A labor risk assessment system based on deep learning judgment

Abstract

The application discloses a labor risk assessment system based on deep learning judgment, belongs to the technical field of artificial intelligence and labor risk assessment, constructs an end-to-end technical architecture of "multi-modal feature fusion-multi-task deep learning-double path fusion analysis", maps multi-source labor data to the same vector space through multi-modal feature fusion, realizes dynamic evolution of multi-dimensional risk assessment standards by combining multi-task machine learning and reinforcement update, constructs a pre-deployment and in-service double-path judgment mechanism, fuses the evaluation results of three dimensions of the deployment unit risk, the deployer personal risk and the post risk, outputs recruitment decision suggestions and company overall labor operation risk rating, not only realizes a fundamental leap of labor risk assessment from static rule statistics to dynamic intelligent decision, but also realizes a closed loop coverage of labor risk from "pre-entry" to "in-process monitoring", and provides accurate and differentiated decision support for enterprises at different stages.

Description

Technical Field

[0001] This invention relates to the fields of artificial intelligence and employment risk assessment technology, and more specifically, to an employment risk assessment system based on deep learning judgment. Background Technology

[0002] With the rapid development of new employment models such as flexible employment and platform economy, enterprises are facing increasingly complex employment environments, and the management of employment risks (such as work-related accidents, labor disputes, and employee turnover) is becoming significantly more difficult. Traditional employment risk assessment methods rely heavily on human experience or simple statistical indicator thresholds, which are difficult to cope with the dynamically changing employment environment and complex risk relationships.

[0003] In existing technologies, some employment risk monitoring solutions based on big data have emerged. For example, Chinese patent application CN117035252A discloses a user risk monitoring method and system based on a flexible employment platform. This method collects multi-dimensional information from enterprise users, transforms it into structured events, and matches it with a preset metric table to derive a risk level through weighted calculation. This method automates risk monitoring to some extent, but it is essentially still a method based on expert rules and statistical weighting, and its limitations are:

[0004] First, risk rating relies on manually preset metrics and weights, and cannot automatically learn and adapt to new risk patterns. When the employment scenario or risk characteristics change, manual intervention is required to update the rules, resulting in poor flexibility.

[0005] Secondly, this method mainly processes structured "event" information and has difficulty effectively utilizing deep risk signals in unstructured data (such as safety incident report texts, employee interview videos, and the network of connections between dispatching units).

[0006] Finally, this method fails to distinguish between the two different scenarios of recruitment and on-the-job management, and it is also unable to conduct a detailed evaluation of the three dimensions of dispatching unit, dispatcher, and position.

[0007] Therefore, in response to actual technical defects, a deep learning-based employment risk assessment system is proposed. Summary of the Invention

[0008] The purpose of this invention is to address practical technical deficiencies by providing an employment risk assessment system based on deep learning.

[0009] The objective of this invention can be achieved through the following technical solution: an employment risk assessment system based on deep learning, comprising:

[0010] The employment data acquisition module is configured to acquire multi-source employment data under the target employment scenario, including structured business data of dispatching units, multimodal data of dispatchers consisting of recruitment data of pre-dispatch workers and work data of dispatched workers, and employment dispatch data consisting of pre-dispatch data and real-time dispatch data.

[0011] The feature transformation module is configured to use a deep learning model to transform multi-source employment data into the same vector space, generating a unified multimodal feature vector composed of feature sub-vectors of dispatching units, feature sub-vectors of dispatchers, and feature sub-vectors of employment dispatch nature.

[0012] The multi-dimensional risk assessment module is configured to receive multimodal feature vectors and perform multi-dimensional employment risk assessment tasks using multiple risk assessment models, specifically:

[0013] The risk assessment task of dispatching units is carried out using the dispatching unit risk assessment model, and the values ​​of multiple predicted risk indicators and comprehensive risk rating of dispatching units are output.

[0014] The dispatcher risk assessment model is used to perform a dispatcher risk profiling assessment task, and outputs the job risk label of the pre-dispatcher and the on-the-job behavior risk score of the current dispatcher.

[0015] The risk assessment model for the nature of labor dispatch is used to perform the assessment of the nature of labor dispatch, and the risk level of the proposed dispatch position and the dispatch risk score of the actual position are output.

[0016] The dual-path scenario fusion module is configured to perform dual-path risk fusion analysis of pre-dispatch and on-the-job scenarios, generate a pre-dispatch risk report and output recruitment decision suggestions, and generate an on-the-job employment risk report and output the company's overall employment operation risk rating.

[0017] Furthermore, the structured business data includes personnel size data, job structure data, historical accident data, and claims amount data; the pre-dispatch worker recruitment data includes resume text data and interview audio and video data; the on-the-job dispatch worker work data includes attendance time sequence data and safety accident report text data; the pre-dispatch nature data includes job description text data; and the real-time dispatch nature data includes dynamic information of actual job positions and real-time monitoring data of the work environment.

[0018] Furthermore, the process for the dispatching unit risk assessment model to perform the dispatching unit risk assessment task includes:

[0019] The risk assessment model for dispatching units consists of a first fully connected network and a first output layer. The first fully connected network is used to receive feature sub-vectors of dispatching units.

[0020] The first output layer includes a first classification output unit and a first regression output unit. The first regression output unit uses a linear activation function and outputs multiple predicted risk index values, including at least one of the following: accident rate per 10,000 people, compensation rate, traffic accident rate, disability rate, accident rate of over-age personnel, and the proportion of 4 / 5 types of work. The first classification output unit uses a Softmax activation function and outputs the comprehensive risk rating of the dispatching unit, including low risk, medium risk, high risk, and extremely high risk.

[0021] Furthermore, the process by which the dispatcher risk assessment model performs the dispatcher risk profiling assessment task includes:

[0022] The dispatcher risk assessment model consists of a second fully connected network and a second output layer, wherein the second fully connected network is used to receive the dispatcher feature subvector.

[0023] The second output layer includes a second classification output unit and a second regression output unit;

[0024] For pre-dispatch personnel, the second category output unit outputs the job risk label of the pre-dispatch personnel. The job risk label includes resume risk label and interview risk label. The resume risk label and interview risk label are each independently classified as low risk, medium risk and high risk.

[0025] For on-the-job temporary workers, the second regression output unit outputs an on-the-job behavior risk score.

[0026] Furthermore, the process of the employment dispatch nature risk assessment model to perform the employment dispatch nature assessment task includes: the employment dispatch nature risk assessment model consists of a third fully connected network and a third output layer, the third fully connected network is used to receive employment dispatch nature feature sub-vectors, and the third output layer further includes a pre-dispatch assessment sub-branch and a real-time dispatch assessment sub-branch.

[0027] The pre-dispatch assessment sub-branch outputs the predicted risk level of the proposed dispatch position during the recruitment stage, which is one of low risk, medium risk, high risk, or extremely high risk.

[0028] The real-time dispatch assessment sub-branch outputs a dispatch risk score for the actual position performed during the employment phase.

[0029] Furthermore, for pre-dispatch scenarios, a pre-dispatch risk report is generated by integrating the dispatching unit's comprehensive risk rating, various predicted risk indicator values, the pre-dispatch worker's job risk label, and the job risk level of the proposed dispatch position.

[0030] Furthermore, based on the pre-dispatch risk report, recruitment decision recommendations are output and configured as follows:

[0031] When the overall risk rating of the dispatching unit is low or medium risk, all predicted risk indicators do not exceed the corresponding high risk threshold, and the resume risk label and interview risk label of the prospective dispatcher are both low or medium risk, and the job risk level of the position to be dispatched is low or medium risk, an admission recommendation is generated.

[0032] When the overall risk rating of the dispatching unit is high risk or extremely high risk, or the risk label of the resume or interview of the prospective dispatcher is high risk, or the risk level of the position to be dispatched is extremely high risk, a rejection suggestion is generated; otherwise, a suggestion requiring manual review is generated.

[0033] Furthermore, for the on-the-job scenario, the on-the-job behavior risk score of the dispatched worker and the dispatch risk score of the actual position are integrated to generate an on-the-job employment risk report.

[0034] Furthermore, based on the on-the-job employment risk report, the overall employment operation risk rating is output and configured as follows:

[0035] The statistical characteristics of the on-the-job behavior risk scores of all dispatched workers are obtained as the first risk factor. The statistical characteristics include at least one of the average score, the lowest score, and the proportion of high-risk personnel. The statistical characteristics of the dispatch risk scores of all actual positions are obtained as the second risk factor. The statistical characteristics include at least one of the average score, the highest score, and the proportion of high-risk positions. The first and second risk factors are input into the preset fusion rules to output the overall employment operation risk rating.

[0036] Compared with the prior art, the advantages of this invention are:

[0037] 1. This invention uses a deep learning model to transform the structured business data of dispatching units, multimodal data of dispatchers, and data on the nature of labor dispatch into the same vector space, generating a unified multimodal feature vector. This achieves deep fusion of cross-modal information. At the same time, it combines multi-task machine learning and reinforcement updates to achieve end-to-end multi-dimensional dynamic evolutionary prediction and assessment. On the one hand, it improves the accuracy and comprehensiveness of risk assessment. On the other hand, it enables risk assessment standards to automatically evolve with changes in the business environment, effectively overcoming the shortcomings of relying on static rule bases and being unable to adaptively update, and significantly improving the accuracy and timeliness of risk assessment.

[0038] 2. This invention also divides employment risk assessment into two independent paths: pre-dispatch (recruitment access) and on-the-job (employment monitoring). It integrates the assessment results of three dimensions: risk of dispatching unit, personal risk of dispatcher, and risk of position, to form targeted decision output. Based on the dual-path scenario judgment, it realizes the closed-loop coverage of employment risk from "pre-access" to "in-process monitoring", providing enterprises with accurate and differentiated decision support at different stages. Attached Figure Description

[0039] Figure 1 This is a block diagram illustrating the system module principle of the present invention;

[0040] Figure 2 This is a schematic diagram of the dual-path fusion analysis process of the present invention. Detailed Implementation

[0041] The technical solutions of 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. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.

[0042] Example 1: This invention discloses an employment risk assessment system based on deep learning. Please refer to [link / reference]. Figures 1-2 It includes a labor data acquisition module, a feature conversion module, a multi-dimensional risk assessment module, and a dual-path scenario fusion module;

[0043] The employment data acquisition module is configured to acquire multi-source employment data under the target employment scenario, including structured business data of the dispatching unit;

[0044] Multimodal data on dispatchers; including recruitment data for prospective dispatchers and work data for currently employed dispatchers;

[0045] The data on the nature of employment dispatch includes pre-dispatch data, which describes the nature of the positions to be dispatched during the recruitment stage, and real-time dispatch data, which describes the nature of the positions actually performed during the employment stage.

[0046] More specifically, structured business data includes personnel size data, job structure data, historical accident data, and claims amount data; pre-dispatch worker recruitment data includes resume text data and interview audio and video data; on-the-job dispatch worker work data includes attendance time series data and safety accident report text data; pre-dispatch nature data includes job description text data; and real-time dispatch nature data includes at least dynamic information of the actual job position and real-time monitoring data of the work environment.

[0047] This module achieves comprehensive collection of employment data by connecting to multi-source heterogeneous data interfaces, providing a rich data foundation for subsequent multi-dimensional risk assessment and solving the problems of single data sources and insufficient information in existing technologies.

[0048] The feature transformation module is configured to use a deep learning model to transform multi-source employment data into the same vector space, generating a unified multimodal feature vector;

[0049] Specifically, this module uses the following model to process various types of data:

[0050] Pre-trained Language Model (BERT): Processes text data such as resume text, safety incident reports, and job descriptions, outputting a 768-dimensional text semantic vector; Multimodal Fusion Network: Processes interview audio and video data, extracts visual and audio features and fuses them, outputting an interview feature vector; Recurrent Neural Network (LSTM) or Transformer: Processes attendance time-series data and environmental monitoring data, extracting time-series change features; Embedding Layer: Processes categorical data such as job classification and position category, outputting a category embedding vector;

[0051] The aforementioned feature vectors are concatenated after dimensional alignment to form a unified multimodal feature vector. This vector contains three segments: the dispatching unit feature sub-vector, the dispatcher feature sub-vector, and the employment dispatch nature feature sub-vector. This module maps multiple types of data, such as text, audio and video, and time series, to the same semantic space, enabling subsequent models to learn cross-modal correlation features simultaneously. This overcomes the shortcomings of data fragmentation and inability to conduct joint analysis in existing technologies, and significantly improves the utilization efficiency of risk information.

[0052] The multi-dimensional risk assessment module is configured to receive multimodal feature vectors and perform multi-dimensional employment risk assessment tasks using multiple risk assessment models, specifically:

[0053] The risk assessment task of dispatching units is carried out using the dispatching unit risk assessment model, and the values ​​of multiple predicted risk indicators and comprehensive risk rating of dispatching units are output.

[0054] The dispatcher risk assessment model is used to perform a dispatcher risk profiling assessment task, and outputs the job risk label of the pre-dispatcher and the on-the-job behavior risk score of the current dispatcher.

[0055] The risk assessment model for the nature of labor dispatch is used to perform the assessment of the nature of labor dispatch, and the risk level of the proposed dispatch position and the dispatch risk score of the actual position are output.

[0056] The process for the dispatching unit risk assessment model to perform the dispatching unit risk assessment task includes:

[0057] The risk assessment model for dispatching units consists of a first fully connected network and a first output layer. The first fully connected network is used to receive feature sub-vectors of dispatching units.

[0058] The first output layer includes a first classification output unit and a first regression output unit. The first regression output unit uses a linear activation function and outputs multiple predicted risk indicators, including the accident rate per 10,000 people, the compensation rate, the proportion of traffic accidents, the disability rate, the accident rate of over-age personnel, and the proportion of occupations in categories 4 and 5. The first classification output unit uses a Softmax activation function and outputs the comprehensive risk rating of the dispatching unit, including low risk, medium risk, high risk, and extremely high risk.

[0059] The update mechanism of the risk assessment model for dispatching units is as follows: obtain the actual risk index values ​​of the dispatching units after the prediction period as historical index feedback data, and use the difference between the historical index feedback data and the real-time output predicted risk index values ​​as a reward signal to backpropagate and update the weight matrix and bias vector of the first fully connected network.

[0060] The process of the dispatcher risk assessment model to perform dispatcher risk profiling assessment includes:

[0061] The dispatcher risk assessment model consists of a second fully connected network and a second output layer. The second fully connected network is used to receive the dispatcher feature subvector.

[0062] The second output layer includes a second classification output unit and a second regression output unit;

[0063] For pre-dispatch personnel, the second category output unit outputs the job risk label of the pre-dispatch personnel. The job risk label includes resume risk label and interview risk label. The resume risk label and interview risk label are each independently classified as low risk, medium risk and high risk.

[0064] For employed temporary workers, the second regression output unit outputs the on-the-job behavior risk score of the employed temporary workers, which is a continuous value between 0 and 100.

[0065] The update mechanism of the dispatcher risk assessment model is as follows: obtain the actual accident situation, actual disciplinary situation, and actual resignation situation of the dispatched workers after the prediction period as behavioral feedback data;

[0066] According to the preset mapping rules, the behavioral feedback data is converted into actual behavioral risk scores, actual resume risk tags, and actual interview risk tags.

[0067] Based on the reinforcement update algorithm, the difference between the actual behavior risk score and the on-the-job behavior risk score is used as the first reward signal, the difference between the actual resume risk label and the resume risk label is used as the second reward signal, and the difference between the actual interview risk label and the interview risk label is used as the third reward signal.

[0068] The weighted sum of the first, second, and third reward signals is used as the total reward signal, and the weight matrix and bias vector of the second fully connected network are updated through the backpropagation algorithm.

[0069] The preset mapping rules include: determining the actual behavioral risk score based on the actual severity of the incident, the severity of the violation, and the reason for leaving the company; and mapping the actual resume risk label and the actual interview risk label in reverse based on the overall performance during employment (actual severity of the incident, the severity of the violation, and the reason for leaving the company).

[0070] The process of performing the risk assessment task for the nature of labor dispatch includes:

[0071] The risk assessment model for the nature of labor dispatch consists of a third fully connected network and a third output layer. The third fully connected network is used to receive the feature vector of the nature of labor dispatch, and the third output layer further includes a pre-dispatch assessment sub-branch and a real-time dispatch assessment sub-branch.

[0072] The pre-dispatch assessment sub-branch is used to output the predicted risk level of the proposed dispatch positions during the recruitment stage. The risk level is one of low risk, medium risk, high risk, or extremely high risk.

[0073] The real-time dispatch assessment sub-branch is used to output a dispatch risk score for the actual position performed during the employment phase. The dispatch risk score is a continuous value between 0 and 100.

[0074] The update mechanism for the risk assessment model of the employment dispatch nature is as follows: obtain safety accident data and work injury data that occur in the actual work position as historical dispatch feedback data. Among them, the safety accident data includes accident type, accident severity and accident frequency, and the work injury data includes whether it is identified as a work injury and the level of work injury.

[0075] According to the preset mapping rules, the historical dispatch feedback data is converted into the actual job risk level and the actual dispatch risk score. Based on the reinforcement update algorithm, the difference between the predicted job risk level and the actual job risk level is used as the first reward signal, and the difference between the dynamic risk score and the actual risk score is used as the second reward signal. The weight matrix and bias vector of the employment dispatch nature risk assessment model for each feature sub-vector are updated by backpropagation.

[0076] The preset mapping rules include: determining the actual job risk level based on the severity of the accident and the level of work-related injury, and calculating the actual risk score based on the severity of the accident, the level of work-related injury, and the frequency of the accident.

[0077] This embodiment uses a deep learning model to transform the structured business data of the dispatching unit, the multimodal data of the dispatcher, and the data on the nature of the employment dispatch into the same vector space, generating a unified multimodal feature vector. This achieves deep fusion of cross-modal information. At the same time, it combines multi-task machine learning and reinforcement updates to achieve dynamic evolution of multi-dimensional risk assessment standards. This overcomes the shortcomings of existing technologies that rely on static rule bases and cannot be updated adaptively, and significantly improves the accuracy and timeliness of risk assessment.

[0078] Example 2: Please refer to Figure 2 The dual-path scenario fusion module is configured to perform dual-path risk fusion analysis of pre-dispatch scenarios and on-the-job scenarios. Based on the judgment results, it generates a pre-dispatch risk report and outputs recruitment decision suggestions, as well as an on-the-job employment risk report and outputs the company's overall employment operation risk rating.

[0079] In the pre-dispatch scenario, the dual-path scenario fusion module integrates the dispatching unit's comprehensive risk rating, various predicted risk indicator values, the pre-dispatch worker's job risk label, and the job risk level of the proposed dispatch position to generate a pre-dispatch risk report.

[0080] Based on the pre-dispatch risk report, recruitment decision recommendations are generated and configured as follows:

[0081] Admission criteria determination: When the overall risk rating of the dispatching unit is low or medium risk, and all predicted risk indicators do not exceed the corresponding high risk threshold, and the resume risk label and interview risk label of the prospective dispatcher are both low or medium risk, and the job risk level of the position to be dispatched is low or medium risk, an admission recommendation is generated.

[0082] Rejection criteria: When the overall risk rating of the dispatching unit is high risk or very high risk, or the risk label of the resume or interview of the prospective dispatcher is high risk, or the risk level of the position to be dispatched is very high risk, a rejection suggestion is generated.

[0083] Conversely, suggestions requiring manual review will be generated.

[0084] This module integrates the risks from all three parties and outputs specific recruitment decision-making suggestions, realizing the transformation from "data evaluation" to "business decision-making" and providing enterprises with action guidelines that can be directly adopted in recruitment.

[0085] In the context of employment, the dual-path scenario fusion module integrates the employment behavior risk score of the dispatched worker and the dispatch risk score of the actual position they are engaged in, and generates an employment risk report.

[0086] Based on the current employment risk report, the company's overall employment operation risk rating is output and configured as follows:

[0087] The statistical characteristics of the on-the-job behavior risk scores of all on-the-job dispatched workers are obtained as the first risk factor. The statistical characteristics include at least one of the following: average score, lowest score, and percentage of high-risk personnel.

[0088] The statistical characteristics of the dispatch risk scores of all actual positions are obtained as the second risk factor. The statistical characteristics include at least one of the following: average score, highest score, and percentage of high-risk positions.

[0089] The first risk factor and the second risk factor are input into the preset fusion rules to output the overall employment operation risk rating, including low risk, medium risk, high risk and extremely high risk of on-the-job employment.

[0090] The preset integration rules include: if the proportion of high-risk personnel in the first risk factor exceeds the first threshold, or the proportion of high-risk positions in the second risk factor exceeds the second threshold, then the overall employment operation risk rating is extremely high.

[0091] If the proportion of high-risk personnel in the first risk factor exceeds the third threshold, or the proportion of high-risk positions in the second risk factor exceeds the fourth threshold, the overall employment operation risk rating is high risk.

[0092] If the average score of the first risk factor is lower than the fifth threshold, or the average score of the second risk factor is lower than the sixth threshold, the overall employment operation risk rating is medium risk.

[0093] Otherwise, the overall employment operation risk rating is low.

[0094] This module focuses on the actual performance of employees and the real-time risks of their positions, outputting a rating of the company's overall employment safety status, providing a quantitative basis for enterprises to monitor risks, trigger early warnings, and make corrective decisions.

[0095] It should be added that the article involves comparisons of multiple thresholds. Thresholds, preset values, preset ranges, etc., are set for result comparison and analysis to determine good or bad. The magnitude of these thresholds is determined by a combination of large-scale model analysis of sample data and human experience, and can also be appropriately adjusted based on seasonal or common-sense influencing factors.

[0096] In summary, multimodal feature fusion maps heterogeneous employment data from multiple sources to the same vector space. Combined with multi-task machine learning and reinforcement updates, it achieves dynamic evolution of multi-dimensional risk assessment standards and constructs a dual-path judgment mechanism for pre-dispatch and on-the-job deployment. It integrates assessment results from three dimensions: dispatching unit risk, dispatcher personal risk, and job risk, outputting recruitment decision suggestions and an overall company employment operation risk rating. This not only achieves a fundamental leap in employment risk assessment from static rule statistics to dynamic intelligent decision-making but also realizes closed-loop coverage of employment risk from "pre-entry approval" to "in-process monitoring," providing enterprises with accurate and differentiated decision support at different stages.

[0097] The above description is merely a preferred embodiment of the present invention; however, the scope of protection of the present invention is not limited thereto; any equivalent substitutions or modifications made by those skilled in the art within the technical scope disclosed in the present invention, based on the technical solution and its improved concept, should be covered within the scope of protection of the present invention.

Claims

1. A worker risk assessment system based on deep learning judgment, characterized by: include: The employment data acquisition module is configured to acquire multi-source employment data under the target employment scenario, including structured business data of dispatching units, multimodal data of dispatchers consisting of recruitment data of pre-dispatch workers and work data of dispatched workers, and employment dispatch data consisting of pre-dispatch data and real-time dispatch data. The feature transformation module is configured to use a deep learning model to transform multi-source employment data into the same vector space, generating a unified multimodal feature vector composed of feature sub-vectors of dispatching units, feature sub-vectors of dispatchers, and feature sub-vectors of employment dispatch nature. The multi-dimensional risk assessment module is configured to receive multimodal feature vectors and perform multi-dimensional employment risk assessment tasks using multiple risk assessment models, specifically: The risk assessment task of dispatching units is carried out using the dispatching unit risk assessment model, and the values ​​of multiple predicted risk indicators and comprehensive risk rating of dispatching units are output. The dispatcher risk assessment model is used to perform a dispatcher risk profiling assessment task, and outputs the job risk label of the pre-dispatcher and the on-the-job behavior risk score of the current dispatcher. The risk assessment model for the nature of labor dispatch is used to perform the assessment task of the nature of labor dispatch, and the risk level of the proposed dispatch position and the dispatch risk score of the actual position are output. The dual-path scenario fusion module is configured to perform dual-path risk fusion analysis of pre-dispatch and on-the-job scenarios, generate a pre-dispatch risk report and output recruitment decision suggestions, and generate an on-the-job employment risk report and output the company's overall employment operation risk rating.

2. The employment risk assessment system based on deep learning judgment according to claim 1, characterized in that: The structured business data includes personnel size data, job type structure data, historical accident data, and claims amount data. The pre-dispatch recruitment data includes resume text data and interview audio and video data. The on-the-job dispatch work data includes attendance time sequence data and safety accident report text data. The pre-dispatch nature data includes job description text data. The real-time dispatch nature data includes dynamic information of actual job positions and real-time monitoring data of the work environment.

3. The employment risk assessment system based on deep learning judgment according to claim 2, characterized in that: The process for executing the risk assessment task for dispatching units, as outlined in the dispatching unit risk assessment model, includes: The risk assessment model for dispatching units consists of a first fully connected network and a first output layer. The first fully connected network is used to receive feature sub-vectors of dispatching units. The first output layer includes a first classification output unit and a first regression output unit. The first regression output unit uses a linear activation function and outputs multiple predicted risk index values, including at least one of the following: accident rate per 10,000 people, compensation rate, traffic accident rate, disability rate, accident rate of over-age personnel, and the proportion of 4 / 5 types of work. The first classification output unit uses a Softmax activation function and outputs the comprehensive risk rating of the dispatching unit, including low risk, medium risk, high risk, and extremely high risk.

4. The employment risk assessment system based on deep learning judgment according to claim 3, characterized in that: The process of the dispatcher risk assessment model to perform dispatcher risk profiling assessment includes: The dispatcher risk assessment model consists of a second fully connected network and a second output layer, wherein the second fully connected network is used to receive the dispatcher feature subvector. The second output layer includes a second classification output unit and a second regression output unit; For pre-dispatch personnel, the second category output unit outputs the job risk label of the pre-dispatch personnel. The job risk label includes resume risk label and interview risk label. The resume risk label and interview risk label are each independently classified as low risk, medium risk and high risk. For on-the-job temporary workers, the second regression output unit outputs an on-the-job behavior risk score.

5. The employment risk assessment system based on deep learning judgment according to claim 4, characterized in that: The process of the risk assessment model for the nature of labor dispatch to perform the assessment task includes: the risk assessment model for the nature of labor dispatch consists of a third fully connected network and a third output layer. The third fully connected network is used to receive the feature sub-vector of the nature of labor dispatch. The third output layer further includes a pre-dispatch assessment sub-branch and a real-time dispatch assessment sub-branch. The pre-dispatch assessment sub-branch outputs the predicted risk level of the proposed dispatch position during the recruitment stage, which is one of low risk, medium risk, high risk, or extremely high risk. The real-time dispatch assessment sub-branch outputs a dispatch risk score for the actual position performed during the employment phase.

6. The employment risk assessment system based on deep learning judgment according to claim 5, characterized in that: For pre-dispatch scenarios, a pre-dispatch risk report is generated by integrating the dispatching unit's comprehensive risk rating, various predicted risk indicator values, the pre-dispatch worker's job risk label, and the job risk level of the proposed dispatch position.

7. The employment risk assessment system based on deep learning judgment according to claim 6, characterized in that: Based on the pre-dispatch risk report, recruitment decision recommendations are generated and configured as follows: When the overall risk rating of the dispatching unit is low or medium risk, all predicted risk indicators do not exceed the corresponding high risk threshold, and the resume risk label and interview risk label of the prospective dispatcher are both low or medium risk, and the job risk level of the position to be dispatched is low or medium risk, an admission recommendation is generated. When the overall risk rating of the dispatching unit is high risk or extremely high risk, or the risk label of the resume or interview of the prospective dispatcher is high risk, or the risk level of the position to be dispatched is extremely high risk, a rejection suggestion is generated; otherwise, a suggestion requiring manual review is generated.

8. The employment risk assessment system based on deep learning judgment according to claim 7, characterized in that: For employed scenarios, the system integrates the on-the-job behavior risk score of the dispatched worker with the dispatch risk score of the actual position they are engaged in to generate an on-the-job employment risk report.

9. The employment risk assessment system based on deep learning judgment according to claim 8, characterized in that: Based on the output of the on-the-job employment risk report, the overall employment operation risk rating is configured as follows: The statistical characteristics of the on-the-job behavior risk scores of all dispatched workers are obtained as the first risk factor. The statistical characteristics include at least one of the average score, the lowest score, and the proportion of high-risk personnel. The statistical characteristics of the dispatch risk scores of all actual positions are obtained as the second risk factor. The statistical characteristics include at least one of the average score, the highest score, and the proportion of high-risk positions. The first and second risk factors are input into the preset fusion rules to output the overall employment operation risk rating.

Images