Industrial space intelligent association optimization system based on multi-dimensional data analysis

Abstract

The invention provides an industrial space intelligent association optimization system based on multi-dimensional data analysis, which relates to the technical field of industrial space association and comprises the following modules: a data acquisition module, a data preprocessing module, an association analysis module, an optimization decision module, a prediction and feedback module and an intelligent self-learning module. The multi-source heterogeneous data collection and edge computing technology is adopted, real-time data collection from multiple sources such as a sensor network, satellite remote sensing and a public database is supported, data of all nodes in an industrial chain can be rapidly and accurately obtained through multi-dimensional data collection in a complex industrial environment, and the data of all nodes in the industrial chain can be rapidly and accurately obtained through an efficient data preprocessing module. The system can process data in various formats, including industrial chain data, economic activity data, geographic space data and the like, realizes unified normalization and cleaning of different data sources, and ensures the accuracy and consistency of the data.

Description

Technical Field

[0001] The present invention relates to the field of industrial space association technology, and specifically to an industrial space intelligent association optimization system based on multidimensional data analysis. Background Art

[0002] According to a development zone industrial layout method based on industry-space compatibility disclosed in Chinese patent publication number "CN109102434A", it discloses the division of basic industrial space units based on the industry-space hierarchy system; comparing the industrial categories of the industrial development plan, the demand conditions for industrial development are matched with the various conditions of the industrial zone itself for preliminary selection, and the broad-spectrum industrial clusters of each basic industrial space unit are determined; after the industrial type of each basic industrial space unit is preliminarily determined, the correlation, exclusivity and spatial proximity of the industrial categories and other factors are further considered to optimize the industrial category selection and specific layout location of the basic industrial space unit to form a final industrial layout plan.

[0003] According to the irregular port-side industrial zone layout system and method based on multi-agent evolutionary algorithm disclosed in Chinese patent publication number "CN109934513B Irregular port-side industrial zone layout system and method based on multi-agent evolutionary algorithm", a system and method for irregular port-side industrial zone layout based on multi-agent evolutionary algorithm are disclosed. The system specifically includes: a data acquisition unit, used to acquire map data of the port-side industrial zone to be laid out; a rasterization processing unit, used to rasterize the area of ​​the port-side industrial zone; a raster allocation unit, used to allocate a corresponding number of grids to each port-side industry according to its planned area and to each existing building according to its actual area; a model establishment unit, which takes the sum of the correlations between each grid as the comprehensive correlation, and constructs an irregular port-side industrial layout model with the goal of maximizing the comprehensive correlation of the entire port-side industrial zone; a model solving unit, which is used to solve the irregular port-side industrial layout model by multi-agent evolutionary algorithm, and the result obtained is the port-side industrial layout plan of the port-side industrial zone.

[0004] The above patent documents and prior art have the following technical problems when used:

[0005] 1. In the modern industrial environment, data between different industrial chain nodes often come from a wide range of sources, including internal management systems, external public databases, sensor data, satellite remote sensing data, etc. These data types are diverse and complex in format, making them difficult to analyze and integrate directly. Traditional data processing methods often cannot effectively integrate these multi-source heterogeneous data, resulting in incomplete decision-making information and an inability to accurately reflect the overall dynamics of the industrial chain;

[0006] 2. There are a large number of potential nonlinear associations between industrial nodes. These associations often involve multi-dimensional time series data and are not easy to accurately identify through traditional linear models. Traditional analysis methods have difficulty capturing implicit associations between industrial nodes when processing high-dimensional time series data, resulting in a lack of accurate understanding of the mutual influence between nodes in the decision-making process;

[0007] 3. In actual industrial chain management, optimization decisions often involve multiple goals, such as efficient resource allocation, cost control, and increased production capacity. These goals may conflict with each other, and traditional single-objective or simple multi-objective optimization methods cannot take into account multiple complex goals at the same time, resulting in limited optimization results and difficulty in meeting actual needs. Summary of the invention

[0008] Technical issues solved

[0009] In view of the shortcomings of the prior art, the present invention provides an industrial space intelligent association optimization system based on multidimensional data analysis, which solves the following problems:

[0010] 1. In the modern industrial environment, data between different industrial chain nodes often come from a wide range of sources, including internal management systems, external public databases, sensor data, satellite remote sensing data, etc. These data types are diverse and complex in format, making them difficult to analyze and integrate directly. Traditional data processing methods often cannot effectively integrate these multi-source heterogeneous data, resulting in incomplete decision-making information and an inability to accurately reflect the overall dynamics of the industrial chain;

[0011] 2. There are a large number of potential nonlinear associations between industrial nodes. These associations often involve multi-dimensional time series data and are not easy to accurately identify through traditional linear models. Traditional analysis methods have difficulty capturing implicit associations between industrial nodes when processing high-dimensional time series data, resulting in a lack of accurate understanding of the mutual influence between nodes in the decision-making process;

[0012] 3. In actual industrial chain management, optimization decisions often involve multiple goals, such as efficient resource allocation, cost control, and increased production capacity. These goals may conflict with each other, and traditional single-objective or simple multi-objective optimization methods cannot take into account multiple complex goals at the same time, resulting in limited optimization results and difficulty in meeting actual needs.

[0013] Technical Solution

[0014] To achieve the above objectives, the present invention is implemented through the following technical solutions: an industrial space intelligent association optimization system based on multidimensional data analysis, the association optimization system includes the following modules:

[0015] Data collection module, used to obtain multi-dimensional data related to industrial space, including industrial chain data, geographic space data, economic activity data, technology development trend data and policy and regulation data;

[0016] The data preprocessing module is equipped with a multi-dimensional data denoising processor, a data normalization processor, and a dynamic data structuring processor to clean and normalize data from different sources and formats and generate a standardized data input format;

[0017] The association analysis module uses a nonlinear time series deep embedding algorithm to extract implicit association features from multidimensional data, and realizes intelligent association between industrial chain nodes through multi-level spatial relationship modeling; the nonlinear time series deep embedding algorithm extracts time series association features from multidimensional data and performs multi-level relationship modeling. Time series feature extraction: feature extraction of input multi-dimensional time series data:

[0018] X(t)=[x1(t),x2(t),…,x n (t)]

[0019] X(t) is a multidimensional data vector at time t, representing industry data with n different dimensions; i (t): data of the i-th dimension at time t, i∈{1,2,…,n}

[0020] The embedding space is constructed using the embedding delay method:

[0021] Y(t)=[X(t),X(t-τ),X(t-2τ),…,X(t-mτ)]

[0022] Y(t): time series data after delayed embedding, containing multidimensional data of the past m time points; τ: delay time step, used to control the interval between historical time points; m: embedding dimension, indicating the number of steps of historical data;

[0023] The data is decomposed into time series features at different scales through wavelet transform:

[0024]

[0025] W(s,t): The result of the time series after wavelet transformation when the scale is s; X(t′): The original time series data at time point t′; ψ(t): Wavelet function, also called mother wavelet, used to decompose the signal; s: Scale parameter, indicating the time scale of the analysis (large s corresponds to a long time scale, small s corresponds to a short time scale); t: Current time; ψ * : The complex conjugate function of the mother wavelet; This method is used to capture the changing characteristics of different time scales; Deep embedding structure: Construct a nonlinear deep embedding neural network and model the time series data through a recurrent neural network (RNN):

[0026] h t =f(W·Y t+U·h t-1 +b)

[0027] h t : The hidden state vector at the current moment, containing the embedded time series features; W: Input weight matrix, used to process the input data Y t ; Y t : The multidimensional time series data input at the current moment. U: Recursive weight matrix, used to process the hidden state h at the previous moment t-1 ;h t-1 : The hidden state of the previous moment; b: Bias term, used to adjust the output of the network; f: Activation function;

[0028] The optimization decision module, based on the deep mining and dynamic analysis of multi-dimensional data, adopts an interactive optimization algorithm based on adaptive multi-objective reinforcement learning to make real-time optimization decisions for each related node to improve the efficiency of industrial resource allocation and the rationality of spatial layout; the interactive optimization algorithm of adaptive multi-objective reinforcement learning includes the following parts:

[0029] Value function:

[0030]

[0031] V(s): expected cumulative reward value in state s; : Expected value operator, used to calculate the weighted average of all possible results; γ: discount factor, ranging from 0<γ≤1, used to control the importance of future rewards (the larger γ, the more important future rewards); R(s t ,a t ): At time t, state s t Take action a t The instant rewards received; t : state at time t; a t : action at time t;

[0032] Particle swarm optimization (PSO) speed update formula:

[0033]

[0034] The velocity of the i-th particle at time t+1; The speed of the i-th particle at time t; w: inertia factor, which controls the influence of the current speed of the particle on the next step; c1, c2: learning factors, which are used to control the particle to the individual optimal position p i and the global optimal position g i Approach speed; r1, r2: random numbers between 0 and 1; p i : The individual optimal position of the i-th particle; gi : The global optimal position of the group; The position of the i-th particle at time t;

[0035] Action strategy:

[0036]

[0037] P(a|s): the probability of choosing action a in state s; Q(s,a): the value function of taking action a in state s; a′: the action in the possible action set;

[0038] Fuzzy target optimization:

[0039]

[0040] μ(g i ) The fuzzy membership of target g_i; g i : the i-th target value; g i0 : Target g i The ideal threshold of ; d: controls the width of the fuzzy set, which is used to adjust the degree of fuzzification;

[0041] The prediction and feedback module is equipped with an adaptive Bayesian network dynamic causal reasoning model to predict the development trend of industrial space, iteratively optimize based on historical data and real-time data, and generate operational adjustment strategies and feedback; the specific contents of the adaptive Bayesian network dynamic causal reasoning model are as follows:

[0042] Bayesian network joint probability distribution:

[0043]

[0044] P(X1,X2,…,X n ): joint probability distribution of n variables in the Bayesian network; X i : the i-th variable; Pa(X i ): the parent node set of the i-th variable;

[0045] Maximum a posteriori estimation (MAP):

[0046]

[0047] P(X t ∣X t-1 ): Conditional probability distribution at time t given the data at time t-1; P(X t-1 |θ): Prior probability based on parameter θ; P(θ|X t-1 ): the posterior distribution of parameter θ; P(X t-1 ): marginal probability of the data at time t-1;

[0048] Modular logic reasoning formula:

[0049]

[0050] μ(y) is the fuzzy membership of the result y after reasoning; μ A (x), μ B (y): fuzzy membership function of input x and output y; f(x): functional relationship defined by fuzzy rules;

[0051] The intelligent self-learning module can update association rules and optimization strategies in real time according to new data and market changes through a multi-level time-varying causal network learning algorithm, ensuring that the system has adaptive learning and adjustment capabilities; the time-varying causal network learning algorithm is as follows:

[0052] Causal network learning objective function:

[0053]

[0054] L: loss function of the causal network; P(X i ∣∣Pa(X i )): conditional probability of the i-th variable given the parent node set; λ: regularization parameter used to control the causal relationship variable; A ij : The adjacency matrix element of the causal network, indicating whether there is a causal relationship between node i and node j;

[0055] Model parameter update formula:

[0056]

[0057] θ t+1 : Model parameters at time t+1; θ t : Model parameters at time t; α: learning rate, controlling the step size of parameter update; The gradient of the loss function with respect to the parameters θ.

[0058] Preferably, the data acquisition module further includes a distributed data acquisition unit, which can collect data in real time from multiple heterogeneous data sources, including sensor networks, satellite remote sensing data, industrial chain internal management systems and public databases, and pre-process the collected data in real time through edge computing technology.

[0059] Preferably, the multidimensional data denoising processor of the data preprocessing module adopts an adaptive multi-scale wavelet denoising algorithm, which can automatically select the best denoising parameters according to the noise characteristics of different types of data to improve the validity and accuracy of the data.

[0060] Preferably, the nonlinear time series deep embedding algorithm in the association analysis module is further combined with multi-scale and multi-level regression analysis, which can identify the interactive relationship between industrial chain nodes at different time scales while extracting time series association features.

[0061] Preferably, the optimization decision module adopts an adaptive multi-objective reinforcement learning algorithm based on hybrid particle swarm optimization and deep belief network to dynamically adjust the priority of the optimization objectives and achieve rapid convergence when computing resources are limited.

[0062] Preferably, the adaptive Bayesian network dynamic causal reasoning model of the prediction and feedback module is further combined with a fuzzy logic reasoning system, which can improve the accuracy and stability of prediction in situations with high uncertainty.

[0063] Preferably, the intelligent self-learning module adopts a progressive deep learning framework, combined with an online learning algorithm, and periodically updates the training model to ensure that the system still has a high degree of responsiveness and adaptability in a rapidly changing industrial environment.

[0064] Preferably, the working steps of the association optimization system are as follows:

[0065] Sp1. Obtain multi-dimensional industry-related data through the data acquisition module and input it into the data preprocessing module for data cleaning and structuring;

[0066] Sp2, the data enters the association analysis module, and the nonlinear time series deep embedding algorithm (NLTDE) extracts hidden association features from it, and multi-level relationship modeling is performed based on the spatial and temporal dimensions of each industry node;

[0067] Sp3, the optimization decision module adopts the adaptive multi-objective reinforcement learning (AMORL) algorithm to optimize the industrial layout under the constraints of multiple objectives and output an executable optimization plan;

[0068] Sp4, prediction and feedback module predicts development trends and provides feedback optimization strategies based on the adaptive Bayesian network dynamic causal reasoning model (ABNDCR);

[0069] Sp5, the intelligent self-learning module uses the time-varying causal network learning algorithm (MTCNL) to adaptively learn and update association rules and models.

[0070] Preferably, the nonlinear time series deep embedding algorithm (NLTDE) is combined with a time series analysis method based on density clustering.

[0071] Preferably, the adaptive multi-objective reinforcement learning (AMORL) algorithm is combined with fuzzy objective optimization technology to dynamically adjust the fuzzy set of constraint conditions.

[0072] Beneficial Effects

[0073] The present invention provides an industrial space intelligent association optimization system based on multidimensional data analysis. It has the following beneficial effects:

[0074] 1. The present invention adopts multi-source heterogeneous data collection and edge computing technology, and supports real-time data collection from multiple sources such as sensor networks, satellite remote sensing, public databases, etc. This multi-dimensional data collection can quickly and accurately obtain data from each node in the industrial chain under complex industrial environments. Through efficient data preprocessing modules, the system can process data in various formats, including industrial chain data, economic activity data, geographic spatial data, etc., to achieve unified normalization and cleaning of different data sources, ensure data accuracy and consistency. The introduction of edge computing technology reduces data transmission delays, ensures the real-time nature of data, and effectively improves the response speed of the system.

[0075] 2. The association analysis module of the present invention adopts a nonlinear time series deep embedding algorithm (NLTDE), which combines time series analysis and clustering technology. It can extract time series correlation features from multidimensional data, and identify the deep correlation relationship between industrial chain nodes through multi-scale and multi-level regression analysis. Through embedded time series and wavelet transform technology, the system can perform time series analysis of different scales in the time dimension, accurately identify the complex nonlinear correlation between industrial chain nodes, especially in multi-dimensional time series data, and capture the potential interaction patterns between different nodes. Through dimensionality reduction technologies such as t-SNE, the system can effectively perform dimensionality reduction processing in high-dimensional data space, reduce data redundancy, and identify mutually related nodes in the industrial chain through clustering analysis. This analysis provides important correlation information for subsequent optimization decisions.

[0076] 3. The present invention adopts the adaptive multi-objective reinforcement learning (AMORL) algorithm, combined with fuzzy target optimization technology, and can adaptively optimize the industrial layout under complex multi-objective constraint environments. Through reinforcement learning, the system can dynamically adjust the optimization strategy and find the optimal industrial chain layout plan. The system can handle multiple optimization goals at the same time, such as improving the resource allocation efficiency of the industrial chain, reducing costs, and increasing production capacity. In each optimization process, the system accelerates the search for the optimal solution through particle swarm optimization (PSO), ensuring that it can quickly converge to the optimal decision-making plan when computing resources are limited. In an environment with high uncertainty, fuzzy optimization technology ensures that the system can flexibly adjust the priorities of different goals, which is particularly suitable for industrial layout scenarios with multiple constraints and uncertain factors. BRIEF DESCRIPTION OF THE DRAWINGS

[0077] Figure 1 It is a normalized multi-dimensional data graph of the industrial chain nodes of the present invention;

[0078] Figure 2 It is the industrial chain node correlation matrix diagram of the present invention;

[0079] Figure 3 Embedding scatter plot of node association features of the present invention;

[0080] Figure 4 It is the industrial chain node clustering tree of the present invention;

[0081] Figure 5 The optimal industrial layout path map for AMORL optimization of the present invention;

[0082] Figure 6 is the normalized future trend prediction matrix of the present invention;

[0083] Figure 7 It is the self-learning module dynamic adjustment diagram of the present invention (time step 1);

[0084] Figure 8 It is the self-learning module dynamic adjustment diagram of the present invention (time step 20);

[0085] Fig. 9 It is a dynamic change diagram of multi-dimensional data of the industrial chain of the present invention;

[0086] Fig.10 It is a diagram showing the final optimized layout path of the present invention. DETAILED DESCRIPTION

[0087] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention. Specific embodiment one:

[0089] An industrial space intelligent association optimization system based on multidimensional data analysis, the association optimization system includes the following modules:

[0090] Data collection module, used to obtain multi-dimensional data related to industrial space, including industrial chain data, geographic space data, economic activity data, technology development trend data and policy and regulation data;

[0091] The data preprocessing module is equipped with a multi-dimensional data denoising processor, a data normalization processor, and a dynamic data structuring processor to clean and normalize data from different sources and formats and generate a standardized data input format;

[0092] The association analysis module uses a nonlinear time series deep embedding algorithm to extract implicit association features from multidimensional data, and realizes intelligent association between industrial chain nodes through multi-level spatial relationship modeling; the nonlinear time series deep embedding algorithm extracts time series association features from multidimensional data and performs multi-level relationship modeling. Time series feature extraction: feature extraction of input multi-dimensional time series data:

[0093] X(t)=[x1(t),x2(t),…,x n (t)]

[0094] X(t) is a multidimensional data vector at time t, representing industry data with n different dimensions; i (t): data of the i-th dimension at time t, i∈{1,2,…,n}

[0095] The embedding space is constructed using the embedding delay method:

[0096] Y(t)=[X(t),X(t-τ),X(t-2τ),…,X(t-mτ)]

[0097] Y(t): time series data after delayed embedding, containing multidimensional data of the past m time points; τ: delay time step, used to control the interval between historical time points; m: embedding dimension, indicating the number of steps of historical data;

[0098] The data is decomposed into time series features at different scales through wavelet transform:

[0099]

[0100] W(s,t): The result of the time series after wavelet transformation when the scale is s; X(t′): The original time series data at time point t′; ψ(t): Wavelet function, also called mother wavelet, used to decompose the signal; s: Scale parameter, indicating the time scale of the analysis (large s corresponds to a long time scale, small s corresponds to a short time scale); t: Current time; ψ * : The complex conjugate function of the mother wavelet; This method is used to capture the changing characteristics of different time scales; Deep embedding structure: Construct a nonlinear deep embedding neural network and model the time series data through a recurrent neural network (RNN):

[0101] h t =f(W·Y t +U·h t-1 +b)

[0102] h t : The hidden state vector at the current moment, containing the embedded time series features; W: Input weight matrix, used to process the input data Y t ; Yt : The multidimensional time series data input at the current moment. U: Recursive weight matrix, used to process the hidden state h at the previous moment t-1 ;h t-1 : The hidden state of the previous moment; b: Bias term, used to adjust the output of the network; f: Activation function;

[0103] The optimization decision module, based on the deep mining and dynamic analysis of multi-dimensional data, adopts an interactive optimization algorithm based on adaptive multi-objective reinforcement learning to make real-time optimization decisions for each related node to improve the efficiency of industrial resource allocation and the rationality of spatial layout; the interactive optimization algorithm of adaptive multi-objective reinforcement learning includes the following parts:

[0104] Value function:

[0105]

[0106] V(s): expected cumulative reward value in state s; Expected value operator, used to calculate the weighted average of all possible results; γ: discount factor, ranging from 0<γ≤1, used to control the importance of future rewards (the larger γ, the more important future rewards); R(s t ,a t ): At time t, state s t Take action a t The instant rewards received; t : state at time t; a t : action at time t;

[0107] Particle swarm optimization (PSO) speed update formula:

[0108]

[0109] The velocity of the i-th particle at time t+1; The speed of the i-th particle at time t; w: inertia factor, which controls the influence of the current speed of the particle on the next step; c1, c2: learning factors, which are used to control the particle to the individual optimal position p i and the global optimal position g i Approach speed; r1, r2: random numbers between 0 and 1; p i : The individual optimal position of the i-th particle; g i : The global optimal position of the group; The position of the i-th particle at time t;

[0110] Action strategy:

[0111]

[0112] P(a|s): the probability of choosing action a in state s; Q(s,a): the value function of taking action a in state s; a′: the action in the possible action set;

[0113] Fuzzy target optimization:

[0114]

[0115] μ(g i ) The fuzzy membership of target g_i; g i : the i-th target value; g i0 : Target g i The ideal threshold of ; d: controls the width of the fuzzy set, which is used to adjust the degree of fuzzification;

[0116] The prediction and feedback module is equipped with an adaptive Bayesian network dynamic causal reasoning model to predict the development trend of industrial space, iteratively optimize based on historical data and real-time data, and generate operational adjustment strategies and feedback; the specific contents of the adaptive Bayesian network dynamic causal reasoning model are as follows:

[0117] Bayesian network joint probability distribution:

[0118]

[0119] P(X1,X2,…,X n ): joint probability distribution of n variables in the Bayesian network; X i : the i-th variable; Pa(X i ): the parent node set of the i-th variable;

[0120] Maximum a posteriori estimation (MAP):

[0121]

[0122] P(X t ∣X t-1 ): Conditional probability distribution at time t given the data at time t-1; P(X t-1 |θ): Prior probability based on parameter θ; P(θ|X t-1 ): the posterior distribution of parameter θ; P(X t-1 ): marginal probability of the data at time t-1;

[0123] Modular logic reasoning formula:

[0124]

[0125] μ(y) is the fuzzy membership of the result y after reasoning; μ A (x), μB (y): fuzzy membership function of input x and output y; f(x): functional relationship defined by fuzzy rules;

[0126] The intelligent self-learning module can update association rules and optimization strategies in real time according to new data and market changes through a multi-level time-varying causal network learning algorithm, ensuring that the system has adaptive learning and adjustment capabilities; the time-varying causal network learning algorithm is as follows:

[0127] Causal network learning objective function:

[0128]

[0129] L: loss function of the causal network; P(X i ∣∣Pa(X i )): conditional probability of the i-th variable given the parent node set; λ: regularization parameter used to control the causal relationship variable; A ij : The adjacency matrix element of the causal network, indicating whether there is a causal relationship between node i and node j;

[0130] Model parameter update formula:

[0131]

[0132] θ t+1 : Model parameters at time t+1; θ t : Model parameters at time t; α: learning rate, controlling the step size of parameter update; The gradient of the loss function with respect to the parameters θ.

[0133] Data acquisition module: Sensor network: including environmental sensors, location sensors, economic activity monitoring sensors, etc. Satellite communication unit: used to receive satellite remote sensing data in real time. Network interface module: connected to various public databases and enterprise management systems, supporting multiple communication protocols such as Ethernet, 5G, WiFi, etc. Distributed computing node: performs preliminary data processing through edge computing devices to reduce the burden of data transmission. This module is responsible for collecting and aggregating data from multidimensional data sources, and realizes real-time connection with external data sources through sensors, satellite communication units, and network interfaces to ensure the timeliness and integrity of the data. Distributed computing nodes are used to process massive data and transmit the preprocessed data to the data preprocessing module.

[0134] Data preprocessing module: High-performance CPU / GPU processor: used for complex data processing tasks. Storage unit: equipped with large-capacity solid-state hard disk and RAM for fast data access and temporary storage. Data normalization hardware unit: used to standardize data from different sources to ensure consistency. Wavelet denoising hardware processor: specially used for multi-scale wavelet denoising processing, combined with FPGA to accelerate data processing. After the data enters the data preprocessing module, the wavelet denoising hardware processor is used to denoise the data from different sources, the GPU accelerates the normalization process, and the processed data is temporarily stored in the storage unit. The high-performance processor is responsible for real-time processing and scheduling of data streams to ensure data continuity.

[0135] Correlation analysis module: Deep learning accelerator: includes dedicated AI chips (such as TPU or FPGA) to accelerate the training and reasoning of recursive neural networks (RNN). Image processing unit: used to analyze satellite images or other spatial data for feature extraction. Time series data memory: cache module for storing and reading time series data. The system uses a deep learning accelerator and RNN to process multi-dimensional time series data and extract time series correlation features. The AI ​​chip improves the execution speed of the nonlinear time series deep embedding algorithm (NLTDE). The image processing unit is used to analyze geospatial data from satellites and integrate it with other data streams.

[0136] Optimization decision module: Embedded AI processor: supports the calculation of reinforcement learning algorithms, accelerates the calculation of value functions and strategy optimization. GPU cluster: handles multi-objective optimization tasks and supports large-scale parallel computing. Communication interface: used to communicate with external systems and supports the output and feedback of optimization solutions. The embedded AI processor is responsible for real-time calculation of the value function and combines the particle swarm optimization (PSO) algorithm to make optimization decisions. The GPU cluster supports parallel processing of multi-objective optimization tasks and speeds up decision-making. The optimization solution is sent to the execution layer of the industrial system through the communication interface to achieve dynamic adjustment of the industrial layout.

[0137] Prediction and feedback module: Bayesian reasoning chip: specially used for adaptive Bayesian network reasoning, supporting dynamic causal analysis. Data processing accelerator: combined with fuzzy logic operation unit, processing data in a high uncertainty environment. Storage device: local database or cloud storage for storing historical data and feedback data. Bayesian reasoning chip is used for dynamic causal reasoning, and real-time update of industry development trend prediction model. Fuzzy logic operation unit generates optimized feedback information through reasoning and feedback mechanism. Storage device records historical data and optimization feedback for future iterative analysis.

[0138] Intelligent self-learning module: Progressive deep learning processor: A deep learning processor based on an online learning algorithm that can quickly adjust model parameters. Large-capacity cache: used to store temporary update data of the model in the self-learning module. High-frequency communication module: supports frequent data interaction with the data acquisition and analysis module. The progressive deep learning processor uses the time-varying causal network learning algorithm (MTCNL) to update and optimize the model based on new data input in real time, ensuring that the model can respond quickly to market changes. The high-frequency communication module ensures effective data interaction with other modules, and the self-learning module quickly starts the adjustment process when new data is received.

[0139] The data acquisition module further includes a distributed data acquisition unit, which can collect data in real time from multiple heterogeneous data sources, including sensor networks, satellite remote sensing data, internal management systems of the industrial chain and public databases, and pre-process the collected data in real time through edge computing technology. Sensor cluster: sensors installed in the industrial area, including temperature, humidity, location and other sensors; edge computing devices: edge computing devices deployed on site to process sensor data; real-time data stream processor: realize real-time analysis and fusion of different types of data sources; distributed data acquisition units collect data in real time through sensor clusters and satellite remote sensing devices, and edge computing devices perform pre-processing locally to reduce data transmission delays; real-time data stream processors stream and preliminarily filter data from multiple heterogeneous data sources, and then transmit them to the core module of the system for further processing and analysis.

[0140] The multidimensional data denoising processor of the data preprocessing module adopts an adaptive multi-scale wavelet denoising algorithm, which can automatically select the best denoising parameters according to the noise characteristics of different types of data to improve the effectiveness and accuracy of the data. Wavelet denoising accelerator: a dedicated hardware processor based on FPGA, used to accelerate the wavelet denoising algorithm; adaptive parameter selection unit: dynamically adjusts the parameters in the denoising algorithm; high-performance computing node: parallel processing of large-scale data denoising tasks through GPU or FPGA; the wavelet denoising accelerator performs wavelet transforms of different scales on the data, and dynamically adjusts the parameters of the denoising algorithm in combination with the adaptive parameter selection unit to improve the accuracy and effectiveness of the data; the high-performance computing node accelerates the denoising processing of multidimensional data through parallel computing to ensure that the system can process large-scale data sets in real time.

[0141] The nonlinear time series deep embedding algorithm in the association analysis module is further combined with multi-scale and multi-level regression analysis, which can identify the interaction relationship between industrial chain nodes at different time scales while extracting time series association features. Time series data analysis hardware accelerator: FPGA / ASIC chip that supports multi-scale analysis, used to accelerate time series regression analysis; deep learning processor: dedicated AI processing chip for multi-scale feature extraction and regression analysis; hardware application logic: the time series data analysis hardware accelerator is responsible for multi-level regression analysis of large-scale industrial chain time series data, extracting features of different time scales, and analyzing nonlinear associations between nodes through deep learning processors; the system realizes comprehensive identification and modeling of complex associations by processing data at multiple time scales in parallel.

[0142] The optimization decision module adopts an adaptive multi-objective reinforcement learning algorithm based on hybrid particle swarm optimization and deep belief network to dynamically adjust the priority of optimization objectives and achieve rapid convergence under limited computing resources.

[0143] The adaptive Bayesian network dynamic causal reasoning model of the prediction and feedback module is further combined with the fuzzy logic reasoning system, which can improve the accuracy and stability of prediction in situations with high uncertainty.

[0144] The intelligent self-learning module adopts a progressive deep learning framework, combined with an online learning algorithm, and periodically updates the training model to ensure that the system remains highly responsive and adaptable in a rapidly changing industrial environment. Specific embodiment 2:

[0146] The working steps of the industrial space intelligent association optimization system based on multidimensional data analysis are as follows:

[0147] Sp1. Obtain multi-dimensional data related to the industry through the data acquisition module. The hardware facilities of this module include sensor networks, satellite remote sensing systems, industrial chain management systems, and public database connection units. These data sources provide the system with various types of data, including industrial chain data, geospatial data, economic activity data, technology development trend data, and policy and regulatory data. In order to ensure the real-time nature of data and the integration of multi-source heterogeneous data, the system deploys edge computing devices to reduce data transmission delays and complete preliminary processing locally.

[0148] After the data are collected, they are transferred to the data preprocessing module. This module uses a multidimensional data denoising processor to clean the data, remove noise and ensure the accuracy of the data. An adaptive multi-scale wavelet denoising algorithm is used to handle the noise characteristics of different types of data, and the processing efficiency and effectiveness are improved by dynamically adjusting the denoising parameters. The data is also standardized through normalization and structured processors to ensure that the data formats from different sources are consistent, which is convenient for subsequent analysis;

[0149] Sp2. The cleaned and standardized data are input into the association analysis module. The core of this module uses the nonlinear time series deep embedding algorithm (NLTDE), which can extract hidden time series correlation features from multidimensional data. NLTDE combines the time series analysis method based on density clustering, which can not only identify dynamic changes in the time dimension, but also find implicit correlations between nodes in different industries through clustering analysis. The system uses recursive neural network (RNN) modeling to process multi-level time series relationships between nodes in the industrial chain. In the time series, the system first uses the delayed embedding method to construct the historical data into a multidimensional data matrix through time series embedding, and then uses wavelet transform to decompose the data into time series features at different time scales. Through these methods, the system can identify the interactive relationships between industrial nodes at different time scales, so as to better understand the internal correlation and influence of the industrial chain;

[0150] Sp3. After the system extracts implicit correlation features through the correlation analysis module, the data enters the optimization decision module for further processing. This module adopts the adaptive multi-objective reinforcement learning algorithm (AMORL), which aims to achieve real-time optimization of industrial spatial layout and resource allocation through multi-objective optimization. AMORL combines fuzzy objective optimization technology to optimize objectives under different constraints by dynamically adjusting fuzzy sets. Specifically, fuzzy objective optimization technology adapts to complex multi-objective environments by fuzzifying different optimization objectives. Based on reinforcement learning, the AMORL algorithm dynamically calculates the optimal strategy in the current state to ensure that the system can strike a balance between multiple objectives. In this optimization process, the system uses the particle swarm optimization (PSO) algorithm to accelerate target search, using multiple particles to simultaneously explore the solution space and continuously approach the optimal solution. Through such optimization, the system can generate an operational optimization decision plan for industrial layout and output execution instructions in real time;

[0151] Sp4. After the optimization decision is completed, the prediction and feedback module will take over the next operation of the system. This module predicts the future development trend of the industrial space based on the adaptive Bayesian network dynamic causal reasoning model (ABNDCR). The Bayesian network constructs the causal relationship between each industrial node, and derives the mutual influence of each node in the industrial space through conditional probability distribution. In order to enhance the prediction ability in an uncertain environment, the system further combines the fuzzy logic reasoning system to fuzzy the causal relationship to ensure that accurate prediction results can be generated even in situations with high uncertainty. The ABNDCR model iteratively optimizes the relationship between nodes in the industrial chain through the combination of historical data and real-time data, and provides early warning of possible changes in the future. At the same time, the module can feed back the prediction results to the optimization decision module to adjust and re-optimize the strategy;

[0152] Sp5. After the prediction and feedback module completes its task, the system's intelligent self-learning module will start working. This module is based on the time-varying causal network learning algorithm (MTCNL), and continuously updates the association rules and optimization strategies through a self-learning mechanism. The MTCNL algorithm can update and adjust the model in real time according to market changes and new data, ensuring that the system is always adaptive and highly responsive. It uses a progressive deep learning framework to dynamically adjust model parameters when new data is received to improve the association prediction of industrial nodes. In order to achieve self-learning, the system uses an online learning algorithm to periodically update the training model to continuously improve the system's optimization capabilities and maintain adaptability to the ever-changing industrial environment.

[0153] The whole system works together through the above five main steps to form a closed-loop optimization process. In the first step (Sp1), the system obtains standardized input data through multi-dimensional data collection and cleaning; then in the second step (Sp2), the association analysis module extracts the deep association characteristics of the industrial chain through advanced time series feature extraction and cluster analysis; then, in the third step (Sp3), the system dynamically adjusts the optimization target and generates strategies based on the reinforcement learning algorithm; in the fourth step (Sp4), the prediction and feedback module predicts the trend of future industrial development based on Bayesian network and fuzzy logic reasoning, and provides real-time feedback to the system; finally, in the fifth step (Sp5), the self-learning module continuously optimizes the association rules through the time-varying causal network learning algorithm to ensure that the system can adaptively adjust the optimization strategy.

[0154] It should be noted that, in this article, relational terms such as first and second, etc. are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Moreover, the terms "include", "comprise" or any other variants thereof are intended to cover non-exclusive inclusion, so that a process, method, article or device including a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, method, article or device. In the absence of further restrictions, the elements defined by the statement "comprising a reference structure" do not exclude the existence of other identical elements in the process, method, article or device including the elements.

[0155] Although embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions and variations may be made to the embodiments without departing from the principles and spirit of the present invention, and that the scope of the present invention is defined by the appended claims and their equivalents.

Claims

1. An industrial space intelligent correlation optimization system based on multi-dimensional data analysis, characterized by: The association optimization system includes the following modules: Data collection module, used to obtain multi-dimensional data related to industrial space, including industrial chain data, geographic space data, economic activity data, technology development trend data and policy and regulation data; The data preprocessing module is equipped with a multi-dimensional data denoising processor, a data normalization processor, and a dynamic data structuring processor to clean and normalize data from different sources and formats and generate a standardized data input format; The association analysis module uses a nonlinear time series deep embedding algorithm to extract implicit association features from multidimensional data and realizes intelligent association between industrial chain nodes through multi-level spatial relationship modeling; The nonlinear time series deep embedding algorithm extracts time series correlation features from multidimensional data and performs multi-level relationship modeling. Time series feature extraction: feature extraction of input multidimensional time series data: X(t)=[x1(t),x2(t),...,x n (t)] X(t) is a multidimensional data vector at time t, representing industry data with n different dimensions; i (t): data of the i-th dimension at time t, i∈{1, 2, ..., n} The embedding space is constructed using the embedding delay method: Y(t)=[X(t),X(t-τ),X(t-2τ),...,X(t-mτ)] Y(t): time series data after delayed embedding, including multidimensional data of the past m time points; τ: delay time step, used to control the interval between historical time points; m: embedding dimension, indicating the number of steps of historical data; The data is decomposed into time series features at different scales through wavelet transform: W(s, t): the result of the time series after wavelet transformation when the scale is s; X(t′): the original time series data at time point t′; ψ(t): wavelet function, also called mother wavelet, used to decompose the signal; s: scale parameter, indicating the time scale of analysis (large s corresponds to long time scale, small s corresponds to short time scale); t: current time; ψ*: complex conjugate function of mother wavelet; this method is used to capture the changing characteristics of different time scales; deep embedding structure: construct a nonlinear deep embedding neural network, and model the time series data through a recurrent neural network (RNN): h t =f(W·Y t +U·h t-1 +b) h t : The hidden state vector at the current moment, containing the embedded time series features; W: Input weight matrix, used to process the input data Y t ; Y t : The multidimensional time series data input at the current moment. U: Recursive weight matrix, used to process the hidden state h at the previous moment t-1 ;h t-1 : The hidden state of the previous moment; b: Bias term, used to adjust the output of the network; f: Activation function; The optimization decision module, based on the deep mining and dynamic analysis of multi-dimensional data, adopts an interactive optimization algorithm based on adaptive multi-objective reinforcement learning to make real-time optimization decisions for each related node to improve the efficiency of industrial resource allocation and the rationality of spatial layout; the interactive optimization algorithm of adaptive multi-objective reinforcement learning includes the following parts: Value function: V(s): expected cumulative reward value in state s; Expected value operator, used to calculate the weighted average of all possible results; γ: discount factor, ranging from 0<γ≤1, used to control the importance of future rewards (the larger γ, the more important future rewards); R(s t , a t ): At time t, state s t Take action a t The instant rewards received; t : state at time t; a t : action at time t; Particle swarm optimization (PS0) speed update formula: The velocity of the i-th particle at time t+1; The speed of the i-th particle at time t; w: inertia factor, which controls the influence of the current speed of the particle on the next step; c1, c2: learning factors, which are used to control the particle to the individual optimal position p i and the global optimal position g i Approach speed; r1, r2: random numbers between 0 and 1; p i : The individual optimal position of the i-th particle; g i : The global optimal position of the group; The position of the i-th particle at time t; Action strategy: P(a|s): the probability of selecting action a in state s; Q(s, a): the value function of taking action a in state s; a′: the action in the possible action set; Fuzzy target optimization: μ(g i ) The fuzzy membership of target g_i; g i : the i-th target value; g i0 : Target g i The ideal threshold of ; d: controls the width of the fuzzy set, which is used to adjust the degree of fuzzification; The prediction and feedback module is equipped with an adaptive Bayesian network dynamic causal reasoning model to predict the development trend of industrial space, iteratively optimize based on historical data and real-time data, and generate operational adjustment strategies and feedback; the specific contents of the adaptive Bayesian network dynamic causal reasoning model are as follows: Bayesian network joint probability distribution: P(X1, X2, ..., X n ): joint probability distribution of n variables in the Bayesian network; X i : the i-th variable; Pa(X i ): the parent node set of the i-th variable; Maximum a posteriori estimation (MAP): P(X t |X t-1 ): Conditional probability distribution at time t given the data at time t-1; P(X t-1 |θ): Prior probability based on parameter θ; P(θ|X t-1 ): the posterior distribution of parameter θ; P(X t-1 ): marginal probability of the data at time t-1; Fuzzy logic reasoning formula: μ(y) is the fuzzy membership of the result y after reasoning; μ A (x), μ B (y): fuzzy membership function of input x and output y; f(x): functional relationship defined by fuzzy rules; The intelligent self-learning module can update association rules and optimization strategies in real time according to new data and market changes through a multi-level time-varying causal network learning algorithm, ensuring that the system has adaptive learning and adjustment capabilities; the time-varying causal network learning algorithm is as follows: Causal network learning objective function: L: loss function of the causal network; P(X i |Pa(X i )): conditional probability of the i-th variable given the set of parent nodes; λ: regularization parameter used to control the causal relationship variable; A ij : The adjacency matrix element of the causal network, indicating whether there is a causal relationship between node i and node j; Model parameter update formula: θ t+1 : Model parameters at time t+1; θ t : Model parameters at time t; α: learning rate, controlling the step size of parameter update; The gradient of the loss function with respect to the parameters θ.

2. The industrial space intelligent association optimization system based on multidimensional data analysis according to claim 1 is characterized by: The data acquisition module further includes a distributed data acquisition unit, which can collect data in real time from multiple heterogeneous data sources, including sensor networks, satellite remote sensing data, internal management systems of the industrial chain and public databases, and pre-process the collected data in real time through edge computing technology.

3. The industrial space intelligent association optimization system based on multidimensional data analysis according to claim 1 is characterized in that: The multidimensional data denoising processor of the data preprocessing module adopts an adaptive multi-scale wavelet denoising algorithm, which can automatically select the best denoising parameters according to the noise characteristics of different types of data to improve the effectiveness and accuracy of the data.

4. The industrial space intelligent association optimization system based on multidimensional data analysis according to claim 1 is characterized by: The nonlinear time series deep embedding algorithm in the association analysis module is further combined with multi-scale and multi-level regression analysis, which can identify the interactive relationship between industrial chain nodes at different time scales while extracting time series association features.

5. The industrial space intelligent association optimization system based on multidimensional data analysis according to claim 1 is characterized by: The optimization decision module adopts an adaptive multi-objective reinforcement learning algorithm based on hybrid particle swarm optimization and deep belief network to dynamically adjust the priority of optimization objectives and achieve rapid convergence under limited computing resources.

6. The industrial space intelligent association optimization system based on multidimensional data analysis according to claim 1 is characterized by: The adaptive Bayesian network dynamic causal reasoning model of the prediction and feedback module is further combined with a fuzzy logic reasoning system, which can improve the accuracy and stability of predictions in situations with high uncertainty.

7. The industrial space intelligent association optimization system based on multidimensional data analysis according to claim 1 is characterized by: The intelligent self-learning module adopts a progressive deep learning framework, combined with an online learning algorithm, and periodically updates the training model to ensure that the system remains highly responsive and adaptable in a rapidly changing industrial environment.

8. The industrial space intelligent association optimization system based on multidimensional data analysis according to any one of claims 1 to 7, characterized in that: The working steps of the association optimization system are as follows: Sp1. Obtain multi-dimensional industry-related data through the data acquisition module and input it into the data preprocessing module for data cleaning and structuring; Sp2, the data enters the association analysis module, and the nonlinear time series deep embedding algorithm (NLTDE) extracts hidden association features from it, and multi-level relationship modeling is performed based on the spatial and temporal dimensions of each industry node; Sp3, the optimization decision module adopts the adaptive multi-objective reinforcement learning (AMORL) algorithm to optimize the industrial layout under the constraints of multiple objectives and output an executable optimization plan; Sp4, prediction and feedback module predicts development trends and provides feedback optimization strategies based on the adaptive Bayesian network dynamic causal reasoning model (ABNDCR); Sp5, the intelligent self-learning module uses the time-varying causal network learning algorithm (MTCNL) to adaptively learn and update association rules and models.

9. The working steps of the industrial space intelligent association optimization system based on multidimensional data analysis according to claim 7 are characterized by: The nonlinear time series deep embedding algorithm (NLTDE) is combined with a time series analysis method based on density clustering.

10. The working steps of the industrial space intelligent association optimization system based on multidimensional data analysis according to claim 7 are characterized by: The adaptive multi-objective reinforcement learning (AMORL) algorithm combines fuzzy objective optimization technology by dynamically adjusting the fuzzy set of constraints.

Images