Enterprise risk dynamic score assignment and grade intelligent determination method

Abstract

The invention discloses an enterprise risk dynamic score assignment and grade intelligent determination method, and the method comprises the steps: customizing a risk reference and a score assignment rule through a rule setting module, achieving the complete-period monitoring of risks through combining with a region inherent risk assessment module, a dynamic risk assessment module and an existing risk assessment module, and innovatively introducing a core algorithm module, the problem of evaluation deviation caused by fixed weight and linear accumulation in the prior art is solved; according to the core algorithm, the weight of risk factors is optimized in real time through an LSTM model, a differential nonlinear function is adopted to adapt to the influence rules of different risks, the coupling relation between factors is corrected, and the risk assessment accuracy is improved by 20% or above; the system automatically updates a risk value every 20 minutes, multi-region edge calculation deployment is supported, and the operation delay is less than or equal to 5 seconds; according to the method, the defects of high subjectivity and poor real-time performance of a traditional method are overcome, personalized risk management and control requirements of enterprises in different industries are met, and accurate risk data support can be provided for supervision departments.

Description

Technical Field

[0001] This invention belongs to the field of data processing systems or methods, and is a method for dynamic scoring and intelligent determination of enterprise risk levels. It is applicable to data processing for administrative, commercial, financial, management, supervision or prediction purposes, so as to achieve accurate assessment and management of enterprise risks. Background Art

[0002] Driven by both enterprise digital transformation and business expansion, risk factors are characterized by "multi-source, dynamic, and coupled" features. Traditional risk assessment methods and existing technological systems are no longer sufficient to meet the needs of precise and real-time risk management. Specific shortcomings are as follows: 1. The core deficiency of existing technologies: the limitations of static evaluation and linear calculation. Traditional risk assessment methods and most existing systems generally adopt a calculation logic of "fixed weights + linear accumulation," which cannot adapt to the dynamic changes and complex relationships of risk factors. Fixed weights lead to poor adaptability: Existing technologies often pre-determine the weights of risk factors (such as 30% for special operations and 20% for personnel numbers), without considering the differences in risk characteristics between different stages of enterprise development and different regions. For example, the impact of "special operation risk" in chemical enterprises is drastically different from that of "transaction anomaly alarm risk" in financial enterprises, but existing systems cannot dynamically adjust the weights, resulting in an assessment result that deviates from the actual risk by more than 30%.

[0003] Linear accumulation ignores the non-linear growth law of risk: The impact of risk factors such as overloading and the superposition of hidden dangers does not increase linearly. When the overloading rate exceeds 20%, the accident rate will increase exponentially. However, the existing technology still calculates based on "adding a fixed score for each person overloaded", which cannot reflect the rapid changes after the "risk threshold". This can easily lead to high risk being misjudged as medium risk.

[0004] The coupling effect of risk factors is not considered: In real-world scenarios, risk factors such as "hot work" and "combustible gas alarm," and "core system upgrade" and "system vulnerabilities" are strongly correlated, and the risk is significantly amplified when they are combined. However, existing technologies simply add up the scores of each factor individually, without correcting for duplicate scores (such as the superposition of weakly correlated factors) or reflecting the synergistic risks of strongly correlated factors, resulting in an assessment distortion rate of over 25%.

[0005] 2. Pain points in industry practice: Insufficient real-time performance and accuracy Lack of real-time monitoring capabilities: Traditional manual assessments take 1-7 days. While existing systems support automatic calculations, most update cycles exceed one hour, making them unable to handle acute risks such as sudden equipment failures or instantaneous personnel overload. For example, if a transformer temperature alarm in a power substation is not included in the risk assessment within 10 minutes, the window for fault handling may be missed, leading to equipment damage or power outages.

[0006] Data anomaly interference assessment results: Real-time collected risk data (such as the number of personnel and the number of alarms) is susceptible to anomalies due to sensor malfunctions, data transmission delays, etc. (e.g., a momentary display of "100 personnel" far exceeds the actual base of 30 personnel). Existing technologies lack effective anomaly identification and processing mechanisms, often leading to misjudgments of risk levels due to a single abnormal data point (e.g., low risk being misjudged as high risk), interfering with enterprise decision-making.

[0007] Poor adaptability to multi-regional deployment: Large enterprises (such as cross-regional oil refineries and chain banks) need to conduct distributed monitoring of multiple risk areas, but existing systems mostly rely on centralized cloud computing, with data transmission volume reaching 100MB / hour / region. In scenarios with limited network bandwidth (such as remote substations), the computing latency exceeds 30 seconds, which cannot meet the real-time management and control requirements.

[0008] 3. The urgent needs of regulation and industry development The need for more refined regulation is increasing: government regulatory departments (such as the Ministry of Emergency Management and the State Financial Supervision and Administration Bureau) need to implement differentiated regulation based on the real-time risk level of enterprises—increasing the frequency of inspections for high-risk enterprises and reducing intervention for low-risk enterprises. However, existing technologies, due to insufficient accuracy in assessment, cannot provide regulatory departments with a precise correspondence between "risk level and control measures," leading to a waste of regulatory resources or regulatory loopholes.

[0009] The differentiated needs of different industries have not been met: the risk characteristics of industries such as chemical, power and finance are significantly different. Chemical companies are concerned about the "coupling risks of special operations and hidden dangers", while financial companies are concerned about the "trading alarms and system operation risks". However, most existing systems are general-purpose designs and lack the ability to adapt algorithms to different industries. Companies need to invest more than 50% more in customization costs, and the adaptation effect is limited.

[0010] In summary, existing technologies lack core algorithmic support such as "dynamic weight optimization, nonlinear scoring, and risk correlation correction," which fails to address the issues of "poor assessment accuracy, insufficient real-time performance, and weak industry adaptability." Therefore, this application proposes a technical solution to address these technical problems. Summary of the Invention

[0011] The purpose of this invention is to provide a method for dynamic risk scoring and intelligent risk level determination for enterprises, thereby solving the aforementioned technical problems.

[0012] The technical solution adopted in this invention is as follows: A method for dynamic risk scoring and intelligent risk level determination for enterprises includes the following modules: Rule setting module: includes risk benchmark setting unit, special operation scoring rule unit, personnel quantity scoring rule unit, hidden danger scoring rule unit, and critical alarm quantity scoring rule unit; Regional Inherent Risk Assessment Module: Assigns inherent risk scores to each risk area; Regional dynamic risk assessment module: The system automatically calculates and generates dynamic risk values, and generates unmodifiable scoring records; The existing risk assessment module in the region automatically calculates the current risk value and generates an unmodifiable risk assessment record. Risk Dynamics in One Chart Module: Statistically displays the status of risk areas and influencing factors; It also includes a core algorithm module, which is used to optimize the accuracy of risk value calculation, including dynamic weight optimization algorithm, nonlinear scoring model, and risk correlation correction algorithm; The dynamic weight optimization algorithm is based on historical risk records and real-time risk change data. It uses a long short-term memory network (LSTM) model to predict the impact of each risk factor (special operation, number of personnel, hidden dangers, critical alarms) on the final risk and adjusts the weight of each factor in real time. The nonlinear scoring model employs differentiated nonlinear functions based on the characteristics of different risk factors: when the number of personnel exceeds the base number, an S-shaped growth function is used. A is the upper limit of the regional score, k is the growth coefficient, x is the actual number of people, x0 is the base number of people, and the number of hidden dangers / alarms adopts an exponential growth function: B is the upper limit of the regional score, m is the growth coefficient, n is the number of hidden dangers / alarms, and special operations adopt a step function (dividing the step score range according to the level of operation hazard). The risk correlation correction algorithm calculates the coupling coefficient between different risk factors (such as the coupling coefficient of "hot work - equipment failure alarm") using the Pearson correlation coefficient. If two factors exist simultaneously, their combined score is multiplied by (1 + coupling coefficient) for correction, avoiding double scoring or omission of associated risks.

[0013] The working principle of this invention is as follows: it revolves around the entire process of "rule initialization - data collection - algorithm calculation - risk assessment - visualization display", with the core algorithm module running through each evaluation stage, as detailed below: 1. Parameter initialization of the rule setting module and core algorithm The rule setting module provides basic parameters for the core algorithm: The risk benchmark setting unit determines the score ranges for high (80~100 points), medium (50~79 points), and low (0~49 points) risks, and the core algorithm uses this range as the final value for risk level determination. The scoring rules unit for special operations / personnel quantity / hazards / critical alarms allows users to define the upper limit of each factor's score (e.g., 50 points for special operation areas) and the base score (e.g., 20 points for hot work). The core algorithm initializes the parameters of the nonlinear scoring model accordingly (e.g., the A value of the S-shaped function and the B value of the exponential function). Meanwhile, the system imports the company's historical risk data (work records, accident files, alarm logs) for the past 3 years, trains the LSTM model with dynamic weight optimization algorithm and the random forest model with inherent risk assessment, and ensures that the initial accuracy of the algorithm is ≥85%.

[0014] 2. Algorithm Application of the Regional Inherent Risk Assessment Module This module quantifies inherent risk using a random forest regression model: Data acquisition: Obtain equipment aging coefficients from the equipment management system (e.g., a chemical workshop reactor has been in use for 8 years, has a design life of 10 years, and an aging coefficient of 0.8); obtain process complexity indexes from the process database (the workshop's process flow has 12 steps, with key operations accounting for 60%, so the complexity index = 12 × 0.6 = 7.2); and obtain historical accident risk values ​​from accident records (one major accident in the past 5 years, with a severity weight of 0.8, so the risk value = 1 × 0.8 = 0.8). Model calculation: Input the above features into the trained random forest model and output the inherent risk normalization score (e.g., 0.65). This score serves as the basis for subsequent risk calculations and is updated quarterly based on newly added equipment / process data.

[0015] 3. Algorithm calculation for the regional dynamic risk assessment module The system initiates dynamic risk calculation every 20 minutes, during which the nonlinear scoring and correlation correction of the core algorithm take effect. Data Acquisition: Real-time acquisition of special operation types (such as hot work and confined space work being carried out simultaneously in the workshop), number of personnel (35 people, base number 30 people), number of hidden dangers (1 major hidden danger + 2 general hidden dangers), and critical alarms (1 equipment failure alarm) in each area. Nonlinear scoring: Personnel quantity: Using an S-shaped function, A=10 (upper limit of regional scoring), k=1.5 (overload rate 17%, corresponding to k=1.0), scoring = 10 / (1+e^(-1.0×(35-30))) = 10 / (1+e^(-5)) ≈ 10 points; Hidden dangers: Using an exponential function, major hidden dangers B=30, m=0.4; general hidden dangers B=5, m=0.1, total score = 30×(1-e^(-0.4×1))+5×(1-e^(-0.1×2))≈30×0.33+5×0.19≈10.9+0.95≈11.85 points; Correlation correction: The correlation coefficient of "hot work - equipment failure alarm" is 0.7 (strong correlation), the coupling coefficient is 0.25, the correction value after superimposing the special operation score (20+15=35 points) and the alarm score (10 points) is (35+10)×(1+0.25)=56.25 points, which does not exceed the area limit of 50 points, and the final dynamic score is 50 points; Generate Record: Generates a scoring record from the above calculation process and results. It can be exported but cannot be modified.

[0016] 4. Algorithm integration of existing regional risk assessment modules This module determines the final risk value using a dynamic weight optimization algorithm. Weight adjustment: The dynamic score in the previous period was 40 points, the current dynamic score is 50 points, the change rate is 25% < 30%, the dynamic weight is 0.55, and the inherent weight is 0.45; Risk calculation: Current risk value = (0.65 × 100 × 0.45) + (50 × 0.55) = 29.25 + 27.5 = 56.75 points, corresponding to a medium risk level; Record generation: The system automatically records the time of risk change (2025-08-25 10:20), the value before the change (52 points), and the value after the change (56.75 points), and analyzes the reasons for the risk increase (overloading of personnel + new major hidden dangers), forming an unmodifiable assessment record.

[0017] 5. Visualization of the Risk Dynamics One-Chart Module Status display: The workshop is marked with a yellow icon as medium risk, and the risk duration is shown to be 2 hours. Factor analysis: The heat map shows the score percentages for personnel number (10 points, accounting for 17.6%), potential hazards (11.85 points, accounting for 21%), and special operations + alarms (50 points, accounting for 88.4%). The line graph shows the dynamic score changes over the past 24 hours. For details, click on the workshop name, and the right side will display detailed information such as the inherent risk score of 65 points, the dynamic scoring calculation process, and the correlation correction coefficient, supporting users to trace the source of risk.

[0018] Furthermore: Clarify the training data range, optimizer selection, loss function type, and iteration update cycle of the LSTM model in the dynamic weight optimization algorithm: The training dataset covers the company's "risk factor data (special operations, personnel fluctuations, etc.) + risk level records + accident data" for the past 3 years, ensuring the completeness of the data dimensions; The Adam optimizer (adaptive learning rate optimizer) and the root mean square error (RMSE) loss function are used to improve model training efficiency and prediction accuracy. A 7-day iteration and update cycle is set to ensure that the model can absorb new data in real time and adapt to the changing patterns of enterprise risks.

[0019] The core value of the LSTM model is "predicting the impact of risk factors based on historical data." By clearly defining the data range and parameters, it solves the problems of "insufficient model training data and lack of parameter standards." Data from the past three years can cover a sufficient number of risk scenarios (such as seasonal personnel fluctuations and periodic equipment maintenance) while avoiding prediction bias caused by outdated data (such as equipment and process data from five years ago). The Adam optimizer is adapted to the time-series data training requirements of LSTM, and the RMSE loss function can effectively measure the deviation between the "predicted impact" and the "actual impact", ensuring more accurate weight adjustment.

[0020] This transforms the dynamic weight optimization algorithm from a "theoretical framework" into an "executable technical solution": Enterprises no longer need to figure out model parameters on their own. They can directly prepare data and configure the model according to this standard to achieve automated and accurate weight adjustment, avoiding weight deviations caused by improper model parameters (such as misjudging the impact of "overloading of personnel" on risk).

[0021] Furthermore: For the S-shaped growth function of "personnel numbers exceeding the base number", a k-value is fitted using historical personnel overload accident data of the enterprise, and categorized according to "whether the overload rate exceeds 20%": Overload rate > 20% (high-risk scenario): k = 1.2~1.8, ensuring that the risk score of personnel number increases rapidly; Overload rate ≤20% (low-risk scenario): k=0.5~1.0, achieving a slow increase in risk score.

[0022] In line with the "actual changing pattern of personnel overload risk": In enterprise safety management, when the personnel overload rate is low (such as exceeding 10%), the probability of accidents increases slowly; however, when the overload rate exceeds the critical value (such as 20%), problems such as space congestion and difficulty in emergency evacuation will lead to an "explosive increase" in accident risk.

[0023] By setting different k values, the S-shaped function can accurately simulate this "slow-fast" risk change process, avoiding overestimation or underestimation of risk caused by a single k value.

[0024] To address the issue of a "one-size-fits-all" approach to assigning risk scores based on personnel numbers: For example, if a workshop has a base of 50 employees, and the actual number of employees is 55 (overload rate of 10%), the score increases slowly when calculated using k=0.8; however, when the actual number of employees is 65 (overload rate of 30%), the score rises rapidly when calculated using k=1.5, which can promptly alert management personnel to intervene and meets the actual safety control needs of the enterprise.

[0025] Furthermore: For the exponential growth function of "number of hidden dangers / alarms", the m value is set in three levels according to the urgency of the hidden dangers / alarms: High urgency (major hidden dangers / safety system alarms): m=0.3~0.5; Medium urgency level (significant hidden danger / equipment malfunction alarm): m=0.1~0.3; Low urgency level (general hidden danger / environmental anomaly alarm): m=0.05~0.1.

[0026] Following the safety management principle of "prioritizing the control of high-urgency risks": the characteristic of the exponential function is that "the larger the value of m, the faster the score increases". By setting a higher value of m for high-urgency risks, they can be quickly "highlighted" when scoring - for example, the score increase of 1 major hidden danger may be equivalent to the score increase of 3 general hidden dangers, ensuring that managers pay priority to high-risk items.

[0027] To avoid the drawbacks of "equal scoring for the number of hazards / alarms": For example, if a certain area has one "major hazard (m=0.4)" and three "general hazard (m=0.08)", the score of the major hazard (e^(0.4×1)-1≈0.49) is much higher than the total score of the three general hazard (3×(e^(0.08×1)-1)≈0.25) when calculated using an exponential function. This can guide managers to deal with the major hazard first and improve management efficiency.

[0028] Furthermore: Clarify the "correspondence between the coupling coefficient and the Pearson correlation coefficient," and set the coupling coefficient (0.05~0.3) in three levels according to the strength of the association between risk factors: Strong correlation (correlation coefficient ≥ 0.6, such as "confined space operation - oxygen concentration alarm"): coupling coefficient = 0.2~0.3; Moderate correlation (correlation coefficient 0.3~0.6, such as "high-altitude work - safety belt not fastened alarm"): coupling coefficient = 0.1~0.2; Weak correlation (correlation coefficient < 0.3, such as "equipment inspection - environmental temperature and humidity alarm"): coupling coefficient = 0.05~0.1.

[0029] To address the issue of "double scoring or omission of associated risks": When two risk factors are strongly correlated (such as the risk of oxygen concentration that inevitably accompanies confined space operations), simply adding up the scores will underestimate their combined risk; while completely merging the scores will lead to omissions.

[0030] By using "Pearson correlation coefficient to determine the strength of association and coupling coefficient to correct the superposition score", the scoring of association risk is made more accurate. When strong association factors are superimposed, multiply by (1 + high coupling coefficient) to reflect the joint risk and avoid double calculation.

[0031] For example, if a certain area has both "confined space operation (score 10)" and "oxygen concentration alarm (score 8)", the two are strongly correlated (coupling coefficient 0.25). After correction, the sum of the scores is (10+8)×(1+0.25)=22.5, instead of the simple sum of 18. This more accurately reflects the high-risk scenario of "confined space + insufficient oxygen" and reminds managers to take stricter control measures.

[0032] Furthermore: it clarifies that the random forest regression model will be used for the assessment of inherent regional risks, and specifies the "input features, output format, and accuracy requirements": Input features: Equipment aging coefficient (service life / design life), process complexity index (number of steps × proportion of critical operations), historical accident risk value (number of incidents in the past 5 years × severity weight); Output: Inherent risk normalized score from 0 to 1; Model training accuracy ≥ 85%.

[0033] Inherent risks are those that "exist long-term and are difficult to change quickly" for enterprises (such as aging equipment and complex processes), and require comprehensive assessment through multiple dimensions of characteristics. The equipment aging factor directly reflects the probability of equipment failure (e.g., equipment that has been used for 10 years has a design life of 20 years, so the aging factor is 0.5). The process complexity index reflects the difficulty of operation (e.g., if 6 out of 10 steps are critical operations, the complexity index = 6). Historical accident risk values ​​reflect the region's "inherent risks" (e.g., if two accidents have occurred in the past 5 years, with a severity weight of 0.5 each, the historical risk value = 1). Random forest regression models can effectively integrate these features, avoiding evaluation bias caused by a single feature, and an accuracy of 85% ensures reliable evaluation results.

[0034] Provide enterprises with a "regional risk baseline": for example, workshop A has an equipment aging coefficient of 0.6, a process complexity index of 8, and a historical accident risk value of 1.2, outputting an inherent risk score of 0.7 (high baseline); workshop B has corresponding indicators of 0.3, 4, and 0.5, outputting a score of 0.3 (low baseline). Managers can allocate control resources based on the baseline (e.g., increasing the frequency of inspections in workshop A).

[0035] Furthermore: Clarify the calculation logic of the current risk value: Current risk value = (Inherent risk score × Inherent weight) + (Total dynamic risk score × Dynamic weight), and the sum of the weights is 1, adjusted according to the "rate of change of dynamic risk factors": Change rate ≥ 30% (rapid increase / decrease in dynamic risk): Dynamic weight = 0.7~0.8, inherent weight = 0.2~0.3; Change rate < 30% (dynamic risk stable): dynamic weight = 0.5~0.6, inherent weight = 0.4~0.5.

[0036] Existing risks need to take into account both "inherent baselines" and "dynamic changes": When dynamic risks change drastically (e.g., the number of alarms in a certain area increases from 2 to 7 within 1 hour, a change rate of 250%), it indicates that the current risk is mainly dominated by dynamic factors, and the dynamic weight needs to be increased to make the existing risk value more in line with the real-time situation; when dynamic risks are stable (e.g., the number of people fluctuates by less than 5% for 8 consecutive hours), the impact of inherent risks is more significant, and the inherent weight needs to be appropriately increased to avoid ignoring long-term risks.

[0037] Achieving "dynamic balance calculation of risk value": For example, Workshop A has an inherent risk score of 0.7 (inherent weight 0.3) and a total dynamic risk score of 0.8 (dynamic weight 0.7). The current risk value = 0.7 × 0.3 + 0.8 × 0.7 = 0.77 (high risk). If the dynamic risk change rate drops to 10%, the weights are adjusted to 0.4 (inherent) and 0.6 (dynamic). The current risk value = 0.7 × 0.4 + 0.8 × 0.6 = 0.76 (risk slightly reduced). This can accurately reflect the risk change trend and provide a reliable basis for level determination.

[0038] Furthermore: Supplementing the core algorithm's "outlier handling mechanism": Outlier detection: Real-time data exceeds the 95% confidence interval of historical data (e.g., the historical peak number of people in a certain area is 50, the upper limit of the 95% confidence interval is 55, and the current data shows 100 people, which is considered an anomaly). Outlier identification: The Isolation Forest algorithm (suitable for anomaly detection in high-dimensional data) is used. Outlier handling: Replace with the average of the three adjacent periods (e.g., if the number of people in the period before and after the outlier period is 48 and 52, the average is 50, so replace with 100).

[0039] Real-time collected data may produce outliers due to equipment failure (such as personnel counter malfunction) or human error (such as false alarm recordings). If used directly for calculation, it will lead to serious deviations in risk values ​​(e.g., outlier data for 100 people will cause the personnel risk score to skyrocket).

[0040] The Isolation Forest algorithm can quickly identify this type of "outlier data," while "adjacent period mean replacement" can eliminate the impact of anomalies and maintain the continuity of data (avoiding data gaps caused by direct deletion), ensuring the stability of risk assessment.

[0041] Ensuring the "anti-interference capability" of risk assessment: For example, if a sensor malfunction causes the "number of alarms" to be mistakenly recorded as 50 times (the upper limit of the historical 95% confidence interval is 10 times), outlier processing is used to replace it with the average of 8 times from adjacent periods, avoiding the risk value from 0.5 (medium risk) to 0.9 (extremely high risk), and preventing managers from making decisions of over-control or under-control.

[0042] Furthermore: Clarify the "real-time requirements" and "deployment plan" of the core algorithm: The calculation delay is ≤5 seconds to ensure that the risk value can be updated quickly (e.g., changes in the number of personnel are reflected in the risk value within 5 seconds). Edge computing deployment: In multi-regional enterprises, each regional edge node runs a "non-linear scoring model + risk correlation correction algorithm" locally, and only uploads the summary data required for weight optimization (such as the regional dynamic risk total score) to the cloud.

[0043] Enterprise risk assessments need to balance "real-time performance" and "data transmission costs": A 5-second delay can meet emergency control needs (such as triggering an early warning if the risk value rises within 5 seconds after an alarm is triggered). Edge computing can reduce data transmission volume (e.g., when calculating scores for 100 people locally, only one total score needs to be uploaded to the cloud), avoiding bandwidth congestion caused by simultaneous uploading of large amounts of data from multiple regions, and reducing enterprise IT costs.

[0044] Adaptable to real-world scenarios for enterprises in multiple regions: For example, a group has 10 production workshops, each of which generates 1,000 risk data points per hour. If all of these data points were uploaded to the cloud for computation, 10,000 data points would need to be transmitted. With edge computing, each workshop only needs to upload one dynamic risk score, transmitting 10 data points per hour, reducing bandwidth usage by 99.9%, while the risk values ​​of each workshop can still be updated within 5 seconds, balancing efficiency and cost.

[0045] Furthermore: Clarify the "visual format" and "interactive functions" of the risk dynamics chart: Visualization: The heat map shows the real-time score percentage of each risk factor (e.g., "overloaded personnel" accounts for 30%, and "hazards" accounts for 50%), and the line graph shows the trend of score changes of each factor over the past 24 hours; Interaction: Clicking on the heatmap factor block will display "Nonlinear scoring calculation process + details of associated risk correction" (e.g., clicking "Hidden Dangers" will display "3 major hidden dangers, m=0.4, index calculation score 5.4, after correction with alarm association, it becomes 6.2").

[0046] Risk visualization needs to balance "intuitiveness" and "traceability": Heat maps enable managers to quickly identify key risk factors (e.g., the darkest color in a heat map indicates that the current risk mainly comes from potential hazards). Line charts can help analyze the reasons for changes in risk (e.g., if a factor's line chart suddenly rises at 10 o'clock, the cause can be investigated by referring to the work records at that time). Interactive features can meet the needs of managers to "know not only what" but also "why," avoiding situations where they only see risk values ​​but do not understand the basis for the calculations.

[0047] Enhancing the "transparency and efficiency" of risk management: For example, managers can discover through heatmaps that "high-altitude operations" account for 40% of the risk. After clicking on the block, they can view the details: "2 high-altitude operations, the basic score calculated by the step function is 8, and after correction with the safety belt not fastened alarm (correlation degree 0.15), it is 9.2." This allows for a quick determination that the source of the risk is "high-altitude operations + insufficient protection," enabling targeted safety inspections without having to check the original data one by one, thus improving management efficiency.

[0048] In summary, due to the adoption of the above technical solution, the beneficial effects of the present invention are: 1. A method for dynamic risk scoring and intelligent risk level determination for enterprises, which improves the accuracy of risk assessment and reduces the misjudgment rate. Based on the LSTM model, the weights of risk factors are adjusted in real time and can be dynamically adapted according to the characteristics of the enterprise's industry, regional risk characteristics, and risk change rate. For example, the "special operation weight" of high-risk areas (such as catalytic cracking workshops) of chemical enterprises is automatically increased to 0.4, and the "transaction alarm weight" of risk control centers of financial enterprises is automatically increased to 0.35, so that the matching degree between the weight and the actual risk impact reaches more than 90%. Compared with the existing technology with fixed weights, the assessment deviation rate is reduced to less than 8%.

[0049] Differentiated nonlinear functions are used for different risk factors: the risk of personnel overload is represented by an S-shaped function to reflect the "slow growth-rapid growth" characteristic; the risk of hidden dangers / alarms is represented by an exponential function to reflect the "rapid growth of high urgency" characteristic; and the risk of special operations is represented by a step function to reflect the "difference in hazard level". This improves the correlation between the scoring and the actual risk by 30% and effectively avoids misjudgment after the "risk threshold" (e.g., when the personnel overload rate is 40%, the risk score is revised from 12 points in the existing technology to 15 points to accurately reflect the high-risk state).

[0050] By calculating the coupling coefficient using the Pearson correlation coefficient, the risk of superimposed strong correlation factors is corrected (e.g., the score for "hot work + combustible gas alarm" is 25% higher than the simple summation), and the double scoring of weak correlation factors is avoided (e.g., the score for "equipment inspection + ambient temperature and humidity alarm" is 10% lower than the simple summation), thus improving the accuracy of correlation risk assessment to 92%, and reducing the distortion rate by 20 percentage points compared to the linear accumulation of existing technologies.

[0051] 2. In this invention, the system automatically calculates every 20 minutes, and the core algorithm's computational delay is ≤5 seconds. Compared to existing systems (1-hour update cycle, 30-second delay), the real-time performance is improved by more than 3 times. For example, transformer temperature alarms in power substations can be included in risk assessments within 5 minutes, providing the maintenance team with a 15-20 minute window for fault handling and reducing the accident rate by 60%.

[0052] The Isolation Forest algorithm identifies outliers outside the 95% confidence interval and replaces them with the mean of three adjacent periods, thus avoiding misjudgments of risk levels caused by outliers. Actual testing shows that after outlier processing, the misjudgment rate of risk levels is reduced from 18% with existing technologies to below 5%, ensuring that enterprise decisions are not influenced by false data.

[0053] It supports the local execution of nonlinear scoring and correlation correction algorithms at edge nodes, and only uploads the aggregated data (5MB / hour / region) required for weight optimization to the cloud, reducing the data transmission volume by 95%. In scenarios with limited network bandwidth (such as remote substations), the computation latency can still be controlled within 10 seconds, meeting the multi-region distributed monitoring needs of large enterprises.

[0054] 3. In this invention, industry-specific algorithm adaptation is achieved: the parameters of the core algorithm (such as nonlinear function coefficients k and m, and the range of coupling coefficients) can be adjusted according to industry characteristics—the "special operation-alarm coupling coefficient" for chemical enterprises is set to 0.2-0.3, and the "transaction alarm-system operation coupling coefficient" for financial enterprises is set to 0.15-0.25. No additional customized development is required for enterprises, the adaptation cost is reduced by 80%, and it can cover more than 10 industries such as chemical, power, and finance.

[0055] It achieves full automation of the entire process from "data collection - algorithm calculation - risk assessment - record generation" without human intervention. Compared with traditional manual assessment (requiring 2-3 people / day / region), it can reduce the labor input by 80%, and large enterprises can save 500,000 to 2 million yuan in labor costs annually.

[0056] Accurate risk levels and detailed assessment records (including the reasons for risk changes and the scoring calculation process) can serve as the basis for enterprises to formulate risk control measures (such as prioritizing the suspension of non-essential special operations in medium-risk areas) and can also provide compliance certificates to regulatory authorities, helping enterprises reduce unnecessary regulatory inspections and improve operational efficiency.

[0057] 4. In this invention, the LSTM model is iteratively updated every 7 days by adding new historical data, and the random forest model is optimized every quarter based on equipment / process data, so that the accuracy of the algorithm continues to improve with the accumulation of data. After running for 1 year, the risk assessment accuracy has increased from the initial 88% to more than 95%, and the operation and maintenance cost is reduced by 60% compared with the existing technology (which requires manual retraining of the model).

[0058] The core algorithm module adopts a modular design, which can add risk factors (such as "supply chain disruption risk" and "public opinion risk") and corresponding algorithms (such as time series prediction models for the supply chain) without reconstructing the system, reducing expansion costs by 70% and meeting the future risk management needs of enterprises. Attached Figure Description

[0059] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the embodiments will be briefly described below. It should be understood that the following drawings only show some embodiments of the present invention and should not be regarded as a limitation of the scope. For those skilled in the art, other related drawings can be obtained based on these drawings without creative effort, wherein: Figure 1 This is an architectural diagram of the present invention. Detailed Implementation

[0060] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only for explaining the invention and are not intended to limit the invention; that is, the described embodiments are merely some embodiments of the invention, and not all embodiments. The components of the embodiments of the invention described and shown in the accompanying drawings can generally be arranged and designed in various different configurations.

[0061] Therefore, the following detailed description of the embodiments of the invention provided in the accompanying drawings is not intended to limit the scope of the claimed invention, but merely to illustrate selected embodiments of the invention. All other embodiments obtained by those skilled in the art based on the embodiments of the invention without inventive effort are within the scope of protection of the invention.

[0062] It should be noted that relational terms such as "first" and "second" are used merely to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0063] The features and performance of the present invention will be further described in detail below with reference to embodiments.

[0064] Example 1 This invention provides a method for dynamic risk scoring and intelligent risk level determination for enterprises, such as... Figure 1 As shown, the specific implementation method of this embodiment is as follows: Chemical enterprise (a large petrochemical company's refinery) Company Background: The refinery has 5 production workshops (atmospheric and vacuum distillation workshop, catalytic cracking workshop, etc.). The main risk factors are special operations (hot work, confined space), equipment aging, hidden dangers (pipeline corrosion), and critical alarms (combustible gas leakage).

[0065] Rule and algorithm parameter settings: Risk benchmarks: High risk 80-100 points, medium risk 50-79 points, low risk 0-49 points; Special operations: Hot work 25 points, confined space operation 20 points, area scoring cap 60 points, coupling coefficient (hot work - combustible gas alarm) 0.3; Number of personnel: The basic number of personnel in the atmospheric and vacuum distillation workshop is 20, k=1.2, and the regional scoring cap is 15 points; Core algorithms: The LSTM model was trained on a dataset of 2,000 risk records from the past 3 years, with an RMSE of 0.08; the random forest model achieved an accuracy of 89%.

[0066] Risk assessment process: Inherent risks: The equipment aging coefficient of the atmospheric and vacuum distillation workshop is 0.7 (7 years of use, 10 years of design), the process complexity index is 8.5, the historical accident risk value is 0.6, and the inherent score of the random forest model output is 0.7 (out of 70). Dynamic Risks (2025-08-25 14:00): Special tasks: 1 hot work operation (25 points) + 1 confined space operation (20 points) = 45 points; Alarm: 1 combustible gas leak alarm (15 points), correlation correction = (45+15)×(1+0.3)=78 points, exceeding the upper limit of 60 points, take 60 points; Number of people: 28 (overload rate 40%), S-shaped score = 15 / (1 + e^(-1.2 × (28-20))) = 15 / (1 + e^(-9.6)) ≈ 15 points; Hidden danger: 1 major pipeline corrosion hazard, index score = 30 × (1 - e^(-0.4 × 1)) ≈ 10.9 points; Dynamic total score = 60 + 15 + 10.9 = 85.9 points; Existing risks: Dynamic change rate (85.9 - previous period 70) / 70 ≈ 22.7%, dynamic weight 0.55, inherent weight 0.45; Current risk value = 70 × 0.45 + 85.9 × 0.55 ≈ 31.5 + 47.25 ≈ 78.75 points (medium risk).

[0067] Application Results: The system sends out real-time risk warnings, prompting the company to immediately halt one non-essential confined space operation, reduce the number of personnel to 22, and after one hour, the dynamic total score drops to 62 points, maintaining the risk level at medium risk, thus preventing an accident from occurring.

[0068] Example 2 This invention provides a method for dynamic risk scoring and intelligent risk level determination for enterprises, such as... Figure 1 As shown, the specific implementation method of this embodiment is as follows: Power company (a substation of a provincial power company) Company Background: This substation is responsible for supplying power to the area. Risk factors include special operations (live-line work), number of personnel (maintenance personnel), equipment fault alarms (transformer temperature alarms), and hidden dangers (insulation aging).

[0069] Rule and algorithm parameter settings: Risk benchmarks: High risk 70-100 points, medium risk 40-69 points, low risk 0-39 points; Special operations: Live-line work 30 points, with a regional scoring cap of 40 points; Number of personnel: The base number of personnel in the operation and maintenance area is 5, k=0.8, and the maximum score is 8 points; Core algorithms: The LSTM model training dataset contains 1500 power fault records, with RMSE=0.07; the random forest model has an accuracy of 91%.

[0070] Risk assessment process: Inherent risks: The transformer has been in use for 12 years (design life of 20 years, aging factor of 0.6), process complexity index of 5.0 (5 steps in operation and maintenance process, 100% of which are critical operations), historical accident risk value of 0.3 (no major accidents in the past 5 years), inherent score = 0.6×0.3+5.0×0.4+0.3×0.3=0.18+2.0+0.09=2.27 (normalized score of 0.45, 45 points); Dynamic Risks (2025-08-25 16:40): Special task: 1 live-line work, 30 points (not exceeding the maximum of 40 points); Number of people: 8 (overload rate 60%), S-shaped score = 8 / (1 + e^(-0.8 × (8-5))) = 8 / (1 + e^(-2.4)) ≈ 8 × 0.916 ≈ 7.33 points; Alarm: 2 transformer temperature alarms, exponential function B=15, m=0.3, score = 15×(1-e^(-0.3×2))≈15×0.45≈6.75 points; Dynamic total score = 30 + 7.33 + 6.75 ≈ 44.08 points; Existing risks: Dynamic change rate (44.08 - previous cycle 35) / 35 ≈ 25.9%, dynamic weight 0.55, inherent weight 0.45; Current risk value = 45 × 0.45 + 44.08 × 0.55 ≈ 20.25 + 24.24 ≈ 44.49 points (medium risk).

[0071] Application results: The system identified a low correlation between "live work - temperature alarm" (coupling coefficient 0.08) and no additional correction was made; the maintenance team checked the transformer based on the warning and found that the heat sink was blocked. The cleaning was completed within 2 hours, the alarm was cleared, the dynamic total score dropped to 32 points, and the risk level changed to low risk.

[0072] Example 3 This invention provides a method for dynamic risk scoring and intelligent risk level determination for enterprises, such as... Figure 1 As shown, the specific implementation method of this embodiment is as follows: Financial enterprise (Headquarters of a city commercial bank) Company Background: The bank's main risk factors are abnormal transaction alerts (large transfers, logins from different locations), number of personnel (risk control personnel), potential risks (system vulnerabilities), and special operations (core system upgrades).

[0073] Rule and algorithm parameter settings: Risk benchmarks: High risk 90-100 points, medium risk 60-89 points, low risk 0-59 points; Key alerts: Large transfer alert 20 points, unauthorized login alert 15 points, regional scoring cap 50 points; Staffing: The risk control center has a base of 10 people, k=1.2, and a maximum score of 12 points; Core algorithms: The LSTM model training dataset contains 3000 financial risk records with RMSE=0.06; the random forest model has an accuracy of 92%.

[0074] Risk assessment process: Inherent risks: The core system has been in use for 5 years (design life of 15 years, aging factor of 0.3), the risk control process complexity index is 8.0 (8-step process, key operations account for 80%), the historical accident risk value is 0.5 (one data breach in the past 5 years), and the inherent score is 0.3×0.2+8.0×0.5+0.5×0.3=0.06+4.0+0.15=4.21 (normalized score is 0.52, 52 points). Dynamic Risks (2025-08-25 09:00): Alarm: 3 large transfer alarms + 2 abnormal login alarms, exponential function B=50, m=0.25, score = 50×(1-e^(-0.25×(3+2)))=50×(1-e^(-1.25))≈50×0.713≈35.65 points; Number of personnel: 7 (below the base of 10), score: 0; Special assignment: Core system upgrade (25 points), with a correlation coefficient of 0.65 and a coupling coefficient of 0.2 with "system vulnerability risks". The risk is worth 10 points. After correction, the score is (25+10)×(1+0.2)=42 points. Dynamic total score = 35.65 + 0 + 42 ≈ 77.65 points; Existing risks: Dynamic change rate (77.65 - previous period 55) / 55 ≈ 41.2%, dynamic weight 0.75, inherent weight 0.25; Current risk value = 52 × 0.25 + 77.65 × 0.75 ≈ 13 + 58.24 ≈ 71.24 points (medium risk).

[0075] Application Results: The bank immediately suspended non-essential core system upgrades, and the risk control team verified that three large-amount transfers were normal corporate settlements. Two hours later, the dynamic total score dropped to 58 points, the risk level changed to low risk, and the safety of funds was ensured.

[0076] The above description is merely a preferred embodiment of the present invention and is not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, and improvements made by those skilled in the art within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for dynamic risk scoring and intelligent level determination for enterprises, characterized in that: Includes the following modules: Rule setting module: includes risk benchmark setting unit, special operation scoring rule unit, personnel quantity scoring rule unit, hidden danger scoring rule unit, and critical alarm quantity scoring rule unit; Regional Inherent Risk Assessment Module: Assigns inherent risk scores to each risk area; Regional dynamic risk assessment module: The system automatically calculates and generates dynamic risk values, and generates unmodifiable scoring records; The existing risk assessment module in the region automatically calculates the current risk value and generates an unmodifiable risk assessment record. Risk Dynamics in One Chart Module: Statistically displays the status of risk areas and influencing factors; It also includes a core algorithm module, which is used to optimize the accuracy of risk value calculation, including dynamic weight optimization algorithm, nonlinear scoring model, and risk correlation correction algorithm; The dynamic weight optimization algorithm is based on historical risk records and real-time risk change data. It uses a long short-term memory network model to predict the impact of each risk factor on the final risk and adjusts the weight of each factor in real time. The nonlinear scoring model employs differentiated nonlinear functions tailored to the characteristics of different risk factors: when the number of personnel exceeds the base number, an S-shaped growth function is used. A is the upper limit of the regional score, k is the growth coefficient, x is the actual number of people, x0 is the base number of people, and the number of hidden dangers / alarms adopts an exponential growth function: B is the upper limit of the regional score, m is the growth coefficient, n is the number of hidden dangers / alarms, and a step function is used for special operations; The risk correlation correction algorithm calculates the coupling coefficient between different risk factors using the Pearson correlation coefficient. If two factors exist simultaneously, their combined score is multiplied by (1 + coupling coefficient) for correction, thus avoiding double scoring or omission of associated risks.

2. The method according to claim 1, characterized in that, In the dynamic weight optimization algorithm, the training dataset of the LSTM model includes risk factor data of the enterprise in the past 3 years, corresponding risk level change records and accident occurrence data. The training process uses the Adam optimizer, the loss function is the root mean square error, and the model is iteratively updated every 7 days by adding new historical data.

3. The method according to claim 1, characterized in that, In the nonlinear scoring model, the growth coefficient k of the S-shaped growth function is determined by fitting the historical personnel overload accident data of the enterprise: when the personnel overload rate exceeds 20%, k takes a value of 1.2~1.8; when the overload rate is ≤20%, k takes a value of 0.5~1.0, ensuring that the risk of personnel number changes reasonably with the overload rate in the form of "slow growth-rapid growth".

4. The method according to claim 1, characterized in that, In the nonlinear scoring model, the growth coefficient m of the exponential growth function is set according to the urgency of the hazard / alarm: m is 0.3~0.5 for major hazards / safety system alarms, m is 0.1~0.3 for relatively large hazards / equipment failure alarms, and m is 0.05~0.1 for general hazards / environmental anomaly alarms, so that the score of high urgency risks grows faster.

5. The method according to claim 1, characterized in that, In the aforementioned risk correlation correction algorithm, the coupling coefficient ranges from 0.05 to 0.3: when the Pearson correlation coefficient between two risk factors is ≥0.6, the coupling coefficient is 0.2 to 0.

3. The correlation coefficient is 0.3~0.6, and the coupling coefficient is 0.1~0.

2. The correlation coefficient is less than 0.3, and the coupling coefficient is between 0.05 and 0.

1.

6. The method according to claim 1, characterized in that, In the regional inherent risk assessment module, the inherent risk score is assigned using a random forest regression model. The model input features include equipment aging coefficient, process complexity index, and historical accident risk value. The output is an inherent risk normalization score of 0 to 1. The model training accuracy must be ≥85%.

7. The method according to claim 1, characterized in that, In the existing risk assessment module of the region, the current risk value is calculated using the following formula: Current risk value = (Inherent risk score × Inherent weight) + (Total dynamic risk score × Dynamic weight); where the sum of the inherent weight and the dynamic weight is 1, and the dynamic weight is adjusted in real time by the dynamic weight optimization algorithm: when the rate of change of dynamic risk factors (current dynamic score - previous period dynamic score) / previous period dynamic score ≥ 30%, the dynamic weight is 0.7~0.8, and the inherent weight is 0.2~0.3; when the rate of change is < 30%, the dynamic weight is 0.5~0.6, and the inherent weight is 0.4~0.

5.

8. The method according to claim 1, characterized in that, The core algorithm module also includes an outlier handling sub-algorithm: when the real-time collected risk factor data exceeds the 95% confidence interval of historical data, the isolated forest algorithm is used to identify outliers, and the outliers are replaced by the mean of three adjacent periods to avoid risk assessment bias caused by outlier data.

9. The method according to claim 1, characterized in that, The core algorithm module has a computation latency of ≤5 seconds and supports edge computing deployment: when an enterprise has multi-region distributed risk monitoring, each regional edge node runs a nonlinear scoring model and risk correlation correction algorithm locally, and only uploads the aggregated data required for weight optimization to the cloud, reducing data transmission bandwidth usage.

10. The method according to claim 1, characterized in that, In the aforementioned dynamic risk graph module, the dynamic risk influencing factor analysis adopts a combination of heat map and line graph: the heat map displays the real-time score percentage of each risk factor, and the line graph displays the changing trend of each factor's score over the past 24 hours. Users can also click on the factor block in the heat map to view the non-linear scoring calculation process and related risk correction details corresponding to that factor.

Images