Material business data-oriented AI training and reasoning method and application

Abstract

The invention relates to the technical field of artificial intelligence and data processing, in particular to a material business data AI-oriented training and reasoning method and application, and the method comprises the steps: carrying out the heterogeneous data alignment based on a multi-dimensional space-time interpolation grid; reasoning optimization based on dynamic grouping cutting and cold and hot weight routing; the invention relates to adaptive updating based on dynamic drift monitoring and experience playback. According to the method, through multi-dimensional space-time alignment and delay compensation, the data quality and the model training basis are improved; intelligent on-demand allocation of computing resources is realized through dynamic grouping and differentiated reasoning strategies; and through online monitoring and a double-path updating mechanism, the continuous effectiveness of the model in a dynamic environment is ensured.

Description

Technical Field

[0001] This application relates to the fields of artificial intelligence and data processing technology, specifically to a method and application for AI training and inference of material business data. Background Technology

[0002] With the increasing sophistication of modern supply chain management, the scale of material business data has exploded. Transactional data generated by Enterprise Resource Planning (ERP) systems and real-time sensing data generated by Internet of Things (IoT) systems together constitute the core assets of material business. Deep mining of this multi-source heterogeneous data using artificial intelligence (AI) technology has become a key approach to achieving accurate inventory forecasting and automated allocation. However, existing technologies still have limitations in practical applications. On the one hand, the low-frequency discrete data from ERP systems and the high-frequency time-series data from IoT systems differ significantly in time scale. Simple rule alignment is insufficient to handle business logic deviations caused by system latency, resulting in poor training sample quality. On the other hand, facing massive material inventory units, the uniform weight inference model ignores the differences in call frequency and prediction error rate between different materials, leading to a waste of computational resources. Furthermore, the material business environment is greatly affected by market and policy fluctuations, and existing models lack an effective online adaptive update mechanism when concept drift occurs, causing the model's prediction accuracy to decay rapidly over time. Summary of the Invention

[0003] The objective of this application is to provide a method for AI training and inference of material business data, including: heterogeneous data alignment based on a multi-dimensional spatiotemporal interpolation grid: acquiring discrete transaction data from an Enterprise Resource Planning (ERP) system and real-time time-series data from an Internet of Things (IoT) system, constructing a virtual spatial time axis based on unified time synchronization, and mapping the low-frequency discrete transaction data from the ERP system to the high-frequency sampling grid of the IoT system using a Lagrange interpolation algorithm; establishing a material association model based on a knowledge graph, semantically mapping the material codes from the ERP system, pallet numbers from the warehousing system, and waybill numbers from the logistics process through the material association model, transforming heterogeneous data from different business processes into a unified-dimensional feature tensor representation; extracting the temporal association features of heterogeneous data from different sources using a self-attention mechanism, identifying business logic offsets caused by system delays, and pre-compensating for time lag conflicts of heterogeneous data from different sources based on the business logic offsets, thus completing the multi-dimensional alignment of heterogeneous data. Spatiotemporal alignment generates an aligned sample set for AI model training; inference optimization based on dynamic grouping pruning and hot / cold weight routing: according to material call frequency and prediction error rate, the material inventory units in the aligned sample set are dynamically divided into three grouping types: active, stable, and dormant. Differentiated computational pruning strategies are executed for different grouping types. Combined with hardware resource load status, the model sub-weights of the corresponding material categories are routed from the storage medium to the GPU memory of the computing unit through model parameter sharding loading technology, performing incremental or full inference, and outputting material business prediction results; adaptive update based on dynamic drift monitoring and experience replay: a statistical monitor is deployed in the AI ​​inference pipeline to calculate the prediction residual between the material business prediction results and the actual business feedback values, and to quantify the degree of concept drift based on statistical distribution changes; a dual-path online update mechanism is triggered according to the degree of concept drift, and the model weights are iteratively updated through online gradient descent or incremental training fused with the experience replay cache, combined with an adaptive forgetting factor.

[0004] By adopting the above technical solutions, a highly consistent aligned sample set was generated through multi-dimensional spatiotemporal alignment and latency compensation, improving data quality and the foundation for model training. Through dynamic grouping and differentiated inference strategies, limited computing resources (especially GPU memory) were prioritized for high-value (active) materials, while low-value (dormant) materials were downclocked. This resulted in a several-fold increase in overall system inference throughput and a significant reduction in average memory overhead and latency per inference, achieving intelligent on-demand allocation of computing resources. Through online monitoring and a dual-path update mechanism, the AI ​​model was able to continuously learn and adapt to changes in the business environment, reducing the performance recovery time after concept drift from several hours / days required for traditional retraining to minutes / hours, maintaining long-term prediction accuracy, and ensuring the model's continued effectiveness in dynamic environments.

[0005] Optionally, the inference optimization based on dynamic grouping and pruning and hot / cold weighted routing specifically includes: real-time statistics of the call frequency and prediction error rate of each material inventory unit within a preset period; identifying material inventory units with a call frequency higher than a first threshold and prediction error rate fluctuations exceeding a second threshold as active states, and performing full deep learning inference on them; identifying material inventory units with a call frequency between the first and third thresholds and stable prediction error rates as stationary states, and performing only lightweight inference based on incremental deviation correction on them; identifying material inventory units with a call frequency lower than the third threshold as dormant states, and triggering low-frequency batch inference tasks on them.

[0006] By adopting the above technical solutions and using clearly defined threshold indicators, the classification of material states is no longer based on manual experience, but rather on data-driven objective classification, achieving quantification and automation of grouping. Clear and differentiated calculation strategies (full volume, lightweight, and batch) are defined for the three types of materials, ensuring a precise match between computational resource investment and the business value and forecasting needs of the materials. This reduces the overall system computational load by 30%-50% while maintaining the forecasting accuracy for high-value materials, achieving refined scheduling of computational resources.

[0007] Optionally, the lightweight inference for incremental deviation correction is implemented as follows: extracting the hidden layer state features generated by the stationary material in the previous inference cycle as a reference benchmark; calculating the feature change vector of the current input business data relative to the data of the previous cycle in the feature space; using a small-scale lightweight residual network to perform feature mapping only on this tiny feature change vector, and fusing and superimposing the mapped residual increment with the hidden layer state of the previous cycle, thereby quickly deriving the current prediction result.

[0008] By adopting the above technical solution, since there is no need to recalculate the entire deep network and only handles minor feature changes, the computational load and time for inference of stationary materials can be reduced by an order of magnitude, significantly improving inference efficiency. By reusing historical hidden states and superimposing incremental corrections, historical information is inherited while capturing the latest changes. While significantly improving efficiency, the loss in prediction accuracy can be controlled within an acceptable range, maintaining the consistency and accuracy of predictions.

[0009] Optionally, the model parameter sharding loading technique specifically includes: deconstructing the global artificial intelligence model into multiple independent model weight shards according to material category characteristics; constructing a memory prediction routing model, pre-calculating the memory usage based on the grouping attributes and computation priority of the inventory units of the material to be inferred; before the inference task starts, loading only the model weight shards corresponding to the target group into the computing core through the direct memory access channel, and using eight-bit integer quantization technology to compress the model weight shards to reduce the memory overhead of a single inference.

[0010] By adopting the above technical solutions, through fragmentation and component-based loading, a single inference task only needs to occupy a small portion of the global model's GPU memory, reducing peak GPU memory usage by 70%-90%. This supports deploying larger or more models in resource-constrained environments, enabling dynamic on-demand use of GPU memory. The application of direct memory access loading and octet quantization significantly reduces the amount and time of data transfer from storage to GPU memory, lowers I / O bottlenecks, improves the startup speed and overall throughput of inference tasks, and reduces data transfer overhead.

[0011] Optionally, the process of transforming heterogeneous data from different business processes into a unified-dimensional feature tensor representation specifically includes: extracting static attributes, dynamic flow paths, and environmental features from the material association model; using an encoder to map unstructured text descriptions and structured numerical codes to the same high-dimensional vector space; concatenating feature vectors of different dimensions into a fixed-length feature tensor through a feature fusion layer; and using a standardization layer to eliminate dimensional differences between different heterogeneous systems.

[0012] By adopting the above technical solution, the encoder unifies the vector space and the standardization layer eliminates dimensional differences, enabling subsequent AI models to learn in a consistent feature space. This avoids model bias caused by differences in data sources and achieves deep fusion and standardization of heterogeneous information. The fixed-length feature tensor not only contains the original data but also the semantic relationships between materials provided by the knowledge graph, giving the model a more powerful feature representation capability. This helps improve the accuracy and generalization ability of subsequent prediction models and generates semantically rich structured features.

[0013] Optionally, the dual-path online update mechanism includes a fast path update mode and a deep path update mode. The adaptive update based on dynamic drift monitoring and experience replay specifically includes: using statistical inequalities to monitor the data distribution during the inference process in real time; when the detected statistical deviation exceeds a preset warning threshold but does not reach an action threshold, entering the fast path update mode, using the currently generated new business data to fine-tune the top-level weights of the model through an online gradient descent algorithm; when the detected statistical deviation exceeds a preset action threshold, entering the deep path update mode, extracting representative historical samples from the experience replay cache, and mixing the historical samples with the current new business data to trigger incremental training of the artificial intelligence model.

[0014] By adopting the above technical solutions, the fast path can quickly adapt to minor changes with extremely low computational cost; the deep path can cope with fundamental changes and prevent model failure. This mechanism, while ensuring model adaptability, significantly reduces the computational cost of model maintenance, achieving intelligent and efficient update response. The introduction of historical samples into the experience replay cache in the deep path can prevent catastrophic forgetting of new data, ensuring that the updated model can both adapt to new trends and retain past effective experience, thus improving the stability and effectiveness of model updates.

[0015] Optionally, the statistical inequality is the Hofding inequality, and its monitoring process includes: setting a confidence parameter for the predicted residuals; continuously calculating the difference between the mean of the predicted residuals within the current sliding window and the historical benchmark mean; if the difference satisfies the probability deviation limit defined by the Hofding inequality, then it is determined that a concept drift has occurred in the material business environment.

[0016] By adopting the above technical solution and using the classic probability theory tool, Hofding's inequality, the determination of "concept drift" has a clear mathematical explanation and a controllable false alarm rate. The determination based on statistical confidence reduces the reliance on absolute threshold settings, enabling the system to maintain stable and reliable monitoring performance under different business scenarios and noise levels, thus achieving objectivity and stability in drift detection.

[0017] Optionally, the management mechanism of the experience replay cache includes: clustering historical business data based on time correlation and feature diversity; retaining representative samples near each cluster center and extreme samples with high prediction difficulty; and automatically removing outdated historical samples based on sample freshness index as new samples are stored to ensure the effectiveness of samples in the cache.

[0018] By adopting the above technical solutions, cluster sampling ensures that the samples cover different historical business patterns (diversity), while retaining central and extreme samples guarantees the representativeness of the patterns and the learning challenge, respectively, making replay training more efficient and optimizing the quality and diversity of cached samples. Automatically eliminating old samples using a freshness index ensures that the information in the cache always has a certain relevance to the current business environment, avoiding waste of storage resources and interference from erroneous knowledge, improving the effectiveness and speed of incremental training, and achieving dynamic self-maintenance of the cache.

[0019] Optionally, the adaptive forgetting factor is implemented in the following ways: constructing a functional relationship positively correlated with the degree of concept drift, and dynamically adjusting the weight decay coefficient during model training; when severe concept drift is detected, increasing the weight decay coefficient to accelerate the fading of old ineffective patterns; when the business environment is in a stable period, decreasing the weight decay coefficient to maintain the model's ability to remember long-term patterns.

[0020] By adopting the above technical solution, the forgetting factor is linked to the degree of quantified drift, enabling the model to intelligently determine how much old experience to forget. This allows for rapid adaptation during periods of rapid environmental change and memory consolidation during periods of stability, achieving precise control over the intensity of forgetting. This dynamic balancing mechanism effectively mitigates the catastrophic forgetting problem in continuous learning, ensuring that the model retains a certain level of memory for business patterns across different historical stages even after multiple updates. This results in more robust overall performance and enhances the model's long-term learning robustness.

[0021] The second objective of this application is to provide an application of the aforementioned AI training and inference method for material business data in a material supply chain management system. The application includes: applying the aligned sample set to supply chain demand forecasting to achieve accurate prediction of demand for massive material inventory units; using the optimized results of the inference to guide the automatic allocation of materials in the logistics and warehousing system and optimize inventory turnover; and capturing demand changes caused by market fluctuations and policy adjustments through the adaptive update mechanism to achieve dynamic correction of material supply plans.

[0022] By adopting the above-mentioned technical solutions, advanced AI data processing, inference optimization, and adaptive learning capabilities are directly applied to core business processes such as demand forecasting, inventory allocation, and plan correction, achieving intelligent and precise supply chain management. Through accurate forecasting, safety stock levels are reduced by 10%-30%; through optimized allocation, inventory turnover is increased by 15%-25%; and through rapid dynamic correction, the risk of stockouts or overstocking due to sudden demand changes is reduced, enhancing supply chain resilience and generating quantifiable commercial value. Attached Figure Description

[0023] Figure 1This is a schematic diagram of the method and process for AI training and inference of material business data in this application. Detailed Implementation

[0024] To enable those skilled in the art to better understand the present application, the technical solutions in the embodiments of the present application will be clearly and completely described below in conjunction with the embodiments of the present application. Obviously, the described embodiments are only some embodiments of the present application, and not all embodiments. Based on the embodiments of the present application, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of the present application.

[0025] like Figure 1 As shown in the figure, this application provides a method for AI training and inference of material business data, including the following steps.

[0026] S01: Heterogeneous Data Alignment Based on Multidimensional Spatiotemporal Interpolation Grid: Discrete transaction data from the Enterprise Resource Planning (ERP) system and real-time time-series data from the Internet of Things (IoT) system are acquired. A virtual spatial time axis based on unified time synchronization is constructed. The low-frequency discrete transaction data from the ERP system is temporally mapped using the Lagrange interpolation algorithm to match it to the high-frequency sampling grid of the IoT system. A knowledge graph-based material association model is established. Through this model, material codes from the ERP system, pallet numbers from the warehousing system, and waybill numbers from the logistics process are semantically mapped, transforming heterogeneous data from different business processes into a unified-dimensional feature tensor representation. A self-attention mechanism is used to extract temporal association features from heterogeneous data from different sources. Business logic offsets caused by system delays are identified. Based on these offsets, pre-compensation calculations are performed to address time lag conflicts in heterogeneous data from different sources, completing the multidimensional spatiotemporal alignment of heterogeneous data and generating an alignment sample set for training artificial intelligence models.

[0027] Understandably, within the technological context of intelligent supply chains and modern logistics management, the AI ​​training and inference methods for material business data provided in this application are primarily applied to material business data processing, artificial intelligence model training, and real-time inference and prediction scenarios in industrial production environments. Specifically, in large manufacturing enterprises or warehousing centers, enterprise resource planning systems continuously generate macro-level business data on material flow. This data covers information on inbound applications, outbound confirmations, order changes, real-time inventory balances, and changes in bills of materials. Key fields typically include the material's unique identification code, production batch number, business operation type, and initial storage location coordinates in the physical warehouse. Simultaneously, sensors in IoT systems deployed on production lines, pallets, and transport vehicles continuously generate real-time time-series data on the micro-state of materials, such as real-time temperature and humidity data monitoring environmental quality. Data includes vibration frequency data reflecting the operating status of mechanical equipment, and location flow trajectory data characterizing the dynamic movement of materials within the storage area. However, since Enterprise Resource Planning (ERP) systems are essentially discrete recording systems based on transaction triggers, their data recording frequency is often limited by the completion time of business processes, typically using days, hours, or minutes as time scales. In contrast, IoT sensing systems are continuous monitoring systems based on high-frequency sampling, with sampling accuracy typically reaching the second or even millisecond level. This significant difference in time granularity makes it difficult to match macroscopic business logic with microscopic physical states in the time domain. Furthermore, due to inconsistent material coding rules, inconsistent physical space coordinate system definitions, or the complexity of nested logical levels between different systems, it is difficult to construct complete and strongly correlated feature tensors when training artificial intelligence models, forming a heterogeneous data correlation bottleneck that restricts the intelligent transformation of material business.

[0028] Specifically, addressing the aforementioned complex application scenarios and data alignment bottlenecks, this embodiment first achieves accurate alignment of heterogeneous data in step S01 by constructing a multi-dimensional spatiotemporal interpolation grid. In the specific implementation process, the system can establish a virtual spatial time axis based on a globally unified time reference. The axis of this virtual spatial time axis is divided into equally spaced time grids with a unit of 1 second, serving as a unified coordinate system for all heterogeneous data. For discrete transaction data from the Enterprise Resource Planning (ERP) system, the system first identifies the start and end timestamps of the business transaction and maps them to the corresponding second-level grid start point. However, since business operations often have a long time gap between two recording points, the system cannot directly treat discrete values ​​as continuous states. Therefore, low-frequency transaction records need to be upsampled in the time domain. To accurately reflect the dynamic changes in material states when completing missing time data, this embodiment can use a text-based Lagrange interpolation algorithm for time domain mapping completion, i.e., in... At each second-level grid point where values ​​need to be filled, the system automatically selects several known enterprise resource planning (ERP) system record points before and after that moment as reference benchmarks. A polynomial function is constructed to approximate the trajectory of material state changes. Specifically, for each second-level time point to be solved, the system calculates a set of weighting coefficients derived from known sampling points. The numerator of each weighting coefficient is the product of the differences between the time to be solved and the times of all known points except the current reference point, while the denominator is the product of the differences between the time of the current reference point and the times of all known points except itself. By multiplying the business value of each known point with the corresponding weighting coefficient and summing them, the logical prediction value at that second-level grid point can be obtained. This ensures that the inventory balance or order progress data in the ERP system can be evenly distributed in the second-level time grid in a smooth manner that conforms to the business logic, thereby achieving second-by-second alignment with the real-time high-frequency sampling points of IoT sensors in the time dimension.

[0029] After aligning the time dimension, this application further addresses the semantic consistency of material names and specifications across different business systems using knowledge graph technology. In actual industrial environments, enterprise resource planning systems may identify a part as a code with a specific combination of characters, while warehouse management systems may associate it with a specific pallet number. IoT positioning systems, on the other hand, can only perceive the coordinates of the material's location. This multi-source, heterogeneous identification system leads to severe semantic silos. To address this, this application constructs a knowledge graph model encompassing basic material attributes, business relationships, and physical topology. This knowledge graph model defines material codes, batch numbers, specifications, pallet numbers, and waybill numbers as core entities in the graph and utilizes entity alignment algorithms to identify the different systems pointing to them. When transforming heterogeneous texts and structured information into inputs that can be understood by artificial intelligence models, the embodiments of this application utilize a pre-trained semantic encoder to map the node information and its neighborhood relationships in the knowledge graph to a high-dimensional semantic vector space. By performing nonlinear transformations in the vector space, the originally isolated material names and specifications are transformed into numerical tensors with fixed dimensions, so that materials with similar physical characteristics or business connections are closer to each other in the vector space. This unified semantic vector representation not only eliminates the ambiguity caused by polysemy or homosemansis, but also concatenates the static semantic information of the material with the aforementioned dynamic physical feature tensor after time-domain alignment through a feature fusion layer, forming a multi-dimensional feature matrix covering the entire life cycle of the material.

[0030] To further eliminate business logic offsets caused by system communication delays or manual input lags, this application's embodiments introduce a self-attention mechanism for conflict compensation during the generation of the alignment sample set. Considering that the recording time of the enterprise resource planning system is often later than the actual time when IoT sensors perceive material movement, the system uses a self-attention model to slide and scan within a time grid, calculating the correlation scores between different feature channels to identify the phase difference between the business logic curve and the physical state curve. Once the business logic offset caused by system delay is identified, the system can pre-compensate for time lag conflicts based on the statistical characteristics of the offset, through virtual space... The business logic features are adaptively adjusted ahead or behind on the timeline, ensuring that the feature tensors within each second-level grid accurately reflect the true overall state of the materials at that moment. Finally, the data, after temporal interpolation completion, spatial semantic alignment, and logical offset compensation, are integrated into a standardized aligned sample set. These aligned sample sets are input into the subsequent artificial intelligence neural network in a unified tensor format, which greatly improves the training convergence speed and inference accuracy of the model when handling tasks such as material demand forecasting, inventory early warning, and automatic allocation instruction generation. This fundamentally solves the technical bottleneck in industrial material business where data heterogeneity makes it difficult to achieve intelligent monitoring and decision-making throughout the entire process.

[0031] S02: Inference optimization based on dynamic grouping and pruning and hot / cold weight routing: According to the material call frequency and prediction error rate, the material inventory units in the aligned sample set are dynamically divided into three grouping types: active, stable, and dormant. Differentiated calculation pruning strategies are executed for different grouping types. Combined with the hardware resource load status, the model sub-weights of the corresponding material categories are routed from the storage medium to the video memory of the computing unit through model parameter sharding loading technology. Incremental inference or full inference is performed to output the material business prediction results.

[0032] Specifically, firstly, a multi-dimensional business metric model is established to conduct real-time heat assessment of material inventory units. The system continuously monitors key indicators in the enterprise resource planning system, such as material call frequency, inventory turnover rate, prediction error rate, and historical prediction error rate. By setting a sliding time window, the activity level of each inventory unit within a specific period is statistically analyzed. Material inventory units with a call frequency higher than a preset first threshold and a prediction error rate fluctuation exceeding a second threshold within a unit of time are accurately identified as active materials. These materials typically correspond to the company's core high-frequency turnover products, whose demand is greatly affected by market environment, promotional activities, and supply chain fluctuations. Therefore, the highest priority computing resources need to be allocated and a full-scale deep learning inference algorithm needs to be executed to capture subtle patterns of demand changes using complex nonlinear feature extraction capabilities. The reasoning process for active materials involves not only complete deep neural network calculations but also dynamic weight correction based on real-time logistics environment characteristics. By calling all model parameters in the high-performance computing cluster, the system performs deep fusion calculations of multilayer perceptrons and recurrent neural networks on all time-series feature tensors related to this type of material. This ensures that when material demand fluctuates drastically, the system can output highly confident prediction results through high-precision end-to-end mapping. In addition, to cope with the computational pressure brought by massive active inventory units, the system can enable the parallel acceleration mechanism of the computing core at the hardware level, distributing active material data from different batches to multiple processing units for parallel processing. This minimizes inference latency while ensuring prediction accuracy, enabling real-time demand perception and dynamic adjustment of inventory levels for highly volatile materials.

[0033] For inventory units whose call frequency falls between the first and third thresholds and whose prediction error rate remains stable, the system classifies them as stationary materials. These materials are characterized by strong periodicity or trends in their demand patterns, and their future business performance is highly correlated with their historical state. Therefore, instead of repeatedly performing time-consuming full neural network forward computation in every inference cycle, a lightweight inference strategy based on incremental bias correction is adopted. This strategy extracts the hidden layer state features of the stationary materials from the previous inference cycle as a reference benchmark, calculates the feature change vector of the current input business data relative to the previous cycle's data in the feature space, and uses a small, lightweight residual network to perform feature mapping on this tiny feature change vector. The resulting residual increment is then fused and superimposed with the hidden layer state of the previous cycle. This allows for the rapid derivation of current forecast results, significantly reducing floating-point computation while maintaining relatively high forecast accuracy. When processing feature change vector mapping of stationary materials, the system automatically masks static attribute features unrelated to demand forecasting, focusing on dynamic and sensitive dimensions such as inventory changes, price adjustments, and seasonal factors. Through the attention gating mechanism within the lightweight residual network, it automatically learns and assigns contribution weights to changes in different dimensions, enabling the model to capture trend shifts in material demand with minimal computational cost. This incremental inference architecture not only leverages deep learning's ability to model complex patterns but also reduces redundant feature extraction by introducing historical state memory. It is particularly suitable for large-scale industrial parts or daily consumables with relatively regular demand, effectively balancing the contradiction between forecast accuracy and system throughput.

[0034] For inventory units whose call frequency is consistently below the third threshold, the system identifies them as dormant materials. These materials are typically obsolete inventory, long-tail products, or spare parts. Their business data is sparse and demand occurs very infrequently. Performing high-frequency real-time predictions on them would result in a significant waste of computing resources. Therefore, the system assigns their inference tasks to a low-priority background batch processing queue, triggering low-frequency batch inference tasks. A one-time prediction task is only triggered when the material undergoes actual inbound or outbound transactions or reaches a preset long-cycle inference node. Simple statistical models or shallow machine learning models are used for processing, thereby freeing up valuable computing power for materials in active and stable states. This method, based on material business attributes... The hot and cold weight routing mechanism dynamically decouples computational tasks at the business logic level, greatly optimizing the task orchestration efficiency of the AI ​​inference system in a large-scale warehousing environment. To further improve inference speed in resource-constrained environments, this application introduces hardware-level optimization technology based on model parameter sharding and octet integer quantization during the model deployment stage. First, the trained global AI model is structurally deconstructed according to the category characteristics of materials, splitting it into multiple model weight shards with independent logical functions. Each model weight shard corresponds to one or more categories of materials with similar physical attributes or business logic. Based on the usage frequency of these model weight shards in actual inference, a memory prediction routing model is constructed. The predictive routing model can pre-determine the size of the required model weight fragments and their corresponding memory usage based on the grouping attributes and computational priorities of the material inventory units in the task queue. By establishing a priority-based memory replacement algorithm, it ensures that frequently accessed model weight fragments reside in the high-speed memory of the computing unit, while low-frequency fragments reside in system memory or high-speed solid-state drives, and are scheduled in real-time according to predictive needs. During the actual migration of model weight fragments from storage media to the computing core's memory, the system utilizes direct memory access channels to bypass the cumbersome instruction processing of the central processing unit, achieving high-speed data transmission on the hardware bus, greatly improving the bandwidth utilization of model weight fragment loading. Simultaneously, it targets… To address the uneven distribution of material business data tensors, this application designs an adaptive eight-bit integer quantization algorithm. During the model conversion stage, a large-scale sampling and statistical analysis is performed on the material feature tensor and the model weight tensor to find a saturation cutoff threshold that can cover more than 99% of the data distribution. The scaling factor required for converting floating-point numbers to eight-bit signed integers is calculated. The calculation process of this scaling factor strictly follows the symmetric quantization strategy. By finding the maximum absolute value in the tensor and mapping it to the numerical boundary of the signed integer, the original 30-bit floating-point number is compressed into an eight-bit integer. This not only reduces the storage volume of the model to one-quarter of the original, but also enables the computation core to use a dedicated integer operation unit for vectorization acceleration.To compensate for rounding errors and precision loss during 8-bit integer quantization, the system introduces an offset correction factor at the end of the quantization layer. By performing multiple forward inferences on the calibration dataset, the system statistically analyzes the distribution of activation value deviations before and after quantization. Taylor expansion is then used to linearly compensate for this activation value deviation, ensuring that the decrease in material prediction accuracy after quantization is controlled within 0.3%. For intermediate tensors generated during calculation, the system also employs dynamic quantization technology, dynamically adjusting the scaling factor based on the current numerical range of the calculation layer. This ensures that the numerical expression remains within the high-precision range of integer bit width. The number of inventory unit prediction tasks processed by a single calculation core per unit time is increased several times, meeting the stringent performance requirements of large e-commerce or manufacturing enterprises for concurrent inference of millions of inventory units.

[0035] To address the memory sharding logic in model parameter sharding loading, the system constructs a multi-level buffer scheduling framework. This framework divides the computational core's memory into a resident area, a swap area, and a prefetch area. The resident area stores the most common model base layer weights. The swap area updates specific material prediction shards in real-time based on hot / cold routing results. The prefetch area, based on the subsequent distribution of the current task queue, utilizes idle bus bandwidth to pre-fetch the next potentially triggered model weight shard from memory. This effectively masks the physical latency between disk and memory and the video memory. Even in complex business scenarios involving frequent material type switching, the efficient routing algorithm avoids computational pauses caused by frequent memory swapping, ensuring smooth operation of the AI ​​prediction pipeline and efficient closed-loop utilization of computational resources. In a specific inference task execution sequence, the scheduler dynamically generates a directed acyclic graph (DAG) based on the material's grouping attributes and computational priority. This DAG clearly defines the model weight shard number, data input address, and expected memory release time required for each prediction task. When an inference task for an active material is activated... The scheduler immediately checks whether the model weight slices of the material category are already in memory. If not, it performs forced prefetching through a high-priority direct memory access channel and automatically releases the space of model weight slices that are dormant or have low processing frequency based on the results of the memory prediction routing model. Combined with the high throughput brought by octet integer quantization, the system can simultaneously support several times more model instances than traditional solutions on a single server, significantly reducing the overall cost of ownership of the material prediction system. To further optimize the real-time performance of large-scale material inventory unit prediction, this application arranges material tasks with similar demand characteristics and business relevance on the time axis in a compact manner, maximizing the reuse of model weight slices already loaded into memory and reducing the power consumption and latency caused by repeated loading. When processing supply chain material combinations with strong relevance, the system splices the feature tensors of these materials into a larger batch processing array. Utilizing the single instruction multiple data stream characteristic of the quantization acceleration unit, it completes the preliminary prediction of multiple material inventory units in one clock cycle, and then performs minor residual corrections based on the differentiated characteristics of different materials.In the implementation details of eight-bit integer quantization, this application can adopt a dynamic threshold selection method based on histogram statistics. By analyzing the changes in the value range of the material characteristic tensor over different time periods, the system automatically adjusts the calculation logic of the scaling factor. When a significant long-tail characteristic is detected in the data distribution, the asymmetric quantization mode is automatically activated. By introducing a non-zero offset reference point, the originally unevenly distributed floating-point values ​​are mapped more evenly to the integer range of 0 to 255, thereby preserving more low-level feature details. This flexibility ensures that the model can adapt to huge numerical differences between different material categories. For example, for low-value consumables with extremely stable demand and high-value precision equipment with drastic demand fluctuations, the system can automatically switch quantization strategies to find the balance between performance and accuracy. The optimal balance point is achieved to ensure the stability of inference results. In the memory sharding loading logic, this application also introduces a model weight sharding compression storage technology. During the loading of model weight shards from disk to memory, a lossless compression algorithm is used to reduce the amount of data transmitted. In the final stage before entering memory, the hardware decompression engine restores the data in real time, further reducing loading latency. This allows the system to adjust the model distribution in memory with a microsecond-level response speed based on real-time fluctuations in material activity. This dynamic response capability enables the system to quickly route active weight shards when facing sudden material business peaks, such as annual large-scale promotions or seasonal inventory adjustments, ensuring that the prediction tasks of core business materials always receive sufficient computing power support and avoiding business decision delays caused by system overload.

[0036] To verify the effectiveness of the aforementioned inference optimization scheme in actual material prediction scenarios, the system incorporates a performance monitoring module. This module continuously tracks the inference latency, memory usage, computing unit utilization, and quantized precision offset for each material group. If the prediction residual for a certain type of material exceeds the preset safety boundary due to quantization, the performance monitoring module immediately sends an early warning command to the routing scheduler, triggering a dynamic rollback from eight-bit integer inference mode to sixteen-bit floating-point mode, or increasing the loading frequency of model weight slices to obtain more complete weight information. This closed-loop self-monitoring and feedback mechanism ensures that while pursuing ultimate performance, the inference optimization strategy always maintains a balance with... The rigorous protection of business accuracy provides a solid technical guarantee for the efficient management of large-scale industrial inventory units. By deeply integrating the above-mentioned strategies based on dynamic grouping and hot / cold weighted routing with underlying hardware optimization, this application successfully solves the technical challenge of traditional artificial intelligence models in massive material scenarios, which struggle to simultaneously balance real-time prediction and model complexity. It not only achieves refined prediction of active materials but also greatly saves overall system computing resources through differentiated processing of stable and dormant materials. This enables a unified artificial intelligence architecture to serve tens of thousands of materials with different attributes simultaneously, significantly improving the digitalization level and response speed of supply chain management.

[0037] S03: Adaptive Update Based on Dynamic Drift Monitoring and Experience Replay: A statistical monitor is deployed in the AI ​​inference pipeline to calculate the prediction residual between the material business forecast results and the actual business feedback values, and to quantify the degree of concept drift based on changes in statistical distribution. A dual-path online update mechanism is triggered based on the degree of concept drift, using online gradient descent or incremental training fused with an experience replay buffer, combined with an adaptive forgetting factor to iteratively update the model weights. The dual-path online update mechanism includes a fast path update mode and a deep path update mode.

[0038] To achieve a keen awareness of changes in material demand patterns, this embodiment of the application deploys a statistical monitor at the output end of the artificial intelligence inference pipeline. This monitor continuously collects actual business feedback values—such as actual warehouse outbound order quantities, procurement delivery cycles, and actual inventory occupancy rates—calculates the prediction residual sequence between these values ​​and the prediction results, and maps it to a continuous statistical space. Considering that material business data often contains a large amount of random noise, such as inventory reductions due to occasional material damage or temporary warehousing caused by administrative intervention, the system utilizes statistical inequalities to monitor the data during the inference process in real time. To address model drift, a robust automatic model drift identification mechanism is established, particularly through the engineering implementation of the Hofding inequality. This mechanism constructs a probabilistic confidence interval for fluctuations in the prediction residuals without assuming the original data follows a normal distribution. In the specific engineering implementation of the Hofding inequality, the system first sets a global confidence parameter for the prediction residuals. This parameter characterizes the system's tolerance for false alarm drift. Then, a sliding time window is used to continuously track the mean residual within the current business cycle, and this mean is differiated from the historical baseline mean recorded during model training. The system then calculates the difference based on the sliding time window. The system calculates a dynamic probability deviation limit in real time by considering the number of samples within the window and the theoretical lower and lower limits of residual values. When the observed statistical deviation exceeds this theoretical threshold derived from the inequality, the system has a very high probability of determining that the current material distribution has undergone a fundamental logical shift, rather than a normal fluctuation influenced by random noise. To address fluctuations of varying degrees in material turnover, this application designs a two-layer monitoring threshold logic: a warning threshold and an action threshold. When the statistical monitor detects that the statistical deviation exceeds the warning threshold but has not yet reached the action threshold, the system determines that the material environment is in a potential drift period. For example, if the demand for cold-weather supplies increases slowly due to seasonal changes or the procurement frequency is slightly adjusted due to routine industry price adjustments, the system will not trigger large-scale model retraining. Instead, it will automatically enter the fast path update mode. In this mode, the system will lock the underlying parameters in the model that are responsible for general feature extraction and use only the newly generated business data tensor within the current sliding window to fine-tune the top prediction mapping layer of the model through an online gradient descent algorithm. This allows the model to quickly fit these subtle distribution changes with extremely low computational cost, ensuring that the prediction results will not have obvious systematic biases.Once the statistical deviation exceeds the preset action threshold, it signifies a drastic conceptual shift in the material business environment. This typically corresponds to a major transformation in corporate business strategy, structural reorganization following supply chain disruptions, or dramatic changes in the macroeconomic environment. In this case, the original model weights cannot be fine-tuned to recover the accuracy loss. The system will automatically activate the deep path update mode, triggering a complete or incremental model iteration task, and simultaneously activate the experience playback mechanism to prevent the model from completely losing its memory of long-term historical patterns during the learning of new patterns. This application establishes a screening mechanism with dynamic clustering labels. By performing feature engineering extraction on historical business data, the system calculates the position coordinates of each sample in the multidimensional business space and uses a real-time clustering algorithm to divide the samples into different business scenario clusters, such as high-turnover promotion clusters, long-tail stagnation clusters, or sudden anomaly clusters. The system will sample the most representative samples from the vicinity of each cluster center, while specifically retaining those extreme samples with large residuals in historical predictions. These extreme samples often represent the boundaries of business logic. The system mixes representative historical samples with newly generated business samples in a certain proportion before feeding them into the training pipeline. This effectively acts as an anchor point during model weight updates, preventing catastrophic forgetting caused by overfitting to current short-term volatility. To further refine the management of the experience playback cache, the system introduces a dual elimination strategy based on sample freshness and feature diversity. Each business sample stored in the cache is assigned a weight score that decays linearly over time. The system also periodically evaluates the feature distribution entropy of samples in the cache. If newly added samples highly overlap with existing samples in the feature space, the system will eliminate old samples based on the freshness priority principle, or trigger a re-retrieval of historical cold data warehouse when sample diversity is insufficient. This ensures that the dataset used for model updates always maintains optimal coverage of material business logic, enabling the AI ​​model to identify current abnormal trends and refer to the handling logic of similar historical events when facing complex business games such as stockpiling behavior caused by raw material price fluctuations.To address the balance between knowledge transfer and updating during model weight iteration, this application proposes an adaptive forgetting factor adjustment algorithm. This algorithm uses the degree of concept drift output by the statistical monitor as the independent variable, constructs a function positively correlated with the degree of concept drift, and dynamically adjusts the weight decay coefficient during model training. When severe concept drift is detected, the system increases the weight decay coefficient, guiding the model to allocate more weight to the loss function of the new samples in the optimization objective, thereby accelerating the fading of old ineffective patterns and enabling the model to capture new market patterns. When the business environment returns to a stable period, i.e., when the prediction residuals return to a statistically controlled state, the system rapidly decreases the weight decay coefficient to maintain stability. By leveraging the model's ability to memorize long-term patterns and reinforcing existing knowledge reserves, the system ensures that when dealing with highly predictive materials such as daily consumables, it can output highly stable prediction curves, avoiding unnecessary inventory level fluctuations caused by parameter oscillations. In specific engineering applications, after the model completes weight iteration via fast or deep paths, the newly generated model version can enter a gray-scale validation period through a traffic switching logic. The system will assign a small portion of non-core material prediction tasks to the new model and continuously compare its performance with the old model during the same business period. Only when the new model significantly outperforms the old model in key business indicators—such as the improvement in prediction accuracy and the reduction in inventory turnover days—will the model be validated. When good stability is achieved, the routing scheduler will smoothly switch the inference tasks for the remaining materials to the new model; this gives the system a self-healing and evolutionary ability similar to a biological nervous system. Through deep integration of material business characteristics, such as identifying inventory anomalies as high-value samples in experience replay and quantifying material turnover fluctuations as the physical basis for statistical offsets, the embodiments of this application not only improve the survivability of artificial intelligence in complex industrial supply chain environments, but also significantly reduce the operational complexity and computational cost of large-scale heterogeneous material prediction systems through refined orchestration of computing and storage resources, providing an indispensable underlying technical foundation for building a highly intelligent and flexible supply chain system; for materials with extreme seasonal characteristics For example, in certain regions, the adaptive update mechanism of the system demonstrates strong predictive ability for flood control materials or agricultural fertilizers. On the eve of seasonal changes, the statistical monitor often detects and identifies small residual deviations caused by advance stocking, thereby triggering warning thresholds in advance and initiating rapid path fine-tuning. This allows the predictive model to always stay ahead of market trends, rather than making passive adjustments after inventory backlog or stockouts occur. In addition, when dealing with structural changes within an enterprise, such as changes in material codes, supplier changes, or warehouse location changes, this application can also complete business logic reconstruction and alignment in a very short time through its powerful semantic mapping and drift monitoring logic, greatly shortening the data integration period for enterprises during the information technology upgrade process.In a multi-branch supply network, this application can even leverage the drift correlation between different branch warehouses for collaborative updates. Specifically, when a material in a regional warehousing center experiences significant conceptual drift, the system automatically assesses whether this conceptual drift exhibits diffusion characteristics and pre-adjusts the confidence threshold of the corresponding material models in other related warehouses. This achieves cross-regional risk warning and synchronous model optimization, not only ensuring the accuracy of individual material forecasts but also enhancing the enterprise's resilience to black swan events in the supply chain from a global perspective, truly realizing an intelligent transformation towards data-driven decision-making.

[0039] This application also discloses the application of the above-mentioned AI training and inference method for material business data in a material supply chain management system. The application includes: applying the aligned sample set to supply chain demand forecasting to achieve accurate prediction of demand for massive material inventory units; using the optimized results of the inference to guide the automatic allocation of materials in the logistics and warehousing system and optimize inventory turnover; and capturing demand changes caused by market fluctuations and policy adjustments through the adaptive update mechanism to achieve dynamic correction of material supply plans.

[0040] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the above-described division of functional modules is used as an example. In practical applications, the above functions can be assigned to different functional modules as needed, that is, the internal structure of the device can be divided into different functional modules to complete all or part of the functions described above. The specific working process of the system, device, and unit described above can be referred to the corresponding process in the foregoing method embodiments, and will not be repeated here.

[0041] In the several embodiments provided in this application, it should be understood that the disclosed systems, apparatuses, and methods can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative; for instance, the division of modules or units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be through some interfaces, or indirect coupling or communication connection between apparatuses or units, and may be electrical, mechanical, or other forms.

[0042] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0043] Furthermore, the functional units in the various embodiments of this application can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit.

[0044] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) or processor to execute all or part of the steps of the methods described in the various embodiments of this application. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory, random access memory, magnetic disks, or optical disks.

[0045] The above are all preferred embodiments of this application and are not intended to limit the scope of protection of this application. Any feature disclosed in this specification (including the abstract and drawings) may be replaced by other equivalent or similar features unless specifically stated otherwise. That is, unless specifically stated otherwise, each feature is only one example of a series of equivalent or similar features.

Claims

1. A method for AI training and inference based on material business data, characterized in that, include: Heterogeneous data alignment based on multidimensional spatiotemporal interpolation grid: Discrete transaction data from the Enterprise Resource Planning (ERP) system and real-time time-series data from the Internet of Things (IoT) system are acquired. A virtual spatial time axis based on unified time synchronization is constructed. The low-frequency discrete transaction data from the ERP system is time-domain mapped using the Lagrange interpolation algorithm to match it to the high-frequency sampling grid of the IoT system. A material association model based on knowledge graph is established. Through the material association model, the material codes from the ERP system, the pallet numbers from the warehousing system, and the waybill numbers from the logistics process are semantically mapped, transforming heterogeneous data from different business processes into a unified-dimensional feature tensor representation. The self-attention mechanism is used to extract the temporal correlation features of heterogeneous data from different sources, identify the business logic offset caused by system delay, and perform pre-compensation calculation for the time lag conflict of heterogeneous data from different sources based on the business logic offset, thereby completing the multi-dimensional spatiotemporal alignment of heterogeneous data and generating an alignment sample set for training artificial intelligence models. Inference optimization based on dynamic grouping and pruning and hot / cold weight routing: According to the material call frequency and prediction error rate, the material inventory units in the aligned sample set are dynamically divided into three grouping types: active, stable, and dormant. Differentiated calculation pruning strategies are executed for different grouping types. In combination with the hardware resource load status, the model sub-weights of the corresponding material categories are routed from the storage medium to the video memory of the computing unit through model parameter sharding loading technology. Incremental inference or full inference is performed to output the material business prediction results. Adaptive updates based on dynamic drift monitoring and experience replay: Deploy a statistical monitor in the artificial intelligence inference pipeline to calculate the prediction residual between the material business forecast results and the actual business feedback values, and quantify the degree of concept drift based on changes in statistical distribution. The dual-path online update mechanism is triggered based on the degree of concept drift. Incremental training is performed through online gradient descent or by fusing the experience replay buffer, and the model weights are iteratively updated in conjunction with an adaptive forgetting factor.

2. The method for AI training and inference of material business data according to claim 1, characterized in that, The inference optimization based on dynamic group pruning and hot / cold weighted routing specifically includes: Real-time statistics on the call frequency and prediction error rate of each material inventory unit within a preset period; Material inventory units whose call frequency is higher than the first threshold and whose prediction error rate fluctuates beyond the second threshold are identified as active and full deep learning inference is performed on them. Material inventory units with call frequency between the first and third thresholds and stable prediction error rates are identified as stationary states, and only lightweight inference based on incremental deviation correction is performed on them. Material inventory units with a call frequency below the third threshold are identified as dormant and low-frequency batch inference tasks are triggered for them.

3. The method for AI training and inference of material business data according to claim 2, characterized in that, The lightweight inference for incremental bias correction is implemented as follows: Extract the hidden layer state features generated by the steady-state material in the previous inference cycle as a reference benchmark; Calculate the feature change vector of the current input business data relative to the previous period's data in the feature space; By using a small, lightweight residual network, feature mapping is performed on this tiny feature change vector, and the resulting residual increment is fused and superimposed with the hidden layer state of the previous cycle, thereby quickly deriving the current prediction result.

4. The method for AI training and inference of material business data according to claim 1, characterized in that, The model parameter piecewise loading technique specifically includes: The global artificial intelligence model is decomposed into multiple independent model weight slices according to the characteristics of material categories; Construct a video memory prediction routing model and pre-calculate the video memory usage based on the grouping attributes and calculation priority of the inventory units of the materials to be inferred. Before the inference task starts, only the model weight slices corresponding to the target group are loaded into the computing core through the direct memory access channel, and the model weight slices are compressed using octet integer quantization technology to reduce the memory overhead of a single inference.

5. The method for AI training and inference of material business data according to claim 1, characterized in that, The process of transforming heterogeneous data from different business processes into a unified-dimensional feature tensor representation specifically includes: Extract static attributes, dynamic flow paths, and environmental features from the material association model; The encoder maps unstructured text descriptions and structured numerical codes to the same high-dimensional vector space. The feature fusion layer concatenates feature vectors of different dimensions into a fixed-length feature tensor, and the normalization layer eliminates the dimensional differences between different heterogeneous systems.

6. The method for AI training and inference of material business data according to claim 1, characterized in that, The dual-path online update mechanism includes a fast path update mode and a deep path update mode. The adaptive update based on dynamic drift monitoring and experience playback specifically includes: Utilize statistical inequalities to monitor data distribution in real time during the reasoning process; When the detected statistical deviation exceeds the preset warning threshold but does not reach the action threshold, the system enters the fast path update mode, using the newly generated business data to fine-tune the top-level weights of the model through an online gradient descent algorithm. When the detected statistical deviation exceeds the preset action threshold, the system enters the deep path update mode, extracts representative historical samples from the experience replay cache, and mixes the historical samples with the current new business data to trigger incremental training of the artificial intelligence model.

7. The method for AI training and inference of material business data according to claim 6, characterized in that, The statistical inequality is the Hofding inequality, and its monitoring process includes: Set the confidence level parameter for the predicted residuals; Continuously calculate the difference between the mean of the predicted residuals within the current sliding window and the historical baseline mean; If the difference satisfies the probability deviation limit defined by the Hofding inequality, then it is determined that a concept drift has occurred in the material business environment.

8. The method for AI training and inference of material business data according to claim 6, characterized in that, The management mechanism of the experience replay cache includes: Historical business data are clustered and sampled based on temporal relevance and feature diversity; Retain representative samples near each cluster center and extreme samples that are difficult to predict; As new samples are added, outdated historical samples are automatically removed based on the sample freshness index to ensure the validity of samples in the buffer area.

9. The method for AI training and inference of material business data according to claim 1, characterized in that, The adaptive forgetting factor is implemented in the following way: Construct a functional relationship that is positively correlated with the degree of concept drift, and dynamically adjust the weight decay coefficient during model training; When severe concept drift is detected, the weight decay coefficient is increased to accelerate the fading of old invalid patterns; When the business environment is in a stable period, reduce the weight decay coefficient to maintain the model's ability to remember long-term patterns.

10. The application of the method for AI training and inference of material business data as described in any one of claims 1 to 9 in a material supply chain management system, characterized in that, The applications include: applying the aligned sample set to supply chain demand forecasting to achieve accurate prediction of demand for massive material inventory units; using the optimized results of the inference to guide the automatic allocation of materials in the logistics and warehousing system and optimize inventory turnover; and capturing demand changes caused by market fluctuations and policy adjustments through the adaptive update mechanism to achieve dynamic correction of material supply plans.

Images