Big data-driven enterprise operation credit trend prediction method and system

Abstract

The invention relates to the technical field of enterprise credit risk prediction and big data risk control, in particular to a big data-driven enterprise operation credit trend prediction method and system, and the method comprises the following steps: a digital twinborn construction module integrates the operation data of an enterprise group into a twinborn snapshot; a wavelet manifold extraction module performs multi-band splitting on time sequence indexes in the twin snapshots and generates time-frequency manifold feature vectors in combination with manifold or persistent coherence analysis; the chaotic pulse training module deepens perception of extreme situations through chaotic mapping, and outputs enterprise risk state vectors to the Holonic multi-agent for grading and scoring; the quantum search decision module searches and executes various linkage schemes in parallel when the risk is high or abnormal; and the increment updating module quickly corrects the network weight in a small-batch online training mode after the default occurs, so that the risk vector adapts to the latest operating condition, and the accuracy and timeliness of enterprise credit risk prediction are remarkably improved.

Description

Technical Field

[0001] The present invention relates to the field of enterprise credit risk prediction and big data risk control technology, and in particular to a big data-driven enterprise business credit trend prediction method and system. Background Art

[0002] In the context of big data, massive amounts of financial, transactional, and supply chain data are generated every second during business operations. If they cannot be integrated and dynamically trended in a timely manner, it is very easy to miss opportunities for major risk warnings. When facing investment, financing, or upstream and downstream cooperation, if corporate credit risks are not discovered early, it may cause chain reactions such as large-scale defaults or supply chain disruptions. Although existing technologies use methods such as static regression and regular batch training (Chinese invention patent, publication number: CN118153939A, name: Prediction method, device, electronic device, and storage medium for corporate credit risk), which have a certain degree of accuracy in normal operating environments, they have obvious limitations when dealing with frequently iterated operating data and capturing extreme abnormal events.

[0003] On the one hand, the lack of adaptive updating for emerging risk events after a single training session causes the model to be out of sync with real-world business dynamics. On the other hand, the use of general statistical features, simple clustering, or traditional regression algorithms often struggles to capture hidden correlations or strong fluctuation signals in high-dimensional space. As a result, existing methods are prone to underreporting or lags in real-time identification and early prevention of corporate credit risk, creating potential risks for overall corporate credit management and financial institutions' risk control. Summary of the Invention

[0004] To address the numerous issues with the aforementioned existing technologies, the present invention provides a big data-driven method and system for predicting corporate credit trends. This method first uses a digital twin model to generate a snapshot of corporate operational data. Wavelet and manifold coherence are then used to extract multi-band and high-dimensional structural features. These features are then trained using a chaotic pulse network to output a risk vector. If the risk score is high, a quantum-inspired search is used to select the optimal strategy from multiple strategies in parallel. After a default is confirmed, the network weights can be adjusted using an incremental update mechanism. This method comprehensively captures multi-scale anomalies and dynamic changes, shortening decision latency and significantly improving the accuracy of corporate credit predictions.

[0005] A big data-driven enterprise credit trend prediction method includes the following steps:

[0006] Build a digital twin model in a computing environment, import the operational data of the enterprise group and generate a digital twin snapshot;

[0007] Performing wavelet analysis on the time series indicators in the digital twin snapshot to extract multi-band features; combining manifold or persistent homology analysis to obtain high-dimensional distribution information, and fusing the obtained multi-band features with the high-dimensional distribution information to generate a time-frequency manifold feature vector;

[0008] The time-frequency manifold feature vector is trained using a chaotic pulse neural network, and a chaotic map is configured in the network's hidden layer to output an enterprise risk state vector. The risk state vector is then assigned to a Holonic multi-agent environment to calculate a risk score. If the risk score exceeds a threshold or an operational anomaly is detected, a quantum heuristic search is performed in parallel and a linkage strategy is issued.

[0009] When a default is confirmed or a serious risk event occurs, the corresponding information is transmitted back to the chaotic pulse neural network for incremental updates based on the latest records. Through the digital twin model, the enterprise risk state vector can adapt to new risk conditions in a dynamic environment.

[0010] Preferably, the step of performing wavelet analysis on the time series indicators includes dividing the enterprise operation data into at least two scale intervals, extracting frequency band coefficients using multi-channel discrete wavelet transform, and obtaining multi-band features after statistically fusing the frequency band coefficients.

[0011] Preferably, the step of obtaining high-dimensional distribution information in combination with manifold or persistent homology analysis includes constructing a high-dimensional complex structure according to preset adjacency rules, and forming a topological vector by calculating zero-dimensional and one-dimensional homology elements, and splicing the topological vector and the multi-band feature into a time-frequency manifold feature vector.

[0012] Preferably, the process of configuring a chaotic map in the hidden layer of the network includes applying a dynamic perturbation to the excitation threshold of the pulse neural network using a chaotic sequence, so that the network can more easily identify extreme risk states during training.

[0013] Preferably, the Holonic multi-agent environment is composed of enterprise agents and superior agents. The enterprise agents simulate and score the operation data according to the enterprise risk state vector, and the superior agents summarize and analyze the scoring results of the enterprise agents.

[0014] Preferably, the threshold is used to determine whether abnormal fluctuations occur in the enterprise operation indicators in consecutive time periods. After the scoring result exceeds the threshold, the superior agent executes the linkage decision-making process.

[0015] Preferably, the quantum heuristic search defines at least two initial superposition states, attenuates and amplifies each candidate solution in the probability amplitude space, and selects a linkage strategy that meets the target conditions after reaching a preset number of iterations.

[0016] Preferably, the linkage strategy includes issuing temporary limits, working capital adjustments or upstream and downstream collaborative management measures to corporate agents, and further verification by the superior agent based on execution feedback.

[0017] Preferably, a small-batch online training method is used based on the incremental update of the latest records, the samples of defaulting enterprises are incorporated into the learning process of the chaotic pulse neural network, and the subsequent operating conditions are simulated through the digital twin model to correct the risk state vector.

[0018] A big data-driven enterprise credit trend prediction system for implementing the method, characterized in that the system includes:

[0019] A digital twin building module is used to build a digital twin model in a computing environment and import the operational data of the enterprise group to generate a digital twin snapshot;

[0020] A wavelet manifold extraction module is used to perform wavelet analysis on the time series indicators in the digital twin snapshot to extract multi-band features, and combine manifold or persistent homology analysis to obtain high-dimensional distribution information, and fuse the multi-band features with the high-dimensional distribution information to form a time-frequency manifold feature vector;

[0021] a chaotic pulse training module for training the time-frequency manifold feature vector using a chaotic pulse neural network, outputting an enterprise risk state vector by configuring a chaotic map in a hidden layer of the network, and assigning the risk state vector to a Holonic multi-agent environment to calculate a risk score;

[0022] The quantum search decision module is used to conduct quantum heuristic search and issue linkage strategies to enterprise agents according to parallel schemes when the risk score exceeds the threshold or an operational anomaly is detected;

[0023] The incremental update module is used to transmit the corresponding information back to the chaotic pulse neural network after confirming a default or a serious risk event, and perform incremental updates based on the latest records through the digital twin model, so that the enterprise risk state vector can remain adaptable to new risk conditions in a dynamic environment.

[0024] Compared with the prior art, the advantages and beneficial effects of the present invention are:

[0025] This invention uses a digital twin model to mirror high-frequency operational data, achieving global control of the enterprise's real-time or quasi-real-time operating status, and solving the problem of the existing technology being unable to dynamically update enterprise data.

[0026] The present invention combines wavelet manifolds with chaotic pulse training to achieve accurate identification of multi-scale fluctuations and extreme risk events, overcoming the limitations of existing technologies that only use a single regression model and are difficult to adapt to a small number of anomalies.

[0027] This invention uses quantum heuristic search to perform parallel evaluation of multiple strategies, significantly reducing decision-making time and providing more efficient optimal solutions for risk linkage strategies.

[0028] The present invention introduces small-batch online training based on the latest records through an incremental update mechanism. After a default is confirmed, newly emerging high-risk samples can be integrated into the chaotic pulse network to maintain continuous perception of the dynamic risk environment, which is superior to the static model method of the existing technology. BRIEF DESCRIPTION OF THE DRAWINGS

[0029] Figure 1 Schematic diagram of the method flow of the present invention;

[0030] Figure 2 Schematic diagram of the interaction between the digital twin environment and enterprise operation data in the present invention;

[0031] Figure 3 Schematic diagram of the fusion of wavelet analysis and high-dimensional manifold information in the present invention;

[0032] Figure 4 Schematic diagram of the distributed scoring of chaotic pulse neural network and Holonic multi-agent in the present invention;

[0033] Figure 5 A schematic diagram of the multi-strategy parallel quantum heuristic search in the present invention;

[0034] Figure 6 It is a schematic block diagram of the system structure of the present invention. DETAILED DESCRIPTION

[0035] Hereinafter, embodiments of the present disclosure will be described with reference to the accompanying drawings. However, it should be understood that these descriptions are merely exemplary and are not intended to limit the scope of the present disclosure. In the following detailed description, for ease of explanation, many specific details are set forth to provide a comprehensive understanding of the embodiments of the present disclosure.

[0036] like Figure 1 As shown in FIG, a big data-driven enterprise credit trend prediction method includes the following steps:

[0037] Build a digital twin model in a computing environment, import the operational data of the enterprise group and generate a digital twin snapshot;

[0038] like Figure 2As shown, in the present invention, the digital twin model is used to establish a dynamic mirror of the real operating behavior of the enterprise. By periodically generating "twin snapshots", a high-frequency state set in time sequence is formed, so that the business data of the enterprise in different time periods are comparable, traceable and parallelizable. The model is deployed on a programmable high-performance computing platform (such as a container cluster based on Kubernetes), supporting concurrent data access, task scheduling and structured snapshot generation. The twin snapshot created by the system in each time period t is denoted as S t , its core data set X t Represents the key operating indicators of the enterprise group during the period.

[0039] In practice, we set the total number of enterprises to N and divide them into k channels for parallel processing based on resource scale and scheduling load, where 2≤k≤64. Each channel runs data cleaning tasks using container instances, performing type standardization and missing value imputation on fields such as revenue, cash flow, and inventory.

[0040] When the number of parallel channels is set to k = 16, in tests conducted on a compute node equipped with an 8-core CPU and 16GB of memory, a single round of aggregation completes consistently within 10 minutes, supporting data inflows from N = 50 to 200N enterprises. In a stress test where the enterprise size doubled, increasing k to 32 maintained the system's ability to complete aggregation within 15 minutes, demonstrating excellent scalability.

[0041] Each twin snapshot S t Contains operational indicator set X t , defined as follows:

[0042] X t ={R t , F t ,I t ,…}

[0043] Where: R t represents the total income of the enterprise group in time t; F t I represents the net cash flow of the enterprise group in time t; t Represents the total inventory of the enterprise group at time t.

[0044] All fields are collected by the system from the corresponding business systems of each enterprise. When the data is written into the twin snapshot, it is hierarchically organized by business domain (finance, logistics, contract, etc.). If the fluctuation range of the newly added data batch exceeds the set threshold within the snapshot period (such as the growth rate of the number of data items exceeds 2 times), the system automatically triggers the intermediate snapshot. Generate to ensure that the risk monitoring granularity is fine enough.

[0045] Through the present invention, the snapshot generation delay is shortened by about 60%, from an average of 25 minutes to 10 minutes; the data extraction bandwidth utilization is improved by about 25%; the snapshot loss rate of the system is controlled to 0 under peak load, and it has industrial-level stability.

[0046] Twin snapshots serve as a unified input interface for wavelet analysis, manifold modeling, and chaotic pulse neural network training, significantly reducing subsequent processing initialization and caching time. In actual deployments, the transition time from snapshot generation to wavelet feature extraction has been reduced from three minutes in the original solution to less than 30 seconds with this method.

[0047] Example 1 (Basic): Deploy a concurrent data import container in a cluster with 8 CPU cores and 16GB of RAM, set k = 16, and generate twin snapshots every 15 minutes. The system processes 20,000 data records per second, and the snapshots are directly used by the wavelet analysis module. The prediction model is trained using ReLU activation, 128 hidden nodes, and a learning rate of α = 0.01. The loss converges within 20 rounds, and the prediction accuracy reaches 91.3%.

[0048] Example 2 (Extended): Based on Example 1, the "burst update mechanism" of the twin snapshot is enabled, that is, when the growth rate of three consecutive indicators is greater than 20%, a twin snapshot is generated in advance. This mechanism improves the ability to capture sudden changes in cash flow, increases warning time by 12%, and reduces the delay standard deviation by 35%, but increases memory overhead by approximately 18%.

[0049] like Figure 3 As shown, wavelet analysis is performed on the time series indicators in the digital twin snapshot to extract multi-band features; high-dimensional distribution information is obtained by combining manifold or persistent homology analysis, and the obtained multi-band features are fused with the high-dimensional distribution information to generate a time-frequency manifold feature vector;

[0050] This method addresses the fluctuations in enterprise operational data across different time scales, as well as the hidden connectivity or ring structures between enterprises. It first uses discrete wavelet transforms to decompose time series indicators into multiple frequency bands. It then uses manifold or persistent homology analysis to extract high-dimensional topological features. Finally, the two are combined to generate a time-frequency manifold feature vector. Compared to single time series analysis methods, this method is more capable of identifying rapid fluctuations, localized clustering, or potential chain reactions.

[0051] Wavelet analysis, let: X t ={x1,x2,…,x n} represents a time series observation (such as cash flow or inventory sequence) in time period t, and a discrete wavelet basis (such as "db4") is used to transform X t Perform J-layer decomposition to obtain multiple sets of frequency band coefficients. These coefficients are fused at the statistical level to obtain the multi-band feature M tIf J=4, the fluctuation information from high frequency to lower frequency range can be captured respectively.

[0052] Manifold or persistent homology analysis: In the same snapshot, if there are N companies, each with multi-dimensional attributes (financial, supply chain, etc.), an adjacency rule can be established: if the distance d(A, B) between any two companies A and B is less than a threshold θ, then A and B are connected into a high-dimensional complex structure. The generation and extinction time of zero-dimensional and one-dimensional elements is calculated through the persistent homology algorithm and recorded as a topological vector T t In the test, θ is set in the interval 0.01≤θ≤0.05 to strike a balance between ensuring the sensitivity of outlier detection and network connectivity.

[0053] After completing the wavelet feature M t With the topological vector T t Then, concatenate it into:

[0054] V t =[M t ||T t ]

[0055] It is called the time-frequency manifold eigenvector. t The chaotic pulse neural network is used for processing, which enables it to have a higher recognition rate for short-period oscillations or localized cluster anomalies.

[0056] Preferably, the step of performing wavelet analysis on the time series indicators includes dividing the enterprise operation data into at least two scale intervals, extracting frequency band coefficients using multi-channel discrete wavelet transform, and obtaining multi-band features after statistically fusing the frequency band coefficients.

[0057] like Figure 4 As shown, the present invention aims at the fact that business operating data often have multi-level fluctuations in different time periods (such as daily cash flow, weekly order volume, quarterly financial settlement, etc.). It is difficult to accurately capture short-term drastic changes or periodic characteristics by relying solely on a single scale or original time domain indicators. In order to improve the sensitivity of risk prediction, wavelet analysis is used here. First, the time series data is divided into several scale intervals according to different time lengths or different periods. Then, a discrete wavelet transform is performed on each interval, and its frequency band coefficients are extracted. Statistical fusion is performed to obtain the final multi-band characteristics. Subsequently, it can be combined with other analysis modules (such as manifold analysis, chaotic pulse neural network, etc.) to perform more refined identification of business operating risks in multi-scale time dimensions.

[0058] In practical applications, data partitioning and multi-channel discrete wavelet are used to segment the collected enterprise operation time series data (such as revenue, number of transactions, inventory, etc.) into at least two scale intervals after preliminary cleaning to reflect the fluctuations of long cycles (such as one week, one month) and short cycles (such as one day, half a day) or smaller granularity. The time series data sequence is recorded as:

[0059] X={x1,x2,…,x n}

[0060] where x i represents the i-th observation value (such as net capital inflow, order increment, etc.), and n is the length of the time series.

[0061] For each segmented interval, the discrete wavelet transform method is used to layer the sequence according to the "multi-channel" idea. That is, several wavelet basis function operations are performed in each interval to extract detail coefficients and approximate coefficients of different frequency bands. If "wavelet basis = db4, number of layers = 4" is used, a set of 4 layers of coefficients can be obtained, which decreases from high frequency to low frequency, and is recorded as:

[0062]

[0063] Where W j Represents the set of wavelet coefficients at layer j. If more precise detection of short-period fluctuations is required, the number of layers can be increased to 6 or 8. However, testing has found that exceeding 8 layers significantly increases the computational complexity and increases sensitivity to low-amplitude noise.

[0064] Statistical fusion of frequency band coefficients: quantify or count the coefficients of each layer to form the required multi-band features. Specific methods include:

[0065] (a) Calculate the mean, variance, and maximum value of each layer's coefficients;

[0066] (b) If it is necessary to distinguish special mutations, the proportion of abnormal points can be additionally calculated in the high-level coefficients;

[0067] (c) Combine the statistics of each layer in the same interval into vector M k , and then perform the same operation on multiple intervals (at least two) to obtain {M1,M2,…}. The resulting multi-band features can be subsequently aggregated or reduced in dimension to form the input for chaotic networks or quantum searches.

[0068] In segmented parallel wavelet mode, the system can simultaneously capture intraday, weekly, and even longer-term fluctuations in business operating data. By setting different levels and basis functions, it can identify more granular patterns, such as half-day fluctuations or large weekly purchases. Compared to single-scale wavelet methods, it significantly improves accuracy for both short-term, high-intensity fluctuations and longer, steady, slow increases and decreases, with up to a 3% increase in accuracy.

[0069] Tests have shown that when the stratification setting is 4 layers (db4) and the target is 1.5×10 7When concurrently processing enterprise operational data for a group of records, each aggregation takes less than 10 minutes, seamlessly integrating with the digital twin snapshot relay module. Compared to traditional multi-stage Fourier analysis, wavelet analysis can reduce computational complexity by 15-20% through localized processing, while maintaining minimal peak memory requirements under parallel scheduling.

[0070] In real-world industrial scenarios, corporate financial or order data often contains some noise (such as sudden product returns and temporary payment delays). Multi-channel discrete wavelet analysis has been shown to be less sensitive to noise when using 2 to 4 layers, with a signal principal component retention rate exceeding 95%, maintaining the ability to capture true cyclical changes.

[0071] Basic implementation example: Deployed on a set of nodes with 8-core CPUs and 16GB of memory, the system can process about 1.5×10 7 After importing the data into the digital twin snapshot through parallel channels, the time series indicators of each enterprise are divided into two scales (daily and half-day), and a four-layer discrete wavelet transform (db4) is performed in each scale. The mean and variance of each layer coefficient are calculated and then spliced into a vector M k Finally, {M1,M2} is fused into the system's multi-band feature matrix. Compared to the control scheme with only a single scale, the accuracy of identifying daily short-term sales peaks increased by 2%, and the delay standard deviation decreased by approximately 25%.

[0072] Extended implementation example: Based on the above, the number of layers is expanded from 4 to 6 and a third scale interval (weekly data) is added to make the capture of long-term trends more sensitive. However, the test found that the amount of calculation increased by 20%, and additional container instances were needed to maintain a 15-minute merging cycle. If the learning rate α is <0.001, the convergence speed of network training decreases by 50%. If α>0.1, the loss oscillates under high-level wavelet features, indicating that the critical value test is consistently effective for parallel wavelet processes. In actual calculations, the number of layers exceeding 8 did not bring about a significant improvement in recognition rate, but instead amplified low-amplitude noise.

[0073] By dividing the enterprise operation data into at least two scale intervals and performing a multi-channel discrete wavelet transform, and then statistically fusing the coefficients of each frequency band, multi-band features can be obtained. Combined with the subsequent manifold analysis, chaotic pulse network and other steps of the present invention, it helps to achieve a more accurate and timely credit risk prediction system in terms of multi-period fluctuations and local anomaly identification.

[0074] Preferably, the step of obtaining high-dimensional distribution information in combination with manifold or persistent homology analysis includes constructing a high-dimensional complex structure according to preset adjacency rules, and forming a topological vector by calculating zero-dimensional and one-dimensional homology elements, and splicing the topological vector and the multi-band feature into a time-frequency manifold feature vector.

[0075] For large-scale enterprise operational data, relying solely on time-domain multi-band analysis (such as wavelet decomposition) is insufficient to capture the interconnected relationships between enterprises. By defining adjacency rules, mapping the feature vectors of all enterprises in the same time period onto a high-dimensional complex, and using a persistent homology algorithm to extract zero-dimensional elements (connected components) and one-dimensional elements (rings), we can reveal industry chain "clusters" or potential "closed-loop" supply relationships from a spatial topological perspective. This topological information is vectorized and concatenated with wavelet features to form feature vectors of a time-frequency manifold. This can be used with chaotic pulse networks or quantum search strategies for anomaly detection or credit trend assessment at higher dimensions.

[0076] In practical applications, adjacency rules and high-dimensional complexes, in the obtained digital twin snapshot, if there are N companies, the feature vector of each company in this period can be recorded as:

[0077] U i ={u i,1 ,u i,2 ,…,u i,m}

[0078] where u i,j Denotes the jth business indicator (such as inventory, cash flow or order increment). Define the distance metric d(U i ,U j ) is used to judge the similarity between enterprises i and j.

[0079] d(U i ,U j )<θ

[0080] Then, an adjacency relationship is established for i and j in the complex structure. The parameter θ is generally set in the range of 0.01≤θ≤0.05. If it is too low, the complex will easily split into too many sub-blocks, while if it is too high, all enterprises will be almost connected.

[0081] Persistent homology and topological vectors, based on the constructed high-dimensional complex, use the persistent homology algorithm to calculate the generation and extinction information of zero-dimensional elements (connected components) and one-dimensional elements (rings). These birth and death moments at different thresholds are numerically organized into a fixed-length vector T t , which is the topological vector. For large enterprise clusters, distance matrices can be constructed in parallel blocks and homology results can be merged in a high-performance environment to maintain acceptable computational cycles even at the measured scale of N = 50 or N = 100.

[0082] Splicing into time-frequency manifold feature vectors, first perform wavelet analysis on the enterprise time series data of the same period t to obtain the multi-band feature M t Then the topological vector T of the persistent homology output t Splice with it to form

[0083] Vt =[M t ||T t ]

[0084] Vector V t It includes both the time domain fluctuation characteristics and the topological structure of the high-dimensional space of the enterprise group. Subsequent chaotic pulse neural networks and quantum heuristic search can be based on V t Conduct training and inference to gain higher sensitivity to risks such as cascading defaults and localized abnormal clusters.

[0085] When a loop exists between enterprises (e.g., multiple suppliers signing closed-loop contracts), imbalanced operations at one enterprise can easily lead to cascading risks. Using zero- and one-dimensional factors, we can promptly identify such loops and provide a warning in risk scoring. In testing, when θ = 0.02, persistent homology was applied to the distance matrix, successfully identifying enterprise clusters with looped supply chains as potentially high-risk nodes, issuing warnings 1–2 cycles earlier than in scenarios without topological information.

[0086] Using only wavelet multiband features can only capture cyclical changes in the time domain, but cannot detect static or semi-static spatial correlations between companies. Incorporating topological vectors improved the detection rate of localized supply chain disruptions or capital collaboration failures in a 50-company scenario by approximately 15%, and the F1-score improved by 2%.

[0087] The basic implementation example is deployed in an 8-core CPU and 16GB memory environment and processes 1.5×10 7 Operational data. Perform 4-layer wavelet decomposition ("db4") on the enterprise time series indicators to obtain multi-band features M t , construct a high-dimensional complex in the distance matrix with a threshold θ = 0.02 and perform persistent homology, mapping the zero-dimensional and one-dimensional birth and death information into T t After and M t Spliced into the time-frequency manifold feature vector V t . V t A chaotic pulse network with 128 hidden layer nodes, ReLU activation function, and learning rate α=0.01 was trained, and it was observed that it converged within 20 rounds. The default warning accuracy was improved by about 2% compared with the control method without adding topological vectors.

[0088] Extended embodiment: Based on the above basic embodiment, when the number of enterprises increases to 100 and the amount of data increases by 2 times, θ is adaptively expanded to 0.03 and parallel blocking is used to calculate the distance matrix. The processing time increases by 10%, but the ability to capture local ring-shaped enterprise clusters increases by 1.5%. After 30 rounds of training, the F1-score increases by 2%. When α>0.1, training loss oscillations will occur, so maintaining α=0.01 is the optimal range.

[0089] like Figure 5 As shown, a chaotic pulse neural network is used to train the time-frequency manifold feature vector, and a chaotic map is configured in the network hidden layer to output the enterprise risk state vector; the risk state vector is assigned to the Holonic multi-agent environment to calculate the risk score. If the risk score exceeds the threshold or an operational anomaly is detected, a quantum heuristic search is performed in a parallel scheme and a linkage strategy is issued;

[0090] After completing the fusion of wavelet multi-band features and high-dimensional topological information (such as persistent coherence), the present invention forms a time-frequency manifold feature vector. By modeling this feature vector through a Chaos-based Spiking Neural Network, a chaotic map can be introduced in the hidden layer of the network to produce a stronger excitation response to short-period violent oscillations or local ring correlations, thereby outputting a risk state vector for the enterprise. After the state vector is assigned to the Holonic multi-agent environment, each enterprise agent performs a local judgment based on its own state vector; if the risk score exceeds the preset threshold, a multi-scheme parallel search can be performed through quantum heuristic search, and linkage strategies can be distributed to enterprise agents to quickly respond to possible high-risk events.

[0091] In practical applications, when training chaotic pulse neural networks and configuring chaotic sequences in hidden layer neurons, it is necessary to dynamically perturb the excitation threshold of the pulse neurons. For example, let the excitation threshold be initially set to θ0 and use a chaotic mapping function:

[0092] h n+1 =β·h n (1-h n )

[0093] where h n The nth round of perturbation is represented by β, which is a chaos control parameter (e.g., 3.5 to 4.0). This parameter is then added or multiplied to the threshold to make the network more sensitive to extreme or randomly changing data during training. The present invention recommends using ≥128 hidden layer nodes, a ReLU activation function, and a learning rate α within the range of 0.001 to 0.1, with the optimal value determined through comparative testing. If α is less than 0.001, the convergence rate decreases by 50%, and if α is greater than 0.1, oscillation occurs.

[0094] The enterprise risk state vector is output. After the training is completed, the chaotic pulse neural network is used to input the time-frequency manifold feature vector {V t} is mapped to the output layer to generate the risk state vector corresponding to each enterprise, which is recorded as:

[0095] Y t ={y 1,t ,y 2,t ,…,y N,t}

[0096] where y i,t Represents the risk representation or score of the i-th enterprise in time period t. This vector is assigned to the Holonic multi-agent structure, and the enterprise agent reads and calculates the final risk score. For example, if a certain rule is used to convert y i,t Converted to a score in the [0,1] interval.

[0097] Holonic multi-agent and quantum-inspired search: If an enterprise agent finds a score ≥ a threshold (e.g., 0.6), it reports the anomaly to the upper-level Holonic agent. If multiple enterprises simultaneously exceed the threshold, a parallel quantum-inspired search is employed to define multiple initial superposition states, exploring the probability amplitude of linkage strategies in parallel and selecting a solution with the highest target satisfaction. After issuing the solution, the enterprise agent can execute actions such as restricting credit, freezing some inventory, or implementing upstream and downstream control.

[0098] Preferably, the process of configuring a chaotic map in the hidden layer of the network includes applying a dynamic perturbation to the excitation threshold of the pulse neural network using a chaotic sequence, so that the network can more easily identify extreme risk states during training.

[0099] Conventional spiking neural networks use a fixed excitation threshold for pulse triggering, making it difficult to differentiate rare or extreme samples. Introducing chaotic sequences allows for random and nonlinear adjustments to the excitation threshold at each training round or each batch of data. For example, a chaotic map is defined as:

[0100] h n+1 =β·h n (1-h n )

[0101] where h n represents the perturbation factor of this round, and β is called the chaos control parameter, which is usually set to 3.5≤β≤4.0 to ensure chaotic behavior. When the original excitation threshold of the network hidden layer node is θ0, the system will make corrections before each batch of training:

[0102] θ n =θ0+κ·(h n -0.5)

[0103] κ controls the amplitude of the chaotic perturbation. If κ is set too large, the network will easily fall into an unstable zone; if it is set too small, the chaotic amplification effect will be less pronounced. In offline testing, this paper typically sets κ to ≈ 0.2, which achieves a good compromise when the number of hidden layer nodes is ≥ 128.

[0104] In practical applications, chaotic sequences are injected into the hidden layer of the network, and the chaotic mapping module is bound to the hidden layer of the pulse neural network in a containerized computing environment (for example, orchestrated by Kubernetes). Whenever the input batch is loaded, h is randomly updated. n And make corrections to the excitation threshold. The mapping is given by:

[0105] h n+1 =β·h n (1-h n )

[0106] Given, β needs to be selected experimentally to balance the convergence speed and the amplification of extreme samples. If β < 3.5, the chaos phenomenon is not obvious, and if β > 4, the sequence is prone to divergence.

[0107] Regarding training and learning rate ranges, this invention requires the learning rate α to be in the range [0.001, 0.1]. If α < 0.001, the convergence rate decreases by 50%, while if α > 0.1, the loss function oscillates. Comparative testing shows that a more robust convergence curve is obtained when α = 0.01. The remaining core parameters of the neural network, such as the number of hidden layer nodes (128), the ReLU activation function, and the batch size (32), are determined by the deployment environment and data scale. Typically, it can be run on an 8-core CPU with 16GB of memory.

[0108] To stimulate the dynamic perturbation of the threshold, before each batch of training begins, the original threshold θ0 is:

[0109] θ n =θ0+κ·(h n -0.5)

[0110] Under this threshold, the spiking neuron judges the internal accumulation potential of the input time-frequency manifold feature vector. Once it exceeds θ n This triggers a discharge and propagates the pulse downstream. Because κ causes the threshold to produce nonlinear random jitter, the network is more likely to enter an excited state when faced with high-amplitude or rare anomalies, thereby enhancing the recognition of extreme risks.

[0111] Through the present invention, it is measured that in a company with 50 enterprises and a daily increase of 1.5×10 7 In a sample of records, after enabling the chaotic sequence, the detection rate of sudden surges (such as funding shortages or large-scale order cancellations) increased by about 2.5% compared with the conventional fixed-threshold pulse network. It is more sensitive to extreme events such as sudden drops in inventory and can improve the warning timeliness by about 20% under the condition of the same number of training rounds.

[0112] If β = 3.8, κ = 0.2, and α = 0.01, the loss curve converges stably within 20 training epochs, and convergence speed is accelerated by approximately 30% for abnormal datasets. Increasing α beyond 0.1 may lead to unstable oscillations after the 10th epoch. Compared to a control experiment without the chaotic sequence, the convergence time increases by less than 10%, but the recognition accuracy for high-value defaults and large-value idle orders increases by 2-3%.

[0113] In a basic implementation, a chaotic spiking neural network was deployed on an 8-core CPU with 16GB of RAM. The activation function was ReLU, with 128 hidden nodes, a batch size of 32, β = 3.8, κ = 0.2, and α = 0.01. The time-frequency manifold feature vectors were trained for 20 epochs. The prediction accuracy for scenarios like sudden inventory drops and cash flow decreases increased by 2%. The average loss converged to a stable range after the 15th epoch.

[0114] In the extended implementation, based on the basic implementation, the number of hidden layer nodes is increased to 256 and β is adaptively adjusted between [3.6, 3.9] to handle a scenario where the number of companies doubles. This increases training time by approximately 25%, but allows for earlier detection of the cascading risk of simultaneous defaults by multiple companies, one or two batches of data later, and improves the F1-score by 1.5%. Gradient oscillations occur during training when the learning rate α is greater than 0.1, consistent with critical value experiments.

[0115] In light of the above, the present invention employs a chaotic map in the network's hidden layer, dynamically perturbing the excitation threshold of the spiking neural network through a chaotic sequence. This allows the model to demonstrate greater sensitivity to abnormal scenarios in big data-driven enterprise credit trend forecasting. Field measurements and extended examples demonstrate that this chaotic perturbation significantly improves the accuracy of extreme risk detection and early warning capabilities while maintaining convergence efficiency.

[0116] Preferably, the Holonic multi-agent environment is composed of enterprise agents and superior agents. The enterprise agents simulate and score the operation data according to the enterprise risk state vector, and the superior agents summarize and analyze the scoring results of the enterprise agents.

[0117] This invention introduces a two-tiered structure of enterprise agents and superior agents, reducing reliance on centralized scheduling through distributed collaboration. Enterprise agents receive the enterprise risk state vector output by a chaotic pulse neural network, simulate the enterprise's operational data (such as cash flow and order volume), calculate the risk score, and determine whether it exceeds the set threshold. If a large number of enterprise agents simultaneously report an anomaly, the superior agent uses quantum heuristic search to find the optimal linkage strategy (such as supplier replacement and fund allocation) in parallel solutions. This strategy is then distributed to relevant enterprise agents, enabling rapid intervention and overall scheduling.

[0118] In practical applications, the simulated scoring of the enterprise agent records the risk state vector output by the chaotic pulse neural network as:

[0119] Y t ={y 1,t ,y 2,t ,…,y N,t}

[0120] where y i,t Represents the risk value of the i-th enterprise in period t. The enterprise agent obtains its own weight y i,t Combined with operational indicators (inventory i,t 、Cash flow F i,t etc.) to calculate the pre-set scoring function. If the score exceeds a certain threshold θ s , then the abnormal information and main indicators are immediately sent to the upper-level agent queue. In the present invention, θ s The value is usually in [0.5,0.7]. If θ s <0.5 is prone to excessive false alarms. s >0.7, the detection rate of high-value defaults in the test dropped by nearly 2%.

[0121] Upon receiving an alert from an enterprise agent, the upper-level agent executes a quantum-inspired search in parallel within a containerized cluster. Setting two or more initial superposition states, it calculates the probability and magnitude of different disposal strategies (e.g., temporary financing, contract extension), and selects the optimal solution based on an objective function (e.g., minimizing default losses). Once the upper-level agent determines a solution, it immediately sends a "coordination instruction" to the relevant enterprise agents. This instruction can include specific steps such as temporary fund transfers and supplier changes.

[0122] The parameter range and critical value need to be within the enterprise agent score threshold θ s and the learning rate α. The chaotic pulse network of this invention was tested using α∈[0.001,0.1]. Above 0.1, loss oscillation occurred after 15 rounds, and below 0.001, the convergence rate dropped significantly by 50%. If the quantum superposition state of the parent agent is set to less than 2, parallel search becomes meaningless; if it exceeds 8, search time increases by 25% and the returns tend to level off.

[0123] Enterprise agents can independently complete the scoring in a container orchestration environment (such as deployed on AWS EC2 t3.xlarge instances). For example, when N=50, each round only takes 10 minutes to complete the simulation. Once the number exceeds θ s The alert enters the upper-level agent queue, and the entire collection initiates quantum search within 1 minute and produces a feasible strategy within 2 minutes. Compared with traditional centralized scheduling, the pull-through time is reduced by about 30%.

[0124] Tests have shown that, in extreme cases where multiple companies face simultaneous default risks due to a disrupted industrial chain, the superior agent's quantum search can evaluate multiple funding scheduling options in parallel and select the one with the highest returns, reducing the time required by 20% to 25% compared to sequential evaluation. Compared to conventional distributed methods, this method reduces the overall lag rate in alerting and linkage processing by nearly 15% and improves the F1-score by 2%.

[0125] Basic implementation: Deploy 50 enterprise agent containers and 1 parent agent container in an 8-core CPU and 16GB memory cluster. The enterprise agent has 128 hidden nodes, ReLU activation function, learning rate α = 0.01, and threshold θ. s =0.6. When more than five corporate agents reported an anomaly simultaneously, the superior agent invoked quantum search to solve the problem using two initial superposition states. It found that it took an average of three minutes before a unified policy (such as supplier changes or freezing some credit lines) was issued to the relevant corporate agents. This improved risk control efficiency by 1.5% compared to the control group without quantum search.

[0126] Extended embodiment: Based on the basic embodiment, the initial superposition state of quantum search is increased to 4, and the enterprise agent is allowed to adaptively change θ s The training process fluctuates within the range of [0.55, 0.65]. When the number of enterprises increases to 100 and the number of reports doubles, processing time increases by 10%, but the emergency detection rate increases by 1.8%. If α > 0.1, the training process exhibits gradient oscillation after 10 rounds, which is consistent with the critical value test and indicates that the optimal balance is α = 0.01.

[0127] As can be seen from the above, the Holonic multi-agent environment allows corporate agents to complete independent scoring after obtaining the risk state vector output by the chaotic pulse neural network. If the threshold is exceeded or an anomaly occurs, the upper-level agent uses quantum heuristic search in a parallel scheme to quickly issue a linkage strategy, building a risk management architecture that is efficient, scalable, and rapidly linked in the prediction of large-scale corporate credit trends.

[0128] Preferably, the threshold is used to determine whether abnormal fluctuations occur in the enterprise operation indicators in consecutive time periods. After the scoring result exceeds the threshold, the superior agent executes the linkage decision-making process.

[0129] This invention is based on large-scale enterprise operational data, combined with the risk status output by chaotic pulse networks and Holonic multi-agent evaluation, to monitor key enterprise operational indicators (such as inventory, cash flow, etc.) over a continuous period of time. Once the score exceeds the set threshold, it indicates that the fluctuation is significant or exceeds the established safety range, requiring the superior agent to intervene quickly and execute the linkage decision. The threshold is usually determined in the range of 0.5 to 0.7 based on historical default records, prediction error rate and other indicators. When the threshold is too low, the false alarm rate increases; when the threshold is too high, the false negative rate increases significantly. The optimal balance point can be found through experiments.

[0130] In practical applications, threshold determination and scoring calculation are performed by training the chaotic pulse neural network to obtain the enterprise risk state vector:

[0131] Y t ={y 1,t ,y 2,t ,…,y N,t}

[0132] where y i,t is the risk value of the i-th enterprise in period t. The enterprise agent compares this value with the increment of its own operating indicators in the last 2-3 cycles to form a comprehensive score result. For example, a weighted function can be used to combine recent cash flow and inventory changes for linear or nonlinear superposition. If the score exceeds the threshold θ s , relevant enterprise information (order volume, duration of outstanding payments, etc.) is packaged and reported to the upper-level agent. Tests show that on an edge node with an 8-core CPU and 16GB of memory, scoring and threshold determination can be completed for 50 enterprises within 10 minutes.

[0133] Abnormal fluctuations in continuous periods. To avoid false alarms caused by single abnormal spikes, the system will smooth or accumulate the fluctuation amplitudes monitored in continuous periods on the enterprise agent side. If the fluctuation amplitudes are higher than θ within the set window (for example, 2 time periods), s , the alarm weight is strengthened or the reporting is accelerated. This mechanism can filter out short-term jitter and focus on identifying persistent or significant anomalies.

[0134] Linked decision-making process, when the upper-level agent receives the alarm data of multiple enterprise agents, if the same industrial chain or multiple supply chain nodes are all higher than θ s , parallel solutions such as quantum heuristic search are used to evaluate different intervention strategies (such as emergency credit granting and adjusting supplier quotas). Based on the calculation results, instructions are then issued to enterprise agents for execution. If only a few enterprises are found to have reached the threshold and the impact is limited, localized decision-making can be adopted to save resources.

[0135] Through the solution of the present invention, it is measured in a scenario including 50 enterprises, assuming θ s=0.6, when inventory drops by more than 20% for two consecutive cycles, at least three corporate agents report the anomaly almost simultaneously, and the superior agent promptly executes parallel decisions, reducing the total default loss by about 15%. s <0.5, false positives increase by 30%, and at θ s When it is >0.7, the underreporting rate increases to more than 5%.

[0136] Taking a weighted average of the scores over two consecutive periods can reduce noise interference caused by single peaks. Compared to a control without smoothing, this reduces false alarms by approximately 20% while maintaining the same level of abnormal movement perception. If the learning rate α > 0.1 causes neural network training oscillation, it is recommended to control α to around 0.01 to ensure scoring stability.

[0137] In the basic implementation, 50 enterprise agent containers are deployed on the edge node (8-core CPU, 16GB memory). Each enterprise agent receives the neural network output Y t After that, the operating indicators of the past two periods (such as inventory I t 、Cash flow F t ) Perform linear superposition to form score S i,t If S i,t ≥θ s If the event is triggered for two consecutive periods, an alarm is sent to the upper-level agent. s =0.6 and abnormal inventory of 5 companies appeared in the 12-hour simulation, all of which were captured, with an underreporting rate of <1% and a processing delay of <10 minutes.

[0138] Expanded embodiment, based on the basic embodiment, when the number of enterprises increases to 100 and θ is introduced s The adaptive mechanism (oscillating within the range of [0.55, 0.65]) can improve the alarm accuracy by about 1.5% during peak periods. If α < 0.001, the network training convergence will slow down by 50% and cause the threshold judgment to lag. The quantum-inspired linkage ultimately reduces the loss caused by collective default of interdependent enterprises by about 2%, and increases memory usage by 15%. If θ s When it is >0.7, the underreporting rate of cases of continuous inventory decay is actually found to increase to 5%.

[0139] In summary, by setting the threshold θ s This system determines whether an enterprise's operating indicators fluctuate abnormally over consecutive time periods. If the scoring result exceeds this threshold, it triggers a higher-level agent-linked decision-making process, enabling real-time control of big data-driven enterprise credit trend forecasts. This system has been tested and validated in real-world scenarios involving parallel scoring, asynchronous alerting, and the final quantum search-based decision-making process for a large number of enterprises, demonstrating enhanced perception and response to sudden anomalies.

[0140] Preferably, the quantum heuristic search defines at least two initial superposition states, attenuates and amplifies each candidate solution in the probability amplitude space, and selects a linkage strategy that meets the target conditions after reaching a preset number of iterations.

[0141] When a cluster of enterprises faces potentially high risks from multiple companies simultaneously, a large number of feasible linkage strategies arises. Using traditional greedy or sequential evaluation methods can easily lead to high computational overhead or long decision delays in large-scale scenarios. Quantum-inspired search first defines at least two initial superposition states, assigning a set of probability amplitudes for candidate strategies to each state. After multiple rounds of attenuation and amplification updates, the search converges on the optimal strategy in the probability space. Ultimately, a strategy that satisfies the target condition (e.g., minimizing default losses) is obtained through a single collapse or measurement sampling.

[0142] In practical applications, the initial superposition state configuration is used to record the candidate solution set as {S1, S2, ..., S m Set 2 to 4 initial superposition states, and assign initial probability amplitudes to some strategies in each superposition state, such as:

[0143] ψ1:α1(S1),α1(S2);

[0144] Ψ2:α2(S3),α2(S4)

[0145] Here α k (S i ) indicates that the i Next Strategy S i The initial probability amplitudes need to be normalized to ensure that the sum of the probabilities is 1. This allows multiple strategies to be evaluated in parallel without having to search each one individually.

[0146] Amplitude attenuation and amplification. In each round of iteration, the system evaluates each strategy based on the established objective function (such as reducing the overall corporate default probability). If a strategy S i If the effect is poor, the corresponding probability amplitude will be attenuated. For example, let the attenuation coefficient be η and the negative feedback degree be δ i , you can press:

[0147] α k (S i )←α k (S i )×(1-ηδ i )

[0148] The probability is reduced. For those with better performance, the magnitude is amplified, making it easier to retain or strengthen the strategy in the next round. The number of iterations is usually 5-10 rounds. If η is too large, it is easy to oscillate midway, while too small will slow convergence.

[0149] After reaching a predetermined number of rounds (or yield curve convergence), the final strategy sampling process measures the probability amplitude of each strategy in each superposition state and selects the most likely strategy or the one with the best objective function value as the linkage strategy output. The system then distributes this strategy to the corresponding enterprise agent for execution, enabling them to implement measures such as temporary financing and supply chain restructuring. If subsequent data reflects unsatisfactory results, the search process can be reactivated in the next cycle.

[0150] Compared to a one-by-one evaluation approach, quantum-inspired search can simultaneously attenuate and amplify multiple strategies within the same iteration cycle, shortening the tree-like search path. Tests show that after eight rounds of iteration using five candidate strategies, the average computation time is 30% shorter than a traditional BFS search, and strategy implementation can be completed within 10 minutes on an 8-core CPU with 16GB of memory.

[0151] Comparative experiments on the attenuation coefficient η revealed that if η > 0.1, the probability distribution fluctuates after the fourth round. If η < 0.01, the convergence rate decreases by 50%. A value around η = 0.05 generally achieves a balance between eliminating inefficient solutions and retaining high-quality ones. Offline simulation data also revealed that, when the number of enterprises doubles, increasing the number of initial superposition states or slightly extending the number of iterations can yield a feasible strategy within 15 to 20 minutes.

[0152] Compared to single-state heuristic algorithms, this method can handle larger strategy sets on equivalent hardware and more easily escape local optimality traps through attenuation and amplification mechanisms. Compared to traditional genetic algorithms, it can reduce average computational latency by 25% for multi-constraint linkage strategies (fund allocation + contract extension).

[0153] In the basic implementation example, quantum heuristic search was performed in a container environment with an 8-core CPU and 16GB of memory. Two initial superposition states were defined, each containing 2 to 3 scenario probability components, with an attenuation coefficient of η = 0.05. After six rounds of iteration, strategy combinations such as capital injection and customer replacement were measured and selected, successfully reducing the risk scores of a group of 50 companies by approximately 2%, and the time consumed was reduced by 25% compared to sequential search.

[0154] The extended implementation example expands the basic implementation example to 100 companies and adds four initial superposition states, with an adaptive attenuation coefficient between 0.02 and 0.08. After 10 iterations of sampling, processing time increases by 15%, but the decision-making speed for the multi-link supply model of the industrial chain is still controlled at around 15 minutes, and the F1-score improves by 2%. When η > 0.1, the probability amplitude tends to oscillate violently, making convergence difficult.

[0155] This demonstrates that quantum-inspired search, by simultaneously defining at least two initial superposition states and performing probability amplitude attenuation and amplification, can quickly identify a desired linkage strategy after reaching a preset number of iterations. This approach effectively addresses the bottleneck of concurrent strategy evaluation in big data-driven corporate credit trend forecasting, significantly improving the efficiency and accuracy of decision-making for sudden or multi-corporate risks.

[0156] When a default is confirmed or a serious risk event occurs, the corresponding information is transmitted back to the chaotic pulse neural network for incremental updates based on the latest records. Through the digital twin model, the enterprise risk state vector can adapt to new risk conditions in a dynamic environment.

[0157] In large-scale enterprise data scenarios, chaotic pulse neural networks lack an update mechanism after conventional training, making it difficult to promptly absorb sudden defaults or major risk events. To overcome this shortcoming, the present invention collects the latest abnormal data through a digital twin model each time a serious event is confirmed, and incrementally inputs it into the network, executing a small-batch or online training process. This enhances the network's ability to perceive extreme situations while maintaining existing parameters, avoiding the high cost and long latency required for full retraining, while ensuring high accuracy and rapid responsiveness for subsequent risk assessments.

[0158] In actual applications, default event feedback is combined with data. When an enterprise agent or a higher-level agent confirms that a company has defaulted, experienced a funding shortage, or suffered inventory losses, the time period t′ of the event and key indicators (such as cash shortfall and order default rate) are recorded, and the latest status of the enterprise node is updated in the digital twin model. This incremental data is then packaged into small batches of input, recorded as:

[0159] X inc ={x i,t′ |i∈Γ}

[0160] Where Γ is the set of enterprises where serious incidents occur, x i,t′ Contains necessary upstream and downstream information (such as supply chain dependencies). These incremental samples are treated differently from regular historical data and are only used in the incremental training phase.

[0161] For online training of chaotic pulse neural networks, in a container environment with 8-core CPU and 16GB memory (such as AWS EC2t3.xlarge and orchestrated by Kubernetes), the system will incMini-batch training was performed, maintaining key network structures (e.g., 128 hidden layer nodes and ReLU activation function). The learning rate α must be within the range [0.001, 0.1]. If α < 0.001, the training convergence rate will significantly decrease by 50%. If α > 0.1, oscillations and stability will occur after approximately 10 iterations. The use of chaotic sequences in the hidden layer excitation threshold remains to enhance the network's response to outliers or burst magnitude increases.

[0162] The risk state vector is updated synchronously with the digital twin. After completing incremental training, the chaotic pulse neural network will output the updated enterprise risk state vector:

[0163] Y′ t+1 ={y 1,t′+1 ,…,y N,t′+1}

[0164] This vector is then written back to the digital twin model. Subsequent wavelet analysis or Holonic multi-agents, when invoking enterprise risk status, will automatically retrieve this vector to adapt to new conditions (such as the impact of supply chain disruptions), thereby maintaining the real-time and accurate risk prediction of large-scale enterprise groups.

[0165] Preferably, the linkage strategy includes issuing temporary limits, working capital adjustments or upstream and downstream collaborative management measures to corporate agents, and further verification by the superior agent based on execution feedback.

[0166] In the Holonic multi-agent environment, after collecting risk scores from corporate agents, the superior agent, if it identifies any companies with high risk indicators, selects a set of optimal or suboptimal coordinated measures through quantum-inspired search or other parallel decision-making mechanisms. These measures may include: temporarily imposing funding or transaction limits on risky companies to prevent further debt escalation; redistributing working capital to ensure core operations; and coordinating contracts or supply and marketing processes with upstream and downstream companies to mitigate the risk of cascading defaults from a systemic perspective. After implementation, the superior agent collects feedback on execution effectiveness and default status based on the digital twin model to further iterate or revise the plan.

[0167] In practice, temporary limits are imposed on new transactions or outstanding debts at the enterprise agent level, either by a certain amount or time limit, to prevent further arrears risk. For example, if a company is found to be experiencing severe inventory shortages and its financial score exceeds a threshold of 0.7, the company's agent will be instructed to: "Subsequent new transactions exceeding 30% of the order value require authorization from a higher authority." This approach can quickly mitigate risk exposure in the short to medium term and is suitable for scenarios where inventory or funding shortages have already occurred but the company's survival is still desired.

[0168] Working capital adjustment: If the superior agent determines that multiple companies may be dragged down by the default of a core enterprise, funds can be allocated to the core enterprise or other key nodes. For example:

[0169] Partial interest reduction or grace period for repayments; short-term advances can be injected into the enterprise from an internal funding pool. If funding is limited, quantum heuristic search is required to select the optimal allocation combination to maintain the operation of key nodes at the lowest cost. If the machine learning module determines that the risk cannot be reversed after the capital injection, the present invention may be abandoned.

[0170] Upstream and downstream collaborative management and control: When upstream and downstream companies have a high degree of coupling, corporate agents can collaborate to implement supply contract extensions, price adjustments, etc. If a company is short of inventory and cannot complete delivery within 7 days after calculation, temporary upstream and downstream "collaborative management and control" can be initiated, such as:

[0171] Reduce non-critical orders and have the superior agent coordinate and dispatch other suppliers or alternative parts. This link relies on the digital twin model to obtain complete link information so that a unified evaluation can be made on multi-dimensional indicators (capital occupation, production cycle, etc.) before approval.

[0172] Execution feedback verification from the upper-level agent: After deploying the linkage strategy for a period of time, the upper-level agent will again use the digital twin model to observe changes in the operating indicators of the enterprise agent or the risk score of the chaotic pulse network. If the risk score decreases significantly, the strategy is confirmed to be effective. If there is no improvement or the situation deteriorates, supplementary strategies can be initiated or higher-level capital transfers or parallel searches can be triggered.

[0173] Compared with the conventional indiscriminate quota limits or fund allocation for enterprises, the present invention prioritizes temporary quota limits or fund increases for high-risk enterprises, thus avoiding excessive economic losses caused by blind large-scale blockades. Tests show that when the number of enterprises is 50 and the daily incremental transaction records are 1.5×10 7 Through this hierarchical linkage strategy, the default rate is reduced by about 1-2% and the amount of manual intervention is reduced by 20%.

[0174] The digital twin model and the enterprise agent continuously monitor enterprise metrics and quantify the actual effects of the "limit + funding adjustment" policy. If the score still exceeds the threshold of 0.6 after 2-3 cycles, a quantum heuristic search for a corrective strategy can be performed. Compared to passively waiting after fixed intervention measures, this method offers greater iteration efficiency and scalability.

[0175] Under the traditional centralized "risk control and approval" model, a blanket credit limit cut is often applied to multiple companies facing successive risk outbreaks. While this can help stem losses in the short term, it can also lead to the misjudgment of companies with growth potential. This invention, based on quantum search and scoring differentiation, provides capital turnover or upstream and downstream collaboration measures for potentially resuscitable companies, limiting overall losses while minimizing foregone profits. This has been shown to improve revenue recovery by 1.5% in high-risk event scenarios.

[0176] In a basic implementation, enterprise agents (50 containers) and a superior agent (1 container) were deployed on edge nodes with 8-core CPUs and 16GB of memory. Once a company's score reached ≥0.7, a quantum-inspired search was triggered to initialize two superposition states and define four linkage strategies: temporary funding injection, 30% cap, upstream and downstream price negotiation, and waiver of support. After six iterations, the probability converged on the "30% cap + partial funding support" combination. This invention was then distributed to enterprise agents for implementation. Over the next two cycles, the default rate decreased by 2%, and processing time was reduced by 30% compared to a conventional first-instance trial.

[0177] In an extended implementation example, based on the above, the number of enterprises was expanded to 100, and four superposition states were set. The decay coefficient η = 0.05 was used. After eight iterations, a joint strategy of upstream and downstream coordinated management and control plus a small capital injection was selected. During the actual deployment, the learning rate α = 0.01 was used. If α > 0.1, iterative loss oscillation occurred; if α < 0.001, incremental convergence was delayed by 50%, affecting timely decision-making. The final F1-score improved by 2% for severe risk scenarios, and memory usage increased by 15%, demonstrating the scalability and effectiveness of the strategy.

[0178] To sum up, when determining the linkage strategy, the present invention can issue temporary limits, working capital adjustments, or upstream and downstream collaborative management operations for corporate agents, and the superior agent can re-verify or revise the strategy based on subsequent execution feedback, thereby achieving effective hierarchical management and continuous optimization of sudden risks of corporate groups driven by big data, avoiding excessive losses or waste of resources caused by centralized one-size-fits-all approaches.

[0179] Preferably, a small-batch online training method is used based on the incremental update of the latest records, the samples of defaulting enterprises are incorporated into the learning process of the chaotic pulse neural network, and the subsequent operating conditions are simulated through the digital twin model to correct the risk state vector.

[0180] Faced with constantly updated business operation data and sudden defaults, relying solely on traditional one-time models will fail to provide timely feedback on new risks, which may lead to delayed or misjudgment of prediction results. After detecting serious events (such as high arrears, supply chain collapse), the present invention packages the newly collected enterprise features into incremental data, and uses small-batch online training to locally update the chaotic pulse network, thereby retaining existing parameters while quickly integrating new abnormal features into the model weights, making the network's risk control assessment for subsequent time periods closer to real dynamics. The digital twin model provides an online simulation scenario for this incremental process, helping the network to more accurately grasp the possible evolution of defaulting companies in the next time period.

[0181] In actual applications, incremental sample collection and small-batch online training are used. When a company in the enterprise group is identified as defaulting or facing serious risk, the system records its main indicators during this period (such as capital chain interruption amount, inventory changes, transaction failure rate, etc.) and extracts linkage information of surrounding upstream and downstream companies from the digital twin model. This new abnormal data set is recorded as:

[0182] X inc ={x i,t′ |i∈Γ}

[0183] Where Γ represents the set of defaulting enterprises or the affiliated enterprises significantly affected by the default. inc , using a small batch input method to the chaotic pulse neural network for online training, generally selecting 5 to 10 iterations, and the recommended learning rate α is in the range of [0.001, 0.1]. If α < 0.001, the convergence speed decreases by 50%, while when α > 0.1, the loss curve tends to fluctuate within about 10 iterations.

[0184] The chaotic pulse network update works in conjunction with the digital twin model. During incremental training, only the core weights and chaotic parameters (such as activation threshold perturbations) of the network are fine-tuned, avoiding the high computing power and long latency required for full retraining. After training is completed, a new version of the enterprise risk state vector is obtained:

[0185] Y′ t+1 ={y 1,t′+1 ,…,y N,t′+1}

[0186] And write it back to the digital twin model, which can perform risk assessment based on the latest weights when simulating operational situations in the future, providing enterprise agents or superior agents with status predictions that are closer to current reality.

[0187] Small batch strategy and containerized deployment. In actual projects, the online update module can be deployed in a containerized form on the Kubernetes cluster. incBlock-based parallel processing. If multiple companies default during a peak period, the system will launch multiple copies of the online update container, complete incremental training, and then scale down based on cluster load. In testing, five rounds of training on incremental data from 50 companies typically took no more than 15 minutes on an 8-core CPU with 16GB of memory.

[0188] Compared to waiting for the next full training round, this method can update network weights within 10-15 minutes, taking into account new default characteristics. Tests have shown that it can improve the accuracy of risk scores for key enterprises by 2% and reduce underreporting in high-value debt scenarios by 1.5%.

[0189] Because of the small-batch online training, the model doesn't completely discard its previous understanding of normal data, but rather strengthens its perception of radical or extreme events within a limited number of training rounds. Compared to a completely fixed network, this approach increases the false negative rate by 3-4% during multiple bursts, while this incremental mechanism maintains the false negative rate at 1.5%.

[0190] Full retraining on large datasets can take 30 to 60 minutes or even longer, and require pausing or slowing down regular risk control processes. This incremental solution requires only 30% to 40% of the computing resources to process new defaults and update parameters, allowing the digital twin model to continuously maintain the latest understanding of the enterprise population, significantly reducing resource and time costs.

[0191] like Figure 6 As shown, a big data-driven enterprise credit trend prediction system is used to implement the method described above, characterized in that the system includes:

[0192] The Digital Twin Construction Module is used to build a digital twin model within a computing environment and import operational data from a group of businesses to generate a digital twin snapshot. In practical applications, the Digital Twin Construction Module is used to create a dynamic mapping for a group of businesses. It continuously receives and integrates data from multiple sources, such as financial, inventory, and order data, within a high-performance computing environment to quickly generate a "digital twin snapshot." Business managers or risk control systems can use this "twin snapshot" to obtain a comprehensive overview of recent operational trends, supporting subsequent analysis and decision-making.

[0193] The wavelet manifold extraction module is used to perform wavelet analysis on the time series indicators in the digital twin snapshots to extract multi-band features, and combine manifold or persistent homology analysis to obtain high-dimensional distribution information, and fuse the multi-band features with the high-dimensional distribution information to form a time-frequency manifold feature vector; the wavelet manifold extraction module will extract time series indicators from these "twin snapshots", first use wavelet analysis to decompose the original data into multiple frequency bands, capture short-term and long-term fluctuation characteristics, and then combine manifold or persistent homology analysis to detect the aggregation or ring structure of enterprise groups in high-dimensional space, forming a time-frequency manifold feature vector, providing more refined feature input for identifying complex fluctuations and potential chain risks.

[0194] The Chaotic Pulse Training module trains the time-frequency manifold feature vector using a chaotic pulse neural network. By configuring a chaotic map in the network's hidden layer, it outputs a corporate risk state vector. This risk state vector is then distributed to the Holonic multi-agent environment to calculate a risk score. Building on this, the Chaotic Pulse Training module constructs a chaotic pulse neural network, leveraging a "chaotic map" mechanism to provide a more robust response to extreme or rare situations. The output corporate risk state vector is then distributed to each corporate agent in the Holonic multi-agent environment to generate a risk score for each company. This ensures a stable perception of regular operating data while providing more rapid warnings when significant anomalies occur.

[0195] The quantum search decision module is used to perform quantum heuristic search and issue linkage strategies to corporate agents according to a parallel scheme after the risk score exceeds the threshold or an operating anomaly is detected; the quantum search decision module is used to perform quantum heuristic search through a parallel scheme after discovering that the scores of certain companies exceed the threshold or operating anomalies occur, and parallelly calculate multiple linkage strategy combinations (such as working capital allocation, supply chain integration, etc.) in the probability amplitude space, and then quickly select countermeasures with better returns or security, and issue them to relevant corporate agents for execution, reducing single-point decision-making delays.

[0196] The incremental update module is used to transmit the corresponding information back to the chaotic pulse neural network after confirming a default or a serious risk event, and to perform incremental updates based on the latest records through the digital twin model, so that the enterprise risk state vector can adapt to new risk conditions in a dynamic environment. The last incremental update module is used to transmit the newly generated data back to the chaotic pulse network again after a default or serious risk event actually occurs, update the model weights in small batches based on the latest records, and simulate subsequent scenarios in the digital twin model, so that the enterprise risk state vector can continue to adapt to the latest risk conditions, thereby ensuring that the risk control system will not be "blinded" by a serious event and lose accurate control of subsequent fluctuations.

[0197] The above are merely embodiments of the present application and are not intended to limit the present application. For those skilled in the art, the present application may have various modifications and variations. Any modifications, equivalent replacements, improvements, etc. made within the spirit and principles of the present application should be included within the scope of the claims of the present application.

Claims

1. A big data driven enterprise credit trend prediction method, characterized by: The following steps are involved: Build a digital twin model in a computing environment, import the operational data of the enterprise group and generate a digital twin snapshot; Performing wavelet analysis on the time series indicators in the digital twin snapshot to extract multi-band features; combining manifold or persistent homology analysis to obtain high-dimensional distribution information, and fusing the obtained multi-band features with the high-dimensional distribution information to generate a time-frequency manifold feature vector; The time-frequency manifold feature vector is trained using a chaotic pulse neural network, and a chaotic map is configured in the network's hidden layer to output an enterprise risk state vector. The risk state vector is then assigned to a Holonic multi-agent environment to calculate a risk score. If the risk score exceeds a threshold or an operational anomaly is detected, a quantum heuristic search is performed in parallel and a linkage strategy is issued. When a default is confirmed or a serious risk event occurs, the corresponding information is transmitted back to the chaotic pulse neural network for incremental updates based on the latest records. Through the digital twin model, the enterprise risk state vector can adapt to new risk conditions in a dynamic environment.

2. The method according to claim 1, characterized in that The steps of performing wavelet analysis on time series indicators include dividing the enterprise operation data into at least two scale intervals, extracting frequency band coefficients using multi-channel discrete wavelet transform, and obtaining multi-band features after statistically fusing the frequency band coefficients.

3. The method according to claim 1, characterized in that The step of obtaining high-dimensional distribution information by combining manifold or persistent homology analysis includes constructing a high-dimensional complex structure according to preset adjacency rules, forming a topological vector by calculating zero-dimensional and one-dimensional homology elements, and splicing the topological vector and the multi-band feature into a time-frequency manifold feature vector.

4. The method according to claim 1, wherein The process of configuring chaotic mapping in the network's hidden layer involves using chaotic sequences to dynamically perturb the excitation threshold of the spiking neural network, making it easier for the network to identify extreme risk states during training.

5. The method according to claim 1, wherein The Holonic multi-agent environment is composed of enterprise agents and superior agents. The enterprise agents simulate and score the operation data according to the enterprise risk state vector, and the superior agents summarize and analyze the scoring results of the enterprise agents.

6. The method according to claim 1, characterized in that The threshold is used to determine whether the enterprise operation indicators fluctuate abnormally in continuous time periods. When the scoring result exceeds the threshold, the upper-level agent executes the linkage decision-making process.

7. The method according to claim 1, characterized in that The quantum heuristic search defines at least two initial superposition states, attenuates and amplifies each candidate solution in the probability amplitude space, and selects a linkage strategy that meets the target conditions after reaching a preset number of iterations.

8. The method according to claim 1, characterized in that The linkage strategy includes issuing temporary limits, working capital adjustments or upstream and downstream collaborative management measures to corporate agents, and is further verified by the superior agent based on execution feedback.

9. The method according to claim 1, characterized in that The incremental update based on the latest records adopts a small-batch online training method, incorporates the samples of defaulting enterprises into the learning process of the chaotic pulse neural network, and simulates subsequent operating conditions through the digital twin model to correct the risk state vector.

10. A big data driven enterprise credit trend prediction system, used to implement the method according to any one of claims 1 to 9, characterized in that: The system includes: A digital twin building module is used to build a digital twin model in a computing environment and import the operational data of the enterprise group to generate a digital twin snapshot; A wavelet manifold extraction module is used to perform wavelet analysis on the time series indicators in the digital twin snapshot to extract multi-band features, and combine manifold or persistent homology analysis to obtain high-dimensional distribution information, and fuse the multi-band features with the high-dimensional distribution information to form a time-frequency manifold feature vector; a chaotic pulse training module for training the time-frequency manifold feature vector using a chaotic pulse neural network, outputting an enterprise risk state vector by configuring a chaotic map in a hidden layer of the network, and assigning the risk state vector to a Holonic multi-agent environment to calculate a risk score; The quantum search decision module is used to conduct quantum heuristic search and issue linkage strategies to enterprise agents according to parallel schemes when the risk score exceeds the threshold or an operational anomaly is detected; The incremental update module is used to transmit the corresponding information back to the chaotic pulse neural network after confirming a default or a serious risk event, and perform incremental updates based on the latest records through the digital twin model, so that the enterprise risk state vector can remain adaptable to new risk conditions in a dynamic environment.

Images