CN-121998098-A - Material business data-oriented AI training and reasoning method and application

CN121998098ACN 121998098 ACN121998098 ACN 121998098ACN-121998098-A

Abstract

The application relates to the technical field of artificial intelligence and data processing, in particular to a method for training and reasoning AI (advanced technology) of material business data and application thereof. The application improves the data quality and model training basis through multidimensional space-time alignment and delay compensation, realizes intelligent on-demand distribution of computing resources through dynamic grouping and differential reasoning strategies, and ensures the continuous effectiveness of the model in a dynamic environment through an online monitoring and dual-path updating mechanism.

Inventors

ZHANG GUANGBIN
CHENG HUAJUN
Cai Chunjiu

Assignees

数治云(北京)科技有限责任公司

Dates

Publication Date: 20260508
Application Date: 20260211

Claims (10)

1. The method for training and reasoning the material business data AI is characterized by comprising the following steps: The method comprises the steps of obtaining discrete transaction data from an enterprise resource planning system and real-time sequence data from an Internet of things system, constructing a virtual space time axis based on unified time service, carrying out time domain mapping on low-frequency discrete transaction data in the enterprise resource planning system through a Lagrange interpolation algorithm to enable the low-frequency discrete transaction data to be matched into a high-frequency sampling grid of the Internet of things system, establishing a knowledge-graph-based material association model, carrying out semantic mapping on material codes in the enterprise resource planning system, tray numbers in a warehousing system and waybill numbers in a logistics process through the material association model, converting heterogeneous data of different business processes into characteristic tensor representation of unified dimension, extracting time sequence association characteristics of heterogeneous data of different sources through a self-attention mechanism, identifying business logic offset caused by system delay, carrying out pre-compensation calculation on time conflicts of heterogeneous data of different sources according to the business logic offset, and completing multi-dimensional time-space alignment of heterogeneous data to generate an alignment sample set for artificial intelligent model training; Dynamically dividing material stock units in the aligned sample set into three grouping types of active state, steady state and dormant state according to material calling frequency and prediction error rate indexes, executing differentiated calculation pruning strategies aiming at different grouping types, combining hardware resource load states, routing model sub-weights of corresponding material types from storage media to a display memory of a calculation unit through a model parameter slicing loading technology, executing incremental reasoning or full reasoning, and outputting a material service prediction result; The self-adaptive updating based on dynamic drift monitoring and experience playback comprises the steps of deploying a statistical monitor in an artificial intelligent reasoning pipeline, calculating a prediction residual error between a material service prediction result and an actual service feedback value, quantifying concept drift degree based on statistical distribution change, triggering a dual-path online updating mechanism according to the concept drift degree, and carrying out iterative updating on model weights by means of online gradient descent or incremental training of a fused experience playback buffer zone and matching with a self-adaptive forgetting factor.
2. The method for training and reasoning the material business data AI according to claim 1, wherein the reasoning optimization based on dynamic grouping clipping and hot and cold weight routing specifically comprises: Counting the calling frequency and the predicted error rate of each material stock unit in a preset period in real time; identifying a stock of material with a calling frequency higher than a first threshold value and a predicted error rate fluctuation exceeding a second threshold value as active state, and executing full-scale deep learning reasoning on the stock of material; identifying a stock-keeping unit of materials with a calling frequency between a first threshold value and a third threshold value and a predicted error rate stable as a stable state, and only performing light weight reasoning based on increment deviation correction on the stock-keeping unit of materials; and identifying the material stock units with the calling frequency lower than the third threshold as dormant state, and triggering low-frequency batch reasoning tasks on the material stock units.
3. The method for training and reasoning according to claim 2, wherein the lightweight reasoning for incremental deviation correction is implemented by the following way: Extracting hidden layer state characteristics generated in the previous reasoning period of the steady-state material as a reference standard; calculating a characteristic change vector of currently input service data in a characteristic space relative to the previous period data; and carrying out feature mapping on the tiny feature change vector by using a lightweight residual error network with smaller scale, and carrying out fusion superposition on the residual error increment obtained by mapping and the hidden layer state of the previous period, thereby rapidly deducing the current prediction result.
4. The method for training and reasoning according to claim 1, wherein the model parameter slicing loading technique specifically comprises: The global artificial intelligent model is deconstructed into a plurality of independent model weight fragments according to the material category characteristics; Constructing a display memory prediction route model, and pre-calculating the occupation amount of the display memory according to grouping attribute and calculation priority of the stock units of the materials to be inferred; Before the reasoning task is started, only the model weight fragments corresponding to the target groups are loaded to a computing core through a direct storage access channel, and the model weight fragments are compressed by adopting an eight-bit integer quantization technology, so that the video memory overhead of single reasoning is reduced.
5. The method for training and reasoning the material business data AI according to claim 1, wherein the feature tensor representation for converting heterogeneous data of different business processes into uniform dimensions specifically comprises: extracting static properties, dynamic circulation paths and environment association features in the material association model; mapping the unstructured text description and the structured numerical code to the same high-dimensional vector space using an encoder; and splicing the feature vectors with different dimensions into feature tensors with fixed lengths through a feature fusion layer, and eliminating dimension differences among different heterogeneous systems by utilizing a standardization layer.
6. The method for training and reasoning of material business data AI according to claim 1, wherein the dual-path online update mechanism comprises a fast path update mode and a deep path update mode, and the adaptive update based on dynamic drift monitoring and experience playback specifically comprises: Monitoring data distribution in the reasoning process in real time by using a statistical inequality; When the monitored statistic deviation exceeds a preset warning threshold and does not reach an action threshold, entering a rapid path updating mode, and performing fine adjustment on the top layer weight of the model by using new service data generated at present through an online gradient descent algorithm; And when the monitored statistic deviation exceeds a preset action threshold, entering a deep path updating mode, extracting a representative historical sample from an experience playback buffer area, mixing the historical sample with current new service data, and triggering incremental training of the artificial intelligent model.
7. The method of claim 6, wherein the statistical inequality is Huo Fuding inequality, and the monitoring process comprises: setting confidence parameters of prediction residues; continuously calculating the difference value between the average value of the prediction residual error in the current sliding window and the historical reference average value; If the difference value meets the probability deviation limit defined by Huo Fuding inequality, judging that the material service environment has conceptual drift.
8. The method for training and reasoning about material business data AI of claim 6, wherein the management mechanism of the experience playback buffer comprises: clustering and sampling historical service data according to the time correlation and the feature diversity; Reserving representative samples near the center of each cluster and extreme samples with higher prediction difficulty; And along with the storage of the new sample, automatically eliminating the past historical sample according to the freshness index of the sample, and ensuring the validity of the sample in the buffer area.
9. The method for training and reasoning according to claim 1, wherein the adaptive forgetting factor is implemented by: constructing a functional relation positively correlated with the concept drift degree, and dynamically adjusting a weight attenuation coefficient during model training; when severe concept drift is detected, increasing the weight decay factor to accelerate the fade of the legacy inactive mode; When the business environment is in a stationary phase, the weight decay coefficient is reduced to maintain the memory capacity of the model for long-term laws.
10. An application of the method for training and reasoning the material business data AI according to any one of claims 1 to 9 in a material supply chain management system is characterized in that the application comprises the steps of applying the alignment sample set to supply chain demand prediction to realize accurate pre-judgment of the demand of mass material stock units, guiding automatic material allocation of a logistics storage system by utilizing the reasoning and optimizing result to optimize the stock turnover rate, capturing market fluctuation and demand change caused by policy adjustment through the self-adaptive updating mechanism, and realizing dynamic deviation correction of a material supply plan.

Description

Material business data-oriented AI training and reasoning method and application Technical Field The application relates to the technical field of artificial intelligence and data processing, in particular to a method for training and reasoning AI (advanced technology attachment) of material business data and application thereof. Background Along with the improvement of the modern supply chain management refinement degree, the scale of material business data is explosively increased, business logic offset caused by system delay is difficult to process by simple rule alignment, so that training sample quality is poor, on the other hand, a mass material stock unit is faced, the difference of calling frequencies and prediction error rates among different materials is ignored by a unified weight reasoning mode, so that calculation resource waste is caused, in addition, the material business environment is greatly influenced by market and fluctuation, and an existing policy lacks an effective online self-adaptive updating mechanism when concept drifting occurs, so that model prediction accuracy is fast attenuated along with time. Disclosure of Invention The application aims to provide a method for training and reasoning material business data AI, which comprises the steps of acquiring discrete transaction data from an enterprise resource planning system and real-time sequence data from an Internet of things system based on heterogeneous data alignment of a multidimensional space-time interpolation grid, constructing a virtual space time axis based on unified time service, and carrying out time domain mapping on low-frequency discrete transaction data in the enterprise resource planning system through a Lagrange interpolation algorithm to enable the low-frequency discrete transaction data to be matched into a high-frequency sampling grid of the Internet of things system; establishing a knowledge graph-based material association model, carrying out semantic mapping on material codes in an enterprise resource planning system, tray numbers in a warehouse system and waybill numbers in a logistics process through the material association model, converting heterogeneous data in different business processes into characteristic tensor characterization of unified dimensions, extracting time sequence association characteristics of heterogeneous data of different sources by using a self-attention mechanism, identifying business logic offset caused by system delay, carrying out precompensation calculation on time delay conflict of heterogeneous data of different sources according to the business logic offset, completing multidimensional space-time alignment of heterogeneous data, generating an aligned sample set for training an artificial intelligent model, carrying out reasoning optimization based on dynamic grouping clipping and cold-hot weight routing, dynamically dividing material inventory units in the aligned sample set into three grouping types of active state, steady state and dormant state according to material calling frequency and predicted error rate indexes, executing differentiated calculation pruning strategies for different grouping types, displaying model weights of corresponding material categories from a storage medium to a calculation unit by combining hardware resource loading technology, the method comprises the steps of performing incremental reasoning or full reasoning, outputting a material service prediction result, adaptively updating based on dynamic drift monitoring and experience playback, deploying a statistical monitor in an artificial intelligent reasoning pipeline, calculating a prediction residual between the material service prediction result and an actual service feedback value, quantifying a concept drift degree based on statistical distribution change, triggering a dual-path online updating mechanism according to the concept drift degree, and iteratively updating model weights by means of online gradient descent or incremental training of a fused experience playback buffer zone and matching with an adaptive forgetting factor. By adopting the technical scheme, a high-consistency alignment sample set is generated through multidimensional space-time alignment and delay compensation, the data quality and model training basis are improved, limited computing resources (especially GPU video memory) are preferentially distributed to high-value (active) materials through dynamic grouping and differential reasoning strategies, the low-value (dormant) materials are subjected to frequency reduction processing, so that the system reasoning throughput is improved by several times as a whole, the average video memory overhead and delay of single reasoning are obviously reduced, intelligent on-demand distribution of computing resources is realized, an AI model can continuously learn and adapt to service environment changes through an online monitoring and dual-path updating mechanism, the performance r