Intelligent enterprise operation decision optimization method and system based on neural network

Abstract

The invention relates to the technical field of neural networks, in particular to an enterprise operation decision intelligent optimization method and system based on a neural network. The method comprises the following steps: obtaining enterprise operation data corresponding to each operation link in an enterprise and enterprise external data, carrying out box washing and normalization processing, constructing each sub-neural network module, and carrying out time sequence state machine training output by adopting a long-short term memory network structure to obtain an enterprise operation time sequence state machine vector; constructing a main neural network module, and performing model decision prediction and prediction deviation accounting on the vector of the enterprise operation time sequence state machine by using a convolutional neural network structure to obtain the deviation between the actual operation of the enterprise and the model prediction; and performing dynamic optimization adjustment based on the deviation between the actual operation of the enterprise and the model prediction to generate a multi-module neural network optimization model, and outputting a corresponding enterprise operation optimization decision result. The method can help an enterprise make a more efficient decision in a complex and changeable market scene.

Description

Technical Field

[0001] The present invention relates to the field of neural network technology, and in particular to a method and system for intelligent optimization of enterprise operation decision-making based on neural network. Background Art

[0002] In recent years, the rapid development of artificial intelligence technology, especially neural networks, has provided new solutions for enterprise operational decision-making. Neural networks, as a computing model that imitates the biological nervous system, have strong learning and adaptive capabilities, can process complex, high-dimensional data, and capture nonlinear relationships in the system. By training neural networks, enterprises can conduct in-depth analysis of historical data, identify potential rules and trends in operations, and thus provide more accurate and dynamic support for enterprise decision-making. However, existing neural network optimization methods often rely on a large amount of historical data and efficient computing power. However, in actual applications, due to the diversity and complexity of enterprise operational data, traditional neural network models have the problems of long training time and a lot of noise data, which leads to errors in internal operational decisions of enterprises. Summary of the invention

[0003] Based on this, it is necessary for the present invention to provide a method and system for intelligent optimization of enterprise operation decisions based on neural networks to solve at least one of the above technical problems.

[0004] To achieve the above purpose, a neural network-based enterprise operation decision intelligent optimization method includes the following steps: Step S1: Obtain enterprise operation data corresponding to each operation link within the enterprise and enterprise external data, and perform box cleaning and normalization processing on the enterprise operation data corresponding to each operation link and enterprise external data to obtain operation normalized data corresponding to each operation link and external normalized data; Step S2: construct each sub-neural network module, and use the long short-term memory network structure according to each sub-neural network module to perform time series state machine training and output on the operation normalized data corresponding to each operation link and the external normalized data, and obtain the enterprise operation time series state machine vector corresponding to each sub-neural network module; Step S3: construct a main neural network module, and based on the main neural network module, use the convolutional neural network structure to perform model decision prediction on the enterprise operation time series state machine vector corresponding to each sub-neural network module, generate a multi-module neural network model, and output the corresponding enterprise operation status decision recommendation results; Step S4: Obtain the actual operation decision results of the enterprise, and calculate the prediction deviation based on the actual operation decision results of the enterprise and the decision recommendation results of the enterprise's operating status to obtain the deviation between the actual operation of the enterprise and the model prediction; based on the deviation between the actual operation of the enterprise and the model prediction, dynamically optimize and adjust the multi-module neural network model to generate a multi-module neural network optimization model, and output the corresponding enterprise operation optimization decision results.

[0005] Furthermore, the various operation links described in step S1 include production, sales and human resources operation links.

[0006] Further, step S1 includes the following steps: Obtain enterprise operation data corresponding to each operation link within the enterprise and external data of the enterprise, where the enterprise operation data corresponding to each operation link includes production plan, output and equipment operation status data corresponding to the production operation link, sales volume, sales and customer sales feedback data corresponding to the sales operation link, employee attendance, performance and training status data corresponding to the human resources operation link, and external data of the enterprise includes industry dynamics and competition information data; Perform neighborhood noise identification and labeling on the enterprise operation data corresponding to each operation link and the enterprise external data, so as to obtain the operation noise labeling data and external noise labeling data corresponding to each operation link; The noise anomaly of the box plot is cleaned for the noise points and non-noise points in the operation noise marked data corresponding to each operation link and the external noise marked data, respectively, to obtain the operation cleaned data and external cleaned data corresponding to each operation link; Obtain the enterprise operation mean and the enterprise external mean corresponding to each operation link, and perform mean missing filling processing on the operation cleaning data and external cleaning data corresponding to each operation link based on the enterprise operation mean and the enterprise external mean corresponding to each operation link, so as to obtain the operation filling data and external filling data corresponding to each operation link; The min-max normalization method is used to normalize the operation filling data and external filling data corresponding to each operation link, so as to obtain the operation normalized data and external normalized data corresponding to each operation link.

[0007] Furthermore, the process of identifying and marking the neighborhood noise of the enterprise operation data and the enterprise external data corresponding to each operation link includes the following steps: The enterprise operation data corresponding to each operation link and the operation data or external data corresponding to each time node in the enterprise external data are respectively taken as a data node to obtain the operation data point set or external data point set corresponding to each operation link; Perform dimensional statistics analysis on the operational data point set or external data point set corresponding to each operational link to obtain operational dimensional statistics or external dimensional statistics corresponding to each operational link, wherein the corresponding dimensional statistics include mean, median and mode; Based on the operation dimension statistics or external dimension statistics corresponding to each operation link, a data density statistical analysis is performed on the corresponding operation data point set or external data point set to obtain the operation dimension data density or external dimension data density corresponding to each operation link; Based on the operational dimension data density or external dimension data density corresponding to each operational link, the noise neighborhood radius calculation formula is used to calculate the noise neighborhood radius of the operational data point set or external data point set corresponding to each operational link, so as to obtain the enterprise data noise neighborhood radius corresponding to each operational link; Based on the enterprise data noise neighborhood radius corresponding to each operation link, the corresponding operation data point set or external data point set is marked for neighborhood noise identification, and the neighborhood circle is drawn using the mean under the corresponding dimension as the center point combined with the enterprise data noise neighborhood radius. If a data point is within or on the neighborhood circle, it is regarded as a normal point; if a data point is not within the neighborhood circle, it is regarded as a noise point and marked, so as to obtain the operation noise marked data and external noise marked data corresponding to each operation link.

[0008] Furthermore, the noise neighborhood radius calculation formula is specifically as follows: ; In the formula, is the noise neighborhood radius of enterprise data, is the data space area range, is the total number of data points corresponding to the operational data point set or the external data point set, is an exponential function, The first data points, The first data points, σ ​​is the Gaussian kernel standard deviation, For the The noise weight coefficient corresponding to the data point is For operational dimension data density or external dimension data density, is the neighborhood radius distance contribution coefficient, is the maximum possible distance between data points, and η is the correction coefficient of the noise neighborhood radius of enterprise data.

[0009] Further, step S2 includes the following steps: Step S21: construct each sub-neural network module, including an input layer, a hidden layer and an output layer; Step S22: performing time series synchronization processing on the normalized operation data and external normalized data corresponding to each operation link, and obtaining the time series of operation data and external data corresponding to each operation link in the same time series range; Step S23: According to each sub-neural network module, the long short-term memory network structure is used to perform timing state machine training and output on the operation data time series and external data time series corresponding to each operation link, so as to obtain the enterprise operation timing state machine vector corresponding to each sub-neural network module.

[0010] Further, step S23 includes the following steps: Step S231: According to each sub-neural network module, a long short-term memory network structure is used to perform time series state network module training on the operation data time series and the external data time series corresponding to each operation link, so as to use the input layer to receive the operation data time series and the external data time series corresponding to each operation link, and the hidden layer includes a number of LSTM network units to capture the enterprise operation changes corresponding to the short-term window of external environment changes and the long-term trend of enterprise operation corresponding to the long-term window of external environment changes, including an upward trend, a downward trend and a stable trend, and calculate the corresponding autocorrelation coefficient and partial correlation coefficient under each time window to reconstruct the data features of the input layer, so as to reconstruct and output the operation reconstruction sequence and the external reconstruction sequence corresponding to each operation link; Step S232: Input the operation reconstruction sequence and external reconstruction sequence corresponding to each operation link into the corresponding output layer in the sub-neural network module for forward propagation state machine calculation, and use unique hot encoding to convert the state machine prediction result corresponding to the output layer into vector form, so as to obtain the enterprise operation timing state machine vector corresponding to each sub-neural network module, so as to reflect the corresponding change characteristics of the enterprise operation under different external environment timing states.

[0011] Further, step S3 includes the following steps: Step S31: construct a main neural network module, which includes an input layer, a convolutional layer, a pooling layer, a fully connected layer and an output layer; Step S32: Based on the main neural network module, a convolutional neural network structure is used to construct a model for the enterprise operation time series state machine vector corresponding to each sub-neural network module, so as to use the input layer to receive the enterprise operation time series state machine vector corresponding to each sub-neural network module, the convolution layer extracts the time series correlation features between each operation link through a 3x3 convolution kernel, the pooling layer performs dimensionality reduction processing on the time series correlation features between each operation link, the fully connected layer maps the reduced dimensionality features to the decision suggestion space, and the output layer gives the decision suggestion output corresponding to the enterprise operation status, thereby generating a multi-module neural network model; Step S33: Use the multi-module neural network model to perform model training, decision prediction and output on the enterprise operation timing state machine vector corresponding to each sub-neural network module, so as to divide the enterprise operation timing state machine vector corresponding to each sub-neural network module into training set, verification set and test set according to the ratio of 70%, 15% and 15%, and train the training set through the Adam algorithm, set the learning rate to 0.001, the number of training rounds to 100, and evaluate the multi-module neural network model using the verification set after every 10 training rounds. Adjust the model parameters according to the evaluation results to prevent overfitting, and perform performance evaluation on the output layer corresponding to the multi-module neural network model by using the test set to calculate the corresponding accuracy and recall rate to determine the model evaluation results, and output the corresponding enterprise operation status decision recommendation results.

[0012] Further, step S4 includes the following steps: Step S41: Regularly obtain the actual operational decision results of the enterprise; Step S42: performing forecast deviation calculation based on the actual operation decision results of the enterprise and the decision suggestion results of the enterprise operation status, so as to obtain the deviation between the actual operation of the enterprise and the model prediction; Step S43: Compare and judge the deviation between the actual operation of the enterprise and the model prediction according to the preset model deviation threshold. If the deviation between the actual operation of the enterprise and the model prediction is less than the preset model deviation threshold, use the new actual operation data of the enterprise to fine-tune the multi-module neural network model to update the corresponding model hyperparameters; if the deviation between the actual operation of the enterprise and the model prediction is greater than or equal to the preset model deviation threshold, re-collect and process the corresponding enterprise operation data, and retrain the multi-module neural network model to generate a multi-module neural network optimization model, and output the corresponding enterprise operation optimization decision results.

[0013] Furthermore, the present invention also provides a neural network-based enterprise operation decision intelligent optimization system, which is used to execute the neural network-based enterprise operation decision intelligent optimization method as described above. The neural network-based enterprise operation decision intelligent optimization system includes: The operation link data preprocessing module is used to obtain the enterprise operation data corresponding to each operation link within the enterprise and the enterprise external data, and perform box cleaning and normalization processing on the enterprise operation data corresponding to each operation link and the enterprise external data, so as to obtain the operation normalized data corresponding to each operation link and the external normalized data; The submodule state machine training output module is used to construct each sub-neural network module, and according to each sub-neural network module, the long short-term memory network structure is used to perform time-series state machine training and output on the operation normalized data corresponding to each operation link and the external normalized data, so as to obtain the enterprise operation time-series state machine vector corresponding to each sub-neural network module; The network model decision prediction module is used to construct a main neural network module, and based on the main neural network module, use the convolutional neural network structure to perform model decision prediction on the enterprise operation time series state machine vector corresponding to each sub-neural network module, generate a multi-module neural network model, and output the corresponding enterprise operation status decision recommendation results; The model dynamic optimization and adjustment module is used to obtain the actual operation decision results of the enterprise, and to calculate the prediction deviation based on the actual operation decision results of the enterprise and the decision recommendation results of the enterprise's operating status, so as to obtain the deviation between the actual operation of the enterprise and the model prediction; based on the deviation between the actual operation of the enterprise and the model prediction, the multi-module neural network model is dynamically optimized and adjusted to generate a multi-module neural network optimization model, and the corresponding enterprise operation optimization decision results are output.

[0014] Beneficial effects of the present invention: 1. The intelligent optimization method for enterprise operation decision-making based on neural network proposed in the present invention, compared with the prior art, has the beneficial effect of collecting internal data including data of various operation links such as production, sales and human resources by obtaining various operation links and external data within the enterprise. These data can reflect the operation efficiency and status of various departments of the enterprise. At the same time, external data such as market conditions, competitor data, etc. can also affect the operation decision of the enterprise. The introduction of these external factors helps to ensure the comprehensiveness and accuracy of the model, and can help enterprises identify the impact of changes in the external environment on their own operations. However, these data often have outliers and noise, and direct application will affect the accuracy of the analysis results. Therefore, by setting certain upper and lower limits, the extreme noise point values ​​in the data are filtered to ensure the stability of the data. Then, the normalization process can convert data of different dimensions into comparable numerical intervals, eliminate the impact of dimensional differences between different features on model training, and ensure the fairness and effectiveness of model training. The data normalization process not only improves the efficiency of subsequent model training, but also accelerates the convergence speed and improves the accuracy of the final prediction, thereby providing a good input for the neural network model training in the subsequent steps. Secondly, by constructing a sub-neural network module, the long short-term memory (LSTM) network is used for time series modeling. The LSTM neural network structure has significant advantages and can process data with time series relationships, especially in tasks with long-term dependencies. In the context of enterprise operations, by using normalized operational data and external data as input, the sub-neural network module can learn the time series characteristics of each operational link of the enterprise and predict future states. For example, the changing trend of sales data and the dynamic changes in inventory can be obtained through the time series learning of the LSTM network. At this time, the LSTM model plays the role of a "state machine", analyzes historical data, and predicts possible future state changes, thereby providing enterprises with more accurate operational decision-making suggestions. This state machine training based on historical data can improve the accuracy and timeliness of the decision-making process and reduce operational uncertainty. Then, the main neural network is used to conduct a comprehensive analysis of the outputs of each sub-neural network module to achieve the final operational decision prediction. The main neural network module adopts a convolutional neural network (CNN) structure. CNN can effectively capture local features in the input data and combine them. It is suitable for processing and integrating time-series state machine vectors from multiple sub-neural networks. The convolution operation of CNN can extract and fuse features of input data at multiple levels and improve the accuracy of the final decision prediction, so as to help enterprises make accurate decision predictions from multiple dimensions. In this way, the interaction between different modules can be taken into account to obtain a more accurate enterprise operation status decision recommendation, thereby improving the overall operational efficiency and benefits.Finally, by comparing with the actual operating results, the forecast deviation is calculated. The deviation between the accuracy of the forecast model and the actual operation reflects the applicability and effectiveness of the model in the actual environment. By comparing the actual operating results of the enterprise and the model prediction results, it is possible to analyze which parts have larger forecast deviations, thereby finding the potential deficiencies of the model. In this way, the actual performance of the model can be continuously monitored and feedback can be provided, and the model can be dynamically optimized and adjusted to ensure that it maintains efficient forecasting capabilities in a changing operating environment, thereby making targeted adjustments and improvements to the model. This dynamic optimization mechanism can greatly enhance the long-term stability and adaptability of the model, enabling it to continuously improve the decision-making level of the enterprise. This continuously optimized feedback mechanism enables the enterprise to always maintain keen decision-making capabilities in a highly competitive market environment, thereby avoiding errors in the operation of the enterprise.

[0015] 2. The enterprise operation decision-making intelligent optimization system based on neural network proposed in the present invention is composed of an operation link data preprocessing module, a sub-module state machine training output module, a network model decision prediction module and a model dynamic optimization adjustment module as a whole. It can realize any enterprise operation decision-making intelligent optimization method based on neural network described in the present invention, and is used to combine the operations between computer programs running on various modules to realize the enterprise operation decision-making intelligent optimization method based on neural network. The internal structure of the system cooperates with each other, which can greatly reduce duplication of work and manpower investment, and can quickly and effectively provide a more accurate and efficient enterprise operation decision-making intelligent optimization process based on neural network, thereby simplifying the operation process of the enterprise operation decision-making intelligent optimization system based on neural network. BRIEF DESCRIPTION OF THE DRAWINGS

[0016] Other features, objects and advantages of the present invention will become more apparent from the detailed description of non-limiting embodiments thereof made with reference to the following drawings: Figure 1 A schematic diagram of the steps of the enterprise operation decision-making intelligent optimization method based on a neural network of the present invention; Figure 2 for Figure 1 Detailed step flow chart of step S2 in FIG. DETAILED DESCRIPTION

[0017] The technical method of the present invention is described clearly and completely below in conjunction with the accompanying drawings. Obviously, the described embodiments are 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 technicians in this field without creative work are within the scope of protection of the present invention.

[0018] In addition, the accompanying drawings are only schematic illustrations of the present invention and are not necessarily drawn to scale. The same reference numerals in the figures represent the same or similar parts, and their repeated description will be omitted. Some of the block diagrams shown in the accompanying drawings are functional entities and do not necessarily correspond to physically or logically independent entities. The functional entities can be implemented in software form, or implemented in one or more hardware modules or integrated circuits, or implemented in different networks and / or processor methods and / or microcontroller methods.

[0019] It should be understood that, although the terms "first", "second", etc. may be used herein to describe various units, these units should not be limited by these terms. These terms are used only to distinguish one unit from another unit. For example, without departing from the scope of the exemplary embodiments, the first unit may be referred to as the second unit, and similarly the second unit may be referred to as the first unit. The term "and / or" used herein includes any and all combinations of one or more of the listed associated items.

[0020] To achieve this, please refer to Figure 1 to Figure 2 The present invention provides a method for intelligent optimization of enterprise operation decision-making based on neural network, the method comprising the following steps: Step S1: Obtain enterprise operation data corresponding to each operation link within the enterprise and enterprise external data, and perform box cleaning and normalization processing on the enterprise operation data corresponding to each operation link and enterprise external data to obtain operation normalized data corresponding to each operation link and external normalized data; Step S2: construct each sub-neural network module, and use the long short-term memory network structure according to each sub-neural network module to perform time series state machine training and output on the operation normalized data corresponding to each operation link and the external normalized data, and obtain the enterprise operation time series state machine vector corresponding to each sub-neural network module; Step S3: construct a main neural network module, and based on the main neural network module, use the convolutional neural network structure to perform model decision prediction on the enterprise operation time series state machine vector corresponding to each sub-neural network module, generate a multi-module neural network model, and output the corresponding enterprise operation status decision recommendation results; Step S4: Obtain the actual operation decision results of the enterprise, and calculate the prediction deviation based on the actual operation decision results of the enterprise and the decision recommendation results of the enterprise's operating status to obtain the deviation between the actual operation of the enterprise and the model prediction; based on the deviation between the actual operation of the enterprise and the model prediction, dynamically optimize and adjust the multi-module neural network model to generate a multi-module neural network optimization model, and output the corresponding enterprise operation optimization decision results.

[0021] In the embodiment of the present invention, please refer to Figure 1 FIG. 1 is a schematic diagram of the steps of the enterprise operation decision intelligent optimization method based on a neural network of the present invention. In this example, the enterprise operation decision intelligent optimization method based on a neural network includes the following steps: Step S1: Obtain enterprise operation data corresponding to each operation link within the enterprise and enterprise external data, and perform box cleaning and normalization processing on the enterprise operation data corresponding to each operation link and enterprise external data to obtain operation normalized data corresponding to each operation link and external normalized data; In an embodiment of the present invention, relevant operational data, such as production data, sales data, inventory data, financial data, etc., are obtained from various operational links within the enterprise. In addition, various types of data from the outside, such as market trends, macroeconomic data, industry dynamics, etc., are collected. These data will be preprocessed separately, including detection and removal of outliers (i.e., the "box washing" process), and data normalization. Data normalization uses standardized methods (such as z-score normalization) or Min-Max normalization to convert operational data of different dimensions and ranges into a unified numerical range (such as [0,1]) to ensure that the data is subsequently analyzed on the same scale. The normalized data will be used to train the neural network model and flow into the subsequent sub-neural network module as input data, and finally the operational normalized data corresponding to each operational link and the external normalized data are obtained.

[0022] Step S2: construct each sub-neural network module, and use the long short-term memory network structure according to each sub-neural network module to perform time series state machine training and output on the operation normalized data corresponding to each operation link and the external normalized data, and obtain the enterprise operation time series state machine vector corresponding to each sub-neural network module; In an embodiment of the present invention, an independent sub-neural network module is constructed for each operation link, and each sub-network is based on a long short-term memory network (LSTM) structure. LSTM can effectively capture the long-term dependency of time series data. In each sub-neural network module, the input data is the normalized time series data corresponding to each operation link, including internal enterprise operation data and external influencing factor data. Through LSTM model training, each sub-neural network will learn the dynamic change rules of different factors in the time series and their impact on enterprise operations, and generate corresponding time series state machine vectors. The time series state machine vectors contain not only time information, but also the operation status characteristics of each link. These vectors will be used as inputs for subsequent decision predictions, and finally the enterprise operation time series state machine vectors corresponding to each sub-neural network module are obtained.

[0023] Step S3: construct a main neural network module, and based on the main neural network module, use the convolutional neural network structure to perform model decision prediction on the enterprise operation time series state machine vector corresponding to each sub-neural network module, generate a multi-module neural network model, and output the corresponding enterprise operation status decision recommendation results; In an embodiment of the present invention, a main neural network module is constructed as the decision layer of all sub-neural network modules. The main neural network module adopts a convolutional neural network (CNN) structure so that the time series state machine vectors output by each sub-network are processed through a convolution layer to extract deeper features and potential laws. The CNN structure can extract local features in the data through convolution operations and reduce computational complexity through pooling operations. The time series state machine vectors output by each sub-network will be used as the input of CNN. After calculations of several convolution layers and pooling layers, a decision output layer is formed to predict the overall status of the enterprise operation. The result generated by the model is the decision-making recommendation for the enterprise operation status, which helps the management optimize decisions and improve operational efficiency. Finally, a multi-module neural network model is generated and the corresponding decision-making recommendation results for the enterprise operation status are output.

[0024] Step S4: Obtain the actual operation decision results of the enterprise, and calculate the prediction deviation based on the actual operation decision results of the enterprise and the decision recommendation results of the enterprise's operating status to obtain the deviation between the actual operation of the enterprise and the model prediction; based on the deviation between the actual operation of the enterprise and the model prediction, dynamically optimize and adjust the multi-module neural network model to generate a multi-module neural network optimization model, and output the corresponding enterprise operation optimization decision results.

[0025] In an embodiment of the present invention, after implementing the enterprise operation decision, the actual operation decision results are collected and compared with the model prediction results, and the deviation between the two is calculated. This step can quantify the accuracy of the model prediction and the possible deviation by comparing the error between the prediction result and the actual situation, thereby obtaining the deviation between the actual operation of the enterprise and the model prediction. Based on the deviation, the model is further optimized. The optimization process uses the backpropagation algorithm (Backpropagation) to minimize the prediction deviation by adjusting the weights and biases in the neural network. Through continuous iterative optimization, the adjusted multi-module neural network model will be more in line with the actual operation needs of the enterprise, thereby providing more accurate operation optimization decisions. The decision suggestions output by the optimized model are more feasible, can respond to changes in the enterprise's operating status in real time, and self-adjust in a dynamic environment. Finally, the multi-module neural network optimization model is adjusted to generate the corresponding enterprise operation optimization decision results.

[0026] Furthermore, the various operation links described in step S1 include production, sales and human resources operation links.

[0027] Further, step S1 includes the following steps: Obtain enterprise operation data corresponding to each operation link within the enterprise and external data of the enterprise, where the enterprise operation data corresponding to each operation link includes production plan, output and equipment operation status data corresponding to the production operation link, sales volume, sales and customer sales feedback data corresponding to the sales operation link, employee attendance, performance and training status data corresponding to the human resources operation link, and external data of the enterprise includes industry dynamics and competition information data; In the embodiment of the present invention, by collecting data from various operation links of the enterprise (including production, sales and human resource operation links), the relevant data of the production operation link includes production plan, output data and equipment operation status information, the production plan data can be extracted through the enterprise's internal production management system (such as ERP system), the output data is obtained through sensors connected to the production equipment, and the equipment operation status data is obtained through the equipment monitoring system (such as SCADA). The relevant data of the sales operation link includes sales volume, sales revenue and customer sales feedback data, the sales volume and sales revenue can be automatically counted through the enterprise's sales management system (such as CRM system), and the customer sales feedback data can be collected through the customer relationship management platform or survey feedback system, the relevant data of the human resource operation link includes employee attendance, performance and training status data, these data can be extracted through the human resource management system (such as HRM system), the attendance data and performance data are obtained through attendance punching equipment and performance evaluation tools, and the training status is recorded through the enterprise training management system, and the external data of the enterprise includes industry dynamics and competitive information data, which are extracted through industry reports, market analysis platforms, and channels disclosed by competitors, and finally the enterprise operation data corresponding to each operation link within the enterprise and the external data of the enterprise are obtained.

[0028] Preferably, the enterprise operation data corresponding to each operation link and the enterprise external data are marked for neighborhood noise identification to obtain operation noise marking data and external noise marking data corresponding to each operation link; In an embodiment of the present invention, a distance-based noise detection algorithm is used to identify noise in the internal operation data and external data of an enterprise. In this step, the enterprise operation data and the external data set are first standardized and preprocessed to ensure that the data dimensions are consistent. For each data point, the algorithm calculates the similarity between it and the adjacent data points, and marks the noise data points by setting a distance radius threshold. Noise data usually appears as an outlier that is significantly different from other surrounding data points. In this process, based on statistical principles, a neighborhood density comparison method is used to identify potential noise, thereby assigning a noise label to each data point. The noise in the production link, sales link, human resources link and external data are identified separately by this method, and finally the operation noise label data and external noise label data corresponding to each operation link are obtained.

[0029] Preferably, the operation noise mark data corresponding to each operation link and the noise points and non-noise points in the external noise mark data are respectively cleaned by box plot noise anomalies to obtain the operation cleaned data and external cleaned data corresponding to each operation link; In an embodiment of the present invention, a box plot method is used to clean data points marked as noise. For each operation link (such as production, sales, human resources) and external data, a corresponding box plot is first constructed according to the distribution of its data. The box plot identifies outliers by calculating the upper quartile (Q3), lower quartile (Q1) and interquartile range (IQR) of the data. Any data point that exceeds the upper limit (Q3 + 1.5IQR) or the lower limit (Q1 - 1.5IQR) is considered to be an outlier (i.e., a noise point). According to this standard, the operating system automatically filters out these noise points and removes them, retaining only normal data. The cleaned data can provide more accurate basic data for subsequent analysis and decision-making, and finally obtains the operation cleaned data corresponding to each operation link and the external cleaned data.

[0030] Preferably, the enterprise operation mean and the enterprise external mean corresponding to each operation link are obtained, and the operation cleaning data and the external cleaning data corresponding to each operation link are filled with the mean missing values ​​based on the enterprise operation mean and the enterprise external mean corresponding to each operation link, so as to obtain the operation filling data and the external filling data corresponding to each operation link; In an embodiment of the present invention, the data mean of each operation link is first calculated for the cleaned data. The mean is calculated by summing the data of each dimension and then dividing it by the total number of data points. The mean of each operation link (such as production, sales, human resources, etc.) and external data is calculated by this method. For missing values ​​in the cleaned data, mean filling technology is used for processing. The specific operation is to fill the missing values ​​with the mean of the link or external data where the data is missing, so as to ensure data integrity. Missing value filling ensures the continuity of the data set and avoids the adverse impact of missing data on model training. The data processing method used in this step ensures that the cleaned data can seamlessly enter the subsequent analysis and decision-making model, and finally obtains the operation filling data and external filling data corresponding to each operation link.

[0031] Preferably, the operation filling data and external filling data corresponding to each operation link are normalized using the min-max normalization method to obtain the operation normalized data and external normalized data corresponding to each operation link.

[0032] In the embodiment of the present invention, all operational filling data and external filling data need to be normalized to eliminate the influence of different data dimensions, and the Min-Max normalization method is used for processing. First, the minimum value (Min) and the maximum value (Max) of each dimension of data are calculated, and then the data is normalized according to the formula: ;in, is the original data point, is the minimum value in the data, is the maximum value in the data, The normalized data is the data after normalization. Through this method, all operational data (such as production, sales, and human resources data) and external data are converted into a unified range, usually between [0, 1]. The normalized data can be better input into the subsequent neural network model for intelligent optimization analysis, because this method effectively eliminates the dimensional differences between different data dimensions, improves the training efficiency and accuracy of the model, and finally obtains the operational normalized data corresponding to each operational link and the external normalized data.

[0033] Furthermore, the process of identifying and marking the neighborhood noise of the enterprise operation data and the enterprise external data corresponding to each operation link includes the following steps: The enterprise operation data corresponding to each operation link and the operation data or external data corresponding to each time node in the enterprise external data are respectively taken as a data node to obtain the operation data point set or external data point set corresponding to each operation link; In an embodiment of the present invention, by collecting data from various internal operating links of the enterprise (such as production, sales, human resources, etc.) and external environmental data (such as market dynamics, competitor information, etc.), the operating data or external data of each operating link at each time node is obtained. Each data point represents the specific value of a certain operating link or external data in a specific time period. For enterprise operating data, such as production quantity in the production link, sales volume in the sales link, inventory level in the inventory link, etc., it can be collected and recorded in real time through the enterprise's internal information management system (such as ERP system), while external data can be obtained through a public data interface to capture market prices, macroeconomic indicators, competitor data, etc. It is necessary to uniquely identify each data node and finally obtain a set of operating data points or external data points corresponding to each operating link.

[0034] Preferably, a dimensional statistics analysis is performed on the operation data point set or the external data point set corresponding to each operation link to obtain the operation dimensional statistics or the external dimensional statistics corresponding to each operation link, wherein the corresponding dimensional statistics include the mean, the median and the mode; In an embodiment of the present invention, by performing dimensional statistics analysis on a set of enterprise operation data points or an external data point set, dimensional statistics corresponding to each operation link and external data are obtained. Common statistics include mean, median, and mode, etc. The purpose is to reveal the typical values ​​and distribution characteristics in the data set. First, the NumPy library in Python or the built-in functions of the R language are used to calculate the mean, median, and mode of the data set. For example, for data in the production link, numpy.mean() is used to calculate the mean, numpy.median() is used to calculate the median, and scipy.stats.mode() is used to calculate the mode. For external data, similar statistical calculations can also be performed through standard statistical tools. In this step, special attention should be paid to data cleaning and removal of outliers to ensure the accuracy of the calculation results, and finally the operation dimensional statistics or external dimensional statistics corresponding to each operation link are obtained.

[0035] Preferably, based on the operation dimension statistics or external dimension statistics corresponding to each operation link, a data density statistical analysis is performed on the corresponding operation data point set or external data point set to obtain the operation dimension data density or external dimension data density corresponding to each operation link; In an embodiment of the present invention, based on the previously obtained operational dimension statistics or external dimension statistics, a data density statistical analysis is performed on the data set corresponding to each operational link or the external data set. Data density refers to the distribution of data in a specific area or interval, and is often used to analyze whether the data is concentrated or dispersed. To achieve this goal, a density estimation method such as Gaussian kernel density estimation (Kernel Density Estimation, KDE) can be used. In actual operation, the scipy.stats.gaussian_kde function of Python or the density() function in the R language can be used to perform kernel density estimation on each data set to obtain the density distribution of each operational link or external data set, that is, , and finally obtain the operational dimension data density or external dimension data density corresponding to each operational link.

[0036] Preferably, based on the operational dimension data density or external dimension data density corresponding to each operational link, the noise neighborhood radius calculation formula is used to calculate the noise neighborhood radius of the operational data point set or external data point set corresponding to each operational link, so as to obtain the enterprise data noise neighborhood radius corresponding to each operational link; In an embodiment of the present invention, a suitable acoustic neighborhood radius calculation formula is constructed by combining the data space area range, the total number of data points corresponding to the operation data point set or the external data point set, the data points within the operation data point set or the external data point set, the Gaussian kernel standard deviation, the noise weight coefficient corresponding to the data point, the operation dimension data density or the external dimension data density, the neighborhood radius distance contribution coefficient, the maximum possible distance between data points and related parameters to calculate the noise neighborhood radius, so as to quantitatively calculate the noise neighborhood radius and finally obtain the enterprise data noise neighborhood radius corresponding to each operation link.

[0037] Preferably, based on the enterprise data noise neighborhood radius corresponding to each operation link, the corresponding operation data point set or external data point set is marked for neighborhood noise identification, so as to use the mean under the corresponding dimension as the center point and draw a neighborhood circle in combination with the enterprise data noise neighborhood radius. If a data point is within or on the neighborhood circle, it is regarded as a normal point; if a data point is not within the neighborhood circle, it is regarded as a noise point and marked, so as to obtain the operation noise marking data and external noise marking data corresponding to each operation link.

[0038] In an embodiment of the present invention, based on the noise neighborhood radius of the enterprise data corresponding to each operation link, the neighborhood circle method is used to perform noise identification and marking. The specific implementation is that in each operation link or external data point set, the neighborhood circle is drawn in combination with the previously calculated mean and noise neighborhood radius, that is, the mean is the center, the radius is the noise neighborhood radius, all data points in or on the neighborhood circle are regarded as normal points, and those data points outside the neighborhood circle are regarded as noise points and marked. During the operation, graphical tools such as Python's Matplotlib or Seaborn library are usually used to draw the distribution of data points and neighborhood circles. In this way, the difference between noise points and normal data points can be more intuitively displayed. After the data processing is completed, the noise points will be marked, and a data set with noise marking will be generated, and finally the operation noise marking data corresponding to each operation link and the external noise marking data will be obtained.

[0039] Furthermore, the noise neighborhood radius calculation formula is specifically as follows: ; In the formula, is the noise neighborhood radius of enterprise data, is the data space area range, is the total number of data points corresponding to the operational data point set or the external data point set, is an exponential function, The first data points, The first data points, σ ​​is the Gaussian kernel standard deviation, For the The noise weight coefficient corresponding to the data point is For operational dimension data density or external dimension data density, is the neighborhood radius distance contribution coefficient, is the maximum possible distance between data points, and η is the correction coefficient of the noise neighborhood radius of enterprise data.

[0040] The present invention obtains a noise neighborhood radius calculation formula by using a specific mathematical model and verifying it, which is used to calculate the noise neighborhood radius of the operation data point set or external data point set corresponding to each operation link. The noise neighborhood radius calculation formula can help identify abnormal data points (noise points) in the data set. By analyzing the neighborhood of each data point, it can be determined which data points deviate from the distribution of other points, thereby marking them as noise. This noise identification is crucial for data cleaning and data preprocessing, because noise points affect the training and prediction accuracy of subsequent models. By combining the dimensional data density, the noise neighborhood radius formula can reflect the distribution of data points in multidimensional space. If the neighborhood radius of a point is small and the data density is low, it indicates that the point is an abnormal point, which is particularly important for analyzing the distribution characteristics of different operation links or external data. The formula calculates the distance and weight between data points, combined with the effect of the Gaussian kernel, so that the similarity between discrete data points is smoother, thereby improving the sensitivity of noise identification, which also helps detect data points outside the normal data pattern and further optimizes the quality of the data set. The use of Gaussian kernel makes the noise neighborhood radius calculation more smooth and robust, which has a positive effect on avoiding misjudging data points as noise. By using a weighted neighborhood radius for each data point, the calculation can be adjusted according to the relative importance of each point, thereby improving the model's tolerance to noise. Multiple parameters in the noise neighborhood radius formula can be adjusted according to the specific situation of the data set. This flexibility allows the formula to adapt to different types of data sets and is suitable for the analysis of various operational links and external data. For dynamically changing data sets, the noise neighborhood radius can be dynamically calculated by adjusting these parameters to improve the accuracy and timeliness of the analysis. The exponential function and distance weight in the formula make the noise neighborhood radius more sensitive when considering the relative distance between data points, and can better capture those abnormal data that are significantly different from the overall data distribution. By combining the distance, weight and density analysis between data points, noise points can be more accurately identified. By effectively marking the noise data points, the analysis model of the enterprise operation data is further preprocessed, which can effectively remove the outliers that cause errors, thereby improving the accuracy of subsequent analysis and reducing the impact of external interference on the model. In summary, the formula fully considers the noise neighborhood radius of enterprise data , data space area range , the total number of data points corresponding to the operational data point set or external data point set , exponential function , the first Data points , the first Data points , Gaussian kernel standard deviation σ, The noise weight coefficient corresponding to the data point , operational dimension data density or external dimension data density , neighborhood radius distance contribution coefficient , the maximum possible distance between data points , the correction coefficient η of the enterprise data noise neighborhood radius, and a functional relationship is formed based on the correlation between the enterprise data noise neighborhood radius and the above parameters This formula can realize the noise neighborhood radius calculation process for the operation data point set or external data point set corresponding to each operation link. At the same time, by introducing the correction coefficient η of the noise neighborhood radius of enterprise data, it can be adjusted according to the error situation occurring in the calculation process, thereby improving the accuracy and applicability of the noise neighborhood radius calculation formula.

[0041] Further, step S2 includes the following steps: Step S21: construct each sub-neural network module, including an input layer, a hidden layer and an output layer; Step S22: performing time series synchronization processing on the normalized operation data and external normalized data corresponding to each operation link, and obtaining the time series of operation data and external data corresponding to each operation link in the same time series range; Step S23: According to each sub-neural network module, the long short-term memory network structure is used to perform timing state machine training and output on the operation data time series and external data time series corresponding to each operation link, so as to obtain the enterprise operation timing state machine vector corresponding to each sub-neural network module.

[0042] As an embodiment of the present invention, refer to Figure 2 As shown, Figure 1 Detailed step flow diagram of step S2 in FIG. 1 , in this embodiment, step S2 includes the following steps: Step S21: construct each sub-neural network module, including an input layer, a hidden layer and an output layer; In an embodiment of the present invention, by constructing each sub-neural network module, which includes an input layer, a hidden layer and an output layer, in this step, the function and data flow of each sub-neural network module are first determined. For example, in the enterprise operation decision optimization problem, the relevant data of each operation link can be input into the corresponding sub-neural network module. The input layer receives the data of each operation link, including company sales data, inventory data, employee performance data, etc. Then, feature extraction and data processing are performed through multiple hidden layers. Each hidden layer contains a number of neurons, and nonlinear mapping is performed through activation functions. The design of the hidden layer takes into account the complexity and multidimensionality of the operation link data, and uses common activation functions such as ReLU and sigmoid to enhance the performance of the model. The output layer is designed according to the task requirements, and the output is the trained operation decision information, such as the future prediction data or status label of each operation link, which is used for subsequent decision analysis. The structure of each sub-neural network module can be optimized by deep learning methods such as convolutional neural network (CNN) and recurrent neural network (RNN), and finally each sub-neural network module is obtained.

[0043] Step S22: performing time series synchronization processing on the normalized operation data and external normalized data corresponding to each operation link, and obtaining the time series of operation data and external data corresponding to each operation link in the same time series range; In an embodiment of the present invention, time series synchronization is performed on the normalized operation data corresponding to each operation link and the external normalized data. First, relevant data of each operation link of the enterprise (such as production, sales, logistics, etc.) is collected, and the data is normalized to convert data of different dimensions into a unified standard range, such as a numerical range of 0 to 1. External data such as macroeconomic data and industry development trends are also normalized. Next, the data is time-series aligned through timestamps to ensure the synchronization of each data series in the time dimension. If the data collection time intervals are different, interpolation or time resampling is required to adjust the data of different time series to a unified time scale. The specific time series synchronization method can be processed using algorithms such as interpolation and linear interpolation to ensure the consistency of the data time series, and finally the operation data time series and external data time series corresponding to each operation link in the same time series range are obtained.

[0044] Step S23: According to each sub-neural network module, the long short-term memory network structure is used to perform timing state machine training and output on the operation data time series and external data time series corresponding to each operation link, so as to obtain the enterprise operation timing state machine vector corresponding to each sub-neural network module.

[0045] In an embodiment of the present invention, the time series of operation data and the time series of external data corresponding to each operation link are trained and output by using a long short-term memory network structure according to each sub-neural network module. The long short-term memory network (LSTM) is a classic recurrent neural network (RNN) model with the advantage of processing long-term dependency problems. In this step, the synchronized time series of operation data and the time series of external data are first input into the LSTM network. LSTM models the time series data through its unique gating mechanism (input gate, forget gate, output gate), extracts the intrinsic relationship between different time points, and captures the long-term dependency in the data. Each sub-neural network module is trained through the LSTM network according to different operation links and external data to obtain the enterprise operation time series state machine vector corresponding to the sub-network. These time series state machine vectors can effectively represent the dynamic state of enterprise operation and serve as the input for subsequent decision optimization. The optimization goal in the training process is to minimize the error function. The network weights are updated using the back propagation algorithm to ensure that the output vector can accurately predict the enterprise operation state, and finally obtain the enterprise operation time series state machine vector corresponding to each sub-neural network module.

[0046] Further, step S23 includes the following steps: Step S231: According to each sub-neural network module, a long short-term memory network structure is used to perform time series state network module training on the operation data time series and the external data time series corresponding to each operation link, so as to use the input layer to receive the operation data time series and the external data time series corresponding to each operation link, and the hidden layer includes a number of LSTM network units to capture the enterprise operation changes corresponding to the short-term window of external environment changes and the long-term trend of enterprise operation corresponding to the long-term window of external environment changes, including an upward trend, a downward trend and a stable trend, and calculate the corresponding autocorrelation coefficient and partial correlation coefficient under each time window to reconstruct the data features of the input layer, so as to reconstruct and output the operation reconstruction sequence and the external reconstruction sequence corresponding to each operation link; In an embodiment of the present invention, the time series state network module is trained according to the time series of operation data of each operation link and the time series of external data, wherein the time series of operation data of each operation link (such as production, sales, human resources and other links) and the time series of external data (such as macroeconomic indicators, industry changes, competitor dynamics, etc.) need to be properly cleaned and normalized to ensure data consistency and standardization. To this end, a standardization method (such as Z-score or Min-Max standardization) can be used to pre-process the time series data to remove outliers and ensure that the data is on the same scale to avoid unsatisfactory training results due to dimensional differences. In terms of network structure, the input layer will receive pre-processed data from each operation link and external data, and the hidden layer contains multiple LSTM units. LSTM (Long Short-Term Memory Network) is a network structure suitable for processing time series data. It can capture long-term dependencies and short-term dependencies in the data. Through multi-layer LSTM units, enterprise operation changes and long-term trends under different time windows can be extracted. In the training process of the LSTM network, by The autocorrelation coefficient and partial correlation coefficient of the data in each time window are calculated to evaluate the correlation of the time series data in each time period. This process can be achieved through statistical methods. The autocorrelation and partial correlation functions in the statsmodels library in Python or the pandas library are used to calculate the data. The autocorrelation coefficient is used to measure the linear relationship between a certain time point and the past moment, and the partial correlation coefficient can reflect the relationship between the data after removing the intermediate variables, thereby helping the network to better reconstruct the features. Then, based on the autocorrelation coefficient and partial correlation coefficient, the network reconstructs the features of the input data. Through feature extraction and data reconstruction, the change trend of the operation data time series of each operation link and the external data time series is optimized to more accurately reflect the short-term and long-term changes in the enterprise operation process. The LSTM network continuously adjusts the weights during the training process through the back propagation algorithm to optimize the reconstruction effect, thereby reconstructing and outputting the operation reconstruction sequence and external reconstruction sequence corresponding to each operation link, which can reflect the relationship and trend changes between the operation link data and external data in each time window.

[0047] Step S232: Input the operation reconstruction sequence and external reconstruction sequence corresponding to each operation link into the corresponding output layer in the sub-neural network module for forward propagation state machine calculation, and use unique hot encoding to convert the state machine prediction result corresponding to the output layer into vector form, so as to obtain the enterprise operation timing state machine vector corresponding to each sub-neural network module, so as to reflect the corresponding change characteristics of the enterprise operation under different external environment timing states.

[0048] In the embodiment of the present invention, the operation reconstruction sequence corresponding to each operation link and the external reconstruction sequence obtained previously are input into the output layer of the corresponding sub-neural network module for forward propagation. The process uses the reconstruction sequence as input data and uses the neural network to perform state machine calculation to extract the time series state characteristics of the enterprise operation under different external environments. The network gradually derives the state output at each moment by adjusting the weights and biases and combining the activation function. The state machine prediction result of the network output layer is usually expressed in the form of categories. In order to facilitate subsequent processing and optimization, the output category needs to be one-hot encoded (One-Hot Encoding) conversion, by mapping the category predicted by each state machine into a vector representation, a high-dimensional sparse vector can be obtained, which represents the characteristics of enterprise operations under different states and can effectively reflect the transformation characteristics of time series data. The implementation of one-hot encoding can be completed through the OneHotEncoder class in Python's scikit-learn library, or a custom encoding method. After completing the one-hot encoding conversion, the state machine vectors output by each sub-neural network module are integrated into the final enterprise operation time series state machine vector. These vectors will reflect the changing characteristics of enterprise operations under different external environment time series states, such as the performance of enterprises under different external economic and market environment changes. Finally, the enterprise operation time series state machine vectors corresponding to each sub-neural network module are obtained to reflect the corresponding changing characteristics of enterprise operations under different external environment time series states.

[0049] Further, step S3 includes the following steps: Step S31: construct a main neural network module, which includes an input layer, a convolutional layer, a pooling layer, a fully connected layer and an output layer; In an embodiment of the present invention, a main neural network module is constructed, which consists of multiple layers, including an input layer, a convolution layer, a pooling layer, a fully connected layer and an output layer. The input layer is used to receive the enterprise operation time series state machine vectors generated by each sub-neural network module. These vectors contain the time series feature information of each operation link corresponding to the enterprise operation in the external environment. The convolution layer uses a 3x3 convolution kernel to perform a convolution operation on the input time series data to extract the time series correlation features between each operation link. The pooling layer uses a maximum pooling strategy to reduce the dimension of the time series features extracted by the convolution layer, reduce the feature dimension, and reduce the computational burden. Then, the reduced-dimensional features will be mapped to the decision recommendation space through the fully connected layer to form a multi-dimensional feature representation. In the final output layer, decision recommendations for the enterprise operation status are output based on the mapped features, thereby generating a complete neural network module, and finally constructing the main neural network module.

[0050] Step S32: Based on the main neural network module, a convolutional neural network structure is used to construct a model for the enterprise operation time series state machine vector corresponding to each sub-neural network module, so as to use the input layer to receive the enterprise operation time series state machine vector corresponding to each sub-neural network module, the convolution layer extracts the time series correlation features between each operation link through a 3x3 convolution kernel, the pooling layer performs dimensionality reduction processing on the time series correlation features between each operation link, the fully connected layer maps the reduced dimensionality features to the decision suggestion space, and the output layer gives the decision suggestion output corresponding to the enterprise operation status, thereby generating a multi-module neural network model; In an embodiment of the present invention, based on the constructed main neural network module, a convolutional neural network (CNN) structure is used to further construct a model for the enterprise operation time series state machine vector corresponding to each sub-neural network module. Each sub-neural network module receives the corresponding enterprise operation time series state machine vector as input data. After receiving the data through the input layer, the convolution layer uses a 3x3 convolution kernel to extract features from the time series state machine vector. The convolution layer captures the time series correlation features between each operation link through multiple convolution operations. These features include but are not limited to key information such as the time series trend and the amplitude of change between the data. The pooling layer reduces the dimension of the extracted time series features through the maximum pooling method, thereby reducing the complexity of the features and avoiding the occurrence of overfitting. The fully connected layer is responsible for mapping the reduced-dimensional features to the decision suggestion space to provide sufficient information support for subsequent decision outputs. The output layer generates a decision suggestion for the enterprise operation based on the output data of the fully connected layer. The output layer not only considers the current operation status of the enterprise, but also generates possible optimization paths in the future. Through the above structure, the model can effectively capture the time series dependency between the operation links, generate targeted operation decisions, and finally generate a multi-module neural network model.

[0051] Step S33: Use the multi-module neural network model to perform model training, decision prediction and output on the enterprise operation timing state machine vector corresponding to each sub-neural network module, so as to divide the enterprise operation timing state machine vector corresponding to each sub-neural network module into training set, verification set and test set according to the ratio of 70%, 15% and 15%, and train the training set through the Adam algorithm, set the learning rate to 0.001, the number of training rounds to 100, and evaluate the multi-module neural network model using the verification set after every 10 training rounds. Adjust the model parameters according to the evaluation results to prevent overfitting, and perform performance evaluation on the output layer corresponding to the multi-module neural network model by using the test set to calculate the corresponding accuracy and recall rate to determine the model evaluation results, and output the corresponding enterprise operation status decision recommendation results.

[0052] In an embodiment of the present invention, based on the constructed multi-module neural network model, the enterprise operation time series state machine vector corresponding to each sub-neural network module is trained and the decision prediction output is performed. In order to train the model, it is first necessary to divide the time series state machine vector data of each sub-neural network module into a training set, a validation set and a test set in proportion, wherein the training set accounts for 70%, the validation set accounts for 15%, and the test set accounts for 15%. The training set is trained by the Adam optimization algorithm, the learning rate is set to 0.001, and the number of training rounds is specified to be 100. During the training process, after every 10 rounds of training, the validation set is used to evaluate the current model, mainly to evaluate the accuracy of the model. In order to improve the accuracy, generalization ability and prevent overfitting, the parameters of the model can be adjusted based on the results of the validation set evaluation to ensure its performance consistency on different data sets. After the training is completed, the output layer of the multi-module neural network model is evaluated using the test set, and the corresponding accuracy, recall rate and other indicators are calculated. According to the evaluation results, the final performance of the model can be determined, and then the corresponding enterprise operation status decision-making recommendations can be output to help enterprises achieve intelligent optimization. Through the above steps, the efficiency and accuracy of the multi-module neural network model can be ensured, and the operation strategy can be dynamically adjusted according to the time series data of the enterprise operation, and finally the corresponding enterprise operation status decision-making recommendation results can be output.

[0053] Further, step S4 includes the following steps: Step S41: Regularly obtain the actual operational decision results of the enterprise; In an embodiment of the present invention, by collecting actual operational decision results from various departments of the enterprise, these data can be automatically collected through the enterprise's management information system (MIS), or manually entered by operations personnel from various departments such as production, sales, human resources, and marketing on a regular basis. During the collection process, focus is placed on actual business execution results, such as key performance indicators (KPIs) such as sales, production progress, market feedback, and customer satisfaction. These data should be updated according to a predetermined cycle (such as daily, weekly, or monthly) to ensure the timeliness of operational decisions. In addition, to ensure the accuracy of the data, all collected data should undergo a strict data cleaning process to remove invalid data and outliers, maintain data consistency and reliability, ensure real-time updating of all operational data, and ultimately obtain the actual operational decision results of the enterprise.

[0054] Step S42: performing forecast deviation calculation based on the actual operation decision results of the enterprise and the decision suggestion results of the enterprise operation status, so as to obtain the deviation between the actual operation of the enterprise and the model prediction; In an embodiment of the present invention, a neural network model is used to make predictions based on the historical data of the enterprise to generate model prediction results. This process trains the parameters of the neural network model by comparing the historical data and the prediction results so that it can accurately reflect the operating trends of the enterprise. Then, the actual operating decision results of the enterprise collected are compared with the prediction results generated by the neural network model to calculate their deviations. Specifically, the deviation can be calculated using common error evaluation methods such as mean square error (MSE) or absolute error (AE). Through this calculation process, a quantitative indicator can be obtained to represent the gap between the predictive ability of the model and the actual operating results of the enterprise. If the deviation is large, it means that the prediction effect of the model is not ideal and further optimization and adjustment are needed to finally obtain the deviation between the actual operation of the enterprise and the model prediction.

[0055] Step S43: Compare and judge the deviation between the actual operation of the enterprise and the model prediction according to the preset model deviation threshold. If the deviation between the actual operation of the enterprise and the model prediction is less than the preset model deviation threshold, use the new actual operation data of the enterprise to fine-tune the multi-module neural network model to update the corresponding model hyperparameters; if the deviation between the actual operation of the enterprise and the model prediction is greater than or equal to the preset model deviation threshold, re-collect and process the corresponding enterprise operation data, and retrain the multi-module neural network model to generate a multi-module neural network optimization model, and output the corresponding enterprise operation optimization decision results.

[0056] In an embodiment of the present invention, whether the prediction accuracy of the current neural network model meets the actual needs of the enterprise is determined according to a set deviation threshold. First, a deviation threshold is defined. For example, the deviation standard is set to an error range of less than 5%. If the deviation between the actual operation and the model prediction is less than the threshold, it means that the prediction result of the neural network model can better reflect the operating status of the enterprise. At this time, there is no need to make large-scale adjustments to the model. At this time, the neural network model is fine-tuned using the latest operating data of the enterprise. The fine-tuning process involves detailed adjustments to the model hyperparameters (such as learning rate, number of layers, number of neurons, etc.) in order to optimize the model to continue to provide effective predictions in future decision-making processes. The fine-tuning process can be achieved through gradient descent or other optimization algorithms to ensure that the model can make decisions more accurately based on new data. However, if the deviation is greater than or equal to the preset model deviation threshold, it means that the prediction effect of the model deviates significantly from the actual operating conditions of the enterprise, and the model needs to be retrained. In this process, the relevant enterprise operation data is first collected and processed again, and data cleaning and feature engineering are performed to ensure that the quality of the data used meets the training requirements. Then, the neural network model is retrained using these updated data to generate an optimized multi-module neural network model. At this time, different network architectures such as multi-layer perceptron (MLP), convolutional neural network (CNN) or recurrent neural network (RNN) can be selected according to the needs of the problem. During the training process, the cross-validation method is used to detect the generalization ability of the model to ensure that the final model can better predict future enterprise operation results. After the training is completed, the new enterprise operation optimization decision results are output, and finally a multi-module neural network optimization model is generated, and the corresponding enterprise operation optimization decision results are output.

[0057] Furthermore, the present invention also provides a neural network-based enterprise operation decision intelligent optimization system, which is used to execute the neural network-based enterprise operation decision intelligent optimization method as described above. The neural network-based enterprise operation decision intelligent optimization system includes: The operation link data preprocessing module is used to obtain the enterprise operation data corresponding to each operation link within the enterprise and the enterprise external data, and perform box cleaning and normalization processing on the enterprise operation data corresponding to each operation link and the enterprise external data, so as to obtain the operation normalized data corresponding to each operation link and the external normalized data; The submodule state machine training output module is used to construct each sub-neural network module, and according to each sub-neural network module, the long short-term memory network structure is used to perform time-series state machine training and output on the operation normalized data corresponding to each operation link and the external normalized data, so as to obtain the enterprise operation time-series state machine vector corresponding to each sub-neural network module; The network model decision prediction module is used to construct a main neural network module, and based on the main neural network module, use the convolutional neural network structure to perform model decision prediction on the enterprise operation time series state machine vector corresponding to each sub-neural network module, generate a multi-module neural network model, and output the corresponding enterprise operation status decision recommendation results; The model dynamic optimization and adjustment module is used to obtain the actual operation decision results of the enterprise, and to calculate the prediction deviation based on the actual operation decision results of the enterprise and the decision recommendation results of the enterprise's operating status, so as to obtain the deviation between the actual operation of the enterprise and the model prediction; based on the deviation between the actual operation of the enterprise and the model prediction, the multi-module neural network model is dynamically optimized and adjusted to generate a multi-module neural network optimization model, and the corresponding enterprise operation optimization decision results are output.

[0058] Therefore, the embodiments should be regarded as illustrative and non-restrictive from all points, and the scope of the present invention is limited by the appended claims rather than the above description, and it is therefore intended that all changes falling within the meaning and range of equivalent elements of the application documents are included in the present invention.

[0059] The above description is only a specific embodiment of the present invention, so that those skilled in the art can understand or implement the present invention. Various modifications to these embodiments will be apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the present invention. Therefore, the present invention will not be limited to the embodiments shown herein, but should conform to the widest scope consistent with the principles and novel features invented herein.

Claims

1. A method for intelligent optimization of enterprise operation decision-making based on neural network, characterized in that: The following steps are involved: Step S1: Obtain enterprise operation data corresponding to each operation link within the enterprise and enterprise external data, and perform box cleaning and normalization processing on the enterprise operation data corresponding to each operation link and enterprise external data to obtain operation normalized data corresponding to each operation link and external normalized data; Step S2: construct each sub-neural network module, and use the long short-term memory network structure according to each sub-neural network module to perform time series state machine training and output on the operation normalized data corresponding to each operation link and the external normalized data, and obtain the enterprise operation time series state machine vector corresponding to each sub-neural network module; Step S3: construct a main neural network module, and based on the main neural network module, use the convolutional neural network structure to perform model decision prediction on the enterprise operation time series state machine vector corresponding to each sub-neural network module, generate a multi-module neural network model, and output the corresponding enterprise operation status decision recommendation results; Step S4: Obtain the actual operation decision results of the enterprise, and perform forecast deviation accounting based on the actual operation decision results of the enterprise and the decision suggestion results of the enterprise operation status, so as to obtain the deviation between the actual operation of the enterprise and the model prediction; Based on the deviation between the actual operation of the enterprise and the model prediction, the multi-module neural network model is dynamically optimized and adjusted to generate a multi-module neural network optimization model and output the corresponding enterprise operation optimization decision results.

2. The enterprise operation decision intelligent optimization method based on neural network according to claim 1 is characterized in that: The various operation links described in step S1 include production, sales and human resources operation links.

3. The enterprise operation decision intelligent optimization method based on neural network according to claim 2 is characterized in that: Step S1 includes the following steps: Obtain enterprise operation data corresponding to each operation link within the enterprise and external data of the enterprise, where the enterprise operation data corresponding to each operation link includes production plan, output and equipment operation status data corresponding to the production operation link, sales volume, sales and customer sales feedback data corresponding to the sales operation link, employee attendance, performance and training status data corresponding to the human resources operation link, and external data of the enterprise includes industry dynamics and competition information data; Perform neighborhood noise identification and labeling on the enterprise operation data corresponding to each operation link and the enterprise external data, so as to obtain the operation noise labeling data and external noise labeling data corresponding to each operation link; The noise anomaly cleaning of the box plot is performed on the operation noise marked data corresponding to each operation link and the noise points and non-noise points in the external noise marked data, respectively, to obtain the operation cleaned data and external cleaned data corresponding to each operation link; Obtain the enterprise operation mean and the enterprise external mean corresponding to each operation link, and perform mean missing filling processing on the operation cleaning data and external cleaning data corresponding to each operation link based on the enterprise operation mean and the enterprise external mean corresponding to each operation link, so as to obtain the operation filling data and external filling data corresponding to each operation link; The min-max normalization method is used to normalize the operation filling data and external filling data corresponding to each operation link, so as to obtain the operation normalized data and external normalized data corresponding to each operation link.

4. The enterprise operation decision intelligent optimization method based on neural network according to claim 3 is characterized in that: The process of identifying and marking the neighborhood noise of the enterprise operation data and the enterprise external data corresponding to each operation link comprises the following steps: The enterprise operation data corresponding to each operation link and the operation data or external data corresponding to each time node in the enterprise external data are respectively taken as a data node to obtain the operation data point set or external data point set corresponding to each operation link; Perform dimensional statistics analysis on the operational data point set or external data point set corresponding to each operational link to obtain operational dimensional statistics or external dimensional statistics corresponding to each operational link, wherein the corresponding dimensional statistics include mean, median and mode; Based on the operation dimension statistics or external dimension statistics corresponding to each operation link, a data density statistical analysis is performed on the corresponding operation data point set or external data point set to obtain the operation dimension data density or external dimension data density corresponding to each operation link; Based on the operational dimension data density or external dimension data density corresponding to each operational link, the noise neighborhood radius calculation formula is used to calculate the noise neighborhood radius of the operational data point set or external data point set corresponding to each operational link, so as to obtain the enterprise data noise neighborhood radius corresponding to each operational link; Based on the enterprise data noise neighborhood radius corresponding to each operation link, the corresponding operation data point set or external data point set is marked for neighborhood noise identification, and the neighborhood circle is drawn using the mean under the corresponding dimension as the center point combined with the enterprise data noise neighborhood radius. If a data point is within or on the neighborhood circle, it is regarded as a normal point; if a data point is not within the neighborhood circle, it is regarded as a noise point and marked, so as to obtain the operation noise marked data and external noise marked data corresponding to each operation link.

5. The enterprise operation decision intelligent optimization method based on neural network according to claim 4 is characterized in that: The noise neighborhood radius calculation formula is specifically: ; In the formula, is the noise neighborhood radius of enterprise data, is the data space area range, is the total number of data points corresponding to the operational data point set or the external data point set, is an exponential function, The first data points, The first data points, σ ​​is the Gaussian kernel standard deviation, For the The noise weight coefficient corresponding to the data point is For operational dimension data density or external dimension data density, is the neighborhood radius distance contribution coefficient, is the maximum possible distance between data points, and η is the correction coefficient of the noise neighborhood radius of enterprise data.

6. The enterprise operation decision intelligent optimization method based on neural network according to claim 1 is characterized in that: Step S2 includes the following steps: Step S21: construct each sub-neural network module, including an input layer, a hidden layer and an output layer; Step S22: performing time series synchronization processing on the normalized operation data and external normalized data corresponding to each operation link, and obtaining the time series of operation data and external data corresponding to each operation link in the same time series range; Step S23: According to each sub-neural network module, the long short-term memory network structure is used to perform timing state machine training and output on the operation data time series and external data time series corresponding to each operation link, so as to obtain the enterprise operation timing state machine vector corresponding to each sub-neural network module.

7. The enterprise operation decision intelligent optimization method based on neural network according to claim 6 is characterized in that: Step S23 includes the following steps: Step S231: According to each sub-neural network module, a long short-term memory network structure is used to perform time series state network module training on the operation data time series and the external data time series corresponding to each operation link, so as to use the input layer to receive the operation data time series and the external data time series corresponding to each operation link, and the hidden layer includes a number of LSTM network units to capture the enterprise operation changes corresponding to the short-term window of external environment changes and the long-term trend of enterprise operation corresponding to the long-term window of external environment changes, including an upward trend, a downward trend and a stable trend, and calculate the corresponding autocorrelation coefficient and partial correlation coefficient under each time window to reconstruct the data features of the input layer, so as to reconstruct and output the operation reconstruction sequence and the external reconstruction sequence corresponding to each operation link; Step S232: Input the operation reconstruction sequence and external reconstruction sequence corresponding to each operation link into the corresponding output layer in the sub-neural network module for forward propagation state machine calculation, and use unique hot encoding to convert the state machine prediction result corresponding to the output layer into vector form, so as to obtain the enterprise operation timing state machine vector corresponding to each sub-neural network module, so as to reflect the corresponding change characteristics of the enterprise operation under different external environment timing states.

8. The enterprise operation decision intelligent optimization method based on neural network according to claim 1 is characterized in that: Step S3 includes the following steps: Step S31: construct a main neural network module, which includes an input layer, a convolutional layer, a pooling layer, a fully connected layer and an output layer; Step S32: Based on the main neural network module, a convolutional neural network structure is used to construct a model for the enterprise operation time series state machine vector corresponding to each sub-neural network module, so as to use the input layer to receive the enterprise operation time series state machine vector corresponding to each sub-neural network module, the convolution layer extracts the time series correlation features between each operation link through a 3x3 convolution kernel, the pooling layer performs dimensionality reduction processing on the time series correlation features between each operation link, the fully connected layer maps the reduced dimensionality features to the decision suggestion space, and the output layer gives the decision suggestion output corresponding to the enterprise operation status, thereby generating a multi-module neural network model; Step S33: Use the multi-module neural network model to perform model training, decision prediction and output on the enterprise operation timing state machine vector corresponding to each sub-neural network module, so as to divide the enterprise operation timing state machine vector corresponding to each sub-neural network module into training set, verification set and test set according to the ratio of 70%, 15% and 15%, and train the training set through the Adam algorithm, set the learning rate to 0.001, the number of training rounds to 100, and evaluate the multi-module neural network model using the verification set after every 10 training rounds. Adjust the model parameters according to the evaluation results to prevent overfitting, and perform performance evaluation on the output layer corresponding to the multi-module neural network model by using the test set to calculate the corresponding accuracy and recall rate to determine the model evaluation results, and output the corresponding enterprise operation status decision recommendation results.

9. The enterprise operation decision intelligent optimization method based on neural network according to claim 1 is characterized in that: Step S4 includes the following steps: Step S41: Regularly obtain the actual operational decision results of the enterprise; Step S42: performing forecast deviation calculation based on the actual operation decision results of the enterprise and the decision suggestion results of the enterprise operation status, so as to obtain the deviation between the actual operation of the enterprise and the model prediction; Step S43: Compare and judge the deviation between the actual operation of the enterprise and the model prediction according to the preset model deviation threshold. If the deviation between the actual operation of the enterprise and the model prediction is less than the preset model deviation threshold, use the new actual operation data of the enterprise to fine-tune the multi-module neural network model to update the corresponding model hyperparameters; if the deviation between the actual operation of the enterprise and the model prediction is greater than or equal to the preset model deviation threshold, re-collect and process the corresponding enterprise operation data, and retrain the multi-module neural network model to generate a multi-module neural network optimization model, and output the corresponding enterprise operation optimization decision results.

10. An intelligent optimization system for enterprise operation decision-making based on neural network, characterized in that: Used to execute the enterprise operation decision intelligent optimization method based on neural network as claimed in claim 1, the enterprise operation decision intelligent optimization system based on neural network includes: The operation link data preprocessing module is used to obtain the enterprise operation data corresponding to each operation link within the enterprise and the enterprise external data, and perform box cleaning and normalization processing on the enterprise operation data corresponding to each operation link and the enterprise external data, so as to obtain the operation normalized data corresponding to each operation link and the external normalized data; The submodule state machine training output module is used to construct each sub-neural network module, and according to each sub-neural network module, the long short-term memory network structure is used to perform time-series state machine training and output on the operation normalized data corresponding to each operation link and the external normalized data, so as to obtain the enterprise operation time-series state machine vector corresponding to each sub-neural network module; The network model decision prediction module is used to construct a main neural network module, and based on the main neural network module, use the convolutional neural network structure to perform model decision prediction on the enterprise operation time series state machine vector corresponding to each sub-neural network module, generate a multi-module neural network model, and output the corresponding enterprise operation status decision recommendation results; The model dynamic optimization and adjustment module is used to obtain the actual operation decision results of the enterprise, and to calculate the prediction deviation based on the actual operation decision results of the enterprise and the decision recommendation results of the enterprise's operating status, so as to obtain the deviation between the actual operation of the enterprise and the model prediction; based on the deviation between the actual operation of the enterprise and the model prediction, the multi-module neural network model is dynamically optimized and adjusted to generate a multi-module neural network optimization model, and the corresponding enterprise operation optimization decision results are output.

Images