Safety support system, safety support method, and safety support program

Abstract

To provide a safety support system, a safety support method, and a safety support program for efficiently and accurately supporting safety assurance.SOLUTION: A support device 20 includes: a knowledge base storage unit 23 recording analysis results of elements included in a written report; a learning result storage unit 24 recording risk prediction information for predicting a risk based on the analysis results; and a control unit 21 connected to a user terminal 10. Based on the elements recorded in the knowledge base storage unit 23, the control unit 21 generates risk prediction information for predicting the risk caused by work and a probability of occurrence, and records it in the learning result storage unit 24. In addition, in response to acquisition of new work contents from the user terminal 10, the control unit 21 uses the risk prediction information to identify risk candidates and output them to the user terminal 10.SELECTED DRAWING: Figure 1

Description

Technical Field

[0001] The present invention relates to a safety support system, a safety support method, and a safety support program for assisting in ensuring safety in work.

Background Art

[0002] An accident prevention management method for determining accident recurrence prevention measures by making use of lessons learned from past accidents has been studied (see, for example, Patent Document 1). The accident prevention management method disclosed in this document classifies data on a large number of accidents that have occurred in the past according to the accident situation at the time of occurrence, and stores in a computer the risk level corresponding to the degree of influence and the occurrence frequency of each accident. Further, the hazard factors of each accident, the important management points in the work process for reducing / eliminating them, and the accident recurrence prevention measures to be taken at the important management points in the work process of each accident are stored in the computer. Then, when performing work under the same situation as the accident situation, the accident recurrence prevention measures corresponding to the important management points in the work process of the work are extracted and determined.

Prior Art Documents

Patent Documents

[0003]

Patent Document 1

Summary of the Invention

Problems to be Solved by the Invention

[0004] However, classifying a large number of accident data according to the accident situation and inputting risk levels, important management points in the work process, etc. are time-consuming and also require a large burden for system construction and maintenance. Further, in the target industries that involve large-scale and complex relationships, it is difficult to systematize and control the classification by manual work and the provision of various information such as risk levels and important management points.

[0005] In addition, when using a uniform measure for risks as an index, it does not suggest information that should be noted in various on-site situations and among workers, and there are limitations in preventing industrial accidents in advance.

Means for Solving the Problems

[0006] The safety support system for solving the above problems includes a knowledge base storage unit that records the analysis results of the elements included in the report, a learning result storage unit that records risk prediction information for predicting risks based on the analysis results, and a control unit connected to the user terminal. The control unit includes a learning unit that generates risk prediction information for predicting the risks and the probabilities of occurrence caused by the work based on the elements recorded in the knowledge base storage unit and records it in the learning result storage unit, and a support unit that, when acquiring the content of a new work from the user terminal, specifies risk candidates using the risk prediction information and outputs them to the user terminal.

Effects of the Invention

[0007] According to the present invention, it is possible to efficiently and accurately support ensuring safety.

Brief Description of the Drawings

[0008]

Figure 1

Figure 2

Figure 3

Figure 4

Figure 5

Figure 6

Figure 7

Figure 8

Figure 9

Figure 10

Figure 11

Figure 12

Figure 13

Figure 14

Figure 15

Figure 16

Figure 17

Figure 18

Figure 19

Figure 20

Figure 21

Figure 22

Figure 23

Figure 24

Mode for Carrying Out the Invention

[0009] Hereinafter, an embodiment of a safety support system, a safety support method, and a safety support program will be described with reference to FIGS. 1 to 24. In this embodiment, a safety support system used for supporting safety management when using a large number of past accident reports at a building construction site will be described. In this embodiment, as shown in FIG. 1, a user terminal 10 and a support device 20 connected via a network or the like are used.

[0010] (Description of Hardware Configuration) Using FIG. 2, the hardware configuration of the information processing device H10 that constitutes the user terminal 10 and the support device 20 will be described. The information processing device H10 includes a communication device H11, an input device H12, a display device H13, a storage device H14, and a processor H15. Note that this hardware configuration is an example, and it is also possible to be realized by other hardware.

[0011] The communication device H11 is an interface that establishes a communication path with other devices and executes data transmission and reception, such as a network interface or a wireless interface.

[0012] The input device H12 is a device that accepts input of various information, such as a mouse or a keyboard. The display device H13 is a display or the like that displays various information. The storage device H14 is a storage device that stores data and various programs for executing various functions of the user terminal 10 and the support device 20. Examples of the storage device H14 include ROM, RAM, and hard disks.

[0013] The processor H15 controls each process in the user terminal 10 and the support device 20 using the programs and data stored in the storage device H14. Examples of the processor H15 include a CPU and an MPU. This processor H15 expands the programs stored in ROM or the like into RAM and executes various processes for each process.

[0014] Processor H15 is not limited to performing software processing for all processes it executes. For example, processor H15 may include a dedicated hardware circuit (e.g., an application-specific integrated circuit: ASIC) that performs hardware processing for at least a part of the processes it executes. That is, processor H15 may be configured as a circuit including [1] one or more processors that operate according to a computer program, [2] one or more dedicated hardware circuits that execute at least a part of various processes, or [3] a combination thereof. The processor includes a CPU and memories such as a RAM and a ROM, and the memories store program codes or instructions configured to cause the CPU to execute processes. The memory, that is, the computer-readable medium, includes any available medium that can be accessed by a general-purpose or dedicated computer.

[0015] (System Configuration) Next, each function of the safety support system will be described with reference to FIG. 1. The user terminal 10 is a computer terminal used by users such as administrators and workers of the safety support system.

[0016] The support device 20 is a computer system that supports the safety management of workers. This support device 20 includes a control unit 21, a report storage unit 22, a knowledge base storage unit 23, and a learning result storage unit 24.

[0017] The control unit 21 performs safety support processing described later. For this purpose, it functions as an analysis unit 211, a learning unit 212, and a support unit 213 by a safety support application.

[0018] The analysis unit 211 executes a process of analyzing reports and constructing each knowledge base. The learning unit 212 learns the knowledge base and generates risk prediction information for predicting risks that may occur. The support unit 213 generates information for supporting safety management.

[0019] As shown in FIG. 3, the report storage unit 22 stores files of a large number of past reports (industrial accident reports 22a, near miss reports 22b, safety patrol record reports 22c).

[0020] The industrial accident report 22a includes data regarding the victim, age, experience, weather, job type, type of work where the accident occurred, causative object, date and time of occurrence, type of accident, circumstances of occurrence, cause of occurrence, and countermeasures. In each data area for the victim, age, and experience, the name, age, and years of experience of the victim are recorded respectively.

[0021] In the weather data area, data for specifying the weather at the time of the accident (e.g., "sunny") is recorded. In the job type data area, data for specifying the job type of the victim (e.g., "pole worker") is recorded.

[0022] In the type of work where the accident occurred data area, data for specifying the type of work where the disaster occurred (e.g., "scaffold disassembly") is recorded. In the causative object data area, data regarding the material that caused the disaster (e.g., "scaffold") is recorded.

[0023] In the date and time of occurrence data area, data regarding the date and time of the accident is recorded. In the type of accident data area, data for specifying the classification of the accident (e.g., "fall") is recorded.

[0024] In the circumstances of occurrence data area, data regarding the circumstances in which the disaster occurred is recorded. The circumstances of occurrence are described in text (unstructured information) such as the work location, work content, work situation, details of the accident, and safety measure situation.

[0025] In the cause of occurrence data area, data regarding the cause of the accident is recorded. For example, "the main rope was slack, so the safety belt was not used", "the foot area was not sufficiently checked", etc. are described in text (unstructured information).

[0026] In the countermeasure data area, data related to the countermeasures for preventing this disaster is recorded. For example, "Before work, check the condition of safety equipment such as the main rope. If any insecurity is confirmed, take fall prevention measures", "Before moving, carefully check the area around your feet such as the condition of the scaffolding board and then move" etc. are recorded as text (unstructured information).

[0027] The near-miss report 22b includes data related to the report date, reporter, workplace name, work content, near-miss situation, cause, and countermeasures. The near-miss report is recorded when an event that did not lead to a major disaster or accident but has the potential to directly cause a disaster or accident is recognized.

[0028] In each data area for the report date, reporter, and workplace name, data for the report date of the near-miss event, the name of the reporter of the near-miss event, and data for identifying the site where the near-miss event occurred are recorded respectively.

[0029] In the work content data area, data for identifying the work in which the near-miss event occurred is recorded. In each data area for the near-miss situation, cause, and countermeasures, data related to the situation, cause, and countermeasures of the near-miss event are recorded as text (unstructured information) respectively.

[0030] The safety inspection record report 22c includes data related to the report date, reporter, workplace name, checker, inspection result, corrective plan, corrective result, improvement instruction items, and improvement status. In each data area for the report date, reporter, and workplace name, data for the report date of the issues pointed out during the safety inspection, the name of the reporter of the issues pointed out, and data related to the site where the issues occurred are recorded respectively.

[0031] In the checker data area, data related to the name of the checker of the issues pointed out is recorded. In the patrol result data area, data related to the results of safety patrols is recorded. As pointed-out matters, "materials protruding onto the work passageway", "the standing horses near the pile opening are placed at a position close to the opening", "it is possible for a vehicle to drive onto the corner floor plate", "a welding machine without permission to bring it in is placed", etc. are recorded in text (unstructured information).

[0032] In each data area for the correction plan and correction results, data related to improvement instructions and improvement status is recorded. As improvement instructions, "tidying up the work passageway", "moving the standing horses more than 2 m away from the opening end", "posting a no-vehicle-entry sign", "obtaining permission to bring it in", etc. are described in text (unstructured information). Note that in this embodiment, the report storage unit 22 is provided in the support device 20, but it may be provided in another server.

[0033] As shown in FIG. 4, in the knowledge base storage unit 23, a probability distribution knowledge base 23a, an association rule knowledge base 23b, a causative agent knowledge base 23c, a generalized causative agent knowledge base 23d, an action / function knowledge base 23e, and a generalized action / function knowledge base 23f are recorded.

[0034] In the probability distribution knowledge base 23a, the probability distribution calculated by cross-analysis is recorded. For example, a cross-analysis of "age group" (e.g., age division in 10-year units) of the victims and "type of accident" is performed, and the cross-analysis results including the probability distribution calculated for each combination of elements of "age group" and "type of accident" are recorded. Also, the cross-analysis results are recorded for probability distributions of combinations of elements such as "number of years of experience" and "type of accident", and "occupation type" and "type of accident".

[0035] In the association rule knowledge base 23b, the results of the association analysis are recorded. This association analysis result includes an antecedent (IF component), a consequent (THEN component), and a lift value. The antecedent consists of a combination of attribute items (antecedents) of the structure information (job type, causative agent, age, work type, causative agent, time zone, etc.) included in each report. The consequent consists of the type of accident. The lift value is the strength of the rule for the co-occurrence of the antecedent and the consequent. For example, through association analysis, it is possible to derive regularities regarding the likelihood of accidents due to combinations of multiple attribute items, such as "when a 'steelworker' in their '30s' is engaged in 'S body'-related construction work, they are likely to have a 'falling or tumbling' accident". In this embodiment, rules and normalized lift values formatted in a form that can be used by a computer are recorded.

[0036] In the causative agent knowledge base 23c, vocabulary related to the causative agents extracted from each report is recorded. Here, the causative agent is the object that is the cause of the accident. In the analysis results, this causative agent appears as a noun phrase in the nominative or accusative case. In this embodiment, for generalized vocabulary, a "#" is attached at the beginning for explanation. For example, as causative agents co-occurring with the generalized vocabulary "#falling", "standing horse", "roof terrace", and "workbench" are recorded. Also, as causative agents co-occurring with the generalized vocabulary "#dropping", "single-pipe", "bolt", and "nut" are recorded.

[0037] In the generalized causative agent knowledge base 23d, generalized vocabulary that aggregates the same type of causative agents into a single expression is recorded in a form that can be processed by a computer. For example, the generalized vocabulary "#mobile scaffold" is recorded as an aggregation of the expressions "standing horse", "roof terrace", and "workbench". The generalized vocabulary "#construction materials" is recorded as an aggregation of the expressions "single-pipe", "bolt", and "nut".

[0038] In the action / function knowledge base 23e, predicates representing actions and functions leading to disasters are recorded in a form that can be processed by a computer based on the results of the case frame analysis. For example, in the case of a "falling or tumbling" accident, predicates representing actions and functions leading to disasters such as "falling", "tumbling", and "plunging" are recorded.

[0039] In the generalization behavior / function knowledge base 23f, generalization vocabulary is recorded for aggregating predicates representing behaviors and functions leading to disasters for each vocabulary group with the same meaning. For example, "#fall" is recorded as the generalization vocabulary for "fall", "topple", and "plunge". Also, "#drop" is recorded as the generalization vocabulary for "drop" and "cause to drop".

[0040] As shown in FIG. 5, in the learning result memory unit 24, a risk factor association graph 24a and a risk prediction learning model 24b are recorded as risk prediction information for predicting risks. The risk factor association graph 24a is a directed graph (association information) that associates risk factors (causative agents, behaviors / functions / statuses, etc.) leading to "accident types" based on "occupation type" and "work type", and has as its end the risks that can result therefrom. For example, in the occupation type "interior decorator" and work type "painting work", when the causative agent is "#mobile scaffold" and there are "#fall" and "#inappropriately move", it is a graph that links the accident type "fall / topple" and the risk "insufficient risk awareness".

[0041] The risk prediction learning model 24b is a learning result generated by machine learning using all the derived or extended learning data sets as the learning data set. This risk prediction learning model 24b is a neural network that outputs predicted risks (accident types, causes) from a set of risk factors derived from "occupation type" and "causative agent".

[0042] (Outline of processing during learning) The outline of the processing during learning will be described with reference to FIG. 6. The details of each process will be described later. First, the control unit 21 of the support device 20 executes a labor accident knowledge extraction process S10. Specifically, the analysis unit 211 of the control unit 21 systematically converts various risk elements into knowledge using the reports recorded in the report storage unit 22. Here, a labor accident data analysis process S11 and a labor accident data update process S12 are executed.

[0043] Next, the control unit 21 of the support device 20 executes a determination process S15 on whether to end. Specifically, the analysis unit 211 of the control unit 21 determines to end if the labor accident knowledge extraction process S10 has been performed on all the reports recorded in the report storage unit 22.

[0044] If it is determined that it is not the end (in the case of "NO"), the control unit 21 of the support device 20 repeats the labor accident knowledge extraction process S10. On the other hand, if it is determined that it is the end (in the case of "YES"), the control unit 21 of the support device 20 executes the labor accident knowledge storage process S20. Specifically, the learning unit 212 of the control unit 21 converts the systematized knowledge into a form that can be used by the computer during risk prediction and stores it in the knowledge base storage unit 23. Here, the risk prediction data expansion process S21 and the risk prediction learning model generation process S22 are executed.

[0045] (Overview of support-time processing) Next, the overview of the support-time processing will be described with reference to FIG. 7. The details of each process will be described later. First, the control unit 21 of the support device 20 executes a risk candidate derivation process S30. Specifically, the support unit 213 of the control unit 21 uses the learning result storage unit 24 to derive risk candidates that may occur under the given conditions.

[0046] Next, the control unit 21 of the support device 20 executes a dynamic risk level measurement process S40. Specifically, the support unit 213 of the control unit 21 dynamically quantifies the degree of risk for the derived risk candidates according to the operator and the situation.

[0047] Next, the control unit 21 of the support device 20 executes a dynamic risk information presentation process S50. Specifically, the control unit 21 of the support device 20 provides selective information according to the work based on the risk level (quantified degree of risk) and the risk candidates. Here, the risk arousal information expansion process S51 and the risk arousal information generation process S52 are executed.

[0048] (Labor accident data analysis process) Next, with reference to FIG. 8, the industrial accident data analysis process S11 performed in the industrial accident knowledge extraction process S10 will be described.

[0049] In the industrial accident data analysis process S11, the analysis unit 211 executes a quantitative analysis process a1 and a qualitative analysis process b1. The quantitative analysis process a1 is a process of analyzing the structural information (such as "age" to "type of accident", etc.) included in the industrial accident report 22a. In the present embodiment, the quantitative analysis process a1 includes a cross-analysis process a111 and a correlation analysis process a121. Note that the analysis method is not limited to cross-analysis and correlation analysis, and any method included in data science can be applied.

[0050] First, the analysis unit 211 executes the cross-analysis process a111. Here, the tendency between an accident and any item is grasped. For example, the age of the victim is divided into 10-year units and classified into "age groups", and a cross-analysis of "age group" and "type of accident" is performed. Here, it is tabulated for each "age group" to calculate the probability distribution of the "type of accident". Similarly, by performing cross-analyses of "years of experience" and "type of accident", "occupation" and "type of accident", etc., the respective probability distributions are calculated.

[0051] Next, the analysis unit 211 executes the probability distribution knowledge derivation process a112. Here, the results of the cross-analysis are formatted into a form that can be used by a computer and stored in the probability distribution knowledge base 23a.

[0052] In addition, the analysis unit 211 executes the correlation analysis process a121. Here, the combined tendency of a plurality of items with respect to an accident is grasped. For example, by correlation analysis, a lift value, which is an index indicating the correlation between the antecedent due to the combination of attributes, the consequent due to the type of accident, and the antecedent and the consequent, is calculated. For example, in particular, rules regarding the likelihood of an accident due to a combination of a plurality of attribute items, such as "when a '30s' 'ironworker' is performing 'S body' related work, a 'fall' accident is likely to occur", are derived. Here, the antecedent: "occupation = ironworker, causative object = S body, age = 30s", "fall", the consequent: "fall", and the lift value "8.65" are calculated.

[0053] Next, the analysis unit 211 executes an association rule knowledge derivation process a122. Here, the results of the association analysis are formatted into a form that can be used by a computer and stored in the association rule knowledge base 23b.

[0054] On the other hand, the qualitative analysis process b1 is a process of analyzing text information of unstructured information (such as "occurrence situation", "cause of occurrence", etc.) included in the industrial accident report 22a. In the present embodiment, the qualitative analysis process b1 includes a situation analysis process b111 for analyzing the "occurrence situation" and a cause analysis process b121 for analyzing the "cause of occurrence".

[0055] First, the analysis unit 211 executes the situation analysis process b111. Here, the description of the occurrence situation is processed sentence by sentence. In this case, known morphological analysis processing and dependency analysis processing are performed. In the morphological analysis processing, words are divided by part of speech. In the dependency analysis processing, the dependency relationship between words is specified. Further, in the present embodiment, case structure analysis for specifying the subject (nominative case), object (accusative case), and predicate is performed. For example, for the description of the occurrence situation "When the victim was painting the wall surface, he missed his step and fell from the standing horse", by case structure analysis, the nominative case "victim", accusative case "standing horse", and predicate "fall" are specified. Also, for the description of the occurrence situation "The worker on the upper floor slipped and dropped a single-pipe and hit the victim", the nominative case "worker", accusative case "single-pipe", and predicate "drop" are specified.

[0056] Next, the analysis unit 211 executes the language knowledge compilation process b112. Here, vocabulary knowledge for aggregating into an expression of generalized vocabulary is compiled for each group of vocabulary having the same meaning. For example, according to the results of the case structure analysis, in the case of a "falling·toppling" accident, predicates such as "fall", "topple", and "plunge" that represent actions or functions leading to disasters tend to be used. Therefore, the actually used predicate vocabulary is stored in the action·function knowledge base 23e in a form that can be processed by a computer. Further, the generalized vocabulary related to the predicate is stored in the generalized action·function knowledge base 23f in a form that can be processed by a computer.

[0057] In addition, the object that causes an accident is called the causative object. In the analysis results, it appears as a noun phrase in the nominative or accusative case. For example, as causative objects co-occurring with the generalized vocabulary of the predicate "#fall", there are "standing horse", "rooftop", and "workbench", and as causative objects co-occurring with "drop", there are "single-pipe", "bolt", and "nut". Based on this result, furthermore, the same type of causative objects are aggregated into a single expression. For example, "standing horse", "rooftop", and "workbench" are aggregated into the expression "#mobile scaffold". Also, "single-pipe", "bolt", and "nut" are aggregated into "#construction materials".

[0058] Then, the vocabulary related to the causative objects actually used in the text is stored in the causative object knowledge base 23c in a form that can be processed by a computer. Furthermore, the generalized vocabulary related to the causative objects is stored in the generalized causative object knowledge base 23d in a form that can be processed by a computer.

[0059] In addition, the analysis unit 211 executes the cause analysis process b121. Here, it is performed to clarify the risks to be predicted from the perspective of why an accident occurred through text analysis of the cause description sentence when an accident occurs.

[0060] Next, the analysis unit 211 executes the similar cause aggregation process b122. Here, for the same content, the description sentences described in different expressions by each reporter are aggregated. For this reason, in the similar cause aggregation process b122, in order to aggregate description sentences with the same content even if they are in different expressions, cluster analysis is performed. For example, using the singular value decomposition method, the feature quantities of the sentences are expressed as vectors, and the similarity between the vectors is measured to aggregate the sentences with the same content. In this case, the analysis unit 211 forms clusters of the description sentences by the method of cluster analysis. Then, the administrator assigns a predicted risk label to each cluster using the user terminal 10. An identifier called a predicted risk ID is assigned to this predicted risk label. This predicted risk label is used for the terminal nodes of the risk factor association graph.

[0061] (Industrial accident knowledge storage process) Next, the industrial accident knowledge memory process S20 will be described. In this industrial accident knowledge memory process S20, the learning unit 212 analyzes the situation description in the industrial accident report using the causative agent knowledge base 23c, the generalized causative agent knowledge base 23d, the action / function knowledge base 23e, and the generalized action / function knowledge base 23f. Then, the learning unit 212 generates a risk factor association graph 24a and a risk prediction learning model 24b as knowledge bases for risk prediction.

[0062] First, the learning unit 212 creates a risk factor association graph 24a. As shown in FIG. 9, the document D1 of the accident case E01 is related to an accident in which an interior finisher fell from a scaffold during painting work.

[0063] As shown in FIG. 10, the learning unit 212 generates a risk factor association graph G01 from the accident case in FIG. 9. In the risk factor association graph G01, the learning unit 212 uses the generalized causative agent knowledge base 23d to abstract a specific causative agent as a generalized causative agent and map it to a node. Also, the learning unit 212 uses the generalized action / function knowledge base 23f to abstract a specific action as a generalized action and map it to a node. The terminal node of the graph represents the predicted risk. This terminal node has size information as an attribute. This size information uses the cluster size obtained from the cause cluster analysis (here, the number of cases "30").

[0064] Furthermore, the control unit 21 of the support device 20 executes a risk prediction data expansion process S21 as a process accompanying the industrial accident knowledge memory process S20. Here, data expansion is achieved from quantitative and qualitative viewpoints. In this embodiment, the following two methods are used.

[0065] (a) A method of expanding the original case by extracting requirements corresponding to risks from the descriptions in near-miss reports and safety inspection reports other than industrial accident reports. (b) A method of expanding the whole based on a predetermined probability distribution. In method (a), compared with the comprehensive information obtained from the industrial accident report, the requirements corresponding to risks usually appear as partial information. For example, in the near-miss report 22b, detailed information such as age, experience, and causative agent included in the industrial accident report 22a is not included. Also, since it is not an actual accident, there is no accident type column, and the information is limited. In addition, in the safety patrol record report 22c, since it is the situation before work, the information is limited accordingly.

[0066] Therefore, the learning unit 212 generates a partial risk factor association graph obtained within the range of partial or limited information from the near-miss report 22b and the safety patrol record report 22c. As shown in FIG. 11, the partial risk factor association graph G10 generated from the near-miss report 22b is subsumed in the overall risk factor association graph G11. The learning unit 212 determines whether a certain overall graph subsumes a partial graph by comparing partial nodes. When it is determined that the risk factor association graph G11 subsumes the risk factor association graph G10, the learning unit 212 adds "1" to the original number of cases "28" to make it "29", thereby expanding the number of data.

[0067] Also, as shown in FIG. 12, the partial risk factor association graph G20 obtained from the safety patrol record report 22c is subsumed in the overall risk factor association graphs G21 and G22. The risk factor association graph G21 is an example where when a "roof worker" is performing "rebar assembly", they stumble over "# materials" scattered on the "# passage", leading to a "fall" accident. The risk factor association graph G22 is an example where a "geotechnical worker" has a "fall" accident while performing "transportation work". In such a case, the pointed-out situation of the risk factor association graph G20 may lead to both accidents. Therefore, the learning unit 212 expands the number of data by adding "1" to the number of cases of both.

[0068] In method (b), Heinrich's law is utilized. Heinrich's law states that for every one actual accident case, there are 29 minor events that did not lead to an accident behind it, and 300 abnormalities (near misses) in the background. The basic number of extended cases is generated using the number of these minor events. In this case, when the original number of cases is small (for example, 5 cases), the number of cases remains small (for example, 50 cases) even after expansion. Therefore, to ensure the significance as learning data, a threshold value is set to further perform data expansion. Specifically, an expansion threshold value (for example, 500 cases) is set for the minority cases, and cases with a number of cases less than the expansion threshold value are expanded to reach the expansion threshold value.

[0069] Also, for this expansion, it is also possible to pseudo-generate and expand cases that could potentially occur but did not actually occur. For example, for the work type "steel frame assembly", assume the case where three job types, namely "reinforcement worker", "carpenter", and "earthwork worker", are engaged.

[0070] Figure 13 shows the risk factor association graph G30 of "reinforcement worker" in the work type "steel frame assembly". Here, assume that "reinforcement worker" and "carpenter" have reports of encountering the same type of accident, while "earthwork worker" has not encountered the accidents encountered by these two job types. In terms of possibility, it is assumed that since they are engaged in the same type of work, there is a possibility of encountering the same type of accident. To enable risk prediction including such possibilities, data expansion is carried out based on similar examples.

[0071] Here, as shown in the frequency - probability distribution table T10 of Figure 14, the occurrence frequency and probability distribution by cause of the "collision / being collided" accidents actually encountered by "earthwork worker" are used, using the labor accident report 22a recorded in the report storage unit 22. This does not include those with the reason of "insufficient hazard prediction" encountered by "reinforcement worker" and "carpenter". Therefore, to quantitatively reflect the possibility of inclusion, data expansion is performed using the bias function shown in formula (1). freq(y)=a·min〔freq(xi)〕…(1) Here, "a" is the bias coefficient, "xi" is the cause factor i, and "freq(x)" is the frequency value.

[0072] Equation (1) is an equation for calculating the number of hypothetical accidents assuming that an accident that has not occurred occurs based on the minimum value in the original probability distribution. For example, when the bias coefficient is "0.8", the bias coefficient is multiplied by the minimum frequency value ("10" for "carelessness at the feet") to assume that the unoccurred "insufficient danger prediction" is "8 (= 10 * 0.8)".

[0073] As shown in FIG. 15, based on the assumed values, a frequency - probability distribution table T11 including the hypothetical occurrence frequency and the hypothetical occurrence probability is generated. In this case, as shown in FIG. 16, a virtual risk factor association graph G31 for "earthwork" is generated.

[0074] Furthermore, based on Heinrich's law, until the total number of No1 - No5 in the frequency - probability distribution table T11 reaches the threshold of 500 cases, additional hypothetical probability distributions are added to perform data expansion that virtually includes unoccurred cases. Thereby, the comprehensiveness of attention - arousal can be enhanced.

[0075] Next, the control unit 21 of the support device 20 executes the risk prediction learning model generation process S22. Specifically, the learning unit 212 of the control unit 21 creates learning data for prediction risk learning. As shown in FIG. 17, the configuration of the learning data V01 for prediction risk learning is composed of a prediction pattern vector and a correct label. For example, in a fully - connected neural network, the prediction pattern vector is composed of a causative agent region (L - dimensional), an action / function / state region (M - dimensional), and an accident type region (N - dimensional). Furthermore, by assigning a correct label indicating the correct answer, a set of learning data is generated. The correct label also has dimensions corresponding to the types of causes.

[0076] In each area of the prediction pattern vector, a vector is generated in which the corresponding elements are "1" and the non-corresponding elements are "0". In the risk factor association graph G01 of FIG. 10, only the positions of the elements corresponding to "#mobile scaffold", "#falling", "#moving inappropriately", and "falling / tripping" become "1". Also, this prediction pattern vector is generated in the number of numerical values of the prediction risk, here 30. The correct label becomes "1" only at the position of the element corresponding to the cause "insufficient risk awareness".

[0077] Then, as shown in FIG. 18, the learning unit 212 generates a risk prediction learning model 24b by performing neural network learning using the generated set of learning data as a learning data set, and records it in the learning result storage unit 24.

[0078] (Industrial accident data update process) Next, the industrial accident data update process S12 shown in FIG. 6 will be described. In the industrial accident knowledge extraction process S10, various knowledge bases that contribute to quantitative and qualitative risk prediction are constructed by analyzing past industrial accident reports at a certain point in time. If new industrial accidents continue to occur after a certain point in time, the number of industrial accident reports 22a will inevitably increase.

[0079] Therefore, similar to the industrial accident data analysis process S11, when the analysis unit 211 of the control unit 21 of the support device 20 detects the registration of a new industrial accident report 22a, the industrial accident data analysis process S11 is performed. Then, the learning unit 212 analyzes the new situation description text and the cause description text, and generates a risk factor association graph in the same manner as described above.

[0080] In this case, the learning unit 212 searches in the learning result storage unit 24 for a graph corresponding to the newly generated risk factor association graph. And when the learning result storage unit 24 extracts an existing risk factor association graph 24a equivalent to the newly generated risk factor association graph, the learning unit 212 adds "1" to the number of cases at the end of this risk factor association graph 24a. On the other hand, when no equivalent risk factor association graph 24a is extracted, the learning unit 212 newly registers it in the learning result storage unit 24 as a new risk factor association graph 24a.

[0081] Also, the administrator may manually update the new risk factor association graph 24a using the user terminal 10. For example, assume a case where a "drone" is used as a new piece of equipment (unknown equipment). Since there is no accident case E02 in FIG. 19 and the generalization target of the "drone" is unknown, a risk factor association graph cannot be generated.

[0082] Therefore, as shown in FIG. 20, the "drone" is mapped to the known generalization vocabulary "#surveying equipment". This "#surveying equipment" is also a generalization vocabulary of "level". As a result, as shown in FIG. 21, a graph equivalent to the risk factor association graph G41 corresponding to a known accident, for example, the accident case E03 in FIG. 19, is obtained. In this case, the update is completed by adding "1" to the number of cases N of the risk factor association graph G41. That is, even when the risk factor association graph 24a in the learning result storage unit 24 is not updated, an appropriate risk prediction for unknown equipment can be made.

[0083] Note that the update of the risk factor association graph 24a is performed separately from the update of the in-use risk prediction learning model 24b. And at a predetermined learning timing, the risk factor association graph 24a is reflected in the risk prediction learning model 24b. As this learning timing, an arbitrary time point determined by a person, the time point when the change amount of the risk factor association graph reaches a determined threshold value, etc. can be used.

[0084] (Risk candidate derivation process) Next, the risk candidate derivation process S30 in the support-time process shown in FIG. 7 will be described. Here, the support unit 213 derives risk candidates from the given conditions using the risk factor association graph 24a and the risk prediction learning model 24b recorded in the learning result storage unit 24. For example, assume that the given conditions are the job type "pole worker" and the work type "scaffold disassembly and assembly". Using the risk factor association graph 24a based on these elements, all corresponding risk factors are acquired.

[0085] Next, the support unit 213 generates a risk prediction vector for all corresponding risk factors. This risk prediction vector is composed of a causative agent area (L dimensions), an action / function / state area (M dimensions), and an accident type (N dimensions), similar to the risk prediction pattern vector in the industrial accident knowledge storage process S20.

[0086] Then, the support unit 213 calculates the probability of the predicted risk by inputting the generated risk prediction vector into a neural network inference device using the risk prediction learning model 24b. And as the predicted risk, the element with the highest probability is selected.

[0087] (Dynamic risk level measurement process) Next, the dynamic risk level measurement process S40 in the support-time process will be described. Here, for the predicted risk obtained in the risk candidate derivation process S30, the risk level is dynamically quantified according to various on-site situations, and the risks that should be noted are made apparent.

[0088] In this case, the support unit 213 of the control unit 21 uses the penalty function shown in Equation (2) to dynamically quantify the risk level. P = a1·P1 + a2·P2 + a3·P3 + a4·P4 + az·Pz…(2) This penalty function is composed of a plurality of penalty terms. an (n = 1, 2, 3, 4, z) are coefficients of 0 or more.

[0089] The reference penalty term P1 is a probability value for the frequency of the predicted risk held at the end of the risk factor association graph 24a, and means the basic likelihood of occurrence of this risk. The penalty term P2 is a term based on the probability distribution calculated by the probability distribution knowledge derivation process a112. This penalty term P2 is a probability value obtained by referring to the probability distribution knowledge base 23a, and represents the degree of the relationship between each attribute and the "type of accident".

[0090] The penalty term P3 is a term based on the association rule calculated by the association rule knowledge derivation process a122. This penalty term P3 is a probability value obtained by referring to the association rule knowledge base 23b, and represents the degree of the relationship between the combination of a plurality of attribute items and the "type of accident".

[0091] The penalty term P4 is a term based on environmental conditions. This penalty term P4 is a probability value obtained by referring to the probability distribution knowledge base 23a for cross-cutting items such as weather conditions not included in the penalty term P2.

[0092] Furthermore, the penalty term Pz is an arbitrary term given from the outside. This penalty term Pz is used, for example, when deliberately excluding a specific "predicted risk". For example, when "the risk of flying or falling due to solo work" is predicted, and the given conditions are originally premised on "being solo work", a function that lowers the probability value is given to exclude it from the predicted risk.

[0093] As shown in FIG. 22, in the sequence P11 of the 20s age group with respect to the sequence P10 based on the basic penalty value, "flying / falling due to intrusion into the lower part of the suspended load", "flying / falling due to incomplete work procedures", and "flying / falling due to insufficient maintenance of equipment / tools" rank higher. That is, in the 20s age group, the situations that require particular attention due to the lack of knowledge about risks are becoming apparent. Also, in the sequence P12 of the 50s age group, "falling / tumbling due to non-use of safety belts", "falling / tumbling due to insufficient confirmation of work procedures", and "falling / tumbling due to taking shortcuts" rank higher. That is, in the 50s age group, the situations that require particular attention due to risks caused by habit are becoming apparent.

[0094] This exemplification is an example of measuring the risk level for individual attribute items. However, when identifying the targets that require attention regarding the combination of multiple attribute items, the association rule knowledge base 23b is used.

[0095] (Dynamic risk information presentation process) Next, the dynamic risk information presentation process S50 in the support-time process will be described. Specifically, the "predicted risk" obtained by the risk candidate derivation process S30 and the dynamic risk level measurement process S40 is presented as attention-calling information according to the user's situation.

[0096] Here, as shown in FIG. 23, the support unit 213 of the control unit 21 outputs a support screen 500 as a user interface to the user terminal 10. First, the user selects the item values of the occupation candidate column 501 and the work type candidate column 502 from the selectable lists respectively. The support unit 213 derives the causative agent candidates using the risk factor association graph 24a recorded in the learning result storage unit 24 starting from the selected "occupation" and "work type". Then, the support unit 213 displays the derived causative agent candidates in the causative agent candidate column 503 of the support screen 500.

[0097] Next, the user selects the causative agent to be the target of risk prediction from the causative agent candidate column 503 and displays it in the causative agent column 504. After the selection of the causative agent is completed, if the "Predict" button is pressed, the support unit 213 displays the predicted risk in the prediction result column 505.

[0098] This prediction result column 505 is composed of columns for the predicted risk, risk level, and caution information. In the risk level column, the penalty value corresponding to each risk is displayed. In the caution information column, marks that should be individually noted are displayed.

[0099] As shown in FIG. 24, assume that "roofing worker" is selected in the job type candidate column 501 of the support screen 500 and "scaffold disassembly and assembly" is selected in the work type candidate column 502. In this state, when the "Acquire" button is pressed, causative agent candidates such as "construction materials" and "fixed scaffold" are displayed in the causative agent candidate column 503. When the user selects the target causative agent from the causative agent candidate column 503 and presses the "Select" button, the causative agent selected by the user is displayed in the selected causative agent column 504. Further, when the "Predict" button is pressed, the predicted risk and attached information (risk level and caution information) are displayed in the prediction result column 505.

[0100] In the caution information of the prediction result column 505, the penalty value (risk level) calculated using the penalty function is high, and the content that should be particularly noted is displayed as an icon. For example, [fall due to "insufficient foot check"] should be particularly noted because it is particularly likely to occur when it is raining, so "rain" is presented. Also, [flying / falling due to "working up and down"] should be particularly noted by "young" manual workers, so "young" is presented. Also, [falling / tripping due to "not using a safety belt"] should prompt re-recognition by skilled workers, so "skilled" is presented.

[0101] 7, the control unit 21 of the support device 20 performs a risk elicitation information expansion process S51 and a risk elicitation information generation process S52. The risk elicitation information expansion process S51 is a process for presenting typical disaster cases related to a predicted risk. The risk elicitation information generation process S52 is a process for generating and presenting an image (reference image) of a typical disaster case related to the predicted risk.

[0102] In the risk alert information expansion process S51, typical disaster cases and reference images are output to the display device H13 of the user terminal 10. For example, as shown in FIG. 24, when any predicted risk in the prediction result field 505, for example, a fall or trip due to "not using safety belt", is "selected", one typical accident case related to the risk is displayed in the typical accident case field 506. For example, "Fall due to not using safety belt in a place where the braces have been removed" is displayed. In this case, an image of the typical accident case may be displayed in the reference image field 507. When there are multiple typical accident cases related to the risk, the forward and backward buttons may be used to display the typical accident cases in order.

[0103] Next, the control unit 21 of the support device 20 executes a risk elicitation information generation process S52. Specifically, the support unit 213 of the control unit 21 generates and displays an image of a risk that may occur due to the corresponding "causing object." The support unit 213 generates various accident images that may be caused by the causing object using a deep learning GAN (generative adversarial network) or the like, and displays them in the reference image field 507. For example, in the case of "foundation work," if the causing object "drag shovel" is identified, the support unit 213 generates and outputs images of accidents such as a worker being caught or caught in the drag shovel, an accident in which the worker is hit by the machine, an accident in which the worker falls from the machine, and the like.

[0104] According to this embodiment, the following effects can be obtained. (1) In this embodiment, the control unit 21 of the support device 20 executes the industrial accident knowledge extraction process S10. Here, it executes the industrial accident data analysis process S11. As a result, using the reports recorded in the report storage unit 22, the existing data such as past industrial accident reports can be effectively utilized. And by suppressing manual input and the like, the burden of initial system construction can be reduced and the efficiency of system maintenance can be improved. Therefore, various risk elements can be systematically turned into knowledge. And even when there are the diversity of objects, the large number of attribute items, convergence, etc., situations that can be regarded as the same for preventing recurrence based on past accident situations can be efficiently identified. For example, under the situation of the diversity and scale of buildings and manufactured products targeted by industries such as the construction industry and the automobile industry, the large number of involved job types and work types, and the enormity of causative objects that are the causes of accidents, data on individual accident cases can be classified. Also, the arbitrariness of each input person can be eliminated.

[0105] (2) In this embodiment, the control unit 21 of the support device 20 executes the industrial accident knowledge storage process S20. As a result, even when there are the diversity of objects, the large number of attribute items, convergence, etc., information for preventing recurrence based on past accident situations can be generated. Different from the case of using a uniform scale for risks as the only indicator, information that should be noted for individual workers in different situations and characteristics placed at the site can be suggested, and information for preventing industrial accidents in advance can be generated. Different from the case where a person arbitrarily assigns keywords, in an object industry that includes large-scale and complex relationships, systematization and control can be achieved. And the quality of appropriate search results can be ensured.

[0106] (3) In this embodiment, the control unit 21 of the support device 20 executes risk prediction data expansion processing S21. Generally, neural network learning requires a relatively large amount of learning data. However, in the field of industrial accidents, where the quality of data is also an important requirement, there are variations in the number of accident cases with respect to the expected risks in terms of quantity, and it may not be possible to ensure a sufficient amount of learning. Also, in terms of quality, simply converting reported cases into data faithfully may narrow the range of predicted risks. Even in this case, the data required for neural network learning can be ensured from both quantitative and qualitative perspectives.

[0107] (4) In this embodiment, the control unit 21 of the support device 20 executes risk candidate derivation processing S30. As a result, according to the job type and work type, risk factors that may cause disasters can be identified using the risk factor association graph 24a. Furthermore, using the risk prediction learning model 24b, the level of risk occurrence (risk degree) corresponding to the risk factors can be calculated.

[0108] (5) In this embodiment, the control unit 21 of the support device 20 executes dynamic risk level measurement processing S40. As a result, for risk candidates, the degree of risk can be dynamically quantified according to the worker and the situation.

[0109] (6) In this embodiment, the control unit 21 of the support device 20 executes dynamic risk information presentation processing S50. For example, the accidents that young workers are likely to encounter are different from those that experienced workers are likely to encounter. Therefore, using evaluation indicators according to each occurrence frequency, it is possible to alert workers to the risks that should be avoided as the parties affected by the disaster.

[0110] (7) In this embodiment, the control unit 21 of the support device 20 executes risk alert information expansion processing S51. As a result, it is possible to prompt the worker to be alerted so that the risk is felt more closely.

[0111] (8) In this embodiment, the control unit 21 of the support device 20 executes the risk arousal information generation process S52. As a result, it is possible to flexibly arouse attention regarding the possibility of risks for any target person and situation.

[0112] This embodiment can be implemented with the following modifications. This embodiment and the following modification examples can be implemented in combination with each other as long as they do not technically conflict. · In the above embodiment, it is used when supporting safety assurance by using a large number of past labor accident reports at a building construction site. The application target of the present invention is not limited to labor accidents at construction sites.

[0113] · In the above embodiment, each file of the labor accident report 22a, near miss report 22b, and safety patrol record report 22c is used. As long as it is a report in which the attributes of the victim, the occurrence status of the injury, the cause of occurrence, etc. are recorded, it is not limited to these.

[0114] · In the above embodiment, a knowledge base is generated using the information recorded in the report storage unit 22. The information used for generating the knowledge base is not limited to these. The control unit 21 of the support device 20 may acquire the information lacking in the report. For example, when the weather information is lacking, the control unit 21 identifies the address of the work place and acquires the weather information from a weather information site publicly available on the Internet or the like.

[0115] · In the above embodiment, the control unit 21 of the support device 20 executes the risk candidate derivation process S30. In this case, the occupation and work type input to the user terminal 10 are received. In addition to this, risks may be specified based on the user using the user terminal 10. In this case, the user is identified in the login authentication when using the user terminal 10. Then, user attributes (for example, age, skill, etc.) are acquired from the user management server, and the risk level corresponding to the user attributes is calculated.

[0116] In addition, the method of inputting various types of information into the user terminal 10 is not limited to using text. For example, a two-dimensional code image or the like that holds the work type may be captured by the camera of the user terminal 10. Also, at the work site, by photographing actual objects such as cranes, trucks, reinforcing bars, and earth and sand, related equipment and materials may be identified as causative objects by image recognition. In addition, the method of presenting information in the dynamic risk information presentation process is not limited to presenting on the screen of a case. For example, a checklist including countermeasures against risks may be generated and displayed. Also, an attention call list may be read out by voice.

[0117] ·In the above embodiment, the control unit 21 of the support device 20 executes the dynamic risk information presentation process S50. Here, an evaluation is performed based on the occurrence frequency. Further, an evaluation may be performed according to the degree of influence. In this case, a penalty term corresponding to the degree of influence is added to the penalty function. In this case, in the industrial accident report, the degree of influence is added and cross-analysis and correlation analysis are performed. For example, the same accident situation may result in a minor injury with 0 days off work for one worker and a serious injury with 30 days off work for another worker. Therefore, it is possible to call attention to the risks that should be avoided for the workers who may be the victims of the disaster, using an evaluation index according to the degree of influence and the occurrence frequency.

[0118] Next, the technical ideas that can be grasped from the above embodiment and the alternative example are added below together with their effects. (a) An industrial accident knowledge extraction process for extracting the analysis results obtained by the cooperative analysis of humans and computers by data analysis means based on data science from industrial accident reports reported in the past, An industrial accident knowledge storage process for storing the analysis results as industrial accident knowledge in a computer, A risk candidate derivation process for deriving risk candidates that are predicted to occur under arbitrary conditions based on the industrial accident knowledge, A dynamic risk level derivation process for calculating the risk level according to individual situations for the risk candidates A safety support system characterized by executing a dynamic risk information presentation process for presenting risk information according to individual situations during preventive safety activities.

[0119] (b) The safety support system according to (a), wherein the computer executes, in the industrial accident knowledge extraction process, an industrial accident data automatic analysis process for automatically analyzing a newly generated industrial accident report, and an industrial accident knowledge automatic update process for reflecting the new analysis result in existing industrial accident knowledge. Thereby, industrial accident knowledge can be gradually expanded.

[0120] (c) The safety support system according to (a) or (b), wherein the computer executes, in the industrial accident knowledge storage process, a risk prediction data expansion process for expanding learning data for generating a risk prediction learning model used in the risk candidate derivation process, and a risk prediction learning model generation process for generating a risk prediction learning model from the expanded risk prediction data. Thereby, a learning function for improving the accuracy and comprehensiveness of risk prediction can be realized.

[0121] (d) The safety support system according to any one of (a) to (c), wherein the computer executes, in the dynamic risk information presentation process, a risk evoking information expansion process for selectively providing any information that gives a specific awareness of the predicted risk, and a risk evoking information generation process for generating an image for the predicted risk. Thereby, a support function for improving the sensitivity to risks can be realized.

[0122] (e) The safety support system according to claim 1, wherein the control unit generates risk prediction information including association information that associates the work, the causative agent, and the type of accident by specifying the causative agent and the type of accident used in the work using elements obtained by generalizing the words included in the report.

[0123] (f) The safety support system according to (e), wherein the control unit generates risk prediction information including a risk prediction learning model that predicts the probability of the causative agent and the type of accident from the work using the association information.

[0124] (g) The safety support system according to claim 1, (e), and (f), wherein when the control unit acquires a new report, the control unit generates risk prediction information based on the analysis result of analyzing the new report and records the risk prediction information in the knowledge base storage unit.

Explanation of Signs

[0125] 10... User terminal, 20... Support device, 21... Control unit, 211... Analysis unit, 212... Learning unit, 213... Support unit, 22... Report storage unit, 23... Knowledge base storage unit, 24... Learning result storage unit

Claims

1. A knowledge base storage unit that records the probability distribution of a combination of elements including occupation type, work type, age, causative agent, accident type, cause type, and occurrence status as an analysis result obtained by analyzing using the elements included in the report; A learning result storage unit that records a risk factor association graph, which is risk prediction information for predicting a risk that is a cause type corresponding to an accident type, and a risk prediction learning model, based on the analysis result; A safety support system comprising a control unit connected to a user terminal, wherein the control unit uses the occupation type, work type, causative agent, action / function / status, accident type, and cause type recorded in the knowledge base storage unit to connect an occupation node and a work type node corresponding to the occupation type and the work type respectively as starting nodes, intermediate nodes corresponding to the causative agent, the action / function / status, and terminal nodes corresponding to the cause type corresponding to the accident type, thereby generating a risk factor association graph; generates a risk prediction learning model learned using a learning data set consisting of correct labels corresponding to the causative agent, action / function / status, accident type, and cause type; a learning unit that records the risk factor association graph and the risk prediction learning model in the learning result storage unit; when obtaining the occupation type and work type regarding the content of a new work from the user terminal, specifying the causative agent and action / function as risk factors using the risk factor association graph from the occupation type and work type, and specifying the cause type corresponding to the accident type as a risk candidate using the risk prediction learning model from the specified causative agent and action / function; a support unit that outputs to the user terminal a risk level indicating the height of risk occurrence, calculated using the probability distribution of a combination of elements including age, causative agent, and occurrence status recorded in the knowledge base storage unit as a penalty term for the cause type corresponding to the accident type. A safety support system characterized by comprising the above.

2. A knowledge base storage unit that records the probability distribution of a combination of elements including occupation type, work type, age, causative agent, accident type, cause type, and occurrence status as an analysis result obtained by analyzing using the elements included in the report; A learning result storage unit that records a risk factor association graph, which is risk prediction information for predicting a risk that is a cause type corresponding to an accident type, and a risk prediction learning model, based on the analysis result; A method for providing support for safety management using a safety support system comprising a control unit connected to a user terminal, The control unit uses the occupation type, work type, causative agent, action / function / status, accident type, and cause type recorded in the knowledge base storage unit to connect the occupation node and work type node corresponding to the occupation type and work type respectively as starting nodes, the intermediate nodes corresponding to the causative agent, action / function / status, and the terminal nodes corresponding to the cause type according to the accident type, thereby generating a risk factor association graph, generates a risk prediction learning model learned using a learning data set consisting of correct labels corresponding to the causative agent, action / function / status, accident type, and cause type, a learning-time process of recording the risk factor association graph and the risk prediction learning model in the learning result storage unit, when obtaining the occupation type and work type regarding the content of a new operation from the user terminal, uses the risk factor association graph based on the occupation type and work type to identify the causative agent, action / function as risk factors, and uses the risk prediction learning model from the identified causative agent, action / function to identify the cause type according to the accident type as a risk candidate, a support-time process of outputting to the user terminal a risk degree indicating the height of risk occurrence, which is calculated using, as a penalty term, the probability distribution of a combination of elements including age, causative agent, and occurrence situation recorded in the knowledge base storage unit for the cause type according to the accident type. A safety support method is characterized by executing the above.

3. A knowledge base storage unit that records the probability distribution of a combination of elements including occupation type, work type, age, causative agent, accident type, cause type, and occurrence situation as an analysis result analyzed using the elements included in the report, A learning result storage unit that records a risk factor association graph and a risk prediction learning model, which are risk prediction information for predicting a risk that is a cause type according to the accident type, based on the analysis result, A program for performing safety management support using a safety support system including a control unit connected to a user terminal, The control unit uses the occupation type, work type, causative agent, action / function / status, accident type, and cause type recorded in the knowledge base storage unit to connect the occupation node and work type node corresponding to the occupation type and work type respectively as starting nodes, the intermediate nodes corresponding to the causative agent, action / function / status, and the terminal nodes corresponding to the cause type according to the accident type, thereby generating a risk factor association graph, Generate a risk prediction learning model trained using a learning data set consisting of correct labels corresponding to the causative agent, action / function / state, type of accident, and cause type. A learning unit that records the risk factor association graph and the risk prediction learning model in the learning result storage unit. When the occupation and work type regarding the content of a new task are acquired from the user terminal, from the occupation and work type, using the risk factor association graph, identify the causative agent, action / function as risk factors, and from the identified causative agent, action / function, use the risk prediction learning model to identify the cause type corresponding to the type of accident as a risk candidate. A support unit that outputs to the user terminal a risk level indicating the likelihood of risk occurrence, calculated using, as a penalty term, the probability distribution of a combination of elements including age, causative agent, and occurrence situation recorded in the knowledge base storage unit for the cause type corresponding to the type of accident. A safety support program characterized by functioning as such.

Images