Multi-base inventory optimization method and system based on deep learning

Abstract

The invention discloses a deep learning-based multi-base inventory optimization method and system, and the method comprises the steps: collecting multi-base historical inventory data, carrying out the data preprocessing, and constructing time series data, and taking the time series data as the input characteristics of inventory optimization; introducing a parallel multi-thread mechanism to improve an LSTM network model, training the improved LSTM network model by using the preprocessed time sequence data, adjusting model parameters through back propagation to improve prediction precision, and preliminarily obtaining an ideal model for predicting and outputting a multi-feature inventory prediction result; and on the basis of the prediction result of the LSTM model, an optimized inventory management strategy is generated, and linkage feedback with a material control end of a production field is further carried out, so that intelligent inventory monitoring management and material allocation are realized. According to the method, the material inventory data of each process of the production line on the production site is collected, the preprocessing of the multi-source and multi-mode heterogeneous data is combined, the LSTM model is used for demand prediction, and the optimal inventory management strategy is provided in combination with the optimization algorithm, so that the efficiency and accuracy of multi-base inventory management are improved.

Description

Technical Field

[0001] The present invention belongs to the technical field of deep learning, and particularly relates to a multi-site inventory optimization method for deep learning. Background Art

[0002] With the development of industry and manufacturing, in order to improve the utilization rate of equipment and production efficiency, enterprises often reserve a large amount of inventory at multiple sites. However, current multi-site inventory management faces many challenges, making it difficult for enterprises to efficiently manage and optimize inventory.

[0003] Inaccurate demand forecasting is a major problem in multi-site inventory management. Traditional inventory management methods usually rely on historical data and manual experience for demand forecasting, but the volatility of market demand makes it difficult for these methods to accurately capture trends. Inaccurate demand forecasting often leads to excessive or shortage of inventory, bringing unnecessary inventory backlog costs or supply interruption risks to enterprises.

[0004] Difficulty in optimizing inventory levels is also an important challenge. In a multi-site environment, the inventory demands of different sites vary greatly and may be interdependent. For example, a change in the inventory demand of one site may affect the inventory strategies of other sites. Manual management methods are difficult to consider the inventory levels of multiple sites and their mutual influences simultaneously, lacking a globally optimal inventory management strategy. This local optimization makes it difficult for enterprises to coordinate inventory at the overall level, thus increasing management complexity and operating costs.

[0005] The problem of low information processing efficiency is also particularly prominent in multi-site inventory management. Since managers need to process a large amount of inventory data, demand forecasting data, and replenishment records, traditional manual management methods are inefficient and error-prone when dealing with large-scale data processing. Especially when the market changes rapidly, the reaction speed of manually adjusting inventory strategies often cannot keep up with the changes in actual demand.

[0006] Multi-site inventory management also faces the challenge of coping with emergencies. The weak ability to cope with emergencies makes it difficult for enterprises to adjust inventory levels in a timely manner when encountering emergencies such as equipment failures and sudden increases in market demand. Traditional manual management methods usually react slowly and fail to make rapid adjustments when emergencies occur, resulting in production interruptions or increased operating costs.

[0007] Based on the above problems, the complexity of multi-site inventory management has become increasingly prominent. The complexity of demand forecasting is particularly notable. In a multi-site or multi-location operating environment, the demand patterns and trends vary across different locations, making it increasingly difficult to accurately predict demand. Although traditional inventory management methods can solve problems to a certain extent, in the face of increasingly complex and volatile market demands, traditional rules of thumb and simple statistical models are inadequate. Therefore, modern enterprises need to leverage deep learning technologies to address these complex challenges.

[0008] On this basis, specifically in combination with the material inventory management of each process on the production line at the production site, the existing problems in this field are further explored. These problems not only exist commonly in multi-site inventory management but are also particularly evident in actual production scenarios. The following is a specific problem analysis and the deficiencies of current methods. The material inventory management of the production line at the production site, especially in a multi-site environment, faces multiple challenges such as difficult demand forecasting, inventory imbalance, lagging information processing, insufficient ability to respond to emergencies, and difficult supply chain coordination.

[0009] In the specific application of multi-site inventory management, especially in the material management of each process on the production line, the deficiencies of traditional methods are more significant. There is a high degree of dependence between the production processes on the production line. The material demand of a certain process not only depends on the consumption of that process but is also affected by the production rhythm of other processes and the upstream raw material supply situation. This causes frequent fluctuations in the material inventory of each process and significantly increases the difficulty of demand forecasting. Traditional inventory management methods mainly rely on historical data and simple statistical models to predict material demand, but these methods are difficult to capture dynamic changes and complex time-series dependencies. For example, on a certain production line, if the production speed of the downstream process suddenly increases, it may lead to a rapid increase in the material demand of the upstream process. However, due to the lack of real-time adjustment ability, traditional inventory management systems often struggle to respond in a timely manner, resulting in problems such as inventory shortages or surpluses.

[0010] In addition to the difficulties brought about by demand fluctuations, the phenomena of overstock and shortage are also common problems in the production line. Due to the tendency of traditional inventory management methods to optimize locally, each process or site manages inventory independently, and this approach cannot take into account the coordination and allocation of the overall inventory. For example, during the production process in a multi-site factory, one site may face the problem of inventory backlog, while another site is in a dilemma of material shortage. This imbalance not only increases logistics and warehousing costs but also affects the continuity and efficiency of production. Therefore, relying solely on the independent decisions of each process or site and lacking a global scheduling and optimization inventory management strategy, it is difficult to effectively solve the problem of material inventory imbalance.

[0011] Multi-site inventory management also faces challenges in information processing and decision-making response lags. Due to differences in material consumption and demand changes among various sites and processes, managers need to process a large amount of inventory data in real time. Traditional manual management methods usually rely on manual input and processing, with untimely information updates, which easily lead to overly long decision-making chains and an inability to quickly respond to changing demands in production. For example, when the material consumption of a certain process is too fast and the inventory is approaching depletion, the manual management method may not be able to make a replenishment decision in time, thereby leading to the risk of production stagnation or material shortage. Therefore, in the face of a highly variable production environment, traditional inventory management systems are difficult to ensure real-time response capabilities.

[0012] The ability to respond to emergencies is also a major challenge in material management of production lines. During the production process, sudden surges in market demand, supply chain disruptions, or equipment failures will all put pressure on material inventories. Traditional inventory management methods are usually based on preset inventory thresholds and fixed replenishment rules, and they are slow to respond and unable to adjust flexibly when dealing with emergencies. For example, when market demand suddenly rises, the inventory management system cannot immediately increase material replenishment, resulting in limited production capacity and the enterprise missing market opportunities. Therefore, traditional methods lack emergency response plans and dynamic adjustment capabilities, and appear passive and lagging when dealing with emergencies.

[0013] The coordination problem between the supply chain and multiple sites also limits the effectiveness of existing inventory management methods. In multi-site inventory management, the coordination of production and the supply chain is crucial. However, existing management methods are usually based on independent management of each site, making it difficult to monitor and optimize inventory scheduling among various sites in real time. For example, when the material inventory of a certain site is exhausted, it is impossible to transfer materials from other sites in time, or due to the lag in information transmission, delays occur during the transfer process, affecting the production schedule. Therefore, traditional inventory management systems have obvious deficiencies in multi-site coordination and are unable to achieve efficient utilization of resources. Summary of the Invention

[0014] Aiming at problems such as inaccurate demand forecasting, difficult inventory optimization, low information processing efficiency, and insufficient ability to respond to emergencies in existing multi-site inventory management technologies, the present invention provides a multi-site inventory optimization method based on a long short-term memory network (LSTM). This method collects material inventory data of each process on the production site assembly line, combines the preprocessing of multi-source multi-modal heterogeneous data, uses the LSTM model for demand forecasting, and combines optimization algorithms to provide the best inventory management strategy, thereby improving the efficiency and accuracy of multi-site inventory management.

[0015] The technical solution adopted by the present invention to achieve the above object is:

[0016] A deep learning-based multi-site inventory optimization method, comprising the following steps:

[0017] Collect the historical inventory data of multiple bases for data preprocessing, and construct time series data as the input features for inventory optimization;

[0018] Introduce a parallel multi-threaded mechanism to improve the LSTM network model, use the preprocessed time series data to train the improved LSTM network model, adjust the model parameters through backpropagation to improve the prediction accuracy, and initially obtain an ideal model for predicting and outputting multi-feature inventory prediction results;

[0019] Based on the prediction results of the LSTM model, generate an optimized inventory management strategy, and further link and feedback with the material control end at the production site to achieve intelligent inventory monitoring management and material allocation.

[0020] The historical inventory data of the multiple bases is multi-source and multi-modal heterogeneous data obtained from various sensors and production control systems in each process of the industrial production site assembly line, including multiple dimensions of production and logistics.

[0021] The multi-source and multi-modal heterogeneous data includes: inventory management system data representing inventory levels, inbound and outbound records, and replenishment situations; production line sensor data representing material consumption rates, production speeds, and equipment status; logistics system data representing material transportation times and inventory transfer records; and market demand data representing historical order data and market demand forecasts.

[0022] The preprocessing is as follows:

[0023] Data cleaning: Clean noise data and fill in missing values to ensure data integrity;

[0024] Normalization processing: Unify the scales of data from different sources through normalization or standardization processing to eliminate the difference in measurement units;

[0025] Time series construction: Align multi-modal data from different sources according to time to construct unified time series data to capture the temporal dependence relationships between variables;

[0026] Feature selection and dimensionality reduction: Extract the features that have the most influence on inventory management through feature selection algorithms to reduce the data dimension and improve the model training efficiency.

[0027] The introduction of a parallel multi-threaded mechanism to improve the LSTM network model is to set the number of multi-threads according to the dimension of the input features, enabling parallel calculation of different features or data dimensions; each thread is used to process one feature, and the calculation tasks are executed simultaneously in a multi-threaded environment to improve the calculation efficiency of the model.

[0028] The LSTM network model is a long short-term memory network model; the improved LSTM network model adopts a multi-layer structure, sets multiple parallel LSTM layers, and is followed by a fully connected layer;

[0029] The multi-layer LSTM structure is used to process complex time series relationships and accurately capture the material demand trends of different processes in multi-site and multi-process scenarios; several LSTM units are stacked in each parallel single thread, and each LSTM unit interacts with its corresponding time step data to capture the short-term and long-term dependencies of the data and transmit them to the LSTM unit of the next time step;

[0030] The fully connected layer is used to summarize and analyze the data processed by multiple threads, perform in-depth learning and weight allocation on the relationships between these features, and generate the final demand prediction output.

[0031] Training the improved LSTM network model and adjusting the model parameters through backpropagation to improve the prediction accuracy includes:

[0032] Dividing the time series data into a training set, a validation set, and a test set in proportion; the training set is used for training, the validation set is used to verify the model accuracy, and the test set is used for testing;

[0033] Set the sliding window length and model training parameters for time series prediction; through the iterative update of the sliding window, the model is trained or updated based on new historical data to gradually predict future data requirements;

[0034] After each sliding window update, the model continuously reduces the error through continuous adjustment and iterative optimization;

[0035] After comparing the predicted value in the new window with the true value, the weights are updated based on the feedback and retrained to improve the prediction accuracy.

[0036] The inventory management strategy and dynamic allocation based on the prediction results are used to combine prediction, execution, monitoring, and feedback to achieve automated and intelligent inventory management and allocation, effectively improving the efficiency and response speed of inventory management; including:

[0037] Generate inventory replenishment suggestions: According to the prediction of future demand by the LSTM model and combined with the real-time inventory levels of each base, automatically generate an inventory replenishment strategy and generate monthly material replenishment suggestions to ensure that the inventory level of each base can meet future production needs and avoid overstocking or material shortages;

[0038] Real-time allocation of the material control end: Through data linkage with the material sensor devices at the production site, monitor the material consumption situation and equipment operation status in real time, and dynamically adjust the material supply and allocation to ensure that the material supply matches the production rhythm;

[0039] Closed-loop feedback and optimization: After implementing the inventory management strategy, through real-time data feedback, the actual consumption is compared with the predicted result, the error is calculated and fed back to the LSTM model. The model continuously optimizes the prediction accuracy through retraining and parameter adjustment, forming a closed-loop feedback mechanism to ensure more accurate future inventory predictions.

[0040] The multi-site inventory optimization system based on deep learning includes: a front-end display layer, an application layer, a data layer, a network layer, and a device layer;

[0041] The front-end display layer is set on the client side and is used to set the model training parameters and visually display the generated prediction results;

[0042] The application layer is set on the cloud server and includes a data acquisition and processing module, an inventory prediction model training module, and a production linkage optimization module. It is used to preprocess the historical data collected by the device layer through the data acquisition and processing module to obtain time series features, optimize and train the inventory prediction model of the inventory prediction model training module in combination with the model training parameters input by the user, and further optimize the model parameters in combination with the production linkage optimization module to obtain an ideal model for outputting inventory prediction results;

[0043] The data layer uses a database to store the multi-source heterogeneous data collected by the device layer;

[0044] The network layer uses multiple communication networks to establish a network connection between the client and the server;

[0045] The device layer includes an inventory management system, production line sensors, and production line equipment set at the industrial production site.

[0046] The application layer includes:

[0047] The data acquisition and processing module collects the historical inventory data of multiple sites for data preprocessing and constructs time series data as the input features for inventory optimization;

[0048] The inventory prediction model training module introduces a parallel multi-thread mechanism to improve the LSTM network model, uses the preprocessed time series data to train the improved LSTM network model, adjusts the model parameters through backpropagation to improve the prediction accuracy, and initially obtains an ideal model for predicting and outputting multi-feature inventory prediction results;

[0049] The production linkage optimization module generates an optimized inventory management strategy based on the prediction results of the LSTM model, and further conducts linkage feedback with the material control end at the production site to achieve intelligent inventory monitoring management and material allocation.

[0050] The present invention has the following beneficial effects and advantages:

[0051] 1. Significantly improve the accuracy of demand forecasting: By using the LSTM model to train the historical inventory data of multiple bases, the present invention can accurately capture complex time series features, significantly improve the accuracy of demand forecasting, effectively reduce the problems of inventory backlog or shortage caused by forecasting errors, optimize the balance of material supply, and ensure the rationality of inventory management at each base. By combining optimization algorithms, based on the predicted demand, the best inventory management strategy is determined to ensure that the inventory levels at each base are within a reasonable range, optimize the overall inventory management strategy, and reduce inventory costs.

[0052] 2. Optimize the inventory management strategy and reduce inventory costs: By combining optimization algorithms, based on the predicted demand, the system can automatically generate the best inventory management strategy for each base, ensure that the inventory level is always within a reasonable range, avoid overstocking or understocking, thereby reducing inventory management costs and improving the efficiency of the overall supply chain.

[0053] 3. Efficiently process a large amount of heterogeneous data and achieve real-time adjustment: Through deep learning technology, the present invention can efficiently process multi-source and multi-modal heterogeneous data from multiple bases, avoiding the problems brought by low processing efficiency and human errors in traditional manual management methods. At the same time, the system can respond in real time to inventory changes and demand fluctuations, automatically adjust the replenishment strategy, and ensure that the production plan is not affected.

[0054] 4. Enhance the flexibility to respond to emergencies: The powerful prediction ability of the LSTM model enables the system to quickly respond to emergencies, such as equipment failures and sudden surges in market demand. The system can quickly adjust the inventory and replenishment strategies when an event occurs, reduce the negative impact of emergencies on production, and ensure production continuity and the stability of enterprise operations.

[0055] 5. An inventory management solution applicable to large-scale and complex environments: The present invention fully considers the complexity of multi-base and multi-process inventory management. By comprehensively analyzing and optimizing the inventory data between different bases, it provides an inventory management solution applicable to large-scale and complex environments, effectively improving the management efficiency of the overall supply chain and ensuring the coordination of production and logistics links.

[0056] 6. Reduce manual intervention and improve the level of intelligence: Through automated deep learning algorithms and optimization processes, the present invention reduces the dependence on manual operations in traditional inventory management, reduces the costs and risks of errors caused by manual intervention, significantly improves the level of intelligence of inventory management, and realizes efficient and accurate automated inventory management. Brief Description of the Drawings

[0057] Figure 1 Flowchart of the optimization method based on long short-term memory network (LSTM) in multi-base inventory implementation of the present invention;

[0058] Figure 2 Iterative training phase display diagram based on the sliding window mechanism;

[0059] Figure 3 Inventory forecast diagram of each material code for the next year;

[0060] Figure 4 Comparison of the true values and predicted values of multiple material characteristics;

[0061] Figure 5 Prediction accuracy of Feature 1 and Feature 2 under some material codes. Specific implementation manner

[0062] To make the above objects, features, and advantages of the present invention more obvious and understandable, the following will describe in detail the specific implementation method of the present invention with reference to the accompanying drawings. Many specific details are set forth in the following description to fully understand the present invention. However, the present invention can be implemented in many other ways different from those described herein, and those skilled in the art can make similar improvements without departing from the connotation of the invention. Therefore, the present invention is not limited by the specific implementations disclosed below.

[0063] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by those of ordinary skill in the technical field to which the present invention belongs. The terms used in the specification of the present invention are only for the purpose of describing specific embodiments and are not intended to limit the present invention. The following will further describe the present invention in detail with reference to the accompanying drawings and embodiments.

[0064] The present invention provides an optimization method based on the long short-term memory network (LSTM) in multi-site inventory. The LSTM model is used to more accurately predict future demand and reduce inventory backlog or shortage problems caused by prediction errors. The technical solution of the present invention will be elaborated in detail below with reference to specific embodiments.

[0065] Step 1: Data collection and preprocessing. Use the historical inventory data of multiple sites as input features to construct time series data;

[0066] Data collection in the present invention is mainly applied to the inventory management of materials in each process of the production line in the industrial production site. Heterogeneous data is collected from multiple data sources to ensure the comprehensiveness and accuracy of the model input data. By collecting multi-source and multi-modal data from different sensors and systems, time series data suitable for the LSTM model is constructed. The specific data sources and processing steps are as follows:

[0067] Data collection:

[0068] Inventory management system (such as ERP / MES system): Records the material inventory levels, inbound and outbound operation records of each site.

[0069] Production line sensor data: including material consumption sensors, equipment status monitoring sensors (such as temperature, pressure, vibration, etc.), which can indirectly reflect material requirements and consumption.

[0070] Logistics system data: such as material transportation, delivery time, supply chain transfer information, etc.

[0071] Market demand data: historical order data or future demand forecast information from the market analysis system, which helps to dynamically adjust the inventory strategy according to market changes.

[0072] These multi-source, multi-modal heterogeneous data include structured and unstructured data, with different formats and dimensions, such as time series data, real-time values collected by sensors, static records of logistics and supply chain, etc.

[0073] Data preprocessing:

[0074] To ensure data consistency and quality, before inputting into the LSTM model, it is necessary to perform unified preprocessing on multi-source heterogeneous data. The preprocessing steps include:

[0075] Data cleaning:

[0076] For the collected data, first perform data cleaning to remove noise and outliers. For example, error records in sensor collection, equipment failure data, or abnormal data input by humans need to be filtered or corrected. For missing data, use interpolation methods or context-based prediction algorithms for reasonable filling to ensure the integrity and continuity of time series data.

[0077] Normalization processing:

[0078] Since the data comes from different sources and has different dimensions and units (such as inventory, temperature, transportation time, etc.), normalization processing is performed on the data (such as using MinMax normalization or Zscore standardization) to map different feature values to a unified range between 0 and 1. This can eliminate the scale differences of different data sources and ensure that the LSTM model will not produce biases due to numerical range differences during the training process. The normalized data can better reflect the relative changes of each variable in the time series, thus improving the training effect of the model.

[0079] Time series construction:

[0080] Align the cleaned and normalized data according to the time dimension to construct data input in time series format. For example, the data at each time point contains feature values in multiple dimensions such as timestamp, inventory, material consumption, equipment status, and logistics situation. These multi-dimensional time series data will be input into the LSTM model for it to capture long-term dependencies and complex time series relationships.

[0081] Feature Selection and Dimensionality Reduction:

[0082] Select the most relevant key features from multi-source heterogeneous data as the input variables of the model to ensure the effectiveness and computational efficiency of the model. The key features include historical inventory, inbound quantity, outbound quantity, operating status of production equipment, changes in market demand, etc. For redundant or unnecessary features, use feature selection algorithms (such as principal component analysis PCA) for dimensionality reduction to reduce the model complexity and improve the training efficiency.

[0083] Innovation and Differences from the Prior Art:

[0084] The preprocessing step of the present invention has significant creativity. The advantages compared with the existing methods include:

[0085] Integration and Processing of Multi-Source Multimodal Heterogeneous Data:

[0086] Different from traditional inventory management methods that only rely on a single data source, the present invention combines multiple data from inventory management systems, sensors, logistics systems, and market data, covering multiple dimensions such as material consumption, supply, and equipment operating status. This comprehensive data integration improves the accuracy of demand forecasting and the comprehensiveness of inventory optimization.

[0087] Deep Application of Normalization Processing: Through the normalization processing of heterogeneous data, the present invention ensures that data from different sources with different dimensions can be effectively processed within a unified framework, reducing the bias caused by differences in data units in traditional methods. The use of normalization improves the accuracy and stability of the LSTM model when processing complex time series data.

[0088] Automated Feature Selection and Time Series Construction: Through feature selection algorithms and the standardized construction of time series, the present invention can automatically extract the features most relevant to inventory management and convert them into multi-dimensional time series suitable for LSTM processing. This processing method enables the model to better capture the dynamic laws of inventory changes and achieve more accurate prediction and optimization.

[0089] Through the above data collection and preprocessing, the present invention effectively improves the quality of data input, laying a foundation for subsequent LSTM model training and inventory management strategy optimization.

[0090] Step 2: Use the preprocessed data to train the LSTM model and improve the prediction accuracy by adjusting the model parameters;

[0091] Design and construct an LSTM model, and determine parameters such as the number of model layers, the number of neurons in each layer, and the activation function. Divide the preprocessed dataset into a training set, a validation set, and a test set to evaluate the model performance. Use the training set to train the LSTM model, and adjust the model parameters through the backpropagation algorithm to minimize the prediction error. Use the validation set to perform hyperparameter tuning on the model, select the optimal learning rate, batch size and other parameters to improve the generalization ability of the model. Use the test set to evaluate the trained model, calculate the prediction error metrics, and ensure the prediction accuracy of the model.

[0092] The specific implementation process is as follows:

[0093] Design and construction of the LSTM model

[0094] Design and construct an LSTM model suitable for multi-base inventory demand forecasting. The design of the model includes determining parameters such as the number of LSTM layers, the number of neurons in each layer, and the activation function. The present invention particularly designs a multi-layer LSTM architecture, and the number of neurons in each layer is optimized according to experimental results and data complexity to ensure that each layer can capture different levels of dependencies in the time series. Commonly used activation functions such as ReLU (Rectified Linear Unit) are adopted, combined with a suitable dropout mechanism to prevent model overfitting.

[0095] Multi-layer LSTM architecture: Adopt a multi-layer stacked LSTM unit structure. The first layer is responsible for processing short-term dependencies, and subsequent layers further extract long-term dependencies.

[0096] Number of neurons: According to the input feature dimension and the complexity of the training data, set the number of neurons in each layer of LSTM, and set 128 neurons in each layer.

[0097] Activation function: To improve the expression ability of non-linear features, use ReLU as the activation function to enable the model to capture more complex time series changes.

[0098] Dataset division and preparation:

[0099] Before model training, divide the preprocessed dataset into a training set, a validation set, and a test set. Usually, it is divided according to the ratio of 70% for the training set, 15% for the validation set, and 15% for the test set. The training set is used to adjust the model weights, the validation set is used for hyperparameter tuning and model performance evaluation, and the test set is used for the final model performance evaluation.

[0100] Training set: Used for the initial training of the LSTM model to adjust the model weights through a large amount of data.

[0101] Validation set: Used to evaluate the model performance during training, helping to adjust hyperparameters (such as learning rate, batch size, etc.) and improve the generalization ability of the model.

[0102] Test set: After the model training is completed, the test set is used for the final evaluation to ensure the prediction accuracy of the model in actual applications.

[0103] Model training and parameter tuning:

[0104] The LSTM model is trained using the training set. During the training process, the backpropagation algorithm is adopted, combined with gradient descent optimization algorithms (such as Adam or RMSprop) to adjust the weight parameters of the model, gradually minimizing the prediction error. The goal of training is to continuously reduce the loss function value (such as mean absolute error MAE) through backpropagation to improve the prediction accuracy of the model.

[0105] Loss function: The mean absolute error MAE is adopted as the objective function to measure the difference between the predicted value and the actual value.

[0106] Optimization algorithm: The Adam optimizer is used, which can adaptively adjust the learning rate, effectively accelerating convergence and avoiding the local optimum problem.

[0107] Hyperparameter tuning and optimization:

[0108] To further improve the generalization ability of the LSTM model, the validation set is used to tune the hyperparameters of the model. During this process, the tuned parameters include learning rate, batch size, number of LSTM layers, and number of neurons, etc. Through repeated experiments and validations, the optimal combination of hyperparameters is selected.

[0109] Learning rate: Different learning rates (such as 0.001, 0.01, 0.1, etc.) are tried to find the best balance between the convergence speed and accuracy.

[0110] Batch size: The batch size (such as 32, 64, 128, etc.) is adjusted to optimize the relationship between training time and accuracy.

[0111] Number of LSTM layers and number of neurons: By adjusting the number of LSTM layers and the number of neurons in each layer, ensure that the model has sufficient expressive ability but does not overfit.

[0112] The goal of hyperparameter tuning is to improve the generalization ability of the model and ensure that it can maintain a high prediction accuracy on unseen data.

[0113] Model Evaluation and Accuracy Verification: After the model training and hyperparameter tuning are completed, the trained LSTM model is evaluated using the test set. By calculating prediction error metrics (such as MSE or MAE), the prediction performance of the model in practical applications is evaluated. Ensure that the model can accurately predict future material requirements in multi-site inventory management.

[0114] Evaluation Metrics: Usually, the Mean Absolute Error (MAE) is used to measure the prediction accuracy of the model.

[0115] Model Performance Evaluation: By comparing the real data in the test set with the prediction results of the model, the prediction effect of material requirements in different bases and processes is evaluated.

[0116] Innovation Point: Parallel Multi-threaded Training

[0117] To further improve the training efficiency and computational performance of the LSTM model, the present invention introduces a parallel multi-threaded training mechanism. When the traditional LSTM model processes large-scale multi-modal data, the sequential training method often leads to low computational efficiency. By introducing a multi-threaded mechanism, the present invention realizes parallel processing of different feature dimensions. Specifically, during training, each thread is responsible for processing different time series features (such as inventory levels, incoming quantities, outgoing quantities, etc.) and performing parallel calculations simultaneously, thus significantly reducing the training time and improving the training speed of the model.

[0118] Introduction of the Parallel Multi-threaded Mechanism: Parallel multi-threading allows different feature data to be passed to the LSTM unit in parallel, thus accelerating the model training process, especially suitable for data scenarios with a large number of time series and high-dimensional features.

[0119] Data Flow Processing: Each thread is responsible for processing data streams of different dimensions. After gradual calculations through the LSTM network layer, the calculation results of each thread are aggregated through the fully connected layer to output the final prediction value.

[0120] The innovation point of parallel multi-threading not only improves the training speed but also ensures that the model can maintain high accuracy when processing multi-dimensional heterogeneous data, significantly enhancing the prediction ability, especially suitable for complex production environments with multiple bases and multiple processes.

[0121] Through this step, the improvements in the LSTM model design, training, parameter optimization, and parallel processing of the present invention significantly improve the accuracy and efficiency of multi-site inventory demand prediction, providing strong technical support for intelligent inventory management in complex industrial scenarios.

[0122] Step 3: Implementation Method of Iterative Training

[0123] In the present invention, the iterative training step is based on a sliding window mechanism, which gradually inputs time series data into the model and continuously improves the prediction accuracy through real-time data update and model retraining, as Figure 2 . The specific implementation steps are as follows:

[0124] Data preprocessing and feature engineering:

[0125] Obtaining of original data: First, multi-source heterogeneous data for model training is extracted from the original data collected from the inventory management systems, sensors, etc. of each base. Depending on the specific application scenario, the data may include historical inventory levels, inbound quantities, outbound quantities, equipment operating status, market demand changes, etc.

[0126] Feature extraction and data splitting: After cleaning and normalizing the original data, it is split according to a ratio to form a training set and a label set. For example, 80% of the data (Lenth = 0.8) is used as the input training feature set, and the remaining 20% of the data (Lenth = 0.1) is used as the label set of the model for the verification and prediction of model training.

[0127] Time series construction: In the feature engineering stage, the data is reorganized into a time series format to ensure that the model can effectively capture the temporal dependencies of the input data.

[0128] Initial model training:

[0129] Model input: The historical data processed by feature engineering is input into the LSTM or XGBoost model, and the model starts to learn from the training set to capture the change rules of inventory levels, inbound quantities, and other features at different time steps.

[0130] Backpropagation and parameter adjustment: The model uses the backpropagation algorithm and an optimizer (such as Adam or RMSprop) to adjust its internal parameters to minimize the loss function (such as the mean absolute error MAE). After several training iterations, the model completes the initial learning of the historical data and can initially generate demand predictions.

[0131] Iterative prediction in the sliding window mechanism:

[0132] Sliding window setting: After the initial model training is completed, the sliding window mechanism is used to predict the data. Specifically, each window covers a period of historical data (such as a length of Lenth = 1.0) to predict the demand for the next time period (a length of Lenth = 0.1). After each prediction, the sliding window moves forward by a certain time period (such as a length of 0.1) and uses the data in the new window for prediction again.

[0133] Prediction Process: After the historical data in the window is input into the model, the LSTM model will generate demand prediction results for a future period (Length = 0.1) based on previous training. This result includes the material requirements for each base or process within the future time period, which is used for subsequent inventory management and allocation decisions.

[0134] Real-time Data Update and Model Iterative Optimization:

[0135] Real-time Data Input: As production and inventory continue, new data will continuously enter the system. These new data points will be added to the time series to become new historical data for the next prediction. After the window slides, the new data points replace the earliest data points to ensure that the prediction model is always based on the latest production and inventory information.

[0136] Error Feedback and Model Retraining: Each time the prediction result generated by the model is compared with the actual inventory demand to calculate the error (such as the Mean Absolute Error MAE). Based on the error feedback, the model further optimizes its internal parameters through iterative retraining. Through repeated iterative processes, the model can adapt to data changes and improve the prediction accuracy.

[0137] Online Learning and Parameter Update: The system also has the ability of online learning, that is, when new data arrives, the model can quickly adjust its own parameters to ensure that as time goes by and the environment changes, the prediction ability of the model does not degrade. Through continuous online learning and parameter optimization, the stability of the model during long-term operation is ensured.

[0138] Combined Long-term and Short-term Prediction Ability:

[0139] Short-term Prediction: The sliding window mechanism ensures that the model can perform short-term predictions, which usually involve material requirements in the next few days or weeks. This kind of prediction helps enterprises adjust production and material supply in a timely manner to avoid inventory backlog or shortage.

[0140] Long-term Prediction: By extending the time length of the sliding window, the model can also perform predictions for a longer time period (such as material requirements in the next few months). Long-term prediction provides data support for the strategic planning of enterprises, such as formulating procurement plans and optimizing production scheduling.

[0141] Verification and Application of Prediction Results:

[0142] Prediction Result Verification: The demand prediction results generated by the model need to be verified by actual data. The error between the predicted value and the actual value will be used as the feedback for the next round of iterative training to guide the further optimization of the model.

[0143] Inventory Strategy Adjustment: Based on the forecasting results, the system can automatically provide inventory replenishment suggestions for each base or process, ensuring that the inventory level is within a reasonable range and avoiding overstocking or understocking of materials. Meanwhile, enterprises can rationally allocate production resources according to the demand forecasting results to ensure production continuity and cost minimization.

[0144] Key Technical Points: Sliding Window and Iterative Training

[0145] The innovation of the present invention lies in the adoption of a sliding window mechanism combined with iterative training. Through this mechanism, the model can adapt to new data inputs in real time, continuously update the forecasting results, and optimize the parameters. This mechanism is particularly suitable for complex production environments with multiple bases and processes, ensuring that the model maintains high-efficiency and accurate forecasting capabilities when dealing with dynamic demands.

[0146] Sliding Window Mechanism: Through the sliding window, the model can gradually utilize the latest historical data for forecasting, ensuring that the forecasting results are based on the latest production and inventory changes.

[0147] Iterative Training: The model continuously retrains through the continuous update of real-time data, ensuring that the parameters always remain in the optimal state. Through this iterative optimization, the model has high robustness and adaptability when processing large-scale heterogeneous data.

[0148] Through the above iterative training process, the present invention can provide accurate inventory demand forecasting for enterprises, improve the intelligent level of production planning and inventory management, reduce inventory costs, and improve the efficiency of supply chain management.

[0149] Step 4: Inventory Management Strategy and Dynamic Allocation Based on Forecasting Results

[0150] Based on the predicted demand results, corresponding monthly inventory suggestions are given to determine the optimal inventory management strategy; the present invention proposes the application of an ideal model, which is not only the inventory demand forecasting for new input data based on the LSTM model, but also further extended to the material control ends of each process on the production site assembly line, realizing the coordinated allocation of material sensor devices, ensuring that the forecasting results can be real-time feedback and used for the optimization management of inventory and production processes. Through this extension, the system can automatically discover problems, propose solutions, and continuously adjust through the feedback mechanism to form a closed-loop management, ensuring the dynamic balance of inventory and production.

[0151] Generation of Inventory Suggestions and Strategies Based on Forecasts

[0152] The present invention forecasts the inventory demand for a future period (such as one month) through the LSTM model, and the system generates corresponding inventory management strategies for each base and each process according to the forecasting results. This strategy includes the following contents:

[0153] Monthly Inventory Suggestion: Based on the predicted demand, combined with the current inventory level, replenishment cycle, production plan, etc., the system provides monthly inventory replenishment suggestions for each base to ensure that materials can be in place in a timely manner without shortages or surpluses.

[0154] Inventory Replenishment Strategy: Through continuous monitoring of inventory and demand data, the system dynamically adjusts the replenishment plan. According to demand fluctuations, the system can increase or decrease the replenishment frequency of materials to avoid problems such as material accumulation or production stoppages.

[0155] Coordination of the Material Control End and Sensor Deployment

[0156] To ensure the effective implementation of the inventory suggestions predicted by the model, the system has a deep linkage with the material control end at the production site and coordinates and deploys various types of sensor devices to ensure the precise matching of material management and production rhythm. The specific steps are as follows:

[0157] Material Sensor Monitoring: The system integrates various types of sensor devices, such as weight sensors, inventory location sensors, RFID sensors, etc., to monitor the real-time inventory levels of each process. When the inventory is below the threshold or the consumption rate accelerates, the system will automatically schedule replenishment according to the prediction results to ensure the smooth progress of production.

[0158] Equipment Status Sensors: Such as temperature sensors, vibration sensors, etc., monitor the operating status of production equipment in real time. For example, if a certain piece of equipment is operating abnormally, the system can automatically reduce or suspend the material supply for the corresponding process and allocate the excess materials to other processes to prevent production stoppages.

[0159] Multi-Process Coordination: When multiple processes share a certain type of material, the system can dynamically adjust the material scheduling priority according to the real-time material consumption of each process to ensure the efficient coordination of material supply for each process. For example, when the consumption rate of a certain process exceeds expectations, the system can reduce the material supply for low-priority processes and give priority to ensuring the production needs of key processes.

[0160] Dynamic Inventory Management and Real-Time Deployment

[0161] During the actual operation process, the system can dynamically adjust the inventory management strategy according to the latest data by continuously monitoring inventory and demand data and combining the prediction results of the LSTM model, realizing closed-loop inventory management and real-time deployment. The specific steps are as follows:

[0162] Real-Time Monitoring and Data Collection: The system collects real-time material consumption data, equipment status information, and market demand changes of each process on the production line through sensor devices. Combining these real-time data, the prediction results of the LSTM model are continuously compared with the actual data to adjust the replenishment plan in real time.

[0163] Dynamic Allocation and Feedback Mechanism: When it is found that the material consumption of a certain process exceeds the expectation or the inventory is insufficient, the system will automatically trigger the allocation mechanism to adjust the replenishment plan in real time. For example, if the inventory demand of Process A suddenly increases, the system will allocate materials from Process B or other bases in advance according to the prediction results and actual data to ensure the uninterrupted production.

[0164] Automatic Feedback and Re-optimization: After the material supply strategy is executed, the system verifies the prediction accuracy according to the actual situation. If there is a large deviation between the actual demand and the prediction result, the system will re-optimize based on the new data to further improve the prediction accuracy of the LSTM model, forming an automatic feedback mechanism to continuously optimize the prediction and inventory allocation strategies.

[0165] Closed-loop Feedback Mechanism and Adaptive Adjustment

[0166] Through the collection and real-time feedback of sensor data on the production site, the system realizes closed-loop management to ensure the dynamic optimization of inventory management. The specific closed-loop feedback mechanism is as follows:

[0167] Error Analysis and Feedback Adjustment: The system regularly compares the predicted data with the actual demand, and calculates the prediction error (such as the Mean Absolute Error MAE). When the error exceeds the preset threshold, the system will automatically feedback to the LSTM model to trigger the retraining of the model to adapt to the new production and inventory changes.

[0168] Adaptive Adjustment: Through continuous learning and optimization, the system can adjust the model parameters according to the real-time data to achieve adaptive prediction and allocation capabilities. As the amount of data increases, the prediction accuracy of the model will be further improved to ensure more accurate and reliable future demand prediction.

[0169] Case Study: Inventory Prediction and Error Verification

[0170] To better illustrate the effectiveness of the present invention, a specific case study is provided as follows:

[0171] Original Data: Collect historical inventory levels, material inbound and outbound quantities, equipment status data of production lines, etc. of a certain base.

[0172] Prediction Results: The LSTM model predicts the material demand for the next month based on the input historical data.

[0173] Error Verification: Compare the model prediction results with the actual material demand, and calculate the prediction error (such as the Mean Absolute Error MAE). If the error is large, the system automatically feedbacks to the model for retraining to improve the accuracy of the next round of prediction.

[0174] Replenishment Adjustment: Based on the predicted demand results, the system generates a monthly inventory replenishment plan and dynamically adjusts the replenishment strategy in combination with real-time monitoring data to ensure sufficient inventory without waste.

[0175] Innovation Points: Intelligent Material Control and Closed-loop Management

[0176] Through deep linkage with on-site material sensors and control devices in production, the present invention realizes intelligent material control and a closed-loop feedback mechanism, which can not only predict future material requirements but also conduct real-time allocation according to the actual production and inventory situations. This innovation point ensures the accuracy and timeliness of material supply, avoiding the lag and over-reliance on manual operations in traditional inventory management.

[0177] Intelligent Control: The system can automatically regulate the material supply for each process, reducing manual intervention and improving management efficiency.

[0178] Dynamic Allocation: Through real-time linkage with sensors, the system can flexibly respond to sudden changes in material requirements.

[0179] Closed-loop Optimization: The feedback mechanism ensures that the system can continuously optimize predictions and strategies according to actual situations, forming an adaptive closed-loop management.

[0180] The ideal model application of the present invention realizes intelligent and automated inventory management and material allocation through deep linkage with the production site. The system can not only accurately predict future demands based on the LSTM model but also dynamically adjust the replenishment strategy through real-time monitoring of sensors and devices, forming a closed-loop feedback and re-optimization mechanism. This intelligent inventory management system effectively improves production efficiency, reduces inventory costs, and ensures the stability and flexibility of the supply chain.

[0181] Based on the predicted demand results, corresponding monthly inventory suggestions are given to determine the optimal inventory management strategy;

[0182] Taking the demand results predicted by the LSTM model as input, combining with the inventory levels and inventory replenishment strategies of each base. During actual operation, inventory and demand data are monitored in real time, and the inventory management strategy is dynamically adjusted according to the latest data.

[0183] The present invention also provides a multi-base inventory optimization system for deep learning, including: a front-end display layer, an application layer, a data layer, a network layer, and a device layer;

[0184] The front-end display layer is set on the client side and is used to set model training parameters and visually display the generated prediction results;

[0185] The application layer is set on the cloud server and includes a data collection and processing module, an inventory prediction model training module, and a production linkage optimization module. It is used to preprocess the historical data collected by the device layer through the data collection and processing module to obtain time series features, optimize and train the inventory prediction model of the inventory prediction model training module in combination with the model training parameters input by the user, and further optimize the model parameters in combination with the production linkage optimization module to obtain an ideal model for outputting inventory prediction results;

[0186] The data layer stores multi-source heterogeneous data collected by the device layer using a database;

[0187] The network layer establishes a network connection between the client and the server using multiple communication networks;

[0188] The device layer includes an inventory management system, production line sensors, and production line equipment set at the industrial production site.

[0189] The application layer includes:

[0190] The data collection and processing module collects historical inventory data from multiple bases for data preprocessing and constructs time series data as input features for inventory optimization;

[0191] The inventory prediction model training module introduces a parallel multi-thread mechanism to improve the LSTM network model, trains the improved LSTM network model using the preprocessed time series data, adjusts the model parameters through backpropagation to improve the prediction accuracy, and initially obtains an ideal model for predicting and outputting multi-feature inventory prediction results;

[0192] The production linkage optimization module generates an optimized inventory management strategy based on the prediction results of the LSTM model, and further conducts linkage feedback with the material control end at the production site to achieve intelligent inventory monitoring management and material allocation.

[0193] To better illustrate the effectiveness of the present invention, a specific numerical example is provided below to show the original data, prediction results, and error analysis.

[0194] The original data contains the following fields: material_code, raw_material_code, material_name, material_type, material_model, material_pice, material_inventory_amount, material_buying_amount, material_delivery_amount, material_time.

[0195] Prediction Results: The preprocessed data is used to train the LSTM model, and the prediction accuracy is improved by adjusting the model parameters. According to the prediction results, inventory prediction charts for each material code in the next year are drawn. Taking Figure 3 as an example, the figure shows the inventory prediction values for the material code 20240509181620_5F1125AFAF88E23F in the next year. Based on these prediction results, inventory optimization can be further carried out.

[0196] To evaluate the accuracy of the prediction results, the mean absolute error (MAE) and mean absolute error (MAE) are used for error analysis. Finally, the prediction accuracy of some material codes can reach 90%.

[0197] The model will give inventory suggestions for March, June, and December according to the prediction results. The following is an example:

[0198] Suggested inventory for the 3rd month: 0. The material in this inventory needs to replenish the inventory by 0.17255593836307526

[0199] Suggested inventory for the 6th month: 1. The material in this inventory needs to replenish the inventory by 1.046560876071453

[0200] Suggested inventory for the 12th month: 0. The material in this inventory needs to replenish the inventory by 0.0035219741985201836

[0201] Step 5: Model Evaluation;

[0202] In the model evaluation step, the present invention comprehensively evaluates by comparing each error index of this algorithm with other existing algorithms to ensure the accuracy of the LSTM model in inventory prediction and the reliability in practical applications. We use the mean absolute error (MAE) and the percentage of errors within a specific range as the main indicators for model evaluation. The detailed calculation process is as follows:

[0203] Calculation Process of Precision Index:

[0204] For each feature, first calculate the absolute error between the predicted value and the true value. The formula is:

[0205] |Predicted Value - True Value|

[0206] Calculate the mean absolute error (MAE) of all data points. The formula is:

[0207]

[0208] This index reflects the overall average error level of the model.

[0209] Next, calculate the number of data points whose error is within the set threshold N (e.g., N = 5 units), that is, the number of data points satisfying: |predicted value - true value| ≤ N. Divide this number by the total number of data points and multiply by 100% to obtain a percentage representing the prediction accuracy of the model within the allowable error range.

[0210]

[0211] Flexibly adjust the error threshold: The threshold of the absolute error can be adjusted according to the accuracy requirements of different application scenarios, for example, adjusted from 5 units to 3 units or 10 units. By adjusting different thresholds, the performance of the model in actual applications can be evaluated more flexibly.

[0212] Compare the evaluation results of algorithms

[0213] In the evaluation, we compare the LSTM model with other classical algorithms (such as XGBoost, traditional ARIMA model) to demonstrate the performance advantages of the model in multi-base inventory forecasting. The following Table 1 lists the comparison of MAE and the percentage of errors within a specific range for different algorithms:

[0214] Table 1

[0215]

[0216] It can be seen from the table that the LSTM model has the lowest MAE and performs the best among all models, with an average absolute error of 3.24. At the same time, the LSTM model reaches 87.5% in the percentage of errors ≤ 5, which is significantly higher than other models, indicating that its prediction results are more accurate within a smaller error range. When the error threshold is set to 10, the prediction accuracy of the LSTM model further increases to 95.3%, demonstrating good generalization ability and robustness.

[0217] Prediction results and evaluation

[0218] Analyze the prediction performance of the LSTM model under different material codes and characteristics through graphs and tables, and

[0219] Figure 4 Show:

[0220] The blue line in the figure represents the true value of each material, and the red line is the predicted value of the LSTM model.

[0221] It can be seen from the figure that although there are significant differences between the predicted values and the true values under some material codes (such as 99B8E67C4C5CE33F), in most cases, the predictions of the LSTM model can well follow the changes in actual demand, especially in the scenario of material demand with less fluctuation.

[0222] From Figure 5 it can be seen that:

[0223] The prediction performances of Feature 1 and Feature 2 of materials such as 327D3FCE98BEBF3F, FCC82B107EE4953F, and 4303238AA181D13F are very good, with the MAE all lower than 1 and the prediction accuracy reaching 100%. This shows that the LSTM model can predict the demand very accurately under certain features.

[0224] However, for material 99B8E67C4C5CE33F, the prediction error is relatively large, with the MAE being 6.40 and the accuracy only being 33.33%. Such a large error may be due to the large fluctuations in the feature data, and the model fails to fully capture its complexity.

[0225] Summary and Optimization Directions

[0226] From the above evaluation, it can be obtained that the LSTM model performs excellently in multi-site inventory management. Especially in scenarios with low prediction errors, it can achieve a high prediction accuracy. However, for some features with large errors, the overall prediction performance can be improved by further optimizing the model structure, enhancing the feature extraction ability, or combining other algorithms (such as hybrid models).

[0227] Future optimization directions can include:

[0228] Deeper feature extraction: Automatically extract more valuable features through deep learning to further improve the learning ability of the model.

[0229] Model integration and hybridization: Combine models such as LSTM and XGBoost to form a hybrid model, make full use of the advantages of both, and further reduce the prediction error in complex scenarios.

[0230] The above are only the preferred embodiments of the present invention and do not impose any limitations on the present invention. Any simple modifications, changes, and equivalent structural changes made to the above embodiments according to the technical essence of the present invention still fall within the protection scope of the technical solution of the present invention.

Claims

1. A multi-base inventory optimization method based on deep learning, characterized in that: The following steps are involved: Collect historical inventory data from multiple bases for data preprocessing and construct time series data as input features for inventory optimization; The parallel multi-threading mechanism is introduced to improve the LSTM network model. The improved LSTM network model is trained using the preprocessed time series data. The model parameters are adjusted through back propagation to improve the prediction accuracy. The ideal model is initially obtained to predict and output multi-feature inventory prediction results. Based on the prediction results of the LSTM model, an optimized inventory management strategy is generated, and further linked feedback is carried out with the material control end at the production site to achieve intelligent inventory monitoring management and material allocation.

2. The multi-base inventory optimization method based on deep learning according to claim 1, characterized in that: The multi-base historical inventory data are multi-source, multi-modal, heterogeneous data acquired from a variety of sensors and production control systems in various processes of industrial production lines, including multiple dimensions of production and logistics.

3. The multi-base inventory optimization method based on deep learning according to claim 2, characterized in that: The multi-source, multi-modal, heterogeneous data include: inventory management system data representing inventory levels, inbound and outbound records, and replenishment status; production line sensor data representing material consumption rate, production speed, and equipment status; logistics system data representing material transportation time and inventory transfer records; and market demand data representing historical order data and market demand forecasts.

4. The multi-base inventory optimization method based on deep learning according to claim 1, characterized in that: The pre-processing is as follows: Data cleaning: clean up noisy data and fill in missing values ​​to ensure data integrity; Normalization: normalize or standardize data from different sources to unify the scale to eliminate dimensional differences; Time series construction: align multimodal data from different sources in time to construct unified time series data to capture the temporal dependencies between variables; Feature selection and dimensionality reduction: The most influential features for inventory management are extracted through feature selection algorithms to reduce data dimensions and improve model training efficiency.

5. The multi-base inventory optimization method based on deep learning according to claim 1, characterized in that: The parallel multi-threading mechanism is introduced to improve the LSTM network model, which is to set the number of multi-threads according to the dimension of the input feature, so that different features or data dimensions can be calculated in parallel; each thread is used to process one feature, and the calculation tasks are executed simultaneously in a multi-threaded environment, thereby improving the calculation efficiency of the model.

6. The multi-base inventory optimization method based on deep learning according to claim 1 or 5, characterized in that: The improved LSTM network model adopts a multi-layer structure, with multiple parallel LSTM layers and a fully connected layer thereafter; The multi-layer LSTM structure is used to process complex time series relationships and accurately capture the material demand trends of different processes in multi-base and multi-process scenarios. Several LSTM units are stacked in each parallel single thread. Each LSTM unit interacts with its corresponding time step data to capture the short-term and long-term dependencies of the data and pass it to the LSTM unit of the next time step. The fully connected layer is used to summarize and analyze the data after multi-threaded processing, perform deep learning and weight assignment on the relationships between these features, and generate the final demand forecast output.

7. The multi-base inventory optimization method based on deep learning according to claim 1, characterized in that: The improved LSTM network model is trained to adjust the model parameters through back propagation to improve the prediction accuracy, including: Divide the time series data into training set, validation set, and test set in proportion; Set the sliding window length and model training parameters for time series prediction; through iterative updates of the sliding window, the model is trained or updated based on new historical data to gradually predict future data needs; After each sliding window update, the model continuously reduces the error through continuous adjustment and iterative optimization; After comparing the predicted value in the new window with the true value, the weights are updated and retrained based on the feedback to improve the prediction accuracy.

8. The multi-base inventory optimization method based on deep learning according to claim 1, characterized in that: The inventory management strategy and dynamic allocation based on the forecast results are used to combine forecasting, execution, monitoring and feedback to achieve automated and intelligent inventory management and allocation, effectively improving the efficiency and response speed of inventory management; including: Generate inventory replenishment suggestions: Based on the LSTM model's prediction of future demand and the real-time inventory level of each base, automatically generate inventory replenishment strategies and monthly material replenishment suggestions to ensure that the inventory level of each base can meet future production needs and avoid excessive inventory or material shortages; Real-time allocation of material control: through data linkage with material sensor equipment at the production site, real-time monitoring of material consumption and equipment operation status, dynamic adjustment of material supply and allocation, to ensure that material supply matches the production rhythm; Closed-loop feedback and optimization: After the inventory management strategy is executed, the actual consumption is compared with the predicted results through real-time data feedback, and the error is calculated and fed back to the LSTM model. The model continuously optimizes the prediction accuracy through retraining and parameter adjustment, forming a closed-loop feedback mechanism to ensure more accurate inventory forecasts in the future.

9. A multi-base inventory optimization system based on deep learning, characterized by: include: Front-end display layer, application layer, data layer, network layer, and device layer; The front-end display layer is set up on the client side and is used to set model training parameters and visualize the generated prediction results; The application layer is set up on the cloud server and includes a data collection and processing module, an inventory forecasting model training module, and a production linkage optimization module. The application layer is used to pre-process the historical data collected by the equipment layer through the data collection and processing module to obtain time series features, optimize the inventory forecasting model of the inventory forecasting model training module in combination with the model training parameters input by the user, and optimize the model parameters in combination with the production linkage optimization module to obtain an ideal model for outputting inventory forecasting results; The data layer uses multi-source heterogeneous data collected by the database storage device layer; The network layer uses a variety of communication networks to establish network connections between clients and servers; The equipment layer includes inventory management systems, production line sensors, and production line equipment installed at industrial production sites.

10. The deep learning multi-base inventory optimization system according to claim 9, characterized in that: The application layer includes: The data collection and processing module collects historical inventory data from multiple bases for data preprocessing and constructs time series data as input features for inventory optimization; The inventory forecasting model training module introduces a parallel multi-threading mechanism to improve the LSTM network model. The improved LSTM network model is trained using preprocessed time series data. The model parameters are adjusted through back propagation to improve the forecasting accuracy. The ideal model is initially obtained to predict and output multi-feature inventory forecasting results. The production linkage optimization module generates an optimized inventory management strategy based on the prediction results of the LSTM model, and further links feedback with the material control end of the production site to achieve intelligent inventory monitoring management and material allocation.

Images