Multi-modal prediction and inventory intelligent control method based on big data fusion

Abstract

The present application relates to the technical field of inventory control, and more particularly to a multi-modal prediction and inventory intelligent control method based on big data fusion, comprising: collecting multi-source heterogeneous data in a cross-border e-commerce supply chain, sequentially performing preprocessing and feature extraction, determining the rationality of the configuration of multi-modal feature resources based on the effective extraction algorithm consumption proportion of multi-modal fusion features; training an initial model, analyzing multi-source heterogeneous data, and statistically analyzing environmental changes; determining the environmental adaptability of the deep learning model, and determining whether to adjust the traversal boundary of counterfactual deduction; generating inventory control instructions, comparing the matching results of target linkage actions and actual execution actions; determining the linkage and coordination capability of prediction and control, and reconstructing the continuous perception prediction anchor point drift of the execution body control. The present application improves the stability of inventory control.

Description

Technical Field

[0001] This invention relates to the field of inventory control technology, and in particular to a multimodal forecasting and intelligent inventory control method based on big data fusion. Background Technology

[0002] In existing technologies, inventory forecasting and control often rely on single historical sales data and traditional statistical algorithms for modeling. The models only learn from historical data and cannot cope with sudden disturbances in new supply chains. Forecasting and inventory control are disconnected, and there are no linkage mismatch quantitative correction methods. Inventory preparation calculations only consider market demand and ignore warehouse capacity limitations. They cannot dynamically correct forecast deviations, making it difficult to meet the stability requirements of inventory control.

[0003] Chinese Patent Publication No. CN121707444A discloses a closed-loop management method and system for reverse logistics in cross-border e-commerce of apparel, integrating big data and intelligent sorting. The method includes: a dynamic prediction step for return risks based on machine learning, wherein, during the user order generation stage, multi-dimensional user front-end behavior data, historical transaction data, and product attribute data are collected and integrated in real time; a pre-trained machine learning prediction model is invoked to calculate a quantified return probability value for the products in the order; and an intelligent identification and sorting step for returned goods based on multimodal perception information, wherein, when the returned package enters the automated processing line, SKU information is identified using RFID technology; subsequently, machine vision technology is used... The returned goods undergo status quality inspection; and based on the SKU information identification results and the status quality inspection results, the returned goods are automatically physically sorted; a deep analysis and closed-loop feedback step based on structured return data is implemented, wherein the process data generated in the intelligent identification and sorting steps is collected in real time and structured and stored to form a return reason database; the database is analyzed using data mining algorithms to identify the root causes of high return rates; and, after identifying the root causes of high return rates, control instructions containing correction parameter suggestions are automatically generated and pushed to the corresponding positive supply chain information system module through a data interface to automatically correct the front-end business logic. Therefore, the aforementioned closed-loop management method and system for reverse logistics in cross-border e-commerce of apparel, which integrates big data and intelligent sorting, suffers from a lack of stability in inventory control due to relying solely on historical return data for risk prediction, failing to dynamically adjust inventory preparation benchmarks in real time based on warehouse capacity constraints and supply-demand timeline deviations, and lacking a predictive anchor drift perception and reconstruction mechanism. Summary of the Invention

[0004] To address this issue, the present invention provides a multimodal forecasting and intelligent inventory control method based on big data fusion, which overcomes the problem in the prior art that the inventory control stability decreases because the inventory preparation relies solely on historical sales forecasts and cannot be dynamically adjusted based on warehouse capacity constraints and supply and demand timeline deviations.

[0005] To achieve the above objectives, this invention provides a multimodal forecasting and intelligent inventory control method based on big data fusion, comprising: Collect multi-source heterogeneous data from the cross-border e-commerce supply chain, perform preprocessing and feature extraction on the multi-source heterogeneous data in sequence to obtain multimodal fusion features, and determine the rationality of multimodal feature resource allocation based on the effective extraction computing power consumption ratio of multimodal fusion features. The initial model is trained based on the multimodal fusion features to obtain a deep learning model. The multi-source heterogeneous data is analyzed based on the deep learning model to obtain prediction results. Simultaneously, environmental anomalies in the supply chain are monitored, and the environmental anomalies are statistically analyzed to obtain the environmental anomaly perception response coverage. The environmental adaptability of the deep learning model is determined based on the perceived response coverage of the environmental anomalies. If the environmental adaptability of the deep learning model does not meet the requirements, it is determined whether to adjust the traversal boundary of the counterfactual inference. Based on the prediction results, an inventory control instruction is generated, and an inventory control action is executed based on the inventory control instruction. The matching results of the target linkage action and the actual executed action are compared to obtain the mismatch ratio of the linkage action. The coordination capability of prediction and control is determined based on the mismatch ratio of the aforementioned linkage actions. If the coordination capability of prediction and control does not meet the requirements, the continuous perception prediction anchor point drift of the actuator regulation is reconstructed.

[0006] Furthermore, the rationality of multimodal feature resource allocation is determined based on the proportion of effective extraction computing power consumption for multimodal fusion features, including: If the effective extraction computing power consumption ratio of the multimodal fusion features is less than or equal to the preset ratio, then the configuration of multimodal feature resources is determined to be reasonable and meets the requirements. If the effective extraction computing power consumption ratio of the multimodal fusion features is greater than the preset ratio, then the configuration of multimodal feature resources is determined to be unreasonable.

[0007] Furthermore, if the configuration of the multimodal feature resources does not meet the requirements, the environmental adaptability of the deep learning model is determined based on the coverage of the perception response to environmental changes.

[0008] Furthermore, the environmental adaptability of the deep learning model is determined based on the coverage of the perceived response to environmental anomalies, including: If the coverage rate of the perception response to the environmental anomaly is greater than or equal to the preset first coverage rate, then the environmental adaptability of the deep learning model is determined to meet the requirements. If the coverage rate of the perception response to environmental anomalies is less than the preset first coverage rate, then it is determined that the environmental adaptability of the deep learning model does not meet the requirements.

[0009] Further, determine whether to adjust the ergodic boundaries of the counterfactual inference, including: If the coverage of the perception response to the environmental anomaly is less than the preset first coverage but greater than or equal to the preset second coverage, then it is determined that the traversal boundary of the counterfactual inference needs to be increased. If the coverage of the perceived response to the environmental anomaly is less than the preset second coverage, then it is determined that there is no need to increase the traversal boundary of the counterfactual inference.

[0010] Furthermore, the increase in the traversal boundary of the counterfactual inference is determined by the difference between the preset first coverage rate and the perceived response coverage rate of the environmental anomaly.

[0011] Furthermore, based on the condition that the sensing response coverage of the environmental anomaly is less than the preset second coverage, the determination of whether the linkage coordination capability of prediction and control meets the requirements is based on the mismatch ratio of the linkage action.

[0012] Furthermore, the determination of whether the coordination capability between prediction and control meets the requirements is based on the mismatch ratio of the linked actions, including: If the mismatch ratio of the linked action is less than or equal to the preset ratio, then the linkage coordination capability of prediction and control is determined to meet the requirements. If the mismatch ratio of the linked action is greater than the preset ratio, it is determined that the linkage coordination capability of prediction and control does not meet the requirements, and the continuous perception prediction anchor point drift of the actuator regulation is reconstructed.

[0013] Furthermore, the mismatch ratio of the linkage action is determined based on the ratio of the number of linkage mismatch actions per unit time to the total number of linkage executions.

[0014] Furthermore, the continuous perception prediction anchor point drift of the execution body regulation is reconstructed by reconstructing the timing constraints of the regulation command and the regulation execution strategy based on the degree of deviation of the prediction benchmark value.

[0015] Compared with the prior art, the beneficial effects of the present invention are as follows: the method described in the embodiments of the present invention determines whether the configuration of multimodal feature resources meets the requirements based on the effective extraction computing power consumption ratio of multimodal fusion features, eliminates redundant feature operations, balances computing power load to ensure real-time prediction, adjusts the traversal boundary of counterfactual inference based on the perception response coverage of environmental anomalies, expands the extreme mutation simulation scenario, improves the model's predictive adaptability to unknown supply chain emergencies, and reconstructs the continuous perception prediction anchor point drift of the execution body control based on the mismatch ratio of linkage actions, offsets the deviation caused by system latency, and enables the prediction results to align with actual inventory operations, thereby improving the stability of inventory control.

[0016] Furthermore, the method described in this invention determines whether the configuration of multimodal feature resources meets the requirements by setting a preset proportion. Since all multimodal fusion features are extracted in full, a large number of unstructured features are invalidally calculated, the system's computing power load exceeds the standard, and the online real-time prediction latency exceeds the limit. By determining the configuration of multimodal feature resources, redundant calculations with low marginal contribution can be eliminated, the released computing power resources can be redistributed, and the system's computing power load can be controlled within the carrying capacity of edge devices, thereby further improving the stability of inventory control.

[0017] Furthermore, the method described in this invention adjusts the traversal boundary of counterfactual inference by setting a preset first coverage rate and a preset second coverage rate. Since the model learns patterns from historical data, it cannot identify entirely new events that have never appeared in history. When a sudden change occurs in the supply chain environment outside the distribution, the model cannot make a correct prediction. By adjusting the traversal boundary of counterfactual inference, the simulation coverage can be expanded, and extreme sudden event scenarios not included in historical samples can be supplemented. This strengthens the ability to adapt to and predict unknown supply chain changes, and further improves the stability of inventory control.

[0018] Furthermore, the method described in this invention reconstructs the drift of the continuous perception prediction anchor point of the execution body control by setting a preset proportion. Due to the inconsistency between the adjustment pace of the dynamic data backtracking time domain and the trend constraint strength, there is a delay in instruction transmission and action execution, which causes the prediction output to be misaligned and mismatched with the inventory control action. By reconstructing the drift of the continuous perception prediction anchor point of the execution body control, the prediction benchmark offset can be corrected in real time, the timing constraints and execution strategies can be dynamically adapted, and the coordination deviation caused by system delay can be offset, so that the prediction result can be aligned with the actual inventory operation action, thereby further improving the stability of inventory control. Attached Figure Description

[0019] Figure 1 This is an overall flowchart of the multimodal prediction and intelligent inventory control method based on big data fusion, as described in an embodiment of the present invention. Figure 2 This is a flowchart illustrating the process of determining the rationality of multimodal feature resource allocation in the multimodal prediction and intelligent inventory control method based on big data fusion according to an embodiment of the present invention. Figure 3 This is a flowchart illustrating the logic of determining whether to adjust the traversal boundary process of counterfactual inference in an embodiment of the present invention's multimodal prediction and intelligent inventory control method based on big data fusion. Figure 4 This is a flowchart illustrating the process of reconstructing the anchor point drift of continuous perception prediction for the control of the execution entity using the multimodal prediction and intelligent inventory control method based on big data fusion, as described in an embodiment of the present invention. Detailed Implementation

[0020] To make the objectives and advantages of the present invention clearer, the present invention will be further described below with reference to embodiments; it should be understood that the specific embodiments described herein are merely for explaining the present invention and are not intended to limit the present invention.

[0021] Preferred embodiments of the present invention will now be described with reference to the accompanying drawings. Those skilled in the art should understand that these embodiments are merely illustrative of the technical principles of the present invention and are not intended to limit the scope of protection of the present invention.

[0022] Please see Figure 1 As shown, it is the overall flowchart of the multimodal prediction and intelligent inventory control method based on big data fusion in an embodiment of the present invention.

[0023] This invention provides a multimodal forecasting and intelligent inventory control method based on big data fusion, comprising: Step S1: Collect multi-source heterogeneous data from the cross-border e-commerce supply chain, perform preprocessing and feature extraction on the multi-source heterogeneous data in sequence to obtain multimodal fusion features, and determine the rationality of multimodal feature resource allocation based on the effective extraction computing power consumption ratio of multimodal fusion features. Step S2: Train the initial model based on the multimodal fusion features to obtain a deep learning model; analyze the multi-source heterogeneous data based on the deep learning model to obtain prediction results; and simultaneously monitor environmental anomaly events in the supply chain; and statistically analyze the environmental anomaly events to obtain the environmental anomaly perception response coverage rate. Step S3: Determine the environmental adaptability of the deep learning model based on the perceived response coverage of the environmental anomaly. If the environmental adaptability of the deep learning model does not meet the requirements, determine whether to adjust the traversal boundary of the counterfactual inference. Step S4: Generate inventory control instructions based on the prediction results, execute inventory control actions based on the inventory control instructions, and compare the matching results of the target linkage action and the actual executed action to obtain the mismatch ratio of the linkage action. Step S5: Determine the linkage coordination capability of prediction and control based on the mismatch ratio of the linkage action. If the linkage coordination capability of prediction and control does not meet the requirements, reconstruct the continuous perception prediction anchor point drift of the actuator control. Specifically, multi-source heterogeneous data includes order cancellations, inventory balance, and order placement time.

[0024] Specifically, preprocessing includes cleaning, noise reduction, fusion, and normalization.

[0025] Specifically, the multimodal fusion features include the proportion of order cancellations to total orders, current available inventory, and the time interval between adjacent orders.

[0026] Specifically, the process of training an initial model based on multimodal fusion features to obtain a deep learning model involves constructing a time-series sample set of multimodal fusion features and dividing it into a training set, a validation set, and a test set. The training set is then input into the initial model for forward training and parameter iteration. Model parameters are tuned based on the validation set, and model performance is verified based on the test set to obtain the deep learning model.

[0027] Specifically, the initial model is a basic model framework that already has the ability to analyze and interpret multimodal fusion features but has not yet undergone parameter iteration optimization using data feature samples.

[0028] Specifically, the deep learning model can be a deep convolutional neural network model, a long short-term memory neural network model, or a gated recurrent unit model, with the preferred embodiment being a long short-term memory neural network model.

[0029] Specifically, the process of analyzing multi-source heterogeneous data based on a deep learning model to obtain prediction results involves inputting the multi-source heterogeneous data into the deep learning model, performing fusion calculations on the multimodal fusion features, analyzing the current state, and outputting the corresponding prediction results.

[0030] Specifically, the forecast results include the expected number of order cancellations in the next week, the estimated time when stockouts will occur, and the total number of orders during peak periods.

[0031] Specifically, the process of adjusting inventory based on prediction results involves comparing and integrating the prediction results output by the deep learning model with the current real-time inventory situation item by item, generating replenishment instructions to adjust inventory in real time.

[0032] Specifically, replenishment instructions include SKU identifier, replenishment quantity, and expected delivery time.

[0033] Specifically, the traversal boundary of counterfactual deduction is used to traverse the maximum range of values ​​for simulated supply chain disturbance parameters, demand fluctuation ranges, and instruction timing combinations without changing the actual historical business.

[0034] In implementation, the method described in this embodiment of the invention determines whether the configuration of multimodal feature resources meets the requirements based on the effective extraction computing power consumption ratio of multimodal fusion features, eliminates redundant feature operations, balances computing power load to ensure real-time prediction, adjusts the traversal boundary of counterfactual inference based on the perception response coverage of environmental anomalies, expands the extreme mutation simulation scenario, improves the model's predictive adaptability to unknown supply chain emergencies, and reconstructs the continuous perception prediction anchor point drift of the execution body control based on the mismatch ratio of linkage actions, offsetting the deviation caused by system latency, so that the prediction results can be aligned with the actual inventory operation actions, thereby improving the stability of inventory control.

[0035] Please continue reading. Figure 2 As shown, it is a logical flowchart of the process of determining the rationality of multimodal feature resource allocation in the multimodal prediction and intelligent inventory control method based on big data fusion according to an embodiment of the present invention.

[0036] In this embodiment, the proportion of computing power consumed in the effective extraction of multimodal fusion features is used to quantify the rationality of the allocation of multimodal feature resources. The proportion of computing power consumed for effective extraction of multimodal fusion features is determined by the ratio of computing power consumed for effective feature extraction to the total computing power consumed for feature extraction. Effective extraction refers to feature extraction in which the extracted features are ultimately used by the model for inference and decision-making, and which make a quantifiable positive contribution to the prediction results.

[0037] Specifically, the determination of whether the allocation of multimodal feature resources meets the requirements is based on the proportion of effective extraction computing power consumption of multimodal fusion features, including: If the effective extraction computing power consumption ratio of the multimodal fusion features is less than or equal to the preset ratio, then the configuration of multimodal feature resources is determined to be reasonable and meets the requirements. If the effective extraction computing power consumption ratio of the multimodal fusion features is greater than the preset ratio, then the configuration of multimodal feature resources is determined to be unreasonable.

[0038] It is understandable that the effective extraction computing power consumption ratio of multimodal fusion features is a value determined by those skilled in the art based on the statistical distribution of the effective extraction computing power consumption ratio of multimodal fusion features in historical successful control records, combined with the configuration of multimodal feature resources corresponding to different ratios.

[0039] Preferably, the proportion of computational power consumed in the effective extraction of multimodal fusion features is negatively correlated with the rationality of the allocation of multimodal feature resources, wherein, The preset percentage represents the effective computing power consumption ratio for multimodal fusion feature extraction, corresponding to the resource configuration meeting the qualification requirements in historical statistical distribution. When the effective computing power consumption ratio for multimodal fusion feature extraction is less than or equal to this value, the resource configuration is basically consistent with the theoretical configuration, and the effective computing power consumption ratio for multimodal fusion feature extraction can accurately reflect the actual situation of resource configuration. Therefore, the rationality of multimodal feature resource configuration meets the requirements. When the effective computing power consumption ratio for multimodal fusion feature extraction is greater than this value, there is a large deviation between the resource configuration and the theoretical configuration, indicating that the system's computing power load exceeds the standard and the online real-time prediction latency exceeds the limit. Therefore, the continuity of control command output does not meet the requirements.

[0040] The specific value of the preset percentage can be set according to the baseline of computing power consumption in historical control and the requirements of those skilled in the art for the rationality of multimodal feature resource configuration, and will not be elaborated here; Among them, the baseline for setting computing power consumption in historical control is a standardized reference value pre-determined by those skilled in the art based on the average computing power consumption level under stable operation under historical successful control, taking into account the total computing power capacity of the equipment hardware, the normalized operating load extracted from multimodal features, the maximum real-time prediction latency allowed by the business, and the standard theoretical resource allocation.

[0041] In implementation, the method described in this embodiment of the invention determines whether the configuration of multimodal feature resources meets the requirements by setting a preset proportion. Since all multimodal fusion features are extracted in full, a large number of unstructured features are invalidally calculated, the system's computing power load exceeds the standard, and the online real-time prediction latency exceeds the limit. By determining the configuration of multimodal feature resources, redundant calculations with low marginal contribution can be eliminated, the released computing power resources can be redistributed, and the system's computing power load can be controlled within the carrying capacity of edge devices, further improving the stability of inventory control.

[0042] Please continue reading. Figure 3 As shown, it is a logical flowchart of the process of determining whether to adjust the traversal boundary of the counterfactual inference in the multimodal prediction and intelligent inventory control method based on big data fusion according to an embodiment of the present invention.

[0043] Specifically, if the configuration of the multimodal feature resources does not meet the requirements, the environmental adaptability of the deep learning model is determined based on the coverage of the perception response to environmental changes.

[0044] Specifically, determining whether the environmental adaptability of a deep learning model meets the requirements is based on the coverage of the perceived response to environmental changes, including: If the coverage rate of the perception response to the environmental anomaly is greater than or equal to the preset first coverage rate, then the environmental adaptability of the deep learning model is determined to meet the requirements. If the coverage rate of the perception response to environmental anomalies is less than the preset first coverage rate, then it is determined that the environmental adaptability of the deep learning model does not meet the requirements.

[0045] Specifically, determining whether to adjust the ergodic boundaries of counterfactual inference includes: If the coverage of the perception response to the environmental anomaly is less than the preset first coverage but greater than or equal to the preset second coverage, then it is determined that the traversal boundary of the counterfactual inference needs to be increased. If the coverage of the perceived response to the environmental anomaly is less than the preset second coverage, then it is determined that there is no need to increase the traversal boundary of the counterfactual inference.

[0046] In implementation, the preset first coverage rate is greater than the preset second coverage rate, where: If the coverage rate of the perception response to environmental changes is greater than or equal to the preset first coverage rate, the corresponding situation is: the environmental adaptability of the deep learning model meets the requirements. If the coverage of the perception response to environmental anomalies is less than the preset first coverage but greater than or equal to the preset second coverage, the corresponding situation is: because the model learns patterns through historical data, it cannot identify entirely new events that have never appeared in history. When a sudden change occurs in the supply chain environment that is outside the distribution, the model cannot make a correct prediction. If the coverage rate of the perception response to environmental anomalies is less than the preset second coverage rate, the corresponding situation is as follows: due to the inconsistency between the adjustment pace of the dynamic data backtracking time domain and the trend constraint strength, there is a delay in instruction transmission and action execution, which causes the prediction output and inventory control actions to be misaligned and mismatched.

[0047] It is understood that, given the preset first coverage rate and preset second coverage rate, which are statistical distributions of the perceived response coverage rate of environmental anomalies in historical records of successful prediction and control by a person skilled in the art, and values ​​determined by combining the environmental adaptability corresponding to different coverage rates, preferably, the perceived response coverage rate of environmental anomalies is positively correlated with the environmental adaptability of the deep learning model, wherein, The preset first coverage rate is the perception response coverage rate of environmental anomalies corresponding to the surge in the proportion of computing power consumption for effective extraction of multimodal fusion features in historical statistical distribution. When the coverage rate is greater than this value, the extraction of multimodal features meets the basic resource allocation, and the proportion of computing power consumption for effective extraction of multimodal fusion features surges. The surge inflection point is determined based on the effective extraction of computing power consumption ratio data from the perception response coverage of environmental anomalies recorded synchronously in statistical history and the multimodal fusion features. The computing power growth rate is calculated in segments and the nodes of sudden growth are identified. The result is determined after verification with a preset computing power consumption base. The computing power consumption baseline is a standard reference value pre-set by those skilled in the art based on the upper limit of system hardware computing power, the conventional extraction of load from multimodal features, and operating costs.

[0048] The preset second coverage rate is the environmental anomaly perception response coverage rate corresponding to the saturation point of the effective extraction computing power consumption ratio of multimodal fusion features in historical statistical distribution. When the required coverage rate is greater than this value, the extraction of multimodal features is close to complete, and the effective extraction computing power consumption ratio of multimodal fusion features reaches the expected upper limit. The saturation point is determined based on data matching the coverage of the perception response to environmental anomalies recorded in historical synchronous records with the proportion of computing power consumed in the effective extraction of multimodal fusion features.

[0049] Specifically, the process of determining the data pairing between the sensing response coverage of environmental anomalies recorded in historical synchronous records and the effective extraction computing power consumption ratio of multimodal fusion features is as follows: historical data are arranged in ascending order according to the sensing response coverage of environmental anomalies, and the growth rate of the effective extraction computing power consumption ratio of multimodal fusion features is statistically analyzed segment by segment. When the growth rate of computing power ratio approaches zero after a certain coverage node, the computing power occupancy reaches the upper limit of computing power corresponding to the preset computing power consumption base, and this node is the saturation point.

[0050] The size of the preset first coverage rate and the preset second coverage rate depends on the number of historical statistical samples and the baseline for setting the effective extraction computing power consumption ratio of multimodal fusion features by those skilled in the art. They can be set according to the actual situation and will not be elaborated here.

[0051] Specifically, the environmental anomaly perception and response coverage rate is the ratio of the number of environmental anomalies that were successfully perceived and responded to to the total number of environmental anomalies that occurred.

[0052] Specifically, the increase in the traversal boundary of the counterfactual deduction is determined by the difference between the preset first coverage rate and the perceived response coverage rate of the environmental anomaly.

[0053] Specifically, the increase in the counterfactual inference traversal boundary is calculated by obtaining a preset first coverage rate and the perceived response coverage rate of environmental anomalies, calculating the difference, and using the difference as the benchmark for the increase in the counterfactual inference traversal boundary. If the difference is greater than 0, it indicates that there is a gap in the perceived response coverage of the current environmental anomalies. The traversal boundary is widened outward according to the magnitude corresponding to the difference. The larger the difference value, the higher the expansion of the traversal boundary. If the difference is equal to or less than 0, it indicates that the perceived response coverage of the current environmental anomalies meets the preset standard. The traversal boundary does not perform the amplification operation and maintains the original range unchanged.

[0054] Specifically, the step of expanding the traversal boundary outward according to the amplitude corresponding to the difference is to use the difference between the preset first coverage rate and the perception response coverage rate of environmental anomalies as the basic amplitude benchmark, pre-establish a mapping rule between the difference value and the boundary expansion amplitude, substitute the obtained difference value into the mapping relationship, and match to obtain a unique corresponding boundary expansion amplitude value.

[0055] In practice, the method described in this embodiment of the invention adjusts the traversal boundary of counterfactual inference by setting a preset first coverage rate and a preset second coverage rate. Since the model learns patterns from historical data, it cannot identify entirely new events that have never appeared in history. When a sudden change occurs in the supply chain environment outside the distribution, the model cannot make a correct prediction. By adjusting the traversal boundary of counterfactual inference, the simulation coverage can be expanded, and extreme sudden event scenarios not included in historical samples can be supplemented. This strengthens the ability to adapt to and predict unknown supply chain changes and further improves the stability of inventory control.

[0056] Please continue reading. Figure 4 The above is a logical flowchart of the process of reconstructing the anchor point drift of the continuous perception prediction of the execution body regulation in the multimodal prediction and intelligent inventory control method based on big data fusion according to an embodiment of the present invention.

[0057] Specifically, under the condition that the sensing response coverage of the environmental anomaly is less than the preset second coverage, the determination of whether the linkage coordination capability of prediction and control meets the requirements is based on the mismatch ratio of the linkage action.

[0058] Specifically, the determination of whether the coordination capability of prediction and control meets the requirements is based on the mismatch ratio of the linked actions, including: If the mismatch ratio of the linked action is less than or equal to the preset ratio, then the linkage coordination capability of prediction and control is determined to meet the requirements. If the mismatch ratio of the linked action is greater than the preset ratio, it is determined that the linkage coordination capability of prediction and control does not meet the requirements, and the continuous perception prediction anchor point drift of the actuator regulation is reconstructed.

[0059] In implementation, if the mismatch ratio of the linkage action is less than or equal to the preset ratio, the corresponding situation is: the linkage coordination capability of prediction and control meets the requirements. If the mismatch ratio of the linkage action is greater than the preset ratio, the corresponding situation is: due to the inconsistent adjustment pace between the dynamic data backtracking time domain and the trend constraint strength, there is a delay in instruction transmission and action execution, which causes the prediction output and inventory control action to be misaligned and mismatched.

[0060] Understandably, in multimodal forecasting and intelligent inventory control methods based on big data fusion, the mismatch ratio of linked actions is used to characterize the linkage coordination capability between forecasting and control. The core logic is to directly reflect the linkage coordination capability between forecasting and control as the mismatch ratio of linked actions, transform the linkage coordination capability between forecasting and control into a quantitative ratio indicator, and compare it with the preset ratio to determine whether the linkage coordination capability has achieved the expected effect, thereby providing a quantitative basis for judging the linkage coordination capability between forecasting and control.

[0061] The preset percentage is a value determined by those skilled in the art based on the statistical distribution of the mismatch percentage of linked actions in historical successful prediction and control records using deep learning models, combined with the linkage coordination situation corresponding to different percentages. Preferably, the mismatch percentage of linked actions is negatively correlated with the linkage coordination capability of prediction and control. The preset percentage is the mismatch percentage of the linkage action corresponding to the saturation point of the preset second coverage rate in the historical statistical distribution. When the percentage is less than this value, the linkage coordination capability meets the basic requirements and the preset second coverage rate reaches the upper limit.

[0062] The size of the preset percentage depends on the statistical distribution of the mismatch percentage of all historical successful predictive control cases, and the historical mismatch percentage threshold corresponding to the preset second coverage saturation point under the distribution. The preset percentage can be set according to the actual situation, and will not be elaborated here.

[0063] Specifically, the mismatch ratio of the linkage action is determined based on the ratio of the number of mismatched linkage actions per unit time to the total number of linkage executions.

[0064] Specifically, the continuous perception prediction anchor point drift of the execution body regulation is reconstructed by reconstructing the timing constraints of the regulation command and the regulation execution strategy based on the degree of deviation of the prediction benchmark value.

[0065] Specifically, the process of reconstructing the timing constraints and execution strategies of control instructions based on the degree of deviation of the prediction benchmark value involves the inventory control execution body using the predicted value output by the model as a fixed prediction anchor point, comparing the real-time deviation between the model prediction value and the supply chain field data, dynamically quantifying the deviation magnitude and drift trend of the prediction anchor point, and correcting the timing matching rules between the prediction domain and the control domain based on the actual drift amount of the anchor point when the prediction anchor point continues to deviate from the actual state.

[0066] Specifically, the prediction anchor is the baseline demand value output by the model, which is used by the inventory control execution entity to carry out inventory preparation scheduling.

[0067] Specifically, the corrected prediction domain is a dynamic prediction range that adapts to the current supply chain situation after dynamically correcting the prediction benchmark, updating the counterfactual inference boundary and multimodal feature computing power configuration based on the real-time drift of the prediction anchor point.

[0068] Specifically, the control domain is the real-time control execution space that relies on the dynamic demand benchmark of the modified prediction domain, combined with the physical constraints of warehouse capacity, system computing power load and linkage action matching rules, to perform inventory replenishment, material allocation and instruction timing correction.

[0069] In practice, the method described in this embodiment of the invention reconstructs the drift of the continuous perception prediction anchor point of the execution body control by setting a preset proportion. Due to the inconsistency between the adjustment pace of the dynamic data backtracking time domain and the trend constraint strength, there is a delay in instruction transmission and action execution, which causes the prediction output to be misaligned and mismatched with the inventory control action. By reconstructing the drift of the continuous perception prediction anchor point of the execution body control, the prediction benchmark offset can be corrected in real time, the timing constraints and execution strategies can be dynamically adapted, and the coordination deviation caused by system delay can be offset, so that the prediction result can be aligned with the actual inventory operation action, further improving the stability of inventory control.

[0070] The technical solution of the present invention has been described above with reference to the preferred embodiments shown in the accompanying drawings. However, it will be readily understood by those skilled in the art that the scope of protection of the present invention is obviously not limited to these specific embodiments. Without departing from the principles of the present invention, those skilled in the art can make equivalent changes or substitutions to the relevant technical features, and the technical solutions after these changes or substitutions will all fall within the scope of protection of the present invention.

Claims

1. A multimodal forecasting and intelligent inventory control method based on big data fusion, characterized in that, include: Collect multi-source heterogeneous data from the cross-border e-commerce supply chain, perform preprocessing and feature extraction on the multi-source heterogeneous data in sequence to obtain multimodal fusion features, and determine the rationality of multimodal feature resource allocation based on the effective extraction computing power consumption ratio of multimodal fusion features. The initial model is trained based on the multimodal fusion features to obtain a deep learning model. The multi-source heterogeneous data is analyzed based on the deep learning model to obtain prediction results. Simultaneously, environmental anomalies in the supply chain are monitored, and the environmental anomalies are statistically analyzed to obtain the environmental anomaly perception response coverage. The environmental adaptability of the deep learning model is determined based on the perceived response coverage of the environmental anomalies. If the environmental adaptability of the deep learning model does not meet the requirements, it is determined whether to adjust the traversal boundary of the counterfactual inference. Based on the prediction results, an inventory control instruction is generated, and an inventory control action is executed based on the inventory control instruction. The matching results of the target linkage action and the actual executed action are compared to obtain the mismatch ratio of the linkage action. The coordination capability of prediction and control is determined based on the mismatch ratio of the aforementioned linkage actions. If the coordination capability of prediction and control does not meet the requirements, the continuous perception prediction anchor point drift of the actuator regulation is reconstructed.

2. The multimodal forecasting and intelligent inventory control method based on big data fusion according to claim 1, characterized in that, The determination of whether the allocation of multimodal feature resources meets the requirements is based on the proportion of effective extraction computing power consumption of multimodal fusion features, including: If the effective extraction computing power consumption ratio of the multimodal fusion features is less than or equal to the preset ratio, then the configuration of multimodal feature resources is determined to be reasonable and meets the requirements. If the effective extraction computing power consumption ratio of the multimodal fusion features is greater than the preset ratio, then the configuration of multimodal feature resources is determined to be unreasonable.

3. The multimodal forecasting and intelligent inventory control method based on big data fusion according to claim 2, characterized in that, If the configuration of the multimodal feature resources does not meet the requirements, the environmental adaptability of the deep learning model is determined based on the coverage of the perception response to environmental changes.

4. The multimodal forecasting and intelligent inventory control method based on big data fusion according to claim 3, characterized in that, The environmental adaptability of a deep learning model is determined by the coverage of its perceived response to environmental anomalies, including: If the coverage rate of the perception response to the environmental anomaly is greater than or equal to the preset first coverage rate, then the environmental adaptability of the deep learning model is determined to meet the requirements. If the coverage rate of the perception response to environmental anomalies is less than the preset first coverage rate, then it is determined that the environmental adaptability of the deep learning model does not meet the requirements.

5. The multimodal forecasting and intelligent inventory control method based on big data fusion according to claim 4, characterized in that, Determine whether to adjust the ergodic boundaries of the counterfactual deduction, including: If the coverage of the perception response to the environmental anomaly is less than the preset first coverage but greater than or equal to the preset second coverage, then it is determined that the traversal boundary of the counterfactual inference needs to be increased. If the coverage of the perceived response to the environmental anomaly is less than the preset second coverage, then it is determined that there is no need to increase the traversal boundary of the counterfactual inference.

6. The multimodal forecasting and intelligent inventory control method based on big data fusion according to claim 5, characterized in that, The increase in the traversal boundary of the counterfactual deduction is determined by the difference between the preset first coverage rate and the perceived response coverage rate of the environmental anomaly.

7. The multimodal forecasting and intelligent inventory control method based on big data fusion according to claim 6, characterized in that, When the coverage of the perception response to the environmental anomaly is less than the preset second coverage, the determination of whether the linkage coordination capability of prediction and control meets the requirements is based on the mismatch ratio of the linkage action.

8. The multimodal forecasting and intelligent inventory control method based on big data fusion according to claim 7, characterized in that, The determination of whether the coordination capability of prediction and control meets the requirements is based on the mismatch ratio of linked actions, including: If the mismatch ratio of the linked action is less than or equal to the preset ratio, then the linkage coordination capability of prediction and control is determined to meet the requirements. If the mismatch ratio of the linked action is greater than the preset ratio, it is determined that the linkage coordination capability of prediction and control does not meet the requirements, and the continuous perception prediction anchor point drift of the actuator regulation is reconstructed.

9. The multimodal forecasting and intelligent inventory control method based on big data fusion according to claim 8, characterized in that, The mismatch ratio of the linkage actions is determined based on the ratio of the number of mismatched linkage actions per unit time to the total number of linkage executions.

10. The multimodal forecasting and intelligent inventory control method based on big data fusion according to claim 9, characterized in that, The reconstruction of the continuous perception prediction anchor point drift of the execution control is to reconstruct the timing constraints of the control instructions and the control execution strategy based on the degree of deviation of the prediction benchmark value.

Images