Systems, integrated management devices, edge devices

Abstract

To improve the inference accuracy of an edge device.SOLUTION: A system according an embodiment comprises: one or more edge devices; and an integrated management device for managing the edge devices. The edge devices include light detection devices. The integrated management device manages a first trained model in association with a first condition that sets the conditions of the light detection devices used to generate the first trained model. The integrated management device is configured to be able to deliver the first trained model and the first condition to the edge devices.SELECTED DRAWING: Figure 1

Description

[Technical Field]

[0001] The present invention relates to an integrated management device and a system including the integrated device. [Background technology]

[0002] Image recognition and classification processing using machine learning and deep learning are used in various technical fields. Patent Document 1 describes a repository service for managing and distributing machine learning algorithm data (trained models) in a container. [Prior art documents] [Patent documents]

[0003] [Patent Document 1] US Patent Application Publication No. 2019 / 0278640 Summary of the Invention [Problem to be solved by the invention]

[0004] The management system described in Patent Document 1 facilitates the operation of machine learning by providing a machine learning repository service. However, when using one or more edge devices, each of which includes a photodetector, it is expected that the conditions set for each edge device will be different. Since the conditions set for each edge device are different, there is a possibility that the expected inference accuracy will not be achieved when considering distribution to the edge devices.

[0005] The present invention aims to improve the inference accuracy in edge devices. [Means for solving the problem]

[0006] A system according to one embodiment includes one or more edge devices and an integrated management device that manages the edge devices, wherein the edge devices include a photodetector, and the integrated management device manages a first trained model in association with a first condition that sets the conditions of the photodetector used to generate the first trained model, and the integrated management device is configured to be able to distribute the first trained model and the first condition to the edge devices.

[0007] A system in one embodiment includes two or more edge devices and an integrated management device that manages the two or more edge devices, wherein the edge devices include a photodetection device, and the integrated management device manages a first trained model in association with a first condition that sets the conditions of the photodetection device used to generate the first trained model.

[0008] In one embodiment, an edge device includes a photodetector, and a first trained model and conditions set when generating the first trained model are delivered to the edge device. [Effects of the Invention]

[0009] According to the present invention, it is possible to improve the inference accuracy in edge devices. [Brief explanation of the drawings]

[0010] [Figure 1] FIG. 1 is a block diagram showing an edge device management system according to a first embodiment. [Figure 2] FIG. 1 is a block diagram showing a first embodiment and a comparative example. [Figure 3] FIG. 2 is a flow chart showing the operation of the first embodiment. [Figure 4] 1 is an example of a container according to the first embodiment. [Figure 5] FIG. 10 is a block diagram showing an edge device management system according to a second embodiment. [Figure 6] FIG. 10 is a block diagram showing an edge device management system according to a third embodiment. [Figure 7]FIG. 10 is a block diagram showing an edge device management system according to a fourth embodiment. [Figure 8] FIG. 10 is a conceptual diagram showing an application of the fifth embodiment. [Figure 9] FIG. 2 is a diagram illustrating an example of a container according to the first embodiment. [Figure 10] FIG. 4 is a diagram showing another example of the container of the first embodiment. [Figure 11] FIG. 10 is a flow diagram illustrating an example of a deployment flow to an edge device according to the first embodiment. DETAILED DESCRIPTION OF THE INVENTION

[0011] The embodiments shown below are intended to embody the technical concept of the present invention and are not intended to limit the present invention. Note that the size and positional relationship of components shown in each drawing may be exaggerated for clarity. In the following description, the same components are designated by the same reference numerals and their description will be omitted.

[0012] In the following, when the explanation is common, the subscripts a, b, etc. will be omitted.

[0013] (Embodiment 1) An edge device management system (hereinafter referred to as "system") according to a first embodiment will be described with reference to FIGS. 1 to 4 and 9 to 11. FIG.

[0014] FIG. 1 is a diagram showing the basic configuration of the system. The system includes one or more edge devices 300 and an integrated management device 200. The edge devices 300 and the integrated management device 200 are connected so that trained models managed by the integrated management device 200 can be deployed to the edge devices 300. The edge devices and the integrated management device may be located in physically separate locations. It is only necessary that data such as trained models can be deployed from the integrated management device 200 to the edge devices 300 using wireless or wired communication. For example, the edge devices 300 may be located in a factory, and the integrated management device 200 may be located in a management center located in a remote location. Furthermore, one of the edge devices 300 and the integrated management device 200 may be located in one country, and the other may be located in a different country.

[0015] The system can be used, for example, as an inspection system. The following describes a case where the system is an inspection system. The systems according to each embodiment can be employed in various systems in addition to inspection systems. Examples of various systems include an image recognition system that identifies whether or not a specific object exists in image data, and an automatic distribution system for a distribution center.

[0016] The system according to this embodiment will be described below with reference to FIG.

[0017] (Edge Device 300) The edge device 300 includes a light detection device 301 and a computer device 302. The computer device 302 includes at least an input unit, a storage unit, and an output unit. Information related to the control of the light detection device 301 is transmitted from the integrated management device 200 to the input unit of the computer device 302.

[0018] For example, an image sensor, a photometric sensor, or a distance measuring sensor can be used as the light detection device 301. In the following, a case where the light detection device 301 is an image sensor will be described.

[0019] The computer device 302 controls the light detection device 301. The computer device 302 can further store the trained model distributed from the integrated management device 200 in a storage unit. The computer device 302 can also store the setting conditions for creating the trained model and the obtained data in a storage unit. In FIG. 1, the integrated management device 200 manages the trained model and imaging conditions in a container, and the container is deployed to the computer device 302. The light detection device 301 is controlled based on the container information. A container refers to a virtually constructed "execution environment for a specific application." In this embodiment, a container is an execution environment for capturing images using the light detection device 301 and inferring trained data. By introducing a container, this execution environment can be quickly and easily constructed, reducing the burden of environment management. Specifically, when not managed using a container as in this case, multiple applications are linked. Therefore, updating a specific application may cause problems in other applications. By introducing a container, the possibility of problems occurring can be reduced and the management burden can be reduced. In the following, we will explain the case where a container is used, but methods other than containers may be used as long as the imaging conditions and the trained model can be linked and managed.

[0020] (Integrated management device 200) The integrated managing device 200 controls one or more edge devices 300. In FIG. 1 , the integrated managing device 200 controls two or more edge devices 300, but may also control one edge device 300. The integrated managing device 200 includes at least an edge device environment providing unit 210. In FIG. 1 , the integrated managing device 200 further includes a trained model database 220, an imaging condition database 230, an integrated managing device endpoint 240, and a training execution unit 250.

[0021] The trained model database 220 manages multiple trained models generated under multiple imaging conditions, such as a first trained model of a first object (first work) generated under first imaging conditions and a second trained model of a first work generated under second imaging conditions. In other words, it includes a first trained model and a second trained model generated from the same work under different imaging conditions. The same work does not necessarily have to be an identical work. For example, if the work is product A, multiple product A's correspond to the same work. Specifically, if product A is a red ink tank, a red ink tank may be used as the work, and the same type of red ink tank may be used as the same work. The trained model database 220 may also manage a third trained model of a second object (second work) generated under first imaging conditions. In other words, the trained model database 220 may include multiple trained models obtained by imaging the same object under different imaging conditions, or multiple trained models obtained by imaging different objects under the same imaging conditions. Images of different objects may also be captured under different imaging conditions. It may also contain both.

[0022] The imaging condition database 230 manages each imaging condition when generating a trained model managed in the trained model database 220. The integrated management device 200 is configured to be able to distribute the trained model and the imaging conditions when generating the trained model.

[0023] The edge device environment providing unit 210 distributes the associated trained model and imaging conditions to the edge device 300. In FIG. 1, the trained model and imaging conditions are associated and managed by the container 211, and the container 211 is deployed to the edge device 300. That is, the container 211 manages the trained model and the imaging conditions for capturing the trained model in association with each other. For example, in FIG. 1, the containers 211a to 211c are each captured under the same imaging conditions. The imaging conditions include, for example, exposure time and gain. Specifically, the imaging conditions are an exposure time of 1 ms and a double gain, and are managed by the containers 211a to 211c. The containers 211a to 211c may be trained models of the same workpiece, or may be trained models of different workpieces as described in the second embodiment. The imaging conditions may also include various conditions such as ISO sensitivity setting, F-number setting, whether or not a high dynamic range is used, and white balance adjustment. The imaging conditions are preferably, but not limited to, the conditions of a photodetector included in an edge device managed by the integrated management device. For example, a trained model may be formed based on an image obtained by a photodetector included in an edge device not managed by the integrated management device, and the imaging conditions and trained model at this time may be managed in a container.

[0024] 9, the container 211 may include an inference application 308 and a machine learning library 310 in addition to imaging parameters 307 and a trained model 309. When the container 211 is deployed in the computer device 302, the edge device environment acquisition unit 311 acquires the imaging conditions, firmware version, installation information, etc. of the edge device 300 through the imaging application 312 and the imaging library 313. Then, the imaging parameters 307 in the container are referenced, and consistency is confirmed before the container 211 is deployed.

[0025] In this case, even if an inconsistency occurs, the information in the edge device 300 can be overwritten and used in accordance with the flow shown in FIG.

[0026] FIG. 11 shows an example of a deployment flow. In step S1, deployment of the container 211 to the edge device computer 302 is initiated. In step S2, the integrated management device 200 acquires edge device information through the edge device environment acquisition unit 311. Then, in step S3, consistency between the edge device information and the container information is confirmed. This information includes, for example, the installation information, library, and firmware version of the imaging device. If consistency is confirmed in step S3, the process proceeds to container deployment (step S6) and ends (step S7). If a mismatch is found in step S3, the process confirms whether to proceed with the update procedure to match the edge device with the container (step S4). If the operator does not agree, the process ends (step S7). If the operator agrees, the edge device information is updated (step S5). Update methods include automatic updates via electronic media or communication, as well as manual updates. If a manual update is required, a method of notifying the operator and issuing an instruction to do so is included. When the edge device update is complete, the container is deployed to the edge device (step S6) and ends (step S7).

[0027] Note that the container 211 may also be configured without including a machine learning library to reduce its weight, as shown in FIG. 10 . That is, the container 211 includes the inference application 308 and the trained model 309, and the machine learning library 310 is located outside the container 211. In this case, the inference application 308 and the trained model 309 must be consistent with the machine learning library 310, which can be resolved by adding library parameters 314. Library parameters generally have a smaller capacity than the machine learning library 310. Therefore, even if the container 211 includes the library parameters 314 instead of the machine learning library 310, the container 211 can be made lighter in weight than when it includes the machine learning library 310.

[0028] When deploying a container to an edge device, if there is no identical container information and no inconsistency between the container information and the edge device information, similar container information or optimal container information may be selected and distributed.

[0029] When the operator 700 inputs information about a workpiece to be inspected by each edge device 300 into the integrated management device endpoint 240 of the integrated management device 200, the workpiece information is transmitted from the integrated management device endpoint 240 to the edge device environment providing unit 210. In other words, the operator 700 inputs information for changeover into the integrated management device endpoint 240 of the integrated management device 200. Then, a trained model of the target workpiece and the imaging conditions under which the trained model was generated are transmitted to the edge device environment providing unit 210, where they are linked and managed by the container 211. The target container 211 is then distributed to the corresponding edge device 300. For example, when the same workpiece is inspected by edge devices 300a, 300b, and 300c, information about the container 211a is input to the computers 302a, 302b, and 302c. Alternatively, the container 211a may be delivered to the computer device 302a, and the containers 211b and 211c that manage the same information as the container 211a may be delivered to the computers 302b and 302c, respectively.

[0030] Simply inputting a trained model corresponding to a workpiece to be inspected into the edge device 300 may result in inaccurate inspection due to differences in settings such as imaging conditions between the edge device that generated the trained model and the edge device that performs the inspection. In this embodiment, not only the trained model but also the imaging conditions for imaging the trained model are input from the integrated management device 200 to the edge device 300. This allows the trained model to be distributed to the edge device 300, and when inference is performed, imaging can be performed under the same imaging conditions as when the trained model was generated, improving inference accuracy.

[0031] The effects of this embodiment will be described using FIG. 2. It is assumed that a workpiece has a scratch 100 and that the workpiece is to be inspected for defects. In each figure, it is assumed that the workpiece is moving from top to bottom. For example, FIG. 2(a) shows an image captured by edge device 300a that generated a trained model. FIG. 2(b) shows an image captured by edge device 300b, to which a trained model has been input, but no imaging conditions have been input. FIG. 2(c) shows an image captured by edge device 300c, to which a trained model has been input, but no imaging conditions have been input. FIG. 2(d), FIG. 2(e), and FIG. 2(f) show images captured by edge devices 300a, 300b, and 300c, to which a trained model and imaging conditions have been input, respectively. Because FIG. 2(a) was captured under appropriate conditions, the scratch 100a on the workpiece can be detected. However, because FIG. 2(b) was captured under improper exposure conditions and blurred, the scratch 100b cannot be correctly detected. In addition, in Fig. 2(c), the gain is inappropriate and the image is overexposed, making it impossible to detect the scratch 100c. On the other hand, as shown in Fig. 2(d) to Fig. 2(f), an edge device that has input the trained model and imaging conditions can accurately capture images and detect scratches.

[0032] Optimal conditions vary depending on the workpiece. For example, in FIG. 1, a case will be described in which the imaging conditions when the trained model was generated were condition A and the standard imaging conditions were condition B. In this case, edge device 300a can capture images under optimal conditions because it has condition A when the trained model was generated and the trained model. On the other hand, edge device 300b is under condition B, so it cannot achieve optimal conditions even if only the trained model is input, and inspection accuracy may be lower than that of edge device 300a. In contrast, in this embodiment, the optimal conditions for product inspection and the trained model of the workpiece are managed in association with each other, and both are input to the edge device. Therefore, each edge device can achieve inspection accuracy equivalent to that of edge device 300a.

[0033] Next, an execution flow when a trained model trained by the edge device 300a is used by other edge devices 300b and 300c will be described with reference to FIGS. 1 and 3.

[0034] As shown in Figure 3, the execution flow is divided into a learning flow and a deployment flow.

[0035] First, the learning flow will be described. An image of a workpiece is captured under a first condition using the photodetector 301a of the edge device 300a (step S101). The first condition may be, for example, an exposure time of 1 ms and a gain of 2 times. The captured image is sent to the learning execution unit 250 of the integrated management device 200.

[0036] Learning is performed using the sent image, and a trained model is generated (step S102). The trained model can be generated by machine learning. When an image of a workpiece is input during inspection, the trained model determines whether there is a defect, determines whether it is good or bad, and outputs the result. The trained model is stored in the trained model database 220. The first condition is also stored in the imaging condition database 230. At this time, the trained model database 220 and the imaging condition database 230 are linked by a relational database or the like, and managed by the edge device environment providing unit 210 (step S103).

[0037] Specific machine learning algorithms that may be used include nearest neighbor methods, naive Bayes methods, decision trees, and support vector machines. Deep learning, which uses a neural network to generate features and connection weighting coefficients for learning, may also be used. For example, a CNN (Convolutional Neural Network) model may be used as a deep learning model.

[0038] When creating multiple trained models, the above steps S101 to S103 are repeated by changing the imaging conditions and the workpiece.

[0039] Next, the deployment flow will be described. First, the operator 700 accesses the integrated management device endpoint 240 using a REST API or the like and issues a command so that the trained model of the edge device 300a can be deployed to the other edge devices 300b and 300c (step S201). Note that instead of the operator 700 accessing the integrated management device endpoint 240, it is also possible to set up the system so that the trained model of the edge device 300a is deployed to the other edge devices 300b and 300c by programming. For example, the trained model may be scheduled using a script or the like and deployed every weekend.

[0040] Upon receiving a command from the integrated management device endpoint 240, the edge device environment providing unit 210 imports a trained model suitable for the edge device from the trained model database and the imaging conditions into the container as a set (step S202). At this time, the trained model is copied from the trained model database 220, and the imaging conditions are copied from the imaging condition database 230 and imported into the container 211. A container orchestration system for deploying, scaling, and managing containerized applications can be used for the mechanism and orchestration of the container 211. Any mechanism can be used as the container orchestration system as long as it can perform similar tasks.

[0041] Note that the container in which the trained model is stored and the inference container may be managed separately, and each container may be linked and managed.

[0042] Next, the central managing device 200 distributes the containers 211a, 211b, and 211c corresponding to the edge devices 300a, 300b, and 300c, respectively (step S203).

[0043] Then, each edge device performs imaging and inference (step S301). If the predetermined inference accuracy is ensured in step S301, the inspection is performed using the workpiece to be inspected. If the predetermined inference accuracy is not ensured in step S301, the imaging conditions are changed and steps S101 to S301 are executed again, and steps S101 to S301 are executed again until the predetermined inference accuracy is achieved.

[0044] According to the above execution flow, the trained model trained by the edge device 300a can be used by the other edge devices 300b and 300c.

[0045] FIG. 4 shows an example of a case where imaging conditions and a trained model are managed together in a container, but this is not limiting. As shown in FIG. 4, a system administrator executes Deployment.yaml, which manages imaging conditions and a trained model, via an API or the like. This yaml file is then used to deploy to an edge device. Specifically, in FIG. 4, imaging conditions such as exposure time, gain, lighting intensity of the lighting device, and lighting device angle are linked to the trained model and managed. Although the lighting device is not shown in FIG. 1, the angle and lighting intensity are input to the lighting device included in the edge device as shown in FIG. 8. The above imaging conditions and trained model are then input to the edge device, and object estimation is performed.

[0046] In the above description, the container 211a is also distributed from the integrated management device 200 to the edge device 300a. However, since the edge device 300a is the edge device used to generate the trained model, the container 211a does not need to be distributed.

[0047] In addition, an example has been described in which a trained model and imaging conditions are managed by a container and the container is distributed, but it is also possible to manage the trained model in the container and manage the imaging conditions separately from the container by linking them to the trained model of the container.

[0048] According to this embodiment, a trained model and the conditions for generating the trained model are linked and managed by the integrated management device 200. The trained model and conditions are then distributed to the edge device, which then performs imaging and inference. This improves the inference accuracy of the edge device. Furthermore, the dead time before the start of trained inspection can be shortened during workpiece inspection.

[0049] (Embodiment 2) An edge device management system according to a second embodiment will be described with reference to Fig. 5. The system according to the second embodiment differs from the first embodiment in that the multiple edge devices managed by the central management device 200 include an edge device that inspects a first work and an edge device that inspects a second work that is different from the first work. The configuration other than that described below is the same as that of the first embodiment, and therefore the description may be omitted.

[0050] As shown in FIG. 5, the integrated management device 200 manages edge devices 300a and 300c that inspect a first workpiece (product A) and edge devices 300b and 300d that inspect a second workpiece (product B). Containers suitable for the workpieces to be imaged by the edge devices are distributed to each edge device. For example, in FIG. 5, containers 211a and 211c that manage the trained model and imaging conditions corresponding to the first workpiece are distributed to the edge devices 300a and 300c. Containers 211b and 211d that manage the trained model and imaging conditions corresponding to the second workpiece are distributed to the edge devices 300b and 300d.

[0051] The method for creating the trained model and the mechanism for managing the trained model and imaging conditions are the same as in embodiment 1, so explanations will be omitted.

[0052] According to this embodiment, the time required to complete a changeover when the product to be inspected by the edge device changes can be shortened compared to when only a trained model is input. For example, an edge device may inspect product A for a certain period of time, and then inspect product B for a certain period of time using the same edge device. Because the optimal imaging conditions differ for each product, simply inputting a trained model corresponding to the product to be inspected into the edge device takes time for the imaging conditions of the edge device to become optimal. Therefore, time is required to complete a changeover each time the product to be inspected changes. In contrast, according to this embodiment, the trained model and imaging conditions are distributed to the edge device, thereby shortening the time required to achieve optimal imaging conditions and improving inference accuracy. Furthermore, it becomes possible to inspect different workpieces with each edge device.

[0053] (Embodiment 3) An edge device management system according to a third embodiment will be described with reference to Fig. 6. The system according to the third embodiment differs from the second embodiment in that each edge device 300 includes a learning execution unit 250. The configuration other than that described below is the same as that of the second embodiment, and therefore the description may be omitted.

[0054] In the system according to the third embodiment, each edge device 300a, 300b, and 300c includes a corresponding learning execution unit 250a, 250b, and 250c. The learning execution unit 250 included in the edge device 300 generates a trained model. Specifically, a trained model obtained from a light detection device 301 is input to the learning execution unit 250, and the trained model is generated by the learning execution unit 250.

[0055] The trained models generated by each edge device 300 are input to the trained model database 220 and the imaging condition database 230 of the integrated management device 200. The trained models of the multiple edge devices 300 and the imaging conditions under which the trained models were generated are collectively managed by the integrated management device 200.

[0056] An example of a flow in this embodiment is described below. First, the operator 700 issues an instruction via the integrated management device endpoint 240 as to whether or not learning should be performed in the learning execution unit 250, and as to whether or not to update the trained model.

[0057] Next, an image captured by the photodetector 301 of the edge device 300 is sent to the learning execution unit 250, where learning is executed.

[0058] Once learning has progressed to a certain extent and a predetermined accuracy has been achieved, the trained model is stored in the trained model database 220 of the integrated management device 200. At this time, if a trained model that was stored before the training was performed exists, it may be updated to the newly generated trained model. The predetermined accuracy can be set as appropriate, but for example, it is an accuracy that results in a pass / fail judgment probability of 80% or more in the inspection process.

[0059] Simultaneously with, or before or after, storing a trained model in the trained model database 220, the imaging conditions used to generate the trained model are also stored in the imaging condition database 230. The trained model database 220 and the imaging condition database 230 manage the corresponding trained model and imaging conditions in association with each other.

[0060] According to this embodiment, the trained model and the imaging conditions are managed as a set, and the trained model and the imaging conditions are distributed to the edge device, thereby improving the inference accuracy. Also, as in this embodiment, the learning execution unit 250 is configured as part of each edge device 300, and re-learning is performed in the edge device, thereby improving the accuracy of the trained model.

[0061] (Embodiment 4) An edge device management system according to the fourth embodiment will be described with reference to FIG. 7. The system according to the fourth embodiment differs from the first embodiment in that data from the photodetector 301 is input from the edge device 300 to the learning execution unit 250 via the preprocessing unit 303. The system also differs from the first embodiment in that data is input from the learning execution unit 250 to the trained model database 220 via the postprocessing unit 304. The system also differs from the first embodiment in that data is input from the preprocessing unit 303b to the inference execution unit 260 in the inference flow. Configurations other than those described below are the same as those of the second embodiment, and therefore descriptions thereof may be omitted.

[0062] 7, the learning flow is from the photodetector 301a to the post-processing unit 304a, and the inference flow is from the photodetector 301b to the post-processing unit 304b. The learning flow and the inference flow differ in the configuration in which data is input from the pre-processing unit.

[0063] The pre-processing unit 303 and the post-processing unit 304 may be included in the central managing device 200 or the edge device 300 .

[0064] The preprocessing unit 303 performs trimming, target region identification, inversion, correction, etc. on the image data obtained by the photodetection device 301. Corrections include, for example, averaging, brightness correction, and contrast correction. Edge emphasis may also be performed.

[0065] During learning in the learning execution unit 250, hyperparameters that control the machine learning algorithm may be additionally managed.

[0066] The post-processing unit 304 can include information on the pass / fail judgment of the inspection and production line information.

[0067] According to this embodiment, a system for managing imaging conditions and trained models can manage a combination of pre-processing and post-processing. This can improve inference accuracy and reproducibility. Additional learning can also be performed during training. Furthermore, the inclusion of a post-processing unit can be useful for annotation processing, etc. Therefore, it becomes easier to link with various management systems, such as industrial control systems (Supervisory Control and Data Acquisition) and manufacturing execution systems.

[0068] (Embodiment 5) An edge device management system according to the fifth embodiment will be described with reference to Fig. 8. The system according to the fifth embodiment differs from the first embodiment in that the edge devices include a lighting device 305 and a robot 306, and the integrated management device 200 manages not only the imaging conditions but also the operating conditions of devices other than the light detection device. The configuration other than that described below is the same as that of the first embodiment, and therefore the description may be omitted.

[0069] 8 shows a conceptual diagram of an edge device management system used as an inspection application. The management system includes a light detection device 301, a lighting device 305, a robot 306, and a production line 600.

[0070] The edge device 300 inspects the workpieces transported along the production line 600. The photodetector 301 of the edge device 300 captures an image of the workpiece. Light is irradiated into the imaging range of the photodetector by an illumination device 305. The integrated management device 200 determines whether the workpiece is defective based on the image data from the edge device 300. The robot 306 removes any workpieces determined to be defective from the production line.

[0071] When inspecting a workpiece, various conditions are attached to each component, such as the operating speed of the production line 600, the lighting intensity, angle, and color temperature of the lighting device 305, and the range of motion, operating angle, and angular velocity of the robot 306. At least one of these conditions is managed by the integrated management device 200. When distributing a trained model from the integrated management device 200 to the edge device 300, these conditions are distributed in addition to the imaging conditions, thereby enabling accurate imaging. Furthermore, by uniformly managing the range of motion, operating speed, etc. of the robot, it is possible to reduce dead time in the workpiece inspection process.

[0072] The features described in the above first to fifth embodiments can be combined as appropriate.

[0073] Furthermore, it is also possible to adopt the first embodiment in which the same workpiece is inspected by multiple edge devices during a predetermined period, and the second embodiment in which each edge device inspects a different workpiece during other periods. In other words, each edge device does not always inspect the same workpiece, but can change the workpieces and conditions selected as appropriate depending on the period.

[0074] In addition, in the above-described first to fifth embodiments, the light detection device 301 is an image sensor, and the integrated management device 200 manages the trained model and the imaging conditions. However, the conditions are not limited to the imaging conditions. The image parameters in FIG. 7 may be linked to the trained model and managed. In addition, if the light detection device 301 is a distance measurement sensor, conditions such as the pulse interval for measuring distance can be distributed to the trained model. [Explanation of symbols]

[0075] 200 Integrated management device 300 Edge Devices

Claims

1. A system comprising: a first edge device; a second edge device; and an integrated management device that manages the first edge device and the second edge device; the first edge device and the second edge device include a light detection device; the integrated management device manages a first trained model and first imaging conditions that set conditions of a light detection device used to generate the first trained model in a first container in association with each other; the first imaging condition is at least one of an exposure time, a gain, an ISO sensitivity setting, an F-number setting, whether or not a high dynamic range is used, and an adjustment of a white balance; The system, wherein the first container is delivered to the first edge device or the second edge device.

2. The system according to claim 1, characterized in that the integrated management device further manages a second trained model in association with a second imaging condition when generating the second trained model.

3. The first trained model and the second trained model are trained models based on the same object, 3. The system according to claim 2, wherein the first imaging condition and the second imaging condition are different imaging conditions.

4. The first trained model and the second trained model are trained models based on different objects, 3. The system according to claim 2, wherein the first imaging condition and the second imaging condition are the same imaging condition.

5. The first trained model and the second trained model are trained models based on different objects, 3. The system according to claim 2, wherein the first imaging condition and the second imaging condition are different imaging conditions.

6. the first container is delivered from the integrated management device to the first edge device; The system described in any one of claims 3 to 5, characterized in that a second container managing the second trained model and the second imaging conditions is delivered from the integrated management device to the second edge device.

7. the light detection device is an image sensor, The system according to any one of claims 1 to 6, wherein the first trained model is generated based on an image obtained from the image sensor of the first edge device.

8. The system of any one of claims 1 to 7, wherein the first container does not include a machine learning library.

9. the integrated management device includes a learning execution unit, The system according to any one of claims 1 to 8, characterized in that an image obtained from the first edge device is sent to the learning execution unit, and the first trained model is generated in the learning execution unit.

10. the integrated management device includes a first database and a second database; The first database manages a plurality of trained models including the first trained model and the second trained model; 7. The system according to claim 2, wherein the second database manages a plurality of conditions including the first imaging condition and the second imaging condition.

11. The system includes a production line; 11. The system according to claim 1, wherein the first imaging condition includes an operating speed of the production line.

12. the system comprises a robot; The system of claim 11 , further comprising: moving a workpiece determined to have a defect by the first trained model.

Images