Industrial chain multi-modal data intelligent analysis system

Abstract

The invention discloses an industrial chain multi-modal data intelligent analysis system, which fuses an improved multi-layer perceptron and a Q-learning model, performs efficient collection on multi-modal data, and improves the multi-layer perceptron architecture, increases a residual error connection structure and adjusts activation function parameters, so that the multi-modal data analysis efficiency is improved. According to the method, accurate feature extraction of multi-modal data and fusion based on an attention mechanism are realized, an industrial chain topological graph is constructed in combination with parameters such as industrial chain productivity and logistics transportation time, multi-modal fusion feature vectors are mapped to nodes in the graph, association between the data and the industrial chain parameters is established, a Q-learning algorithm is applied, and long-term accumulated income maximization is taken as a target. A dynamic strategy set is generated based on an industrial chain topological graph and a fusion feature vector, feedback data is executed through a strategy, optimization iteration is carried out on improved multi-layer perceptron network parameters and a Q value table of a Q-learning algorithm, intelligent analysis processing of industrial chain multi-modal data is achieved, and intelligent upgrading of the industrial chain is promoted.

Description

Technical Field

[0001] The present invention relates to the field of industrial chain data analysis, and in particular to an industrial chain multimodal data intelligent analysis system. Background Art

[0002] In today's digital and globalized economy, the scale and complexity of industrial chains are constantly increasing. The intertwined nature of various industry sectors, encompassing a wide range of businesses, from raw material supply and product production to logistics and transportation and marketing, generates massive amounts of data in diverse formats, encompassing text, images, numerical values, and other modalities. Accurately and efficiently analyzing this multimodal data is crucial for companies across the industry chain to accurately grasp market trends, optimize production processes, rationally allocate resources, and enhance overall competitiveness.

[0003] However, existing industry chain data analysis technologies have significant flaws. For one thing, traditional data processing methods lack effective integration methods for multimodal data. Data from different modalities is often processed in isolation, making it difficult to uncover deep connections between them. For example, text-based market research reports and numerical sales data cannot be organically combined, making it impossible to comprehensively analyze the inherent connections between market demand and sales performance, making it difficult for companies to obtain comprehensive and accurate information support for decision-making.

[0004] On the other hand, existing technologies have limitations in strategy formulation. Most analytical systems are based on fixed algorithmic models and are unable to adjust strategies in real time based on dynamic changes in the supply chain. When key parameters within the supply chain, such as logistics transportation times and raw material prices, fluctuate, adaptive response strategies cannot be generated in a timely manner. This makes it difficult for companies to make flexible decisions in complex and volatile market environments, leading to missed development opportunities and even operational difficulties. Summary of the Invention

[0005] In order to overcome the shortcomings and deficiencies of the existing technology, the present invention provides an industrial chain multimodal data intelligent analysis system.

[0006] The technical solution adopted by the present invention is an industrial chain multimodal data intelligent analysis system, which includes:

[0007] Multi-source heterogeneous industry chain data acquisition adaptation module: This module is used to establish communication connections with multimodal data acquisition terminals distributed in different links and entities in the industry chain. Based on the preset data interface protocol, it collects multimodal data and converts the collected data into a unified data structure format based on the data type and source.

[0008] Improved multi-layer perceptron feature extraction and fusion module: This module is used to receive the converted multimodal data, extract features from different modal data separately through the improved multi-layer perceptron architecture, and fuse the extracted multimodal features to form a multimodal fusion feature vector;

[0009] Industrial chain parameter association modeling module: This module is used to construct an industrial chain topology map based on the production processes and product structure parameters of different links in the industrial chain. The map contains nodes representing different links in the industrial chain and edges representing the relationships between links. The module maps the multimodal fusion feature vectors to the corresponding nodes of the industrial chain topology map and establishes the association between multimodal data and industrial chain parameters.

[0010] Q-learning strategy generation module: Based on the industry chain topology map and the mapped multimodal fusion feature vector, this module uses the Q-learning algorithm to generate a strategy set for data processing and analysis at different links in the industry chain, with the goal of maximizing long-term cumulative benefits;

[0011] Strategy execution and feedback module: This module is used to receive the generated strategy set, execute the corresponding strategy during the industry chain data processing and analysis, and collect feedback data after the strategy execution. The feedback data includes the data changes after the strategy execution and the impact on the node status of the industry chain topology map;

[0012] System optimization iteration module: This module is used to adjust the network parameters of the improved multi-layer perceptron according to the feedback data, and update the Q value table of the Q-learning algorithm to complete the optimization iteration of the system.

[0013] Furthermore, the improved multi-layer perceptron feature extraction and fusion module, the feature extraction process of the improved multi-layer perceptron is implemented by the following formula:

[0014] H l =σ l (W l H l-1 +b l +R l-1 )

[0015] Among them, H l represents the hidden layer output of the lth layer, l = 1, 2, ..., L, L is the total number of layers of the improved multilayer perceptron; σ l represents the activation function of the lth layer, which is the adjusted LeakyReLU function, and its parameters are set according to the distribution characteristics of the industrial chain data; W l is the weight matrix of the first layer, and its dimension is determined according to the characteristic dimension of data in different links of the industrial chain; b l is the bias vector of the lth layer; R l-1is the residual connection of the l-1 layer, which is used to alleviate the gradient disappearance problem. Its calculation formula is R l-1 =H l-1 .

[0016] Furthermore, the improved multi-layer perceptron feature extraction and fusion module is implemented by the following formula based on the fusion strategy of the attention mechanism:

[0017]

[0018] Among them, A i represents the attention weight of the i-th modality feature, i = 1, 2, ..., N, N is the total number of modalities; e i is the calculated attention score of the i-th modal feature, which is calculated based on the importance parameters of different links in the industrial chain; V i represents the feature vector of the i-th modality; F is the multimodal fusion feature vector after fusion.

[0019] Furthermore, in the industrial chain parameter association modeling module, the construction process of the industrial chain topology map is combined with the production capacity parameters and logistics transportation time parameters in the industrial chain, and the weight of the edge between nodes is calculated by the following formula:

[0020] w ij =α×C ij +β×T ij

[0021] Among them, w ij represents the weight of the edge between node i and node j; C ij is the capacity matching parameter between node i and node j, and its value is determined according to the capacity ratio relationship of the links represented by the two nodes; T ij is the logistics transportation time parameter between node i and node j; α and β are weight coefficients, which are set according to the actual operation of the industrial chain, and α + β = 1.

[0022] Furthermore, the Q-learning strategy generation module and the Q-learning algorithm's Q-value update formula are as follows:

[0023]

[0024] Among them, Q(s t , a t ) means in state s t Next, perform action a t Q value, s t is the node state combination of the industry chain topology at time t, a tis the operation performed on the node data at time t; γ is the discount factor, which is used to balance current rewards and future rewards. Its value is set according to the dynamic change speed of the industrial chain; In state s t+1 The maximum Q value of all possible actions.

[0025] Furthermore, the strategy execution and feedback module also collects feedback data in combination with inventory parameters in the industry chain. The changes in inventory data after strategy execution are quantified using the following methods:

[0026]

[0027] Among them, ΔI k It represents the inventory change of the kth link before and after the implementation of the strategy; is the inventory quantity at the kth stage after executing the strategy; is the inventory quantity at the kth stage before executing the strategy.

[0028] Furthermore, the system optimization iteration module adopts an adaptive learning rate algorithm to adjust the network parameters of the improved multi-layer perceptron. The adjustment formula of the learning rate η is:

[0029]

[0030] Among them, η t represents the learning rate at time t; η0 is the initial learning rate, which is set according to the complexity of the industry chain data; λ is the learning rate decay parameter; and t is the number of iterations.

[0031] Furthermore, the Q-learning strategy generation module, when generating a strategy set, combines the cost parameters in the industrial chain and incorporates the cost parameters into the action selection process. The probability formula for action selection is:

[0032]

[0033] Among them, P(a t ∣s t ) means in state s t Next select action a t ; τ is the temperature parameter, which is used to control the randomness of action selection, and its value is set according to the cost fluctuation of the industrial chain.

[0034] Furthermore, when extracting features from the industrial chain data in text format, the improved multi-layer perceptron feature extraction and fusion module combines the keyword frequency parameters in the text and performs weighted processing in the following manner:

[0035]

[0036] Among them, H text is the processed text feature vector; M is the number of keywords in the text; f k is the frequency parameter of the kth keyword; h k is the feature vector corresponding to the kth keyword.

[0037] The multimodal data intelligent analysis system of the industrial chain includes the following steps:

[0038] The first step is to establish communication connections with data collection terminals at different links of the industrial chain through the multi-source heterogeneous industrial chain data collection adaptation module according to the preset data interface protocol, collect multimodal data, and convert the collected data into a unified data structure format;

[0039] In the second step, the converted multimodal data is transmitted to the improved multi-layer perceptron feature extraction and fusion module. The improved multi-layer perceptron architecture is used to extract features from different modal data respectively, and a fusion strategy based on the attention mechanism is adopted to fuse the extracted multimodal features into a multimodal fusion feature vector.

[0040] In the third step, the industrial chain parameter association modeling module constructs an industrial chain topology map based on the parameters of different links in the industrial chain, and maps the multimodal fusion feature vectors to the corresponding nodes in the topology map to establish the association between data and industrial chain parameters;

[0041] In the fourth step, the Q-learning strategy generation module uses the Q-learning algorithm based on the industry chain topology map and multimodal fusion feature vectors to generate a strategy set for data processing and analysis at different links in the industry chain with the goal of maximizing long-term cumulative benefits.

[0042] In the fifth step, the strategy execution and feedback module receives the strategy set and executes it during the industry chain data processing and analysis process, while also collecting feedback data after the strategy is executed;

[0043] In the sixth step, the system optimization iteration module adjusts the network parameters of the improved multi-layer perceptron based on the feedback data, updates the Q value table of the Q-learning algorithm, and performs system optimization iteration.

[0044] Beneficial effects: The present invention proposes an intelligent analysis system for multimodal data of the industrial chain. The system deeply integrates the improved multi-layer perceptron and Q-learning model to carry out all-round processing and analysis of the multimodal data of the industrial chain, significantly improving the operating efficiency and competitiveness of the industrial chain. The system can efficiently collect multi-source heterogeneous data such as contract terms, product appearance images, and production process parameters, and use the improved multi-layer perceptron to accurately extract and fuse features of different modal data, and deeply explore the value of data. Combined with key parameters such as industrial chain production capacity, logistics transportation time, inventory, and cost, an industrial chain topology map is constructed to intuitively present the relationship between each link. At the same time, based on this, the Q-learning algorithm is used to generate a strategy set with the goal of maximizing long-term cumulative benefits, providing a scientific decision-making basis for industrial chain data processing and analysis. During the operation of the system, by collecting feedback data after the execution of the strategy, the improved multi-layer perceptron network parameters and the Q value table of the Q-learning algorithm are dynamically adjusted to continuously optimize system performance so that it can better adapt to the dynamic changes of the industrial chain. The system realizes the intelligent analysis and processing of multimodal data of the industrial chain, helping enterprises to accurately grasp the operating status of the industrial chain, improve resource allocation efficiency, reduce operating costs, and promote the development of the industrial chain towards intelligence and efficiency, thereby enhancing the overall competitiveness and economic benefits of the industrial chain. BRIEF DESCRIPTION OF THE DRAWINGS

[0045] Figure 1 It is a diagram of the system module composition of the present invention;

[0046] Figure 2 This is a flow chart of the system operation of the present invention. DETAILED DESCRIPTION

[0047] It should be noted that, unless there is a conflict, the embodiments in this application and the features described in the embodiments can be combined with each other. The application is further described in detail below with reference to the accompanying drawings and specific embodiments.

[0048] like Figure 1 As shown, the industry chain multimodal data intelligent analysis system includes:

[0049] Multi-source heterogeneous industry chain data acquisition adaptation module: This module is used to establish communication connections with multimodal data acquisition terminals distributed in different links and entities in the industry chain. Based on the preset data interface protocol, it collects multimodal data such as contract terms data in text format, product appearance data in image format, and production process parameter data in numerical format. It then converts the collected data into a unified data structure format based on the data type and source.

[0050] Specifically, the multi-source heterogeneous industrial chain data acquisition and adaptation module is the fundamental entry point for the entire system to acquire data. Its core function is to enable the collection and format conversion of the widely distributed multimodal data within the industrial chain. Within the industrial chain environment, data comes from a wide variety of sources. For example, in the automotive manufacturing industry, raw material ingredient description documents provided by upstream raw material suppliers constitute text data; during the production process, quality inspection images of product parts constitute image data; and real-time monitoring values such as temperature and pressure generated during the operation of production equipment constitute numerical data.

[0051] This module establishes a communication connection with the data acquisition terminals in various links of the industrial chain through a preset data interface protocol. Common data interface protocols such as HTTP and TCP / IP can be selected according to the characteristics of different data acquisition terminals. Taking the sensor equipment in the smart factory as an example, it usually uses the Modbus protocol for data transmission. This module can adapt to the Modbus protocol, exchange data with the sensor equipment, and obtain various parameter data of the equipment operation. In terms of data format conversion, for text data, different encoding formats (such as UTF-8, GBK) will be uniformly converted into a standard encoding format that the system can recognize; for image data, different formats such as BMP and JPEG will be uniformly adjusted to a format that is convenient for subsequent processing, and parameters such as image resolution and color mode will be standardized; for numerical data, parameters such as data precision and units will be unified.

[0052] The significance of this module lies in breaking down barriers to data collection within the industry chain, enabling standardized collection of multimodal data and providing an accurate and unified data foundation for subsequent data analysis and processing. Without this module's data adaptation and processing, data of varying formats and standards would be unable to effectively flow and analyze within the system, rendering the entire intelligent analysis system ineffective.

[0053] Improved multi-layer perceptron feature extraction and fusion module: connected to the multi-source heterogeneous industrial chain data acquisition adaptation module, receives the converted multimodal data, and extracts features from different modal data separately through the improved multi-layer perceptron architecture, wherein the improved multi-layer perceptron includes an added residual connection structure and adjusted activation function parameters, and then adopts a fusion strategy based on the attention mechanism to fuse the extracted multimodal features to form a multimodal fusion feature vector;

[0054] Specifically, the improved multi-layer perceptron feature extraction and fusion module is a key link in the system's implementation of deep data analysis. Traditional multi-layer perceptrons suffer from problems such as vanishing gradients and insufficient feature extraction when processing complex multimodal data. This module improves these issues by adding a residual connection structure and adjusting the activation function parameters. The introduction of the residual connection structure enables the network to more effectively transfer gradients during training, avoiding performance degradation as the number of network layers increases. For example, when processing market research reports and product sales numerical data containing large amounts of text information, as the number of network layers increases, the residual connection can ensure the network's ability to extract deep features and prevent information loss due to vanishing gradients.

[0055] The adjustment of the activation function parameters is optimized based on the distribution characteristics of the industry chain data. In the industry chain data, numerical data often has a large numerical range and distribution differences. For example, product sales may range from a few hundred yuan to over 100 million yuan. By adjusting the activation function parameters, it is possible to better adapt to this data distribution and improve the network's ability to express data features. After the feature extraction is completed, a fusion strategy based on the attention mechanism is used to fuse the multimodal features. This strategy assigns different weights to the features of different modalities based on the importance of each link in the industry chain. For example, in the electronic product industry chain, product performance parameters (numerical data) are crucial for product quality assessment, while product appearance design (image data) has an important impact on market sales. Based on the attention mechanism, the system will assign different weights to the features of these two modalities according to specific business needs, thereby generating a more representative multimodal fusion feature vector.

[0056] This module can extract deep and valuable features from multimodal data and effectively integrate them, providing high-quality data support for subsequent association modeling and strategy generation based on industry chain parameters, and improving the system's analysis capabilities and accuracy of industry chain data.

[0057] Industrial chain parameter association modeling module: This module is connected to the improved multi-layer perceptron feature extraction and fusion module. Based on the production process, product structure and other parameters of different links in the industrial chain, it constructs an industrial chain topology map with nodes representing different links in the industrial chain and edges representing the relationship between links. The multimodal fusion feature vector is mapped to the corresponding node of the industrial chain topology map, and the association relationship between multimodal data and industrial chain parameters is established.

[0058] Specifically, the industry chain parameter association modeling module is primarily responsible for constructing the industry chain topology and establishing the association between multimodal data and industry chain parameters. When constructing the industry chain topology, parameters such as the production process and product structure of each link in the industry chain are fully considered. For example, in the clothing industry chain, from raw material procurement, fabric processing, garment design, cutting and sewing, finished product inspection, to warehousing and logistics, and marketing, each link is represented as a node in the topology graph, and the edges connecting the nodes represent the business relationships between the links.

[0059] At the same time, the weighting of edges between nodes is determined based on factors such as production capacity and logistics transportation time. For example, the weight of the edge between a fabric processing plant and a garment manufacturing plant is determined based on the degree of match between the fabric processing plant's production capacity and the garment manufacturing plant's demand, as well as the logistics transportation time between the two. Links with high production capacity matching and short logistics transportation times have relatively high edge weights, indicating a closer business connection between the two links. Once the multimodal fusion feature vector is generated, the module maps it to the corresponding node in the industry chain topology. For example, the feature vector obtained after feature extraction and fusion of clothing style design image data is mapped to the node corresponding to the clothing design link, thereby establishing a connection between the data and the actual links in the industry chain.

[0060] In this way, the industrial chain parameter association modeling module can combine abstract data features with specific industrial chain business processes, intuitively display the relationship between data and each link in the industrial chain, and provide a clear logical framework and data foundation for subsequent strategy generation based on the industrial chain topology structure, enabling the system to perform data analysis and decision-making from the perspective of the overall industrial chain.

[0061] Q-learning strategy generation module: This module is connected to the industrial chain parameter association modeling module. Based on the industrial chain topology map and the mapped multimodal fusion feature vector, it uses the Q-learning algorithm to generate a strategy set for data processing and analysis at different links of the industrial chain, with the goal of maximizing long-term cumulative benefits. The state space of the Q-learning algorithm is the node state combination of the industrial chain topology map, and the action space is the set of operations on the data of each node.

[0062] Specifically, the Q-learning strategy generation module utilizes the Q-learning algorithm to generate a set of strategies for data processing and analysis at each stage of the industry chain, based on the industry chain topology map and multimodal fusion feature vectors. In this module, the state space of the Q-learning algorithm is defined as the combination of node states in the industry chain topology map. For example, in the food processing industry chain topology map, the inventory status of the raw material supply node, the equipment operation status of the production and processing node, and the delivery progress status of the logistics and transportation node combine to form a complete state space. The action space is the set of operations performed on the data at each node. For example, at the raw material supply node, actions can include adjusting procurement plans and changing suppliers; at the production and processing node, actions can include adjusting production process parameters and increasing equipment maintenance frequency.

[0063] For example, in a household appliance manufacturing industry chain, when the system detects that the inventory of a key raw material at a raw material supply node is below a safety threshold (state), it uses the Q-learning algorithm to evaluate the potential benefits of different actions (such as emergency procurement and finding alternative materials) in the action space. Through continuous trial and error and learning, it selects the optimal action with the goal of maximizing long-term cumulative benefits. This action, along with its corresponding state and expected benefit, is recorded in a Q-value table. As the system continuously analyzes and learns from industry chain data, the Q-value table is continuously updated, ultimately generating a set of optimal strategies for each industry chain link under different states.

[0064] These strategy sets can provide a scientific decision-making basis for the operation and management of the industrial chain, helping enterprises to adjust production, procurement, sales and other strategies in a timely manner when facing complex and changing market environments and dynamic changes in the industrial chain, so as to achieve optimal resource allocation and maximize economic benefits.

[0065] Strategy execution and feedback module: connects to the Q-learning strategy generation module, receives the generated strategy set, executes the corresponding strategy during the industry chain data processing and analysis process, and collects feedback data after the strategy execution. The feedback data includes the data changes after the strategy execution and the impact on the node status of the industry chain topology map;

[0066] Specifically, the Strategy Execution and Feedback Module is responsible for implementing the strategies developed by the Q-learning Strategy Generation Module and collecting data on the effectiveness of these strategies. Once the system generates a strategy for a specific link in the supply chain, this module transmits the strategy to the corresponding execution terminal. For example, in the agricultural product supply chain, if the system generates a strategy to adjust the storage temperature of agricultural products to extend their shelf life, the Strategy Execution and Feedback Module will transmit this strategy to the temperature control system of the storage link, ensuring that it operates according to the new temperature parameters.

[0067] During policy execution, this module collects feedback data in real time. This data covers changes in data after policy execution and the impact on the status of nodes in the industry chain topology. For example, in the e-commerce industry chain, after the system executes a promotional policy for a specific product, the policy execution and feedback module collects changes in data such as sales volume, sales volume, and inventory quantity. It also monitors changes in the status of the product's sales node in the industry chain topology, such as whether the promotion has accelerated inventory turnover, thereby affecting the workload of logistics and distribution nodes.

[0068] This feedback data is crucial for system optimization and iteration. By analyzing this feedback data, we can evaluate the effectiveness and rationality of strategies, identify problems in strategy execution, and provide accurate data support for the system's optimization and iteration modules. This enables the system to continuously improve and refine its strategy generation and execution mechanisms, ultimately enhancing the intelligent management of the industry chain.

[0069] System optimization iteration module: connected to the improved multi-layer perceptron feature extraction and fusion module, Q-learning strategy generation module and strategy execution and feedback module respectively, adjusts the network parameters of the improved multi-layer perceptron according to the feedback data, and updates the Q value table of the Q-learning algorithm to complete the optimization iteration of the system.

[0070] Specifically, the system optimization and iteration module is a core module that ensures the industry chain's multimodal data intelligent analysis system continuously adapts to dynamic changes in the industry chain, improving analysis accuracy and decision effectiveness. It receives feedback from the strategy execution and feedback module and adjusts the network parameters of the improved multilayer perceptron and the Q-value table of the Q-learning algorithm.

[0071] For the improved multi-layer perceptron, feedback data can reveal deficiencies in the network's feature extraction and fusion processes. For example, if the system's decision-making results in data analysis of a particular industrial chain link deviate significantly from the actual situation, this suggests that the features extracted by the network may not accurately reflect the inherent patterns of the data. In this case, the system's optimization and iteration module adjusts the multi-layer perceptron's network parameters, such as the weight matrix and bias vector, based on the feedback data, optimizing the network structure and improving its ability to extract and fuse multimodal data features.

[0072] Feedback data from the Q-learning algorithm's Q-value table verifies the effectiveness of a strategy. When a strategy fails to achieve its intended profit target during actual execution, the system's optimization and iteration module updates the Q-value table for the corresponding state-action pair based on the feedback data, reassessing the strategy's value under different states. Through continuous adjustments and updates, the Q-learning algorithm generates a set of strategies that better reflect actual conditions.

[0073] Through continuous optimization and iteration, the system can continuously adapt to dynamic factors such as changes in market demand, technological updates, and supply chain adjustments in the industrial chain, maintain efficient analysis and accurate decision-making capabilities for multimodal data in the industrial chain, and provide reliable technical support for the stable operation and development of the industrial chain.

[0074] Preferably, in the improved multi-layer perceptron feature extraction and fusion module, the feature extraction process of the improved multi-layer perceptron is implemented by the following formula:

[0075] H l =σ l (W l H l-1 +b l +R l-1 )

[0076] Among them, H l represents the hidden layer output of the lth layer, l = 1, 2, ..., L, L is the total number of layers of the improved multilayer perceptron; σ l represents the activation function of the lth layer, which is the adjusted LeakyReLU function, and its parameters are set according to the distribution characteristics of the industrial chain data; W l is the weight matrix of the first layer, and its dimension is determined according to the characteristic dimension of data in different links of the industrial chain; b l is the bias vector of the lth layer; R l-1 is the residual connection of the l-1 layer, which is used to alleviate the gradient disappearance problem. Its calculation formula is R l-1 =H l-1 .

[0077] Specifically, the optimization is carried out to address the problems of gradient vanishing and insufficient feature extraction in traditional multi-layer perceptrons when processing multimodal data. By introducing a residual connection structure, a direct information transmission channel is constructed between network layers, and the output of the previous layer is directly added to the output of the current layer. This structural design effectively alleviates the problem of gradient attenuation during deep network training, ensuring that when the network increases the number of layers to extract more complex features, the gradient can still stably backpropagate, maintaining the network's training efficiency and performance. At the same time, based on the numerical distribution characteristics of industrial chain data, the activation function parameters are adjusted in a targeted manner. Industrial chain data often has large numerical spans and uneven distributions. Adjusting the activation function parameters can enable the network to better adapt to the data distribution characteristics and enhance the ability to express the nonlinear characteristics of the data, thereby more accurately extracting representative features from multimodal data, providing a high-quality data foundation for subsequent multimodal feature fusion and analysis, and improving the system's processing capabilities and analysis accuracy for industrial chain data.

[0078] Preferably, the fusion strategy based on the attention mechanism is implemented by the following formula:

[0079]

[0080] Among them, A i represents the attention weight of the i-th modality feature, i = 1, 2, ..., N, N is the total number of modalities; e i is the calculated attention score of the i-th modal feature, which is calculated based on the importance parameters of different links in the industrial chain; V i represents the feature vector of the i-th modality; F is the multimodal fusion feature vector after fusion.

[0081] Specifically, a fusion strategy based on the attention mechanism aims to address the issue of varying importance of different modal data to different links in the multimodal data fusion process within an industry chain. This strategy assigns weights to features of different modalities based on the business characteristics and actual needs of each link in the industry chain. In the actual operation of an industry chain, key influencing factors vary across links, and the corresponding multimodal data importance also varies. Through in-depth analysis of the industry chain's business processes, the modal data that requires focus in each link is identified, and these modal features are assigned higher weights, while less important modal features are assigned lower weights. During the fusion process, a weighted summation of the different modal features is performed, enabling the fused multimodal feature vector to highlight data features closely related to the current link's business, while suppressing interference from irrelevant or minor features, thereby generating a more targeted and effective feature representation. This fusion strategy accurately mines data value, providing information support that better aligns with actual business needs for industry chain decision-making, and improving the scientificity and accuracy of decision-making.

[0082] Preferably, in the industrial chain parameter association modeling module, the process of constructing the industrial chain topology map combines the production capacity parameters and logistics transportation time parameters in the industrial chain, and the weights of the edges between nodes are calculated by the following formula:

[0083] w ij =α×C ij +β×T ij

[0084] Among them, w ij represents the weight of the edge between node i and node j; C ij is the capacity matching parameter between node i and node j, and its value is determined according to the capacity ratio relationship of the links represented by the two nodes; T ij is the logistics transportation time parameter between node i and node j; α and β are weight coefficients, which are set according to the actual operation of the industrial chain, and α + β = 1.

[0085] Specifically, when constructing the industry chain topology, edge weights are determined by comprehensively considering capacity parameters and logistics transportation time parameters. Capacity parameters reflect the degree of production capacity matching between upstream and downstream links in the industry chain, including factors such as production scale and supply capacity. A high degree of capacity matching indicates better coordination between upstream and downstream links in production capacity, ensuring the stable operation of the industry chain. Logistics transportation time parameters reflect the efficiency of material flow between links. Short transportation times indicate timely and reliable logistics distribution, helping to reduce inventory backlogs and production delays. By weighting these two parameters, capacity matching and logistics transportation time are converted into edge weights, visually demonstrating the closeness of business connections between links. Higher weights indicate closer links and greater interdependence. Based on this, enterprises can clearly grasp the structural relationships in the industry chain and prioritize collaboration with higher-weighted links in decision-making processes such as resource allocation and supplier selection, optimizing the industry chain layout and improving overall operational efficiency and economic benefits.

[0086] Preferably, in the Q-learning strategy generation module, the Q-value update formula of the Q-learning algorithm is:

[0087]

[0088] Among them, Q(s t , a t ) means in state s t Next, perform action a t Q value, s t is the node state combination of the industry chain topology at time t, a t is the operation performed on the node data at time t; γ is the discount factor, which is used to balance current rewards and future rewards. Its value is set according to the dynamic change speed of the industrial chain; In state s t+1 The maximum Q value of all possible actions.

[0089] Specifically, the Q-value update process of the Q-learning algorithm is a key mechanism for dynamic strategy generation within the industrial chain. In the complex and volatile operating environment of the industrial chain, factors such as market demand, raw material prices, and policies and regulations are constantly changing, requiring the system to dynamically adjust its strategy to adapt to these changes. The Q-learning algorithm aims to maximize long-term cumulative returns by optimizing its strategy through trial and error and learning. For each decision, the algorithm selects an action based on the current industrial chain state and obtains the actual return after execution. The actual return is compared with the expected return before the action is executed, and the Q-value of the corresponding state-action pair is updated based on the difference. If the actual return is higher than the expected return, the Q-value of the state-action pair is increased, indicating that the action has a better return in that state, making it more likely to be chosen in the same or similar states. Otherwise, the Q-value is decreased. As the system continuously collects industrial chain data and learns, the Q-value table is continuously updated and refined, gradually determining the optimal set of strategies for various states. These strategies can make scientific decisions based on the real-time state of the industrial chain, helping enterprises navigate complex and volatile market environments, optimize resource allocation, and maximize economic benefits.

[0090] Preferably, in the strategy execution and feedback module, the collection of feedback data is also combined with inventory parameters in the industrial chain, and the changes in inventory data after the strategy is executed are quantified using the following method:

[0091]

[0092] Among them, ΔI k It represents the inventory change of the kth link before and after the implementation of the strategy; is the inventory quantity at the kth stage after executing the strategy; is the inventory quantity at the kth stage before executing the strategy.

[0093] Specifically, the strategy execution and feedback module quantifies changes in inventory parameters, which is crucial for inventory management within the supply chain. As a critical link in supply chain operations, inventory status directly impacts a company's capital turnover, production plans, and market supply. During strategy execution, the system monitors changes in inventory parameters in real time, including inventory quantity and inventory turnover. By accurately calculating inventory changes, the system quantifies dynamic changes in inventory, clearly demonstrating the impact of strategy execution on inventory. For example, after implementing a promotional strategy, calculating the decrease in inventory quantity and the change in inventory turnover allows for a direct assessment of the promotion's impact on product sales and inventory consumption. Furthermore, inventory changes can trigger chain reactions throughout the supply chain. For example, insufficient inventory can cause production to shut down while waiting for materials, while accumulated inventory increases storage costs and capital utilization. Quantifying changes in inventory parameters provides the system with accurate feedback, enabling it to promptly identify inventory management issues and providing data support for subsequent optimization and iteration. This helps companies adjust production, procurement, and sales strategies, achieve rational inventory control and optimized management, and ensure the smooth operation of the supply chain.

[0094] Preferably, in the system optimization iteration module, an adaptive learning rate algorithm is used to adjust the network parameters of the improved multi-layer perceptron, and the adjustment formula of the learning rate η is:

[0095]

[0096] Among them, η t represents the learning rate at time t; η0 is the initial learning rate, which is set according to the complexity of the industry chain data; λ is the learning rate decay parameter; and t is the number of iterations.

[0097] Specifically, an adaptive learning rate algorithm is used to adjust the network parameters of the improved multi-layer perceptron to adapt to the dynamic nature of industrial chain data. During industrial chain data processing, the size, distribution, and characteristics of the data continuously change as production and business activities progress. The adaptive learning rate algorithm dynamically adjusts the learning rate based on the statistical characteristics of the data and the training process. In the early stages of training, when data uncertainty is high, the algorithm sets a higher learning rate to enable rapid network convergence, accelerate the learning process of the data features, and improve training efficiency. As training progresses, the network gradually learns the data's characteristic patterns. At this point, the algorithm reduces the learning rate, making the network parameter updates more refined. This avoids missing the optimal solution due to excessively high learning rates, thereby improving the network's training accuracy and generalization ability. By dynamically adjusting the learning rate, the improved multi-layer perceptron can better adapt to the dynamic changes in industrial chain data, continuously optimize network parameters, and enhance its ability to extract and fuse features from different stages and types of data, ensuring that the system maintains efficient and accurate data analysis capabilities in the complex and changing industrial chain environment.

[0098] Preferably, in the Q-learning strategy generation module, when generating the strategy set, the cost parameters in the industrial chain are also combined to integrate the cost parameters into the action selection process. The probability formula for action selection is:

[0099]

[0100] Among them, P(a t ∣s t ) means in state s t Next select action a t ; τ is the temperature parameter, which is used to control the randomness of action selection, and its value is set according to the cost fluctuation of the industrial chain.

[0101] Specifically, cost parameters are incorporated into the action selection process of the Q-learning strategy generation module to optimize costs across the industry chain. Cost factors permeate every aspect of industry chain operations, directly impacting a company's economic benefits and market competitiveness. During the action selection phase of the Q-learning algorithm, the system not only considers the degree to which an action achieves business objectives but also comprehensively calculates the costs involved in executing each action, including raw material procurement costs, production and processing costs, logistics and transportation costs, and labor costs. Through a cost-benefit analysis of different actions, the value of each action in its current state is determined. Based on this, a probabilistic selection mechanism prioritizes actions with lower costs that meet business requirements. This ensures that the generated strategy set achieves business objectives while effectively controlling costs. This integration of cost parameters into action selection helps companies fully consider cost factors when formulating industry chain operation strategies, optimize resource allocation, avoid unnecessary cost waste, improve cost management and economic efficiency, and enhance their competitive advantage in the market.

[0102] Preferably, when extracting features from the industrial chain data in text format, the improved multi-layer perceptron feature extraction and fusion module combines the keyword frequency parameters in the text and performs weighted processing in the following manner:

[0103]

[0104] Among them, H text is the processed text feature vector; M is the number of keywords in the text; f k is the frequency parameter of the kth keyword; h k is the feature vector corresponding to the kth keyword.

[0105] Specifically, the improved multi-layer perceptron (MLP) processes text-based industry chain data by weighting it with keyword frequency, improving the effectiveness of feature extraction from text data. Within the industry chain, text data such as market research reports, technical documents, and contract terms contain a wealth of information, but the importance of different content varies. Keywords, as the core summary of text content, reflect the importance and attention of the relevant information to a certain extent through their frequency. When extracting features from text data, the system calculates the frequency of each keyword in the text and assigns weights to the corresponding feature vectors based on their frequency. Feature vectors of keywords with high frequency are assigned higher weights and carry a larger weight in the multimodal feature fusion process; feature vectors of keywords with low frequency receive relatively lower weights. This weighted processing method highlights key information in the text data and suppresses interference from secondary information, enabling the extracted text features to more accurately reflect the core content of the text. When fused with other modal data, the weighted text features can more effectively contribute valuable information, improving the quality of the multimodal fusion feature vectors. This provides more accurate and valuable text information support for industry chain decision-making, helping companies better grasp market dynamics and industry trends.

[0106] like Figure 2 As shown, the industry chain multimodal data intelligent analysis system includes the following steps:

[0107] The first step is to establish communication connections with data collection terminals at different links of the industrial chain through the multi-source heterogeneous industrial chain data collection adaptation module according to the preset data interface protocol, collect multimodal data, and convert the collected data into a unified data structure format;

[0108] In the second step, the converted multimodal data is transmitted to the improved multi-layer perceptron feature extraction and fusion module. The improved multi-layer perceptron architecture is used to extract features from different modal data respectively, and a fusion strategy based on the attention mechanism is adopted to fuse the extracted multimodal features into a multimodal fusion feature vector.

[0109] In the third step, the industrial chain parameter association modeling module constructs an industrial chain topology map based on the parameters of different links in the industrial chain, and maps the multimodal fusion feature vectors to the corresponding nodes in the topology map to establish the association between data and industrial chain parameters;

[0110] In the fourth step, the Q-learning strategy generation module uses the Q-learning algorithm based on the industry chain topology map and multimodal fusion feature vectors to generate a strategy set for data processing and analysis at different links in the industry chain with the goal of maximizing long-term cumulative benefits.

[0111] In the fifth step, the strategy execution and feedback module receives the strategy set and executes it during the industry chain data processing and analysis process, while also collecting feedback data after the strategy is executed;

[0112] In the sixth step, the system optimization iteration module adjusts the network parameters of the improved multi-layer perceptron based on the feedback data, updates the Q value table of the Q-learning algorithm, and performs system optimization iteration.

[0113] The system uses an improved multi-layer perceptron feature extraction and fusion module to first accurately extract features from different modal data, introduce a residual connection structure to address the vanishing gradient problem in deep networks, and adjust activation function parameters to adapt to the data distribution of the industrial chain. Then, based on the attention mechanism, it assigns weights to different modal features based on the business needs of each link in the industrial chain, achieving efficient fusion. In this way, the system can deeply explore the intrinsic value of multimodal data, transforming originally scattered and isolated data into information with comprehensive analytical value, providing comprehensive and accurate data support for industrial chain decision-making.

[0114] In terms of strategy generation, the system utilizes the Q-learning model combined with the actual parameters of the industrial chain to solve the problem of the lack of dynamic adaptability in existing technology strategy formulation. Traditional systems are based on fixed algorithm models and are unable to cope with dynamic changes in market demand, raw material prices, and other factors within the industrial chain. The Q-learning strategy generation module of this system is based on the industrial chain topology map, integrating parameters such as production capacity and logistics transportation time into the state space and action space construction, and generating strategies with the goal of maximizing long-term cumulative benefits. During the execution process, multi-dimensional feedback data such as inventory and cost are collected through the strategy execution and feedback module, and the Q-value table of the Q-learning algorithm is updated based on the data to continuously optimize the strategy. This dynamic learning and adjustment mechanism enables the system to adapt to changes in the industrial chain environment in real time, providing enterprises with scientific decision-making strategies that fit the actual situation and avoiding operational risks caused by strategy lags.

[0115] From the perspective of the system's overall operation, its optimization and iteration mechanism ensures continuous improvement in system performance. The adaptive learning rate algorithm adjusts the parameters of the improved multi-layer perceptron network based on the dynamic changes in the scale, distribution, and characteristics of industry chain data, achieving rapid convergence in the early stages of training and fine-tuning in the later stages to ensure the network's efficient extraction and integration of data features. Cost parameters are integrated into the action selection process of the Q-learning strategy generation module to achieve optimized control of industry chain costs and avoid resource waste. The system optimization and iteration module integrates feedback data from various aspects and continuously adjusts the parameters of the improved multi-layer perceptron and Q-learning model, forming a virtuous cycle. This allows the system to maintain efficient and accurate data analysis and decision-making capabilities in the complex and changing industry chain environment, comprehensively improving the intelligent management level and economic benefits of the industry chain.

[0116] In the description of the present invention, it should be noted that, unless otherwise expressly specified or limited, the terms "disposed," "installed," "connected," "connected," and "fixed" should be understood in a broad sense. For example, they may refer to fixed connections, detachable connections, or integral connections; they may refer to mechanical connections or electrical connections; they may refer to direct connections or indirect connections through an intermediate medium; and they may refer to internal communication between two components. Those skilled in the art will understand the specific meanings of the above terms in the present invention based on specific circumstances.

[0117] While embodiments of the present invention have been shown and described, it will be understood by those skilled in the art that various equivalent changes, modifications, substitutions and variations may be made to the embodiments without departing from the principles and spirit of the invention, and that the scope of the invention is defined by the appended claims and their equivalents.

Claims

1. An industrial chain multimodal data intelligent analysis system, characterized by: include: Multi-source heterogeneous industry chain data acquisition adaptation module: This module is used to establish communication connections with multimodal data acquisition terminals distributed in different links and entities in the industry chain. Based on the preset data interface protocol, it collects multimodal data and converts the collected data into a unified data structure format based on the data type and source. Improved multi-layer perceptron feature extraction and fusion module: This module is used to receive the converted multimodal data, extract features from different modal data separately through the improved multi-layer perceptron architecture, and fuse the extracted multimodal features to form a multimodal fusion feature vector; Industrial chain parameter association modeling module: This module is used to construct an industrial chain topology map based on the production processes and product structure parameters of different links in the industrial chain. The map contains nodes representing different links in the industrial chain and edges representing the relationships between links. The module maps the multimodal fusion feature vectors to the corresponding nodes of the industrial chain topology map and establishes the association between multimodal data and industrial chain parameters. Q-learning strategy generation module: Based on the industry chain topology map and the mapped multimodal fusion feature vector, this module uses the Q-learning algorithm to generate a strategy set for data processing and analysis at different links in the industry chain, with the goal of maximizing long-term cumulative benefits; Strategy execution and feedback module: This module is used to receive the generated strategy set, execute the corresponding strategy during the industry chain data processing and analysis, and collect feedback data after the strategy execution. The feedback data includes the data changes after the strategy execution and the impact on the node status of the industry chain topology map; System optimization iteration module: This module is used to adjust the network parameters of the improved multi-layer perceptron according to the feedback data, and update the Q value table of the Q-learning algorithm to complete the optimization iteration of the system.

2. The multimodal data intelligent analysis system for the industrial chain according to claim 1 is characterized in that: The improved multi-layer perceptron feature extraction and fusion module, the feature extraction process of the improved multi-layer perceptron is implemented by the following formula: H l =σ l (W l H l-1 +b l +R l-1 ) Among them, H l represents the hidden layer output of the lth layer, l = 1, 2, ..., L, L is the total number of layers of the improved multilayer perceptron; σ l represents the activation function of the lth layer, which is the adjusted LeakyReLU function, and its parameters are set according to the distribution characteristics of the industrial chain data; W l is the weight matrix of the first layer, and its dimension is determined according to the characteristic dimension of data in different links of the industrial chain; b l is the bias vector of the lth layer; R l-1 is the residual connection of the l-1 layer, which is used to alleviate the gradient disappearance problem. Its calculation formula is R l-1 =H l-1 .

3. The multimodal data intelligent analysis system for the industrial chain according to claim 1 is characterized in that: The improved multi-layer perceptron feature extraction and fusion module is implemented by the following formula based on the fusion strategy of the attention mechanism: Among them, A i represents the attention weight of the i-th modality feature, i = 1, 2, ..., N, N is the total number of modalities; e i is the calculated attention score of the i-th modal feature, which is calculated based on the importance parameters of different links in the industrial chain; V i represents the feature vector of the i-th modality; F is the multimodal fusion feature vector after fusion.

4. The industrial chain multimodal data intelligent analysis system according to claim 1 is characterized in that: The industrial chain parameter association modeling module and the industrial chain topology map construction process combine the production capacity parameters and logistics transportation time parameters in the industrial chain. The weight of the edge between nodes is calculated by the following formula: In ij =α×C ij +β×T ij Among them, w ij represents the weight of the edge between node i and node j; C ij is the capacity matching parameter between node i and node j, and its value is determined according to the capacity ratio relationship of the links represented by the two nodes; T ij is the logistics transportation time parameter between node i and node j; α and β are weight coefficients, which are set according to the actual operation of the industrial chain, and α + β = 1.

5. The industrial chain multimodal data intelligent analysis system according to claim 1 is characterized in that: The Q-learning strategy generation module and the Q-learning algorithm Q-value update formula are as follows: Among them, Q(s t , a t ) means in state s t Next, perform action a t Q value, s t is the node state combination of the industry chain topology at time t, a t is the operation performed on the node data at time t; γ is the discount factor, which is used to balance current rewards and future rewards. Its value is set according to the dynamic change speed of the industrial chain; In state s t+1 The maximum Q value of all possible actions.

6. The industrial chain multimodal data intelligent analysis system according to claim 1 is characterized in that: The strategy execution and feedback module also collects feedback data based on inventory parameters in the industry chain. The changes in inventory data after strategy execution are quantified using the following methods: Among them, ΔI k It represents the inventory change of the kth link before and after the implementation of the strategy; is the inventory quantity at the kth stage after executing the strategy; is the inventory quantity at the kth stage before executing the strategy.

7. The industrial chain multimodal data intelligent analysis system according to claim 1 is characterized in that: The system optimization iteration module uses an adaptive learning rate algorithm to adjust the network parameters of the improved multi-layer perceptron. The adjustment formula of the learning rate η is: Among them, η t represents the learning rate at time t; η0 is the initial learning rate, which is set according to the complexity of the industry chain data; λ is the learning rate decay parameter; and t is the number of iterations.

8. The industrial chain multimodal data intelligent analysis system according to claim 1 is characterized in that: The Q-learning strategy generation module combines the cost parameters in the industrial chain when generating the strategy set and incorporates the cost parameters into the action selection process. The probability formula for action selection is: Among them, P(a t ∣s t ) means in state s t Next select action a t ; τ is the temperature parameter, which is used to control the randomness of action selection, and its value is set according to the cost fluctuation of the industrial chain.

9. The industrial chain multimodal data intelligent analysis system according to claim 1 is characterized in that: When extracting features from text-formatted industry chain data, the improved multi-layer perceptron feature extraction and fusion module combines the keyword frequency parameters in the text and performs weighted processing in the following manner: Among them, H text is the processed text feature vector; M is the number of keywords in the text; f k is the frequency parameter of the kth keyword; h k is the feature vector corresponding to the kth keyword.

10. The industrial chain multimodal data intelligent analysis system according to any one of claims 1 to 9, characterized in that: The system operation includes the following steps: The first step is to establish communication connections with data collection terminals at different links of the industrial chain through the multi-source heterogeneous industrial chain data collection adaptation module according to the preset data interface protocol, collect multimodal data, and convert the collected data into a unified data structure format; In the second step, the converted multimodal data is transmitted to the improved multi-layer perceptron feature extraction and fusion module. The improved multi-layer perceptron architecture is used to extract features from different modal data respectively, and a fusion strategy based on the attention mechanism is adopted to fuse the extracted multimodal features into a multimodal fusion feature vector. In the third step, the industrial chain parameter association modeling module constructs an industrial chain topology map based on the parameters of different links in the industrial chain, and maps the multimodal fusion feature vectors to the corresponding nodes in the topology map to establish the association between data and industrial chain parameters; In the fourth step, the Q-learning strategy generation module uses the Q-learning algorithm based on the industry chain topology map and multimodal fusion feature vectors to generate a strategy set for data processing and analysis at different links in the industry chain with the goal of maximizing long-term cumulative benefits. In the fifth step, the strategy execution and feedback module receives the strategy set and executes it during the industry chain data processing and analysis process, while also collecting feedback data after the strategy is executed; In the sixth step, the system optimization iteration module adjusts the network parameters of the improved multi-layer perceptron based on the feedback data, updates the Q value table of the Q-learning algorithm, and performs system optimization iteration.

Images