A method, medium, and system for constructing a problem-priority decision-making model that considers the prediction of operational impact.
By performing pre-spatial orthogonal alignment and high-order spatiotemporal manifold topology transmission network processing on multi-source heterogeneous operational data, and combining lock-free ring queues and minimum cost maximum flow resource scheduling algorithms, the resource scheduling and priority decision-making problem in high-concurrency scenarios is solved, achieving globally optimal and traceable decision results.
Patent Information
- Application Number
- CN202611115334.2
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-07-27
- Publication Date
- 2026-11-17
AI Technical Summary
Existing technologies cannot achieve globally optimal resource scheduling and priority decision collaborative reasoning for multi-source heterogeneous business data in high-concurrency scenarios, resulting in computational deadlock, resource conflicts, and lack of traceability in decision-making.
By collecting multi-source heterogeneous operational data and performing pre-spatial orthogonal alignment processing, a high-order spatiotemporal manifold topology transmission network is constructed. Combined with dynamic pruning of sliding memory windows and lock-free circular queues, a spatiotemporal evolution-constrained sparse activation hybrid expert decision-making network is driven. The minimum cost maximum flow resource scheduling algorithm is used for optimization, and combined with adaptive parameter updates and low-rank matrix decomposition, on-demand computing power allocation and dynamic backtracking correction are achieved.
It achieves collaborative reasoning for global optimal resource scheduling and priority decision-making in high-concurrency scenarios, improves the accuracy of impact prediction, ensures the traceability of decisions and the rational allocation of resources, and avoids computational deadlock and resource conflicts.
Smart Images

Figure CN123117589A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of large model technology, specifically, it relates to a method, medium and system for constructing a problem priority decision model that considers the prediction of business impact. Background Technology
[0002] In the field of enterprise management, problem prioritization relies on the comprehensive analysis of multi-source heterogeneous operational data, such as financial indicators, customer flow data, and public opinion data. Existing technologies typically employ rule engines or single machine learning models to independently score various operational indicators, followed by manual aggregation and sorting to determine problem priorities, and allocation of enterprise resources in an experience-driven manner. In high-concurrency real-time data stream scenarios, due to the inconsistent dimensions and scales of various heterogeneous data sources, feature alignment failures and computational deadlocks exist during multi-threaded parallel processing. Simultaneously, traditional models lack the ability to model the nonlinear coupling relationship between temporal dynamics and spatial topology, resulting in low prediction accuracy. Furthermore, manual resource allocation lacks global optimization guarantees, easily leading to local optima and resource conflicts, and the decision-making process lacks traceable quantitative evidence. In other words, existing technologies suffer from the technical problem of failing to achieve globally optimal resource scheduling and collaborative reasoning for priority decisions in high-concurrency scenarios with multi-source heterogeneous operational data. Summary of the Invention
[0003] In view of this, the present invention provides a method, medium and system for constructing a problem priority decision-making model that considers the prediction of operational impact, which can solve the technical problem in the prior art that global optimal resource scheduling and priority decision collaborative reasoning cannot be achieved under high-concurrency scenarios with multi-source heterogeneous operational data.
[0004] The present invention is implemented as follows: The first aspect of the present invention provides a method for constructing a problem-priority decision-making model that considers the prediction of operational impact, comprising the following steps:
[0005] Collect multi-source heterogeneous operational data, including financial indicators, passenger flow data and public opinion data. After pre-spatial orthogonalization alignment processing, construct a unified feature tensor and send it into a high-order spatiotemporal manifold topology transmission network to complete spatiotemporal feature decoupling and encoding.
[0006] A sliding memory window dynamic pruning mechanism is used to perform cascaded pre-dimensionality reduction processing on real-time streaming business data. Combined with physical isolation technology of memory ring buffer, hot data is kept in the background and cold data is swapped out in seconds, ensuring non-blocking access to high-concurrency data streams.
[0007] A lock-free circular queue and asynchronous multi-threaded parallel distribution architecture is constructed. A work-stealing algorithm is used to schedule various heterogeneous business indicators, driving a spatiotemporal evolution-constrained sparse activation hybrid expert decision network to complete parallel inference of multiple indicators and output the impact prediction tensor of each business problem.
[0008] The minimum cost maximum flow resource scheduling algorithm with influence prediction tensor input is used to establish a directed arc network with capacity and cost, taking each operational crisis as the source and the enterprise's allocable resources as the sink. The optimal resource scheduling scheme is solved and the priority sequence of each problem is derived in reverse.
[0009] Based on the adaptive lightweight parameter update strategy, and combined with the low-rank matrix factorization operator to compress the forward propagation computation of the spatiotemporal evolution constraint sparse activation hybrid expert decision network, only the core inference branch is deployed at the edge, and cascade gating is used to intercept worthless inference flow to achieve on-demand computing power allocation.
[0010] The priority decision sequence and resource scheduling scheme are output to the operation and management platform to dynamically trigger the priority update loop. When the real-time prediction error rate exceeds the backtracking trigger threshold, the internal state of the gating network is dynamically backtracked to roll back the gating parameters of the previous time step to correct the decision.
[0011] Specifically, the pre-space orthogonal alignment process refers to projecting heterogeneous indicators with inconsistent dimensions, such as financial indicators, passenger flow data, and public opinion data, onto the same orthogonal feature space before multi-threaded computation. This eliminates differences in dimensions and magnitudes, ensuring that each indicator has a consistent feature dimension representation in subsequent parallel computation, thereby avoiding deadlock and waiting in multi-threaded computation due to dimension misalignment.
[0012] Specifically, the high-order spatiotemporal manifold topology transmission network refers to a network structure that maps geospatial business entities and their dynamic temporal changes to a high-dimensional orthogonal feature space. It uses multidimensional difference operators to adaptively remove spatiotemporal coupling redundancy, thereby achieving spatiotemporal feature decoupling. It jointly models the time axis and spatial topology of business data into a high-dimensional manifold, and then propagates the features between nodes through a topology transmission mechanism.
[0013] Specifically, the sliding memory window dynamic pruning mechanism refers to dynamically adjusting the window length to extract the data segments needed for current processing from the continuously incoming real-time streaming business data, discarding expired data outside the window, and controlling the instantaneous memory usage within a preset upper limit. The window length value is determined by stress testing at different concurrency pressure levels on the target hardware platform and analyzing the memory peak curve.
[0014] Specifically, the memory ring buffer physical isolation technology refers to dividing the memory into several fixed-size physical isolation partitions to form a ring-shaped buffer pool. Hot data preferentially occupies the partitions closer to the computing core, while cold data is automatically mapped to the outer partitions and swapped out when the access interval exceeds the cold data swap-out time threshold. The cold data swap-out time threshold is determined by statistically analyzing the processing blocking rate under each swap-out time threshold.
[0015] Specifically, the lock-free circular queue refers to a first-in-first-out data structure that does not rely on mutex locks and maintains read and write pointers with atomic operations. Multiple producer threads and consumer threads complete data enqueueing and dequeueing through atomic comparison and exchange operations, completely eliminating waiting and deadlock caused by lock contention between threads.
[0016] The input tensor dimension of the spatiotemporal evolution-constrained sparse activation hybrid expert decision network is . ,in For batch size, For time steps, For the number of nodes, The feature dimension is given; the feature difference between adjacent time steps is calculated through temporal difference layering to obtain the spatial topological dynamics velocity tensor, with a dimension of . It is used to characterize the instantaneous rate of change of the operating status of each node.
[0017] The gated routing module uses the instantaneous cash flow gap and resource-constrained physical boundary of each node as independent variables, inputs them into the multilayer perceptron to generate gated routing coefficients, and the routing weight calculation formula is as follows: ,in For supply and demand conservation mapping function, For learnable route weight matrix, This is the physical constraint strength coefficient; the gated routing selectively activates the top 3 expert subnetworks with the highest physical sensitivity based on the routing weight.
[0018] The spatiotemporal evolution-constrained sparse activation hybrid expert decision network internally constructs a long-span residual dense jump matrix, enabling input features to bypass inactive expert subnetworks and achieve direct cross-layer flow of features. It also embeds a dynamic routing weight search algorithm based on Fibonacci golden section optimization, which uses the Fibonacci sequence to generate golden section ratio sampling points and performs efficient single-peak search in the routing weight solution space, with a cumulative iteration count of no more than 7 times.
[0019] Specifically, the minimum cost maximum flow resource scheduling algorithm iteratively finds the shortest augmenting path, pushes the maximum feasible flow along the shortest augmenting path each time, until there is no augmenting path from the source to the sink in the network. The resulting flow is the maximum flow, and the corresponding total cost is the minimum total operating loss. The flow allocation on each arc directly gives the optimal resource scheduling scheme, and the flow satisfaction of each source point is used to deduce the priority sequence of each operating crisis.
[0020] Specifically, the low-rank matrix decomposition operator refers to decomposing the weight matrix with a large number of parameters in the spatiotemporal evolution-constrained sparse activation hybrid expert decision network into the product of two low-rank matrices, so that the floating-point operation required for forward propagation is reduced from the quadratic level of the original matrix dimension to the linear level of the low-rank dimension. The value of the low-rank dimension is determined by testing the inference latency and accuracy loss curves under different low-rank dimensions on the target edge device.
[0021] This includes an adaptive dynamic activation intensity adjustment function, which calculates a comprehensive adjustment index based on the prediction error rate in the current batch, the average information entropy of the input feature tensor, and the proportion of available video memory in the current device. Based on the comparison between the comprehensive adjustment index and the activation interval boundary threshold, the upper bound parameter of the activation intensity is dynamically adjusted to control the activation number and cascaded gating truncation behavior of the expert subnetwork.
[0022] The steps for establishing the training dataset for the spatiotemporal evolution-constrained sparse activation hybrid expert decision-making network are as follows: extract financial indicators, passenger flow data, and public opinion data from the historical business database, divide them into time series samples according to time steps, perform pre-spatial orthogonal alignment processing on each time series sample, and label the priority sequence of each business issue and the corresponding resource scheduling scheme as supervision labels based on manual annotation or historical decision records.
[0023] Among them, the training of the spatiotemporal evolution constrained sparse activation hybrid expert decision network uses the Adam optimizer for backpropagation. The loss function is composed of the weighted sum of priority ranking loss and resource scheduling cost loss. The learning rate is dynamically adjusted through learning rate warm-up and cosine annealing scheduling strategies. Training stops when the prediction error rate of the validation set is lower than the convergence judgment threshold and no longer decreases for several consecutive rounds.
[0024] The window length ranges from 256 to 4096 data sample points; the cold data swapping time threshold ranges from 1 to 5 seconds; and the physical constraint strength coefficient... The values range for the following parameters: 0.1–2.0; 5%–20% for the backtracking trigger threshold; 0.01–0.1 for the lower bound threshold of activation intensity; 10%–30% for the low-rank dimension of the original matrix; 1%–5% for the convergence judgment threshold; 0.1–0.3 for the lower bound threshold of the activation interval of the comprehensive adjustment index; and 0.5–0.8 for the upper bound threshold. (The weighting coefficient is also mentioned.) The value range is 0.3 to 0.5. The value range is 0.2 to 0.4. The value range is 0.1 to 0.3.
[0025] A second aspect of the present invention provides a computer-readable storage medium storing program instructions that, when executed in a computer, are used to perform the above-described method for constructing a problem priority decision model that considers business impact prediction.
[0026] A third aspect of the present invention provides a problem priority decision-making model construction system that considers the prediction of operational impact, comprising the aforementioned computer-readable storage medium, wherein the system is a computer, the computer-readable storage medium is disposed within the system, and the system is provided with a microprocessor that executes program instructions stored in the computer-readable storage medium.
[0027] This invention eliminates the dimensional misalignment problem in multi-threaded parallel computing by employing pre-spatial orthogonal alignment processing to project heterogeneous operating indicators onto the same orthogonal feature space. It utilizes a high-order spatiotemporal manifold topological transmission network to jointly model the temporal and spatial topology of operating data, capturing spatiotemporal nonlinear dependencies and improving the accuracy of influence prediction. Furthermore, it transforms the priority ranking problem into a network flow optimization problem with strict mathematical guarantees by using a minimum-cost maximum-flow resource scheduling algorithm for the prediction tensor input. This minimizes overall operating losses under limited resource constraints and provides traceable quantitative evidence for priority decisions. In summary, this invention solves the technical problem mentioned in the background art of failing to achieve globally optimal resource scheduling and collaborative reasoning for priority decisions in high-concurrency scenarios with multi-source heterogeneous operating data. Attached Figure Description
[0028] Figure 1 This is a flowchart of the method of the present invention.
[0029] Figure 2 A schematic diagram of a spatiotemporal evolution-constrained sparse activation hybrid expert decision network structure.
[0030] Figure 3 This is a graph showing the relationship between the comprehensive adjustment index and the cascaded gating cutoff path ratio.
[0031] Figure 4 This diagram illustrates the time-series changes in real-time prediction error rates and the effect of backtracking corrections.
[0032] Figure 5 This is a diagram illustrating the data flow in the decision-making network.
[0033] Figure 6 This is a comparison chart of the distribution of feature components of each index before and after the pre-space orthogonalization alignment process. Detailed Implementation
[0034] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below.
[0035] like Figure 1 The diagram shown is a flowchart of a method for constructing a problem priority decision-making model that considers the prediction of operational impact, provided by the first aspect of this invention. This method includes the following steps:
[0036] S01. Collect multi-source heterogeneous business data, including financial indicators, customer flow data and public opinion data. After pre-spatial orthogonalization alignment processing, construct a unified feature tensor and send it into a high-order spatiotemporal manifold topology transmission network to complete spatiotemporal feature decoupling and encoding.
[0037] S02. A sliding memory window dynamic pruning mechanism is used to perform cascaded pre-dimensionality reduction processing on real-time streaming business data. Combined with physical isolation technology of memory ring buffer, hot data is kept in the resident state and cold data is swapped out in seconds, ensuring non-blocking access to high-concurrency data streams.
[0038] S03. Construct a lock-free circular queue and asynchronous multi-threaded parallel distribution architecture, perform work-stealing algorithm scheduling on various heterogeneous business indicators, drive the spatiotemporal evolution constraint sparse activation hybrid expert decision network to complete multi-indicator parallel reasoning, and output the impact prediction tensor of each business problem.
[0039] S04. Input the operational impact prediction tensor into the minimum cost maximum flow resource scheduling algorithm. With each operational crisis as the source and the enterprise's available resources as the sink, establish a directed arc network with capacity and cost, solve the optimal resource scheduling scheme, and deduce the priority sequence of each problem in reverse.
[0040] S05. Based on the adaptive lightweight parameter update strategy, and combined with the low-rank matrix factorization operator to compress the forward propagation computation of the spatiotemporal evolution constraint sparse activation hybrid expert decision network, only the core inference branch is deployed at the edge to intercept the worthless inference flow through cascade gating, so as to achieve on-demand computing power allocation.
[0041] S06. Output the priority decision sequence and resource scheduling scheme to the operation and management platform, dynamically trigger the priority update loop, and when the real-time prediction error rate exceeds the preset backtracking trigger threshold, start the dynamic backtracking of the internal state of the gating network, roll back the gating parameters of the previous time step to correct the decision.
[0042] Among them, the high-order spatiotemporal manifold topology transmission network is a network structure that maps geospatial business entities and their temporal dynamic changes to a high-dimensional orthogonal feature space. It uses multidimensional difference operators to adaptively remove spatiotemporal coupling redundancy and achieves spatiotemporal feature decoupling. Its core idea is to jointly model the time axis and spatial topology of business data into a high-dimensional manifold, and then propagate features between nodes through the topology transmission mechanism, thereby capturing spatiotemporal nonlinear dependencies that are difficult to express by traditional models.
[0043] Among them, the pre-space orthogonalization alignment process refers to projecting heterogeneous indicators with inconsistent dimensions, such as financial indicators, passenger flow data, and public opinion data, onto the same orthogonal feature space before multi-threaded computation. This eliminates differences in dimensions and magnitudes, ensuring that each indicator has a consistent feature dimension representation in subsequent parallel computation, thereby avoiding deadlock and waiting in multi-threaded computation due to dimension misalignment.
[0044] The sliding memory window dynamic pruning mechanism refers to dynamically adjusting the window length to extract the data segments needed for current processing from the continuously incoming real-time streaming business data, discarding expired data outside the window, thereby controlling the instantaneous memory usage within a preset upper limit. The recommended range for the window length is 256 to 4096 data sample points, and the specific value is determined by stress testing on the target hardware platform at different concurrency pressure levels and analyzing the memory peak curve.
[0045] Among them, the memory ring buffer physical isolation technology refers to dividing the memory into several fixed-size physical isolation partitions to form a ring-shaped buffer pool. Hot data preferentially occupies the partitions closer to the computing core, while cold data is automatically mapped to the outer partitions and swapped out when the access interval exceeds the cold data swap-out time threshold. The recommended range for the cold data swap-out time threshold is 1 to 5 seconds. The specific value is determined by conducting multiple iterative experiments under real business data flow and statistically analyzing the processing blocking rate under each swap-out time threshold.
[0046] Among them, cascaded pre-dimensionality reduction processing refers to the process of sequentially compressing the feature dimensions of a high-concurrency input data stream through multiple cascaded dimensionality reduction modules. Each dimensionality reduction module retains only the feature components that contribute the most to subsequent inference, thereby significantly reducing the data volume of a single processing without losing key prediction information.
[0047] Among them, the lock-free circular queue refers to a first-in-first-out data structure that does not rely on mutex locks and maintains read and write pointers with atomic operations. Multiple producer threads and consumer threads complete data enqueueing and dequeueing through atomic comparison and exchange operations, completely eliminating waiting and deadlock caused by lock contention between threads.
[0048] The work-stealing algorithm refers to each computing thread maintaining its own double-ended task queue. When a thread runs out of tasks, it automatically steals tasks from the tail of other threads' task queues, achieving dynamic load balancing among threads and preventing some threads from waiting for a long time, which would cause global computing resources to be suspended.
[0049] Among them, the spatiotemporal evolution-constrained sparse activation hybrid expert decision network (hereinafter referred to as the decision network) is based on deep learning, adopts a hybrid expert model structure, and has an input tensor dimension of [missing information]. ,in For batch size, For time steps, For the number of nodes, The feature dimension is [number]. First, the feature difference between adjacent time steps is calculated through temporal difference layering to obtain the spatial topological dynamics velocity tensor, with dimension [number]. This is used to characterize the instantaneous rate of change in the operational status of each node. The gated routing module uses the instantaneous cash flow gap and resource-constrained physical boundaries of each node as independent variables, inputting them into the multilayer perceptron to generate gated routing coefficients. The routing weight calculation formula is... ,in For supply and demand conservation mapping function, For learnable route weight matrix, Physical constraint strength coefficient The recommended range is 0.1 to 2.0, determined by taking the value corresponding to the fastest convergence of the loss function through multiple comparative experiments on the validation set. Gated routing selectively activates the top three expert subnetworks with the highest physical sensitivity based on route weights, and the output dimension is... ,in This represents the feature output dimension of the expert subnetwork. A long-span residual dense jump matrix is constructed within the network, enabling input features to bypass inactive expert subnetworks and achieve direct cross-layer flow at the feature level, preserving all input feature information. At the neuron activation state level, the decision network embeds a dynamic routing weight search algorithm based on Fibonacci golden ratio optimization: during the model inference pre-stage, the algorithm uses the Fibonacci sequence to generate a set of golden ratio sampling points, performing an efficient single-peak search in the routing weight solution space. The temporary weights obtained from each generation of search are directly used as the bias scaling factor for routing gating, with a cumulative iteration count not exceeding 7 times. When a better path is found, the algorithm instantly rewrites the gating activation state of the corresponding neuron, achieving second-level routing adaptation and avoiding the lengthy optimization in spatiotemporal topology required by traditional backpropagation. In terms of information flow control, if the real-time prediction error rate exceeds the backtracking trigger threshold, the internal state of the gating network is dynamically backtracked, rolling back the gating parameters of the previous time step. The recommended range for the backtracking trigger threshold is 5%–20%, and the optimal value is determined by multi-round cross-validation on historical business data and statistical analysis of the joint distribution of decision accuracy and backtracking frequency under different thresholds. Regarding the weight allocation of connections between neurons, the decision network dynamically allocates the information transmission bandwidth weights between neurons based on the proportion of output activation intensity of each expert subnetwork. Connection paths with activation intensity below the lower bound threshold are cascaded and truncated. The recommended range for the lower bound threshold is 0.01–0.1, determined by analyzing the changes in inference accuracy and computational cost before and after truncation through ablation experiments. In terms of inter-layer memory allocation, the decision network dynamically allocates the upper limit of memory usage for each layer proportionally based on the real-time variance statistics of the feature maps of each layer. Layers with higher variance receive more memory allocation, thus prioritizing the storage of high-information-content features under overall memory constraints.
[0050] The technical effects of the decision network are reflected in the following aspects: Gated routing embeds physical constraints such as supply and demand conservation into the routing decision, ensuring that the activation path of the expert subnetwork always aligns with the actual physical laws of operation, rather than being driven solely by data statistics; the long-span residual dense jump matrix ensures that feature information on inactive expert paths is not discarded, thus maintaining the model's ability to perceive the global operational status under sparse activation conditions; the Fibonacci golden section optimization algorithm approximates the global optimum in the routing weight solution space with very few iterations, enabling the gated decision to quickly converge to accurate priority prediction results even under limited computing power at the edge.
[0051] The specific steps for establishing the training dataset for the decision-making network include: extracting financial indicators, passenger flow data, and public opinion data from historical operational databases across multiple time periods and geographical locations, and organizing them by time step. The sample is divided into time-series samples; each time-series sample undergoes pre-spatial orthogonalization alignment to construct a dimension of... The feature tensor; based on manual annotation or historical decision records, each sample is labeled with the priority sequence of each business problem and the corresponding resource scheduling scheme as a supervision label; 80% to 90% of the samples are divided into training set and the remaining samples are divided into validation set, and the proportion is determined by comparing the validation set loss under different division proportions through experiments.
[0052] The training steps for the decision network specifically include: inputting the training set into the decision network in batches; forward propagation sequentially through the temporal difference layering, the gated routing module, and the activated expert subnetwork; outputting the operational impact prediction tensor; calculating the loss function value between the predicted priority sequence and the labeled priority sequence, the loss function being a weighted sum of the priority ranking loss and the resource scheduling cost loss, with the weight coefficients determined through validation set experiments; and using the Adam optimizer for backpropagation updates. And the parameters of each expert subnetwork, the recommended range of learning rate is: ~ The learning rate is dynamically adjusted through preheating and cosine annealing scheduling strategies. After each training round, the prediction error rate is evaluated on the validation set. Training is stopped when the error rate is lower than the convergence threshold and does not decrease for several consecutive rounds. The recommended range for the convergence threshold is 1% to 5%, which is determined by statistically analyzing the inflection point of the loss curve through multiple experiments.
[0053] The minimum-cost maximum flow resource scheduling algorithm is a classic network flow algorithm in graph theory. It abstracts operational crises at each level as source nodes in a directed network and the company's available rescue resources as sink nodes. The capacity limit of each directed arc in the network corresponds to the resource carrying capacity limit, and the unit cost of the arc corresponds to the additional operational losses caused by rescue delays. The algorithm iteratively finds the shortest augmenting path (with cost as the path length), pushing the maximum feasible flow along the shortest augmenting path each time until there is no more augmenting path from the source to the sink in the network. The resulting flow is the maximum flow, and the corresponding total cost is the minimum total operational loss. The algorithm then solves for the maximum flow... Simultaneously, the flow allocation on each arc directly provides the optimal resource scheduling scheme, and the flow satisfaction of each source point is used to deduce the priority sequence of each operational crisis. This enables resource scheduling and priority decision-making to be globally optimized within the same algorithm framework, completely eliminating the local optima and resource conflict problems in manual allocation. The technical effect of this algorithm is that it transforms the priority ranking problem, which originally relied on experience-based judgment, into a network flow optimization problem with strict mathematical guarantees. Under the constraint of limited resources, it minimizes the overall operational loss and provides a traceable quantitative basis for priority decision-making, improving the global consistency and interpretability of the decision.
[0054] Among them, the adaptive dynamic activation intensity adjustment function The upper bound parameter used to adjust the activation intensity of each expert subnetwork in the decision network is based on the prediction error rate in the current batch. Average information entropy of input feature tensors Real-time available video memory ratio of the current device Three data points are used to calculate the comprehensive adjustment index. ,in , , These are weighting coefficients, and the sum of the three is 1. Their specific values are determined through grid search experiments on the validation set, and the recommended range is as follows: , , ;when When the activation intensity upper bound parameter is adjusted to the high activation range, more expert subnetworks are allowed to participate in inference; when At this time, the upper bound parameter of activation intensity remains in the medium activation range, keeping the current number of activated expert subnetworks unchanged; when When the activation intensity upper bound parameter is lowered to the low activation interval, cascade gating is triggered to truncate the low-contribution expert subnetwork, reducing the forward propagation computation. The boundary values of 0.2 and 0.6 for each interval are determined by conducting multiple comparative experiments on multiple sets of historical operating data, with the joint optimization objective of prediction accuracy and inference latency. The recommended range for the lower bound of 0.2 is 0.1 to 0.3, and the recommended range for the upper bound of 0.6 is 0.5 to 0.8.
[0055] Among them, the low-rank matrix factorization operator refers to decomposing the weight matrix with a large number of parameters in the decision network into the product of two low-rank matrices, so that the floating-point operation required for forward propagation is reduced from the square level of the original matrix dimension to the linear level of the low-rank dimension, thereby compressing the computing power of model inference on edge devices; the recommended range of low-rank dimension is 10% to 30% of the original matrix rank, and the optimal value is determined by testing the inference latency and accuracy loss curves under different low-rank dimensions on the target edge device.
[0056] In this method, all specific values such as the backtracking trigger threshold, cold data swapping time threshold, activation intensity lower bound threshold, and low-rank dimension are not preset with fixed values. Instead, they are selected from the above recommended range to achieve the best overall performance after designing a control experiment on the historical dataset of the target business scenario, conducting multiple iterative tests, and analyzing the joint distribution curve of each indicator.
[0057] The specific implementation of step S01 is as follows: First, financial indicators (including operating revenue, gross profit margin, cash flow gap, etc.), customer flow data (including the number of visitors to each store, conversion rate, etc.), and public opinion data (including the sentiment polarity of user comments, frequency of complaints, etc.) are collected from the business database and external data interfaces. These three types of data show significant differences in dimensions, sampling frequency, and feature dimensions. The core principle of the pre-space orthogonalization alignment process is to uniformly project various indicators onto the same orthogonal feature space through linear or affine transformations, eliminating differences in dimensions and magnitudes. This ensures that the feature vector dimensions processed by each thread in subsequent multi-threaded parallel computation are completely consistent, thereby physically preventing computational deadlock and waiting caused by dimension misalignment. The aligned multi-class features are then concatenated into a unified feature tensor, with the dimension denoted as . ,in For time steps, For the number of operating nodes, The feature dimension is denoted as . This feature tensor is then fed into a high-order spatiotemporal manifold topological propagation network, which jointly models the temporal axis and spatial topology of the operational data as a high-dimensional manifold. Through multidimensional difference operators, it adaptively removes spatiotemporal coupling redundancy and propagates features between nodes via a topological propagation mechanism, thus completing the decoupling and encoding of spatiotemporal features and outputting a high-dimensional feature representation with spatiotemporal semantics.
[0058] The specific implementation of step S02 is as follows: To address the high concurrency characteristics of real-time streaming operational data, a sliding memory window dynamic pruning mechanism is used to preprocess the continuously flowing data stream. This mechanism works by dynamically adjusting the window length to extract the data segment required for the current inference. The recommended window length ranges from 256 to 4096 data sample points. The specific value is determined by stress testing on the target hardware platform at different concurrency pressure levels and analyzing the memory peak curve. In a typical scenario, 512 data sample points can be used. Expired data outside the window is discarded immediately, thus controlling instantaneous memory usage within a preset upper limit. Simultaneously, a memory ring buffer physical isolation technology divides the memory into several fixed-size physical isolation partitions, forming a ring-shaped buffer pool. Hot data (i.e., data accessed frequently recently) preferentially occupies partitions closer to the computing core, while cold data (i.e., data with longer access intervals) is automatically mapped to peripheral partitions. When the cold data access interval exceeds the cold data swapping time threshold (recommended range 1-5s, typical value 2s), it is swapped out, ensuring that hot data remains resident in memory to support low-latency inference. The cascaded pre-dimensionality reduction process compresses the feature dimensions of the data stream step by step through multiple cascaded dimensionality reduction modules after the sliding memory window is truncated. Each dimensionality reduction module retains only the feature components that contribute the most to subsequent inference, thereby significantly reducing the data volume of a single processing without losing key prediction information, and keeping the input scale of subsequent parallel inference controllable.
[0059] The specific implementation of step S03 is as follows: A lock-free circular queue is constructed as the data channel for various heterogeneous operational indicators. This lock-free circular queue does not rely on mutex locks; it maintains read and write pointers using atomic comparison-exchange operations. Multiple producer and consumer threads complete data enqueueing and dequeueing through atomic operations, completely eliminating waiting and deadlocks caused by thread lock contention. An asynchronous multi-threaded parallel distribution architecture distributes the inference tasks for each indicator to multiple computing threads. Each thread maintains its own double-ended task queue. A work-stealing algorithm automatically steals tasks from the tail of other threads' task queues when a thread runs out of tasks, achieving dynamic load balancing among threads. This architecture drives a spatiotemporal evolution-constrained sparse activation hybrid expert decision network to complete multi-indicator parallel inference. The input tensor dimension of this network is... The instantaneous rate of change of the operating status of each node is extracted through time-series difference layering, and the gated routing module is based on the supply and demand conservation mapping function. With learnable route weight matrix Calculate route weights The algorithm selectively activates the top three expert subnetworks with the highest physical sensitivity. The long-span residual dense jump matrix ensures that the feature information on the inactive path is not discarded. The Fibonacci golden section optimization algorithm approaches the global optimal gating state in the routing weight solution space in no more than 7 iterations, and finally outputs the impact prediction tensor of each business problem.
[0060] The specific implementation of step S04 is as follows: The impact prediction tensor output from step S03 is used as the input to the minimum-cost maximum-flow resource scheduling algorithm. Each level of operational crisis is abstracted as a source node in a directed network, and the enterprise's available rescue resources are abstracted as sink nodes. The capacity limit of each directed arc in the network corresponds to the resource carrying capacity limit, and the unit cost of the arc corresponds to the additional operational losses caused by rescue delays. The algorithm iteratively finds the shortest augmenting path (with cost as the path length), pushing the maximum feasible flow along the shortest augmenting path each time until there is no longer an augmenting path from the source node to the sink node in the network. At this point, the maximum flow obtained corresponds to the minimum total operational loss. The flow allocation on each arc directly provides the optimal resource scheduling scheme, and the flow satisfaction degree of each source node inversely derives the priority sequence of each operational crisis, making resource scheduling and priority decision-making globally optimized within the same algorithm framework.
[0061] The specific implementation of step S05 is as follows: Based on the adaptive lightweight parameter update strategy, the weight matrix with a large number of parameters in the spatiotemporal evolution-constrained sparse activation hybrid expert decision network is decomposed using the low-rank matrix factorization operator, and replaced with the product of two low-rank matrices. This reduces the floating-point operation required for forward propagation from the quadratic level of the original matrix dimension to the linear level of the low-rank dimension. The recommended low-rank dimension is 10% to 30% of the original matrix rank. Only the core inference branch is deployed at the edge, and cascade gating is used to truncate connection paths with activation strength below the lower bound threshold of activation strength (recommended range 0.01 to 0.1), intercepting worthless inference flows. The adaptive dynamic activation strength adjustment function calculates a comprehensive adjustment index based on the current batch prediction error rate, the average information entropy of the input feature tensor, and the proportion of real-time available video memory on the device. ,in accordance with The number of activated expert subnetworks is dynamically adjusted based on the comparison results with boundary thresholds (lower bound recommended range 0.1-0.3, upper bound recommended range 0.5-0.8) to achieve on-demand computing power allocation.
[0062] The specific implementation of step S06 is as follows: The priority decision sequence and optimal resource scheduling scheme output in step S04 are pushed to the operation and management platform. The operation and management platform dynamically triggers the priority update loop based on real-time monitored operation data. When the real-time prediction error rate exceeds the backtracking trigger threshold (recommended range 5% to 20%, typical value 10%), the system activates the dynamic backtracking mechanism of the internal state of the gating network, rolls back the gating parameters of the spatiotemporal evolution constraint sparse activation hybrid expert decision network to the state of the previous time step, so as to eliminate the decision deviation caused by noise or abnormal data in the current time step, thereby completing the automatic correction of the priority decision sequence and resource scheduling scheme and ensuring the continuous reliability of the output results.
[0063] It should be noted that the key technologies of this invention include: First, the pre-space orthogonalization alignment process projects all heterogeneous operating indicators onto the same orthogonal feature space before entering the parallel computing process, eliminating the risk of multi-threaded deadlock caused by inconsistent dimensions from the root. This is a structural problem that is common in traditional asynchronous multi-threaded inference frameworks but difficult to solve fundamentally through locking. Orthogonalization alignment allows all subsequent parallel computing modules to share a unified feature dimension benchmark. Second, the spatiotemporal evolution-constrained sparse activation hybrid expert decision network embeds physical constraints such as supply and demand conservation into the gated routing decision, ensuring that the expert activation path is always constrained by the real physical laws of operation, rather than being purely driven by data statistical distribution. Combined with the Fibonacci golden section optimization algorithm, it approaches the global optimum of routing weights in a very small number of iterations, ensuring both the convergence speed and accuracy of decision-making under the condition of limited computing power at the edge. Third, the minimum cost maximum flow resource scheduling algorithm transforms the priority ranking problem into a network flow optimization problem with strict mathematical guarantees, completely eliminating local optima and resource conflicts in manual allocation. The overall effect of the synergistic effect of the three key technologies is as follows: orthogonal alignment ensures the stability of multi-threaded parallel inference, the hybrid expert decision network improves the accuracy of influence prediction, and the network flow algorithm transforms the accurate prediction tensor into a globally optimal scheduling scheme and a traceable priority sequence. The three form a complete closed loop from data access and feature inference to decision output. The overall system has self-consistent technical logic in high-concurrency and resource-constrained edge scenarios.
[0064] It's important to note that in real-time enterprise operations with multiple business nodes, when multiple regions experience simultaneous operational anomalies and the impact and urgency of various issues are highly intertwined, the decision-making system needs to quantify the impact and prioritize a large number of concurrent operational events within a very short time window, while ensuring that limited human, financial, and material resources are not duplicated or overlooked. The reason for this technical problem is that traditional prioritization methods typically treat impact assessment and resource allocation as two independent processes. First, a model or rule engine scores each operational issue individually, then schedulers manually summarize, prioritize, and allocate resources. As the number of concurrent events increases, the delay in manual summarization increases linearly with the number of events, and the coupling relationship between impact scoring and resource constraints cannot be accurately characterized in the manual process. This leads to a disconnect between the final ranking result and the actual availability of resources, resulting in local optima and resource conflicts. A common solution to this technical problem is to introduce heuristic rules or greedy algorithms to approximate the resource allocation order. However, heuristic rules rely on expert experience and cannot be adaptively updated to adapt to changes in business scenarios. Furthermore, greedy algorithms can only guarantee local optima, not global optima. In resource-constrained scenarios, they are prone to the phenomenon of high-priority issues not being responded to due to resource exhaustion, while low-priority issues consume a large amount of resources, resulting in reverse allocation. This invention effectively solves this technical problem. Its core lies in unifying impact prediction and resource scheduling into the minimum cost maximum flow algorithm framework. The impact prediction tensor of each business crisis is directly used as the construction basis for the directed arc cost, and the enterprise resource constraints are directly mapped to the arc capacity limit. In the same network flow solution process, the global optimal resource allocation and the reverse derivation of the priority sequence are completed simultaneously. From the algorithm structure, the strong coupling between impact assessment and resource constraints is guaranteed, eliminating the global suboptimal problem caused by the two-stage separate solution. Moreover, the optimality of the network flow algorithm is guaranteed by strict mathematical theorems, without relying on human experience or heuristic assumptions, and can provide traceable quantitative decision-making basis under any concurrent event scale.
[0065] A second aspect of the present invention provides a computer-readable storage medium storing program instructions that, when executed in a computer, are used to perform the above-described method for constructing a problem priority decision model that considers business impact prediction.
[0066] A third aspect of the present invention provides a problem priority decision-making model construction system that considers the prediction of operational impact, comprising the aforementioned computer-readable storage medium, wherein the system is any one of a computer, a server, or a microcontroller, the computer-readable storage medium is disposed within the system, and the system is provided with a microprocessor that executes the program instructions stored in the computer-readable storage medium.
[0067] Specifically, the principle of this invention is as follows: The fundamental reason why this invention can solve the above-mentioned technical problems lies in the synergistic effect of three core mechanisms. First, the pre-processing spatial orthogonalization alignment unifies heterogeneous indicators with inconsistent dimensions, such as financial indicators, passenger flow data, and public opinion data, onto the same orthogonal feature space before the start of multi-threaded computation. This eliminates differences in dimensions and magnitudes, ensuring that the threads of subsequent parallel computation are fully aligned in feature dimensions, thereby completely eliminating computational deadlock and waiting caused by inconsistent dimensions. This pre-processing is the logical prerequisite for the stable operation of multi-threaded parallel inference in high-concurrency scenarios. Second, the high-order spatiotemporal manifold topology transmission network jointly models the time axis and spatial topology of operational data into a high-dimensional manifold. It uses multi-dimensional difference operators to adaptively remove spatiotemporal coupling redundancy and propagates features between nodes through a topology transmission mechanism, capturing spatiotemporal nonlinear dependencies that are difficult to express by traditional single models. The temporal difference layer extracts the instantaneous rate of change of the operating status of each node. The gated routing module embeds physical constraints such as supply and demand conservation into the routing decision, ensuring that the activation path of the expert subnetwork always aligns with the real physical laws of operation, rather than being driven solely by data statistics. The long-span residual dense jump matrix ensures that feature information on inactive expert paths is not discarded, thus maintaining the model's ability to perceive the global operating status under sparse activation conditions. The Fibonacci golden section optimization algorithm approximates the global optimum in the routing weight solution space with very few iterations, enabling the gated decision to quickly converge to accurate priority prediction results even under limited computing power at the edge. This network structure is the core technical support for improving prediction accuracy. Third, the minimum cost maximum flow resource scheduling algorithm abstracts the operating crises at each level as the source of a directed network and the enterprise's allocatable resources as the sink. It iteratively finds the shortest augmenting path, solves the minimum total operating loss under the constraint of limited resources, and simultaneously derives the priority sequence of each operating problem in reverse. This algorithmic framework transforms the priority ranking problem, which originally relied on empirical judgment, into a network flow optimization problem with rigorous mathematical guarantees. It completely eliminates local optima and resource conflicts in manual allocation and provides traceable quantitative evidence for decision-making. The three mechanisms work sequentially and are logically consistent, together forming a complete closed loop for solving the core technical problem of this invention.
[0068] The following provides a specific embodiment 1 of the present invention, and the specific implementation of each step in this embodiment 1 is described in detail below.
[0069] The specific implementation method of step S01 is as follows.
[0070] After collecting three types of heterogeneous operational data—financial indicators, passenger flow data, and public opinion data—the first step is to perform pre-spatial orthogonal alignment processing. Let the... The original feature vector of the class is Through the projection matrix Mapping it to a unified orthogonal feature space yields the alignment vector. ,in To unify feature dimensions, satisfy , for An identity matrix of order 1. The singular value decomposition of the covariance matrix of various data types is used to obtain the previous data. The left singular vectors are obtained. After alignment, the three types of features are concatenated to construct a unified feature tensor. ,in For batch size, For time steps, For the number of nodes, This represents the feature dimension. Then... The signal is fed into a high-order spatiotemporal manifold topological transmission network, which utilizes multidimensional difference operators. Feature decoupling is performed on the high-dimensional manifold jointly modeled by the time axis and spatial topology, and the temporal difference operator is applied to the time axis. The spatial difference operator acts on the node neighborhood topology, and propagates the decoupled features between nodes through the topology propagation mechanism, and finally outputs a spatiotemporal feature encoding tensor for use by subsequent modules.
[0071] The specific implementation method of step S02 is as follows.
[0072] For the continuous influx of real-time streaming operational data, the sliding memory window dynamic pruning mechanism uses a dynamic length... Extract the current data segment. The value ranges from 256 to 4096 data sample points. The specific value was determined by stress testing on the target hardware platform at different concurrency levels and analyzing the memory peak curve. A circular buffer pool of memory is formed by partitioning into fixed-size partitions. Hot data preferentially occupies the partitions closest to the computing core, while cold data is mapped to peripheral partitions. When the access interval of a partition exceeds the cold data swap-out time threshold... The swap is executed at the appropriate time. The value range is 1 to 5. The blocking rate was determined through multiple iterative experiments under real business data streams, statistically analyzing the processing blocking rate at various swapping thresholds. The cascaded pre-dimensionality reduction processing sequentially processes the input data... A series of sequential dimensionality reduction modules progressively compress the feature dimensions, the first... Level dimensionality reduction output is ,in This is the original input after the window has been cropped. For the first The dimensionality reduction function retains only the feature components that contribute the most to subsequent reasoning.
[0073] The specific implementation method of step S03 is as follows.
[0074] Lock-free circular queues maintain read / write pointers using atomic compare-and-swap operations, completely eliminating lock contention between threads. The work-stealing algorithm assigns a work-to-the-work task to each computation thread. Maintain a dual-end task queue ,when When empty, thread From other threads The algorithm steals tasks from the tail of the queue to achieve dynamic load balancing. Various heterogeneous operational metrics are scheduled by the work-stealing algorithm and then fed into the decision network in parallel. The input tensor dimension is... The temporal difference layer is used to calculate the feature difference between adjacent time steps, and the spatial topological dynamic velocity tensor is obtained. The calculation formula is as follows:
[0075] ;
[0076] In the formula, For batch Middle node In feature dimension The instantaneous rate of change on For batch ,node Feature Dimension At the current time step eigenvalues, For the same position at the previous time step The eigenvalues of the two values have the same dimension, and the dimension of the difference is consistent with that of the eigenvalues. The gated routing module takes the instantaneous cash flow gap and resource-constrained physical boundary of each node as input, and generates gated routing coefficients through a multilayer perceptron. The routing weight calculation formula is as follows:
[0077] ;
[0078] In the formula, For learnable route weight matrix, For the number of expert subnetworks, The physical constraint strength coefficient ranges from 0.1 to 2.0. It is determined by taking the value that results in the fastest convergence of the loss function through multiple comparative experiments on the validation set. As a supply-demand conservation mapping function, the funding gap and resource constraints are mapped to... A dimensionless physical penalty term. The function normalizes the result of a linear combination into a probability distribution. For each node pair The routing weight vector of each expert subnetwork, where the sum of its components is 1 and dimensionless. Gated routing is based on... Selectively activate the top 3 expert subnetworks with the highest physical sensitivity, and the output dimension is... ,in This provides the feature output dimension for the expert subnetwork. A long-span residual dense transition matrix is constructed internally within the network. This allows input features to bypass inactive expert paths and flow directly to the output layer, preserving all input feature information. The Fibonacci golden section optimization algorithm generates a set of golden ratio sampling points during the inference pre-processing phase. ,in For the first The sampling points of the next iteration are based on the golden ratio. An efficient single-peak search is performed in the routing weight solution space. The temporary routing gate bias scaling weight for each generation is:
[0079] ;
[0080] In the formula, For the first The temporary routing gating bias scaling weights obtained in the next iteration, and With the same dimensions, it is directly used as the bias scaling factor for routing gating, with a cumulative iteration count not exceeding 7. When a better path is found, the corresponding neuron's gating activation state is immediately rewritten. Activation intensity is below the lower bound threshold. The connection path is interrupted by cascading gating. The value ranges from 0.01 to 0.1, and was determined through ablation experiments analyzing the changes in inference accuracy and computational complexity before and after truncation. Memory allocation for each layer is based on the real-time variance statistics of the feature map. Dynamically allocated proportionally, layer Video memory quota satisfy:
[0081] ;
[0082] In the formula, For the first Layer memory quota, This is the upper limit of the overall video memory constraint. For the first Real-time variance of layer feature maps For summation indexing, the units on both the right and left sides of the equation are in memory capacity units, ensuring that high-information feature layers receive priority access to storage resources. The final output is the impact prediction tensor for each business problem.
[0083] The specific implementation method of step S04 is as follows.
[0084] By inputting the operational impact prediction tensor into the minimum-cost maximum-flow resource scheduling algorithm, operational crises at each level are abstracted as source nodes of a directed network. Abstracting the enterprise's available rescue resources into a collection point Each directed arc in the network Maximum capacity The unit cost of an arc corresponds to the upper limit of resource capacity. This addresses the additional operational losses caused by delayed rescue efforts. The algorithm iteratively finds the shortest augmenting path (with cost as the path length), pushing the maximum feasible traffic along this path each time, until no further augmenting path exists from the source to the sink. The minimum total operational loss obtained is:
[0085] ;
[0086] In the formula, To minimize total operating loss, For a set of directed arcs, For arc The optimal traffic allocation value on the surface. and The product is expressed in units of operating loss, and the summation of each arc... The unit of measurement is also the unit of operational loss. The degree of flow satisfaction at each source point is used to deduce the priority sequence of each operational crisis, and output the optimal resource scheduling plan.
[0087] The specific implementation method of step S05 is as follows.
[0088] Adaptive dynamic activation intensity adjustment function Based on the current batch prediction error rate Average information entropy of input feature tensors Real-time available video memory ratio of the current device Calculate the comprehensive adjustment index:
[0089] ;
[0090] In the formula, , , These are weighting coefficients; the sum of the three is 1, and their values range from [value range missing]. , , This was determined through grid search experiments on the validation set. The prediction error rate for the current batch. The average information entropy of the input feature tensor. This represents the current real-time available video memory percentage for the device. Indicates the level of video memory usage. To achieve a comprehensive adjustment index, all four items are normalized to have the same dimension, becoming dimensionless quantities. When the activation intensity upper bound parameter is adjusted to the high activation range, more expert subnetworks are allowed to participate in inference; when When, maintain a moderate activation range; when At this point, the upper bound parameter of the activation intensity is lowered to the low activation range, triggering cascaded gating to truncate low-contribution expert subnetworks, reducing the computational cost of forward propagation. The low-rank matrix factorization operator decomposes the weight matrix... Decomposed into:
[0091] ;
[0092] In the formula, , , , These are the row and column dimensions of the weight matrix, respectively. It is a low-rank dimension, with values ranging from 10% to 30% of the original matrix rank. Different dimensions are tested on the target edge device. The inference delay and accuracy loss curves are determined. , All units of measurement are consistent and are network weight units. Only core inference branches are deployed at the edge, using cascading gating to intercept worthless inference streams and achieve on-demand computing power allocation.
[0093] The specific implementation method of step S06 is as follows.
[0094] The priority decision sequence and resource scheduling scheme are output to the operation and management platform, dynamically triggering the priority update loop. This is done while predicting the error rate in real time. Exceeding the backtracking trigger threshold At that time, the internal state dynamic backtracking of the gating network is initiated, and the gating parameters are... Rollback to the previous time step status In order to revise the decision, The value ranges from 5% to 20%. The optimal value is determined by statistically analyzing the joint distribution of decision accuracy and backtracking frequency under different thresholds through multiple rounds of cross-validation on historical business data.
[0095] To better understand and implement this invention, the following is a specific application scenario of this invention, Example 2:
[0096] This embodiment uses a chain retail enterprise with 12 directly operated stores as the test scenario. The method described in this invention is used to construct a problem priority decision model, and priority decision and resource scheduling are verified for multiple types of abnormal business events that occur concurrently within a certain working day.
[0097] The raw data collection covered 12 store nodes across 30 consecutive time steps (each step 10 minutes apart, for a total of 5 hours) for financial indicators (including half-hour sales and cash flow deficit ratio), customer flow data (including visitor count and conversion rate), and public opinion data (including number of complaints and mean sentiment polarity). This resulted in a total of [number of dimensions missing]. The original feature tensor is given, where 18 represents the total feature dimension after concatenating various indicators. After pre-space orthogonalization alignment, various indicators are uniformly projected onto the 18-dimensional orthogonal feature space. After dimensional normalization, the component magnitudes of each indicator tend to be consistent, effectively eliminating the problem of dimensional disparity between public opinion sentiment polarity (original range -1 to 1) and turnover (original range tens of thousands to hundreds of thousands of yuan), ensuring that the feature dimensions processed by each thread in subsequent multi-threaded parallel inference are completely aligned.
[0098] In the real-time streaming data access phase, the sliding memory window length was set to 512 data sample points, corresponding to approximately 85 minutes of operational data in this scenario. The memory circular buffer was divided into 8 physically isolated partitions, and the cold data swapping time threshold was set to 2 seconds. During the test, the average processing blocking rate of data access was 0.3%, meeting the latency requirements of real-time inference. The cascaded pre-dimensionality reduction processing used a 3-stage serial dimensionality reduction module to progressively compress the 18-dimensional features to 12, 8, and 6 dimensions. Each stage of dimensionality reduction retained only the feature components that contributed most to subsequent inference, compressing the data volume to 33% of the original, significantly reducing the computational burden of subsequent inference.
[0099] The structure of a spatiotemporal evolution-constrained sparse activation hybrid expert decision network (hereinafter referred to as the decision network) is as follows: Figure 2 As shown, the input tensor has a dimension of (Batch size) Time step Number of nodes Feature Dimension Temporal difference layering subtracts the feature tensors of adjacent time steps, outputting a spatial topological dynamics velocity tensor with dimension . This captures the instantaneous rate of change in the operational status of each store. In this embodiment, the decision network is configured with 8 expert subnetworks, and the gating routing module determines the routing based on routing weights. (in (Value 0.8) Selectively activate the three expert subnetworks with the highest physical sensitivity, and the output dimension is... ( The long-span residual dense jump matrix directly connects the input features to the output layer, ensuring that feature information on inactive expert paths is not discarded. The Fibonacci golden section optimization algorithm converges to the current optimal gating state after 5 iterations in the route weight solution space, with a convergence speed significantly faster than traditional backpropagation. The output dimension of the impact prediction tensor for each store's operational problems is... (5 types of operational issues), and the predicted impact scores for each type of issue are shown in Table 1.
[0100] Table 1. Evaluation Scoring Table for Predicting the Impact of Operational Issues on Each Store
[0101]
[0102] Table 1 lists the predicted impact scores of five types of operational issues for some stores. The higher the score, the greater the potential impact of the issue on operations at the current moment.
[0103] The impact prediction score tensor from Table 1 is input into the minimum-cost maximum-flow resource scheduling algorithm. Various operational crises are used as the source points of a directed network, and the four types of resources available to the enterprise on a given day (emergency funds, inspection personnel, replenishment vehicles, and public opinion management personnel) are used as the sink points. The capacity limit of each directed arc is set based on the enterprise's available resources on that day, and the unit cost of the arc is converted into additional operational losses due to delayed response based on the impact prediction score. After seven augmenting path iterations, the algorithm reaches maximum flow and outputs the optimal resource scheduling scheme and the priority sequence of each operational problem, as shown in Table 2.
[0104] Table 2 Optimal Resource Scheduling Scheme and Priority Sequence Table
[0105]
[0106] As shown in Table 2, the minimum cost maximum flow algorithm prioritizes public opinion crisis (store 04, score 0.85) as the highest priority under the constraint of total resource amount, which is highly consistent with the distribution trend of the impact prediction score. At the same time, it avoids the conflict phenomenon of repeated allocation of similar resources in the same area, which is common in manual allocation. The decision results have traceable quantitative basis.
[0107] During the edge deployment phase, the low-rank matrix factorization operator will determine the maximum weight matrix (original dimension) in the decision network. Decomposed into dimensions and The product of two low-rank matrices, with a low-rank dimension of 32 (approximately 12.5% of the original matrix's rank), reduces the forward propagation floating-point computation cost from... Second drop This reduces the computational load to 25% of the original. The adaptive dynamic activation intensity adjustment function calculates the comprehensive adjustment index in this round of inference. The network is in a moderate activation range, with the number of activated expert subnetworks remaining at 3. Cascaded gating truncates connection paths with activation strength below 0.05. During inference, truncated, worthless inference flows account for 38% of the total connection paths. Figure 3 As shown, with As the value increases, the proportion of truncated paths gradually increases, the forward propagation time curve tends to flatten, and the efficiency of edge computing power allocation continues to improve.
[0108] Real-time prediction error rate monitoring results, such as Figure 4As shown, the prediction error rate remained below 4% in the first 20 time steps, and the backtracking threshold was not triggered (set to 10% in this scenario). In the 21st time step, a sudden supply chain interruption at store 03 caused a jump in the input data, and the prediction error rate briefly rose to 13%, exceeding the backtracking threshold. The system automatically initiated dynamic backtracking of the gating network's internal state, rolling back the gating parameters to the state of the 20th time step. The error rate recovered to below 5% in the 22nd time step. Figure 4 As shown in the curve trend after backtracking correction, the backtracking mechanism effectively suppressed the interference of abnormal data on the continuity of decision-making.
[0109] In this embodiment, the decision network adopts a hybrid expert structure of time-difference hierarchical structure, gated routing, and sparse activation. The data flow is as follows: input tensor Output after timing difference layering The velocity tensor is then used by the gated routing module to select three expert subnetworks for output. The final output of the fully connected layer affects the prediction tensor. This structure can be replaced by the following scheme: Scheme 1, using a graph attention network to replace the topology transmission part of the high-order spatiotemporal manifold topology transmission network, with the input also being... The graph attention network dynamically calculates the attention weights between nodes at each time step using a tensor, and outputs... The node characteristics are then processed through temporal difference layering, and the subsequent gated routing and expert subnetwork structure remain unchanged. The graph attention mechanism and topology transmission mechanism are highly equivalent in terms of information propagation between nodes, which is an equivalent replacement that can be implemented by ordinary people in the field without creative effort; Scheme 2, uses a temporal convolutional network to replace temporal difference layering, input Tensors are subjected to temporal convolution operations to extract multi-scale temporal features along the time axis, outputting... The feature tensors, temporal convolutional networks, and temporal difference hierarchies all aim to capture dynamic changes on the time axis. The convergence objective of the loss function is consistent with the main scheme, so they can be directly replaced. Figure 5 This is a diagram illustrating the data flow in the decision-making network. Figure 6 This is a comparison chart of the distribution of feature components of each index before and after the pre-space orthogonalization alignment process.
[0110] It should be noted that the business data, such as store traffic data, user sentiment data, and operating financial indicators involved in this application embodiment, have all undergone compliant authorization, legal collection, necessary de-identification, and secure transmission processes in accordance with relevant laws and regulations, and the use of the data complies with the requirements of data protection laws and regulations.
[0111] It should be noted that the variables involved in this invention are explained in detail in Tables 3 and 4.
[0112] Table 3. Variable Explanation Table (Part 1)
[0113]
[0114] Table 4. Variable Explanation Table (Part Two)
[0115]
[0116] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in the present invention should be included within the scope of protection of the present invention.
Claims
1. A method for constructing a problem-priority decision-making model that considers the prediction of operational impact, characterized in that, Includes the following steps: Collect multi-source heterogeneous operational data, including financial indicators, passenger flow data and public opinion data. After pre-spatial orthogonalization alignment processing, construct a unified feature tensor and send it into a high-order spatiotemporal manifold topology transmission network to complete spatiotemporal feature decoupling and encoding. A sliding memory window dynamic pruning mechanism is used to perform cascaded pre-dimensionality reduction processing on real-time streaming business data. Combined with physical isolation technology of memory ring buffer, hot data is kept in the background and cold data is swapped out in seconds, ensuring non-blocking access to high-concurrency data streams. A lock-free circular queue and asynchronous multi-threaded parallel distribution architecture is constructed. A work-stealing algorithm is used to schedule various heterogeneous business indicators, driving a spatiotemporal evolution-constrained sparse activation hybrid expert decision network to complete parallel inference of multiple indicators and output the impact prediction tensor of each business problem. The minimum cost maximum flow resource scheduling algorithm with influence prediction tensor input is used to establish a directed arc network with capacity and cost, taking each operational crisis as the source and the enterprise's allocable resources as the sink. The optimal resource scheduling scheme is solved and the priority sequence of each problem is derived in reverse. Based on the adaptive lightweight parameter update strategy, and combined with the low-rank matrix factorization operator to compress the forward propagation computation of the spatiotemporal evolution constraint sparse activation hybrid expert decision network, only the core inference branch is deployed at the edge, and cascade gating is used to intercept worthless inference flow to achieve on-demand computing power allocation. The priority decision sequence and resource scheduling scheme are output to the operation and management platform to dynamically trigger the priority update loop. When the real-time prediction error rate exceeds the backtracking trigger threshold, the internal state of the gating network is dynamically backtracked to roll back the gating parameters of the previous time step to correct the decision.
2. The method for constructing a problem-priority decision-making model considering operational impact prediction as described in claim 1, characterized in that, The aforementioned pre-space orthogonal alignment process specifically refers to projecting heterogeneous indicators with inconsistent dimensions, such as financial indicators, passenger flow data, and public opinion data, onto the same orthogonal feature space before multi-threaded computation. This eliminates differences in dimensions and magnitudes, ensuring that each indicator has a consistent feature dimension representation in subsequent parallel computation, thereby avoiding deadlocks and waiting in multi-threaded computations due to dimension misalignment.
3. The method for constructing a problem-priority decision-making model considering operational impact prediction as described in claim 2, characterized in that, The aforementioned high-order spatiotemporal manifold topology transmission network specifically refers to a network structure that maps geographic spatial business entities and their dynamic temporal changes to a high-dimensional orthogonal feature space. It utilizes multidimensional difference operators to adaptively remove spatiotemporal coupling redundancy, thereby achieving spatiotemporal feature decoupling. The time axis and spatial topology of the business data are jointly modeled as a high-dimensional manifold, and then the features are propagated between nodes through a topology transmission mechanism.
4. The method for constructing a problem-priority decision-making model considering operational impact prediction as described in claim 3, characterized in that, The sliding memory window dynamic pruning mechanism specifically refers to the dynamic adjustment of the window length to extract the data segments required for current processing from the continuously incoming real-time streaming business data, discarding expired data outside the window, and controlling the instantaneous memory usage within a preset upper limit. The window length value is determined by stress testing at different concurrency pressure levels on the target hardware platform and analyzing the memory peak curve.
5. The method for constructing a problem-priority decision-making model considering operational impact prediction as described in claim 4, characterized in that, The aforementioned memory ring buffer physical isolation technology specifically refers to dividing memory into several fixed-size physical isolation partitions to form a ring-shaped buffer pool. Hot data preferentially occupies the partitions closer to the computing core, while cold data is automatically mapped to the outer partitions and swapped out when the access interval exceeds the cold data swap-out time threshold. The cold data swap-out time threshold is determined by statistically analyzing the processing blocking rate under each swap-out time threshold.
6. The method for constructing a problem-priority decision-making model considering operational impact prediction as described in claim 5, characterized in that, The lock-free circular queue refers to a first-in-first-out data structure that does not rely on mutex locks and maintains read and write pointers with atomic operations. Multiple producer threads and consumer threads complete data enqueueing and dequeueing through atomic comparison and exchange operations, completely eliminating waiting and deadlock caused by lock contention between threads.
7. The method for constructing a problem-priority decision-making model considering operational impact prediction as described in claim 6, characterized in that, The input tensor dimension of the spatiotemporal evolution constrained sparse activation hybrid expert decision network is [B, T, N, D], where B is the batch size, T is the number of time steps, N is the number of nodes, and D is the number of feature dimensions. By calculating the characteristic difference between adjacent time steps through time difference layering, the spatial topological dynamic velocity tensor with dimensions [B, N, D] is obtained, which is used to characterize the instantaneous change rate of the operating status of each node.
8. The method for constructing a problem-priority decision-making model considering operational impact prediction as described in claim 7, characterized in that, The spatiotemporal evolution-constrained sparse activation hybrid expert decision network internally constructs a long-span residual dense jump matrix, enabling input features to bypass inactive expert subnetworks and achieve direct cross-layer flow of features. It also embeds a dynamic routing weight search algorithm based on Fibonacci golden section optimization, which uses the Fibonacci sequence to generate golden section ratio sampling points and performs efficient single-peak search in the routing weight solution space, with a cumulative iteration count of no more than 7 times.
9. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores program instructions, which, when executed in a computer, are used to perform the problem priority decision-making model construction method considering operational impact prediction as described in any one of claims 1-8.
10. A system for constructing a problem-priority decision-making model that considers the prediction of operational impact, characterized in that, The system comprises the computer-readable storage medium of claim 9, wherein the system is a computer, the computer-readable storage medium is disposed within the system, and the system is provided with a microprocessor that executes program instructions stored in the computer-readable storage medium.
