledge as shown in Figure 6.
  • R 1 , R 2 ,..., R m represent discrete values of different R values.
  • the performance index R can be directly obtained through [I, x 1 , x 2 , x 3 , y 1 , y 2 , z 1 , z 2 ]. Therefore, the user can select a specific value of [I, x 1 , x 2 , x 3 , y 1 , y 2 , z 1 , z 2 ] to obtain an acceptable or best performance index R.
  • the decision tree is automatically generated by the algorithm. Once the customer's data is input, the decision tree is automatically used to make a prediction model and output the prediction result.
  • the decision tree shown in FIG. 6 is generated based on the simulation task characteristics of the foregoing steps and the corresponding parameter groups with different values and their performance indicators through automatic training data of a machine learning algorithm.
  • the resolution of the feature image of the simulation task is distinguished according to the value range of >0.92 and ⁇ 0.92.
  • the image resolution is greater than 0.92
  • the target ratio in the feature image of the simulation task is distinguished according to the value range of> 0.7 and ⁇ 0.7.
  • the image resolution is ⁇ 0.92
  • the average number of feature objects in the simulation task is determined to be> 0.851 It is distinguished from the value range ⁇ 0.851.
  • step S4 the following step is further included: based on the input simulation task characteristics and the numerical value of the parameter group, query the prediction model to obtain the numerical value of the performance index.
  • the prediction model For example, in the embodiment of the grabbing robot B, the simulation task feature I and the parameter group x, y, z input by the customer, therefore, the decision tree shown in Figure 6 can be queried to obtain the corresponding performance index R value, the customer Judgments can be made based on the predictive model.
  • the second aspect of the present invention provides a predictive model learning device of an industrial system, wherein the industrial system executes simulation according to simulation tasks on a simulation platform, and the predictive model learning device of the industrial system: a task feature management device, which uses statistics The index quantifies the simulation task of the industrial system to extract simulation task characteristics; a parameter adjustment management device that extracts a parameter group from the system module of the industrial system that performs the simulation and adjusts the value of the parameter group to trigger the simulation platform Perform simulation on multiple parameter groups of the industrial system based on different values; performance recording device, which records performance indicators according to the values of the parameter groups after adjusting the parameters; training data management device, which is based on simulation task characteristics and corresponding differences Numerical parameter groups and their performance indicators train data through machine learning algorithms; learning management devices, which generate predictive models.
  • a task feature management device which uses statistics The index quantifies the simulation task of the industrial system to extract simulation task characteristics
  • a parameter adjustment management device that extracts a parameter group from the system module of the industrial system that performs the simulation and adjusts the
  • the parameter adjustment management device further includes: a parameter feature group recording device, which extracts parameter groups from the system modules of the industrial system that executes the simulation, and marks the adjustable parameter groups; and a parameter adjustment device, which is based on each The adjustment range corresponding to the adjustable parameter group adjusts the value of the adjustable parameter group, triggering the simulation platform to perform simulation on multiple parameter groups of the industrial system based on different values; a parameter recording device that synchronizes and records the adjusted parameters The combination and value of the parameter group.
  • a parameter feature group recording device which extracts parameter groups from the system modules of the industrial system that executes the simulation, and marks the adjustable parameter groups
  • a parameter adjustment device which is based on each The adjustment range corresponding to the adjustable parameter group adjusts the value of the adjustable parameter group, triggering the simulation platform to perform simulation on multiple parameter groups of the industrial system based on different values
  • a parameter recording device that synchronizes and records the adjusted parameters The combination and value of the parameter group.
  • the performance recording device further includes: a performance registration device that records performance indicators according to the values of the parameter group after adjusting the parameters; a performance recording sub-device that records the performance indicators and simulation results corresponding to each other The value of the performance index.
  • the performance index is a key performance index or determined based on customer needs.
  • the learning management device is also used to query the prediction model to obtain the value of the performance index based on the input simulation task feature and the value of the parameter group.
  • the third aspect of the present invention provides a predictive model learning system for an industrial system, which includes: a processor; and a memory coupled to the processor, the memory having instructions stored therein, and the instructions are executed by the processor.
  • the action includes: S1, using statistical indicators to quantify the simulation task of the industrial system to extract simulation task features; S2, extracting a parameter set from the system module of the industrial system that performs the simulation And adjust the value of the parameter group, trigger the simulation platform to perform simulation on the multiple parameter groups of the industrial system based on different values; S3, record performance indicators according to the value of the parameter group after adjusting the parameters; S4, based on simulation
  • S1 using statistical indicators to quantify the simulation task of the industrial system to extract simulation task features
  • S2 extracting a parameter set from the system module of the industrial system that performs the simulation And adjust the value of the parameter group, trigger the simulation platform to perform simulation on the multiple parameter groups of the industrial system based on different values
  • S4 based
  • the action S2 further includes: S21, extracting a parameter group from the system module of the industrial system that executes the simulation, and marking the adjustable parameter group; S22, adjusting the station according to the adjustment range corresponding to each adjustable parameter group.
  • the value of the adjustable parameter group triggers the simulation platform to perform simulation on the multiple parameter groups of the industrial system based on different values; S23, synchronize and record the parameter group combination and value after adjusting the parameters.
  • the action S3 further includes: S31, recording the performance index according to the value of the parameter group after adjusting the parameter; S32, recording the value of the performance index based on the corresponding recording performance index and the simulation result.
  • the method further includes: querying the prediction model to obtain the value of the performance index based on the input simulation task feature and the value of the parameter group.
  • the fourth aspect of the present invention provides a computer program product, which is tangibly stored on a computer-readable medium and includes computer-executable instructions that, when executed, cause at least one processor to execute The method described in the first aspect of the present invention.
  • the fifth aspect of the present invention provides a computer-readable medium on which computer-executable instructions are stored, and when executed, the computer-executable instructions cause at least one processor to perform the method according to the first aspect of the present invention.
  • the present invention considers the data in the simulation task to train and construct the prediction model.
  • the step of adjusting parameters of the present invention adopts a systematic and automatic way, allowing simulation to generate a large number of and various parameters for data training, while this part of the prior art is often done manually. Therefore, these system parameters can be explored and tested in a structured manner to avoid incomplete consideration of manual methods.
  • simulation data can be converted into online prediction models, which can guide users to dynamically and more efficiently select parameter combinations and values for a complex system.
  • the present invention can simultaneously help customers build an online prediction system.
  • the present invention allows industrial customers to predict their system performance online and dynamically adjust parameters.

Landscapes

  • Engineering & Computer Science (AREA)
  • General Engineering & Computer Science (AREA)
  • Theoretical Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • General Physics & Mathematics (AREA)
  • Manufacturing & Machinery (AREA)
  • Computational Linguistics (AREA)
  • Artificial Intelligence (AREA)
  • Data Mining & Analysis (AREA)
  • Evolutionary Computation (AREA)
  • Computing Systems (AREA)
  • Mathematical Physics (AREA)
  • Software Systems (AREA)
  • Automation & Control Theory (AREA)
  • Quality & Reliability (AREA)
  • Robotics (AREA)
  • Mechanical Engineering (AREA)
  • Management, Administration, Business Operations System, And Electronic Commerce (AREA)

Abstract

本发明提供了工业系统的预测模型学习方法、装置和系统，其中，所述工业系统在一个仿真平台上按照仿真任务执行仿真，所述工业系统的预测模型学习方法包括如下步骤：S1，利用统计指标量化所述工业系统的仿真任务以提取仿真任务特征；S2，从执行仿真的所述工业系统的系统模块中提取参数组并调整所述参数组的数值，触发所述仿真平台对所述工业系统基于不同数值的多个参数组执行仿真；S3，根据调整参数后的参数组的数值来记录性能指标；S4，基于仿真任务特征以及相对应的不同数值的参数组及其性能指标通过机器学习算法训练数据，并生成预测模型。利用该预测模型，用户可以快速预测基于特定场景和参数的系统性能，并通过自动选择系统参数来优化系统。

Description

工业系统的预测模型学习方法、装置和系统 技术领域
本发明涉及工业自动化领域，尤其涉及一种工业系统的预测模型学习方法、装置和系统。
背景技术
工业领域出现了越来越多的复杂系统，例如智能机器人和自动驾驶系统等。预测这些复杂系统的性能能够允许用户采取积极动作，例如当需要满足性能需要时改变系统参数。
然而，预测复杂系统的性能不容易。其中一种方法是通过仿真评估系统性能。在仿真中产生多个情景并执行测试，以产生大量用于执行评估的仿真结果。然而，由于探索一个目标系统的设计空间(例如参数调试)会花费非常多仿真执行时间，此外还需要额外的人工时间来在一个情景中建立和执行仿真。因此，一个典型的仿真平台易于测试和评估复杂系统，但是并不适合在线预测系统性能。
因此，基于真实数据为复杂工业系统建立一个预测模型是更好的方案，然而，为复杂工业系统的所有子系统收集真是数据非常贵，并且有时并不可能。
为了解决上述问题，现有技术的其中一种方案是利用及其学习方法来建立一个刷新机械的培训数据，并从数据库中选择一个预测模型模板以及基于刷新数据调试数据来优化预测模型。然而，这种方案依赖于未数据分析收集真实数据，并非通过仿真机制。此外，其需要一个预测模板库，然后通过额外培训来精炼模板。这种方案的预测目标是和商业相关，尤其是销量。
现有技术的另一个方案是进行可信性评估，其提供了一种灰色理论预测模型(grey prediction model)的完整应用以定性和定量地从有限测试数据发现一个复杂仿真系统的内在规律，以改善复杂系统的可信性评估。这种方案的效果仅限于可信性，而并不是性能指标。
现有技术还提出了一种方案，为控制参数调试机制，其基于仿真控制参数优化。模型错误作为一个问题被解决了，具有更高精确度的控制参数调节能够被识别。然而，这种方案并不是针对复杂系统，也并不是基于仿真。并且，这种方案仅限于伺服电机和参数调试控制器。
现有技术的又一方案是在仿真环境中训练的神经网络机器人，其从一个pybullet模拟器中的算法监督人收集综合演示统计(bin)，然后其通过从仿真器中获取的综合数据学习神经网络政策，其也基于从仿真器中获取的弹道诸元(trajectory data)训练神经网络机器人控制器。然而，这种方案对智能机器人系统和收集的仿真数据有用，但是其重点在训练机器人控制器来，而并不是为了复杂系统和仿真器来增强仿真器来获得一个预测模型产生器。
发明内容
本发明第一方面提供了工业系统的预测模型学习方法，其中，所述工业系统在一个仿真平台上按照仿真任务执行仿真，所述工业系统的预测模型学习方法包括如下步骤：S1，利用统计指标量化所述工业系统的仿真任务以提取仿真任务特征；S2，从执行仿真的所述工业系统的系统模块中提取参数组并调整所述参数组的数值，触发所述仿真平台对所述工业系统基于不同数值的多个参数组执行仿真；S3，根据调整参数后的参数组的数值来记录性能指标；S4，基于仿真任务特征以及相对应的不同数值的参数组及其性能指标通过机器学习算法训练数据，并生成预测模型。
进一步地，所述步骤S2包括如下步骤：S21，从执行仿真的所述工业系统的系统模块中提取参数组，并标记可调参数组；S22，根据每个可调参数组对应的调整范围调整所述可调参数组的数值，触发所述仿真平台对所述工业系统基于不同数值的多个参数组执行仿真；S23，同步并记录调整参数后的参数组组合和数值。
进一步地，所述步骤S3包括如下步骤：S31，根据调整参数后的参数组的数值来记录性能指标；S32，基于相互对应的所述记录性能指标和仿真结果来记录所述性能指标的数值。
进一步地，所述性能指标为关键性能指标或者是基于客户需求确定 的。
进一步地，在所述步骤S4之后还包括如下步骤：基于输入的仿真任务特征和参数组的数值，查询所述预测模型获得性能指标的数值。
进一步地，所述工业系统为情境感知机器人，其中，所述仿真任务特征包括深度彩色图像的多个指标，所述指标包括以下任一项或任多项：图像分辨率；图像中目标比例；物体平均数。
进一步地，所述情境感知机器人为抓取机器人，其系统模块包括:视觉模块，其对输入的深度彩色图像做图像处理；路径规划模块，其通过图像识别得到的抓取点计算出机械臂运动到抓取点的路径；动作控制模块，其根据路径规划模块规划的路径控制机械臂的运动。
本发明第二方面提供了工业系统的预测模型学习装置，其中，所述工业系统在一个仿真平台上按照仿真任务执行仿真，所述工业系统的预测模型学习装置：任务特征管理装置，其利用统计指标量化所述工业系统的仿真任务以提取仿真任务特征；参数调节管理装置，其从执行仿真的所述工业系统的系统模块中提取参数组并调整所述参数组的数值，触发所述仿真平台对所述工业系统基于不同数值的多个参数组执行仿真；性能记录装置，其根据调整参数后的参数组的数值来记录性能指标；训练数据管理装置，其基于仿真任务特征以及相对应的不同数值的参数组及其性能指标通过机器学习算法训练数据；学习管理装置，其生成预测模型。
进一步地，所述参数调节管理装置还包括：参数特征组记录装置，其从执行仿真的所述工业系统的系统模块中提取参数组，并标记可调参数组；参数调整装置，其根据每个可调参数组对应的调整范围调整所述可调参数组的数值，触发所述仿真平台对所述工业系统基于不同数值的多个参数组执行仿真；参数记录装置，其同步并记录调整参数后的参数组组合和数值。
进一步地，所述性能记录装置还包括：性能登记装置，其根据调整参数后的参数组的数值来记录性能指标；性能记录子装置，其基于相互对应的所述记录性能指标和仿真结果来记录所述性能指标的数值。
进一步地，所述性能指标为关键性能指标或者是基于客户需求确定的。
进一步地，在所述学习管理装置还用于基于输入的仿真任务特征和参数组的数值，查询所述预测模型获得性能指标的数值。
本发明第三方面提供了工业系统的预测模型学习系统，其中，包括：处理器；以及与所述处理器耦合的存储器，所述存储器具有存储于其中的指令，所述指令在被处理器执行时使所述电子设备执行动作，所述动作包括：S1，利用统计指标量化所述工业系统的仿真任务以提取仿真任务特征；S2，从执行仿真的所述工业系统的系统模块中提取参数组并调整所述参数组的数值，触发所述仿真平台对所述工业系统基于不同数值的多个参数组执行仿真；S3，根据调整参数后的参数组的数值来记录性能指标；S4，基于仿真任务特征以及相对应的不同数值的参数组及其性能指标通过机器学习算法训练数据，并生成预测模型。
进一步地，所述动作S2还包括：S21，从执行仿真的所述工业系统的系统模块中提取参数组，并标记可调参数组；S22，根据每个可调参数组对应的调整范围调整所述可调参数组的数值，触发所述仿真平台对所述工业系统基于不同数值的多个参数组执行仿真；S23，同步并记录调整参数后的参数组组合和数值。
进一步地，所述动作S3还包括：S31，根据调整参数后的参数组的数值来记录性能指标；S32，基于相互对应的所述记录性能指标和仿真结果来记录所述性能指标的数值。
进一步地，在所述动作S4之后还包括：基于输入的仿真任务特征和参数组的数值，查询所述预测模型获得性能指标的数值。
本发明第四方面提供了计算机程序产品，所述计算机程序产品被有形地存储在计算机可读介质上并且包括计算机可执行指令，所述计算机可执行指令在被执行时使至少一个处理器执行根据本发明第一方面所述的方法。
本发明第五方面提供了计算机可读介质，其上存储有计算机可执行指令，所述计算机可执行指令在被执行时使至少一个处理器执行根据本发明第一方面所述的方法。
本发明考虑了仿真任务中的数据来训练和构建预测模型。本发明的调整参数的步骤采用了系统和自动的方式，允许仿真产生大量和各种参数用于数据训练，而现有技术这部分往往是由人工完成的。因此，这些 系统参数能够结构化地探索和测验，避免人工方式考虑不完整。
利用本发明，仿真数据能够被转换为在线预测模型，其可以指导用户为一个复杂系统动态地和更有效率地选择参数组合和数值。当工业系统执行仿真时，本发明可以同时帮助客户建立一个在线预测系统。本发明允许工业客户在线预测他们的系统性能并动态地调整参数。
附图说明
图1是根据本发明一个具体实施例的工业系统的预测模型学习系统的结构示意图；
图2是抓取机器人从盒子中抓取部件的场景示意图；
图3是根据本发明一个具体实施例的工业系统的预测模型学习系统的参数调节管理装置的结构示意图；
图4是根据本发明一个具体实施例的工业系统的预测模型学习机制的参数调节逻辑图；
图5根据本发明一个具体实施例的工业系统的预测模型学习系统的性能记录装置的结构示意图；
图6是本发明一个具体实施例的工业系统的预测模型学习机制的决策树C5.0的示意图。
具体实施方式
以下结合附图，对本发明的具体实施方式进行说明。
本发明是现有仿真平台的扩展，其能够自动检索需要仿真的复杂系统的参数组并记录主要的仿真数据以为目标工业系统学习一个潜在预测模型(latent prediction model)。仿真过程中，本发明为每个系统模块记录感兴趣的数据，构建场景和调节的参数的特征组，最终应用设备采样来为仿真的系统产生一个预测模型。利用该预测模型，用户可以快速预测基于特定场景和参数的复杂系统的性能，然后通过自动选择系统参数来优化系统，其中，自动选择利用了预测模型来满足性能需求。
当对一个目标系统的性能仿真时，本发明能够同时帮助客户为他们的复杂系统建立一个在线预测系统。其中，所述目标系统包括智能机器人系统，制造单元或者生产线。本发明能够允许客户同时在线预测他们 的系统性能并自动调节参数。本发明能够应用于很多邻居，特别是复杂工业系统，包括工业情境感知制造系统，例如自主设备、工作站和车间等。本发明适用于情景感知装置，例如自动驾驶、智能机器人以及无人机领域等。
本发明提供了一种从仿真中自动学习预测模型的机制，其是对仿真平台功能的扩展为在线执行模型构建，本发明能够轻易整合到现有的仿真器中。图1是根据本发明一个具体实施例的工业系统的预测模型学习系统的结构示意图。具体地，工业系统的预测模型学习系统100包括任务特征管理装置110，参数调节管理装置120、性能记录装置130、训练数据管理装置140和学习管理装置150。本发明提供的工业系统的预测模型学习系统100是现有仿真平台200的功能扩展。仿真平台200输入的是仿真任务，输出的是仿真结果。
下面结合抓取机器人作为工业系统执行仿真为实施例对本发明进行说明。本发明第一方面提供了一种工业系统的预测模型学习方法，其包括如下步骤：
首先执行步骤S1，任务特征管理装置110利用统计指标量化所述工业系统的仿真任务以提取仿真任务特征。具体地，任务特征管理装置110用于管理和载入仿真任务的特征，也就是利用统计指标(statistical metrics)来量化输入仿真任务的特征，并把仿真任务特征数据在仿真执行后提供给训练数据管理装置140。
如图2所示，根据本发明有具体实施例，需要仿真的是工业系统是一个抓取机器人从盒子B’中抓取多个部件的过程。抓取机器人B按照规划的路径运动来转动第一关节j 1、第二关节j 2和第三关节j 3来从盒子B’中抓取多个部件的过程，所述多个部件示例性地包括第一部件p 1、第二部件p 2和第三部件p 3。
如图1所示，仿真平台200执行抓取机器人B上述过程的仿真，其中，所述仿真平台200中包括多个抓取机器人B的系统模块，其中包括视觉模块210、路径规划模块220和动作控制模块230。在本实施例中，仿真平台200输入的是图像，特别地为深度彩色图像(RGBD image)。视觉模块210基于输入的深度彩色图像做图像处理，路径规划模块220通过图像识别得到的抓取点计算出机械臂运动到抓取点的路径，动作控 制模块230根据路径规划模块220规划的路径控制机械臂的运动。
具体地，任务特征管理装置110输入的是图像数据，其计算的仿真任务也就是图像数据的统计指标组。这些指标和特征能够在所述任务特征管理装置110中预先定义并提取输入仿真任务需要执行仿真的特征。例如，在本实施例中，抓取机器人B是情境感知机器人，所述仿真任务特征是一系列的深度彩色图像的多个指标，其中一个指标是特征矢量I：
I＝[图像分辨率，图像中目标比例，物体平均数]
上述特征数值能够基于每个特定仿真执行从任务数据计算并用仿真平台来规定。任务特征数据数值用参数调节管理装置120同步指标并做数据存储，并且稍后作为训练数量集的一部分。
然后执行步骤S2，参数调节管理装置120从执行仿真的所述工业系统的系统模块中提取参数组并调整所述参数组的数值，触发所述仿真平台200对所述工业系统基于不同数值的多个参数组执行仿真。所述步骤S2包括三个子步骤S21、S22和S23。如图3所示，参数调节管理装置120由三个次装置组成，分别为参数特征组记录装置121、参数调整装置122和参数记录装置123。此外，所述参数调节管理装置120还用于调整为多个系统模块自动地记录参数，其能够系统地为学习管理装置150产生仿真数据。
其中，在子步骤S21中，特征组记录装置121从执行仿真的所述工业系统的系统模块中提取参数组，并标记可调参数组。特征组记录装置121将可调参数组发送给参数调整装置122。
根据本发明一个具体实施例，参数特征组记录装置121从每个系统模块收集可调参数组，也就是从视觉模块210、路径规划模块220和动作控制模块230收集可调参数组，上述参数可以标记为“可调”，这些可调参数和其可调范围一样接着被通知到参数调节管理装置120中的参数调整装置122作为后续分析。参数特征组记录装置121的输出为参数特征数值。
在本实施例中，参数特征组记录装置121从输入的参数组中标记可调参数组包括：
Figure PCTCN2020082459-appb-000001
Figure PCTCN2020082459-appb-000002
Figure PCTCN2020082459-appb-000003
其中，perception是从视觉模块210提取的可调参数组，Path Planner是从路径规划模块220提取的可调参数组，Motion Controller从和动作控制模块230提取的可调参数组。其中，从视觉模块210提取的可调参数组perception包括x 1、x 2、x 3，可调参数组x 1、x 2、x 3还分别具有取值范围定义为x 1min到x 1max，x 2min到x 2max，x 3min到x 3max。例如，参数特征组记录装置121利用的是LINEMOD算法，x 1为模板相似度(similarity of templates)，x 2为非极大值抑制(non-maximum suppression)，x 3为对比度(contrast ratio)。从路径规划模块220提取的可调参数组Path Planner包括y 1和y 2，y 1和y 2具有分别被定义为y 11、y 12、y 13和y 21、y 22、y 23的离散值。其中，y 1、y 2是路由选择方面参数，包括起始点和路径点等。此外，从动作控制模块230提取的可调参数组Motion Controller包括z 1和z 2，其分别表示抓取机器人B的端点速度(endpoint velocity)和加速度(acceleration)，z 1和z 2可以是连续变量或者离散值。
在子步骤S22中，参数调整装置122根据每个可调参数组对应的调整范围调整所述参数组的数值，触发所述仿真平台200对所述工业系统基于不同数值的多个参数组执行仿真。参数调整装置122把参数组的数值发送给参数记录装置123。参数调整装置122用于系统性地自动地从参数特征组选出参数特征数值，然后调参并把调整以后的参数推送给仿真平台200，并出发对目标工业系统的仿真来对这些选出特征的在仿真中执行相应地测试。基于应用要求和客户需求应用许多策略来执行参数调整，其中一种是利用特征组中所有可能的数值组合执行，另一种是根据特定分布选取变量数值，例如正态分布。
图4是根据本发明一个具体实施例的工业系统的预测模型学习机制的参数调节逻辑图。如图4所示，参数调整逻辑基本上是先调整参数x，再调整参数y，最后调整参数z。具体地，首先上述参数的初始状态包括：
x 1＝x 1min,x 2＝x 2min,x 3＝x 3min,
y 1＝y 11，y 2＝y 21，
z 1＝z 1min,z 2＝z 2min,
然后判断是否从x 1min到x 1max完成历数数值，如果没有就通过x 1＝x 1+Δx 1继续历数，并判断是否从x 2min到x 2max完成历数数值，如果是的话返回上一步，否则就通过x 2＝x 2+Δx 2继续历数。接着判断是否从x 3min到x 3max完成历数数值，如果是的话返回上一步，否则通过x 3＝x 3+Δx 3继续历数。
接着判断是否从y 11到y 13完成历数数值，如果没有就将y 1设为y 11到y 13的的下一个，然后判断从y 21到y 23完成历数数值，是的话就返回上一步，否则将y 2设为y 21到y 23的的下一个。
最后判断是否从z 1min到z 1max完成历数数值，如果没有就通过z 1＝z 1+Δz 1继续历数，如果是的话就返回从y 21到y 23完成历数数值的判断步骤。然后判断是否从z 2min到z 3max完成历数数值，如果没有就通过z 2＝z 2+Δz 2继续历数，如果是的话就返回上一步。接着，将调整后的参数发送给仿真平台200，然后完成参数调整。
因此，对参数x、y和z的调参完成以后将调整后的参数x、y和z分别发送给视觉模块210、路径规划模块220和动作控制模块230。
在子步骤S23中，参数记录装置123同步并记录调整参数后的参数组组合和数值。
接着执行步骤S3，性能记录装置130根据调整参数后的参数组的数值来记录性能指标，其中所述性能指标为关键性能指标(KPI)。其中，性能记录装置130和参数调节管理装置120配合来记录该工业系统的关键性能指标(KPI)，并与参数特征组的每个数值一致。性能记录装置130进一步地包括两个子模块：性能登记装置131和性能记录子装置132。所述步骤S3包括子步骤S31和子步骤S32。
如图5所示，性能记录装置130进一步地包括两个子模块：性能登记装置131和性能记录子装置132。
在子步骤S31中，性能登记装置131配合根据调整参数后的参数组的数值来记录性能指标。性能登记装置131配合参数调节管理装置120通过性能特征组的每个数值来记录关键性能指标KPI。具体地，性能登记装置131收集作为仿真目标的工业系统的性能标准(performance metrics)，性能指标能够在仿真平台200中标明。通常，性能指标和客户感兴趣的KPI有关。多个指标(metrics)能够被记录在性能记录装置130 中。例如，在情境感知机器人系统中，抓取机器人B的性能指标(performance metric)可选地为“是否成功完成将一个部件(例如第一部件p1、第二部件p2和第三部件p3)从盒子里抓取出来的任务”，因此在性能登记装置131中将上述任务记录为r，设定当r＝1时意味着上述任务完成，当r＝0时意味着上述任务失败。另一个示例性的性能指标可选地为时间t，时间t代表完成整个抓取机器人B抓取过程的时间。
在子步骤S32中，性能记录子装置132基于所述记录性能指标和仿真结果来记录所述性能指标的数值。其中，性能指标数值与参数调节管理装置120的指标有关，其需要和参数调节管理装置120一起完成每一轮数据记录。其中，每个参数组的组合和数值及其仿真结果一一对应，而每个仿真结果也对应每个性能指针。
可选地，所述性能指标为关键性能指标或者基于客户需求确定的。
然后执行步骤S4，基于仿真任务特征以及相对应的不同数值的参数组及其性能指标通过机器学习算法训练数据，并生成预测模型。
具体地，训练数据管理装置140分别从任务特征管理装置110获取仿真任务特征数据，从参数调节管理装置120获取参数组组合和数值数据，以及从性能记录装置130获取性能指标和数值数据，并为学习管理装置150准备训练数据。训练数据管理装置140还用于数据清理和数据预处理任务。
在本实施例中，训练数据管理装置140的输入为：
[仿真任务特征，参数组数值，性能指标数值]＝[I，x 1，x 2，x 3，y 1，y 2，z 1，z 2，r，t]
其中，I是从任务特征管理装置110获取的抓取机器人B的仿真任务特征，即抓取机器人B的图像数据其中一个特征矢量。参数组数值组合包括x 1，x 2，x 3，y 1，y 2，z 1，z 2。R是针对不同的参数组组合和数值的性能指标矩阵，r和t是性能指标R的权重矢量。
具体地，在基于每一轮参数组组合及其数值执行仿真以后会得到仿真结果及其性能指针矩阵R，R示例性地是由两个权重矢量r和t决定的，如下：
R＝w 1r+w 2t
其中，w 1和w 2是衡量每个性能指标的权重，将性能指针合计在一起作为一个指针是为了执行机器算法，因此，一个训练数据包括：
[I，x 1，x 2，x 3，y 1，y 2，z 1，z 2，R]
然后，训练数据管理装置140也执行数据清理功能，例如清理包括缺失数据和错误数据的数据噪音，其也会应用特定标准算法来预处理这些数据，以获得适当的训练数据。
接着，学习管理装置150应用不同的机器学习算法来获得预测模型，也就是潜规则(latent rules)。基于预测模型的产生规则，用户可以预测系统性能并选择能够将整个工业系统的参数优化得更加高效，其能够尽可能地在仿真中检索。
具体地，学习管理装置150能够嵌入现有数据挖掘和及其学习算法，例如决策树C5.0(decision tree C5.0)或者人工神经网络(ANN，Artificial Neural Network)，以从仿真数据中学习图案和暗示知识。仿真数据是标记好的训练数据，学习目标是一个规则，其能获得的给定任务特征和参数组数值的性能指标。获得了这样的规则，用户能够给定任何场景和参数快速预测系统性能，也可以在线调整系统参数来利用预测模型满足性能需求。
示例性地，如图6所示，我们利用了决策树作为从仿真结果获取知识的学习算法。通过将应用决策树训练数据集，我们可以最终获得如图6所示的规则知识。其中，R 1，R 2，...，R m代表不同R数值的离散数值。基于学习到的规则，性能指标R能够直接通过[I，x 1，x 2，x 3，y 1，y 2，z 1，z 2]获得。因此，用于可以选择[I，x 1，x 2，x 3，y 1，y 2，z 1，z 2]的特定数值来获得可接受或者最好的性能指标R。决策树是算法自动生成的，客户的数据一旦输入就自动利用决策树做预测模型，输出预测结果。
具体地，图6所示的决策树是基于前述步骤的仿真任务特征以及相对应的不同数值的参数组及其性能指标通过机器学习算法自动训练数据生成的。如图6所示，在抓取机器人B中，将仿真任务特征图像分辨率按照>0.92和≤0.92的取值范围区分。当所述图像分辨率>0.92时，将仿真任务特征图像中目标比例按照>0.7和≤0.7的取值范围区分，当所述图像分辨率≤0.92时，将仿真任务特征物体平均数按照>0.851和≤0.851的取值范围区分。进一步地，当图像中目标比例按照>0.7，继续将x1按照>0.357和≤0.357的取值范围区分，当图像中目标比例按照≤0.7继续判断x3。当x1>0.357时判断z1，当z1>0.2078时得到性能指标R2，当z1≤0.2078时 得到性能指标R3。当x1≤0.357时判断y2，当y2>0.7943时得到性能指标R5，当y2≤0.7943时得到性能指标R7。进一步地，物体平均数按照≤0.851时，将x2按照>0.316和≤0.316的取值范围区分。当x2>0.316时判断y1，y1>0.1455时将z2>0.4582和≤0.4582的取值范围区分。当z2>0.4582时则获得性能指标R6，当z2≤0.4582时则获得性能指标R4，y1≤0.1455则获得R1，x2≤0.316则换为Rm。
因此，在所述步骤S4之后还包括如下步骤：基于输入的仿真任务特征和参数组的数值，查询所述预测模型获得性能指标的数值。例如，在抓取机器人B的实施例中，客户输入的仿真任务特征I和参数组x、y、z，因此可以查询如图6所示的决策树就能够获得相应的性能指标R数值，客户可以基于预测模型做出判断。
本发明第二方面提供了工业系统的预测模型学习装置，其中，所述工业系统在一个仿真平台上按照仿真任务执行仿真，所述工业系统的预测模型学习装置：任务特征管理装置，其利用统计指标量化所述工业系统的仿真任务以提取仿真任务特征；参数调节管理装置，其从执行仿真的所述工业系统的系统模块中提取参数组并调整所述参数组的数值，触发所述仿真平台对所述工业系统基于不同数值的多个参数组执行仿真；性能记录装置，其根据调整参数后的参数组的数值来记录性能指标；训练数据管理装置，其基于仿真任务特征以及相对应的不同数值的参数组及其性能指标通过机器学习算法训练数据；学习管理装置，其生成预测模型。
进一步地，所述参数调节管理装置还包括：参数特征组记录装置，其从执行仿真的所述工业系统的系统模块中提取参数组，并标记可调参数组；参数调整装置，其根据每个可调参数组对应的调整范围调整所述可调参数组的数值，触发所述仿真平台对所述工业系统基于不同数值的多个参数组执行仿真；参数记录装置，其同步并记录调整参数后的参数组组合和数值。
进一步地，所述性能记录装置还包括：性能登记装置，其根据调整参数后的参数组的数值来记录性能指标；性能记录子装置，其基于相互对应的所述记录性能指标和仿真结果来记录所述性能指标的数值。
进一步地，所述性能指标为关键性能指标或者是基于客户需求确定 的。
进一步地，在所述学习管理装置还用于基于输入的仿真任务特征和参数组的数值，查询所述预测模型获得性能指标的数值。
本发明第三方面提供了工业系统的预测模型学习系统，其中，包括：处理器；以及与所述处理器耦合的存储器，所述存储器具有存储于其中的指令，所述指令在被处理器执行时使所述电子设备执行动作，所述动作包括：S1，利用统计指标量化所述工业系统的仿真任务以提取仿真任务特征；S2，从执行仿真的所述工业系统的系统模块中提取参数组并调整所述参数组的数值，触发所述仿真平台对所述工业系统基于不同数值的多个参数组执行仿真；S3，根据调整参数后的参数组的数值来记录性能指标；S4，基于仿真任务特征以及相对应的不同数值的参数组及其性能指标通过机器学习算法训练数据，并生成预测模型。
进一步地，所述动作S2还包括：S21，从执行仿真的所述工业系统的系统模块中提取参数组，并标记可调参数组；S22，根据每个可调参数组对应的调整范围调整所述可调参数组的数值，触发所述仿真平台对所述工业系统基于不同数值的多个参数组执行仿真；S23，同步并记录调整参数后的参数组组合和数值。
进一步地，所述动作S3还包括：S31，根据调整参数后的参数组的数值来记录性能指标；S32，基于相互对应的所述记录性能指标和仿真结果来记录所述性能指标的数值。
进一步地，在所述动作S4之后还包括：基于输入的仿真任务特征和参数组的数值，查询所述预测模型获得性能指标的数值。
本发明第四方面提供了计算机程序产品，所述计算机程序产品被有形地存储在计算机可读介质上并且包括计算机可执行指令，所述计算机可执行指令在被执行时使至少一个处理器执行根据本发明第一方面所述的方法。
本发明第五方面提供了计算机可读介质，其上存储有计算机可执行指令，所述计算机可执行指令在被执行时使至少一个处理器执行根据本发明第一方面所述的方法。
本发明考虑了仿真任务中的数据来训练和构建预测模型。本发明的调整参数的步骤采用了系统和自动的方式，允许仿真产生大量和各种参 数用于数据训练，而现有技术这部分往往是由人工完成的。因此，这些系统参数能够结构化地探索和测验，避免人工方式考虑不完整。
利用本发明，仿真数据能够被转换为在线预测模型，其可以指导用户为一个复杂系统动态地和更有效率地选择参数组合和数值。当工业系统执行仿真时，本发明可以同时帮助客户建立一个在线预测系统。本发明允许工业客户在线预测他们的系统性能并动态地调整参数。
尽管本发明的内容已经通过上述优选实施例作了详细介绍，但应当认识到上述的描述不应被认为是对本发明的限制。在本领域技术人员阅读了上述内容后，对于本发明的多种修改和替代都将是显而易见的。因此，本发明的保护范围应由所附的权利要求来限定。此外，不应将权利要求中的任何附图标记视为限制所涉及的权利要求；“包括”一词不排除其它权利要求或说明书中未列出的装置或步骤；“第一”、“第二”等词语仅用来表示名称，而并不表示任何特定的顺序。

Claims (18)

  1. 工业系统的预测模型学习方法，其中，所述工业系统在一个仿真平台上按照仿真任务执行仿真，所述工业系统的预测模型学习方法包括如下步骤：
    S1，利用统计指标量化所述工业系统的仿真任务以提取仿真任务特征；
    S2，从执行仿真的所述工业系统的系统模块中提取参数组并调整所述参数组的数值，触发所述仿真平台对所述工业系统基于不同数值的多个参数组执行仿真；
    S3，根据调整参数后的参数组的数值来记录性能指标；
    S4，基于仿真任务特征以及相对应的不同数值的参数组及其性能指标通过机器学习算法训练数据，并生成预测模型。
  2. 根据权利要求1所述的工业系统的预测模型学习方法，其特征在于，所述步骤S2包括如下步骤：
    S21，从执行仿真的所述工业系统的系统模块中提取参数组，并标记可调参数组；
    S22，根据每个可调参数组对应的调整范围调整所述可调参数组的数值，触发所述仿真平台对所述工业系统基于不同数值的多个参数组执行仿真；
    S23，同步并记录调整参数后的参数组组合和数值。
  3. 根据权利要求1所述的工业系统的预测模型学习方法，其特征在于，所述步骤S3包括如下步骤：
    S31，根据调整参数后的参数组的数值来记录性能指标；
    S32，基于相互对应的所述记录性能指标和仿真结果来记录所述性能指标的数值。
  4. 根据权利要求3所述的工业系统的预测模型学习方法，其特征在于，所述性能指标为关键性能指标或者是基于客户需求确定的。
  5. 根据权利要求1所述的工业系统的预测模型学习方法，其特征在于，在所述步骤S4之后还包括如下步骤：基于输入的仿真任务特征和参数组的数值，查询所述预测模型获得性能指标的数值。
  6. 根据权利要求5所述的工业系统的预测模型学习方法，其特征在于，所述工业系统为情境感知机器人，其中，所述仿真任务特征包括深度彩色图像的多个指标，所述指标包括以下任一项或任多项：
    -图像分辨率；
    -图像中目标比例；
    -物体平均数。
  7. 根据权利要求6所述的工业系统的预测模型学习方法，其特征在于，所述情境感知机器人为抓取机器人，其系统模块包括:
    视觉模块，其对输入的深度彩色图像做图像处理；
    路径规划模块，其通过图像识别得到的抓取点计算出机械臂运动到抓取点的路径；
    动作控制模块，其根据路径规划模块规划的路径控制机械臂的运动。
  8. 工业系统的预测模型学习装置，其中，所述工业系统在一个仿真平台上按照仿真任务执行仿真，所述工业系统的预测模型学习装置：
    任务特征管理装置，其利用统计指标量化所述工业系统的仿真任务以提取仿真任务特征；
    参数调节管理装置，其从执行仿真的所述工业系统的系统模块中提取参数组并调整所述参数组的数值，触发所述仿真平台对所述工业系统基于不同数值的多个参数组执行仿真；
    性能记录装置，其根据调整参数后的参数组的数值来记录性能指标；
    训练数据管理装置，其基于仿真任务特征以及相对应的不同数值的参数组及其性能指标通过机器学习算法训练数据；
    学习管理装置，其生成预测模型。
  9. 根据权利要求8所述的工业系统的预测模型学习装置，其特征在于，所述参数调节管理装置还包括：
    参数特征组记录装置，其从执行仿真的所述工业系统的系统模块中提取参数组，并标记可调参数组；
    参数调整装置，其根据每个可调参数组对应的调整范围调整所述可调参数组的数值，触发所述仿真平台对所述工业系统基于不同数值的多个参数组执行仿真；
    参数记录装置，其同步并记录调整参数后的参数组组合和数值。
  10. 根据权利要求8所述的工业系统的预测模型学习装置，其特征在于，所述性能记录装置还包括：
    性能登记装置，其根据调整参数后的参数组的数值来记录性能指标；
    性能记录子装置，其基于相互对应的所述记录性能指标和仿真结果来记录所述性能指标的数值。
  11. 根据权利要求10所述的工业系统的预测模型学习装置，其特征在于，所述性能指标为关键性能指标或者是基于客户需求确定的。
  12. 根据权利要求8所述的工业系统的预测模型学习装置，其特征在于，在所述学习管理装置还用于基于输入的仿真任务特征和参数组的数值，查询所述预测模型获得性能指标的数值。
  13. 工业系统的预测模型学习系统，其中，包括：
    处理器；以及
    与所述处理器耦合的存储器，所述存储器具有存储于其中的指令，所述指令在被处理器执行时使所述电子设备执行动作，所述动作包括：
    S1，利用统计指标量化所述工业系统的仿真任务以提取仿真任务特征；
    S2，从执行仿真的所述工业系统的系统模块中提取参数组并调整所述参数组的数值，触发所述仿真平台对所述工业系统基于不同数值的多个参数组执行仿真；
    S3，根据调整参数后的参数组的数值来记录性能指标；
    S4，基于仿真任务特征以及相对应的不同数值的参数组及其性能指标通过机器学习算法训练数据，并生成预测模型。
  14. 根据权利要求13所述的工业系统的预测模型学习系统，其特征在于，所述动作S2还包括：
    S21，从执行仿真的所述工业系统的系统模块中提取参数组，并标记可调参数组；
    S22，根据每个可调参数组对应的调整范围调整所述可调参数组的数值，触发所述仿真平台对所述工业系统基于不同数值的多个参数组执行仿真；
    S23，同步并记录调整参数后的参数组组合和数值。
  15. 根据权利要求13所述的工业系统的预测模型学习系统，其特征 在于，所述动作S3还包括：
    S31，根据调整参数后的参数组的数值来记录性能指标；
    S32，基于相互对应的所述记录性能指标和仿真结果来记录所述性能指标的数值。
  16. 根据权利要求13所述的工业系统的预测模型学习系统，其特征在于，在所述动作S4之后还包括：基于输入的仿真任务特征和参数组的数值，查询所述预测模型获得性能指标的数值。
  17. 计算机程序产品，所述计算机程序产品被有形地存储在计算机可读介质上并且包括计算机可执行指令，所述计算机可执行指令在被执行时使至少一个处理器执行根据权利要求1至7中任一项所述的方法。
  18. 计算机可读介质，其上存储有计算机可执行指令，所述计算机可执行指令在被执行时使至少一个处理器执行根据权利要求1至7中任一项所述的方法。
PCT/CN2020/082459 2020-03-31 2020-03-31 工业系统的预测模型学习方法、装置和系统 WO2021195970A1 (zh)

Priority Applications (4)

Application Number Priority Date Filing Date Title
EP20928458.7A EP4122654A4 (en) 2020-03-31 2020-03-31 PREDICTIVE MODEL LEARNING METHOD, APPARATUS AND SYSTEM FOR AN INDUSTRIAL SYSTEM
PCT/CN2020/082459 WO2021195970A1 (zh) 2020-03-31 2020-03-31 工业系统的预测模型学习方法、装置和系统
CN202080096949.XA CN115135463A (zh) 2020-03-31 2020-03-31 工业系统的预测模型学习方法、装置和系统
US17/916,234 US20230153640A1 (en) 2020-03-31 2020-03-31 Prediction Model Learning Method, Apparatus and System for an Industrial System

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
PCT/CN2020/082459 WO2021195970A1 (zh) 2020-03-31 2020-03-31 工业系统的预测模型学习方法、装置和系统

Publications (1)

Publication Number Publication Date
WO2021195970A1 true WO2021195970A1 (zh) 2021-10-07

Family

ID=77927568

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/CN2020/082459 WO2021195970A1 (zh) 2020-03-31 2020-03-31 工业系统的预测模型学习方法、装置和系统

Country Status (4)

Country Link
US (1) US20230153640A1 (zh)
EP (1) EP4122654A4 (zh)
CN (1) CN115135463A (zh)
WO (1) WO2021195970A1 (zh)

Cited By (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN116100578A (zh) * 2023-04-13 2023-05-12 深圳市大族机器人有限公司 一种双关节机器人的指令快速响应系统

Families Citing this family (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN115659711A (zh) * 2022-12-29 2023-01-31 北京云道智造科技有限公司 一种仿真方法、装置、电子设备及存储介质
CN117314926B (zh) * 2023-11-30 2024-01-30 湖南大学 激光改质加工设备维护的确认方法、设备及存储介质

Citations (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US20190303755A1 (en) * 2018-04-02 2019-10-03 International Business Machines Corporation Water quality prediction
CN110450153A (zh) * 2019-07-08 2019-11-15 清华大学 一种基于深度强化学习的机械臂物品主动拾取方法
CN110647354A (zh) * 2019-09-30 2020-01-03 东软医疗系统股份有限公司 设备运行控制方法、装置及设备

Family Cites Families (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
KR101980603B1 (ko) * 2016-05-20 2019-05-22 구글 엘엘씨 오브젝트(들)를 캡처하는 이미지(들)에 기초하는 그리고 환경에서의 미래 로봇 움직임에 대한 파라미터(들)에 기초하여 로봇 환경에서의 오브젝트(들)의 모션(들)을 예측하는 것과 관련된 머신 학습 방법들 및 장치
US10773382B2 (en) * 2017-09-15 2020-09-15 X Development Llc Machine learning methods and apparatus for robotic manipulation and that utilize multi-task domain adaptation
CN109800864B (zh) * 2019-01-18 2023-05-30 中山大学 一种基于图像输入的机器人主动学习方法

Patent Citations (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US20190303755A1 (en) * 2018-04-02 2019-10-03 International Business Machines Corporation Water quality prediction
CN110450153A (zh) * 2019-07-08 2019-11-15 清华大学 一种基于深度强化学习的机械臂物品主动拾取方法
CN110647354A (zh) * 2019-09-30 2020-01-03 东软医疗系统股份有限公司 设备运行控制方法、装置及设备

Non-Patent Citations (1)

* Cited by examiner, † Cited by third party
Title
See also references of EP4122654A4 *

Cited By (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN116100578A (zh) * 2023-04-13 2023-05-12 深圳市大族机器人有限公司 一种双关节机器人的指令快速响应系统

Also Published As

Publication number Publication date
EP4122654A1 (en) 2023-01-25
CN115135463A (zh) 2022-09-30
US20230153640A1 (en) 2023-05-18
EP4122654A4 (en) 2023-11-22

Similar Documents

Publication Publication Date Title
WO2021195970A1 (zh) 工业系统的预测模型学习方法、装置和系统
US20190202053A1 (en) Systems and method for robotic learning of industrial tasks based on human demonstration
US20220121183A1 (en) Method and Apparatus for Simulating the Machining on a Machine Tool Using a Self-learning System
EP3924787A1 (en) Creation of digital twin of the interaction among parts of the physical system
CN109978870A (zh) 用于输出信息的方法和装置
CN115438726A (zh) 一种基于数字孪生技术的设备寿命与故障类型预测方法及系统
CN115394143A (zh) 一种基于数字孪生的虚拟仿真教育系统及其交互方法
US20200160210A1 (en) Method and system for predicting a motion trajectory of a robot moving between a given pair of robotic locations
Sheptunov et al. Simulating reliability of the robotic system software on the basis of artificial intelligence
WO2021066796A1 (en) Modeling human behavior in work environments using neural networks
Jain et al. Towards smart manufacturing with virtual factory and data analytics
Boteanu et al. Towards robot adaptability in new situations
TW202006652A (zh) 具輔助判斷功能之產品測試系統及應用於其上的產品測試輔助方法
JP6947029B2 (ja) 制御装置、それを使用する情報処理装置、制御方法、並びにコンピュータ・プログラム
Kuhar et al. Learning to discern: Imitating heterogeneous human demonstrations with preference and representation learning
Afanasyev et al. Methods and tools for the development, implementation and use of the intelligent distance learning environment
CN115205072A (zh) 一种面向长周期测评的认知诊断方法
JP2022186422A (ja) 分類装置、分類方法、及び、分類プログラム
CN110210838B (zh) 一种政务系统设计方法及政务系统
CN114239938A (zh) 一种基于状态的能源数字孪生体的构建方法
CN113159330A (zh) 基于层次任务网规划模型学习的专业学习路径系统及方法
Zheng et al. Assembly sequence planning method for optimum assembly accuracy of complex products based on modified teaching–learning based optimization algorithm
CN110941651B (zh) 一种基于不完备人口信息的交通模式选择分析方法
Zhu et al. A physically-grounded and data-efficient approach to motion prediction using black-box optimization
Allevato Combining simulated predictions and real-world data for efficient robot model adaptation

Legal Events

Date Code Title Description
121 Ep: the epo has been informed by wipo that ep was designated in this application

Ref document number: 20928458

Country of ref document: EP

Kind code of ref document: A1

ENP Entry into the national phase

Ref document number: 2020928458

Country of ref document: EP

Effective date: 20221017

NENP Non-entry into the national phase

Ref country code: DE