An artificial intelligence-based work injury identification method

Abstract

The application relates to the technical field of artificial intelligence, and discloses a work injury identification method based on artificial intelligence, which comprises the following steps: constructing a causal Bayesian network from historical work injury data; receiving an accident text description as input; based on the causal Bayesian network, calculating the posterior probability of a work injury variable by using a Bayesian updating mechanism, and identifying the work injury when the posterior probability exceeds a preset threshold; integrating real-time data, and calculating the intervention probability of injury consequences under a set intervention variable by using intervention simulation; and when new data is received, only the node parameters affected in the causal Bayesian network are incrementally updated by using a Bayesian parameter updating mechanism. By constructing a causal Bayesian network containing six types of nodes, namely, accident trigger factors, job risk, personal health indicators, regulatory requirements, environmental factors and injury consequences, unified modeling of the multi-factor relationship of accidents is realized.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence technology, and in particular to a method for determining work-related injuries based on artificial intelligence. Background Technology

[0002] In the fields of safety production management and workers' compensation insurance, the analysis of accident processes typically requires a comprehensive consideration of multiple factors, including work behavior, environmental conditions, job risks, health status, and relevant regulatory requirements. As enterprises become increasingly data-driven, accident information recording methods are gradually evolving from text-based records to structured data records, among others. Against this backdrop, how to uniformly process and analyze accident-related factors from multiple sources and of multiple types has become a technical task that industrial safety management must address. Currently, accident causal analysis technology is increasingly being introduced into systems used for risk assessment and workers' compensation determination. Its core objective is to reveal the correlations between accident factors through a data-driven approach, thereby assisting in subsequent judgments.

[0003] Based on existing technologies, work injury determination tasks typically rely on rule matching, human experience inference, or simple statistical models. For example, some technical solutions use a fixed rule base to identify keywords in accident texts and directly use the identified words as the basis for judgment; other solutions infer accident-related factors based on the statistical frequency of specific fields.

[0004] However, in the process of realizing the technical solution of this application, the inventors of this application discovered that the above-mentioned technology has at least the following technical problems: the existing technology lacks a unified modeling mechanism that can simultaneously represent the causal dependence between multiple variables such as accident triggering factors, job risks, health indicators, regulatory requirements and environmental factors, which makes it impossible for the system to form a causal model that is both structured and inferable based on historical work injury data, thereby affecting the accuracy and consistency of the probability inference results involved in the work injury determination process. Summary of the Invention

[0005] To overcome the above shortcomings, this invention provides an artificial intelligence-based method for determining work-related injuries, aiming to improve the existing technology's lack of a unified modeling mechanism that can simultaneously represent the causal dependencies between multiple variables such as accident triggering factors, which leads to the inaccuracy and consistency of probability inference results involved in the work-related injury determination process.

[0006] In a first aspect, the present invention provides the following technical solution: a method for determining work-related injuries based on artificial intelligence, comprising the following steps:

[0007] S1. Construct a causal Bayesian network from historical work injury data. The causal Bayesian network is a directed acyclic graph, including a set of nodes and a set of edges. Each node in the set of nodes represents an accident-related variable, and the set of edges represents the causal dependency between accident-related variables.

[0008] S2. Receive the accident text description as input, and use rule-based parsing to map the keywords in the accident text description to the values ​​of corresponding accident-related variables to form a set of evidence variables;

[0009] S3. Based on the causal Bayesian network, the posterior probability of the work injury variable is calculated using the Bayesian update mechanism. When the posterior probability exceeds a preset threshold, it is identified as a work injury, and the maximum probability path from the root node to the work injury variable is output as the causal path explanation.

[0010] S4. Integrate real-time data and use intervention simulation to calculate the probability of injury consequences under set intervention variables, and generate personalized work injury prevention reminders based on the probability difference before and after intervention.

[0011] S5. When new data is received, the Bayesian parameter update mechanism is used to incrementally update only the parameters of the affected nodes in the causal Bayesian network.

[0012] Preferably, the specific process of constructing a causal Bayesian network from historical work injury data in S1 includes:

[0013] The fractional search algorithm is used to search for the optimal network structure by maximizing the network's fractional function, and the conditional probability table for each node is calculated using maximum likelihood estimation.

[0014] When parsing accident text descriptions, semantically similar words are mapped to the same accident-related variable nodes to achieve unified processing of semantic variations.

[0015] Preferably, the specific process of forming the evidence variable set through rule-based parsing in S2 includes:

[0016] Beforehand, mapping rules are established between keywords and accident-related variables. "Slipping," "falling," and "dropping" are uniformly mapped to environmental factor nodes, "working at height" and "hoisting operations" are mapped to accident triggering factor nodes, and "myocardial infarction" and "heatstroke" are mapped to personal health indicator nodes, thereby obtaining a structured set of evidence variables.

[0017] Preferably, the specific process of calculating the posterior probability of the work-related injury variable using the Bayesian update mechanism in S3 includes:

[0018] Based on the conditional probability table of causal Bayesian networks, the prior probability of the work-related injury variable, the likelihood of the evidence variable set, and the marginal probability of the evidence variable set are calculated sequentially.

[0019] The posterior probability of the work injury variable is obtained through the posterior probability calculation formula.

[0020] The preset threshold can be adjusted according to regulatory requirements.

[0021] Preferably, the specific process of outputting the maximum probability path from the root node to the work injury variable as the causal path explanation in S3 includes:

[0022] A depth-first search algorithm is used to traverse all paths leading to the work injury variable in the causal Bayesian network, calculate the path probability of each path, and select the path with the highest path probability as the final causal path explanation.

[0023] Preferably, the specific process of integrating real-time data and employing intervention simulation in S4 includes:

[0024] Personal health records, industry statistics, and seasonal meteorological data are input into a causal Bayesian network as real-time data. Based on do-calculus operations, the intervention variable is set as "implementing protective measures" or "not implementing protective measures". The probability of intervention for the injury consequences in the two situations is calculated and the difference between the two intervention probabilities is compared.

[0025] Preferably, the specific process of generating personalized work injury prevention reminders based on the probability differences before and after intervention in S4 includes:

[0026] When the probability of intervention for the consequences of injury after taking protective measures is significantly lower than the probability of not taking them, a text reminder containing specific protective measures will be automatically generated based on the individual's age, job position, real-time health indicators, and seasonal factors.

[0027] Preferably, the specific process of incremental update using the Bayesian parameter update mechanism in S5 includes:

[0028] A pseudo-count is introduced for each node as a Dirichlet prior. When new data is introduced, only the nodes related to the new data and their conditional probability tables are locally updated, while the structure and parameters of other parts of the network remain unchanged.

[0029] Preferably, the accident-related variables in the node set include six variables: accident triggering factors, job risks, personal health indicators, regulatory requirements, environmental factors, and injury consequences.

[0030] Preferably, the causal Bayesian network forms a closed loop between work injury determination and work injury prevention. The causal path explanation and determination results generated during the work injury determination stage are fed back to S5 as new data for subsequent Bayesian parameter update mechanisms, thereby achieving continuous adaptive optimization of the network.

[0031] The present invention has the following beneficial effects:

[0032] 1. This invention constructs a causal Bayesian network containing six types of nodes: accident triggering factors, job risks, personal health indicators, regulatory requirements, environmental factors, and injury consequences. Based on historical work injury data, it learns the network structure and conditional probability table, thereby achieving unified modeling of the multi-factor relationships in accidents.

[0033] 2. This invention generates a set of evidence variables through rule parsing and semantic normalization, and calculates the posterior probability of the work injury variable by combining it with a Bayesian update mechanism. At the same time, it outputs the maximum probability path from the root node to the work injury variable, so that the work injury determination result has a clear causal link explanation.

[0034] 3. This invention introduces intervention simulation to calculate the intervention probability of injury consequences, and combines it with real-time data to generate personalized prevention reminders. At the same time, it feeds back the work injury determination results and causal path as new data to the Bayesian parameter incremental update mechanism, forming a closed loop between work injury determination and work injury prevention. Attached Figure Description

[0035] Figure 1 This is a flowchart of an artificial intelligence-based work injury determination method proposed in this invention;

[0036] Figure 2 This is a flowchart of the accident text parsing and evidence variable generation process for an artificial intelligence-based work injury determination method proposed in this invention. Detailed Implementation

[0037] The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0038] Reference Figure 1 and Figure 2 In the first embodiment of the present invention, the present invention provides an artificial intelligence-based method for determining work-related injuries, comprising the following steps:

[0039] S1. Construct a causal Bayesian network from historical work injury data. The causal Bayesian network is a directed acyclic graph, which includes a set of nodes and a set of edges. Each node in the set of nodes represents an accident-related variable, and the set of edges represents the causal dependency between accident-related variables.

[0040] S2. Receive the accident text description as input, and use rule-based parsing to map the keywords in the accident text description to the values ​​of corresponding accident-related variables to form a set of evidence variables;

[0041] S3. Based on the causal Bayesian network, the Bayesian update mechanism is used to calculate the posterior probability of the work injury variable. When the posterior probability exceeds the preset threshold, it is identified as a work injury, and the maximum probability path from the root node to the work injury variable is output as the causal path explanation.

[0042] S4. Integrate real-time data and use intervention simulation to calculate the probability of injury consequences under set intervention variables, and generate personalized work injury prevention reminders based on the probability difference before and after intervention.

[0043] S5. When new data is received, the Bayesian parameter update mechanism is used to incrementally update only the parameters of the affected nodes in the causal Bayesian network.

[0044] Specifically, S1 is used to construct a causal Bayesian network from historical work injury data. S2 is used to convert the accident text description into a set of evidence variables. S3 is used to calculate the posterior probability of the work injury variables based on the causal Bayesian network and output the causal path. S4 is used to generate work injury prevention reminders based on intervention calculations. S5 is used to update the node parameters of the causal Bayesian network after receiving new data.

[0045] In S1, accident-related variables are extracted from historical work injury data to construct a causal Bayesian network. The causal Bayesian network is a graph structure with directed acyclic properties, denoted as […]. Among them, the set A set of nodes, where each node represents an accident-related variable; This is a set of edges used to represent causal dependencies between nodes. For each node... The set of its parent nodes is denoted as Given the value of the parent node, the node The conditional probability is expressed as:

[0046] ;

[0047] The above conditional probabilities were determined using maximum likelihood estimation based on historical work injury data. If the node... The number of times the combined value of the node and its parent node appears are respectively and The conditional probability can then be expressed as:

[0048] (1)

[0049] In the formula, This indicates a node in historical data. The number of times it appears simultaneously with its parent node. This represents the number of times the parent node combination occurs. The network structure is determined based on a score-oriented search process, by maximizing the score function.

[0050] (2)

[0051] In the formula, This represents a historical workplace injury dataset. Represents a node Local scores given a parent node, calculated based on the Bayesian information criterion.

[0052] In S2, the accident text description is received as input, and keywords are identified from the text using the rule-based Pasedi parsing method. Based on pre-defined mapping rules between keywords and accident-related variables, words appearing in the text are assigned to the values ​​of corresponding accident variable nodes. For example, words belonging to the environmental factors category, the accident triggering factors category, or the personal health indicators category are mapped to the values ​​of their respective nodes. This parsing process generates a set of evidence variables. Each element corresponds to a specific observation value of a node in a causal Bayesian network.

[0053] In S3, the posterior probability of the work-related injury variable is calculated based on the causal Bayesian network generated in S1 and the set of evidence variables generated in S2. Let the work-related injury variable be... Its prior probability is In the evidence set The posterior probability is:

[0054] (3)

[0055] In the formula, This represents the conditional probability of the evidence set appearing when the value of the work-related injury is a certain value. The marginal probability of the evidence set is represented by the following calculation method:

[0056] (4)

[0057] in, Let be the combination of variables related to evidence in the network. After obtaining the posterior probability, if ... If the accident is not serious, it will be classified as a work-related injury, where... A preset threshold is used. To provide a causal explanation, a depth-first search is performed on the causal Bayesian network to determine a path from the root node to the work injury variable node. Let the edges on the path be as follows: Each edge The starting node is Termination node is The path probability is expressed as:

[0058] (5)

[0059] The path with the highest probability is taken as the causal path of the output.

[0060] In S4, real-time data is input into the causal Bayesian network. Real-time data includes personal health record information, industry statistics, and meteorological data related to the work environment. The intervention variable is set as follows: ,node This represents the consequences of verbal harm. Given an intervention action, the probability of intervention for the harmful consequences is expressed as:

[0061] ;

[0062] in, This represents the set of moderating variables in the network that are unrelated to the intervention. The probability of intervention is calculated for different intervention values, and prevention alerts are generated by comparing the differences in intervention probabilities. The content of the prevention alerts is generated based on changes in the probability of harm consequences and variable information of the current party involved.

[0063] In S5, when new data is input into the network, a Bayesian parameter update mechanism is used to incrementally update the conditional probability tables of the affected nodes. For each node... Set pseudo-count parameters As a Dirichlet prior, the new dataset is denoted as... The updated conditional probability is expressed as:

[0064] (6)

[0065] In the formula, Indicates nodes in the new data The number of times it appears in combination with its parent node. This represents the number of times the parent node combination appears in the new data. Incremental learning is achieved by updating only the node parameters relevant to the new data, while maintaining the stability of the causal Bayesian network structure.

[0066] Furthermore, the specific process of constructing a causal Bayesian network from historical work injury data in S1 includes:

[0067] The fractional search algorithm is used to search for the optimal network structure by maximizing the network's fractional function, and the conditional probability table for each node is calculated using maximum likelihood estimation.

[0068] When parsing accident text descriptions, semantically similar words are mapped to the same accident-related variable nodes to achieve unified processing of semantic variations.

[0069] Specifically, firstly, variable fields for constructing a causal Bayesian network are extracted from historical work injury data. Historical work injury data is stored as multiple accident records, each containing data items such as accident triggering factors, job risks, personal health indicators, regulatory requirements, environmental factors, and injury consequences. For each data item, it is converted into discrete values ​​according to preset discretization rules. Discretization rules include dividing continuous variables into intervals, such as dividing them into different segments based on the range of heart rate indicators, or dividing them into multiple segments based on the range of temperature indicators. The discretized variables serve as the candidate node set for the causal Bayesian network. Secondly, a score-oriented search algorithm is used to determine the structure of the causal Bayesian network. Let the candidate node set be denoted as... The network structure to be searched is ,in, Let be the set of directed edges between nodes. The scoring function during the search process is set using formula (2).

[0070] The local scoring item is calculated based on the Bayesian information criterion as follows:

[0071] ;

[0072] in, For nodes The likelihood function given the parent node. This represents the number of independent parameters in the conditional probability table for that node. The number of historical data samples is denoted as . The search algorithm continuously increases the score function by performing operations such as adding, deleting, and reversing edges on the candidate structure set until the score function stops increasing, thus obtaining the optimal causal structure that satisfies the acyclicity requirement.

[0073] Then, after determining the network structure, a conditional probability table is calculated for each node. The conditional probabilities are calculated using maximum likelihood estimation, based on historical work injury data, by statistically analyzing the frequency of combinations of nodes and their parent nodes. The value of the parent node is... The conditional probability is represented by formula (1). The above method is applied once to all nodes and all parent nodes to obtain a complete set of conditional probability tables.

[0074] Furthermore, the specific process of forming the evidence variable set through rule-based parsing in S2 includes:

[0075] Beforehand, mapping rules are established between keywords and accident-related variables. "Slipping," "falling," and "dropping" are uniformly mapped to environmental factor nodes, "working at height" and "hoisting operations" are mapped to accident triggering factor nodes, and "myocardial infarction" and "heatstroke" are mapped to personal health indicator nodes, thereby obtaining a structured set of evidence variables.

[0076] Specifically, before system deployment, mapping rules between keywords and accident-related variables are pre-established. These mapping rules are constructed based on the node definitions of accident-related variables in a causal Bayesian network. For environmental factor variables, terms like "slippery," "fall," "trip," and "slippery road surface" are encoded as values ​​of the same category for the environmental factor node. For accident triggering factor variables, terms like "working at height," "hoisting operations," "lifting operations," and "working in confined spaces" are encoded as corresponding values ​​for the accident triggering factor node. For personal health indicator variables, terms like "myocardial infarction," "heatstroke," "chest tightness," and "exhaustion" are encoded as corresponding values ​​for the health indicator node. The mapping rules are stored as key-value pairs, allowing the parsing process to directly search based on keywords.

[0077] Secondly, to maintain a consistent input data format when parsing the accident text, word segmentation is performed. A dictionary-based word segmentation algorithm is used to split the text into word sequences, and each word is tagged with its part of speech. Let the accident text be a string. The word segmentation result is represented as

[0078] ;

[0079] in, Indicates the first word after the segmentation. Each word. Further determine whether it belongs to the keyword set in the mapping rule table. If it belongs to the mapping rule table, fill its mapping value into the corresponding accident-related variable node.

[0080] Then, to handle word variations, synonyms, or differences in habitual expressions within the text, a semantic normalization mechanism is introduced during the parsing process. For words not directly included in the mapping rule table, word vector similarity is used to determine their corresponding accident-related variable categories. Let the unknown word vector be represented as... The set of existing word vectors for a certain category in the mapping rule table is represented as follows: The similarity between an unknown word and words of that category is calculated as follows:

[0081] ;

[0082] Here, ∥⋅∥ represents the vector norm. When the similarity between an unknown word and a certain set of words exceeds a preset threshold, the unknown word is assigned to the corresponding value of the variable node related to the accident. This method allows for the unified classification of words not listed in the rule table but with the same accident meaning, ensuring the stability of the generated evidence variables.

[0083] During the parsing process, the text is scanned sequentially, and for each matched word, a corresponding evidence variable is generated. The set of evidence variables is denoted as... Each of the evidence variables This is a formatted key-value pair format, including node identifiers and node values. For example, when the term "high-altitude work" is identified, an evidence variable is generated. ,in, Indicates the nodes that trigger the accident. This represents the discretized values ​​for high-altitude operations. Corresponding environmental factor nodes, health indicator nodes, and job risk nodes are all recorded in the same way.

[0084] Subsequently, to avoid redundancy in the evidence set due to the repetition of the same words in the text, the generated evidence variable set was deduplicated. When the same node has multiple values, a node priority rule is applied. If a node is a single-value node, the first matching value is retained; if a node can have multiple values, a set structure is used to record all values. The node priority rule is consistent with the network node definition.

[0085] After the above analysis process, a set of structured evidence variables corresponding one-to-one with the nodes of the causal Bayesian network is obtained. This set can be directly input into the Bayesian inference module in S3 to calculate the posterior probability of the work injury variable and search for the causal path.

[0086] Furthermore, the specific process of calculating the posterior probability of the work injury variable using the Bayesian update mechanism in S3 includes:

[0087] Based on the conditional probability table of causal Bayesian networks, the prior probability of the work-related injury variable, the likelihood of the evidence variable set, and the marginal probability of the evidence variable set are calculated sequentially.

[0088] The posterior probability of the work injury variable is obtained through the posterior probability calculation formula.

[0089] The preset threshold can be adjusted according to regulatory requirements;

[0090] The specific process of outputting the maximum probability path from the root node to the work injury variable in S3 as a causal path explanation includes:

[0091] A depth-first search algorithm is used to traverse all paths leading to the work injury variable in the causal Bayesian network, calculate the path probability of each path, and select the path with the highest path probability as the final causal path explanation.

[0092] Specifically, firstly, based on the causal Bayesian network constructed in S1, the prior probability of the work-related injury variable is read from the conditional probability table. The work-related injury variable is denoted as... Its prior probability is The aforementioned prior probabilities are obtained from historical data statistics, representing the probability of the occurrence of work-related injury variables under conditions where no observational information is available.

[0093] Secondly, the likelihood value of the evidence is calculated based on the set of evidence variables generated in S2. Let the set of evidence variables be... Each This corresponds to a specific observation value at a node in the network. The likelihood of the evidence set when the work injury variable takes on a specific state can be expressed as:

[0094] ;

[0095] in, This represents the set of parent nodes corresponding to the node of the evidence variable. The conditional probability is determined by the network's conditional probability table. If a certain evidence node has multiple parent nodes, its conditional probability is calculated by looking up the table according to the combination of parent node values.

[0096] Subsequently, the marginal probability of the evidence set is calculated. Marginal probability represents the probability of an evidence variable appearing under all possible combinations of upper-level variables, and can be expressed as formula (4). After obtaining the prior probability, likelihood value, and marginal probability, the posterior probability of the work-related injury variable under the given evidence set is calculated based on Bayes' theorem. The posterior probability is expressed as formula (3). The above calculation is performed in the model inference module to obtain the final probability of the work-related injury variable under the evidence constraints. If the following conditions are met... If so, a positive judgment is made on the work-related injury variable. Threshold The values ​​are configured by the system and set according to applicable laws, industry standards, or management requirements.

[0097] The specific process of outputting the maximum probability path from the root node to the work injury variable in S3 as a causal path explanation includes the following:

[0098] In a causal Bayesian network, there are multiple directed paths from the root node to the work injury variable. To determine the path that best represents the evidence propagation process, the path probability of each path is calculated. Let a path contain a set of edges. Each edge From node Pointing to node The path probability is defined as the product of all conditional probabilities along the path, expressed as formula (5).

[0099] The path probabilities mentioned above are obtained by sequentially searching the conditional probability table. The network traversal method uses depth-first search. The depth-first search algorithm starts from the root node and recursively expands all possible paths until it reaches the node containing the work injury variable, recording the edge sequence of each path during the expansion process.

[0100] After traversing all paths, the path probabilities of each path are compared, and the path with the highest probability value is selected as the causal path explanation. The paths are output as a sequence of nodes, used to demonstrate the causal propagation relationship between accident-related variables. The causal path output and the posterior probability are independent results; the former describes the variable dependency structure, while the latter characterizes the probabilistic inference result.

[0101] By calculating the posterior probability and determining the maximum path probability as described above, both the work injury determination result and the causal link that can be used for explanation are obtained simultaneously within the structured causal model. This result is directly used in subsequent intervention simulation steps.

[0102] Furthermore, the specific process of integrating real-time data and employing intervention simulations in S4 includes:

[0103] Personal health records, industry statistics, and seasonal meteorological data are input into a causal Bayesian network as real-time data. Based on do-calculus operations, the intervention variable is set as "implement protective measures" or "do not implement protective measures". The probability of intervention for the injury consequences in the two situations is calculated and the difference between the two intervention probabilities is compared.

[0104] The specific process of generating personalized work injury prevention reminders based on the probability differences before and after intervention in S4 includes:

[0105] When the probability of intervention for the consequences of injury after taking protective measures is significantly lower than the probability of not taking them, a text reminder containing specific protective measures will be automatically generated based on the individual's age, job position, real-time health indicators, and seasonal factors.

[0106] Specifically, firstly, real-time data is input into a causal Bayesian network. Real-time data sources include personal health records, industry accident statistics, and seasonal meteorological data. Personal health records include indicators such as body temperature, heart rate, blood pressure, and past medical records; industry statistics include the frequency and risk level of recent accidents in similar positions; and seasonal meteorological data includes environmental parameters such as temperature, humidity, and wind speed. The above data are matched to network nodes according to their definitions, corresponding to health indicator nodes, job risk nodes, and environmental factor nodes in the network. Variables not included in the real-time data maintain their default distribution and do not affect the execution of the intervention calculation steps.

[0107] Secondly, define the values ​​for the intervention variables. The intervention variables are denoted as... This indicates whether the corresponding protective measures have been implemented. This indicates that protective measures are being implemented, 0 indicates that no protective measures are implemented. The variable representing the consequences of injury is denoted as... For example, categories such as minor injury and serious injury. Based on the causal Bayesian network structure, and following the do-calculus computational rules, intervention will be performed in conjunction with... Node-related incoming edges are removed to ensure that intervention variables are not affected by their parent nodes during inference. Based on this, the probabilities of harmful consequences are calculated under both intervention conditions.

[0108] When the intervention value is In this case, the probability of intervention for the harmful consequences is expressed as:

[0109] ;

[0110] in, For combinations of upper-level variables that are unrelated to the intervention variable, The value is obtained from the conditional probability table under the given intervention conditions. This represents the joint probability of these higher-level variables. Similarly, when the intervention value is... When the probability of intervention for the harmful consequences is 0, it is expressed as:

[0111] ;

[0112] Both of the above equations are implemented through the network inference module, which unfolds the relevant conditional probabilities sequentially according to the dependencies of network nodes and sums them.

[0113] Subsequently, the probabilities of harmful consequences under the two intervention scenarios were compared. The differences were denoted as...

[0114] ;

[0115] When the difference value When the value exceeds the preset threshold, it indicates that implementing protective measures can significantly reduce the probability of injury.

[0116] The specific process of generating personalized work injury prevention reminders based on the probability differences before and after intervention in S4 includes the following:

[0117] After obtaining the probability differences, the system generates targeted reminders based on the variable information of the current individual. The variable information includes age, job position, health indicators, and seasonal factors. Age is used to determine the range of intensity for protective recommendations. Job position corresponds to the job risk node in the network and is used to select the type of job-related protective measures. Health indicators are used to determine whether additional safety checks or restrictions on work intensity are needed. Seasonal factors determine environment-related protective measures based on the values ​​of meteorological variables, such as adding anti-slip measures reminders in low-temperature conditions.

[0118] When generating the reminder text, the system selects the corresponding measure description based on a template matching method. For example, when the difference value... When the system indicates that protective measures have a significant impact on reducing the consequences of falls in high-altitude work scenarios, it includes safety recommendations corresponding to the high-altitude work in the alert. The alert content is output in a structured format, including specific measures to be taken, such as wearing a safety belt, checking the fall arrestor rope, checking health status, or adding anti-slip devices according to weather conditions. The alert is generated based on real-time data, and each inference is executed independently, without relying on historical alerts.

[0119] By combining intervention probability calculations with variable information, preventative measures relevant to the individual's situation can be obtained based on network inference. The above steps are performed sequentially during execution S4, outputting the final workplace injury prevention reminder information.

[0120] Furthermore, the specific process of incremental updates using the Bayesian parameter update mechanism in S5 includes:

[0121] A pseudo-count is introduced for each node as a Dirichlet prior. When new data is introduced, only the nodes related to the new data and their conditional probability tables are locally updated, while the structure and parameters of other parts of the network remain unchanged.

[0122] Specifically, firstly, after the causal Bayesian network is constructed, a pseudo-count parameter is set for each node in the network. Let the set of nodes be... For any of these nodes Its conditional probability table depends on the set of parent nodes. The possible combinations of values. For each parent node, the combination... Set up a pseudo-count set corresponding to the node values:

[0123] ;

[0124] in, For nodes The parent node has a value of p and the node itself has a value of p. Dirichlet prior pseudo-counts for each category This represents the number of possible discrete values ​​for the node. The pseudo-count is set to provide a non-zero initial probability when statistical data is insufficient, ensuring that subsequent incremental update operations are stable and executable.

[0125] Subsequently, when the system receives new data, it extracts the fields from the accident records in the new data according to the node definitions, and parses each record to obtain the values ​​of the node and its parent node. The new dataset is denoted as... Each data entry It contains some or all of the node values ​​from the node set. For each data entry, it includes the values ​​of nodes that appear. and its parent node set When considering combinations of values, the corresponding count is calculated. The joint count of a node and its parent node in the new data is represented as follows: The count of parent node combinations in the new data is represented as: After obtaining the above counts, the nodes are... The conditional probability table is locally updated. The updated conditional probability is expressed as formula (6), where the numerator represents the sum of the prior pseudo-count and the new data count, and the denominator represents the sum of the pseudo-count and the sum of the counts for all values ​​under the parent node combination. The above update method conforms to the conjugate property of the Dirichlet-multinomial distribution, and can update the local probability without re-estimating the entire network.

[0126] During incremental updates, only the conditional probability representations related to nodes actually appearing in the new data and their corresponding parent node combinations are updated. Nodes not appearing in the new data retain their conditional probabilities, avoiding impact on network parts not involved in the data and ensuring the locality of the incremental update process. This approach avoids the computational cost of rebuilding the entire network and preserves the network structure. Middle-edge set Stability.

[0127] During the update process, if a node exhibits multiple values ​​in the new data, the conditional probabilities for each value are updated according to the formula described above. For nodes with multiple parent node combinations, the update is performed separately for each parent node combination p. The updated conditional probability table, along with the network structure, is used in the next inference step, enabling the network to continuously reflect newly emerging accident characteristics in the data.

[0128] Furthermore, after all updates are completed, the work-related injury determination results and their causal paths identified in the new data are recorded as new learnable data and processed in the same manner as the incremental update process described above. By continuously inputting new accident data and performing local updates, the network parameters will gradually converge to a state consistent with the current data environment, ensuring that the causal Bayesian network is always based on the latest data for work-related injury determination and prevention analysis.

[0129] The above incremental update process has been fully disclosed and can be executed independently during the parameter learning process of causal Bayesian networks, without relying on the overall training process, thus supporting the gradual updating of parameters as data accumulates over time.

[0130] Furthermore, the accident-related variables in the node set include six variables: accident triggering factors, job risks, personal health indicators, regulatory requirements, environmental factors, and injury consequences.

[0131] The causal Bayesian network forms a closed loop between work injury identification and work injury prevention. The causal path explanation and identification results generated during the work injury identification stage are fed back to S5 as new data for subsequent Bayesian parameter update mechanisms, thereby achieving continuous adaptive optimization of the network.

[0132] Specifically, in terms of node design, the set of nodes in the causal Bayesian network is divided into six categories of accident-related variables. Accident triggering factor variables represent the work behaviors or operating methods that directly lead to the accident, denoted as the node set. Typical values ​​include working at height, hoisting operations, lifting operations, and confined space operations. Job-specific risk variables are used to represent the inherent risk level or type of different jobs, denoted as a set of nodes. The values ​​include construction positions, equipment maintenance positions, and warehousing and handling positions, and can be further subdivided into risk level codes. Personal health indicator variables are used to represent the individual's health status and are denoted as a set of nodes. Its values ​​correspond to indicators such as heart rate zones, past cardiovascular history, and heatstroke tendency. Regulatory requirements define variables used to represent the implementation status of applicable legal provisions, industry standards, and internal management systems, denoted as a set of nodes. The values ​​include mandatory protective equipment requirements and medical examination requirements for special occupations. Environmental factor variables are used to represent meteorological conditions and work environment status, and are denoted as a set of nodes. The values ​​include ground slipperiness, temperature range, wind speed, and lighting conditions. The injury consequence variable represents the outcome of the accident and is denoted as a set of nodes. The possible values ​​include categories such as no damage, minor injury, and serious injury.

[0133] In the network instance, six types of variable nodes are connected by a set of causal edges. Dependencies are established; for example, accident triggering factor nodes and environmental factor nodes point to injury consequence nodes, and job risk nodes and regulatory requirement nodes both point to accident triggering factor nodes or protective measure-related nodes. Each node is coded based on historical data and business rules; for example, high-altitude operations are coded as 1, hoisting operations as 2, and ambient temperature is divided into discrete intervals and numbered accordingly. Through this method, various accident-related variables in the node set form a structured, directed acyclic causal network.

[0134] In terms of closed-loop construction, the causal path explanations and determination results generated during the work-related injury determination stage are fed back as new data to the incremental update step. For each accident that has been determined, its set of evidence variables is recorded. Calculated posterior probabilities of work-related injury variables Threshold comparison results and corresponding causal paths. Causal paths are represented as a sequence of nodes, denoted as...

[0135] ;

[0136] The above sequence starts from the root node, passes through intermediate risk nodes sequentially, and finally reaches the injury consequence node or work injury variable node. The system will... It is stored as a new structured record in the incremental dataset.

[0137] When performing Bayesian parameter updates in S5, the aforementioned feedback data, along with new incident data from other data sources, are treated as a unified new dataset D'. For any node When a node appears in the feedback record and its value is definite, it is considered a new observation sample and included in the statistics. If the node's value originates from an intermediate variable in the causal path, it also participates in the counting of the corresponding conditional probability table row. In this way, the output results of the work injury determination stage affect the network parameters in reverse, making the network more consistent with the characteristics of the actual accident chain during subsequent inference.

[0138] Furthermore, during the closed-loop process, no update operation is performed on nodes not covered by the path, ensuring the locality of updates. The network structure remains largely unchanged, with only parameter adjustments reflecting new risk patterns and causal relationship strengths. From an overall process perspective, S1 constructs an initial causal Bayesian network; S2 and S3 use this network to complete work injury identification and causal path generation; S4 performs intervention simulation and prevention alerts based on the current network; and S5 uses the identification results and path information to incrementally update the network parameters. Through multiple iterations, the accident-related variables and their conditional probability tables in the node set are continuously adjusted based on the latest accident samples and inference results, achieving closed-loop processing and adaptive optimization between work injury identification and prevention.

[0139] Finally, it should be noted that the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A work-related injury determination method based on artificial intelligence, characterized in that, Includes the following steps: S1. Construct a causal Bayesian network from historical work injury data. The causal Bayesian network is a directed acyclic graph, including a set of nodes and a set of edges. Each node in the set of nodes represents an accident-related variable, and the set of edges represents the causal dependency between accident-related variables. S2. Receive the accident text description as input, and use rule-based parsing to map the keywords in the accident text description to the values ​​of corresponding accident-related variables to form a set of evidence variables; S3. Based on the causal Bayesian network, the posterior probability of the work injury variable is calculated using the Bayesian update mechanism. When the posterior probability exceeds a preset threshold, it is identified as a work injury, and the maximum probability path from the root node to the work injury variable is output as the causal path explanation. S4. Integrate real-time data and use intervention simulation to calculate the probability of injury consequences under set intervention variables, and generate personalized work injury prevention reminders based on the probability difference before and after intervention. S5. When new data is received, the Bayesian parameter update mechanism is used to incrementally update only the parameters of the affected nodes in the causal Bayesian network.

2. The work-related injury determination method based on artificial intelligence according to claim 1, characterized in that, The specific process of constructing a causal Bayesian network from historical work injury data in S1 includes: The fractional search algorithm is used to search for the optimal network structure by maximizing the network's fractional function, and the conditional probability table for each node is calculated using maximum likelihood estimation. When parsing accident text descriptions, semantically similar words are mapped to the same accident-related variable nodes to achieve unified processing of semantic variations.

3. The work-related injury determination method based on artificial intelligence according to claim 2, characterized in that, The specific process of forming a set of evidence variables through rule-based parsing in S2 includes: Beforehand, mapping rules are established between keywords and accident-related variables. "Slipping," "falling," and "dropping" are uniformly mapped to environmental factor nodes, "working at height" and "lifting operations" are mapped to accident triggering factor nodes, and "myocardial infarction" and "heatstroke" are mapped to personal health indicator nodes, thereby obtaining a structured set of evidence variables.

4. The work-related injury determination method based on artificial intelligence according to claim 3, characterized in that, The specific process of calculating the posterior probability of the work-related injury variable using the Bayesian update mechanism in S3 includes: Based on the conditional probability table of causal Bayesian networks, the prior probability of the work-related injury variable, the likelihood of the evidence variable set, and the marginal probability of the evidence variable set are calculated sequentially. The posterior probability of the work injury variable is obtained through the posterior probability calculation formula. The preset threshold can be adjusted according to regulatory requirements.

5. The work-related injury determination method based on artificial intelligence according to claim 4, characterized in that, The specific process of outputting the maximum probability path from the root node to the work injury variable in S3 as the causal path explanation includes: A depth-first search algorithm is used to traverse all paths leading to the work injury variable in the causal Bayesian network, calculate the path probability of each path, and select the path with the highest path probability as the final causal path explanation.

6. The work-related injury determination method based on artificial intelligence according to claim 5, characterized in that, The specific process of integrating real-time data and employing intervention simulation in S4 includes: Personal health records, industry statistics, and seasonal meteorological data are input into a causal Bayesian network as real-time data. Based on do-calculus operations, the intervention variable is set as "implement protective measures" or "do not implement protective measures". The probability of intervention for the injury consequences in the two situations is calculated and the difference between the two intervention probabilities is compared.

7. The work-related injury determination method based on artificial intelligence according to claim 6, characterized in that, The specific process of generating personalized work injury prevention reminders based on the probability differences before and after the intervention in S4 includes: When the probability of intervention for the consequences of injury after taking protective measures is significantly lower than the probability of not taking them, a text reminder containing specific protective measures will be automatically generated based on the individual's age, job position, real-time health indicators, and seasonal factors.

8. The work-related injury determination method based on artificial intelligence according to claim 7, characterized in that, The specific process of incremental update using the Bayesian parameter update mechanism in S5 includes: A pseudo-count is introduced for each node as a Dirichlet prior. When new data is introduced, only the nodes related to the new data and their conditional probability tables are locally updated, while the structure and parameters of other parts of the network remain unchanged.

9. The work-related injury determination method based on artificial intelligence according to claim 8, characterized in that, The accident-related variables in the node set include six variables: accident triggering factors, job risks, personal health indicators, regulatory requirements, environmental factors, and injury consequences.

10. The work-related injury determination method based on artificial intelligence according to claim 9, characterized in that, The causal Bayesian network forms a closed loop between work injury identification and work injury prevention. The causal path explanation and identification results generated during the work injury identification stage are fed back to S5 as new data for the subsequent Bayesian parameter update mechanism, thereby achieving continuous adaptive optimization of the network.

Images