Enterprise management method and system based on big data

Abstract

The present invention relates to the technical field of enterprise management methods, and more specifically, to an enterprise management method based on big data, comprising the following steps: step 1: data collection and preprocessing; step 2: cross-modal data fusion; step 3: feature compression representation, using a gated semantic encoder g to adaptively encode fused semantics Z and extract key semantic features C; using lossless compression based on dictionary learning to compress C to obtain a dictionary D and a sparse code S, alternately optimizing the above objective function, fixing D to optimize S, and then fixing S to optimize D until convergence; step 4: business insight and risk analysis, based on cross-modal fusion semantics and key feature compression representation, constructing multi-view clustering, time series prediction, anomaly detection and association analysis machine learning models; comprehensively analyzing business status, risk events and influencing factors to form global business insights; and step 5: management decision optimization.

Description

Technical Field

[0001] The present invention relates to the technical field of enterprise management methods, and more specifically, to an enterprise management method and system based on big data. Background Art

[0002] With the rapid development of information technology and the diversification of data collection methods, the data available to enterprises is becoming increasingly large and complex. This includes both structured business data, such as transaction records and customer information, and unstructured data, such as text, images, audio and video. These multi-source heterogeneous data contain all aspects of enterprise operations. If they can be effectively integrated and analyzed, they will be of great benefit to the refined management and scientific decision-making of enterprises.

[0003] However, traditional data processing methods mainly target structured data and are difficult to effectively respond to the needs of fusion analysis of heterogeneous data. Although academia and industry have conducted research on multimodal data fusion, existing technologies still have many limitations. On the one hand, most methods use simple feature splicing or linear combination strategies, which makes it difficult to mine deep semantic associations between different modalities. For example, the invention patent with the publication number CN117494050A, entitled A Multimodal Data Fusion Method Based on Big Data Governance, proposes a multimodal data fusion algorithm that first extracts feature vectors from each modal data and then fuses them by weighted summation. Although this method is simple and intuitive, it does not consider the nonlinear semantic interaction between modalities, and the actual performance improvement is limited. On the other hand, some fusion methods introduce deep learning models, but they often require a large amount of labeled data for training, lack interpretability, and have insufficient generalization capabilities.

[0004] Existing feature selection and compression methods are also difficult to directly apply to multimodal semantic representation. Traditional feature selection is mostly based on predefined statistical criteria, such as variance, correlation coefficient, etc., which makes it difficult to characterize the importance of semantic features. Mainstream data compression technologies, such as principal component analysis (PCA), usually assume that the data follows a Gaussian distribution and are based on linear transformations, making it difficult to deal with nonlinear and non-Gaussian semantic features. The existing two-stage PCA algorithm for multimodal data compression first performs PCA dimensionality reduction on each modality data separately, and then performs cascade PCA on the reduced features. Although this strategy considers the correlation within and between modalities to a certain extent, it is still linear and does not consider the complementarity at the semantic level. Therefore, how to adaptively extract the most critical semantic features from the fused semantic representation and achieve lossless compression is an urgent problem to be solved.

[0005] In summary, the existing multimodal fusion and processing technologies are still insufficient in the following aspects: 1) lack of deep cross-modal semantic representation and interaction; 2) lack of adaptive key feature extraction mechanism for fusion semantics; 3) lack of lossless compression method for semantic features; 4) lack of efficient computing framework to match it. These problems limit the application of multimodal data analysis in enterprise management, and a new technical approach is urgently needed.

[0006] Therefore, there is an urgent need to provide enterprise management methods and systems based on big data to address the shortcomings of existing technologies. Summary of the invention

[0007] The purpose of the present invention is to provide an enterprise management method and system based on big data, a cross-modal semantic fusion method based on deep learning and attention mechanism, and on this basis, innovatively introduces gating mechanism and dictionary learning to achieve adaptive feature compression of fusion semantics.

[0008] The present invention provides an enterprise management method based on big data, comprising the following steps:

[0009] Step 1: Data collection and preprocessing

[0010] Acquire structured and unstructured data from internal business systems, external partners, and public data sources; clean, integrate, transform, and feature engineer data;

[0011] Step 2: Cross-modal data fusion

[0012] For N heterogeneous modal data X1, X2, ..., X N , through an independent deep encoding network f i Extract the respective semantic features E i The formula is:

[0013] E i =f i (X i θ i )

[0014] where i=1,2,...,N,θ i is the i-th modality encoding network f i The parameters of the semantic representation E are transformed using the transformation function T i Mapped to the common semantic space, the semantic representation is adaptively aggregated through the gated attention mechanism to obtain the cross-modal fusion semantics z, the formula is

[0015]

[0016] where α iis the attention weight of the i-th modality, generated by the attention network; end-to-end training semantic encoding network parameters {θ i} and attention network parameters, minimizing the loss function of the fused semantics z;

[0017] Step 3: Feature compression representation

[0018] The gated semantic encoder g is used to adaptively encode the fused semantics Z and extract the key semantic features C. The formula is:

[0019] G=σ(W g Z+b g )

[0020] H=W h Z+b h

[0021] C=G⊙H

[0022] Where G is the gate signal, ⊙ represents element-by-element multiplication, and W g ,W h ,b g ,b h is the gating layer parameter;

[0023] Use lossless compression based on dictionary learning to compress C, and obtain dictionary D and sparse code S. The objective function is:

[0024]

[0025] st∥D j ∥2≤1

[0026] Where ∥S∥1 is the L1 regularization term, λ is the balance coefficient, and ∥D j ∥2 is the L2 norm constraint of the dictionary atoms. The above objective functions are optimized alternately, D is fixed to optimize S, and then S is fixed to optimize D until convergence;

[0027] Step 4: Business Insight and Risk Analysis

[0028] Based on cross-modal fusion semantics and key feature compression representation, we build multi-view clustering, time series prediction, anomaly detection and association analysis machine learning models; we comprehensively analyze business status, risk events and influencing factors to form global business insights;

[0029] Step 5: Management decision optimization

[0030] Integrate business insights, risk analysis and expert knowledge to form differentiated management strategies, dynamic business plans, risk control measures and resource optimization allocation plans; feed back the optimization plans into business processes to form closed-loop management.

[0031] Specifically, the cross-modal data includes heterogeneous data types such as text, image, audio, and video.

[0032] Specifically, in the gated attention mechanism, the attention weight α i The calculation formula is:

[0033]

[0034] where q′W i ,b i are the parameters of the attention network.

[0035] Specifically, in the extraction of key semantic features, high-level semantic features are extracted layer by layer by superimposing L layers of gated semantic encoders.

[0036] Specifically, in the lossless compression, the following strategy is adopted when optimizing the dictionary D and the sparse code S:

[0037] Fixed D and optimized S using LARS algorithm;

[0038] Fix S and optimize D by projected gradient descent;

[0039] The two steps are iterated alternately until convergence.

[0040] Specifically, in the business insight and risk analysis, support vector machines, decision trees, and neural network machine learning algorithms are used in combination to build a prediction model.

[0041] Specifically, a heterogeneous computing platform for cross-modal semantic fusion and feature compression is constructed, including:

[0042] CPU, used for general computing and task scheduling;

[0043] GPU, used for training and reasoning of deep learning models;

[0044] FPGA, for hardware acceleration of computationally intensive operators;

[0045] Multi-level storage system for efficient data storage and access.

[0046] The enterprise management system based on the enterprise management method of big data includes:

[0047] Data collection and preprocessing module, used for the collection, cleaning, integration and feature extraction of heterogeneous data;

[0048] Cross-modal fusion module, used for semantic representation learning and attention fusion of heterogeneous modal data;

[0049] Feature compression module, used for adaptive extraction and lossless compression of key semantic features;

[0050] Business analysis module, used for business status clustering, trend prediction, anomaly detection and association analysis;

[0051] Decision optimization module, used for management strategy optimization, process reengineering and performance evaluation.

[0052] Specifically, the cross-modal fusion module includes a parallel deep semantic encoder and a gated attention network, which respectively perform semantic extraction and adaptive fusion on different modal data.

[0053] Specifically, the feature compression module includes a gated semantic encoder and a dictionary learning unit, which respectively realize key semantic adaptive extraction and lossless compression representation.

[0054] The present invention has the following beneficial effects:

[0055] First, in terms of data fusion, the system adopts a cross-modal semantic cascade strategy and realizes the integrated representation of multi-source heterogeneous data through deep learning. Traditional multimodal fusion is usually a simple splicing or linear combination, which makes it difficult to explore the deep semantic associations between modalities. However, this system first uses an independent encoder to extract the semantic features of each modality, and then uses the attention mechanism to adaptively fuse the semantics of different modalities. In this way, the complementarity of different modal data can be brought into play, and the semantic synergy between modalities can be captured. In addition, the end-to-end semantic encoding and attention network can promote each other to form accurate cross-modal semantic expression. Therefore, this fusion strategy can produce synergistic effects.

[0056] Secondly, in terms of feature learning, the system introduces gating mechanism and dictionary learning to achieve adaptive extraction and lossless compression of key semantics. Traditional feature selection is often predefined and linear, while this system adaptively focuses on the most critical semantic features through data-driven gating units. Multi-layer gating coding gives semantic features a hierarchical structure and greatly enhances the descriptive power. At the same time, dictionary learning can not only compress semantics losslessly, but also remove redundancy between features. It can be said that gating mechanism and dictionary learning play a "double insurance" role in feature selection and feature enhancement, which not only ensures the integrity of semantics, but also greatly compresses storage and computing overhead. The clever combination of the two has produced a multiplier effect.

[0057] Thirdly, this system is also very ingenious in optimizing and solving algorithms. Taking dictionary learning as an example, the optimization goal needs to minimize the reconstruction error, satisfy the dictionary atomic norm constraint, and make the sparse code as sparse as possible. This is actually a multi-objective optimization problem, and there are certain contradictions between different goals. The system cleverly decomposes it into two sub-problems: finding sparse codes with a fixed dictionary, and finding dictionaries with fixed sparse codes. The former sub-problem is solved efficiently using the LARS algorithm, and the latter sub-problem uses projected gradient descent to ensure constraint satisfaction. This optimization strategy alternates between the two sub-problems, strikes a balance between accuracy and efficiency, resolves the conflict of optimization goals, and embodies the wisdom of divide and conquer.

[0058] In addition, the heterogeneous computing platform constructed by this system also played a role in making a big difference. Faced with massive high-dimensional cross-modal data, the computational overhead of the algorithm cannot be underestimated. The traditional CPU computing mode is difficult to respond quickly. However, this platform specifically integrates CPU, GPU, FPGA and multi-level storage to form a multi-granular heterogeneous computing environment. GPU and FPGA arrays accelerate deep learning in parallel, multi-level storage solves the data IO bottleneck, and combined with CPU flexible scheduling, the system throughput can be greatly improved. More importantly, the heterogeneous platform and algorithm design can match each other to form a joint force of software and hardware collaborative optimization. For example, the semantic coding network can be trained directly on the GPU cluster, dictionary learning can use FPGA to accelerate matrix multiplication, and the critical path can be optimized with cache. These "adapt to local conditions" optimization measures allow limited computing resources to play the maximum efficiency and support the computing power requirements of big data analysis.

[0059] In summary, the system of the present invention is a complex of algorithms, models, hardware, and engineering that are intertwined with each other. Cross-modal semantic fusion realizes the synergistic efficiency of data dimensions, gated learning and dictionary learning realize the complementary enhancement of feature dimensions, and algorithm optimization and hardware acceleration provide strong guarantees from the execution dimension. It can be said that this system not only has major innovations in technology, but also can achieve a leapfrog improvement in performance. These unexpected synergies and cross-dimensional optimizations are the core value of the system, and also mark an important breakthrough in enterprise big data analysis technology. It is believed that this invention can provide the industry with new ideas, empower enterprise management decisions, and bring out greater potential in practice. BRIEF DESCRIPTION OF THE DRAWINGS

[0060] Figure 1 A flowchart of the enterprise management method based on big data of the present invention;

[0061] Figure 2 This is a framework diagram of the big data-based enterprise management system of the present invention. DETAILED DESCRIPTION

[0062] In order to further explain the technical means and effects adopted by the present invention to achieve the predetermined invention purpose, the enterprise management method based on big data proposed by the present invention, its specific implementation method, structure, characteristics and effects are described in detail below in combination with the accompanying drawings and preferred embodiments. In the following description, different "one embodiment" or "another embodiment" does not necessarily refer to the same embodiment. In addition, specific features, structures or characteristics in one or more embodiments may be combined in any suitable form.

[0063] Unless defined otherwise, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention belongs.

[0064] The specific scheme of the enterprise management method based on big data provided by the present invention is described in detail below with reference to the accompanying drawings.

[0065] Example 1

[0066] See also Figure 1 The present invention proposes a method for enterprise management based on big data. The method, as described in claim 1, includes five main steps: data collection and preprocessing, cross-modal data fusion, feature compression representation, business insight and risk analysis, and management decision optimization.

[0067] In the data collection and preprocessing of step 1, structured and unstructured data are obtained from various sources inside and outside the enterprise. Structured data usually exists in the form of databases, tables, etc., such as transaction records and customer information of business systems; unstructured data includes rich data forms such as text documents, images, audio, and video. The collected raw data is first cleaned to remove noise, outliers and duplicate data, process missing values, and improve data quality. Then, through data integration, data from different sources and formats are mapped to consistent data models and representations. After that, the data is normalized and standardized to eliminate dimensional and distribution differences. Finally, feature engineering is carried out to construct high-order statistical features, cross-combination features, etc., to explore the intrinsic feature representation of the data and provide a better foundation for subsequent fusion analysis.

[0068] The cross-modal data fusion in step 2 is a major innovation of the present invention. In reality, enterprise data usually contains multiple modalities, such as text, images, audio, video, etc. The semantic information contained in different modal data is both complementary and different. It is difficult for a single modality to fully represent the operating status of an enterprise. Therefore, the present invention proposes a cross-modal semantic cascade fusion method.

[0069] In the cross-modal data fusion of step 2, the core is to achieve semantic encoding and adaptive fusion of different modal data through deep learning and attention mechanism. Specifically, given N heterogeneous modal data X1, X2, ..., X N , for the i-th modal data X i , design an independent deep encoding network f i Perform semantic encoding on it and extract semantic features to represent E i The mathematical expression is:

[0070] E i =f i (X i θ i ),i=1,2,...,N

[0071] Among them, θ i is the i-th modality encoding network f i Parameters of the encoding network f i Mainstream deep learning architectures can be used, such as convolutional neural networks (CNN) to process image data, recurrent neural networks (RNN) to process sequence data, and transformers to process text data. Taking image data as an example, f i It can be designed as a stack of several convolutional layers and pooling layers. The convolution kernel size is generally 3x3 or 5x5. The maximum pooling is often used for pooling. The number of convolutional layers is generally 3-7. Through end-to-end training f i The key semantic features of each modality data can be adaptively extracted.

[0072] After obtaining the semantic encoding E1, E2, ..., E N After that, through the transformation function T i Map them into a common semantic space to achieve the alignment of different modal semantics. i It can be a simple linear transformation, such as matrix multiplication, or a nonlinear transformation, such as a multi-layer perceptron (MLP). After alignment, a gated attention mechanism is designed to adaptively fuse the semantics of different modalities to obtain the cross-modal fusion semantics Z. The fusion process is described by mathematical expression:

[0073]

[0074] where α i is the attention weight of the i-th modality, reflecting the contribution of this modality to the final semantic representation. i Generated by the attention network, the formula is:

[0075]

[0076] Where q,W i ,bi is the parameter of the attention network. The attention network can be a single-layer feedforward neural network, with the input being the semantic encoding of each modality and the output being the attention weight. i ,b i The dimension of E is generally consistent with the dimension of semantic encoding, such as i is a 256-dimensional feature vector, then W i Can be set to a 256x256 matrix, b i and q are 256-dimensional vectors. Through the attention network, the importance of different modal data can be automatically learned and given higher weights to achieve adaptive cross-modal semantic fusion. Finally, by end-to-end training of the entire network, including the semantic encoding network f i , attention network, and minimize the loss function of the fused semantic representation z, such as reconstruction error, classification error, etc. The choice of loss function depends on the downstream task. For example, if z is used to reconstruct the original multimodal data, the mean square error (MSE) can be used as the reconstruction error; if z is used for business scenario classification, the cross-entropy can be used as the classification error. Through gradient backpropagation and optimization algorithms (such as Adam), the semantic encoding network and the attention network can be optimized at the same time to achieve adaptive fusion of cross-modal semantics.

[0077] In the feature compression representation of step 3, the key semantic features are first extracted through the gated semantic encoder, and then lossless compression is performed using dictionary learning. Specifically, the gated semantic encoder takes the fused semantics Z generated in step 2 as input, adaptively selects key semantic information through the gated unit, and obtains the compressed semantic feature C. The mathematical expression is:

[0078] G=σ(W g Z+b g )

[0079] H=W h Z+b h

[0080] C=G⊙H

[0081] Among them, G is the gate signal, σ is the sigmoid activation function, and W g ,W h ,b g ,b h is the gating layer parameter, and ⊙ represents element-by-element multiplication. The gating signal G can adaptively adjust the importance of the semantic feature H, highlight the key semantics, and suppress the secondary semantics. When G is close to 1, the semantic features of the corresponding dimension are selected; when G is close to 0, the features of the corresponding dimension are ignored. By adjusting W g ,b g, can control the density of the gating signal. Generally speaking, W g Initialized to 0, b g Initialize to a large negative value such as -5, so that the initial gating signal is sparse, allowing only a small number of key

[0082] After extracting the key semantic feature C, it is further losslessly compressed through a dictionary learning algorithm to obtain a more compact feature representation. The goal of lossless compression is to represent the original semantic features with as little data as possible without losing information, greatly reducing the storage and computational overhead of the semantic features. Here, a dictionary learning model based on minimization of reconstruction error is used to reconstruct the semantic feature C as the product of the dictionary D and the sparse code S, and minimize the reconstruction error while controlling the complexity of the dictionary and sparse code. The mathematical expression is:

[0083]

[0084] st∥D j ∥2≤1

[0085] Among them, ∥·∥ F represents the Frobenius norm of the matrix, ∥·∥1 represents the L1 norm, and ∥·∥2 represents the L2 norm. λ is the balance coefficient, which controls the weight of sparsity. D j is the jth column of dictionary D, also called dictionary atom. ∥D j ∥2<=1 limits the L2 norm of dictionary atoms to no more than 1, avoiding excessive redundancy of the dictionary. The L1 norm constraint ∥S∥ of the sparse code S makes S as sparse as possible, and uses fewer dictionary atom combinations to reconstruct semantic features. Solving the above dictionary learning problem can be divided into two sub-problems: optimizing S with fixed D, and optimizing D with fixed S. For the first sub-problem, the least angle regression (LARS) algorithm can be used. It is a heuristic greedy algorithm that can quickly solve sparse codes under the condition of a fixed dictionary. For the second sub-problem, the projected gradient descent algorithm can be used to update the dictionary atoms alternately through gradient descent and projection steps to satisfy the L2 norm constraint. These two sub-problems are executed alternately until the reconstruction error converges or the maximum number of iterations is reached.

[0086] Common parameter settings are as follows: the number of dictionary atoms can be 1-2 times the dimension of semantic features, the sparsity weight λ can be adjusted between 0.1 and 1, and the number of iterations of the LARS algorithm and projected gradient descent is generally between 10 and 100. By adjusting these parameters, a balance can be achieved between compression rate and reconstruction quality.

[0087] Through cross-modal semantic encoding and adaptive fusion, a comprehensive and accurate representation of the enterprise status is obtained; through key semantic extraction and dictionary learning, lossless compression of semantic features is achieved without losing information. These two steps complement each other and together form the cornerstone of data-driven enterprise management analysis.

[0088] In step 4, business insight and risk analysis, based on cross-modal fusion semantics and compressed feature representation, machine learning models such as multi-view clustering, time series prediction, anomaly detection and association analysis are constructed to fully explore the business operation rules and risk hazards. Cluster analysis can reveal typical business models and scenarios; time series prediction can determine business development trends; anomaly detection can timely discover abnormal behaviors and risk events; association analysis can reveal risk transmission paths and influencing factors. These analyses make full use of the semantic information of cross-modal data to form global business insights.

[0089] Finally, in the management decision optimization of step 5, data-driven business insights, risk analysis and domain expert knowledge are combined to form a better management decision plan. Specifically, it includes differentiated management strategies, dynamically adjusted business plans, targeted risk control measures, and optimized resource allocation plans. By applying the decision plan to the actual business process, a closed-loop management from data analysis to decision optimization is formed, which can sustainably improve the business management level of the enterprise.

[0090] Considering the computational overhead of cross-modal semantic fusion and feature compression, the present invention also designs a heterogeneous computing platform based on CPU, GPU, FPGA and multi-level storage, which greatly improves system performance through hardware acceleration and parallel computing. Among them, the CPU is responsible for general computing and task scheduling, the GPU accelerates deep learning models, the FPGA customizes and optimizes computationally intensive operators such as matrix multiplication and convolution, and multi-level storage realizes efficient data IO. The coordinated optimization of software and hardware can give full play to the efficiency of heterogeneous computing resources.

[0091] Compared with the existing technology, the above method has the following innovations: 1) Cross-modal semantic cascading, which realizes the semantic representation and adaptive fusion of multimodal data through deep learning and attention mechanism; 2) Adaptive extraction of key semantics, which realizes adaptive compression of semantic features by using gating mechanism and multi-layer encoder; 3) Lossless compression of semantic features, which realizes compressed representation by using dictionary learning, and proposes efficient optimization strategy; 4) Soft-hard collaborative optimization, which builds a heterogeneous computing platform and improves system performance from the data layer, algorithm layer and hardware layer. These innovations effectively solve technical problems such as the difficulty of multimodal data fusion, semantic feature extraction and massive semantic representation, and form a highly competitive technical solution.

[0092] In order to verify the superiority of the present invention, the following embodiments and comparative examples are designed, and they are compared in terms of cross-modal semantic fusion, key feature extraction, semantic compression, computational efficiency, etc.

[0093] Embodiment 1: The cross-modal data management method of the present invention is adopted, including semantic fusion based on attention mechanism, gated adaptive feature selection, dictionary learning lossless compression, and heterogeneous computing optimization.

[0094] Comparative Example 1: Traditional multimodal data fusion method, using feature concatenation and linear combination, without considering deep semantic interaction. Feature selection is based on statistical variance, and data compression uses PCA. Use a single CPU for processing.

[0095] Comparative Example 2: A deep neural network fusion method based on multi-task learning achieves the correlation expression of different modalities by sharing hidden layers. L1 regularization is used for feature selection and compression. GPU is used to accelerate the calculation.

[0096] Comparative Example 3: First perform PCA dimensionality reduction on each modality data, and then perform cascade PCA on the reduced dimensionality features. Feature selection is based on mutual information, without considering the semantic level. Use CPU and GPU for collaborative computing.

[0097] In the experiment, the enterprise's structured business data and unstructured text and image data were used as input to evaluate the performance of different methods in tasks such as semantic fusion, feature compression, and anomaly detection. The specific evaluation indicators and experimental results are shown in the following table:

[0098] index Example 1 Comparative Example 1 Comparative Example 2 Comparative Example 3 Cross-modal fusion accuracy 95.2 76.4 88.6 82.1 Semantic Compression Rate 92.8 62.5 85.4 79.6 Key feature extraction F1 value 93.6 68.3 80.2 74.9 Anomaly Detection AUC Value 98.1 80.5 90.4 87.2 End-to-end computation time (s) 42.5 126.8 65.2 86.4

[0099] Table 1

[0100] The cross-modal fusion accuracy reflects the ability of different methods to integrate semantic information. The semantic relevance of manual annotations is used as a benchmark to calculate the consistency of fused semantics and annotation results. Example 1 achieved an accuracy of 95.2% on this indicator, which is significantly better than other methods. This is mainly due to the adaptive weighting of different modal semantics by the attention mechanism and the enhanced feature expression brought by end-to-end training.

[0101] The semantic compression rate measures the degree of information retention after semantic feature compression. The reconstruction error is used as the evaluation index to calculate the difference between features before and after compression. The semantic compression rate of Example 1 reaches 92.8%, which is significantly ahead of other methods. It can be seen that the combination of gate selection and dictionary learning can compress semantic features to the greatest extent without losing information.

[0102] The F1 value of key feature extraction examines the accuracy and completeness of different methods for selecting semantic features. Based on the manually annotated key features, the accuracy and recall of the extraction results are calculated and synthesized into the F1 value. The F1 value of Example 1 reaches 93.6%, achieving the best balance between accuracy and completeness. It can be seen that the gating mechanism can adaptively focus on the most critical semantic features.

[0103] The AUC value (Area Under Curve) of anomaly detection evaluates the ability of different methods to identify risk events. An anomaly detector is constructed based on fusion semantics, and the true positive rate and false positive rate are counted, the ROC curve is drawn and the AUC is calculated. The AUC value of Example 1 is as high as 98.1%, indicating that fusion semantics can comprehensively characterize the status of the enterprise and has a strong risk identification ability.

[0104] The end-to-end computing time directly examines the execution efficiency of different systems. It can be seen that Example 1 fully utilizes the advantages of the heterogeneous computing platform and controls the computing time to 42.5 seconds through the collaboration of CPU, GPU, FPGA and multi-level storage, which is much lower than other methods.

[0105] In summary, Example 1, the best example of the present invention, is significantly superior to the prior art in terms of cross-modal semantic fusion, key feature extraction, semantic compression, and computational efficiency. Compared with simple feature splicing, the present invention achieves deep interaction of different modal semantics through the attention mechanism; compared with linear compression, the present invention introduces gated selection and dictionary learning, and achieves significant compression gain on the basis of lossless expression; compared with simple end-to-end learning, the present invention combines adaptive extraction of key features with compressed representation, which not only fully retains the core semantics, but also greatly reduces the time and space overhead. At the same time, the present invention also maximizes the performance of computing resources through soft and hard collaborative optimization.

[0106] These experimental results fully demonstrate the superiority of the present invention. The present invention has been collaboratively designed at multiple levels, including data, features, algorithms, and systems, to form a complete cross-modal analysis solution. It can be foreseen that the present invention will have an important impact in both industry and academia, laying a technical foundation for a new generation of enterprise management and decision-making platforms. In the future, the present invention can also be expanded to more fields, such as smart cities, intelligent manufacturing, etc., with broad application prospects and economic value.

[0107] Example 2

[0108] See also Figure 2, the present invention also proposes an enterprise management system compatible with the above method. The system includes a data acquisition and preprocessing module 1, a cross-modal fusion module 2, a feature compression module 3, a business analysis module 4 and a decision optimization module 5, which correspond to the five steps of the method one by one. Among them, the data acquisition and preprocessing module 1 is used for the collection, cleaning, integration and feature extraction of heterogeneous data; the cross-modal fusion module 2 is used for semantic representation learning and attention fusion of heterogeneous modal data; the feature compression module 3 is used for adaptive extraction and lossless compression of key semantic features; the business analysis module 4 is used for business status clustering, trend prediction, anomaly detection and correlation analysis; the decision optimization module 5 is used for management strategy optimization, process reengineering and performance evaluation.

[0109] The cross-modal fusion module 2 includes a parallel deep semantic encoder and a gated attention network, which respectively perform semantic extraction and adaptive fusion on different modal data.

[0110] The feature compression module 3 includes a gated semantic encoder and a dictionary learning unit, which respectively realize key semantic adaptive extraction and lossless compression representation.

[0111] It should be noted that the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the principles of the present invention should be included in the protection scope of the present invention.

Claims

1. An enterprise management method based on big data, characterized in that: The following steps are involved: Step 1: Data collection and preprocessing Acquire structured and unstructured data from internal business systems, external partners, and public data sources; clean, integrate, transform, and feature engineer data; Step 2: Cross-modal data fusion For N heterogeneous modal data X1, X2, ..., X N , through an independent deep encoding network f i Extract the respective semantic features E i The formula is: E i =f i (X i ;θ i ) where i=1, 2,...,N,θ i is the i-th modality encoding network f i Parameters, X i is the i-th modal data, using the transformation function T i The semantic representation E i Mapped to the common semantic space, the semantic representation is adaptively aggregated through the gated attention mechanism to obtain the cross-modal fusion semantics z, the formula is: where α i is the attention weight of the i-th modality, generated by the attention network, T i represents the transformation function; end-to-end training semantic encoding network parameters {θ i } and attention network parameters, minimizing the loss function of the fused semantics z; Step 3: Feature compression representation The gated semantic encoder g is used to adaptively encode the fused semantics Z and extract the key semantic features C. The formula is: G=σ(W g Z+b g ) H=W h Z+b h C=G⊙H Where G is the gate signal, σ is the sigmoid activation function, ⊙ represents element-by-element multiplication, and W g , W h , b g , b h is the gating layer parameter; Use lossless compression based on dictionary learning to compress C, and obtain dictionary D and sparse code S. The objective function is: s.t.∥D j ∥2≤1 Among them ∥·∥ F represents the Frobenius norm of the matrix, ∥S∥1 is the L1 regularization term, λ is the balance coefficient, and D j represents the jth column of dictionary D, ∥D j ∥2 is the L2 norm constraint of the dictionary atoms. The above objective functions are optimized alternately, D is fixed to optimize S, and then S is fixed to optimize D until convergence; Step 4: Business Insight and Risk Analysis Based on cross-modal fusion semantics and key feature compression representation, we build multi-view clustering, time series prediction, anomaly detection and association analysis machine learning models; we comprehensively analyze business status, risk events and influencing factors to form global business insights; Step 5: Management decision optimization Integrate business insights, risk analysis and expert knowledge to form differentiated management strategies, dynamic business plans, risk control measures and resource optimization allocation plans; feed back the optimization plans into business processes to form closed-loop management.

2. The method according to claim 1, characterized in that ,The cross-modal data includes heterogeneous data types such as text, image, audio, and video.

3. The method according to claim 1, characterized in that In the gated attention mechanism, the attention weight α i The calculation formula is: where q′W i ,b i are the parameters of the attention network.

4. The method according to claim 1, characterized in that ,In the extraction of key semantic features, high-level semantic features are ,extracted layer by layer by superimposing L layers of gated ,semantic encoders.

5. The method according to claim 1, characterized in that ,In the lossless compression, the following strategy is adopted when optimizing the dictionary D and the sparse code S: Fixed D and optimized S using LARS algorithm; Fix S and optimize D by projected gradient descent; The two steps are iterated alternately until convergence.

6. The method according to claim 1, characterized in that ,In the business insight and risk analysis, support vector machine, ,decision tree, and neural network machine learning algorithms are ,comprehensively used to construct a prediction model.

7. The method according to any one of claims 1 to 6, characterized in that , build a heterogeneous computing platform for cross-modal semantic fusion and feature compression, including: CPU, used for general computing and task scheduling; GPU, used for training and reasoning of deep learning models; FPGA, for hardware acceleration of computationally intensive operators; Multi-level storage system for efficient data storage and access.

8. An enterprise management system based on the method according to any one of claims 1 to 6, characterized in that ,include: Data collection and preprocessing module, used for the collection, cleaning, integration and feature extraction of heterogeneous data; Cross-modal fusion module, used for semantic representation learning and attention fusion of heterogeneous modal data; Feature compression module, used for adaptive extraction and lossless compression of key semantic features; Business analysis module, used for business status clustering, trend prediction, anomaly detection and association analysis; Decision optimization module, used for management strategy optimization, process reengineering and performance evaluation.

9. The system according to claim 8, characterized in that ,The cross-modal fusion module includes a parallel deep semantic encoder and a gated attention ,network, which respectively perform semantic extraction and adaptive fusion of different modal data.

10. The system according to claim 8, characterized in that ,The feature compression module includes a gated semantic encoder and a dictionary learning unit, which realize ,key semantic adaptive extraction and lossless compression representation respectively.

Images