A data mining-based enterprise operation risk prediction method
By using data mining techniques to periodically slice and slot-arrange enterprise operational risks, and combining improved DLinear network prediction and risk inertia judgment, the problems of accuracy in enterprise operational risk prediction and unclear early warning results are solved, and efficient risk identification and adjustment suggestions are realized.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- TONGHUI TECH TRANSFER (ZAOZHUANG SHANTING DISTRICT) CO LTD
- Filing Date
- 2026-04-17
- Publication Date
- 2026-07-10
AI Technical Summary
Existing technologies are insufficient to fully preserve the evolutionary relationships of risks during business operations, and cannot standardize and uniformly represent multiple types of business states, resulting in inaccurate risk predictions and unclear early warning results.
Using a data mining-based approach, through operating cycle slicing, slot arrangement, improved DLinear network prediction, and risk inertia assessment, abnormal transitions are identified, and operating risk levels and adjustment suggestions are generated.
It improved the accuracy of operational risk forecasting and the ability to identify anomalies, enhanced the clarity of early warning results and the effectiveness of operational adjustments, and reduced the impact of risk spread.
Smart Images
Figure CN122367170A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of enterprise business risk prediction technology, and in particular to an enterprise business risk prediction method based on data mining. Background Technology
[0002] With the increasing complexity of the business environment and the growing demand for business risk early warning, comprehensive analysis and business risk prediction technologies that consider data on business operations, contract performance, financial status, inventory status, supply status, and external disturbances have received widespread attention. Existing methods for identifying or warning of business risks primarily rely on statistical analysis of financial indicators, monitoring of single business data, or simple time-series forecasting to assess business risks. However, these methods generally suffer from the following problems in practical applications: The business records in enterprise management ledgers come from diverse sources and have complex relationships. Different business records often involve multiple status information such as business operations, contract performance, funds, inventory, supply, and external disturbances. Existing methods usually analyze data directly using single data tables or single indicators, making it difficult to establish a unified organizational structure around the business cycle and business items. This makes it difficult to fully preserve the risk evolution relationship in the business process. Data from multiple business cycles suffers from structural inconsistencies, missing slots, and sequence breaks during time segmentation, item merging, and status alignment. Existing technologies often struggle to achieve standardized arrangement and unified representation of multiple types of business statuses, affecting the accuracy and stability of subsequent risk modeling. For business risk sequences with trend changes, fluctuations, and local abnormal transitions, traditional statistical forecasting methods or ordinary linear time series analysis methods cannot simultaneously characterize the continuous evolution trend of business risks and the characteristics of local peaks and troughs. This can easily lead to inaccurate identification of key abnormal cycles and unclear risk level classification, thereby affecting the reliability of enterprise business risk warning results and the effectiveness of business adjustment recommendations.
[0003] Therefore, how to provide a data mining-based method for predicting business risks is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention
[0004] One objective of this invention is to propose a data mining-based method for predicting business risks. This invention fully utilizes business cycle slicing, slot arrangement, slot foldback input mapping, improved DLinear network prediction, and risk inertia judgment and abnormal transition identification methods. It describes in detail the implementation process of predicting business risks, classifying risk levels, and generating business adjustment suggestions. It has the advantages of high risk prediction accuracy, strong anomaly identification capability, and clear and effective early warning results.
[0005] A method for predicting business risks based on data mining according to an embodiment of the present invention includes the following steps: Step 1: Based on the target company's operating cycle identifier and operating event association identifier, perform retrieval and extraction processing on the operating records in the company's operating ledger to obtain operating risk association data; Step 2: Divide the operational risk-related data into periodic slices according to the operational cycle identifier, and arrange them into slots according to the operational item-related identifier to obtain the operational status unit set; Step 3: Based on the set of operational state units, generate an operational risk state tensor through state vector organization and slot return input mapping; Step 4: Input the operational risk state tensor into the improved DLinear network, perform trend decomposition, alternating recursion and peak-valley enhancement prediction processing to obtain the target operational risk prediction sequence; the improved DLinear network includes a sequence decomposition module, a trend-fluctuation alternating recursion module, a basic prediction generation module, an abnormal peak-valley location enhancement module and an output fusion module; Step 5: Based on the target business risk prediction sequence, risk inertia judgment is performed using the rescaled range analysis method, and abnormal transition cycles are identified using the Z-score method to obtain the business risk judgment result; Step Six: Based on the operational risk assessment results, determine the operational risk level of the target company; Step 7: Based on the operational risk level, generate operational adjustment suggestions and output the operational risk assessment results, operational risk level, and operational adjustment suggestions to the risk management terminal.
[0006] Optionally, the operational risk-related data includes operational business data, performance status data, financial status data, inventory status data, supply status data, and external disturbance data.
[0007] Optionally, step two specifically includes: Extract the operating cycle identifier from each operating record and determine the operating cycle boundary according to the operating cycle identifier; Based on the boundaries of the operating cycle, the operating risk-related data is segmented into several operating cycle data segments. The data segments of each operating cycle are processed sequentially according to the order in which they are identified by the operating cycle identifiers. Within each business cycle data segment, extract the business item association identifiers for each business record, and merge the business records according to the business item association identifiers to obtain multiple item record groups; Within each data segment of an operating cycle, a slot structure is established, which includes operating business slots, fulfillment slots, capital slots, inventory slots, supply slots, and external disturbance slots; The business records in each item record group are arranged in slots according to data categories: business data is arranged in the business slot, performance status data is arranged in the performance slot, funds status data is arranged in the funds slot, inventory status data is arranged in the inventory slot, supply status data is arranged in the supply slot, and external disturbance data is arranged in the external disturbance slot. For multiple business records grouped into the same slot within the same item record group, write them sequentially into the corresponding slot according to the order in which the multiple business records are arranged in the corresponding business cycle data segment; For slots without operational records, retain the corresponding slot location and record it as an empty slot. The slot structure in the same operating cycle data segment is combined into units according to the slot order to generate an operating status unit; Arrange the operating status units of each operating cycle according to the operating cycle identifier to obtain the operating status unit set.
[0008] Optionally, step three specifically includes: Based on the set of operating status units, the record sequence of each slot in each operating status unit is processed into a status vector, and slot alignment is performed according to the slot order to obtain a standard slot arrangement vector sequence. The standard slot arrangement vector sequence is processed using slot-folded input mapping: The standard slot arrangement vector sequence is mapped to the positive slot through linear mapping to obtain the positive slot feature vector sequence; The standard slot arrangement vector sequence is rearranged in reverse order according to the slot order, and the slot reversal mapping is performed through linear mapping to obtain the reversal slot feature vector sequence. The forward slot feature vector sequence and the foldback slot feature vector sequence are bidirectionally interleaved and fused to obtain the slot fused feature vector sequence. The slot fusion feature vector sequence is compressed in dimension and linearly mapped to obtain the standard state feature vector sequence of the corresponding business state unit. The standard state feature vector sequence of each operating state unit is stacked into tensors according to the operating cycle identifier order to generate an operating risk state tensor. The structure of the operating risk state tensor includes an operating cycle dimension, a slot structure dimension, and a feature dimension.
[0009] Optionally, the sequence decomposition module and the trend-fluctuation alternating recursion module specifically include: In the sequence decomposition module, for each slot structure dimension and feature dimension, a moving average operation is performed on the operating risk state tensor along the operating cycle dimension to generate a trend risk tensor. Calculate the initial residual tensor based on the operational risk state tensor and the trend risk tensor; For the trend risk tensor, the trend change between adjacent operating cycles is calculated along the operating cycle dimension to generate the trend change tensor. The initial residual tensor and the trend change tensor are concatenated in the feature dimension to obtain the correction input tensor. Then, a local offset correction mapping is performed through linear mapping and GELU activation to generate a local offset correction tensor. The initial residual tensor and the local offset correction tensor are interpolated element-wise to obtain the offset correction residual tensor; the offset correction residual tensor is then subjected to sign-preserving compression mapping to generate the volatility risk tensor. The sign-preserving compression mapping specifically involves: extracting the direction identifier value for each residual element in the offset-corrected residual tensor using the sign function; performing a natural logarithmic compression operation after incrementing the absolute value of each residual element by 1 to obtain the compression amplitude; multiplying each compression amplitude by the corresponding direction identifier value to obtain the fluctuation state value of each residual element; and reorganizing each fluctuation state value according to its original position in the offset-corrected residual tensor to generate a fluctuation risk tensor. In the trend-volatility alternating recursion module, K rounds of alternating recursion are performed on the trend risk tensor and volatility risk tensor, specifically: The trend risk tensor is denoted as the initial trend recursion tensor, and the volatility risk tensor is denoted as the initial volatility recursion tensor. For the alternating recursion of the kth round, the trend recursion tensor of the (k-1)th round is mapped along the operating cycle dimension through a linear mapping to generate the intermediate trend tensor of the kth round. The trend intermediate tensor of the kth round and the fluctuation recursion tensor of the (k-1)th round are concatenated in the feature dimension, and fluctuation reconstruction is performed through linear mapping and GELU activation to generate the fluctuation reconstruction tensor of the kth round. For the volatility reconstruction tensor of the k-th round, calculate the difference tensor between adjacent operating cycles along the operating cycle dimension, and perform linear compression mapping on the difference tensor to generate the volatility recursive tensor of the k-th round. The fluctuation recursion tensor of the k-th round is cumulatively summed along the operating cycle dimension to generate the fluctuation accumulation tensor. The fluctuation accumulation tensor and the trend intermediate tensor of the k-th round are then subjected to the same-dimensional additive write-back mapping to generate the trend reconstruction tensor of the k-th round. The trend reconstruction tensor of the k-th round is used as the trend recursion tensor of the (k+1)-th round. Repeatedly execute trend recursion, fluctuation reconstruction, fluctuation recursion and trend reconstruction until the end of the Kth round, to obtain the trend reconstruction tensor and fluctuation recursion tensor of the Kth round.
[0010] Optionally, the basic prediction generation module, the abnormal peak and valley localization enhancement module, and the output fusion module specifically include: In the basic forecast generation module, for the trend reconstruction tensor and volatility recursion tensor of the Kth round, linear forecast mapping oriented towards the future operating cycle dimension is performed along the operating cycle dimension to generate the trend forecast tensor and volatility forecast tensor; the trend forecast tensor and volatility forecast tensor are added element by element to generate the basic operating risk forecast tensor. In the abnormal peak and valley positioning enhancement module, the basic operational risk prediction tensor is used to calculate the local change response between adjacent future operational cycles along the future operational cycle dimension, and generate the local change response tensor. Based on the switching relationship of the changing directions of the positions of each element in the local change response tensor, the candidate peak position and candidate valley position are located, specifically as follows: If the response value of the previous local change at the current element position is greater than 0 and the response value of the next local change is less than 0, then the current element position is determined as a candidate peak position. If the response value of the previous local change at the current element position is less than 0 and the response value of the next local change is greater than 0, then the current element position is determined as a candidate valley position. Construct a peak-valley identifier tensor by assigning a value of 1 to the element corresponding to the candidate peak position, a value of -1 to the element corresponding to the candidate valley position, and a value of 0 to the remaining positions. Based on the basic operational risk prediction tensor, local prediction segments with a radius of R are extracted along the future operational cycle, centered on the future operational cycle where the candidate peak or candidate valley position is located, to obtain the local prediction segment tensor. The peak-valley identifier tensor and the local prediction segment tensor are concatenated in the feature dimension, and a peak-valley enhanced feature tensor is generated through linear mapping and GELU activation. In the output fusion module, the basic business risk prediction tensor and the peak-valley enhancement feature tensor are aligned in the same dimension and added element by element. Then, linear compression mapping and scalar projection are performed along the feature dimension to obtain the target business risk prediction value corresponding to each future business cycle. The target business risk prediction sequence is generated by arranging the values according to the order of the future business cycles.
[0011] Optionally, step five specifically includes: Calculate the mean and standard deviation of the target business risk prediction series to obtain the mean and standard deviation of business risk; For the target business risk forecast series, the Hurst index is calculated using rescaled range analysis, specifically including: Calculate the difference between the target operating risk forecast and the mean operating risk for each future operating cycle to obtain the mean-adjusted sequence; Based on the mean-adjusted sequence, the cumulative deviation value for each future operating cycle is calculated to obtain the cumulative deviation sequence, and the range of the cumulative deviation sequence is calculated to obtain the range term of the rescaled range analysis method. The ratio of the range term to the standard deviation of operating risk is used as the rescaled range statistic, and a power function relationship is established between the rescaled range statistic and the total number of future operating cycles. By taking the natural logarithm on both sides of the power function relationship, we obtain the inertial fitting formula for operational risk. Based on the operational risk inertia fitting formula, the least squares method is used to perform linear fitting to obtain the operational risk inertia fitting line; the slope of the operational risk inertia fitting line is used as the Hurst exponent of the target operational risk prediction sequence. The target business risk prediction sequence is cyclically shifted according to the future business cycle to generate several locally rearranged reference sequences. The reference Hurst index of each locally rearranged reference sequence is calculated using the rescaled range analysis method. The mean of each reference Hurst index is calculated to obtain the Hurst reference value. If the Hurst index of the target business risk prediction sequence is greater than the Hurst reference value, the target business risk prediction sequence is determined to have risk inertia; otherwise, the target business risk prediction sequence is determined not to have risk inertia, and risk inertia judgment flags are generated respectively. Based on the mean and standard deviation of operating risk, the standard risk value for each future operating cycle is calculated using the Z-score method. The standard risk values for each future operating cycle are sorted according to their numerical magnitude to obtain a standard risk value sequence; the upper quartile and median values of the standard risk value sequence are extracted. Calculate the difference between the upper quartile value and the median value, amplify the difference proportionally, and add it to the upper quartile value to obtain the transition judgment benchmark value; The future operating cycle position where the standard risk value is greater than or equal to the threshold value for determining the transition is used as the candidate abnormal transition cycle. When the target business risk prediction sequence is determined to have risk inertia, each candidate abnormal transition cycle is identified as an abnormal transition cycle. The operational risk assessment result is composed of the risk inertia judgment indicator, the abnormal transition cycle, the target operational risk prediction value corresponding to the abnormal transition cycle, and the standard risk value.
[0012] Optionally, step six specifically includes: Based on the risk inertia assessment criteria, determine whether the target business risk prediction sequence has a continuous evolution trend; If the target business risk prediction sequence does not have a continuous evolution trend and there is no abnormal transition cycle, then the business risk level of the target enterprise is determined to be low risk. If the target business risk prediction sequence has a continuous evolution trend and there is no abnormal transition cycle, then the business risk level of the target enterprise is determined to be medium risk. If the target business risk prediction sequence does not have a continuous evolution trend and there is an abnormal transition cycle, then the business risk level of the target enterprise will be determined as medium risk. If the target business risk prediction sequence has a continuous evolution trend and there is an abnormal transition cycle, then the business risk level of the target enterprise will be determined as high risk.
[0013] Optionally, the operational adjustment recommendations specifically include: If the operational risk level is low, maintain the current operational arrangements and continuously monitor changes in operational risk. If the operational risk level is medium risk, then strengthen the monitoring of abnormal transition cycles and check for abnormal changes in operational data, performance status data, capital status data, inventory status data, supply status data, or external disturbance data. If the operational risk level is high, the operational arrangements will be adjusted, key operational matters corresponding to the abnormal transition cycle will be prioritized for verification, and abnormal changes in financial status data, inventory status data, supply status data, or contract performance data will be prioritized for handling.
[0014] The beneficial effects of this invention are: This invention retrieves and extracts operational records from the enterprise's operational ledger based on the target enterprise's operational cycle identifier and operational event association identifier to obtain operational risk-related data. Furthermore, it performs cycle slicing according to the operational cycle identifier and slot arrangement according to the operational event association identifier to generate an operational status unit set. This unifies operational business data, performance status data, capital status data, inventory status data, supply status data, and external disturbance data within the operational cycle framework. Further, through state vector organization and slot-based input mapping, an operational risk status tensor is generated, enhancing the alignment and joint representation capability between different slot status information and reducing structural fragmentation and information omissions that easily occur when directly splicing multiple operational statuses. In the prediction stage, the operational risk status tensor is input into an improved DLinear network. Through a sequence decomposition module, a trend-fluctuation alternating recursion module, a basic prediction generation module, an anomaly peak and valley location enhancement module, and an output fusion module, it achieves joint prediction of long-term trends, short-term fluctuations, and local anomaly peak and valley changes in operational risks, improving the target operational risk prediction sequence's ability to characterize continuously evolving risks and sudden leap risks. In the assessment phase, risk inertia is assessed using the Hurst index, and anomalous transition cycles are identified using the Z-score method. Furthermore, the risk inertia assessment indicators and anomalous transition cycles are combined to complete the classification of operational risk levels. This ensures that the operational risk identification results not only reflect the strength of the risk but also whether the risk has a continuous diffusion trend and the occurrence cycle of key anomalies. Finally, operational adjustment suggestions are generated based on the operational risk level, and the operational risk assessment results, operational risk level, and operational adjustment suggestions are output to the risk management terminal. This improves the accuracy of enterprise operational risk prediction, the sensitivity of anomalous transition identification, the clarity of risk level classification, and the effectiveness of operational adjustment suggestions. Therefore, this patent is of great significance for improving enterprises' operational risk early warning capabilities, reducing the impact of operational anomaly diffusion, and enhancing the timeliness of enterprise operational decisions. Attached Figure Description
[0015] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings: Figure 1 This is a schematic diagram of a data mining-based enterprise business risk prediction method proposed in this invention; Figure 2 This is a flowchart of the improved DLinear network structure in a data mining-based enterprise business risk prediction method proposed in this invention. Figure 3 This is a flowchart of the business risk assessment process in a data mining-based business risk prediction method proposed in this invention. Detailed Implementation
[0016] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.
[0017] refer to Figures 1-3 A data mining-based method for predicting business risks includes the following steps: Step 1: Based on the target company's operating cycle identifier and operating event association identifier, perform retrieval and extraction processing on the operating records in the company's operating ledger to obtain operating risk association data; Step 2: Divide the operational risk-related data into periodic slices according to the operational cycle identifier, and arrange them into slots according to the operational item-related identifier to obtain the operational status unit set; Step 3: Based on the set of operational state units, generate an operational risk state tensor through state vector organization and slot return input mapping; Step 4: Input the operational risk state tensor into the improved DLinear network, perform trend decomposition, alternating recursion and peak-valley enhancement prediction processing to obtain the target operational risk prediction sequence; wherein, the improved DLinear network includes a sequence decomposition module, a trend-fluctuation alternating recursion module, a basic prediction generation module, an abnormal peak-valley location enhancement module and an output fusion module. Step 5: Based on the target business risk prediction sequence, risk inertia judgment is performed using the rescaled range analysis method, and abnormal transition cycles are identified using the Z-score method to obtain the business risk judgment result; Step Six: Based on the operational risk assessment results, determine the operational risk level of the target company; Step 7: Based on the operational risk level, generate operational adjustment suggestions and output the operational risk assessment results, operational risk level, and operational adjustment suggestions to the risk management terminal.
[0018] In this embodiment, the operational risk-related data includes operational business data, performance status data, capital status data, inventory status data, supply status data, and external disturbance data; The data includes: business operation data (order amount, order quantity, and business change records); performance status data (agreed delivery time, actual delivery time, and default records); financial status data (operating revenue, operating expenses, and cash flow changes); inventory status data (inventory quantity, inventory turnover cycle, and stockout records); supply status data (number of deliveries, delivery interval, and supply interruption records); and external disturbance data (market price fluctuation records, policy change records, and logistics anomaly records).
[0019] In this embodiment, step two specifically includes: Extract the operating cycle identifier from each operating record and determine the operating cycle boundary according to the operating cycle identifier; Based on the boundaries of the operating cycle, the operating risk-related data is segmented into several operating cycle data segments. The data segments of each operating cycle are executed sequentially according to the order of their identification. Within each business cycle data segment, extract the business item association identifiers for each business record, and merge the business records according to the business item association identifiers to obtain multiple item record groups; Within each data segment of an operating cycle, a slot structure is established, which includes operating business slots, fulfillment slots, funding slots, inventory slots, supply slots, and external disturbance slots. The business records in each item record group are arranged in slots according to data categories: business data is arranged in the business slot, performance status data is arranged in the performance slot, funds status data is arranged in the funds slot, inventory status data is arranged in the inventory slot, supply status data is arranged in the supply slot, and external disturbance data is arranged in the external disturbance slot. For multiple business records grouped into the same slot within the same item record group, write them sequentially into the corresponding slot according to the order in which the multiple business records are arranged in the corresponding business cycle data segment; For slots without operational records, retain the corresponding slot location and record it as an empty slot. The slot structure in the same operating cycle data segment is combined into units according to the slot order to generate an operating status unit; Arrange the operating status units of each operating cycle according to the operating cycle identifier to obtain the operating status unit set.
[0020] In this embodiment, step three specifically includes: Based on the set of operating status units, the record sequence of each slot in each operating status unit is processed into a status vector, and slot alignment is performed according to the slot order to obtain a standard slot arrangement vector sequence. The standard slot arrangement vector sequence is processed using slot-folded input mapping: The standard slot arrangement vector sequence is mapped to the positive slot through linear mapping to obtain the positive slot feature vector sequence; The standard slot arrangement vector sequence is rearranged in reverse order according to the slot order, and the slot reversal mapping is performed through linear mapping to obtain the reversal slot feature vector sequence. The forward slot feature vector sequence and the folded-back slot feature vector sequence are bidirectionally interleaved and fused to obtain the slot fused feature vector sequence. Specifically, the bidirectional interleaved splicing and fusion is performed by aligning the forward slot feature vector sequence and the folded-back slot feature vector sequence one by one according to the slot position, then arranging them alternately in the vector dimension and performing fusion mapping compression processing. The slot fusion feature vector sequence is compressed in dimension and linearly mapped to obtain the standard state feature vector sequence of the corresponding business state unit. The standard state feature vector sequences of each operating state unit are stacked into tensors according to the operating cycle identifier order to generate the operating risk state tensor. The structure of the operating risk state tensor includes the operating cycle dimension, the slot structure dimension, and the feature dimension.
[0021] In this embodiment, the sequence decomposition module and the trend-fluctuation alternating recursion module specifically include: In the sequence decomposition module, for each slot structure dimension and feature dimension, a moving average operation is performed on the operating risk state tensor along the operating cycle dimension to generate a trend risk tensor. Calculate the initial residual tensor based on the operational risk state tensor and the trend risk tensor; For the trend risk tensor, the trend change between adjacent operating cycles is calculated along the operating cycle dimension to generate the trend change tensor. The initial residual tensor and the trend change tensor are concatenated in the feature dimension to obtain the correction input tensor. Then, a local offset correction mapping is performed through linear mapping and GELU activation to generate a local offset correction tensor. The initial residual tensor and the local offset correction tensor are interpolated element-wise to obtain the offset correction residual tensor; the offset correction residual tensor is then subjected to sign-preserving compression mapping to generate the volatility risk tensor. The sign-preserving compression mapping is as follows: For each residual element in the offset-corrected residual tensor, the direction identifier value is extracted using the sign function; the absolute value of each residual element is incremented by 1 and then subjected to natural logarithmic compression to obtain the compression amplitude; each compression amplitude is multiplied by the corresponding direction identifier value to obtain the fluctuation state value of each residual element; each fluctuation state value is reorganized according to its original position in the offset-corrected residual tensor to generate the fluctuation risk tensor. In the trend-volatility alternating recursion module, K rounds of alternating recursion are performed on the trend risk tensor and volatility risk tensor, specifically: The trend risk tensor is denoted as the initial trend recursion tensor, and the volatility risk tensor is denoted as the initial volatility recursion tensor. For the k-th round of alternating recursion, the trend recursion tensor of the (k-1)-th round is mapped along the operating cycle dimension using a linear mapping to generate the intermediate trend tensor of the k-th round; where k ranges from 1 to K. The trend intermediate tensor of the kth round and the fluctuation recursion tensor of the (k-1)th round are concatenated in the feature dimension, and fluctuation reconstruction is performed through linear mapping and GELU activation to generate the fluctuation reconstruction tensor of the kth round. For the volatility reconstruction tensor of the k-th round, calculate the difference tensor between adjacent operating cycles along the operating cycle dimension, and perform linear compression mapping on the difference tensor to generate the volatility recursive tensor of the k-th round. The volatility recursion tensor of the k-th round is cumulatively summed along the operating cycle dimension to generate the volatility cumulative tensor. The volatility cumulative tensor is then subjected to the same-dimensional additive write-back mapping with the trend intermediate tensor of the k-th round to generate the trend reconstruction tensor of the k-th round. The trend reconstruction tensor of the k-th round serves as the trend recursion tensor of the (k+1)-th round. Specifically, the same-dimensional additive write-back mapping is as follows: the fluctuation accumulation tensor is used to perform trend write-back through linear mapping to generate the fluctuation accumulation write-back tensor; the fluctuation accumulation write-back tensor is added element-wise to the trend intermediate tensor of the kth round to generate the trend reconstruction tensor of the kth round. Repeatedly execute trend recursion, volatility reconstruction, volatility recursion, and trend reconstruction until the end of the Kth round, obtaining the trend reconstruction tensor and volatility recursion tensor for the Kth round. During implementation, to balance the sufficiency of the alternating recursion between the trend risk tensor and volatility risk tensor with network computational complexity, K is set to 3.
[0022] In this embodiment, the basic prediction generation module, the abnormal peak and valley location enhancement module, and the output fusion module specifically include: In the basic forecast generation module, for the trend reconstruction tensor and volatility recursion tensor of the Kth round, linear forecast mapping oriented towards the future operating cycle dimension is performed along the operating cycle dimension to generate the trend forecast tensor and volatility forecast tensor; the trend forecast tensor and volatility forecast tensor are added element by element to generate the basic operating risk forecast tensor. In the abnormal peak and valley positioning enhancement module, the basic operational risk prediction tensor is used to calculate the local change response between adjacent future operational cycles along the future operational cycle dimension, and generate the local change response tensor. Based on the switching relationship of the changing directions of the positions of each element in the local change response tensor, the candidate peak position and candidate valley position are located, specifically as follows: If the response value of the previous local change at the current element position is greater than 0 and the response value of the next local change is less than 0, then the current element position is determined as a candidate peak position. If the response value of the previous local change at the current element position is less than 0 and the response value of the next local change is greater than 0, then the current element position is determined as a candidate valley position. Construct a peak-valley identifier tensor by assigning a value of 1 to the element corresponding to the candidate peak position, a value of -1 to the element corresponding to the candidate valley position, and a value of 0 to the remaining positions. Based on the basic operational risk prediction tensor, local prediction segments with a radius of R are extracted along the future operational cycle dimension, centered on the future operational cycle where the candidate peak or valley position is located, to obtain the local prediction segment tensor. The extraction range of the local prediction segment is the basic prediction segment corresponding to each of the R future operational cycles before and after the current position. When the extraction range corresponding to the segment radius R exceeds the boundary of the future operational cycle, the basic prediction value corresponding to the boundary position is used to fill in the gap. In the implementation process, in order to balance the coverage of the local prediction segment with the information on the changes before and after the peak and valley positions and the focus of the local enhancement processing, R is set to 2.
[0023] The peak-valley identifier tensor and the local prediction segment tensor are concatenated in the feature dimension, and a peak-valley enhanced feature tensor is generated through linear mapping and GELU activation. In the output fusion module, the basic business risk prediction tensor and the peak-valley enhancement feature tensor are aligned in the same dimension and added element by element. Then, linear compression mapping and scalar projection are performed along the feature dimension to obtain the target business risk prediction value corresponding to each future business cycle. The target business risk prediction sequence is generated by arranging the values according to the order of the future business cycles.
[0024] The improved DLinear network in this invention maintains the same overall modeling approach as the standard DLinear network. Both belong to the linear decomposition modeling structure for time series forecasting, and their core purpose is to separate the long-term variation components and short-term fluctuation components in the input sequence before prediction. The standard DLinear network typically first performs sequence decomposition on the input time series to obtain the trend and fluctuation sequences. Then, it performs linear prediction mapping along the time dimension on the trend and fluctuation sequences respectively. Finally, it fuses the two prediction results to obtain the target prediction sequence. Therefore, the basic structure of the standard DLinear network can be summarized as a sequence decomposition structure + a two-branch linear prediction structure + a result fusion structure. The improved DLinear network in this invention also retains this basic framework, still including the processing logic of trend decomposition of the operational risk state tensor, separate modeling of the trend and fluctuation branches, and fusion output of the prediction results. Therefore, it is consistent with the standard DLinear network in terms of network backbone structure and basic prediction idea.
[0025] The difference lies in the fact that the improved DLinear network does not simply rely on the standard DLinear network's direct two-branch prediction structure after decomposition. Instead, it adds a multi-level improved structure tailored to enterprise business risk scenarios. Specifically: First, in the sequence decomposition stage, the improved DLinear network does not directly divide the original input into trend and seasonal components. Instead, it first generates a trend risk tensor based on the business risk state tensor, then performs local offset correction mapping by combining the trend change tensor and the initial residual tensor to obtain the offset-corrected residual tensor. Finally, it generates a volatility risk tensor through sign-preserving compression mapping, thus preserving not only the magnitude but also the direction of change in the volatility component. Second, the improved DLinear network adds a trend-volatility alternating recursive module to the standard DLinear network's two-branch prediction structure. Instead of the trend branch and volatility branch each completing a prediction independently, multiple rounds of alternating recursion create a mutual reconstruction and same-dimensional write-back relationship between the trend recursive tensor and the volatility recursive tensor. Third, the improved DLinear network adds an abnormal peak and valley localization enhancement module after the basic prediction is generated. Through local change response, candidate peak and valley location localization, local prediction segment extraction, and peak and valley enhancement feature generation, it can target abnormal peaks and valleys in the future business cycle. Finally, the output fusion module generates the target business risk prediction sequence.
[0026] Through the aforementioned improvements, the improved DLinear network is better suited for predicting business risks. By employing local offset correction mapping and sign-preserving compression mapping, the network can more accurately distinguish between fundamental trend changes and direction-sensitive fluctuations in the business risk sequence, avoiding the standard DLinear network's insufficient characterization of local offsets and fluctuation directions. Through the trend-fluctuation alternating recursive module, the network can establish multi-round coupling relationships between trend risk and fluctuation risk, enabling more comprehensive modeling of the impact of short-term fluctuations on long-term trends and the constraints of long-term trends on local fluctuations, thereby enhancing the predictive ability for continuously evolving business risks. Through the abnormal peak and valley location enhancement module, the network can explicitly locate and enhance local peaks and valleys in future business cycles, improving the standard DLinear network's handling of local abrupt changes, abnormal transition risks, and large smoothing errors. Therefore, the improved DLinear network in this invention not only improves the accuracy of business risk trend prediction but also enhances the ability to identify abnormal peak and valley changes and key transition cycles, thereby improving the completeness, sensitivity, and practicality of business risk prediction results.
[0027] In this embodiment, step five specifically includes: Calculate the mean and standard deviation of the target business risk prediction series to obtain the mean and standard deviation of business risk; For the target business risk forecast series, the Hurst index is calculated using rescaled range analysis, specifically including: Calculate the difference between the target operating risk forecast and the mean operating risk for each future operating cycle to obtain the mean-adjusted sequence; Based on the mean-adjusted sequence, the cumulative deviation value for each future operating cycle is calculated to obtain the cumulative deviation sequence, and the range of the cumulative deviation sequence is calculated to obtain the range term of the rescaled range analysis method. The ratio of the range term to the standard deviation of operating risk is used as the rescaled range statistic, and a power function relationship is established between the rescaled range statistic and the total number of future operating cycles. By taking the natural logarithm on both sides of the power function relationship, we obtain the inertial fitting formula for operational risk. Based on the operational risk inertia fitting formula, the least squares method is used to perform linear fitting to obtain the operational risk inertia fitting line; the slope of the operational risk inertia fitting line is used as the Hurst exponent of the target operational risk prediction sequence. The target business risk prediction sequence is cyclically shifted according to the future business cycle to generate several locally rearranged reference sequences. The reference Hurst index of each locally rearranged reference sequence is calculated using the rescaled range analysis method. The mean of each reference Hurst index is calculated to obtain the Hurst reference value. If the Hurst index of the target business risk prediction sequence is greater than the Hurst reference value, the target business risk prediction sequence is determined to have risk inertia; otherwise, the target business risk prediction sequence is determined not to have risk inertia, and risk inertia judgment flags are generated respectively. Based on the mean and standard deviation of operating risk, the standard risk value for each future operating cycle is calculated using the Z-score method. The standard risk values for each future operating cycle are sorted according to their numerical magnitude to obtain a standard risk value sequence; the upper quartile and median values of the standard risk value sequence are extracted. Calculate the difference between the upper quartile value and the median value, amplify the difference proportionally, and add it to the upper quartile value to obtain the transition judgment benchmark value; The future operating cycle position where the standard risk value is greater than or equal to the threshold value for determining the transition is used as the candidate abnormal transition cycle. When the target business risk prediction sequence is determined to have risk inertia, each candidate abnormal transition cycle is identified as an abnormal transition cycle. The operational risk assessment result is composed of the risk inertia judgment indicator, the abnormal transition cycle, the target operational risk prediction value corresponding to the abnormal transition cycle, and the standard risk value.
[0028] In this invention, rescaled range analysis is used to calculate the Hurst exponent, and combined with a locally rearranged reference sequence to obtain a Hurst reference value. This allows for the determination of whether the target business risk prediction sequence exhibits risk inertia from the perspective of overall temporal evolution, thereby identifying whether there is a long-term trend of continuous diffusion, accumulation, or fluctuation in business risk. The Z-score method, based on the mean and standard deviation of business risk, standardizes the degree of risk deviation in each future business cycle and identifies candidate abnormal transition cycles by combining them with a transition judgment benchmark value. This allows for the location of key business cycles with sudden increases in risk from the perspective of local fluctuations. By combining the long-term inertia judgment capability of rescaled range analysis with the local anomaly identification capability of the Z-score method, this invention can not only determine whether a company's business risk has continuous evolutionary characteristics, but also further identify the location of sudden risk mutations in future business cycles. This avoids the problems of "only seeing the overall trend while ignoring local anomalies" or "only identifying local anomalies without judging persistent risks" that occur when relying on only a single method, thereby improving the completeness, accuracy, and practicality of business risk judgment results.
[0029] In this embodiment, step six specifically includes: Based on the risk inertia assessment criteria, determine whether the target business risk prediction sequence has a continuous evolution trend; If the target business risk prediction sequence does not have a continuous evolution trend and there is no abnormal transition cycle, then the business risk level of the target enterprise is determined to be low risk. If the target business risk prediction sequence has a continuous evolution trend and there is no abnormal transition cycle, then the business risk level of the target enterprise is determined to be medium risk. If the target business risk prediction sequence does not have a continuous evolution trend and there is an abnormal transition cycle, then the business risk level of the target enterprise will be determined as medium risk. If the target business risk prediction sequence has a continuous evolution trend and there is an abnormal transition cycle, then the business risk level of the target enterprise will be determined as high risk.
[0030] In this embodiment, the operational adjustment suggestions specifically include: If the operational risk level is low, maintain the current operational arrangements and continuously monitor changes in operational risk. If the operational risk level is medium risk, then strengthen the monitoring of abnormal transition cycles and check for abnormal changes in operational data, performance status data, capital status data, inventory status data, supply status data, or external disturbance data. If the operational risk level is high, the operational arrangements will be adjusted, key operational matters corresponding to the abnormal transition cycle will be prioritized for verification, and abnormal changes in financial status data, inventory status data, supply status data, or contract performance data will be prioritized for handling.
[0031] Example 1: To verify the feasibility of this invention in practice, the method of this invention was applied to a business risk prediction scenario of an equipment manufacturing enterprise. This enterprise simultaneously undertakes the supply of standard parts, processing of customized components, and after-sales support. In its daily operations, it faces problems such as frequent order fluctuations, long delivery cycles, raw material inventory significantly affected by upstream supply, unstable payment cycles from some customers, and occasional logistical anomalies. Previously, the enterprise mainly relied on manual judgment by finance personnel and the planning department based on monthly reports, inventory reports, and supplier performance records. While this could identify some explicit operational anomalies, it often failed to promptly identify the risk inertia accumulated over multiple operating cycles, as well as potential abnormal leaps in future operating cycles. This resulted in delayed procurement adjustments, increased inventory holdings, and expanded order delays, making it difficult for management personnel to effectively intervene before risks escalated.
[0032] In the implementation of this invention, the company's business ledger for 18 consecutive months was selected as the sample data source, with a week as the operating cycle unit. The business ledger includes order amount, order quantity, business change records, agreed delivery time, actual delivery time, default records, operating revenue amount, operating expenditure amount, cash flow change value, inventory quantity, inventory turnover cycle, stockout records, supply frequency, supply interval duration, supply interruption records, market price fluctuation records, policy change records, and logistics anomaly event records. First, based on the operating cycle identifier and the business item association identifier, the operating records in the company's business ledger are retrieved and extracted to obtain operating risk-related data. Subsequently, the operating risk-related data is sliced according to the operating cycle identifier, and then arranged into slots according to the business item association identifier to form a set of operating status units. This allows the business data, performance status data, capital status data, inventory status data, supply status data, and external disturbance data, which were originally scattered in different forms, to be mapped to a unified slot structure within the same operating cycle. Subsequently, the state vector arrangement and slot return input mapping process are performed on the set of operating state units to generate an operating risk state tensor, so that the state changes between different slots have a continuous and comparable tensor expression in the time dimension and slot structure dimension.
[0033] In the risk prediction phase, the operational risk state tensor is input into the improved DLinear network. The network first extracts trend risk and volatility risk tensors through a sequence decomposition module, then alternately reconstructs long-term evolution information and short-term disturbance information through a trend-volatility alternating recursive module, enabling the transmission between persistent operational deterioration trends and local fluctuations. Subsequently, the basic prediction generation module outputs the basic operational risk prediction tensor, and the abnormal peak and valley location enhancement module further identifies the locations of local peaks and valleys and enhances abnormal peaks and valleys. Finally, the output fusion module obtains the target operational risk prediction sequence. Based on this target operational risk prediction sequence, the Hurst index is used to perform risk inertia judgment, and the Z-score method is used to identify abnormal transition cycles, obtaining the operational risk judgment result. Finally, the operational risk level is determined based on the operational risk judgment result, and corresponding operational adjustment suggestions are output to the risk management terminal for timely viewing and processing by operational managers.
[0034] To demonstrate the problem solved by this invention, the actual operating results of the enterprise in the following 12 weeks were selected as the verification period. This invention was then compared with the enterprise's existing manual assessment methods, ordinary linear time series forecasting methods, and the standard DLinear method. The comparison results are shown in Table 1.
[0035] Table 1. Comparison of the effectiveness of different methods in enterprise business risk prediction scenarios.
[0036] As shown in Table 1, the method of this invention outperforms manual assessment methods, ordinary linear time series forecasting methods, and the standard DLinear method in all comparative indicators. Specifically, the risk prediction accuracy of this invention reaches 91.7%, which is 19.1 percentage points higher than manual assessment methods, 11.9 percentage points higher than ordinary linear time series forecasting methods, and 5.8 percentage points higher than the standard DLinear method. This indicates that this invention can more accurately characterize the overall trend of changes in enterprise operational risks and improve the ability to predict future operational risk levels. The high-risk identification rate reaches 89.6%, and the accuracy rate of identifying abnormal transition cycles reaches 86.3%, indicating that this invention can not only more accurately identify the overall operational risk level but also more effectively identify key abnormal change cycles. Furthermore, the false alarm rate of this invention is reduced to 7.4%, significantly lower than the comparative methods, indicating that this invention can effectively reduce invalid warnings while improving identification capabilities. The average warning lead time reaches 3.3 weeks, indicating that this invention can detect operational risks earlier, allowing enterprises more time to adjust. The recall rates for inventory anomaly warnings and fulfillment anomaly warnings reached 88.9% and 90.4% respectively, further demonstrating that this invention has a stronger ability to perceive key operational issues such as inventory fluctuations and fulfillment anomalies. The adoption rate of operational adjustment suggestions reached 86.8%, indicating that the operational adjustment suggestions output by this invention are more aligned with the actual operational needs of enterprises and have good application value and implementation effectiveness.
[0037] The above description is only a preferred embodiment of the present invention, but 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 scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.
Claims
1. A method for predicting business risks based on data mining, characterized in that, Includes the following steps: Step 1: Based on the target company's operating cycle identifier and operating event association identifier, perform retrieval and extraction processing on the operating records in the company's operating ledger to obtain operating risk association data; Step 2: Divide the operational risk-related data into periodic slices according to the operational cycle identifier, and arrange them into slots according to the operational item-related identifier to obtain the operational status unit set; Step 3: Based on the set of operational state units, generate an operational risk state tensor through state vector organization and slot return input mapping; Step 4: Input the operational risk state tensor into the improved DLinear network, perform trend decomposition, alternating recursion and peak-valley enhancement prediction processing to obtain the target operational risk prediction sequence; the improved DLinear network includes a sequence decomposition module, a trend-fluctuation alternating recursion module, a basic prediction generation module, an abnormal peak-valley location enhancement module and an output fusion module; Step 5: Based on the target business risk prediction sequence, risk inertia judgment is performed using the rescaled range analysis method, and abnormal transition cycles are identified using the Z-score method to obtain the business risk judgment result; Step Six: Based on the operational risk assessment results, determine the operational risk level of the target company; Step 7: Based on the operational risk level, generate operational adjustment suggestions and output the operational risk assessment results, operational risk level, and operational adjustment suggestions to the risk management terminal.
2. The enterprise business risk prediction method based on data mining according to claim 1, characterized in that, The operational risk-related data includes operational business data, contract performance data, financial status data, inventory status data, supply status data, and external disturbance data.
3. The enterprise business risk prediction method based on data mining according to claim 1, characterized in that, Step two specifically includes: Extract the operating cycle identifier from each operating record and determine the operating cycle boundary according to the operating cycle identifier; Based on the boundaries of the operating cycle, the operating risk-related data is segmented into several operating cycle data segments. The data segments of each operating cycle are processed sequentially according to the order in which they are identified by the operating cycle identifiers. Within each business cycle data segment, extract the business item association identifiers for each business record, and merge the business records according to the business item association identifiers to obtain multiple item record groups; Within each data segment of an operating cycle, a slot structure is established, which includes operating business slots, fulfillment slots, capital slots, inventory slots, supply slots, and external disturbance slots; The business records in each item record group are arranged in slots according to data categories: business data is arranged in the business slot, performance status data is arranged in the performance slot, funds status data is arranged in the funds slot, inventory status data is arranged in the inventory slot, supply status data is arranged in the supply slot, and external disturbance data is arranged in the external disturbance slot. For multiple business records grouped into the same slot within the same item record group, write them sequentially into the corresponding slot according to the order in which the multiple business records are arranged in the corresponding business cycle data segment; For slots without operational records, retain the corresponding slot location and record it as an empty slot. The slot structure in the same operating cycle data segment is combined into units according to the slot order to generate an operating status unit; Arrange the operating status units of each operating cycle according to the operating cycle identifier to obtain the operating status unit set.
4. The enterprise business risk prediction method based on data mining according to claim 1, characterized in that, Step three specifically includes: Based on the set of operating status units, the record sequence of each slot in each operating status unit is processed into a status vector, and slot alignment is performed according to the slot order to obtain a standard slot arrangement vector sequence. The standard slot arrangement vector sequence is processed using slot-folded input mapping: The standard slot arrangement vector sequence is mapped to the positive slot through linear mapping to obtain the positive slot feature vector sequence; The standard slot arrangement vector sequence is rearranged in reverse order according to the slot order, and the slot reversal mapping is performed through linear mapping to obtain the reversal slot feature vector sequence. The forward slot feature vector sequence and the foldback slot feature vector sequence are bidirectionally interleaved and fused to obtain the slot fused feature vector sequence. The slot fusion feature vector sequence is compressed in dimension and linearly mapped to obtain the standard state feature vector sequence of the corresponding business state unit. The standard state feature vector sequence of each operating state unit is stacked into tensors according to the operating cycle identifier order to generate an operating risk state tensor. The structure of the operating risk state tensor includes an operating cycle dimension, a slot structure dimension, and a feature dimension.
5. The enterprise business risk prediction method based on data mining according to claim 1, characterized in that, The sequence decomposition module and the trend-fluctuation alternating recursion module specifically include: In the sequence decomposition module, for each slot structure dimension and feature dimension, a moving average operation is performed on the operating risk state tensor along the operating cycle dimension to generate a trend risk tensor. Calculate the initial residual tensor based on the operational risk state tensor and the trend risk tensor; For the trend risk tensor, the trend change between adjacent operating cycles is calculated along the operating cycle dimension to generate the trend change tensor. The initial residual tensor and the trend change tensor are concatenated in the feature dimension to obtain the correction input tensor. Then, a local offset correction mapping is performed through linear mapping and GELU activation to generate a local offset correction tensor. The initial residual tensor and the local offset correction tensor are interpolated element-wise to obtain the offset correction residual tensor; the offset correction residual tensor is then subjected to sign-preserving compression mapping to generate the volatility risk tensor. The sign-preserving compression mapping specifically involves: extracting the direction identifier value for each residual element in the offset-corrected residual tensor using the sign function; performing a natural logarithmic compression operation after incrementing the absolute value of each residual element by 1 to obtain the compression amplitude; multiplying each compression amplitude by the corresponding direction identifier value to obtain the fluctuation state value of each residual element; and reorganizing each fluctuation state value according to its original position in the offset-corrected residual tensor to generate a fluctuation risk tensor. In the trend-volatility alternating recursion module, K rounds of alternating recursion are performed on the trend risk tensor and volatility risk tensor, specifically: The trend risk tensor is denoted as the initial trend recursion tensor, and the volatility risk tensor is denoted as the initial volatility recursion tensor. For the alternating recursion of the kth round, the trend recursion tensor of the (k-1)th round is mapped along the operating cycle dimension through a linear mapping to generate the intermediate trend tensor of the kth round. The trend intermediate tensor of the kth round and the fluctuation recursion tensor of the (k-1)th round are concatenated in the feature dimension, and fluctuation reconstruction is performed through linear mapping and GELU activation to generate the fluctuation reconstruction tensor of the kth round. For the volatility reconstruction tensor of the k-th round, calculate the difference tensor between adjacent operating cycles along the operating cycle dimension, and perform linear compression mapping on the difference tensor to generate the volatility recursive tensor of the k-th round. The fluctuation recursion tensor of the k-th round is cumulatively summed along the operating cycle dimension to generate the fluctuation accumulation tensor. The fluctuation accumulation tensor and the trend intermediate tensor of the k-th round are then subjected to the same-dimensional additive write-back mapping to generate the trend reconstruction tensor of the k-th round. The trend reconstruction tensor of the k-th round is used as the trend recursion tensor of the (k+1)-th round. Repeatedly execute trend recursion, fluctuation reconstruction, fluctuation recursion and trend reconstruction until the end of the Kth round, to obtain the trend reconstruction tensor and fluctuation recursion tensor of the Kth round.
6. The enterprise business risk prediction method based on data mining according to claim 1, characterized in that, The basic prediction generation module, the abnormal peak and valley localization enhancement module, and the output fusion module specifically include: In the basic forecast generation module, for the trend reconstruction tensor and volatility recursion tensor of the Kth round, linear forecast mapping oriented towards the future operating cycle dimension is performed along the operating cycle dimension to generate the trend forecast tensor and volatility forecast tensor; the trend forecast tensor and volatility forecast tensor are added element by element to generate the basic operating risk forecast tensor. In the abnormal peak and valley positioning enhancement module, the basic operational risk prediction tensor is used to calculate the local change response between adjacent future operational cycles along the future operational cycle dimension, and generate the local change response tensor. Based on the switching relationship of the changing directions of the positions of each element in the local change response tensor, the candidate peak position and candidate valley position are located, specifically as follows: If the response value of the previous local change at the current element position is greater than 0 and the response value of the next local change is less than 0, then the current element position is determined as a candidate peak position. If the response value of the previous local change at the current element position is less than 0 and the response value of the next local change is greater than 0, then the current element position is determined as a candidate valley position. Construct a peak-valley identifier tensor by assigning a value of 1 to the element corresponding to the candidate peak position, a value of -1 to the element corresponding to the candidate valley position, and a value of 0 to the remaining positions. Based on the basic operational risk prediction tensor, local prediction segments with a radius of R are extracted along the future operational cycle, centered on the future operational cycle where the candidate peak or candidate valley position is located, to obtain the local prediction segment tensor. The peak-valley identifier tensor and the local prediction segment tensor are concatenated in the feature dimension, and a peak-valley enhanced feature tensor is generated through linear mapping and GELU activation. In the output fusion module, the basic business risk prediction tensor and the peak-valley enhancement feature tensor are aligned in the same dimension and added element by element. Then, linear compression mapping and scalar projection are performed along the feature dimension to obtain the target business risk prediction value corresponding to each future business cycle. The target business risk prediction sequence is generated by arranging the values according to the order of the future business cycles.
7. The enterprise business risk prediction method based on data mining according to claim 1, characterized in that, Step five specifically includes: Calculate the mean and standard deviation of the target business risk prediction series to obtain the mean and standard deviation of business risk; For the target business risk forecast series, the Hurst index is calculated using rescaled range analysis, specifically including: Calculate the difference between the target operating risk forecast and the mean operating risk for each future operating cycle to obtain the mean-adjusted sequence; Based on the mean-adjusted sequence, the cumulative deviation value for each future operating cycle is calculated to obtain the cumulative deviation sequence, and the range of the cumulative deviation sequence is calculated to obtain the range term of the rescaled range analysis method. The ratio of the range term to the standard deviation of operating risk is used as the rescaled range statistic, and a power function relationship is established between the rescaled range statistic and the total number of future operating cycles. By taking the natural logarithm on both sides of the power function relationship, we obtain the inertial fitting formula for operational risk. Based on the operational risk inertia fitting formula, the least squares method is used to perform linear fitting to obtain the operational risk inertia fitting line; the slope of the operational risk inertia fitting line is used as the Hurst exponent of the target operational risk prediction sequence. The target business risk prediction sequence is cyclically shifted according to the future business cycle to generate several locally rearranged reference sequences. The reference Hurst index of each locally rearranged reference sequence is calculated using the rescaled range analysis method. The mean of each reference Hurst index is calculated to obtain the Hurst reference value. If the Hurst index of the target business risk prediction sequence is greater than the Hurst reference value, the target business risk prediction sequence is determined to have risk inertia; otherwise, the target business risk prediction sequence is determined not to have risk inertia, and risk inertia judgment flags are generated respectively. Based on the mean and standard deviation of operating risk, the standard risk value for each future operating cycle is calculated using the Z-score method. The standard risk values for each future operating cycle are sorted according to their numerical magnitude to obtain a standard risk value sequence; the upper quartile and median values of the standard risk value sequence are extracted. Calculate the difference between the upper quartile value and the median value, amplify the difference proportionally, and add it to the upper quartile value to obtain the transition judgment benchmark value; The future operating cycle position where the standard risk value is greater than or equal to the threshold value for determining the transition is used as the candidate abnormal transition cycle. When the target business risk prediction sequence is determined to have risk inertia, each candidate abnormal transition cycle is identified as an abnormal transition cycle. The operational risk assessment result is composed of the risk inertia judgment indicator, the abnormal transition cycle, the target operational risk prediction value corresponding to the abnormal transition cycle, and the standard risk value.
8. The enterprise business risk prediction method based on data mining according to claim 1, characterized in that, Step six specifically includes: Based on the risk inertia assessment criteria, determine whether the target business risk prediction sequence has a continuous evolution trend; If the target business risk prediction sequence does not have a continuous evolution trend and there is no abnormal transition cycle, then the business risk level of the target enterprise is determined to be low risk. If the target business risk prediction sequence has a continuous evolution trend and there is no abnormal transition cycle, then the business risk level of the target enterprise is determined to be medium risk. If the target business risk prediction sequence does not have a continuous evolution trend and there is an abnormal transition cycle, then the business risk level of the target enterprise will be determined as medium risk. If the target business risk prediction sequence has a continuous evolution trend and there is an abnormal transition cycle, then the business risk level of the target enterprise will be determined as high risk.
9. The enterprise business risk prediction method based on data mining according to claim 1, characterized in that, The proposed operational adjustment measures specifically include: If the operational risk level is low, maintain the current operational arrangements and continuously monitor changes in operational risk. If the operational risk level is medium risk, then strengthen the monitoring of abnormal transition cycles and check for abnormal changes in operational data, performance status data, capital status data, inventory status data, supply status data, or external disturbance data. If the operational risk level is high, the operational arrangements will be adjusted, key operational matters corresponding to the abnormal transition cycle will be prioritized for verification, and abnormal changes in financial status data, inventory status data, supply status data, or contract performance data will be prioritized for handling.