An artificial intelligence-based supply chain logistics information management system

Abstract

This invention relates to the field of logistics management system technology, specifically disclosing an artificial intelligence-based supply chain logistics information management system, including a data acquisition module, a demand analysis module, a logistics decision-making module, a dynamic perception module, and a supply chain monitoring module. This invention solves the data problems of traditional systems through multi-source data acquisition and standardized preprocessing, laying the foundation for analysis; relying on an improved LSTM model and dynamic thresholds, it accurately predicts demand, warns of risks, and reduces inventory and stockout losses; employing hybrid algorithms, it collaboratively optimizes routes, assesses inventory, and dynamically adjusts for environmental interference, balancing logistics efficiency and cost; using an Attention-LSTM model to generate a resilience index, combined with hierarchical response and visualization, it achieves quantitative risk tracing and rapid handling; and the modules collaborate in a real-time closed loop, with parameter self-optimization and threshold updates, adapting to industry and environmental changes and providing stable and intelligent support for supply chain management.

Description

Technical Field

[0001] This invention relates to the field of logistics management system technology, specifically to an artificial intelligence-based supply chain logistics information management system. Background Technology

[0002] Supply chain logistics information management currently faces multiple challenges. Traditional management systems often rely on manual decision-making or single-dimensional data analysis, making them ill-suited to complex and dynamic market environments. At the data processing level, supply chain data related to orders, inventory, logistics, and the external environment is scattered across different sources and formats, lacking efficient collection and preprocessing mechanisms. This results in inconsistent data quality, impacting the accuracy of subsequent analysis. Regarding demand forecasting, traditional models struggle to capture short-term fluctuations and long-term trends in demand, leading to significant forecasting errors and increasing the risk of inventory buildup or stockouts, thus raising supply chain operating costs.

[0003] In the logistics decision-making process, existing systems often employ single optimization algorithms, failing to adequately integrate route efficiency, inventory levels, and external environmental disturbances. This leads to irrational route planning, unbalanced inventory allocation, and low logistics efficiency. Simultaneously, supply chain risk assessment lacks comprehensive quantitative indicators and dynamic monitoring mechanisms, failing to integrate multi-dimensional factors such as demand fluctuations, logistics efficiency, and environmental disturbances in real time. This makes it difficult to accurately identify risk levels and core triggers, resulting in significant response delays. Furthermore, insufficient data synergy between modules and a disconnect between decision-making and monitoring prevent the formation of closed-loop management, leading to insufficient supply chain resilience. This makes it difficult to cope with sudden market changes, policy adjustments, or extreme weather events, severely impacting the stable operation and sustainable development of the supply chain. Summary of the Invention

[0004] To address the shortcomings of existing technologies, this invention provides an artificial intelligence-based supply chain logistics information management system, which solves the problems mentioned in the background.

[0005] To achieve the above objectives, the present invention provides the following technical solution: an artificial intelligence-based supply chain logistics information management system, comprising a data acquisition module, a demand analysis module, a logistics decision-making module, a dynamic perception module, and a supply chain monitoring module;

[0006] The data acquisition module is used to collect real-time data in the supply chain, including historical order data, current inventory data, in-transit logistics data, and socio-economic data, and generates a multimodal feature set through data preprocessing;

[0007] The demand analysis module is used to predict demand trends and calculate demand fluctuation coefficients based on a multimodal feature set and an improved LSTM time series analysis model. ,in accordance with Determine whether there is demand risk in the supply chain. If so, generate and issue an early warning instruction, which includes the risk level, triggering factor and related data identifier.

[0008] The logistics decision-making module receives early warning instructions, extracts logistics-related features from the multimodal feature set, combines environmental data from the dynamic perception module to optimize logistics routes and inventory levels, and calculates the logistics efficiency coefficient. The optimization process employs a hybrid strategy combining genetic algorithms and greedy algorithms.

[0009] The dynamic sensing module is used to monitor changes in the external environment of the supply chain in real time and calculate the environmental interference coefficient. ,based on The optimization results of the logistics decision-making module are dynamically adjusted, and the adjustment frequency is synchronized with the environmental data update frequency.

[0010] The supply chain monitoring module is used to build a supply chain risk assessment model using Attention-LSTM deep learning technology, and to incorporate logistics efficiency coefficients. Demand fluctuation coefficient and environmental interference coefficient The input model is processed by Min-Max normalization, and the fitted output is the supply chain resilience index. And it is presented in a multi-dimensional visualization format.

[0011] Preferably, the data acquisition module includes a data acquisition unit, a data cleaning unit, and a feature extraction unit, and the collaborative workflow of each unit is as follows:

[0012] The data acquisition unit acquires supply chain data through three methods: enterprise ERP system API interface; IoT sensors deployed at warehousing / transportation nodes; third-party open data interfaces; and the collected data is categorized and stored in a distributed database, supporting real-time access.

[0013] The data cleaning unit performs a three-step cleaning process: Step 1.1: Missing value handling, using linear interpolation to fill missing values ​​for numerical data and mode imputation for categorical data; Step 1.2: Outlier detection, using improved... The rules will exceed Data within the specified range is marked as anomalies and replaced with the 90th percentile of that feature; Step 1.3: Standardize the format by converting all data into JSON structured format, unifying timestamps to UTC time zone, and standardizing numerical units;

[0014] The feature extraction unit extracts a multimodal feature set from the cleaned data. The calculation methods for each feature are as follows:

[0015] Demand characteristics: historical demand sequence Demand forecast Demand deviation ratio ;

[0016] Inventory characteristics: Current inventory level Safety stock level Inventory turnover rate Out-of-stock rate ;

[0017] Logistics characteristics: Transportation route length Transportation time Logistics costs Real-time location data;

[0018] Environmental characteristics: economic indices Weather Impact Index Policy Change Index .

[0019] Preferably, the demand analysis module includes a demand trend analysis unit and a risk preliminary assessment unit, wherein the demand trend analysis unit performs a four-step analysis process:

[0020] Step 2.1: Demand data preprocessing, including historical demand sequences. Smoothing is performed to eliminate short-term fluctuations;

[0021] Step 2.2: Demand forecasting modeling, constructing an improved LSTM model, using the past 90 days... For the training set, nearly 15 days For the validation set, the Adam optimizer and MSE loss function are used for training. The iteration stops when the validation set loss decreases by less than 0.01 for 5 consecutive rounds, and the demand prediction value is output. ;

[0022] Step 2.3: Calculate the demand deviation ratio According to the formula Calculation, where This represents the actual demand value for the day. This is the predicted value for the day. but ;

[0023] Step 2.4: Calculate the demand stability index According to the formula Calculation, where , This represents the average demand over the past 90 days. This represents the standard deviation of demand over the past 90 days. Range of values The smaller the value, the more stable the demand;

[0024] The initial risk assessment unit has a built-in demand trend visualization submodule, which supports displaying the trend in line charts. and The changing trends help managers intuitively judge the fluctuations in demand.

[0025] Preferably, the demand fluctuation coefficient The newly added parameter constraints and their values ​​are explained below:

[0026] formula In the code, the parameters are defined and their values ​​are as follows:

[0027] This is the demand deviation weighting coefficient, with a value range of [0.4, 0.6], adjusted according to industry type, used to emphasize... Impact on fluctuations; The weighting coefficient for the impact of stockouts ranges from [0.4, 0.6]. Complementary ; This is a smoothing constant, fixed at 0.01, used to avoid... The denominator is 0 to ensure the validity of the formula calculation; According to the original formula Calculation, where This refers to the number of times the product has been out of stock in the past 30 days. This represents the total number of requests in the past 30 days. The value range is [0, 100%]; The value ranges from [0, 100], with larger values ​​indicating more drastic demand fluctuations.

[0028] The demand analysis module also has a built-in parameter self-optimization function, which fine-tunes parameters quarterly based on historical demand risk event data using the gradient descent method. ,make sure Calculation accuracy improved by ≥5%.

[0029] Preferably, the standard threshold of the risk assessment unit The determination is made using a dual mechanism of data-driven approach and expert calibration. The specific process is as follows:

[0030] Step 3.1: Data-driven determination of initial thresholds, collecting normal supply chain operation data for the past 3 years, and calculating the thresholds for each time period. Take its 95th percentile as the initial threshold. ;

[0031] Step 3.2: Expert calibration and adjustment. Five supply chain management experts were invited to conduct adjustments based on industry characteristics. Adjustments will be made, and the adjustment range will be as follows: final threshold This is the expert-adjusted mean.

[0032] Step 3.3: Dynamically update the threshold, repeating steps 1-2 every six months to update. Value, adapting to changes in the supply chain environment;

[0033] The response process after risk assessment is optimized as follows:

[0034] like If a high-risk demand condition is identified, an early warning instruction will be issued, which will include the risk level, triggering factors, and recommended measures. This instruction will be simultaneously pushed to supply chain management personnel via system pop-ups and SMS messages.

[0035] like If a demand is deemed to be at risk, the warning instruction will only be pushed to the system backend and highlighted in yellow; no SMS notification is required.

[0036] like Determined to be of low risk, no warning will be issued; only records will be kept. This information is stored in the historical database for subsequent threshold updates.

[0037] Preferably, the environmental monitoring unit and interference analysis unit of the dynamic sensing module add a data verification and dynamic adjustment process:

[0038] The environmental monitoring unit establishes a multi-source data cross-validation mechanism, including the following steps:

[0039] Step 4.1: Collect data from two or more data sources simultaneously for the same environmental indicator; Step 4.2: Calculate the deviation rate of each data source. If the deviation rate is ≤5%, take the average as valid data; if the deviation rate is >5%, remove abnormal data sources and replace them with the average of the remaining data sources; Step 4.3: Perform time alignment on the verified environmental data to ensure synchronization with the supply chain data.

[0040] Interference analysis unit, environmental interference coefficient The supplementary values ​​and calculation explanations for the formula parameters are as follows: ;

[0041] in, The weather impact index ranges from 0 to 100. ; As an economic index, For nearly 12 months average value, The value represents the proportion of economic fluctuations. ; This is a policy change index, with a value of 0 or 1. The weighting coefficients are γ1=0.3, γ2=0.4, and γ3=0.3, respectively, satisfying γ1+γ2+γ3=1; The value ranges from [0,1], with larger values ​​indicating stronger external environmental interference; the dynamic sensing module also has an interference factor tracing function, when At that time, the system automatically identifies the interference factor that contributes the most and marks it in the system.

[0042] Preferably, the path optimization unit and inventory management unit of the logistics decision-making module are equipped with multi-objective optimization and constraint control processes:

[0043] The path optimization unit executes a three-step path optimization process:

[0044] Step 5.1: Route data collection, obtain the current transportation route. Transportation time Nodes along the route and ;

[0045] Step 5.2: Feasible path generation. Based on Dijkstra's algorithm combined with a genetic algorithm, generate 3-5 feasible paths. Constraints: path length. Transportation time ;

[0046] Step 5.3: Path optimization, calculate the path optimization rate for each feasible path. According to the formula ,choose The longest path is the optimal path. ),in To optimize the path length, To optimize the transportation time; The value ranges from [0, 100%], with a larger value indicating better path optimization.

[0047] The inventory management unit executes the inventory health assessment and adjustment process:

[0048] Step 6.1: Calculate the safety stock , This means adding a 5% buffer to the demand forecast.

[0049] Step 6.2: Calculate the inventory health index According to the formula ,in Must meet , This represents the inventory turnover rate over the past 30 days.

[0050] Step 6.3: Inventory adjustment recommendations, if It is recommended to increase the purchase quantity; if We suggest reducing the purchase volume or clearing inventory through promotions;

[0051] The logistics decision-making module also has a built-in cost-efficiency balancing submodule. When optimizing the route leads to an increase in logistics costs of ≥10%, it automatically prompts managers to weigh costs and efficiency to avoid over-optimization.

[0052] Preferably, the logistics efficiency coefficient The calculation process for new parameter calibration and collaborative optimization is as follows:

[0053] ;

[0054] in, The path optimization weight is set to 0.5. The inventory health weight is set to 0.5. The environmental interference coefficient is... This is an environmental disturbance correction term, used to reduce the impact of external disturbances on logistics efficiency assessment; The value ranges from [0,1], with a larger value indicating higher logistics efficiency;

[0055] The real-time coordination coefficient The computation and application process is optimized as follows:

[0056] Step 7.1: Calculation According to the formula ,in For actual logistics costs, Budgeting logistics costs;

[0057] Step 7.2: Collaboration judgment, if If the logistics process is deemed to be well-coordinated, the current decision will be maintained; if If the coordination is deemed average, internal coordination optimization is triggered; if... If poor coordination is identified, a cross-departmental collaborative meeting should be initiated.

[0058] Step 7.3: Track the collaborative effect; recalculate every 24 hours after collaborative optimization. until ;

[0059] The logistics decision-making module also has a historical data backtracking function, which can query data from the past year. , The changing trend is used for logistics efficiency improvement analysis.

[0060] Preferably, the model building unit and evaluation output unit of the supply chain monitoring module include a new model training optimization and index interpretation process:

[0061] The model building unit executes a five-step model training process:

[0062] Step 8.1: Dataset partitioning. Divide the multimodal feature set into training, validation, and test sets in a 7:2:1 ratio. Step 8.2: Model architecture design. Use the Attention-LSTM model. Step 8.3: Model training. Use the AdamW optimizer and MAE loss function. Train for 100 epochs, validating every 10 epochs, and save the model with the smallest MAE on the validation set. Step 8.4: Model testing. Evaluate model performance using the test set, requiring a MAE ≤ 0.05 (otherwise, retrain). Step 8.5: Online model updates. Fine-tune model parameters quarterly using new data, ensuring model accuracy decay ≤ 3%.

[0063] Assessment Output Unit: Supply Chain Resilience Index The calculation and interpretation process is as follows:

[0064] Step 9.1: Parameter Calculation, Inventory Stagnation Index According to the formula Calculation, where For slow-moving inventory, Current inventory Inventory turnover rate over the past 30 days; dynamic adjustment factor for logistics efficiency. According to the formula Calculate, with a value range of [0,1]. The larger The closer to 1; This is the collaborative weighting coefficient, with a value of 0.2.

[0065] Step 9.2: Calculate according to the formula , The value range is [0, 100];

[0066] Step 9.3: Index Interpretation High toughness, Medium toughness Low toughness;

[0067] The supply chain monitoring module also has a built-in risk attribution factor module, when... Automatic analysis The contribution percentage is used to identify the main sources of risk.

[0068] Preferably, the supply chain risk assessment and response process adds tiered handling and enhanced visualization functions, as follows:

[0069] Evaluation threshold Determination: Based on historical risk event data, identify events that led to supply chain disruptions. Minimum value as If there are no historical interruptions, then =60;

[0070] Tiered response process:

[0071] High-risk handling, Immediately triggers audible and visual alarms; automatically pushes emergency plans, including activating backup storage, adjusting transportation methods, and temporarily adding suppliers; updates every hour. until ;

[0072] Normal status monitoring Data is reviewed on a 24-hour cycle; a weekly supply chain resilience report is generated, including... Trend analysis and key factor analysis; if A decrease of ≥5% for three consecutive days triggers an early warning.

[0073] This invention provides an artificial intelligence-based supply chain logistics information management system, which has the following beneficial effects:

[0074] 1. Efficient and accurate data processing: Through multi-source data collection channels and standardized preprocessing processes, the system achieves comprehensive acquisition, cleaning and integration of supply chain data, generating a high-quality multimodal feature set, providing reliable data support for subsequent analysis and decision-making, and effectively solving the problems of scattered and poor-quality data in traditional systems.

[0075] 2. Accurate demand forecasting and risk warning: Based on the improved LSTM model for demand trend analysis, combined with the quantitative calculation of demand fluctuation coefficient and dynamic threshold mechanism, the accuracy of demand forecasting is greatly improved. It can identify different levels of demand risks in advance and push early warnings in a timely manner, reducing inventory backlog and stockout losses.

[0076] 3. Optimized and efficient logistics decision-making: A hybrid strategy combining genetic algorithms and greedy algorithms is adopted to coordinate path optimization and inventory health assessment, while incorporating dynamic adjustment of environmental interference coefficients. This achieves a cost-efficiency balance while ensuring logistics efficiency, thereby improving the rationality of logistics resource allocation.

[0077] 4. Comprehensive Risk Assessment and Timely Response: The supply chain risk assessment system, built using the Attention-LSTM model, integrates multi-dimensional indicators such as logistics efficiency, demand fluctuations, and environmental disturbances to generate a supply chain resilience index, enabling quantitative risk assessment and precise traceability. A tiered response mechanism and real-time visualization ensure rapid handling of high-risk events and dynamic monitoring of normal conditions, significantly improving the supply chain's resilience and emergency response speed.

[0078] 5. Strong System Collaboration and Adaptability: Real-time data collaboration between modules enables the system to monitor and optimize the coordination of each link in the logistics process through real-time collaboration coefficients, forming a closed-loop management system. The self-optimization of model parameters and dynamic threshold updates allow the system to adapt to different industry characteristics and environmental changes, maintaining high operational accuracy over the long term and providing continuous and stable intelligent support for supply chain management.

[0079] In summary, this invention addresses the data problems of traditional systems through multi-source data acquisition and standardized preprocessing, laying the foundation for analysis. Relying on an improved LSTM model and dynamic thresholds, it accurately predicts demand and warns of risks, reducing inventory and stockout losses. A hybrid algorithm is employed to collaboratively optimize routes, assess inventory, and dynamically adjust for environmental interference, balancing logistics efficiency and cost. An Attention-LSTM model generates a resilience index, which, combined with tiered response and visualization, enables quantifiable risk tracing and rapid response. Real-time collaborative closed-loop processing across modules, along with parameter self-optimization and threshold updates, adapts to industry and environmental changes, providing stable and intelligent support for supply chain management. Attached Figure Description

[0080] Figure 1 This is a schematic diagram of a supply chain logistics information management system based on artificial intelligence according to the present invention.

[0081] Figure 2 This is a block diagram illustrating the principle of the data acquisition module of the present invention.

[0082] Figure 3 This is a block diagram illustrating the principle of the requirements analysis module of the present invention.

[0083] Figure 4 This is a block diagram illustrating the principle of the dynamic sensing module of the present invention.

[0084] Figure 5 This is a block diagram illustrating the principle of the logistics decision-making module of the present invention.

[0085] Figure 6 This is a block diagram illustrating the principle of the supply chain monitoring module of the present invention. Detailed Implementation

[0086] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0087] like Figure 1As shown, the present invention provides a technical solution: an artificial intelligence-based supply chain logistics information management system, including a data acquisition module, a demand analysis module, a logistics decision-making module, a dynamic perception module, and a supply chain monitoring module;

[0088] The data acquisition module collects real-time data from the supply chain, including historical order data, current inventory data, in-transit logistics data, and socio-economic data. It then preprocesses this data to generate a multimodal feature set, including data cleaning, outlier removal, and feature standardization. The demand analysis module, based on the multimodal feature set, uses an improved LSTM time series analysis model to predict demand trends and calculate the demand volatility coefficient. ,in accordance with The system assesses whether there is demand risk in the supply chain. If so, it generates and issues an early warning instruction, which includes the risk level, triggering factors, and related data identifiers. Upon receiving the early warning instruction, the logistics decision-making module extracts logistics-related features (transportation routes, inventory levels, and logistics costs) from a multimodal feature set. It then combines this with environmental data from the dynamic perception module to optimize logistics routes and inventory levels, and calculates the logistics efficiency coefficient. The optimization process employs a hybrid strategy combining genetic algorithms and greedy algorithms; the dynamic sensing module is used to monitor changes in the external environment of the supply chain (weather, economy, policy) in real time and calculate the environmental interference coefficient. ,based on The optimization results of the logistics decision-making module are dynamically adjusted, with the adjustment frequency synchronized with the environmental data update frequency (every 15 minutes). The supply chain monitoring module uses Attention-LSTM deep learning technology to build a supply chain risk assessment model, and incorporates logistics efficiency coefficients. Demand fluctuation coefficient and environmental interference coefficient The input model is processed by Min-Max normalization, and the fitted output is the supply chain resilience index. It is displayed in a multi-dimensional visualization format (smart dashboard, risk heat map) to support managers in real-time viewing and decision-making.

[0089] More specifically, such as Figure 2 As shown, the data acquisition module includes a data acquisition unit, a data cleaning unit, and a feature extraction unit. The collaborative workflow of each unit is as follows:

[0090] The data acquisition unit acquires supply chain data through three methods: enterprise ERP system API interface (collecting order and inventory data, with an interface response time of ≤1s); IoT sensors deployed at warehousing / transportation nodes (such as GPS positioning sensors and temperature and humidity sensors, with a sampling frequency of 5Hz and a data transmission latency of ≤500ms); and third-party open data interfaces (such as the National Bureau of Statistics economic data API and meteorological department weather data API, updated daily). The collected data is categorized and stored in a distributed database, supporting real-time access.

[0091] The data cleaning unit performs a three-step cleaning process: Step 1.1: Missing value handling. For numerical data, linear interpolation is used to fill in missing values ​​(e.g., when inventory data is missing, interpolation is based on data from the previous and next 12 hours). For categorical data, the mode is used to fill in missing values ​​(e.g., when order status is missing, it is filled with "pending shipment"). Step 1.2: Outlier detection, using improved... The rule (using box plots for non-normally distributed data) will exceed... Data within the range is marked as anomalies and replaced with the 90th percentile of that feature; Step 1.3: Standardize the format, convert all data to JSON structured format, unify the timestamp to UTC time zone, and unify the numerical units (e.g., weight unit is kg, amount unit is yuan).

[0092] The feature extraction unit extracts a multimodal feature set from the cleaned data. The calculation methods for each feature are as follows:

[0093] Demand characteristics: historical demand sequence (Based on daily demand data for the past 90 days), demand forecasts (Output via LSTM model) Demand deviation ratio (Calculated using the original formula, used to quantify the deviation between actual and predicted values).

[0094] Inventory characteristics: Current inventory level (Real-time warehouse system read), safety stock level (based on (Calculated using the 95th percentile + 5% buffer) Inventory turnover rate (Total outbound shipments in the past 30 days / Average inventory), Stockout rate (Number of out-of-stock times in the last 30 days / Total number of demand times);

[0095] Logistics characteristics: Transportation route length (Actual mileage calculated based on GPS trajectory), transit time (Cumulative time from shipment to present), logistics costs (Includes transportation and storage fees, calculated based on actual bills), real-time location data (latitude and longitude coordinates uploaded in real-time by GPS sensors);

[0096] Environmental characteristics: economic indices (e.g., PMI index, obtained from a third-party interface), Weather Impact Index (0-100 points, 80-100 for heavy rain / snowstorms, 0-20 for sunny days), Policy Change Index (0-1, take 1 if there is a major adjustment to logistics policy, take 0 if there is no adjustment).

[0097] In this embodiment, the data acquisition module, through the collaborative operation of the data acquisition unit, data cleaning unit, and feature extraction unit, effectively solves the core problems of single data sources, inconsistent data quality, and insufficient feature dimensions in traditional supply chain logistics management systems. Its effectiveness is mainly reflected in three aspects: First, the data acquisition unit relies on three channels: enterprise ERP system API interface (response ≤1s), IoT sensors at warehousing / transportation nodes (sampling 5Hz, latency ≤500ms), and third-party open data interfaces (daily updates) to achieve real-time and periodic collection of supply chain order, inventory, logistics, and environmental data across all dimensions. Combined with distributed database storage, this ensures real-time data retrieval, providing a comprehensive and timely data foundation for subsequent system analysis. Second, the data cleaning unit improves upon this by using targeted missing value filling (numerical linear interpolation, categorical mode filling) and precise outlier correction. The three-step process of "rule and 90th percentile replacement" - "format standardization and unification (JSON structure, UTC time zone and unified units)" significantly improves data quality, avoids interference from incomplete or abnormal data on analysis results, and ensures data reliability. Thirdly, the feature extraction unit extracts structured features with clear calculation logic from the cleaned data according to four dimensions: demand, inventory, logistics, and environment (such as Ka based on Dt's 95th percentile and 5% buffer, and Wii quantization by weather level), forming a multimodal feature set. This set can directly provide high-quality input for the LSTM prediction model of the demand analysis module and the optimization algorithm of the logistics decision module, laying a key data foundation for the overall intelligent decision-making accuracy of the system and significantly improving the accuracy and efficiency of supply chain logistics management.

[0098] More specifically, such as Figure 3 As shown, the demand analysis module includes a demand trend analysis unit and a risk preliminary assessment unit, wherein the demand trend analysis unit performs a four-step analysis process:

[0099] Step 2.1: Demand data preprocessing, including historical demand sequences. Smoothing is performed (using a moving average method with a 7-day window) to eliminate short-term fluctuations.

[0100] Step 2.2: Demand forecasting modeling, constructing an improved LSTM model (input layer dimension = 7, two hidden layers each with 128 neurons, output layer dimension = 1), based on the past 90 days. For the training set, nearly 15 days For the validation set, the Adam optimizer (learning rate = 0.001) and MSE loss function are used for training. The iteration stops when the validation set loss decreases by less than 0.01 for 5 consecutive rounds, and the demand prediction value is output. ;

[0101] Step 2.3: Calculate the demand deviation ratio According to the formula Calculation, where This represents the actual demand value for the day. This is the predicted value for the day. but (To avoid a denominator of 0);

[0102] Step 2.4: Calculate the demand stability index According to the formula Calculation, where (Data from the last 90 days) This represents the average demand over the past 90 days. This represents the standard deviation of demand over the past 90 days. Range of values The smaller the value, the more stable the demand;

[0103] The initial risk assessment unit has a built-in demand trend visualization submodule, which supports displaying the trend in line charts. and The changing trends help managers intuitively judge the fluctuations in demand.

[0104] In this embodiment, the demand analysis module, through the collaborative operation of the demand trend analysis unit and the risk preliminary assessment unit, effectively solves the core problems of low forecast accuracy and delayed risk identification in traditional supply chain demand management. Its effectiveness is mainly reflected in three aspects:

[0105] The demand trend analysis unit employs a standardized four-step process: demand data preprocessing, modeling and forecasting, deviation calculation, and stability analysis. It first smooths historical demand sequences using a 7-day moving average method. To eliminate short-term fluctuations, an improved LSTM model with an input layer dimension of 7 and two hidden layers of 128 neurons each was constructed. This model, combined with a training set of nearly 90 days, a validation set of 15 days, and a training strategy using the Adam optimizer (learning rate = 0.001) and the MSE loss function, ensures accurate demand predictions. The high accuracy and the iteration stopping condition (the validation set loss decreases by <0.01 for 5 consecutive rounds) further ensure the stability of the model's predictions;

[0106] By demand deviation ratio (including) =0 The denominator protection mechanism (=0) quantifies the degree of deviation between actual and predicted demand, using the demand stability index. (Based on the average demand over the past 90 days) Standard deviation (Calculations show that smaller values ​​indicate greater stability) assess the patterns of demand fluctuations, providing objective quantitative evidence for risk assessment and avoiding biases from subjective experience-based judgments;

[0107] The built-in demand trend visualization submodule in the initial risk assessment unit displays the trend intuitively using line charts. and The changing trends help managers quickly grasp demand dynamics, and combined with the above quantitative indicators, enable accurate and timely identification of demand risks. This provides forward-looking guidance for subsequent logistics decision-making modules such as path optimization and inventory adjustment, effectively reducing inventory backlog or stockout losses caused by inaccurate demand forecasting, and improving the scientific nature and responsiveness of supply chain demand management.

[0108] More specifically, the demand fluctuation coefficient The newly added parameter constraints and their values ​​are explained below:

[0109] formula In the code, the parameters are defined and their values ​​are as follows:

[0110] This is the demand deviation weighting coefficient, with a value range of [0.4, 0.6], adjusted according to industry type (Fast Moving Consumer Goods industry). Heavy industry =0.4), used for emphasis Impact on fluctuations; The weighting coefficient for the impact of stockouts ranges from [0.4, 0.6]. Complementary Industries frequently experiencing stockouts (such as fresh produce) ; This is a smoothing constant, fixed at 0.01, used to avoid... The denominator is 0 to ensure the validity of the formula calculation; According to the original formula Calculation, where This represents the number of times an order could not be fulfilled in the past 30 days. This represents the total number of requests (all order requests) over the past 30 days. The value range is [0, 100%]; The value ranges from [0, 100], with larger values ​​indicating more drastic demand fluctuations.

[0111] The demand analysis module also has a built-in parameter self-optimization function, which fine-tunes the parameters quarterly based on historical demand risk event data (such as stockouts / overstocking events caused by demand fluctuations) using the gradient descent method. ,make sure Calculation accuracy improved by ≥5%.

[0112] Demand volatility coefficient The calculation achieves precise quantification through explicit parameter constraints and value rules; its core formula is: .in, As a weighting coefficient for demand deviation, its value is limited to [0.4, 0.6], and is adjusted according to industry characteristics. For example, the fast-moving consumer goods industry has a more significant impact on the supply chain due to demand fluctuations. Set to 0.6 to enhance The role of (demand deviation ratio) is that demand in heavy industry is relatively stable. Take 0.4; The weighting coefficient for the impact of stockouts, with a value range of [missing information]. Complementary and satisfying + =1, targeting industries such as fresh produce that frequently experience shortages. The value is set to 0.6 to highlight the contribution of stockouts to demand fluctuations; A smoothing constant fixed at 0.01 effectively avoids... The calculation is abnormal when the (demand stability index) is 0, ensuring the validity of the formula; (Out-of-stock rate) is calculated based on the number of out-of-stock events in the past 30 days. ) and total demand frequency ( The percentage calculation, with a value range of [0, 100%], ensures an objective reflection of the stockout situation. The value range is [0, 100], and the magnitude of the value directly corresponds to the severity of demand fluctuations, providing an intuitive quantitative basis for risk assessment. Simultaneously, the demand analysis module's built-in parameter self-optimization function, based on historical demand risk event data (such as stockouts and inventory backlogs), uses a gradient descent method to optimize the parameters quarterly. , Make fine adjustments and ensure This design improves computational accuracy by ≥5%. The calculations can continuously adapt to changes in the supply chain environment and industry characteristics, further improving the accuracy of demand fluctuation assessment, providing reliable data support for subsequent demand risk classification and early warning, reducing assessment bias caused by fixed parameters, and ensuring the timeliness of the supply chain's response to demand fluctuations and the rationality of decision-making.

[0113] More specifically, the standard threshold of the initial risk assessment unit The determination is made using a dual mechanism of data-driven approach and expert calibration. The specific process is as follows:

[0114] Step 3.1: Data-driven determination of initial thresholds, collecting normal supply chain operation data for the past 3 years (time periods without major demand risk events), and calculating the thresholds for each time period. Take its 95th percentile as the initial threshold. ;

[0115] Step 3.2: Expert calibration and adjustment. Invite 5 supply chain management experts (with ≥5 years of industry experience) to adjust the system based on industry characteristics (such as seasonal fluctuations and market competition). Adjustments will be made, and the adjustment range will be as follows: final threshold This is the expert-adjusted mean.

[0116] Step 3.3: Dynamically update the threshold, repeating steps 1-2 every six months to update. Value, adapting to changes in the supply chain environment;

[0117] The response process after risk assessment is optimized as follows:

[0118] like The demand is deemed high-risk, and the warning instruction includes the risk level (high risk) and triggering factors (such as...). Suggested measures (such as temporarily increasing inventory or adjusting procurement plans) are pushed to supply chain managers via system pop-ups and SMS messages.

[0119] like If a demand is deemed to be at risk, the warning instruction will only be pushed to the system backend and highlighted in yellow; no SMS notification is required.

[0120] like Determined to be of low risk, no warning will be issued; only records will be kept. This information is stored in the historical database for subsequent threshold updates.

[0121] The standard threshold of the risk assessment unit in this system The determination is based on a dual mechanism of data-driven approach and expert calibration. The specific implementation process is as follows:

[0122] First, the initial threshold is obtained through data-driven methods. We collected normal supply chain operation data from the past three years without any major demand risk events. After calculating the demand fluctuation coefficient Xqbd for each time period, we took its 95th percentile as H0 to ensure that the threshold was based on real operational data and had an objective foundation. Subsequently, we conducted expert calibration and adjustment, inviting five supply chain management experts with ≥5 years of experience to adjust the data based on industry seasonal fluctuations and market competition characteristics. Adjustments were made, with the adjustment range strictly controlled within ±10%, and the final standard threshold was the average value adjusted by the experts. This addresses the issue of purely data-driven approaches failing to adequately consider industry-specific attributes; simultaneously, it establishes a dynamic threshold update mechanism, repeating the above two-step process every six months. Values ​​are set to ensure that thresholds can adapt to changes in the supply chain environment.

[0123] Regarding the risk response process, according to and Relationship-based hierarchical handling: When Xqbd>H, it is judged as a high-risk demand. The warning instruction includes the risk level, triggering factor and suggested measures, and is pushed simultaneously through system pop-ups and SMS to ensure rapid response by managers; If a situation is identified as a risk in the demand, a yellow alert will be displayed only in the system backend to balance the timeliness of the warning with information interference. At that time, the demand was deemed low-risk, and only records were made. The data is stored in the historical database for subsequent threshold updates, reducing the consumption of invalid warnings.

[0124] This design, through dual-machine threshold setting and tiered response, ensures both the accuracy and industry adaptability of demand risk assessment, and achieves high efficiency and resource optimization in risk warning through dynamic updates and differentiated responses. It effectively avoids the misjudgment and omission problems caused by traditional fixed thresholds, and improves the efficiency of the supply chain's response to demand risks and the rationality of decision-making.

[0125] More specifically, such as Figure 4 As shown, the environmental monitoring unit and interference analysis unit of the dynamic sensing module have added data verification and dynamic adjustment processes:

[0126] The environmental monitoring unit establishes a multi-source data cross-validation mechanism, including the following steps:

[0127] Step 4.1: For the same environmental indicator (such as weather), collect data from two or more data sources simultaneously (such as meteorological API, local weather station sensors); Step 4.2: Calculate the deviation rate of data from each data source. If the deviation rate is ≤5%, take the average as valid data; if the deviation rate is >5%, remove abnormal data sources (such as those with the largest data fluctuations) and replace them with the average of the remaining data sources; Step 4.3: Perform time alignment on the verified environmental data (matching according to UTC timestamps) to ensure synchronization with supply chain data time;

[0128] Interference analysis unit, environmental interference coefficient The supplementary values ​​and calculation explanations for the formula parameters are as follows: ;

[0129] in, The weather impact index ranges from 0 to 100 (heavy rain / typhoon = 90-100, light rain / strong wind = 50-80, cloudy = 30-49, sunny = 0-29). (Fixed maximum value); This refers to economic indices (such as the manufacturing PMI). For nearly 12 months Average (updated monthly) The value represents the proportion of economic fluctuations. ; This is the policy change index, with a value of 0 or 1 (1 for national / local government issuing logistics restriction policies, 0 for no policy adjustment). The weighting coefficients are γ1=0.3 (weather impact), γ2=0.4 (economic impact), and γ3=0.3 (policy impact), respectively, satisfying γ1+γ2+γ3=1; The value ranges from [0,1], with larger values ​​indicating stronger external environmental interference; the dynamic sensing module also has an interference factor tracing function, when At that time, automatically identify the interference factor that contributes the most (such as (60%), and marked in the system.

[0130] To accurately capture changes in the external environment of the supply chain and quantify the degree of interference, the environmental monitoring unit and interference analysis unit of the dynamic perception module of this system have added standardized processes, which are implemented as follows:

[0131] The environmental monitoring unit constructs a multi-source data cross-validation mechanism. The first step involves simultaneously collecting data from two or more data sources (such as meteorological API and local weather station sensors, official economic data and third-party economic indices) for the same environmental indicator, such as weather and economy. The second step involves calculating the deviation rate of each data source. If the deviation rate is ≤5%, the average is taken as valid data. If the deviation rate is >5%, the abnormal data source with the largest fluctuation is removed, and the average of the remaining data is taken to ensure data authenticity. The third step involves aligning the validated data according to the UTC timestamp to ensure synchronization with data such as supply chain orders and inventory, and to avoid analysis errors caused by time deviation.

[0132] The interference analysis unit uses formulas Quantify environmental interference, among which Use a scale of 0-100 for each weather level (90-100 for heavy rain / typhoon, 0-29 for sunny weather). The volatility ratio is calculated based on the average of the past 12 months. Take 0 or 1 depending on the policy adjustment, and , , (Economic impact has the highest weight). The value [0,1] is used to intuitively reflect the interference intensity; at the same time, the module has the function of tracing the source of interference factors. Automatically identify the factor contributing the most (e.g., during strong interference) (60%) and marked.

[0133] By using multi-source verification to address the issues of single and distorted environmental data, and by employing quantitative formulas and weight allocation to achieve accurate assessment of interference levels, combined with traceability functions and time synchronization mechanisms, the system provides a reliable basis for the logistics decision-making module to dynamically adjust routes and inventory. This effectively reduces the impact of extreme weather, economic fluctuations, and policy adjustments on the supply chain, and enhances the system's adaptability and anti-interference capabilities to the external environment.

[0134] More specifically, such as Figure 5 As shown, the path optimization unit and inventory management unit of the logistics decision-making module have added multi-objective optimization and constraint control processes:

[0135] The path optimization unit executes a three-step path optimization process:

[0136] Step 5.1: Route data collection, obtain the current transportation route. Transportation time Nodes along the route (such as warehouses, toll stations) and ;

[0137] Step 5.2: Feasible path generation. Based on Dijkstra's algorithm combined with a genetic algorithm, generate 3-5 feasible paths. Constraints: path length. Transportation time ;

[0138] Step 5.3: Path optimization, calculate the path optimization rate for each feasible path. According to the formula ,choose The longest path is the optimal path. ),in To optimize the path length, To optimize the transportation time; The value ranges from [0, 100%], with a larger value indicating better path optimization.

[0139] The inventory management unit executes the inventory health assessment and adjustment process:

[0140] Step 6.1: Calculate the safety stock , This means adding a 5% buffer to the demand forecast.

[0141] Step 6.2: Calculate the inventory health index According to the formula ,in Must meet (If the range is exceeded, an inventory adjustment will be triggered.) Inventory turnover rate over the past 30 days ( = Total outbound volume in the past 30 days / Average inventory in the past 30 days);

[0142] Step 6.3: Inventory adjustment recommendations, if (Low inventory health), it is recommended to increase the purchase quantity (purchase quantity = );like (Inventory health is high), it is recommended to reduce the purchase volume or promote sales to clear inventory;

[0143] The logistics decision-making module also has a built-in cost-efficiency balancing submodule. When optimizing the route leads to an increase in logistics costs of ≥10%, it automatically prompts managers to weigh costs and efficiency to avoid over-optimization.

[0144] Route optimization unit implementation process: Through the logistics management subsystem and IoT GPS sensors, key current transportation data is obtained, including the current actual transportation route length. (Round trip route A→B→C→D); Current actual transportation time (Including loading and unloading time); Route nodes: Central warehouse in location A, highway toll station in location B, temporary stop in location C; Environmental interference coefficient output by the dynamic sensing module. .

[0145] use The algorithm (fast shortest path search) combines a genetic algorithm (global optimization to avoid local optima) to satisfy the "path length" condition. Transportation time "Under the constraints, four feasible paths are generated:"

[0146] Path 1: A→C→B→D ;

[0147] Path 2: A→B→D→C ;

[0148] Path 3: A→D→C→B ;

[0149] Path 4: A→C→D→B .

[0150] According to the formula Calculate the optimization rate for each path:

[0151] Path 1: ;

[0152] Path 2: ;

[0153] Path 3: ;

[0154] Path 4: ;

[0155] choose The longest path, 3, is the optimal path. Its path length and transportation time both meet the constraints, and the optimization effect is the best after correction for interference from light rain.

[0156] Inventory Management Unit Implementation Process: Calling the daily demand forecast values ​​from the three distribution outlets output by the demand analysis module. (120 boxes in location B, 80 boxes in location C, and 100 boxes in location D, totaling...) =300 boxes), according to the formula Calculate safety stock: A 5% buffer is provided to cope with short-term demand fluctuations.

[0157] Data obtained from the warehouse management system: Current total inventory level =280 boxes ( ,satisfy constraint);

[0158] Total outbound volume in the past 30 days: 9200 boxes; Average inventory in the past 30 days: 300 boxes; Inventory turnover rate: ;

[0159] Out-of-stock occurrences in the last 30 days = 6 times, total demand occurrences = 120 times, out-of-stock rate ;

[0160] According to the formula ,calculate: (If the value is less than 0.6, the inventory health is considered low).

[0161] because The system automatically generates adjustment suggestions: Purchase quantity The boxes were simultaneously sent to the purchasing department and the central warehouse, with a recommendation to prioritize replenishing frequently out-of-stock items in locations B and D.

[0162] The logistics cost (including fuel and toll fees) of optimized route 3 is 8% higher than the original route (below the 10% threshold), and the system does not need to trigger a prompt. However, if the cost of route 3 increases by 12% due to highway maintenance, the system will immediately pop up a window: "The current optimized route 3 improves transportation efficiency by 10.8%, but the logistics cost increases by 12% compared to the original route. It is recommended to choose the second-best route 4 (cost increases by 5%, efficiency increases by 8.5%) or coordinate with the highway management department to apply for toll discounts," to assist managers in weighing the options and making decisions.

[0163] More specifically, the logistics efficiency coefficient The calculation process for new parameter calibration and collaborative optimization is as follows:

[0164] ;

[0165] in, The route optimization weight is set to 0.5 (fixed, because route and inventory are equally important to logistics efficiency). As a weight for inventory health, a value of 0.5 is assigned (fixed, and...). Complementary); This is the environmental interference coefficient (taken from the dynamic sensing module and updated in real time). This is an environmental disturbance correction term, used to reduce the impact of external disturbances on logistics efficiency assessment; The value ranges from [0,1], with a larger value indicating higher logistics efficiency;

[0166] The real-time coordination coefficient The computation and application process is optimized as follows:

[0167] Step 7.1: Calculation According to the formula ,in This represents the actual logistics cost (cumulative over the past 7 days). Budget logistics costs (based on the annual budget allocated to daily costs);

[0168] Step 7.2: Collaboration judgment, if If the logistics process is deemed to be well-coordinated, the current decision will be maintained; if If the coordination is deemed average, internal coordination optimization is triggered (e.g., coordinating warehousing and transportation schedules); if If poor coordination is identified, initiate a cross-departmental coordination meeting (notify the heads of procurement, warehousing, and transportation departments).

[0169] Step 7.3: Track the collaborative effect; recalculate every 24 hours after collaborative optimization. until ;

[0170] The logistics decision-making module also has a historical data backtracking function, which can query data from the past year. , The changing trend is used for logistics efficiency improvement analysis.

[0171] Logistics efficiency coefficient The calculation of new parameter calibration and co-optimization process is specifically implemented as follows: according to the formula Calculation, where (Path optimization weights) and (Inventory health weight) is fixed at 0.5 to ensure that the evaluation weight of the route and inventory on logistics efficiency is equal; The environmental interference coefficient is taken from the dynamic sensing module and updated in real time. This serves as a correction term to reduce the impact of external disturbances on the evaluation results. The value ranges from [0,1] and the larger the value, the higher the logistics efficiency.

[0172] Real-time coordination coefficient The calculation and application process is optimized as follows: First, according to... ( This represents the cumulative actual logistics cost over the past 7 days. Calculate the average daily budget cost allocated to the annual budget. Press again (Good coordination, maintaining decision-making) (General collaboration triggers internal optimization) (Poor collaboration, initiate a cross-departmental meeting) Assess collaboration in a tiered manner; finally, track the optimized collaboration every 24 hours. Until the standard is met.

[0173] Meanwhile, the logistics decision-making module has nearly 1 year of experience. Historical data review function is used for efficiency improvement analysis.

[0174] pass Multi-dimensional calibration calculations enable accurate assessment of logistics efficiency, leveraging The tiered response and tracking assurance links are coordinated, and historical backtracking is used to provide a basis for optimization, effectively solving the problems of one-sided traditional logistics efficiency assessment and lagging collaborative response, and improving the accuracy and stability of supply chain logistics operations.

[0175] More specifically, such as Figure 6 As shown, the model building unit and evaluation output unit of the supply chain monitoring module now include a new model training optimization and index interpretation process:

[0176] The model building unit executes a five-step model training process:

[0177] Step 8.1: Dataset partitioning. Divide the multimodal feature set into training, validation, and test sets in a 7:2:1 ratio (data volume ≥ 1000 records). Step 8.2: Model architecture design. Use the Attention-LSTM model (input layer dimension = 4, hidden layer 128 neurons, attention layer 8 attention heads, output layer dimension = 1). Step 8.3: Model training. Use the AdamW optimizer (learning rate = 0.0001, weight decay = 0.01), MAE loss function, train for 100 epochs, validate every 10 epochs, and save the model with the smallest MAE on the validation set. Step 8.4: Model testing. Evaluate model performance using the test set, requiring a test set MAE ≤ 0.05 (otherwise retrain). Step 8.5: Online model updates. Fine-tune model parameters quarterly using new data (last 3 months) to ensure model accuracy decay ≤ 3%.

[0178] Assessment Output Unit: Supply Chain Resilience Index The calculation and interpretation process is as follows:

[0179] Step 9.1: Parameter Calculation, Inventory Stagnation Index According to the formula Calculation, where This refers to slow-moving inventory (inventory turnover days > 90 days). Current inventory Inventory turnover rate over the past 30 days; dynamic adjustment factor for logistics efficiency. According to the formula Calculate, with a value range of [0,1]. The larger The closer to 1; This is the synergy weighting coefficient, with a value of 0.2 (fixed, emphasizing the auxiliary role of synergy in resilience).

[0180] Step 9.2: Calculate according to the formula , The value range is [0, 100];

[0181] Step 9.3: Index Interpretation High resilience (strong ability of the supply chain to withstand risks). Medium toughness Low toughness;

[0182] The supply chain monitoring module also has a built-in risk attribution factor module, when... Automatic analysis The contribution percentage is used to identify the main sources of risk.

[0183] Model building unit implementation process: Divide the multimodal feature set (data volume needs to be ≥1000, such as 1200 order, inventory, logistics and environmental data of FMCG companies) into training set, validation set and test set in a 7:2:1 ratio. In this embodiment, the training set has 840 data points, the validation set has 240 data points and the test set has 120 data points to ensure that the data distribution is balanced to avoid model overfitting.

[0184] The Attention-LSTM deep learning model is adopted, with the following structural parameters: input layer dimension = 4 (corresponding to four core features: logistics efficiency, demand fluctuation, inventory health, and environmental interference), hidden layer with 128 neurons (balancing model fitting ability and computational efficiency), attention layer with 8 attention heads (focusing on key features that significantly affect resilience), and output layer dimension = 1 (corresponding to the predicted value of the supply chain resilience index).

[0185] The AdamW optimizer (learning rate = 0.0001, weight decay = 0.01, suppressing parameter oscillations) and MAE (mean absolute error) loss function were used to train 100 [samples / samples]. ; per 10 The model performance is evaluated using a validation set, and the model with the smallest MAE on the validation set is saved. In this embodiment, the optimal model validation set is... .

[0186] To evaluate the model's generalization ability using a test set, the test set must be... (Otherwise, return to step 8.3 and retrain). This example uses the test set. It meets the performance requirements.

[0187] We collect nearly three months of new data each quarter (such as new order fluctuations and logistics cost changes) and fine-tune the model parameters to ensure that the model accuracy decays by ≤3% and adapts to the dynamic changes in the supply chain environment.

[0188] Assessment Output Unit Implementation Process: The assessment output unit calculates and interprets the supply chain resilience index in three steps. To achieve resilience quantification and risk positioning:

[0189] Inventory sluggishness index According to the formula Calculation, where This example defines slow-moving inventory (inventory turnover days > 90 days). box), For the current inventory (in this example) box), Inventory turnover rate over the past 30 days (in this example) (times / months), substituting into... .

[0190] Logistics efficiency dynamic adjustment factor According to the formula calculate( (where is a natural constant), and its value ranges from [0,1]. The larger the (logistics efficiency coefficient) is, the better. The closer to 1, the better in this embodiment. Substituting into .

[0191] Collaborative weighting coefficient : Fixed value of 0.2, emphasizing the auxiliary impact of real-time collaboration coefficient on supply chain resilience.

[0192] According to the formula Calculation, where Demand fluctuation coefficient (in this embodiment) ), Real-time coordination coefficient (in this embodiment) Substituting, we get:

[0193] (Value range [0,100]).

[0194] Toughness rating classification: It has high toughness (strong risk resistance). For medium toughness, Low toughness; this embodiment It was determined to be of low toughness.

[0195] Risk attribution factor module: when Automatic analysis The contribution percentage in this embodiment It accounts for 95% of the contribution, and the main risk is identified as demand fluctuation (such as forecast deviation caused by a sudden drop in order volume).

[0196] This design ensures the accuracy and long-term stability of supply chain resilience assessment through a standardized model training process, avoiding assessment biases caused by the static nature of traditional models. By combining multi-dimensional indicators such as demand, inventory, and collaboration, a quantitative expression of supply chain resilience is achieved, solving the problems of "qualitative and vague" monitoring in traditional methods. The risk attribution module can quickly pinpoint the core causes of low resilience, avoiding blind adjustments. In this embodiment, based on the conclusion that "demand fluctuation is the main risk," the company optimized the training parameters of the improved LSTM demand forecasting model, and within one month... Dropped to 8, The index rose to 82, successfully achieving highly resilient operations, reducing logistics costs by 12% and the out-of-stock rate to 1.5%.

[0197] More specifically, the supply chain risk assessment and response process now includes enhanced tiered handling and visualization features, as detailed below:

[0198] Evaluation threshold Determination: Based on historical risk event data, identify events that led to supply chain disruptions. Minimum value as (like =60), if there are no interruption events in history, then =60 (industry default value);

[0199] Tiered response process:

[0200] High-risk handling, Immediately triggers audible and visual alarms (system backend + administrator mobile app); automatically pushes emergency plans, including activating backup warehouses (listing the 3 nearest backup warehouses), adjusting transportation methods (e.g., switching from road to rail), and temporarily adding suppliers (recommending 2-3 alternative suppliers); updates every hour. until ;

[0201] Normal status monitoring Data is reviewed on a 24-hour cycle (samples are taken every 6 hours, and calculations are performed). A weekly supply chain resilience report is generated, including... Trend analysis and key factor analysis; if A decrease of ≥5% for 3 consecutive days triggers an early warning (medium risk alert).

[0202] Visualization Enhancement:

[0203] Smart Dashboard: Segmented Display (Central Key Indicators) (Surrounding sub-indicators), indicator color code: green ( ),yellow( ),red( );

[0204] Real-time Alert Graph: Displays historical alert events in a timeline format. Clicking on an event allows you to view the triggering cause, response measures, and results.

[0205] Trend analysis chart: Supports selection of 1-12 months. The changing trends can be overlaid with industry averages to help managers predict changes in supply chain resilience.

[0206] The supply chain monitoring module also supports data export, which can export Scrx and related data to Excel / PDF format for report generation and auditing.

[0207] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. An artificial intelligence-based supply chain logistics information management system, characterized by: It includes a data acquisition module, a demand analysis module, a logistics decision-making module, a dynamic sensing module, and a supply chain monitoring module; The data acquisition module is used to collect real-time data in the supply chain, including historical order data, current inventory data, in-transit logistics data, and socio-economic data, and generates a multimodal feature set through data preprocessing; The demand analysis module is used to predict demand trends and calculate demand fluctuation coefficients based on a multimodal feature set and an improved LSTM time series analysis model. ,in accordance with Determine whether there is demand risk in the supply chain. If so, generate and issue an early warning instruction, which includes the risk level, triggering factor and related data identifier. The logistics decision-making module receives early warning instructions, extracts logistics-related features from the multimodal feature set, combines environmental data from the dynamic perception module to optimize logistics routes and inventory levels, and calculates the logistics efficiency coefficient. The optimization process employs a hybrid strategy combining genetic algorithms and greedy algorithms. The dynamic sensing module is used to monitor changes in the external environment of the supply chain in real time and calculate the environmental interference coefficient. ,based on The optimization results of the logistics decision-making module are dynamically adjusted, and the adjustment frequency is synchronized with the environmental data update frequency. The supply chain monitoring module is used to build a supply chain risk assessment model using Attention-LSTM deep learning technology, and to incorporate logistics efficiency coefficients. Demand fluctuation coefficient and environmental interference coefficient The input model is processed by Min-Max normalization, and the fitted output is the supply chain resilience index. And it is presented in a multi-dimensional visual format; The data acquisition module includes a data acquisition unit, a data cleaning unit, and a feature extraction unit. The collaborative workflow of each unit is as follows: The data acquisition unit acquires supply chain data through three methods: enterprise ERP system API interface; IoT sensors deployed at warehousing / transportation nodes; third-party open data interfaces; and the collected data is categorized and stored in a distributed database, supporting real-time access. The data cleaning unit performs a three-step cleaning process: Step 1.1: Missing value handling, using linear interpolation to fill missing values ​​for numerical data and mode imputation for categorical data; Step 1.2: Outlier detection, using improved... The rules will exceed Data within the specified range is marked as anomaly and replaced with the 90th percentile of that feature; Step 1.3: Standardize the format by converting all data into JSON structured format, unifying timestamps to UTC time zone, and standardizing numerical units; The feature extraction unit extracts a multimodal feature set from the cleaned data. The calculation methods for each feature are as follows: Demand characteristics: historical demand sequence Demand forecast Demand deviation ratio ; Inventory characteristics: Current inventory level Safety stock level Inventory turnover rate Out-of-stock rate ; Logistics characteristics: Transportation route length Transportation time Logistics costs Real-time location data; Environmental characteristics: economic indices Weather Impact Index Policy Change Index ; The demand analysis module includes a demand trend analysis unit and a risk assessment unit, wherein the demand trend analysis unit performs a four-step analysis process: Step 2.1: Demand data preprocessing, including historical demand sequences. Smoothing is performed to eliminate short-term fluctuations; Step 2.2: Demand forecasting modeling, constructing an improved LSTM model, using the past 90 days... For the training set, nearly 15 days For the validation set, the Adam optimizer and MSE loss function are used for training. The iteration stops when the validation set loss decreases by less than 0.01 for 5 consecutive rounds, and the demand prediction value is output. ; Step 2.3: Calculate the demand deviation ratio According to the formula Calculation, where This represents the actual demand value for the day. This is the predicted value for the day. but ; Step 2.4: Calculate the demand stability index According to the formula Calculation, where , This represents the average demand over the past 90 days. This represents the standard deviation of demand over the past 90 days. Range of values The smaller the value, the more stable the demand; The initial risk assessment unit has a built-in demand trend visualization submodule, which supports displaying the trend in line charts. and The changing trends help managers intuitively judge the fluctuations in demand.

2. A supply chain logistics information management system based on artificial intelligence according to claim 1, characterized in that, Demand volatility coefficient The newly added parameter constraints and their values ​​are explained below: formula In the code, the parameters are defined and their values ​​are as follows: This is the demand deviation weighting coefficient, with a value range of [0.4, 0.6], adjusted according to industry type, used to emphasize... Impact on fluctuations; The weighting coefficient for the impact of stockouts ranges from [0.4, 0.6]. Complementary ; This is a smoothing constant, fixed at 0.01, used to avoid... The denominator is 0 to ensure the validity of the formula calculation; According to the original formula Calculation, where This refers to the number of times the product has been out of stock in the past 30 days. This represents the total number of requests in the past 30 days. The value range is [0, 100%]; The value ranges from [0, 100], with larger values ​​indicating more drastic demand fluctuations. The demand analysis module also has a built-in parameter self-optimization function, which fine-tunes parameters quarterly based on historical demand risk event data using the gradient descent method. ,make sure Calculation accuracy improved by ≥5%.

3. A supply chain logistics information management system based on artificial intelligence according to claim 2, characterized in that, The standard threshold of the risk assessment unit The determination is made using a dual mechanism of data-driven approach and expert calibration. The specific process is as follows: Step 3.1: Data-driven determination of initial thresholds, collecting normal supply chain operation data for the past 3 years, and calculating the thresholds for each time period. Take its 95th percentile as the initial threshold. ; Step 3.2: Expert calibration and adjustment. Five supply chain management experts were invited to conduct adjustments based on industry characteristics. Adjustments will be made, and the adjustment range will be as follows: final threshold This is the expert-adjusted mean. Step 3.3: Dynamically update the threshold, repeating steps 3.1-3.2 every six months to update. Value, adapting to changes in the supply chain environment; The response process after risk assessment is optimized as follows: like If a high-risk demand condition is identified, an early warning instruction will be issued, which will include the risk level, triggering factors, and recommended measures. This instruction will be simultaneously pushed to supply chain management personnel via system pop-ups and SMS messages. like If a demand is deemed to be at risk, the warning instruction will only be pushed to the system backend and highlighted in yellow; no SMS notification is required. like Determined to be of low risk, no warning will be issued; only records will be kept. This data is stored in the historical database for subsequent threshold updates.

4. A supply chain logistics information management system based on artificial intelligence according to claim 3, characterized in that, The environmental monitoring unit and interference analysis unit of the dynamic sensing module have added data verification and dynamic adjustment processes: The environmental monitoring unit establishes a multi-source data cross-validation mechanism, including the following steps: Step 4.1: Collect data from two or more data sources simultaneously for the same environmental indicator; Step 4.2: Calculate the deviation rate of each data source. If the deviation rate is ≤5%, take the average as valid data; if the deviation rate is >5%, remove abnormal data sources and replace them with the average of the remaining data sources; Step 4.3: Perform time alignment on the verified environmental data to ensure synchronization with the supply chain data. Interference analysis unit, environmental interference coefficient The supplementary values ​​and calculation explanations for the formula parameters are as follows: ; in, The weather impact index ranges from 0 to 100. ; As an economic index, For nearly 12 months average value, The value represents the proportion of economic fluctuations. ; This is a policy change index, with a value of 0 or 1. These are the weighting coefficients, with values ​​of [values ​​to be filled in]. =0.3、 =0.4、 =0.3, which satisfies + + =1; The value ranges from [0,1], with larger values ​​indicating stronger external environmental interference; the dynamic sensing module also has an interference factor tracing function, when At that time, the system automatically identifies the interference factor that contributes the most and marks it in the system.

5. A supply chain logistics information management system based on artificial intelligence according to claim 4, characterized in that, The path optimization unit and inventory management unit of the logistics decision-making module have added multi-objective optimization and constraint control processes: The path optimization unit executes a three-step path optimization process: Step 5.1: Route data collection, obtain the current transportation route. Transportation time Nodes along the route and ; Step 5.2: Feasible path generation. Based on Dijkstra's algorithm combined with a genetic algorithm, generate 3-5 feasible paths. Constraints: path length. Transportation time ; Step 5.3: Path optimization, calculate the path optimization rate for each feasible path. According to the formula ,choose The longest path is the optimal path. ),in To optimize the path length, To optimize the transportation time; The value ranges from [0, 100%], with a larger value indicating better path optimization. The inventory management unit executes the inventory health assessment and adjustment process: Step 6.1: Calculate the safety stock , This means adding a 5% buffer to the demand forecast; Step 6.2: Calculate the inventory health index According to the formula ,in Must meet , This represents the inventory turnover rate over the past 30 days. Step 6.3: Inventory adjustment recommendations, if It is recommended to increase the purchase quantity; if We suggest reducing the purchase volume or clearing inventory through promotions; The logistics decision-making module also has a built-in cost-efficiency balancing submodule. When optimizing the route leads to an increase in logistics costs of ≥10%, it automatically prompts managers to weigh costs and efficiency to avoid over-optimization.

6. A supply chain logistics information management system based on artificial intelligence according to claim 5, characterized in that, The logistics efficiency coefficient The calculation process for new parameter calibration and collaborative optimization is as follows: ； in, The path optimization weight is set to 0.

5. The inventory health weight is set to 0.

5. The environmental interference coefficient is... This is an environmental disturbance correction term, used to reduce the impact of external disturbances on logistics efficiency assessment; The value ranges from [0,1], with a larger value indicating higher logistics efficiency; Real-time coordination coefficient The computation and application process is optimized as follows: Step 7.1: Calculation According to the formula ,in For actual logistics costs, Budgeting logistics costs; Step 7.2: Collaboration judgment, if If the logistics process is deemed to be well-coordinated, the current decision will be maintained; if If the coordination is deemed average, internal coordination optimization is triggered; if... If poor coordination is identified, a cross-departmental collaborative meeting should be initiated. Step 7.3: Track the collaborative effect; recalculate every 24 hours after collaborative optimization. until ; The logistics decision-making module also has a historical data backtracking function, which can query data from the past year. , The changing trend is used for logistics efficiency improvement analysis.

7. A supply chain logistics information management system based on artificial intelligence according to claim 6, characterized in that, The model building unit and evaluation output unit of the supply chain monitoring module now include a new model training optimization and index interpretation process: The model building unit executes a five-step model training process: Step 8.1: Dataset partitioning. Divide the multimodal feature set into training, validation, and test sets in a 7:2:1 ratio. Step 8.2: Model architecture design. Use the Attention-LSTM model. Step 8.3: Model training. Use the AdamW optimizer and MAE loss function. Train for 100 epochs, validating every 10 epochs, and save the model with the smallest MAE on the validation set. Step 8.4: Model testing. Evaluate model performance using the test set. The MAE on the test set should be ≤0.05; otherwise, retrain. Step 8.5: Online model updates. Fine-tune model parameters quarterly using new data to ensure that model accuracy decay is ≤3%. Assessment Output Unit: Supply Chain Resilience Index The calculation and interpretation process is as follows: Step 9.1: Parameter Calculation, Inventory Stagnation Index According to the formula Calculation, where For slow-moving inventory, Current inventory Inventory turnover rate over the past 30 days; dynamic adjustment factor for logistics efficiency. According to the formula Calculate, with a value range of [0,1]. The larger The closer to 1; This is the collaborative weighting coefficient, with a value of 0.

2. Step 9.2: Calculate according to the formula , The value range is [0, 100]; Step 9.3: Index Interpretation High toughness, Medium toughness Low toughness; The supply chain monitoring module also has a built-in risk attribution factor module, when... Automatic analysis The contribution percentage is used to identify the main sources of risk.

8. A supply chain logistics information management system based on artificial intelligence according to claim 7, characterized in that, The supply chain risk assessment and response process now includes enhanced tiered handling and visualization features, as detailed below: Evaluation threshold Determination: Based on historical risk event data, identify events that led to supply chain disruptions. Minimum value as If there are no historical interruptions, then =60; Tiered response process: High-risk handling, Immediately triggers audible and visual alarms; automatically pushes emergency plans, including activating backup storage, adjusting transportation methods, and temporarily adding suppliers; updates every hour. until ; Normal status monitoring Data is reviewed on a 24-hour cycle; a weekly supply chain resilience report is generated, including... Trend analysis and key factor analysis; if A decrease of ≥5% for three consecutive days triggers an early warning.

Images