Method for generating a pre-trained model, a risk estimation system, and a program.

Abstract

This invention provides a method for generating trained models that can facilitate the construction of systems that support hazard prediction activities in the field. [Solution] The method for generating a trained model comprises: a preparation step of preparing training images based on text included in the section for recording the cause of an accident in a report on an accident during on-site work; an assignment step of assigning labels indicating danger to the training images based on text included in the section for recording the cause of an accident; and a learning step of generating a trained model for estimating the degree of danger from images taken at the site using machine learning, with the training images as input data and the labels as training data.

Description

Technical Field

[0001] The present invention relates to a method for generating a learned model, a risk estimation system, and a program.

Background Art

[0002] In Patent Document 1, an image is analyzed to recognize an object, and among the risk information associated with the object, a certain number of risk information is transmitted starting from the ones with high importance. Based on the information about the operator, the importance corresponding to the remaining risk information is recalculated, and a certain number of risk information is transmitted starting from the ones with high importance among the remaining risk information. A system is disclosed.

Prior Art Documents

Patent Documents

[0003]

Patent Document 1

Summary of the Invention

Problems to be Solved by the Invention

[0004] However, in the above prior art, it is necessary to select risk information for each object in advance and judge its importance, etc., which places a heavy burden on the designer and makes it difficult to construct the system.

[0005] The present invention has been made in view of the above problems, and its main object is to provide a method for generating a learned model, a risk estimation system, and a program that can facilitate the construction of a system for supporting risk prediction activities at the site.

Means for Solving the Problems

[0006] To solve the above problems, a method for generating a trained model according to one aspect of the present invention comprises: a preparation step of preparing training images based on text contained in the field for recording the cause of an accident in a report concerning an accident during on-site work; an assignment step of assigning labels indicating danger to the training images based on the text contained in the field for recording the cause of an accident; and a learning step of generating a trained model for estimating the degree of danger from images taken at the site using machine learning, with the training images as input data and the labels as training data. This makes it possible to facilitate the construction of a system that supports hazard prediction activities. In other words, it is possible to generate training images for assigning labels indicating danger using text contained in the field for recording the cause of an accident, thereby reducing the effort required for assigning labels.

[0007] In the above embodiment, the preparation step may involve preparing a learning image based on the text included in the section for filling in countermeasures in the report, and the labeling step may involve labeling the learning image based on the text included in the section for filling in countermeasures with a label indicating safety. This makes it possible to generate a learning image for which a label indicating safety is to be added using the text included in the section for filling in countermeasures, thereby reducing the effort required for labeling.

[0008] In the above embodiment, the preparation step may involve preparing the training images based on the text and the figures included in the report. This makes it possible to generate more suitable training images.

[0009] In the above embodiment, the preparation step may involve preparing the learning images based on the text and reference images from the site. This makes it possible to generate more suitable learning images.

[0010] In the above embodiment, the learning step may use the learning image and process data representing the work content of the on-site work as input data. This makes it possible to train the trained model to estimate the degree of risk according to the work content.

[0011] In the above embodiment, the labeling step may involve labeling a portion of the training image with the label representing danger, and the training step may generate the trained model for estimating areas in the image where the danger level is above a predetermined level. This makes it possible to train the trained model to estimate areas in the image where the danger level is above a predetermined level.

[0012] Furthermore, another embodiment of the present invention provides a risk estimation system comprising a camera that captures images of a site and outputs images, and an estimation unit that estimates the risk level of the site from the output images using a trained model. The trained model is generated by machine learning using training images based on text contained in the section for recording the cause of an accident in a report on an accident during on-site work as input data, and labels representing hazards as training data. This makes it possible to easily construct a system that supports risk prediction activities.

[0013] In the above embodiment, a mobile body equipped with the camera may be further provided. This makes it possible to move the camera and take images of the site.

[0014] In the above embodiment, the trained model is generated using the training image and process data representing the work content of the on-site work as input data, and the estimation unit may estimate the degree of risk from the image and process data representing the work content planned at the site. This makes it possible to estimate the degree of risk according to the work content.

[0015] In addition, a program according to another aspect of the present invention causes a computer to acquire an image of a site and estimate the risk level of the site from the output image using a learned model. The learned model is generated by machine learning with text included in a cause description column of a report on an accident at a site operation as input data and a label indicating danger as teacher data. According to this, it becomes possible to facilitate the construction of a system for supporting risk prediction activities.

Effects of the Invention

[0016] According to the present invention, it becomes possible to facilitate the construction of a system for supporting risk prediction activities at a site.

Brief Description of the Drawings

[0017] [Figure 1] It is a diagram showing an example of a method for generating a learned model. [Figure 2] It is a diagram showing an example of a report. [Figure 3] It is a diagram showing an example of a learning device. [Figure 4] It is a diagram showing an example of a generation step. [Figure 5] It is a diagram showing an example of a learning image. [Figure 6] It is a diagram showing an example of a learning image. [Figure 7] It is a diagram showing an example of a learning step. [Figure 8] It is a diagram showing an example of a learning image. [Figure 9] It is a diagram showing an example of a learning step. [Figure 10] It is a diagram showing an example of a risk level estimation system. [Figure 11] It is a diagram showing an example of a risk level estimation method. [Figure 12] It is a diagram showing an example of an inference step. [Figure 13] It is a diagram showing an example of a screen display. [Figure 14] It is a diagram showing an example of a risk level estimation method. [Figure 15] This figure shows an example of an inference step. [Figure 16] This is a diagram showing an example of the screen display. [Modes for carrying out the invention]

[0018] Embodiments of the present invention will be described below with reference to the drawings. In this specification and in each drawing, elements similar to those described above in relation to previously shown drawings will be denoted by the same reference numerals, and detailed descriptions may be omitted as appropriate.

[0019] [How to generate a pre-trained model] Figure 1 is a flowchart showing an example of the procedure for generating a pre-trained model. The method for generating a pre-trained model comprises the generation step S1, the assignment step S2, and the training step S3 in that order.

[0020] The method for generating a pre-trained model according to this embodiment is a method for generating a pre-trained model for estimating the degree of risk from images taken at a work site. A work site is a workplace where workers carry out their work, such as a construction site, building site, manufacturing site, or logistics site.

[0021] Generation step S1 is the step of generating training images. Generation step S1 is also an example of a preparation step that prepares training images.

[0022] The assignment step S2 is the step of assigning labels to the training images that indicate danger or safety (so-called annotation).

[0023] Learning step S3 is the step in which a trained model is generated using machine learning, with training images as input data and labels as training data.

[0024] In this embodiment, the generation or preparation of training images is performed based on reports of accidents during on-site work. The reports are not limited to reports of accidents that actually occurred during on-site work, but may also be reports of incidents that did not result in an accident but could have developed into one (so-called near misses).

[0025] Figure 2 shows an example of a report RP (Reporting Process). The report RP includes fields for details such as the date and time of the incident, location, parties involved, work performed, accident circumstances, cause of the incident, and countermeasures. These fields are typically filled out by, for example, the parties involved in the accident or the on-site supervisor.

[0026] In the report RP, the cause of the incident (DB) field should contain the cause of the accident as text. The cause of the incident (DB) field may also include a diagram (sketch or explanatory diagram) to illustrate the cause of the incident.

[0027] The countermeasures section SB is where countermeasures to prevent the recurrence of the accident are written in text. The countermeasures section SB may also include diagrams (sketches or explanatory diagrams) to explain the countermeasures.

[0028] For example, if an accident occurs where a worker falls into a hole at a construction site because there was no fence around it, the cause of the accident entry field (DB) will contain text and a diagram explaining that there was a hole at the site and that there was no fence around it.

[0029] In response to this, if it is decided that the solution will be to erect a fence around the hole in the site, the response section SB will contain text and a diagram explaining that a fence will be erected around the hole in the site.

[0030] Figure 3 is a block diagram showing an example configuration of the learning device 1 for realizing a method for generating a trained model. The learning device 1 is a computer that includes a CPU, RAM, ROM, non-volatile memory, and input / output interfaces, etc.

[0031] The CPU performs information processing according to a program loaded into RAM from ROM or non-volatile memory. The program may be provided via a non-temporary storage medium or via a communication line.

[0032] The learning device 1 comprises an input processing unit 11, a set creation unit 12, and a learning processing unit 13. These functional units are realized by the learning device 1 executing information processing according to a program.

[0033] Furthermore, the learning device 1 has access to the report data storage unit 16, the process data storage unit 17, the dataset storage unit 18, and the trained model storage unit 19. These storage units may be provided in the memory of the learning device 1.

[0034] Furthermore, the learning device 1 can access the Generative AI Server 9. The Generative AI Server 9 is equipped with Generative AI (Artificial Intelligence) that generates images from text.

[0035] The report data storage unit 16 stores report data that represents the contents of the report RP. The report data is a document file (electronic data document) of the report RP. However, the report data may also be data containing text and figures read from a paper report, or data containing text converted from recorded audio.

[0036] The process data storage unit 17 stores process data representing the work content of the on-site work. The process data represents the work content specified in the report RP, that is, the work content that was being performed at the time of the accident. The process data may also include the time required for the work, the order of the work, preceding and succeeding work, interval time, start date and time, or end date and time.

[0037] The dataset storage unit 18 stores the training dataset used for machine learning. The training dataset includes training images and the labels assigned to them. The training dataset may also include process data. The trained model storage unit 19 stores the trained model generated by the learning device 1.

[0038] The input processing unit 11, the set creation unit 12, and the learning processing unit 13 each execute the generation step S1, the assignment step S2, and the learning step S3, respectively.

[0039] The input processing unit 11 executes the generation step S1. Specifically, the input processing unit 11 reads the report data from the report data storage unit 16 and, based on the text and other information contained in the report RP, causes the generation AI installed in the generation AI server 9 to generate training images.

[0040] The set creation unit 12 executes the labeling step S2. That is, the set creation unit 12 labels the training images generated by the generation AI server 9 and stores them in the dataset storage unit 18 as a training dataset. The set creation unit 12 may also include process data read from the process data storage unit 17 in the training dataset.

[0041] The learning processing unit 13 executes the learning step S3. Specifically, the learning processing unit 13 reads the training dataset from the dataset storage unit 18, uses the training images and process data as input data, and the labels as training data to generate a trained model by machine learning, and stores the generated trained model in the trained model storage unit 19.

[0042] The generation step S1, the assignment step S2, and the learning step S3 will be explained in detail below.

[0043] [Generation step S1 and Granting step S2] Figure 4 is a diagram illustrating an example of generation step S1. Figures 5 and 6 are diagrams illustrating examples of training images MG generated by generation step S1.

[0044] As shown in Figure 4, in generation step S1, the generating AI generates training images MG based on the text and other elements contained in the report RP. The generating AI may also receive not only the text contained in the report RP, but also the images contained in the report RP.

[0045] Furthermore, the generating AI may receive not only the text included in the report RP, but also reference images RG from the site. Reference images RG from the site are images that can be used as reference for generating training images MG, such as photographs or blueprints of the same or similar sites.

[0046] The generation of training images (MG) is performed by sending prompts to the generating AI that instruct it to generate images. These prompts include text from the report (RP). For example, a prompt such as "Generate an image representing the following text (text from the report RP)" might be used.

[0047] Furthermore, when attaching a diagram from the report RP or a reference image RG from the site to a prompt, a prompt such as "Generate an image representing the following text (text from the report RP) using the attached image (diagram from the report RP or reference image RG from the site) as a reference" may be used.

[0048] Instead of including the report RP text in the prompt, the report RP document file may be attached to the prompt. In this case, a prompt such as "Generate an image representing the contents of the attached document file (a brief file of the report RP)" may be used.

[0049] In the assignment step S2, binary labels representing either danger or safety are assigned to the training image MG. For example, a label representing danger (hereinafter also referred to as the "danger label") is represented by 1, and a label representing safety (hereinafter also referred to as the "safe label") is represented by 0.

[0050] While a human could decide whether to assign a danger label or a safety label by reviewing the training images (MG), using a generative AI to make this decision, as explained below, can reduce the annotation workload.

[0051] In the first example, in the generation step S1, the generation AI is made to generate a training image MG, and also to generate a label to be attached to the training image MG. In the assignment step S2, the generated label is attached to the training image MG.

[0052] In this case, a prompt such as "Generate an image representing the contents of the attached document file (the report RP file), and then generate a hazard or safety label appropriate to that image" might be used.

[0053] In the second example, in generation step S1, a danger label or a safety label is specified, and the generation AI is made to generate a training image MG to which the specified label is to be applied. In application step S2, the specified label is applied to the training image MG.

[0054] In this case, a prompt such as "Generate an image suitable for a hazard label based on the contents of the attached document file (report RP file)" is used, and the hazard label is assigned to the generated training image MG.

[0055] Alternatively, a prompt such as "Generate an image suitable for a safety label based on the contents of the attached document file (report RP file)" may be used, and a safety label will be applied to the generated training image MG.

[0056] In the third example, in generation step S1, the generation AI generates training images MG based on the text and figures included in the cause entry field DB (see Figure 2) of the report RP, and in assignment step S2, a danger label is assigned to the training images MG.

[0057] Specifically, the prompt includes text extracted from the database containing the cause of the incident in the report RP. In this case, a prompt such as "Generate an image representing the following text (text from the cause of the incident database)" is used, and a danger label is assigned to the generated training image MG.

[0058] Alternatively, the generating AI may be instructed to use the text contained in the database for the cause of occurrence entry field in the report RP. In this case, a prompt such as "Generate an image representing the contents of the cause of occurrence entry field in the attached document file (the report RP file)" may be used, and a danger label will be attached to the generated training image MG.

[0059] For example, if the database for recording the cause of an accident contains text explaining that there was no fence around the hole at the site, then based on this text, a training image MG showing no fence around the hole at the site, as shown in Figure 5, will be generated.

[0060] In the fourth example, in generation step S1, the generation AI generates a training image MG based on the text and figures included in the countermeasure entry field SB (see Figure 2) of the report RP, and in assignment step S2, a safety label is assigned to the training image MG.

[0061] Specifically, the prompt includes text extracted from the countermeasures section SB of the report RP. In this case, a prompt such as "Generate an image representing the following text (text from the countermeasures section SB)" is used, and a safety label is assigned to the generated training image MG.

[0062] Alternatively, the generating AI may be instructed to use the text contained in the countermeasures entry field SB of the report RP. In this case, a prompt such as "Generate an image representing the contents of the countermeasures entry field from the attached document file (the report RP file)" may be used, and a safety label will be added to the generated training image MG.

[0063] For example, if the countermeasure entry field SB contains text explaining that a fence should be erected around the hole at the site as a countermeasure to prevent the accident from recurring, then based on this text, a training image MG showing a fence erected around the hole at the site, as shown in Figure 6, will be generated.

[0064] The above describes a method of generating training images (MG) using a generation AI, but the method is not limited to this; training images (MG) based on the text of the report (RP) may also be prepared by means other than the generation AI.

[0065] For example, a computer may select training images (MG) from a large number of candidate images based on factors such as the degree of relevance between the text of the report (RP) and the tags attached to the candidate images, as well as the similarity between the diagrams or reference images (RG) in the report (RP) and the candidate images. Alternatively, training images (MG) may be photographed, drawn, or selected by a person to match the content of the report (RP).

[0066] While it may be up to the individual to decide whether to assign a danger label or a safety label, the annotation burden can be reduced by assigning a danger label to training images MG prepared based on the text contained in the cause entry field DB, as in the third example above, and a safety label to training images MG prepared based on the text contained in the countermeasure entry field SB, as in the fourth example above.

[0067] Furthermore, when providing the generating AI with a diagram of the report RP, a reference image RG of the site, or a file of the report RP in the generation step S1, it is not limited to attaching it to the prompt; for example, a link to the data may be included in the prompt to allow the generating AI to retrieve the data.

[0068] The diagrams in the report (RP) or the reference images (RG) of the site may be captured images or processed images. Images may be captured manually by a person or automatically by a surveillance camera or the like.

[0069] Image processing can include, for example, simplifying complex structures, removing unnecessary information (such as removing power lines from landscape photos), color-coding based on temperature, or adding lines indicating horizontal, vertical, or directional directions.

[0070] The training images (MG) generated by the generation AI may be multiple. For example, multiple training images (MG) of the same scene viewed from different angles may be generated, or multiple training images (MG) of different levels of proximity, such as individual images and a whole image, may be generated.

[0071] Furthermore, the multiple training images (MG) generated by the generation AI may also be videos. For example, a video that pans horizontally across a wide area may be generated, or a video that gradually zooms in to attract attention may be generated.

[0072] [Learning Step S3] Figure 7 is a diagram illustrating an example of learning step S3. In learning step S3, machine learning is performed on the model MD using training images MG and process data as input data, and binary labels representing danger or safety as training data.

[0073] Model MD is an image recognition model such as a convolutional neural network (CNN). The output layer of Model MD is configured to output a numerical value between 0 and 1, for example, using a sigmoid function, and this value is used as the risk level.

[0074] In machine learning, training images (MG) and process data are input to the model (MD). The difference between the risk level and label output by the model (MD) is calculated, and the parameters of the model (MD) are adjusted by performing backpropagation to reduce this difference.

[0075] The model MD may be input to multiple training images MG with different angles or approaches as described above, or it may be input to multiple training images MG that constitute a video.

[0076] According to the embodiments described above, by having the generating AI generate training images MG, and further by using the generating AI to assign labels to the training images MG, it becomes possible to facilitate the generation of a trained model used in the risk estimation system 200 described later.

[0077] [Another learning method] Next, we will describe the generation steps S1 to the learning steps S3 related to a different learning mode. Figure 8 is a diagram showing an example of a learning image MG generated by generation step S1. Figure 9 is a diagram illustrating an example of learning step S3.

[0078] As shown in Figure 8, in generation step S1, the generation AI generates a training image MG to which markings MK indicating dangerous areas are added. Markings MK are, for example, rectangular frames surrounding dangerous areas. However, they are not limited to these; markings MK may also be, for example, arrows, circles, cloud-shaped frames, circular frames, hatching, or speech bubbles.

[0079] The training image MG and marking MK may be generated separately, or the training image MG and a map representing the position of the marking MK (corresponding to the label map described later) may be generated separately.

[0080] To generate training images (MG) with markings (MK), prompts such as "Generate an image that represents the contents of the attached document file (report RP file), with dangerous areas marked" are used.

[0081] For example, if the cause of the accident is recorded in the report RP's cause entry field DB (see Figure 2), and the text explaining that a portion of the scaffolding railing installed at the site was missing is entered as the cause of the accident, then based on this text, a training image MG is generated showing a portion of the scaffolding railing missing, as shown in Figure 8, with a marking MK placed over that area.

[0082] In the assignment step S2, dangerous labels are assigned to the regions of the training image MG that have been marked with marking MK. Specifically, a label map is created in which the regions corresponding to marking MK are assigned the dangerous label "1", and the remaining regions are assigned the safe label "0", as shown in Figure 9.

[0083] The label map is a set of unit regions EL corresponding to each part of the training image MG, and is a binary bitmap in which each unit region EL is assigned either a danger label "1" or a safety label "0". A unit region EL corresponds to a single pixel or a pixel block containing multiple pixels of the training image MG.

[0084] Furthermore, the method is not limited to assigning danger labels to areas marked with markings (MK) by the generating AI; a human may also identify dangerous areas in the training image (MG) and assign danger labels to those areas.

[0085] As shown in Figure 9, in learning step S3, machine learning of the model MD is performed using the training image MG and process data as input data, and the label map as training data, to output a risk map.

[0086] The risk map is a set of unit regions PX corresponding to each part of the training image MG, and is a bitmap of numerical values ​​between 0 and 1 that represent the calculated risk for each unit region PX. The unit regions PX of the risk map correspond to the unit regions EL of the label map.

[0087] The output layer of the Model MD contains elements corresponding to each unit region PX of the risk map, and each element is configured to output a numerical value between 0 and 1 that represents the risk, for example, using a sigmoid function.

[0088] In machine learning, training images (MG) and process data are input to the model (MD). The difference between the risk map and label map output by the model (MD) for each unit area is calculated, and the parameters of the model (MD) are adjusted by performing backpropagation to reduce these differences.

[0089] [Risk Estimation System] Figure 10 is a block diagram showing an example configuration of the hazard estimation system 200. The hazard estimation system 200 is a system that supports hazard prediction activities at the site.

[0090] The risk estimation system 200 comprises an estimation device 2, a camera 3, a mobile unit 4, and a display unit 5. The estimation device 2 is a computer including a CPU, RAM, ROM, non-volatile memory, and input / output interfaces, and performs information processing according to a program.

[0091] Camera 3 is a digital camera that captures images of the scene and outputs them. Camera 3 transmits the images to the estimation device 2 via wireless communication such as Wi-Fi.

[0092] The mobile unit 4 is equipped with a camera 3 and moves the camera 3. The mobile unit 4 may be, for example, an unmanned aerial vehicle such as a drone, an unmanned ground vehicle, or a walking robot. Alternatively, a person may move the camera 3.

[0093] For example, the mobile device 4 patrols the site at a predetermined time, and the camera 3 mounted on the mobile device 4 takes images of various locations at the site and outputs the images to the estimation device 2. The time when the mobile device 4 patrols may be early in the morning or before work begins when there are no workers present, or it may be during work hours, as unsafe actions by workers are also included in the risk estimation.

[0094] The estimation device 2 has access to the process data storage unit 27 and the trained model storage unit 29. These storage units may be provided in the memory of the estimation device 2.

[0095] The process data storage unit 27 stores process data representing the work content scheduled to be performed on site. The trained model storage unit 29 stores the trained model generated by the method described above.

[0096] The estimation device 2 uses a trained model to estimate the level of danger at the site from images of the site captured by the camera 3 and process data representing the planned work at the site, and outputs the estimation result to the display unit 5.

[0097] Figure 11 is a flowchart showing an example of the procedure for the risk estimation method implemented in the risk estimation system 200. The estimation device 2 executes the information processing shown in the figure according to the program. Figure 12 is a diagram illustrating an example of the inference step S13. Figure 13 is a diagram showing an example of the screen display of the display unit 5.

[0098] As shown in Figure 11, first, the estimation device 2 acquires an image SG of the site from the camera 3 (S11), and then acquires process data representing the planned work content at the site from the process data storage unit 27 (S12).

[0099] Next, the estimation device 2 inputs the on-site image SG and process data as input data into the trained model LM and performs an inference step to estimate the risk level, which is expressed as a numerical value between 0 and 1 (S13, see Figure 12).

[0100] Next, the estimation device 2 determines whether the calculated risk level is above a predetermined level (S14). If the risk level is above a predetermined level (S14: YES), the estimation device 2 outputs a warning on the screen of the display unit 5 (S15).

[0101] For example, as shown in Figure 13, the screen of the display unit 5 displays an image SG of the site and a string AL warning that the site shown in the image SG is dangerous.

[0102] In the inference step S13, the image SG input to the trained model LM may be multiple images. For example, multiple image SGs taken from the same location at different angles may be input, or multiple image SGs with different levels of proximity, such as individual images and a whole image, may be input.

[0103] Furthermore, the multiple images SG input to the trained model LM may also be videos. For example, a video panning horizontally across a wide area may be input, or a video gradually zooming in to attract attention may be input.

[0104] The process data input to the trained model LM includes the work scheduled to be done on-site on the current day or on-site during the current week. The process data may also include the time required for each task, the order of tasks, preceding and succeeding tasks, interval times, start date and time, or end date and time.

[0105] [Another inference pattern] Figure 14 is a flowchart showing an example procedure for a risk estimation method related to another inference mode. Figure 15 is a diagram illustrating an example of inference step S16. In this mode, the trained model LM generated in the above-mentioned other learning mode is used. Figure 16 is a diagram showing an example of the screen display of the display unit 5.

[0106] As shown in Figure 14, first, the estimation device 2 acquires an image SG of the site from the camera 3 (S11), and then acquires process data representing the planned work content at the site from the process data storage unit 27 (S12).

[0107] Next, the estimation device 2 inputs the on-site image SG and process data as input data into the trained model LM and performs an inference step to calculate a risk map in which the risk level of each unit region PX is represented by a numerical value between 0 and 1 (S16, see Figure 15).

[0108] Next, the estimation device 2 combines the on-site image SG with the hazard map (S17) and outputs the image SG showing the area with a hazard level of a predetermined value or higher to the display unit 5 (S18).

[0109] For example, as shown in Figure 16, the screen of the display unit 5 displays an image SG in which an area with a danger level above a predetermined level is surrounded by a rectangular frame BB (a so-called boundary box), and a string of characters AL that warns that there is a dangerous area in the site shown in the image SG.

[0110] Not limited to the rectangular frame BB, areas with a risk level above a predetermined level may be indicated by, for example, arrows, circles, cloud-shaped frames, circular frames, hatching, or callouts. Furthermore, the risk levels of each part of image SG may be color-coded and displayed as a heatmap.

[0111] Although embodiments of the present invention have been described above, the present invention is not limited to the embodiments described above, and various modifications are of course possible for those skilled in the art.

[0112] In the above embodiment, the trained model was made to output the overall risk level or risk map. In addition, if the risk level is above a predetermined level, the trained model may further output text that explains why it is dangerous and how to deal with it, providing guidance. This text can be output by the trained model using, for example, the text contained in the cause entry field DB and the countermeasure entry field SB (see Figure 2) of the report RP as training data.

[0113] Furthermore, in the above embodiment, a warning was output by showing the entire image SG of the site or the hazardous area, but the system is not limited to this, and a warning may also be output by matching the hazardous area recognized by the trained model with the construction plan drawing (e.g., a 3D model). The matching of hazardous areas can be performed by identifying the location on the construction plan drawing based on the position information of the mobile body 4. [Explanation of Symbols]

[0114] 1 Learning device, 11 Input processing unit, 12 Set creation unit, 13 Learning processing unit, 16 Report data storage unit, 17 Process data storage unit, 18 Dataset storage unit, 19 Trained model storage unit, 2 Estimation device, 3 Camera, 4 Mobile unit, 5 Display unit, 27 Process data storage unit, 29 Trained model storage unit, 9 Generation AI server, 200 Risk estimation system

Claims

1. Preparation steps include creating learning images based on the text included in the section for recording the cause of an accident in a report on an accident during on-site work, and A step of assigning a label indicating danger to the learning image based on the text contained in the field for entering the cause of the occurrence, A learning step in which a trained model for estimating the degree of danger from images taken at a site is generated by machine learning, using the aforementioned training images as input data and the aforementioned labels as training data, A method for generating a pre-trained model that includes the following features.

2. The aforementioned preparation step involves preparing learning images based on the text included in the section for filling in countermeasures in the aforementioned report, The aforementioned assignment step involves assigning a label indicating safety to the learning image based on the text contained in the input field for the countermeasures. A method for generating a trained model according to claim 1.

3. The preparation step involves preparing the learning images based on the text and the figures included in the report. A method for generating a trained model according to claim 1.

4. The preparation step involves preparing the learning images based on the text and reference images from the site. A method for generating a trained model according to claim 1.

5. The learning step uses the learning image and process data representing the work content of the on-site work as input data. A method for generating a trained model according to claim 1.

6. The aforementioned labeling step involves assigning the label representing danger to a portion of the training image, The learning step generates the trained model for estimating areas in the image where the risk level is above a predetermined level. A method for generating a trained model according to claim 1.

7. A camera that captures images of the site and outputs them, An estimation unit that estimates the degree of danger at the site from the output image using a trained model, Equipped with, The aforementioned trained model is This model was generated using machine learning, with learning images based on text from the accident cause section of field work accident reports as input data, and hazard labels as training data. Risk estimation system.

8. The mobile body further comprises the aforementioned camera, The risk estimation system according to claim 7.

9. The trained model is generated using the training images and process data representing the work content of the field work as input data. The estimation unit estimates the degree of risk from the image and process data representing the planned work content at the site. The risk estimation system according to claim 7.

10. To acquire images taken at the site, and Using a trained model, estimate the degree of danger at the site from the output image. Have the computer run it, The aforementioned trained model is This model was generated using machine learning, with learning images based on text from the accident cause section of field work accident reports as input data, and hazard labels as training data. program.

Images