A control optimization method and system for intelligent manufacturing process

Abstract

The present invention relates to the field of automation technology, and provides a control optimization method and system for intelligent manufacturing processing, which is applied to the edge equipment of the processing production line, and the method includes: obtaining current production environment data and current processing part data; inputting the current production environment data and the current processing part data into a parameter prediction model, and predicting and outputting n parameter prediction results through the parameter prediction model; converting the n parameter prediction results into processing control commands, and controlling the actuator to execute the processing steps related to the processing control commands. In addition, when the processing steps are executed, the sensor monitors the actual finished product quality information in the actual processing process, and adjusts the parameter prediction model based on the actual finished product quality information. The present invention can effectively solve the technical problem of inaccurate control in the intelligent manufacturing processing process of the prior art, and can realize efficient and accurate manufacturing and production, thereby improving production efficiency.

Description

Technical Field

[0001] The present invention relates to the field of automation technology, and in particular to a control optimization method and system for an intelligent manufacturing process. Background Art

[0002] In recent years, with the rapid development of intelligent manufacturing, the automatic control and optimization of the production and processing process have become increasingly important. The quality of processing control plays a decisive role in the yield rate of processed parts and the production efficiency of the factory.

[0003] However, traditional machining control optimization methods have some limitations. They only predict a single machining parameter and lack the ability to comprehensively consider multiple parameters and provide feedback optimization. When faced with complex machining processes that require extremely high precision, existing methods require controllers to be designed for each control parameter. This will require many controllers, resulting in high computational latency, large errors, and inaccurate control. Summary of the invention

[0004] In view of the defects existing in the prior art in the complex and high-precision processing control process mentioned in the above background technology, the present invention provides a processing control optimization method and system for intelligent manufacturing, which can effectively predict the processing parameters so as to better control the actuator to complete the processing and manufacturing, thereby solving the technical problem of insufficiently precise control in the intelligent manufacturing processing process of the prior art.

[0005] In a first aspect, the present invention proposes a control optimization system for an intelligent manufacturing process, the system comprising a parameter prediction module and a middleware control module, wherein the parameter prediction module comprises a parameter prediction model:

[0006] The parameter prediction module is used to input the current production environment data and the current workpiece data acquired by the sensor into the parameter prediction model, and predict and output n parameter prediction results through the parameter prediction model, where n is a positive integer;

[0007] The control module of the middleware is used to convert the n parameter prediction results into processing control commands, and control the execution mechanism to execute the processing steps related to the processing control commands.

[0008] Preferably, the system further comprises a perception module and an optimization module;

[0009] The perception module is used to obtain actual finished product quality information during the actual processing through sensors when the processing step is executed;

[0010] The optimization module is used to adjust the parameter prediction model based on the actual finished product quality information.

[0011] Preferably, the optimization module is further used to compare each group of actual finished product quality information under the conditions of the current production environment data and the current workpiece data, and select the group of data with the best control parameters from the comparison results as the actual optimal parameters; compare the actual optimal parameters with the parameter prediction results to obtain the differences and deviations in the actual processing process; and adjust the parameter prediction model or the processing control command based on the differences and deviations.

[0012] Preferably, the system further comprises a model building module;

[0013] The model building module is used to train and build the parameter prediction model using historical production environment data and historical workpiece data.

[0014] In a second aspect, the present invention provides a control optimization method for an intelligent manufacturing process, the method being applied to an edge device of a processing production line, the method comprising:

[0015] Obtain current production environment data and current workpiece data;

[0016] Inputting the current production environment data and the current workpiece data into a parameter prediction model, and predicting and outputting n parameter prediction results through the parameter prediction model, where n is a positive integer;

[0017] The n parameter prediction results are converted into processing control commands, and the execution mechanism is controlled to execute processing steps related to the processing control commands.

[0018] Preferably, the method comprises:

[0019] When the processing steps are performed, monitoring actual finished product quality information during the actual processing;

[0020] The parameter prediction model is adjusted based on the actual finished product quality information.

[0021] Preferably, adjusting the parameter prediction model based on the actual finished product quality information includes:

[0022] In the case of the current production environment data and the current workpiece data, each group of actual finished product quality information is compared, and the group of data with the best control parameters is selected from the comparison results as the actual optimal parameters;

[0023] Comparing the actual optimal parameters with the parameter prediction results to obtain differences and deviations in the actual processing;

[0024] Based on the differences and deviations, the parameter prediction model or the process control command is adjusted.

[0025] Optionally, converting the n parameter prediction results into a processing control command, and controlling the execution mechanism to execute the processing steps related to the processing control command, comprises:

[0026] Determine the target processing type corresponding to the parameter prediction result;

[0027] Determining a corresponding target actuator based on the target processing type;

[0028] Converting the parameter prediction results into processing control commands executable by the target actuator;

[0029] The processing control command is input into the target execution mechanism to realize processing control.

[0030] Preferably, the parameter prediction model is constructed or adjusted in the following manner:

[0031] Collecting multiple sets of historical processing condition information, wherein the historical processing condition information includes historical production environment data and historical processing part data;

[0032] Based on different historical processing condition information, multiple groups of historical processing control parameters and finished product quality information are collected;

[0033] According to the finished product quality information, determining the optimal control parameters under different conditions of historical production environment data and historical processed part data to obtain a training data set;

[0034] Using a generative model to learn the data distribution of the training data set and generate data with the same distribution, so as to perform data enhancement on the training data set to obtain an enhanced target training data set;

[0035] A parameter prediction model is constructed based on the target training data set.

[0036] Preferably, the parameter prediction model is constructed or optimized in the following manner:

[0037] Step a: constructing a super network including parameters of multiple levels of granularity according to the set neural network search space, wherein the types of the levels include operation level type, filter level type and weight level type;

[0038] Step b: performing enhancement processing on the historical production environment data and the historical processed part data to obtain target production environment data and target processed part data; inputting the target production environment data and the target processed part data into the super network;

[0039] Step c: using the hypernetwork to predict optimal control parameters, calculating the error between the predicted parameters and the actual optimal parameters and the parameter quantity of the current hypernetwork, and updating the loss function;

[0040] Step d: using a gradient descent algorithm to update the granularity parameter values ​​of each level of the hypernetwork and the corresponding granularity parameter discretization functions of each level according to the updated loss function;

[0041] Step e: applying the updated granularity parameter discretization function of each level type to the granularity parameter value of the corresponding level, and pruning and removing the searchable units whose granularity parameter value in the hypernetwork is a preset value;

[0042] Step f: Repeat steps c to e until the parameter amount of the hypernetwork reaches the target parameter amount, and use the hypernetwork that reaches the target as the parameter prediction model.

[0043] The beneficial effects of the present invention are as follows: the present invention can effectively solve the technical problem of inaccurate control existing in the prior art intelligent manufacturing process, can achieve efficient and accurate manufacturing production, and improve the robustness and accuracy of processing control. BRIEF DESCRIPTION OF THE DRAWINGS

[0044] In order to more clearly illustrate the technical solutions in the embodiments of the present application, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present application. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0045] Figure 1 It is a flow chart of the control optimization method for intelligent manufacturing processing provided by the present invention;

[0046] Figure 2 It is a schematic diagram of the architecture of a control optimization system for intelligent manufacturing processing provided by an embodiment of the present invention;

[0047] Figure 3 is a schematic diagram of blocks in a multi-granularity search space provided by an embodiment of the present invention;

[0048] Figure 4 It is a schematic diagram related to pruning with a fixed pruning rate and using an adaptive pruning technology solution in the process of building or optimizing a parameter prediction model provided by an embodiment of the present invention;

[0049] Figure 5 It is a flow chart of an embodiment of a control optimization method for intelligent manufacturing processing provided by the present invention.

[0050] The realization of the purpose, functional features and advantages of the present invention will be further explained in conjunction with embodiments and with reference to the accompanying drawings. DETAILED DESCRIPTION

[0051] It should be understood that the specific embodiments described herein are only used to explain the present invention, and are not used to limit the present invention.

[0052] In the following description, specific details such as specific system structures, technologies, etc. are provided for the purpose of illustration rather than limitation, so as to provide a thorough understanding of the embodiments of the present application. However, it should be clear to those skilled in the art that the present application may also be implemented in other embodiments without these specific details. In other cases, detailed descriptions of well-known systems, devices, and methods are omitted to prevent unnecessary details from obstructing the description of the present application.

[0053] It should be understood that when used in the present specification and the appended claims, the term "comprising" indicates the presence of described features, integers, steps, operations, elements and / or components, but does not exclude the presence or addition of one or more other features, integers, steps, operations, elements, components and / or combinations thereof.

[0054] It should also be understood that the term “and / or” used in the specification and appended claims refers to and includes any and all possible combinations of one or more of the associated listed items.

[0055] In addition, in the description of the present application specification and the appended claims, the terms "first", "second", "third", etc. are only used to distinguish the descriptions and cannot be understood as indicating or implying relative importance.

[0056] References to "one embodiment" or "some embodiments" etc. described in the specification of this application mean that one or more embodiments of the present application include specific features, structures or characteristics described in conjunction with the embodiment. Therefore, the statements "in one embodiment", "in some embodiments", "in some other embodiments", "in some other embodiments", etc. that appear in different places in this specification do not necessarily refer to the same embodiment, but mean "one or more but not all embodiments", unless otherwise specifically emphasized in other ways. The terms "including", "comprising", "having" and their variations all mean "including but not limited to", unless otherwise specifically emphasized in other ways.

[0057] In order to solve the defects in the prior art mentioned in the above background technology, the present invention mainly discloses a control optimization method and system for intelligent manufacturing processing, wherein intelligent manufacturing refers to a manufacturing mode that integrates traditional manufacturing with information technology by using digital, networked and intelligent technologies. It realizes intelligent production, management and service by integrating advanced sensors, computer control systems, automation equipment and data analysis technologies.

[0058] The method of the present invention is applied to edge equipment of a processing production line, and the method mainly comprises:

[0059] Step S10, obtaining current production environment data and current workpiece data;

[0060] It should be noted that the executor of the method of the present invention is an edge device deployed on a processing production line. The edge device may have a built-in processor and may also have onboard analysis capabilities or artificial intelligence analysis capabilities.

[0061] In some real-time scenarios, there may be multiple different processing production lines in a factory, each processing production line can be deployed with an edge device, and there will be multiple processing links in a production line; in a specific embodiment of the present invention, different processing links can be managed by the same edge device (execute processing control), and of course different processing links can correspond to different edge devices for management and control; in addition, different areas of the production line will also be arranged with different types of sensors (such as temperature sensors, pressure sensors, and speed sensors, etc.), and the edge device is connected to each sensor module to obtain production environment data detected or collected by the sensor. The production environment data may include but is not limited to temperature, humidity, pressure, speed, startup time, running time, and downtime data and current workpiece data (length, width, height, thickness of the workpiece, type of material of the workpiece, processing flow of the workpiece, etc.).

[0062] Step S20, inputting the current production environment data and the current workpiece data into a parameter prediction model, and predicting and outputting n parameter prediction results through the parameter prediction model, where n is a positive integer;

[0063] In a specific implementation, for the construction of a lightweight parameter prediction model, the present invention utilizes advanced machine learning technology, such as preferably using a neural network architecture search method, combined with the workpiece data and production environment data in the intelligent manufacturing process, and obtains a lightweight parameter prediction model through learning of a large amount of data, and deploys the parameter prediction model on the edge device. These workpiece data and production environment data are extracted and regressed through a lightweight neural network model (parameter prediction model) to predict and output multiple processing parameters, that is, the parameter prediction model has the function of simultaneously predicting multiple processing parameters (that is, parameter prediction results);

[0064] It should be noted that the "processing parameters (parameter prediction results)" of the present invention refer to the adjustable parameters that affect the processing process and product processing quality. These parameters directly or indirectly affect the operating status and effect of processing equipment, tools and processes. For example: cutting speed, feed speed, cutting depth, cutting lubrication and cooling, processing force, processing temperature, etc.;

[0065] By using the parameter prediction model, the system can comprehensively consider multiple related processing parameters, thereby more accurately predicting the values ​​of each parameter in the processing process.

[0066] Step S30, converting the n parameter prediction results into processing control commands, and controlling the execution mechanism to execute processing steps related to the processing control commands.

[0067] In a specific implementation, the parameter prediction results obtained in step S20 can be converted into processing control commands directly executed by the actuator using middleware technology. The present invention uses middleware technology to convert the parameter prediction results predicted by the parameter prediction model into a control instruction format or signal that can be directly executed by the actuator, so as to communicate and control with the actuator. This ensures that the parameter prediction results can be directly applied to actual processing control, and improves the real-time control of the processing actuator.

[0068] Step S40, when the processing step is executed, monitoring actual finished product quality information during the actual processing; and adjusting the parameter prediction model based on the actual finished product quality information.

[0069] In a specific implementation, the present invention can detect multiple groups of actual finished product quality information in the actual processing process through sensors, and feed back multiple groups of actual finished product quality information to the edge device: in the case of current production environment data and current workpiece data, the edge device compares each group of actual finished product quality information, selects the group of data with the best control parameters from the comparison results as the actual optimal parameters, compares the actual optimal parameters with the parameter prediction results of step S20, and obtains the differences and deviations in the actual processing process; based on the differences and deviations, adjusts the parameter prediction model or the processing control command.

[0070] The present invention can effectively solve the current status and shortcomings of processing control, has the function of simultaneously predicting multiple processing parameters and synchronously feeding back processing control, can efficiently and accurately complete processing, so as to achieve efficient, precise and reliable production, and improve the robustness and accuracy of processing control.

[0071] The system architecture of the present invention:

[0072] The present invention is mainly used in intelligent manufacturing process control systems to predict process parameters and automatically optimize the process control model according to feedback errors.

[0073] The system application architecture of the present invention is as follows: Figure 2 As shown in the figure, the system framework from top to bottom consists of application layer, development tool layer, intelligent algorithm layer, control layer, and hardware layer.

[0074] Specifically, the application layer includes common processing tasks, such as abrasive water jet polishing, laser welding, high-precision diamond cutting, etc. The application layer builds processing applications with different processing tasks by calling programming interfaces. The programming interface is an abstraction of the intelligent algorithm layer, providing a unified and easy-to-use standard for the construction of upper-level applications. The intelligent algorithm layer includes automatic calibration of sensors, data collection and processing, and enhancement processing. The data after collection and enhancement processing will be used in the design and training of prediction models, and the trained parameter prediction model is used to predict processing parameters. These parameters are input to the control layer to convert and further optimize them into control commands that can be executed by the actuator. Finally, they are sent to the processing equipment to control the actuator to perform tasks, such as: robotic arms, welding mechanisms, polishing machines, etc.

[0075] Based on the above system architecture, the present invention proposes a control optimization system embodiment for intelligent manufacturing processing, as embodiment 1:

[0076] In the first embodiment, the system includes a model building module, a parameter prediction module, a middleware control module, a perception module and an optimization module, wherein:

[0077] (1) The model building module is used to train and build the parameter prediction model using historical production environment data and historical workpiece data; the model building module is mainly responsible for building and training the parameter prediction model, which is used to predict key parameters in the processing process, such as temperature, pressure, speed, etc. The task of the model building module is to select a suitable algorithm and model structure, and use training data (historical production environment data and historical workpiece data) to train and optimize the model to achieve accurate parameter prediction. The specific construction method of the parameter prediction model can be referred to the step scheme of the following embodiment 2.

[0078] The training data can be obtained through calibrated sensors, and the training data is formed by performing data enhancement processing on historical production environment data and historical processing part data. Figure 2 , the model building module of the control optimization system of this embodiment will be associated with Figure 2 The "sensor automatic calibration" functional unit, "processing data acquisition" functional unit, and "processing data enhancement" functional unit in the "intelligent algorithm layer" are used to complete the "parameter prediction model design" task, "parameter prediction model training" task, and "parameter prediction model deployment" task.

[0079] (2) The parameter prediction module is used to input the current production environment data and current workpiece data acquired by the sensor into the parameter prediction model, and predict and output n parameter prediction results through the parameter prediction model; once the parameter prediction model has completed training and optimization, the parameter prediction model will be deployed to the edge device on the production line; the parameter prediction module is responsible for inputting the data collected by the sensor in real time (i.e., the current production environment data and the current workpiece data) into the parameter prediction model for reasoning, so as to obtain the parameter prediction results of the theoretically optimal control in the current processing process. Through the reasoning function of the parameter prediction module, the system can obtain the predicted values ​​of key parameters in the processing process in real time.

[0080] (3) The control module of the middleware is used to convert the n parameter prediction results into processing control commands and control the execution mechanism to execute the processing steps related to the processing control commands.

[0081] The control module of the middleware serves to connect the parameter prediction model and the actual processing equipment. It is responsible for communicating with the actual processing equipment and converting the output results of the parameter prediction model into actual processing control commands. The middleware and the control module of the middleware interact with the interface of the processing equipment to ensure the processing control generated in real time based on the parameter prediction results. A processing equipment may have multiple different processing types, and different parameter prediction results correspond to different target processing types. In the process of using the middleware technology, the edge device first determines the target processing type corresponding to the parameter prediction result, and determines the corresponding target actuator based on the target processing type; then converts the parameter prediction result into a processing control command executable by the target actuator; and then inputs the processing control command into the target actuator to realize processing control.

[0082] In the specific implementation, Figure 2 The “control layer” shown, based on the middleware, its control module may include: motion trajectory optimization, component dynamic configuration, component state management, joint control, actuator control, and chassis control and other functions.

[0083] (4) The perception module is used to obtain actual finished product quality information during the actual processing through sensors when the processing step is executed; the optimization module is used to adjust the parameter prediction model based on the actual finished product quality information.

[0084] The perception module includes sensors deployed on the processing production line, which are used to perceive the production environment data and the actual finished product quality information during the processing in real time. In the case of the current production environment data and the current processing part data, each group of actual finished product quality information is compared, and the group of data with the best control parameters is selected from the comparison results as the actual optimal parameters. By comparing the actual optimal parameters with the prediction results of the parameter prediction model, the system can detect the differences and deviations in the processing process in real time. Based on these differences and deviations, the optimization module can perform real-time feedback optimization, adjust the parameter prediction model or processing control command, and achieve more accurate processing control and optimization.

[0085] This embodiment uses advanced machine learning and middleware control technology to simultaneously predict multiple processing parameters (i.e., parameter prediction results) and convert them into control commands that can be directly executed by the actuator. Through real-time online perception and feedback optimization, the system can improve the robustness and accuracy of processing control, thereby improving production efficiency and product quality.

[0086] Furthermore, based on the above-mentioned embodiment 1, the present invention proposes a specific application scenario embodiment of the control optimization system:

[0087] like Figure 3 As shown, the system can be applied to intelligent manufacturing process control scenarios.

[0088] Specifically, the system can be applied to laser welding scenarios. In the algorithm design stage, firstly, calibrated sensors are used to collect multiple sets of historical processing condition information, including historical production environment data (such as temperature, humidity) and historical processing part data (such as material thickness); then, multiple sets of historical processing control parameters (such as laser energy, welding speed, shielding gas flow rate) and finished product quality information (such as weld width, weld penetration) are collected under different production environment conditions and processing part scenarios. According to the finished product quality information in the above historical data, the optimal control parameters under different environmental conditions and processing part scenarios are determined to obtain a training data set. Secondly, a generative model is used to learn the data distribution of the training data set, and identically distributed data is generated to perform data enhancement on the training data set to obtain an enhanced target training data set. Among them, the identically distributed data is data related to the identical distribution of data. It can be understood that in the field of machine learning, data identical distribution (Data from the same distribution) means that the data sets used in the training and testing processes have similar probability distributions;

[0089] After that, for the enhanced target training dataset, call Figure 2The "parameter prediction model design" function of the "intelligent algorithm layer" shown automatically designs a lightweight, high-accuracy control parameter prediction model. Finally, the "parameter prediction model training" function of the "intelligent algorithm layer" is called to train the model, and the "parameter prediction model deployment" function of the "intelligent algorithm layer" is called to deploy the trained parameter prediction model in the edge device on the production line.

[0090] In the process control stage, the calibrated sensors are first used to collect the current production environment data and the current workpiece data in real time, and then input them into the deployed parameter prediction model to predict the optimal control parameters. After that, the proposed middleware technology is used to convert the parameter prediction results into executable process control commands for the welding robot arm, and the process control commands are input into the welding robot arm to realize process control. At the same time, the sensors will continuously collect finished product quality information. When continuous low-quality finished products are detected, the system will fine-tune the control parameter prediction model based on recent historical data, so as to dynamically adapt to changes and fluctuations in the process.

[0091] Further, based on the above-mentioned embodiment 1, the present invention proposes an embodiment of a control optimization method for an intelligent manufacturing process as embodiment 2:

[0092] The control optimization method for the intelligent manufacturing process in the second embodiment can obtain the production environment and workpiece data in the production process by calling the perception module of the first embodiment, and call the optimization module of the first embodiment to build or optimize the parameter prediction model based on the multi-granularity neural network architecture search technology.

[0093] It is understandable that the neural network search methods of the prior art cannot achieve the best trade-off between model performance and model size. The prior art either only performs coarse-grained operation-level search, which leads to redundant model parameters, or restrictively explores fine-grained filter-level and weight-level searchable units with predefined residual ratios, which has the problem of over-pruning. In view of the above defects of the prior art, this embodiment 2 proposes to better balance the performance and parameter amount of the model by comprehensively exploring the searchable units of different granularities in the search space and optimizing the relative remaining number of units of different granularities, so as to deploy the model on edge devices with limited resources in the production line.

[0094] like Figure 3 As shown, the DARTS (Differentiable ARchiTecture Search) search space of the prior art consists of multiple blocks, each of which is a directed acyclic graph containing an ordered sequence of M computing nodes. Represents the feature graph, each directed edge (i, j) represents the i to x jThe information transformation. It is necessary to determine the operation o on each edge, where the operation types include convolution with different kernel sizes, pooling, and skip connections. During the search process, DARTS builds a supernetwork that contains all operations and combines them using a differentiable architectural parameter α. In the supernetwork, the output of each node is the weighted sum of all operations connected to it, which is expressed by the mathematical formula: Where O represents the candidate operation set, o represents the candidate operation, represents the architectural parameters of operation o on edge (i,j), represents the scaled architecture parameters. In this way, the problem of selecting operation o is transformed into optimizing and discretizing the scaled architecture parameters problem.

[0095] On this basis, the second embodiment proposes a multi-granularity search space, which includes not only operation-level searchable units, but also filter-level and weight-level searchable units, such as Figure 4 As shown. Specifically, in addition to the operation type, this embodiment also needs to determine the number of filters and the sparsity pattern in the convolution operation. The problem of searching for a suitable number of filters and sparsity pattern can be viewed as the problem of selecting important subsets of filters and weights from a hypernetwork. The existing differentiable architecture parameter α can be viewed as an operation-level granularity parameter that represents the importance of the operation. On this basis, this embodiment introduces a differentiable filter-level granularity parameter labeled β to represent the importance of each filter. For a single weight, this embodiment directly uses the size ω of the weight itself to represent its importance. Expressed mathematically, for each convolution operation o(x i ) Among them C in is the input feature map, f is the filter, is the feature map generated by f, ω f represents the weight in the filter, β f represents the particle size parameter of f, σ(β f ) is the scaled granularity parameter. In this hypernetwork, the present invention needs to optimize and discretize three levels of scaled granularity parameters, namely σ(α), σ(β) and ω.

[0096] As mentioned above, the main focus of exploring the multi-granularity search space is the joint optimization and discretization of the scaled granularity parameters, namely σ(α), σ(β) and ω. Previous studies have focused on improving joint optimization to accurately compare the importance of units within each granularity level. The prior art adaptively updates the granularity parameters of different levels, taking into account the impact of changes in units at other levels on the importance of different units at this level. However, the prior art independently discretizes the granularity parameters of different levels at a fixed pruning rate, which has the problem of over-pruning and leads to a decrease in model performance. In this case, the embodiment of the present invention focuses on improving the joint discretization through adaptive pruning to adaptively determine the relative number of remaining units at different levels according to the evolving network architecture, so as to better balance the model performance and parameter amount. To achieve this, the present embodiment needs to compare the importance between units of different granularities. Considering that the granularity parameters of different levels are calculated at different scales, direct numerical comparison is not feasible. Therefore, the present embodiment proposes to dynamically learn pruning criteria of different granularities to determine when the unit can be considered unimportant within its own granularity level as the architecture evolves, thereby implicitly realizing the comparison of unit importance across granularities.

[0097] Specifically, this embodiment learns the granularity discretization function M at each level during the super network optimization process. α (·), M β (·), M ω (·), instead of manually defining the pruning rate, such as Figure 4 The discretization function aims to return the scaled granularity parameters σ(α), σ(β), ω to zero when their absolute values ​​are small enough. This embodiment expresses them by the following formula:

[0098] M α (σ(α))=S(|σ(α)|-t α )·σ(α),

[0099] M β (σ(β))=S(|σ(β)|-t β )·σ(β),

[0100] M ω (ω)=S(|ω|-t ω )·ω,

[0101] Where S(·) is a binary step function, t α , t β , t ω is a trainable threshold. Considering the differences between different operations, this embodiment discretizes the function M α (·), M β (·), M ω (·) Refine to the operational level, i.e. Applied to the forward propagation of the hypernetwork, we have

[0102]

[0103] In each convolution operation o, there is

[0104]

[0105] By learning the discretized function M α (·), M β (·), M ω (·), this embodiment adaptively obtains the pruning criterion of each granularity according to the evolving network architecture, so that the unit residual rate can be better determined.

[0106] In order to make the search process more stable, this embodiment uses a step-by-step pruning strategy, and each step will prune and remove the branches that have passed M. α (·), M β (·) or M ω (·) Units whose scaling granularity parameters σ(α), σ(β), or ω become zero after the calculation update. For operation-level units, the pruning operation will be replaced by a parameter-free zero operation that changes the input to zero. Considering the existence of skip connection operations, this embodiment aligns the number of filters in the output layer of each convolution operation in the same block and sets them together to the average number of output layer filters in all convolution operations in the block.

[0107] The parameters, inputs and outputs of the specific algorithm of this method include:

[0108] Input: dataset D, pruning interval e, target model parameter Param min , the initial remaining unit U act ; Parameters: granularity parameters α, β, ω, threshold t α ,t β ,t ω ;

[0109] Output: Network architecture U consisting of the remaining units act ;

[0110] The source code of the relevant algorithm is:

[0111]

[0112] The first line of code is interpreted as "initialize the number of iterations to 0";

[0113] The second line of code is interpreted as "When the total number of parameters of the remaining units is greater than or equal to the target number of parameters:", the third line of code is interpreted as "Increase the number of iterations by one", the fourth line of code is interpreted as "Update the loss function", and the fifth line of code is interpreted as "According to the loss L, use the gradient descent method to update the granularity parameter The explanation of line 6 is “According to the loss L, update the threshold t using the gradient descent method” α ,t β ,t ω ”;

[0114] The explanation of line 7 is “For each operation in the remaining units:”, and the explanation of line 8 is “Use the operation-level discretization function corresponding to this operation Update the scaled operation-level granularity parameter σ(α o )”, the explanation of line 9 is “When the number of iterations is a multiple of the pruning interval e:”, the explanation of line 10 is “If the scaled granularity parameter σ(α o ) is 0, then: ", the interpretation of the 11th line of code is "remove the operation from the remaining units", the interpretation of the 15th line of code is "if the operation is not removed, then for each filter in the operation:", the interpretation of the 16th line of code is "use the filter-level discretization function corresponding to the operation Update the scaled filter-level granularity parameter σ(β f )”, the explanation of line 17 is “When the number of iterations is a multiple of the pruning interval e:”, the explanation of line 18 is “If the scaled granularity parameter σ(β f ) is 0, then: ", the interpretation of the 19th line of code is "remove the filter from the remaining units", the interpretation of the 23rd line of code is "if the filter is not removed, then for each weight in the filter:", the interpretation of the 24th line of code is "use the weight level discretization function corresponding to this operation Update the weight-level granularity parameter corresponding to the weight (here is the size of the weight itself) ω w ”, the explanation of line 25 is “When the number of iterations is a multiple of the pruning interval e:”, the explanation of line 26 is “If the granularity parameter ω w is 0, then: ", the interpretation of line 27 is "remove this weight from the remaining units".

[0115] Accordingly, the operation of the second embodiment includes the following steps:

[0116] Step a: construct a super network with three types of granularity parameters, namely, operation level, filter level, and weight level, according to the set neural network search space;

[0117] Step b: performing enhancement processing on the historical production environment data and the historical processed part data to obtain target production environment data and target processed part data; inputting the target production environment data and the target processed part data into the super network;

[0118] Step c: using the hypernetwork to predict optimal control parameters, calculating the error between the predicted parameters and the actual optimal parameters and the parameter quantity of the current hypernetwork, and updating the loss function;

[0119] Step d: using a gradient descent algorithm to update the granularity parameter values ​​of each level of the hypernetwork and the corresponding granularity parameter discretization functions of each level according to the updated loss function, including:

[0120] The granularity parameter values ​​and discretization functions of the operation level type are updated according to the updated loss function, the granularity parameter values ​​and discretization functions of the filter level type are updated according to the updated loss function, and the granularity parameter values ​​and discretization functions of the weight level type are updated according to the updated loss function.

[0121] Step e: applying the updated granularity parameter discretization function of each level type to the granularity parameter value of the corresponding level, and pruning and removing the searchable units whose granularity parameter value in the hypernetwork becomes a preset value (the preset value in this embodiment is 0 as an example);

[0122] Step f: Repeat steps c to e to determine whether the parameter amount of the hypernetwork is less than the target parameter amount, until the parameter amount of the hypernetwork reaches the target parameter amount, and use the hypernetwork that meets the target as the parameter prediction model.

[0123] The method of this embodiment can obtain a series of lightweight parameter prediction models of different scales, and the parameters can predict more accurate processing parameters. The parameter prediction model can be deployed in edge devices with different storage capacities on the production line to recommend and output optimal control parameters based on real-time production environment data and workpiece data.

[0124] The above content is a further detailed description of the present invention in combination with specific preferred implementation modes, and it cannot be determined that the specific implementation of the present invention is limited to these descriptions. For ordinary technicians in the technical field to which the present invention belongs, without departing from the concept of the present invention, several simple deductions or substitutions can be made, which should be regarded as belonging to the protection scope of the present invention. That is, a processing control optimization method for intelligent manufacturing proposed in the present invention can be applied to other intelligent manufacturing processes. In any scenario involving limited computing power and limited storage memory, the neural network architecture search method mentioned in the present invention can be used to construct a parameter prediction model running on a computing device that meets these constraints.

Claims

1. A control optimization method for intelligent manufacturing process, characterized in that: The method is applied to edge equipment of a processing production line, and the method comprises: Obtain current production environment data and current workpiece data; Inputting the current production environment data and the current workpiece data into a parameter prediction model, and predicting and outputting n parameter prediction results through the parameter prediction model, where n is a positive integer; Convert the n parameter prediction results into processing control commands, and control the execution mechanism to execute processing steps related to the processing control commands; Wherein, the parameter prediction model is constructed or adjusted in the following manner: Collecting multiple sets of historical processing condition information, wherein the historical processing condition information includes historical production environment data and historical processing part data; Based on different historical processing condition information, multiple groups of historical processing control parameters and finished product quality information are collected; According to the finished product quality information, determining the optimal control parameters under different conditions of historical production environment data and historical processed part data to obtain a training data set; Using a generative model to learn the data distribution of the training data set and generate data with the same distribution, so as to perform data enhancement on the training data set to obtain an enhanced target training data set; Building a parameter prediction model based on the target training data set; Wherein, the parameter prediction model is constructed or optimized in the following manner: Step a: constructing a super network including parameters of multiple levels of granularity according to the set neural network search space, wherein the types of the levels include operation level type, filter level type and weight level type; Step b: performing enhancement processing on the historical production environment data and the historical processed part data to obtain target production environment data and target processed part data; inputting the target production environment data and the target processed part data into the super network; Step c: using the hypernetwork to predict optimal control parameters, calculating the error between the predicted parameters and the actual optimal parameters and the parameter quantity of the current hypernetwork, and updating the loss function; Step d: using a gradient descent algorithm to update the granularity parameter values ​​of each level of the hypernetwork and the corresponding granularity parameter discretization functions of each level according to the updated loss function; Step e: applying the updated granularity parameter discretization function of each level type to the granularity parameter value of the corresponding level, and pruning and removing the searchable units whose granularity parameter value in the hypernetwork is a preset value; Step f: Repeat steps c to e until the parameter amount of the hypernetwork reaches the target parameter amount, and use the hypernetwork that reaches the target as the parameter prediction model.

2. The method according to claim 1, characterized in that The method comprises: When the processing steps are performed, monitoring actual finished product quality information during the actual processing; The parameter prediction model is adjusted based on the actual finished product quality information.

3. The method according to claim 2, characterized in that The adjusting the parameter prediction model based on the actual finished product quality information comprises: In the case of the current production environment data and the current workpiece data, each group of actual finished product quality information is compared, and the group of data with the best control parameters is selected from the comparison results as the actual optimal parameters; Comparing the actual optimal parameters with the parameter prediction results to obtain differences and deviations in the actual processing; Based on the differences and deviations, the parameter prediction model or the process control command is adjusted.

4. The method according to claim 1, characterized in that The step of converting the n parameter prediction results into processing control commands and controlling the execution mechanism to execute processing steps related to the processing control commands includes: Determine the target processing type corresponding to the parameter prediction result; Determining a corresponding target actuator based on the target processing type; Converting the parameter prediction results into processing control commands executable by the target actuator; The processing control command is input into the target execution mechanism to realize processing control.

5. A control optimization system for intelligent manufacturing process, characterized in that: The system includes a parameter prediction module and a middleware control module, wherein the parameter prediction module includes a parameter prediction model: The parameter prediction module is used to input the current production environment data and the current workpiece data acquired by the sensor into the parameter prediction model, and predict and output n parameter prediction results through the parameter prediction model, where n is a positive integer; The control module of the middleware is used to convert the n parameter prediction results into processing control commands, and control the execution mechanism to execute the processing steps related to the processing control commands; Wherein, the system further comprises a model building module; The model building module is used to train and build the parameter prediction model using historical production environment data and historical workpiece data; Wherein, the parameter prediction model is constructed or adjusted in the following manner: Collecting multiple sets of historical processing condition information, wherein the historical processing condition information includes the historical production environment data and the historical processing part data; Based on different historical processing condition information, multiple groups of historical processing control parameters and finished product quality information are collected; According to the finished product quality information, determining the optimal control parameters under different conditions of historical production environment data and historical processed part data to obtain a training data set; Using a generative model to learn the data distribution of the training data set and generate data with the same distribution, so as to perform data enhancement on the training data set to obtain an enhanced target training data set; Building a parameter prediction model based on the target training data set; Wherein, the parameter prediction model is constructed or optimized in the following manner: Step a: constructing a super network including parameters of multiple levels of granularity according to the set neural network search space, wherein the types of the levels include operation level type, filter level type and weight level type; Step b: performing enhancement processing on the historical production environment data and the historical processed part data to obtain target production environment data and target processed part data; inputting the target production environment data and the target processed part data into the super network; Step c: using the hypernetwork to predict optimal control parameters, calculating the error between the predicted parameters and the actual optimal parameters and the parameter quantity of the current hypernetwork, and updating the loss function; Step d: using a gradient descent algorithm to update the granularity parameter values ​​of each level of the hypernetwork and the corresponding granularity parameter discretization functions of each level according to the updated loss function; Step e: applying the updated granularity parameter discretization function of each level type to the granularity parameter value of the corresponding level, and pruning and removing the searchable units whose granularity parameter value in the hypernetwork is a preset value; Step f: Repeat steps c to e until the parameter amount of the hypernetwork reaches the target parameter amount, and use the hypernetwork that reaches the target as the parameter prediction model.

6. The system according to claim 5, characterized in that The system also includes a perception module and an optimization module; The perception module is used to obtain actual finished product quality information during the actual processing through sensors when the processing step is executed; The optimization module is used to adjust the parameter prediction model based on the actual finished product quality information.

7. The system according to claim 6, characterized in that: The optimization module is further used to compare each group of actual finished product quality information under the conditions of the current production environment data and the current workpiece data, and select the group of data with the best control parameters from the comparison results as the actual optimal parameters; compare the actual optimal parameters with the parameter prediction results to obtain the differences and deviations in the actual processing process; and adjust the parameter prediction model or the processing control command based on the differences and deviations.

Images