Enterprise operation analysis early warning control system and method

Abstract

The invention relates to the technical field of intelligent analysis, and particularly discloses an enterprise operation analysis early warning control system and method.The enterprise operation analysis early warning control method comprises the steps that firstly, enterprise operation analysis indexes are determined, current data, historical data and industry average data are obtained according to the analysis indexes, a context encoder is used for conducting sufficient semantic understanding on the current data, the historical data and the industry average data; and the current operation state of the enterprise is judged based on the association characteristics of the three. Therefore, the reliability of enterprise operation analysis and the effect of risk early warning are improved, so that an enterprise decision maker is helped to better understand the enterprise operation condition, and the accuracy and efficiency of decision making are improved.

Description

Technical Field

[0001] The present application relates to the field of intelligent analysis technology, and more specifically, to an enterprise operation analysis early warning control system and method. Background Art

[0002] Enterprises will face various potential problems and risks in their operations, such as market changes, competitive pressure, supply chain problems, etc. If the defects of enterprise operations are not discovered in time, the further expansion of the problems will cause huge losses to the enterprise. Therefore, enterprise operation analysis and early warning are crucial for the stable operation and sustainable development of enterprises. Through operation analysis, we can better understand the operation status of enterprises, help enterprises to discover problems in time, take countermeasures, optimize resource allocation, and improve the accuracy and efficiency of decision-making, so as to help enterprises maintain keen market perception and competitive advantage.

[0003] However, existing enterprise operation analysis usually relies on piling up facts from report data without any refinement, resulting in low reliability and poor risk warning effect.

[0004] Therefore, an enterprise operation analysis and early warning system and method are expected. Summary of the invention

[0005] In order to solve the above technical problems, this application is proposed. The embodiment of this application provides an enterprise operation analysis early warning control system and method, which first determines the enterprise operation analysis indicators, obtains current data, historical data and industry average data according to the analysis indicators, uses context encoders to fully understand the semantics of the three, and judges the current operation status of the enterprise based on the correlation characteristics of the three. In this way, the reliability of enterprise operation analysis and the effect of risk early warning are improved, thereby helping enterprise decision makers to better understand the enterprise operation status and improve the accuracy and efficiency of decision-making.

[0006] Accordingly, according to one aspect of the present application, there is provided an enterprise operation analysis early warning control system, which includes:

[0007] An information acquisition module is used to determine enterprise operation analysis indicators and obtain current data, historical data and industry average data based on the enterprise operation analysis indicators;

[0008] A semantic understanding module, for respectively passing the text description of the current data, the text description of the historical data and the text description of the industry average data through a converter-based context encoder to obtain an enterprise operation status semantic feature vector, an enterprise historical operation status semantic feature vector and an industry standard semantic feature vector;

[0009] A correlation feature extraction module, used to arrange the enterprise operation status semantic feature vector, the enterprise historical operation status semantic feature vector and the industry standard semantic feature vector into a two-dimensional input matrix and then pass it through a convolutional neural network model with mutually transposed convolution kernels in adjacent layers to obtain a correlation feature matrix;

[0010] A feature enhancement module, used for passing the correlation feature matrix through a bidirectional attention mechanism model to obtain a classification feature matrix;

[0011] An optimization module, used for performing low-dimensional mask sparseness on the classification feature matrix to obtain an optimized classification feature matrix;

[0012] The analysis result generation module is used to pass the optimized classification feature matrix through a classifier to obtain a classification result, and the classification result is used to indicate whether the current operating status of the enterprise is excellent, normal or at risk.

[0013] In the above-mentioned enterprise operation analysis, early warning and control system, the semantic understanding module includes: a word segmentation unit, which is used to perform word segmentation processing on the text description of the current data to obtain a current data word sequence; an embedding coding unit, which is used to use the embedding layer of the context encoder to embed code each current data word in the current data word sequence respectively to obtain a sequence of current data word embedding vectors; a context coding unit, which is used to input the sequence of current data word embedding vectors into the converter module of the context encoder to obtain multiple current data feature vectors; a cascade unit, which is used to cascade the multiple current data feature vectors to obtain the enterprise operation status semantic feature vector.

[0014] In the above-mentioned enterprise operation analysis and early warning control system, the context encoding unit includes: an arrangement subunit, which is used to arrange the sequence of the current data word embedding vectors into an input vector; a conversion subunit, which is used to convert the input vector into a query vector and a key vector respectively through a learnable embedding matrix; a self-attention association matrix generation subunit, which is used to calculate the product between the query vector and the transposed vector of the key vector to obtain a self-attention association matrix; a standardized self-attention subunit, which is used to standardize the self-attention association matrix to obtain a standardized self-attention association matrix; an attention calculation subunit, which is used to input the standardized self-attention association matrix into a Softmax activation function for activation to obtain a self-attention feature matrix; an attention application subunit, which is used to multiply the self-attention feature matrix with each current data word embedding vector in the sequence of the current data word embedding vectors to obtain the multiple current data feature vectors.

[0015] In the above-mentioned enterprise operation analysis and early warning control system, the associated feature extraction module is used to: use the adjacent layers of the convolutional neural network model using mutually transposed convolution kernels to respectively perform the following on the input data in the forward transfer of the layer: perform convolution processing, pooling processing and nonlinear processing based on a first two-dimensional convolution kernel on the input data to obtain a first activation feature map; perform convolution processing, pooling processing and nonlinear activation processing based on a second two-dimensional convolution kernel on the first activation feature map to obtain a second activation feature map, wherein the first two-dimensional convolution kernel and the second two-dimensional convolution kernel are transposed of each other; wherein the input of the first layer of the convolutional neural network model of the adjacent layers using mutually transposed convolution kernels is the two-dimensional input matrix, and the output of the last layer of the convolutional neural network model of the adjacent layers using mutually transposed convolution kernels is the associated feature matrix.

[0016] In the above-mentioned enterprise operation analysis, early warning and control system, the feature enhancement module includes: a pooling unit, which is used to pool the correlation feature matrix along the horizontal direction and the vertical direction respectively to obtain a first-direction pooling vector and a second-direction pooling vector; an association unit, which is used to perform association encoding on the first-direction pooling vector and the second-direction pooling vector to obtain a bidirectional correlation matrix; a weight calculation unit, which is used to input the bidirectional correlation matrix into a Sigmoid activation function to obtain a bidirectional correlation weight matrix; and a weight application unit, which is used to calculate the position point multiplication between the bidirectional correlation weight matrix and the correlation feature matrix to obtain the classification feature matrix.

[0017] In the above-mentioned enterprise operation analysis and early warning control system, the optimization module includes: a matrix decomposition unit, which is used to perform matrix decomposition based on eigenvalues ​​on the classification feature matrix to obtain multiple eigenvalues ​​and multiple eigenvectors corresponding to the multiple eigenvalues; an affine subspace construction unit, which is used to arrange the multiple eigenvectors in descending order based on the size of the eigenvalues, and extract the eigenvectors corresponding to the K largest eigenvalues ​​from the multiple eigenvectors to form a K-dimensional affine subspace; a mapping unit, which is used to map the classification feature matrix to the affine subspace to obtain an affine density domain matrix; a probabilization unit, which is used to input the affine density domain matrix into the Softmax activation function to obtain a probabilistic affine density domain matrix; a masking unit, which is used to mask the probabilistic affine density domain matrix based on the comparison between the eigenvalues ​​of each position in the probabilistic affine density domain matrix and a predetermined threshold value to obtain a masked probabilistic affine density domain matrix; an optimization unit, which is used to calculate the positional point multiplication between the masked probabilistic affine density domain matrix and the classification feature matrix to obtain the optimized classification feature matrix.

[0018] In the above-mentioned enterprise operation analysis and early warning control system, the analysis result generation module includes: a matrix expansion unit, which is used to expand the optimized classification feature matrix into a classification feature vector according to a row vector or a column vector; a fully connected encoding unit, which is used to use the fully connected layer of the classifier to fully connect the classification feature vector to obtain a fully connected encoded feature vector; a probability acquisition unit, which is used to pass the fully connected encoded feature vector through the Softmax classification function of the classifier to obtain a first probability belonging to the excellent current operation status of the enterprise, a second probability belonging to the normal current operation status of the enterprise, and a third probability belonging to the risk of the current operation status of the enterprise; a classification unit, which is used to determine the classification label corresponding to the largest of the probability values ​​as the classification result.

[0019] According to another aspect of the present application, there is provided an enterprise operation analysis early warning control method, which includes:

[0020] Determine enterprise operation analysis indicators, and obtain current data, historical data and industry average data based on the enterprise operation analysis indicators;

[0021] The text description of the current data, the text description of the historical data and the text description of the industry average data are respectively passed through a converter-based context encoder to obtain an enterprise operation status semantic feature vector, an enterprise historical operation status semantic feature vector and an industry standard semantic feature vector;

[0022] Arranging the enterprise operation status semantic feature vector, the enterprise historical operation status semantic feature vector and the industry standard semantic feature vector into a two-dimensional input matrix and then passing through a convolutional neural network model with mutually transposed convolution kernels in adjacent layers to obtain a correlation feature matrix;

[0023] Pass the correlation feature matrix through a bidirectional attention mechanism model to obtain a classification feature matrix;

[0024] Performing low-dimensional mask sparseness on the classification feature matrix to obtain an optimized classification feature matrix;

[0025] The optimized classification feature matrix is ​​passed through a classifier to obtain a classification result, and the classification result is used to indicate whether the current operating status of the enterprise is excellent, normal, or at risk.

[0026] In the above-mentioned enterprise operation analysis, early warning and control method, the text description of the current data is passed through a converter-based context encoder to obtain a semantic feature vector of the enterprise operation status, including: performing word segmentation processing on the text description of the current data to obtain a current data word sequence; using the embedding layer of the context encoder to embed each current data word in the current data word sequence to obtain a sequence of current data word embedding vectors; inputting the sequence of current data word embedding vectors into the converter module of the context encoder to obtain multiple current data feature vectors; and cascading the multiple current data feature vectors to obtain the semantic feature vector of the enterprise operation status.

[0027] In the above-mentioned enterprise operation analysis and early warning control method, the sequence of the current data word embedding vectors is input into the converter module of the context encoder to obtain multiple current data feature vectors, including: arranging the sequence of the current data word embedding vectors as an input vector; converting the input vector into a query vector and a key vector respectively through a learnable embedding matrix; calculating the product between the query vector and the transposed vector of the key vector to obtain a self-attention association matrix; normalizing the self-attention association matrix to obtain a standardized self-attention association matrix; inputting the standardized self-attention association matrix into a Softmax activation function for activation to obtain a self-attention feature matrix; multiplying the self-attention feature matrix with each current data word embedding vector in the sequence of the current data word embedding vectors to obtain the multiple current data feature vectors.

[0028] Compared with the prior art, the enterprise operation analysis early warning control system and method provided by this application first determines the enterprise operation analysis indicators, obtains current data, historical data and industry average data according to the analysis indicators, uses context encoders to fully understand the semantics of the three, and judges the current operation status of the enterprise based on the correlation characteristics of the three. In this way, the reliability of enterprise operation analysis and the effect of risk early warning are improved, thereby helping enterprise decision makers to better understand the enterprise operation status and improve the accuracy and efficiency of decision-making. BRIEF DESCRIPTION OF THE DRAWINGS

[0029] By describing the embodiments of the present application in more detail in conjunction with the accompanying drawings, the above and other purposes, features and advantages of the present application will become more apparent. The accompanying drawings are used to provide a further understanding of the embodiments of the present application and constitute a part of the specification. Together with the embodiments of the present application, they are used to explain the present application and do not constitute a limitation of the present application. In the accompanying drawings, the same reference numerals generally represent the same components or steps.

[0030] Figure 1 It is a block diagram of an enterprise operation analysis and early warning control system according to an embodiment of the present application.

[0031] Figure 2 Schematic diagram of the architecture of an enterprise operation analysis and early warning control system according to an embodiment of the present application.

[0032] Figure 3 It is a block diagram of the semantic understanding module in the enterprise operation analysis and early warning control system according to an embodiment of the present application.

[0033] Figure 4 It is a block diagram of a context encoding unit in an enterprise operation analysis and early warning control system according to an embodiment of the present application.

[0034] Figure 5 This is a block diagram of an analysis result generation module in an enterprise operation analysis and early warning control system according to an embodiment of the present application.

[0035] Figure 6 Flow chart of the enterprise operation analysis early warning control method according to an embodiment of the present application. DETAILED DESCRIPTION

[0036] Below, the exemplary embodiments according to the present application will be described in detail with reference to the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present application, rather than all the embodiments of the present application, and it should be understood that the present application is not limited to the exemplary embodiments described here.

[0037] Application Overview

[0038] As mentioned above, through the operation analysis of enterprises, we can better understand the operation status of enterprises, help enterprises to find problems in time, take countermeasures, optimize resource allocation, and improve the accuracy and efficiency of decision-making, so as to help enterprises maintain keen market perception and competitive advantage. However, the existing enterprise operation analysis usually relies on the fact piling up of report data without refinement, with low reliability and poor risk warning effect. Therefore, we expect an enterprise operation analysis warning system and method to improve the risk warning effect and help enterprise decision makers better understand the operation status of enterprises.

[0039] At present, deep learning and neural networks have been widely used in computer vision, natural language processing, speech signal processing and other fields. In addition, deep learning and neural networks have also shown a level close to or even beyond that of humans in image classification, object detection, semantic segmentation, text translation and other fields.

[0040] In recent years, the development of deep learning and neural networks has provided new solutions and plans for the analysis and early warning of enterprise operating status.

[0041] Accordingly, when analyzing the operating status of an enterprise, first of all, the key indicators of analysis and the factors for measuring operating conditions should be clarified, such as sales, profit margin, net profit margin, return on assets, cash flow, etc. After establishing the analysis direction, the current analysis indicator data of the enterprise should be compared and analyzed with historical data and industry average data to evaluate the current operating status of the enterprise and conduct risk warning. In other words, the current data and historical data are analyzed and compared to provide the trend and evolution of the operating status of the enterprise, while the industry average data can provide an industry benchmark to evaluate its own market share through market research. Based on this, in the technical solution of the present application, first, the enterprise operation analysis indicators are determined, and the current data, historical data and industry average data are obtained according to the enterprise operation analysis indicators.

[0042] Next, the text description of the current data, the text description of the historical data, and the text description of the industry average data are respectively passed through a converter-based context encoder to extract semantic features from the text information, thereby obtaining a semantic feature vector of the enterprise operating status, a semantic feature vector of the enterprise historical operating status, and an industry standard semantic feature vector. It should be understood that the text description contains a large amount of unstructured information, such as natural language text. However, there are contextual relationships between natural language texts, and traditional rule-based or keyword matching methods are often difficult to fully capture contextual relationships. The converter-based context encoder is a pre-trained deep learning model with strong semantic understanding and representation capabilities. It can capture the semantic associations and contextual information between words by learning the semantic relationships in a large-scale corpus to convert text information into a vector format containing semantic feature information. In addition, vectors usually have lower dimensions, which can greatly reduce the dimensions of model processing and speed up calculation efficiency compared to the original text description.

[0043] Then, after arranging the semantic feature vector of the enterprise operation status, the semantic feature vector of the enterprise historical operation status and the semantic feature vector of the industry standard into a two-dimensional input matrix, a convolutional neural network (CNN) model is used to perform correlation feature analysis and extraction to obtain a correlation feature matrix. It should be understood that a convolutional neural network is a deep learning model widely used in fields such as computer vision and natural language processing. It has good feature extraction capabilities when processing two-dimensional structured data (such as images). It can use convolution kernels to extract features in local areas, and can capture features of different scales and abstract levels through multiple convolution kernels of different scales. In the technical solution of the present application, the semantic feature vector of the enterprise operation status, the semantic feature vector of the enterprise historical operation status and the semantic feature vector of the industry standard are arranged into a two-dimensional input matrix, and each vector corresponds to a row or column of a matrix. Then, a convolution kernel is used to perform a convolution operation on the two-dimensional input matrix to extract the correlation features in the matrix. In particular, in order to better capture the relationship between features, mutually transposed convolution kernels are used here, which can perform convolution operations in different directions, thereby better revealing the mutual relationship between data and the correlation of features, so as to provide more fine-grained feature information and improve the accuracy and effectiveness of operational analysis.

[0044] Furthermore, the association feature matrix is ​​passed through a bidirectional attention mechanism model to obtain a classification feature matrix. The bidirectional attention mechanism is a technology widely used in natural language processing and sequence modeling tasks. It can dynamically allocate attention to different positions in the input sequence to capture key information and perform weighted aggregation. That is, by using the bidirectional attention mechanism, key features are selected and aggregated from the association feature matrix to extract more important features and reduce redundant information. In addition, the bidirectional attention mechanism model simultaneously considers the positive and negative correlations between the features in the matrix. By calculating the positive and negative attention weights of the features and other features respectively, the relationship between the features is comprehensively modeled. Then, the association feature matrix is ​​weighted according to the attention weights to generate a classification feature matrix.

[0045] Finally, the classification feature matrix is ​​passed through a classifier to evaluate the current operating status of the enterprise. The classifier can learn the association between the input features and the predetermined categories to perform classification or prediction tasks. Here, the trained classifier can map the classification feature matrix to the categories pre-set to represent the current operating status of the enterprise, namely "excellent", "normal" and "risky". In this way, the risks or problems that the enterprise may face can be discovered in a timely manner, and at the same time, it also provides important indications about the current operating status for enterprise managers, which helps to take corresponding measures early, improve operating conditions, and optimize business decisions.

[0046] In particular, considering that the classification feature matrix may contain a large number of features, some of which may contribute little to the results of the classification task or even be noise. Through low-dimensional mask sparsification, the weights of those features that contribute little to the classification results can be eliminated or reduced, thereby reducing the dimension of the feature space. Doing so helps to improve the efficiency of the classifier and reduce computational overhead. At the same time, too many features may lead to overfitting problems, that is, the model performs well on the training data, but has poor generalization ability on new data. Through low-dimensional mask sparsification, features that are more important and relevant to the classification task can be screened out, reducing the degree of fitting of the model to noise and irrelevant features, thereby reducing the risk of overfitting. And the classification feature matrix optimized by low-dimensional mask sparsification pays more attention to important features and can better capture the key information in the data. This can improve the performance and accuracy of the classifier, enabling it to better distinguish different categories.

[0047] Specifically, the classification feature matrix is ​​subjected to low-dimensional mask sparsification to obtain an optimized classification feature matrix, including: performing matrix decomposition based on eigenvalues ​​on the classification feature matrix to obtain multiple eigenvalues ​​and multiple eigenvectors corresponding to the multiple eigenvalues; arranging the multiple eigenvectors in descending order based on the size of the eigenvalues, and extracting eigenvectors corresponding to K largest eigenvalues ​​from the multiple eigenvectors to form a K-dimensional affine subspace; mapping the classification feature matrix to the affine subspace to obtain an affine density domain matrix; inputting the affine density domain matrix into a Softmax activation function to obtain a probabilistic affine density domain matrix; based on a comparison between the eigenvalues ​​of each position in the probabilistic affine density domain matrix and a predetermined threshold, masking the probabilistic affine density domain matrix to obtain a masked probabilistic affine density domain matrix; calculating the position-based point multiplication between the masked probabilistic affine density domain matrix and the classification feature matrix to obtain the optimized classification feature matrix.

[0048] The classification feature matrix is ​​probabilized in the affine density domain of the feature space to perform non-parametric clustering based on the density domain on the classification feature matrix, so that the clustering structure in the data distribution of the classification feature matrix can be automatically discovered without specifying the number of clusters. Specifically, in the technical solution of the present application, the affine transformation in the feature space is utilized to map the classification feature matrix to a low-dimensional affine subspace, and then the logistic activation function is used in the space to estimate and calculate the probability density of each data point to obtain a probabilistic affine density domain matrix, and then by setting a probability threshold, the high-density area can be divided into clusters and the low-density area is regarded as noise or anomaly to obtain a masked probabilistic affine density domain matrix, and the masked probabilistic affine density domain matrix is ​​applied to the classification feature matrix to achieve optimal modulation of the feature distribution. In this way, the problem of high-dimensional data and complex distribution can be effectively handled, because it does not need to assume that the data obeys a certain specific distribution, nor does it need to predetermine the parameters or structure of the model; at the same time, the robustness of the data can be improved, because it can adaptively adjust the affine subspace to adapt to the changes in data and the influence of noise.

[0049] After introducing the basic principles of the present application, various non-limiting embodiments of the present application will be described in detail with reference to the accompanying drawings.

[0050] Exemplary Systems

[0051] Figure 1 FIG. 1 is a block diagram of an enterprise operation analysis and early warning control system according to an embodiment of the present application. Figure 1As shown, according to the enterprise operation analysis early warning control system 100 of the embodiment of the present application, it includes: an information acquisition module 110, which is used to determine the enterprise operation analysis index, and obtain the current data, historical data and industry average data according to the enterprise operation analysis index; a semantic understanding module 120, which is used to pass the text description of the current data, the text description of the historical data and the text description of the industry average data through a converter-based context encoder to obtain the enterprise operation status semantic feature vector, the enterprise historical operation status semantic feature vector and the industry standard semantic feature vector; an associated feature extraction module 130, which is used to convert the enterprise operation status semantic feature vector, the historical operation status semantic feature vector and the industry standard semantic feature vector into a context encoder based on a converter. The semantic feature vector of the enterprise's historical operating status and the industry standard semantic feature vector are arranged into a two-dimensional input matrix and then passed through a convolutional neural network model with mutually transposed convolution kernels in adjacent layers to obtain a correlation feature matrix; a feature enhancement module 140 is used to pass the correlation feature matrix through a bidirectional attention mechanism model to obtain a classification feature matrix; an optimization module 150 is used to perform low-dimensional mask sparsification on the classification feature matrix to obtain an optimized classification feature matrix; an analysis result generation module 160 is used to pass the optimized classification feature matrix through a classifier to obtain a classification result, and the classification result is used to indicate that the current operating status of the enterprise is excellent, normal, or at risk.

[0052] Figure 2 FIG. 1 is a schematic diagram of the architecture of the enterprise operation analysis and early warning control system according to an embodiment of the present application. Figure 2 As shown, first, determine the enterprise operation analysis indicators, and obtain current data, historical data and industry average data according to the enterprise operation analysis indicators. Then, the text description of the current data, the text description of the historical data and the text description of the industry average data are respectively passed through the context encoder based on the converter to obtain the enterprise operation status semantic feature vector, the enterprise historical operation status semantic feature vector and the industry standard semantic feature vector. Then, the enterprise operation status semantic feature vector, the enterprise historical operation status semantic feature vector and the industry standard semantic feature vector are arranged into a two-dimensional input matrix and then passed through the convolutional neural network model with mutually transposed convolution kernels in adjacent layers to obtain the associated feature matrix. Secondly, the associated feature matrix is ​​passed through the bidirectional attention mechanism model to obtain the classification feature matrix. Then, the classification feature matrix is ​​subjected to low-dimensional mask sparseness to obtain the optimized classification feature matrix. Finally, the optimized classification feature matrix is ​​passed through the classifier to obtain the classification result, and the classification result is used to indicate that the current operation status of the enterprise is excellent, normal or at risk.

[0053] In the above-mentioned enterprise operation analysis early warning control system 100, the information acquisition module 110 is used to determine the enterprise operation analysis indicators, and obtain current data, historical data and industry average data according to the enterprise operation analysis indicators. As mentioned in the above-mentioned background technology, by conducting operation analysis on the enterprise, we can better understand the operation status of the enterprise, help the enterprise to find problems in time, take countermeasures, optimize resource allocation, and improve the accuracy and efficiency of decision-making, thereby helping the enterprise maintain a keen market perception and competitive advantage. However, existing enterprise operation analysis usually relies on report data to pile up facts without refinement, has low reliability, and poor risk warning effect. Therefore, it is expected to improve the risk warning effect of enterprise operation analysis.

[0054] Accordingly, when analyzing the operating status of an enterprise, first of all, the key indicators of analysis and the factors for measuring operating conditions should be clarified, such as sales, profit margin, net profit margin, return on assets, cash flow, etc. After establishing the analysis direction, the current analysis indicator data of the enterprise should be compared and analyzed with historical data and industry average data to evaluate the current operating status of the enterprise and conduct risk warning. In other words, the current data and historical data are analyzed and compared to provide the trend and evolution of the operating status of the enterprise, while the industry average data can provide an industry benchmark to evaluate its own market share through market research. Based on this, in the technical solution of the present application, first, the enterprise operation analysis indicators are determined, and the current data, historical data and industry average data are obtained according to the enterprise operation analysis indicators.

[0055] In the above-mentioned enterprise operation analysis and early warning control system 100, the semantic understanding module 120 is used to pass the text description of the current data, the text description of the historical data and the text description of the industry average data through the converter-based context encoder to obtain the enterprise operation status semantic feature vector, the enterprise historical operation status semantic feature vector and the industry standard semantic feature vector. It should be understood that the text description contains a large amount of unstructured information, such as natural language text. There is a contextual relationship between natural language texts, and traditional rule-based or keyword matching methods are often difficult to fully capture the contextual relationship. The converter-based context encoder is a pre-trained deep learning model with strong semantic understanding and representation capabilities. It can capture the semantic association and context information between words by learning the semantic relationship in a large-scale corpus to convert text information into a vector format containing semantic feature information. In addition, the vector usually has a lower dimension, which can greatly reduce the dimension of model processing and speed up calculation efficiency compared to the original text description.

[0056] Figure 3 FIG. 1 is a block diagram of a semantic understanding module in an enterprise operation analysis and early warning control system according to an embodiment of the present application. Figure 3As shown, the semantic understanding module 120 includes: a word segmentation unit 121, which is used to perform word segmentation processing on the text description of the current data to obtain a current data word sequence; an embedding coding unit 122, which is used to use the embedding layer of the context encoder to respectively embed the current data words in the current data word sequence to obtain a sequence of current data word embedding vectors; a context coding unit 123, which is used to input the sequence of current data word embedding vectors into the converter module of the context encoder to obtain multiple current data feature vectors; a cascade unit 124, which is used to cascade the multiple current data feature vectors to obtain the enterprise operation status semantic feature vector.

[0057] Figure 4 FIG. 1 is a block diagram of a context encoding unit in an enterprise operation analysis and early warning control system according to an embodiment of the present application. Figure 4 As shown, the context encoding unit 123 includes: an arrangement subunit 1231, which is used to arrange the sequence of the current data word embedding vectors into an input vector; a conversion subunit 1232, which is used to convert the input vector into a query vector and a key vector respectively through a learnable embedding matrix; a self-attention association matrix generation subunit 1233, which is used to calculate the product between the query vector and the transposed vector of the key vector to obtain a self-attention association matrix; a standardized self-attention subunit 1234, which is used to standardize the self-attention association matrix to obtain a standardized self-attention association matrix; an attention calculation subunit 1235, which is used to input the standardized self-attention association matrix into a Softmax activation function for activation to obtain a self-attention feature matrix; an attention application subunit 1236, which is used to multiply the self-attention feature matrix with each current data word embedding vector in the sequence of the current data word embedding vectors to obtain the multiple current data feature vectors.

[0058] In the above-mentioned enterprise operation analysis and early warning control system 100, the associated feature extraction module 130 is used to arrange the enterprise operation status semantic feature vector, the enterprise historical operation status semantic feature vector and the industry standard semantic feature vector into a two-dimensional input matrix and then use a convolutional neural network model with mutually transposed convolution kernels through adjacent layers to obtain an associated feature matrix. It should be understood that a convolutional neural network is a deep learning model widely used in fields such as computer vision and natural language processing. It has good feature extraction capabilities when processing two-dimensional structured data (such as images). It can use convolution kernels to extract features in local areas, and can capture features of different scales and abstract levels through multiple convolution kernels of different scales. In the technical solution of the present application, the enterprise operation status semantic feature vector, the enterprise historical operation status semantic feature vector and the industry standard semantic feature vector are arranged into a two-dimensional input matrix, and each vector corresponds to a row or column of a matrix. Then, the two-dimensional input matrix is ​​convolved using a convolution kernel to extract implicit associated features in the matrix. In particular, in order to better capture the relationship between features, mutually transposed convolution kernels are used here, which can perform convolution operations in different directions, thereby better revealing the mutual relationship between data and the correlation of features, so as to provide more fine-grained feature information and improve the accuracy and effectiveness of operational analysis.

[0059] Correspondingly, in a specific example, the association feature extraction module 130 is used to: use the adjacent layers of the convolutional neural network model using mutually transposed convolution kernels to perform the following on the input data in the forward pass of the layer: perform convolution processing, pooling processing and nonlinear processing based on a first two-dimensional convolution kernel on the input data to obtain a first activation feature map; perform convolution processing, pooling processing and nonlinear activation processing based on a second two-dimensional convolution kernel on the first activation feature map to obtain a second activation feature map, wherein the first two-dimensional convolution kernel and the second two-dimensional convolution kernel are transposed of each other; wherein the input of the first layer of the convolutional neural network model of the adjacent layers using mutually transposed convolution kernels is the two-dimensional input matrix, and the output of the last layer of the convolutional neural network model of the adjacent layers using mutually transposed convolution kernels is the association feature matrix.

[0060] In the above-mentioned enterprise operation analysis early warning control system 100, the feature enhancement module 140 is used to obtain a classification feature matrix by passing the associated feature matrix through a two-way attention mechanism model. It should be understood that the two-way attention mechanism is a technology widely used in natural language processing and sequence modeling tasks, which can dynamically allocate attention to different positions in the input sequence to capture key information and perform weighted aggregation. That is, by using a two-way attention mechanism to select and aggregate key features from the associated feature matrix to extract more important features and reduce redundant information. In addition, the two-way attention mechanism model simultaneously considers the positive and negative correlations between the features in the matrix. By calculating the positive and negative attention weights of the features and other features respectively, the relationship between the features is comprehensively modeled. Then, the associated feature matrix is ​​weighted according to the attention weights to generate a classification feature matrix.

[0061] Accordingly, in a specific example, the feature enhancement module 140 includes: a pooling unit, used to pool the correlation feature matrix along the horizontal direction and the vertical direction respectively to obtain a first-direction pooling vector and a second-direction pooling vector; an association unit, used to perform association encoding on the first-direction pooling vector and the second-direction pooling vector to obtain a bidirectional correlation matrix; a weight calculation unit, used to input the bidirectional correlation matrix into a Sigmoid activation function to obtain a bidirectional correlation weight matrix; and a weight application unit, used to calculate the positional point multiplication between the bidirectional correlation weight matrix and the correlation feature matrix to obtain the classification feature matrix.

[0062] In the above-mentioned enterprise operation analysis early warning control system 100, the optimization module 150 is used to perform low-dimensional mask sparsification on the classification feature matrix to obtain an optimized classification feature matrix. In particular, considering that the classification feature matrix may contain a large number of features, some of which may contribute less to the results of the classification task or even be noise. Through low-dimensional mask sparsification, the weights of those features that contribute less to the classification results can be eliminated or reduced, thereby reducing the dimension of the feature space. Doing so helps to improve the efficiency of the classifier and reduce computational overhead. At the same time, too many features may lead to overfitting problems, that is, the model performs well on training data, but has poor generalization ability on new data. Through low-dimensional mask sparsification, features that are more important and relevant to the classification task can be screened out, reducing the degree of fitting of the model to noise and irrelevant features, thereby reducing the risk of overfitting. And the classification feature matrix optimized by low-dimensional mask sparsification pays more attention to important features and can better capture key information in the data. This can improve the performance and accuracy of the classifier, enabling it to better distinguish different categories.

[0063] Specifically, the optimization module 150 includes: a matrix decomposition unit, which is used to perform matrix decomposition based on eigenvalues ​​on the classification feature matrix to obtain multiple eigenvalues ​​and multiple eigenvectors corresponding to the multiple eigenvalues; an affine subspace construction unit, which is used to arrange the multiple eigenvectors in descending order based on the size of the eigenvalues, and extract the eigenvectors corresponding to the K largest eigenvalues ​​from the multiple eigenvectors to form a K-dimensional affine subspace; a mapping unit, which is used to map the classification feature matrix to the affine subspace to obtain an affine density domain matrix; a probabilization unit, which is used to input the affine density domain matrix into a Softmax activation function to obtain a probabilistic affine density domain matrix; a masking unit, which is used to mask the probabilistic affine density domain matrix based on the comparison between the eigenvalues ​​of each position in the probabilistic affine density domain matrix and a predetermined threshold value to obtain a masked probabilistic affine density domain matrix; and an optimization unit, which is used to calculate the positional point multiplication between the masked probabilistic affine density domain matrix and the classification feature matrix to obtain the optimized classification feature matrix.

[0064] It should be understood that the classification feature matrix is ​​probabilized in the affine density domain of the feature space to perform non-parametric clustering based on the density domain on the classification feature matrix, so that the clustering structure in the data distribution of the classification feature matrix can be automatically discovered without specifying the number of clusters. Specifically, in the technical solution of the present application, the affine transformation in the feature space is used to map the classification feature matrix to a low-dimensional affine subspace, and then the logistic activation function is used in the space to estimate and calculate the probability density of each data point to obtain a probabilistic affine density domain matrix, and then by setting a probability threshold, the high-density area can be divided into clusters and the low-density area is regarded as noise or anomaly to obtain a masked probabilistic affine density domain matrix, and the masked probabilistic affine density domain matrix is ​​applied to the classification feature matrix to achieve optimal modulation of the feature distribution. In this way, high-dimensional data and complex distribution problems can be effectively handled because it does not need to assume that the data obeys a certain specific distribution, nor does it need to predetermine the parameters or structure of the model; at the same time, the robustness of the data can be improved because it can adaptively adjust the affine subspace to adapt to data changes and the influence of noise.

[0065] In the above-mentioned enterprise operation analysis and early warning control system 100, the analysis result generation module 160 is used to pass the optimized classification feature matrix through a classifier to obtain a classification result, and the classification result is used to indicate whether the current operation status of the enterprise is excellent, normal or at risk. The classifier can learn the association between the input features and the predetermined categories to perform classification or prediction tasks. Here, the trained classifier can map the classification feature matrix to the categories pre-set to represent the current operation status of the enterprise, namely "excellent", "normal" and "at risk". In this way, the risks or problems that the enterprise may face can be discovered in a timely manner, and at the same time, it also provides important instructions on the current operation status for enterprise managers, which helps to take corresponding measures as soon as possible, improve the operation status, and optimize business decisions.

[0066] Figure 5 FIG. 1 is a block diagram of an analysis result generation module in an enterprise operation analysis and early warning control system according to an embodiment of the present application. Figure 5 As shown, the analysis result generation module 160 includes: a matrix expansion unit 161, which is used to expand the optimized classification feature matrix into a classification feature vector according to a row vector or a column vector; a fully connected encoding unit 162, which is used to use the fully connected layer of the classifier to fully connect the classification feature vector to obtain a fully connected encoded feature vector; a probability acquisition unit 163, which is used to pass the fully connected encoded feature vector through the Softmax classification function of the classifier to obtain a first probability belonging to the excellent current operating status of the enterprise, a second probability belonging to the normal current operating status of the enterprise, and a third probability belonging to the risk of the current operating status of the enterprise; a classification unit 164, which is used to determine the classification label corresponding to the largest of the probability values ​​as the classification result.

[0067] In summary, the enterprise operation analysis and early warning control system according to the embodiment of the present application is explained, which first determines the enterprise operation analysis indicators, obtains current data, historical data and industry average data according to the analysis indicators, uses the context encoder to fully understand the semantics of the three, and judges the current operation status of the enterprise based on the correlation characteristics of the three. In this way, the reliability of enterprise operation analysis and the effect of risk warning are improved, thereby helping enterprise decision makers to better understand the enterprise operation status and improve the accuracy and efficiency of decision-making.

[0068] Exemplary Methods

[0069] Figure 6 Flow chart of the enterprise operation analysis early warning control method according to the embodiment of the present application. Figure 6As shown, according to the enterprise operation analysis early warning control method of the embodiment of the present application, the steps include: S110, determining the enterprise operation analysis index, and obtaining current data, historical data and industry average data according to the enterprise operation analysis index; S120, respectively passing the text description of the current data, the text description of the historical data and the text description of the industry average data through a converter-based context encoder to obtain an enterprise operation status semantic feature vector, an enterprise historical operation status semantic feature vector and an industry standard semantic feature vector; S130, arranging the enterprise operation status semantic feature vector, the enterprise historical operation status semantic feature vector and the industry standard semantic feature vector into a two-dimensional input matrix and then passing through a convolutional neural network model with mutually transposed convolution kernels in adjacent layers to obtain an associated feature matrix; S140, passing the associated feature matrix through a bidirectional attention mechanism model to obtain a classification feature matrix; S150, performing low-dimensional mask sparsification on the classification feature matrix to obtain an optimized classification feature matrix; S160, passing the optimized classification feature matrix through a classifier to obtain a classification result, and the classification result is used to indicate that the current operation status of the enterprise is excellent, normal or at risk.

[0070] In a specific example, in the above-mentioned enterprise operation analysis, early warning and control method, the step S120, passes the text description of the current data through a converter-based context encoder to obtain a semantic feature vector of the enterprise operation status, including: performing word segmentation processing on the text description of the current data to obtain a current data word sequence; using the embedding layer of the context encoder to embed code each current data word in the current data word sequence to obtain a sequence of current data word embedding vectors; inputting the sequence of current data word embedding vectors into the converter module of the context encoder to obtain multiple current data feature vectors; and cascading the multiple current data feature vectors to obtain the semantic feature vector of the enterprise operation status.

[0071] In a specific example, in the above-mentioned enterprise operation analysis and early warning control method, the sequence of the current data word embedding vectors is input into the converter module of the context encoder to obtain multiple current data feature vectors, including: arranging the sequence of the current data word embedding vectors as an input vector; converting the input vector into a query vector and a key vector respectively through a learnable embedding matrix; calculating the product between the query vector and the transposed vector of the key vector to obtain a self-attention association matrix; normalizing the self-attention association matrix to obtain a standardized self-attention association matrix; inputting the standardized self-attention association matrix into a Softmax activation function for activation to obtain a self-attention feature matrix; multiplying the self-attention feature matrix with each current data word embedding vector in the sequence of the current data word embedding vectors to obtain the multiple current data feature vectors.

[0072] In a specific example, in the above-mentioned enterprise operation analysis and early warning control method, the step S130, after arranging the enterprise operation status semantic feature vector, the enterprise historical operation status semantic feature vector and the industry standard semantic feature vector into a two-dimensional input matrix, obtains a correlation feature matrix through a convolutional neural network model using mutually transposed convolution kernels in adjacent layers, including: using the convolutional neural network model using mutually transposed convolution kernels in the adjacent layers, respectively performing the following on the input data in the forward pass of the layer: performing convolution processing, pooling processing and nonlinear processing based on a first two-dimensional convolution kernel on the input data to obtain a first activation feature map; performing convolution processing, pooling processing and nonlinear activation processing based on a second two-dimensional convolution kernel on the first activation feature map to obtain a second activation feature map, wherein the first two-dimensional convolution kernel and the second two-dimensional convolution kernel are transposed of each other; wherein the input of the first layer of the convolutional neural network model using mutually transposed convolution kernels in the adjacent layers is the two-dimensional input matrix, and the output of the last layer of the convolutional neural network model using mutually transposed convolution kernels in the adjacent layers is the correlation feature matrix.

[0073] In a specific example, in the above-mentioned enterprise operation analysis, early warning and control method, the step S140, passes the correlation feature matrix through a bidirectional attention mechanism model to obtain a classification feature matrix, including: pooling the correlation feature matrix along the horizontal direction and the vertical direction respectively to obtain a first-direction pooling vector and a second-direction pooling vector; associatively encoding the first-direction pooling vector and the second-direction pooling vector to obtain a bidirectional correlation matrix; inputting the bidirectional correlation matrix into a Sigmoid activation function to obtain a bidirectional correlation weight matrix; calculating the positional point multiplication between the bidirectional correlation weight matrix and the correlation feature matrix to obtain the classification feature matrix.

[0074] In a specific example, in the above-mentioned enterprise operation analysis and early warning control method, the step S150 performs low-dimensional masked sparsification on the classification feature matrix to obtain an optimized classification feature matrix, including: performing matrix decomposition based on eigenvalues ​​on the classification feature matrix to obtain multiple eigenvalues ​​and multiple eigenvectors corresponding to the multiple eigenvalues; arranging the multiple eigenvectors in descending order based on the size of the eigenvalues, and extracting the eigenvectors corresponding to the K largest eigenvalues ​​from the multiple eigenvectors to form a K-dimensional affine subspace; mapping the classification feature matrix to the affine subspace to obtain an affine density domain matrix; inputting the affine density domain matrix into the Softmax activation function to obtain a probabilistic affine density domain matrix; based on the comparison between the eigenvalues ​​of each position in the probabilistic affine density domain matrix and a predetermined threshold, masking the probabilistic affine density domain matrix to obtain a masked probabilistic affine density domain matrix; calculating the positional point multiplication between the masked probabilistic affine density domain matrix and the classification feature matrix to obtain the optimized classification feature matrix.

[0075] In a specific example, in the above-mentioned enterprise operation analysis and early warning control method, the step S160, passes the optimized classification feature matrix through a classifier to obtain a classification result, and the classification result is used to indicate that the current operating status of the enterprise is excellent, normal or at risk, including: expanding the optimized classification feature matrix into a classification feature vector according to a row vector or a column vector; using the fully connected layer of the classifier to fully connect encode the classification feature vector to obtain a fully connected encoded feature vector; passing the fully connected encoded feature vector through the Softmax classification function of the classifier to obtain a first probability belonging to the excellent current operating status of the enterprise, a second probability belonging to the normal current operating status of the enterprise, and a third probability belonging to the risk of the current operating status of the enterprise; and determining the classification label corresponding to the largest of the probability values ​​as the classification result.

[0076] Here, those skilled in the art can understand that the specific operations of each step in the above-mentioned enterprise operation analysis early warning control method have been referred to above. Figures 1 to 5 It has been introduced in detail in the description of the enterprise operation analysis and early warning control system, and therefore, its repeated description will be omitted.

Claims

1. An enterprise operation analysis and early warning control system, characterized in that: include: An information acquisition module is used to determine enterprise operation analysis indicators and obtain current data, historical data and industry average data based on the enterprise operation analysis indicators; A semantic understanding module, for respectively passing the text description of the current data, the text description of the historical data and the text description of the industry average data through a converter-based context encoder to obtain an enterprise operation status semantic feature vector, an enterprise historical operation status semantic feature vector and an industry standard semantic feature vector; A correlation feature extraction module, used to arrange the enterprise operation status semantic feature vector, the enterprise historical operation status semantic feature vector and the industry standard semantic feature vector into a two-dimensional input matrix and then pass it through a convolutional neural network model with mutually transposed convolution kernels in adjacent layers to obtain a correlation feature matrix; A feature enhancement module, used for passing the correlation feature matrix through a bidirectional attention mechanism model to obtain a classification feature matrix; An optimization module, used for performing low-dimensional mask sparseness on the classification feature matrix to obtain an optimized classification feature matrix; The analysis result generation module is used to pass the optimized classification feature matrix through a classifier to obtain a classification result, and the classification result is used to indicate whether the current operating status of the enterprise is excellent, normal or at risk.

2. The enterprise operation analysis and early warning control system according to claim 1 is characterized in that: The semantic understanding module comprises: A word segmentation unit, used for performing word segmentation processing on the text description of the current data to obtain a current data word sequence; An embedding coding unit, configured to use the embedding layer of the context encoder to respectively perform embedding coding on each current data word in the current data word sequence to obtain a sequence of current data word embedding vectors; a context encoding unit, configured to input the sequence of current data word embedding vectors into a converter module of the context encoder to obtain a plurality of current data feature vectors; A cascading unit is used to cascade the multiple current data feature vectors to obtain the enterprise operation status semantic feature vector.

3. The enterprise operation analysis and early warning control system according to claim 2 is characterized in that: The context encoding unit comprises: an arrangement subunit, configured to arrange the sequence of the current data word embedding vectors into an input vector; A conversion subunit, used to convert the input vector into a query vector and a key vector respectively through a learnable embedding matrix; A self-attention association matrix generation subunit, configured to calculate the product between the query vector and the transposed vector of the key vector to obtain a self-attention association matrix; A standardized self-attention subunit, used for normalizing the self-attention association matrix to obtain a standardized self-attention association matrix; An attention calculation subunit, used for inputting the standardized self-attention association matrix into a Softmax activation function for activation to obtain a self-attention feature matrix; An attention applying subunit is used to multiply the self-attention feature matrix with each current data word embedding vector in the sequence of current data word embedding vectors to obtain the multiple current data feature vectors.

4. The enterprise operation analysis and early warning control system according to claim 3 is characterized in that: The associated feature extraction module is used to: use each layer of the convolutional neural network model using mutually transposed convolution kernels in the adjacent layers to perform the following on the input data in the forward transfer of the layer: Performing convolution processing, pooling processing and nonlinear processing based on a first two-dimensional convolution kernel on the input data to obtain a first activation feature map; Performing convolution processing, pooling processing, and nonlinear activation processing based on a second two-dimensional convolution kernel on the first activation feature map to obtain a second activation feature map, wherein the first two-dimensional convolution kernel and the second two-dimensional convolution kernel are transposed of each other; Among them, the input of the first layer of the convolutional neural network model in which adjacent layers use mutually transposed convolution kernels is the two-dimensional input matrix, and the output of the last layer of the convolutional neural network model in which adjacent layers use mutually transposed convolution kernels is the correlation feature matrix.

5. The enterprise operation analysis and early warning control system according to claim 4 is characterized in that: The feature enhancement module comprises: A pooling unit, used for pooling the correlation feature matrix along the horizontal direction and the vertical direction respectively to obtain a first-direction pooling vector and a second-direction pooling vector; an association unit, configured to perform association coding on the first-direction pooled vector and the second-direction pooled vector to obtain a bidirectional association matrix; A weight calculation unit, used for inputting the bidirectional association matrix into a Sigmoid activation function to obtain a bidirectional association weight matrix; A weight applying unit is used to calculate the position point multiplication between the bidirectional association weight matrix and the association feature matrix to obtain the classification feature matrix.

6. The enterprise operation analysis and early warning control system according to claim 5 is characterized in that: The optimization module comprises: A matrix decomposition unit, used for performing eigenvalue-based matrix decomposition on the classification feature matrix to obtain a plurality of eigenvalues ​​and a plurality of eigenvectors corresponding to the plurality of eigenvalues; An affine subspace construction unit, used to arrange the multiple eigenvectors in descending order based on the size of the eigenvalues, and extract eigenvectors corresponding to K largest eigenvalues ​​from the multiple eigenvectors to form a K-dimensional affine subspace; A mapping unit, used for mapping the classification feature matrix to the affine subspace to obtain an affine density domain matrix; A probabilization unit, used for inputting the affine density domain matrix into a Softmax activation function to obtain a probabilistic affine density domain matrix; A masking unit, configured to mask the probabilistic affine density domain matrix based on a comparison between the eigenvalues ​​of each position in the probabilistic affine density domain matrix and a predetermined threshold value to obtain a masked probabilistic affine density domain matrix; The optimization unit is used to calculate the position point multiplication between the masked probabilistic affine density domain matrix and the classification feature matrix to obtain the optimized classification feature matrix.

7. The enterprise operation analysis and early warning control system according to claim 6 is characterized in that: The analysis result generating module comprises: A matrix expansion unit, used for expanding the optimized classification feature matrix into a classification feature vector according to row vectors or column vectors; A fully connected encoding unit, configured to perform fully connected encoding on the classification feature vector using the fully connected layer of the classifier to obtain a fully connected encoded feature vector; A probability obtaining unit, used for passing the fully connected encoded feature vector through the Softmax classification function of the classifier to obtain a first probability that the current operating state of the enterprise is excellent, a second probability that the current operating state of the enterprise is normal, and a third probability that the current operating state of the enterprise is at risk; A classification unit is used to determine the classification label corresponding to the largest probability value as the classification result.

8. A method for early warning control of enterprise operation analysis, characterized in that: include: Determine enterprise operation analysis indicators, and obtain current data, historical data and industry average data based on the enterprise operation analysis indicators; The text description of the current data, the text description of the historical data and the text description of the industry average data are respectively passed through a converter-based context encoder to obtain an enterprise operation status semantic feature vector, an enterprise historical operation status semantic feature vector and an industry standard semantic feature vector; Arranging the enterprise operation status semantic feature vector, the enterprise historical operation status semantic feature vector and the industry standard semantic feature vector into a two-dimensional input matrix and then passing through a convolutional neural network model with mutually transposed convolution kernels in adjacent layers to obtain a correlation feature matrix; Pass the correlation feature matrix through a bidirectional attention mechanism model to obtain a classification feature matrix; Performing low-dimensional mask sparseness on the classification feature matrix to obtain an optimized classification feature matrix; The optimized classification feature matrix is ​​passed through a classifier to obtain a classification result, and the classification result is used to indicate whether the current operating status of the enterprise is excellent, normal, or at risk.

9. The enterprise operation analysis early warning control method according to claim 8 is characterized in that: The text description of the current data is passed through a converter-based context encoder to obtain a semantic feature vector of the enterprise operation status, including: Performing word segmentation processing on the text description of the current data to obtain a current data word sequence; Using the embedding layer of the context encoder, respectively, embedding encodes each current data word in the current data word sequence to obtain a sequence of current data word embedding vectors; Inputting the sequence of current data word embedding vectors into a converter module of the context encoder to obtain a plurality of current data feature vectors; The multiple current data feature vectors are cascaded to obtain the enterprise operation status semantic feature vector.

10. The enterprise operation analysis early warning control method according to claim 9, characterized in that: Inputting the sequence of the current data word embedding vectors into the converter module of the context encoder to obtain a plurality of current data feature vectors, including: Arranging the sequence of the current data word embedding vectors as an input vector; Convert the input vector into a query vector and a key vector respectively through a learnable embedding matrix; Calculating the product between the query vector and the transposed vector of the key vector to obtain a self-attention association matrix; Normalizing the self-attention association matrix to obtain a standardized self-attention association matrix; Input the standardized self-attention association matrix into a Softmax activation function for activation to obtain a self-attention feature matrix; The self-attention feature matrix is ​​multiplied by each current data word embedding vector in the sequence of current data word embedding vectors to obtain the multiple current data feature vectors.

Images