A data modeling method based on deep neural networks

Abstract

This invention discloses a data modeling method based on deep neural networks, comprising: S1, obtaining a data modeling sample set; S2, calculating the correlation metric between sample vectors based on the data modeling sample set, and generating a correlation metric matrix; S3, generating a structure threshold sequence based on the correlation metric matrix, and constructing a corresponding structure complex set; S4, calculating the homology structure description result for the structure complex set, and converting it into network connection constraint information; S5, constructing an improved homology neural network model, and limiting the neuron connection set in the model based on the constraint information; S6, under the condition of shared network parameters, driving the model to perform forward computation based on different constraint information to obtain multiple modeling results; S7, filtering the multiple modeling results, determining and outputting the target modeling result. This invention effectively improves the accuracy and robustness of modeling.

Description

Technical Field

[0001] This invention relates to the field of deep neural networks and structured data modeling technology, and in particular to a data modeling method based on deep neural networks. Background Technology

[0002] With the improvement of large-scale data acquisition and computing capabilities, data modeling methods based on deep neural networks have been widely applied in fields such as financial analysis, industrial monitoring, intelligent manufacturing, and prediction of complex systems. Existing technologies typically construct multi-layer neural networks to perform feature mapping and nonlinear fitting on sample data in order to model the inherent laws of the data. These methods have certain advantages in handling high-dimensional data and complex nonlinear relationships, but their modeling process mainly relies on the numerical features themselves, and the network structure often adopts a fixed connection method, lacking an explicit characterization of the overall structural relationship of the data.

[0003] To enhance the ability of models to express complex data relationships, some studies have introduced graph structures or structured data processing methods. These methods construct graph models based on the similarity or distance relationships between samples and then perform graph neural network modeling on this basis. However, these methods typically construct sample relationship structures only at a single scale or with a fixed threshold, making it difficult to simultaneously reflect the structural characteristics of data at different levels of association strength. Furthermore, once the graph structure is constructed, it remains unchanged, making it difficult to effectively link with subsequent modeling processes.

[0004] In recent years, topological data analysis methods have begun to be used for data structure analysis, extracting connectivity and higher-order relation features of data through complex construction and homology computation. However, existing techniques mostly use homology results as additional feature inputs or statistical indicators, failing to directly apply the structure analysis results to the computational structure of deep neural networks themselves. This results in structural information being independent of the network modeling process, making it difficult to reflect the structural constraints of the data at the network level.

[0005] Furthermore, existing deep neural network modeling methods, when faced with complex data, typically employ a single network structure and a single forward propagation path to output modeling results. When multiple reasonable expressions of the data structure or different structural assumptions exist, the model struggles to effectively compare and select the best approach. The modeling results are highly sensitive to the choice of network structure, exhibiting insufficient stability and consistency. While multi-model ensembles or multi-network parallel schemes can alleviate these problems to some extent, they often require the introduction of multiple sets of parameters or model instances, resulting in high computational costs and complex structures.

[0006] Therefore, how to provide a data modeling method based on deep neural networks is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0007] One objective of this invention is to propose a data modeling method based on deep neural networks. This invention constructs an improved homology neural network model by introducing a connection constraint mechanism driven by structural complexes and homology structures. The structural relationships between samples are mapped to neural network connection constraints. Under shared parameter conditions, multiple structurally constrained forward computations are performed, and the differences of multiple sets of modeling results are screened and determined. This achieves deep coupling between data structure information and network computation structure, and improves the stability and consistency of data modeling results without increasing the scale of model parameters.

[0008] A data modeling method based on a deep neural network according to an embodiment of the present invention includes the following steps: S1. Obtain a data modeling sample set, which contains multiple sample vectors, and each sample vector is composed of multiple feature dimensions. S2. Calculate the correlation metric between sample vectors based on the data modeling sample set, and generate a correlation metric matrix; S3. Generate a structure threshold sequence based on the correlation metric matrix, and construct a corresponding structure complex set based on the structure threshold sequence; S4. Calculate the homology structure description results for the set of structural complexes respectively, and convert the homology structure description results into network connection constraint information; S5. Construct an improved homology neural network model, and limit the neuron connection set in the improved homology neural network model according to the network connection constraint information; S6. Under the condition of shared network parameters, the improved homology neural network model is driven to perform forward calculations based on different network connection constraint information to obtain multiple modeling results; S7. Filter the multiple modeling results, determine the target modeling result, and output the target modeling result.

[0009] Optionally, S1 includes: S11. Obtain the original data source, which contains multiple original sample records, and the original sample records consist of numerical data, discrete data, or a combination of both. S12. Perform data cleaning on the original sample records, remove sample records with missing key fields, compare the sample records with abnormal values ​​with the preset value range, and delete the sample records that do not meet the comparison conditions. S13. Perform feature parsing on the cleaned original sample records, and parse each original sample record into a sample vector composed of multiple feature components. Each feature component corresponds to a data dimension in the original sample record. S14. Perform numerical mapping processing on discrete feature components, wherein the numerical mapping processing converts discrete values ​​into numerical representations based on a preset mapping table. S15. Perform numerical scaling on each feature component in the sample vector and linearly transform the numerical range of each feature component. S16. Combine the sample vectors that have completed numerical scaling according to the sample index order to form a data modeling sample set.

[0010] Optionally, S2 includes: S21. Determine the number of feature dimensions for each sample vector in the data modeling sample set; S22. According to the predetermined feature grouping rules, the feature dimension of each sample vector is divided into multiple feature subsets, and each feature subset contains at least one feature dimension. S23. For any two sample vectors in the sample set, calculate the numerical difference between the corresponding feature components one by one in the same feature subset, and perform subset aggregation processing on the numerical difference in the same feature subset to obtain the subset difference. S24. Perform weighted combination processing on the subset differences corresponding to each feature subset according to the preset weights to generate the combined difference between sample vectors; S25. Calculate the correlation metric between sample vectors based on the combined difference; S26. Arrange the correlation metrics between the sample vectors in the data modeling sample set according to the index order of the sample vectors to form a correlation metric matrix.

[0011] Optionally, S3 includes: S31. Obtain the correlation metric matrix and determine the value range of each correlation metric value in the correlation metric matrix; S32. Generate an ordered structural threshold sequence based on the value range, wherein the structural threshold sequence contains multiple thresholds and the thresholds are arranged in numerical order. S33. For each threshold in the structural threshold sequence, select sample index pairs corresponding to the correlation metric values ​​that satisfy the preset threshold comparison rules from the correlation metric matrix; S34. Based on the sample index pairs, construct a node set and an edge set to form a corresponding structure graph; S35. Based on the structure diagram, construct a higher-order simplex according to the predetermined simplex generation rules to form the corresponding structural complex. S36. Collect the structural complexes generated from each threshold in the structural threshold sequence to form a set of structural complexes.

[0012] Optionally, S4 includes: S41. For the set of structural complexes, select each structural complex one by one, and perform homology calculation processing on the zero-dimensional simplex, one-dimensional simplex and simplex above one dimension in each structural complex to obtain the number of connected branches, the number of loop structures and the identifier of higher-order connected structures in the corresponding dimensions, and form a homology structure description result corresponding one-to-one with each structural complex. S42. According to the homology calculation of the corresponding simplex dimension, the homology structure description result is divided into multiple dimension levels, including the connected branch level based on the zero-dimensional simplex, the ring structure level based on the one-dimensional simplex, and the higher-order connected level based on the simplex above one dimension, and the structure descriptor result corresponding to each dimension level is generated respectively. S43. Based on the simplex connectivity reflected in the structural descriptor results, convert the structural descriptor results into binary or multi-valued structural codes; S44. Generate a network connection constraint matrix based on the structure encoding, wherein the row and column indices of the network connection constraint matrix correspond one-to-one with the indices of adjacent neurons in the neural network. S45. Determine the neuron index pairs that are allowed to establish connections in the neural network based on the network connection constraint matrix to form network connection constraint information.

[0013] Optionally, S5 includes: S51. Construct an improved coherent neural network model, wherein the improved coherent neural network model includes multiple sequentially arranged network layers, and each network layer contains multiple neurons with unique index identifiers; S52. For any two adjacent network layers in the improved coherent neural network model, generate an initial connection matrix in fully connected form according to the neuron index order. Each element of the initial connection matrix corresponds to a set of cross-layer neuron index pairs. S53. Obtain the network connection constraint matrix, and align the network connection constraint matrix with the initial connection matrix according to the neuron index mapping relationship; S54. After completing the position alignment, according to the structure encoding value in the network connection constraint matrix, perform element-by-element filtering on the corresponding matrix elements in the initial connection matrix to generate a structure-constrained connection matrix. S55. For the structure-restricted connection matrix, extract the set of allowed neuron index pairs according to the network layer order, and write the set of neuron index pairs into the network structure definition.

[0014] Optionally, S6 includes: S61. Obtain the improved coherent neural network model and lock the network parameter set of the improved coherent neural network model; S62. Obtain multiple sets of network connection constraint matrices and assign a corresponding structural identifier to each network connection constraint matrix. S63. Without changing the set of network parameters, according to the order of the structure identifiers, write each network connection constraint matrix into the connection structure description of the improved homology neural network model in sequence to form multiple structurally constrained computational configurations. S64. For each structure-constrained computational configuration, perform forward propagation computation based on the same input sample vector to generate a modeling result that corresponds one-to-one with the structure identifier. S65. The modeling results are associated with and stored with the corresponding structural identifiers to form a set of structured modeling results.

[0015] Optionally, S7 includes: S71. Obtain a set of structured modeling results, wherein the set of structured modeling results includes multiple modeling results obtained for the same input sample vector and a structure identifier corresponding to each modeling result. S72. For multiple modeling results, calculate the numerical difference between different modeling results one by one according to the position order of the corresponding elements in the output vector, and perform aggregation processing on the numerical difference to obtain the corresponding result difference value. S73. Sort the result difference values ​​corresponding to each modeling result and generate a result difference sorting list; S74. Based on the result difference sorting list, select the modeling results whose result difference values ​​are within the preset sorting range, and determine the corresponding structure identifiers as candidate structure identifiers. S75. Select a modeling result from the modeling results corresponding to the candidate structure identifier according to the preset selection rules as the target modeling result, and output the target modeling result.

[0016] The beneficial effects of this invention are: This invention introduces a connection constraint mechanism driven by structural complexes and homology structures into the deep neural network modeling process, constructing an improved homology neural network model. Addressing the problem that complex data exhibits multi-level and multi-structured sample relationships that are difficult to fully represent in traditional fixed network structures, this invention first generates a structural threshold sequence based on the correlation metric between sample vectors. At different threshold levels, sets of structural complexes are constructed, allowing for a systematic characterization of sample relationships across multiple structural scales. Subsequently, homology calculations are performed on each structural complex to extract structural descriptions such as connected components, loop structures, and higher-order connectivity relationships. These are then processed hierarchically according to the simplex dimension. Through structural encoding and index mapping, the homology structure information at the data level is deterministically converted into a connection constraint matrix for the neural network. During network construction, the set of neuron connections is constrained, forming an improved homology neural network model with structural constraint characteristics. In the model computation process, under the condition of shared network parameters, this invention sequentially loads multiple sets of connection constraint structures and performs multiple forward propagation calculations to obtain multiple sets of modeling results driven by different structural assumptions. At the result level, the numerical differences between the modeling results are calculated, sorted, and filtered to ultimately determine the target modeling result. The above technical solution achieves a close integration of homology structure analysis and neural network computation structure, enabling the improved homology neural network model to effectively utilize the overall structural information of the data without increasing the parameter scale, thereby improving the stability, consistency and repeatability of the modeling results, and making it suitable for high-reliability modeling scenarios with complex structural data. Attached Figure Description

[0017] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings: Figure 1 This is a schematic diagram of the overall process of a data modeling method based on deep neural networks proposed in this invention; Figure 2 This is a schematic diagram of the structure of the improved coherent neural network model in this invention; Figure 3 This is a schematic diagram of generating a structural threshold sequence and constructing a structural complex set based on the correlation metric matrix in this invention. Detailed Implementation

[0018] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0019] refer to Figure 1-3 A data modeling method based on deep neural networks includes the following steps: S1. Obtain a data modeling sample set, which contains multiple sample vectors, and each sample vector is composed of multiple feature dimensions. S2. Calculate the correlation metric between sample vectors based on the data modeling sample set, and generate a correlation metric matrix; S3. Generate a structure threshold sequence based on the correlation metric matrix, and construct a corresponding structure complex set based on the structure threshold sequence; S4. Calculate the homology structure description results for the set of structural complexes respectively, and convert the homology structure description results into network connection constraint information; S5. Construct an improved homology neural network model, and limit the neuron connection set in the improved homology neural network model according to the network connection constraint information; S6. Under the condition of shared network parameters, the improved homology neural network model is driven to perform forward calculations based on different network connection constraint information to obtain multiple modeling results; S7. Filter the multiple modeling results, determine the target modeling result, and output the target modeling result.

[0020] In this embodiment, S1 includes: S11. Obtain the original data source, which contains multiple original sample records, and the original sample records consist of numerical data, discrete data, or a combination of both. S12. Perform data cleaning on the original sample records, remove sample records with missing key fields, compare the sample records with abnormal values ​​with the preset value range, and delete the sample records that do not meet the comparison conditions. S13. Perform feature parsing on the cleaned original sample records, and parse each original sample record into a sample vector composed of multiple feature components. Each feature component corresponds to a data dimension in the original sample record. S14. Perform numerical mapping processing on discrete feature components, wherein the numerical mapping processing converts discrete values ​​into numerical representations based on a preset mapping table. S15. Perform numerical scaling on each feature component in the sample vector and linearly transform the numerical range of each feature component. S16. Combine the sample vectors that have completed numerical scaling according to the sample index order to form a data modeling sample set.

[0021] In this invention, before entering the subsequent modeling process, the original sample records are first transformed into a sample vector form with a consistent structure through a unified data cleaning and feature parsing step. Discrete features are converted into deterministic numerical codes through a pre-stored mapping table, so that different types of features can be expressed in the same numerical space. Then, a linear scaling transformation is performed on each feature component to make the sample vectors comparable in numerical distribution.

[0022] In this embodiment, S2 includes: S21. Determine the number of feature dimensions for each sample vector in the data modeling sample set; S22. According to the predetermined feature grouping rules, the feature dimension of each sample vector is divided into multiple feature subsets, and each feature subset contains at least one feature dimension. S23. For any two sample vectors in the sample set, calculate the numerical difference between the corresponding feature components one by one in the same feature subset, and perform subset aggregation processing on the numerical difference in the same feature subset to obtain the subset difference. S24. Perform weighted combination processing on the subset differences corresponding to each feature subset according to the preset weights to generate the combined difference between sample vectors; S25. Calculate the correlation metric between sample vectors based on the combined difference; S26. Arrange the correlation metrics between the sample vectors in the data modeling sample set according to the index order of the sample vectors to form a correlation metric matrix.

[0023] In this invention, the feature grouping rules are pre-defined based on the semantic category, data source, or statistical distribution of each feature dimension in the sample vector, ensuring that the feature components within the same feature subset maintain consistency in data structure. For the difference calculation between sample vectors, the difference aggregation is first performed within the feature subset, and then the aggregation results of different feature subsets are combined according to their corresponding weights, thereby introducing hierarchical structural constraints in the correlation metric calculation process. This calculation process forms a stable correlation expression at the sample level, ensuring that the resulting correlation metric matrix has consistent characteristics in both numerical distribution and structural hierarchy, and maintaining data connectivity with the subsequent construction processes of the structural threshold sequence and structural complex.

[0024] In this embodiment, S3 includes: S31. Obtain the correlation metric matrix and determine the value range of each correlation metric value in the correlation metric matrix; S32. Generate an ordered structural threshold sequence based on the value range, wherein the structural threshold sequence contains multiple thresholds and the thresholds are arranged in numerical order. This invention performs statistical analysis on all correlation metric values ​​in the correlation metric matrix to determine the minimum and maximum values ​​of the correlation metric values. Based on the value interval formed by the minimum and maximum values, this interval is divided into several sub-intervals, and the boundary values ​​of each sub-interval are used as structural thresholds to generate an ordered sequence of structural thresholds. For example, when the minimum correlation metric value is 0.12 and the maximum value is 0.85, the interval can be divided into 6 sub-intervals, corresponding to the generated structural threshold sequence {0.12, 0.27, 0.42, 0.57, 0.71, 0.85}. S33. For each threshold in the structural threshold sequence, select sample index pairs corresponding to the correlation metric values ​​that satisfy the preset threshold comparison rules from the correlation metric matrix; The structural threshold sequence is determined based on the quantiles of the correlation metric values ​​in the correlation metric matrix. Specifically, the 10%, 30%, 50%, 70%, and 90% quantiles of the correlation metric value distribution can be selected as structural thresholds to generate a structural threshold sequence that satisfies a monotonically increasing relationship. For example, the corresponding structural thresholds could be {0.18, 0.33, 0.49, 0.64, 0.79}. S34. Based on the sample index pairs, construct a node set and an edge set to form a corresponding structure graph; S35. Based on the structure diagram, construct a higher-order simplex according to the predetermined simplex generation rules to form the corresponding structural complex. S36. Collect the structural complexes generated from each threshold in the structural threshold sequence to form a set of structural complexes.

[0025] In this invention, the structural threshold sequence is not a single threshold setting, but rather an ordered set of thresholds generated based on the value distribution of the correlation metric matrix. This allows sample relationships to form multiple structural expressions at different threshold levels. Within each threshold level, a basic structural graph is constructed using sample index pairs, and further expanded into a structural complex containing higher-order relationships according to predetermined simplex generation rules. This results in a set of structural complexes with hierarchical differences under a unified computational process. This construction method extends structural information from simple pairwise sample relationships to multi-sample combination relationships, and completes the structured representation at the data level, providing a stable and repeatable structural input foundation for subsequent homology structure calculations and network connection constraints.

[0026] In this embodiment, S4 includes: S41. For the set of structural complexes, select each structural complex one by one, and perform homology calculation processing on the zero-dimensional simplex, one-dimensional simplex and simplex above one dimension in each structural complex to obtain the number of connected branches, the number of loop structures and the identifier of higher-order connected structures in the corresponding dimensions, and form a homology structure description result corresponding one-to-one with each structural complex. S42. According to the homology calculation of the corresponding simplex dimension, the homology structure description result is divided into multiple dimension levels, including the connected branch level based on the zero-dimensional simplex, the ring structure level based on the one-dimensional simplex, and the higher-order connected level based on the simplex above one dimension, and the structure descriptor result corresponding to each dimension level is generated respectively. S43. Based on the simplex connectivity reflected in the structural descriptor results, convert the structural descriptor results into binary or multi-valued structural codes; S44. Generate a network connection constraint matrix based on the structure encoding, wherein the row and column indices of the network connection constraint matrix correspond one-to-one with the indices of adjacent neurons in the neural network. S45. Determine the neuron index pairs that are allowed to establish connections in the neural network based on the network connection constraint matrix to form network connection constraint information.

[0027] In this invention, the structural description results obtained from homology computation are not directly applied to the neural network. Instead, they are first decomposed into multiple levels of structural descriptor results according to the simplex dimension, allowing the sample relationships reflected by different dimensions to be distinguished from each other at the data level. Subsequently, based on the connection relationships between simplexes in each level, the structural descriptor results are converted into discrete structural codes and further mapped into connection constraint matrices consistent with the indices of neurons in adjacent layers of the neural network. Through this mapping process, high-order structural relationships at the sample level are deterministically converted into connection constraints in the neural network computational structure. Thus, without changing the form of network parameters, the structural information of the data is introduced into the network connection layers, forming a stable and repeatable computational constraint mechanism.

[0028] In this embodiment, S5 includes: S51. Construct an improved coherent neural network model, wherein the improved coherent neural network model includes multiple sequentially arranged network layers, and each network layer contains multiple neurons with unique index identifiers; S52. For any two adjacent network layers in the improved coherent neural network model, generate an initial connection matrix in fully connected form according to the neuron index order. Each element of the initial connection matrix corresponds to a set of cross-layer neuron index pairs. S53. Obtain the network connection constraint matrix, and align the network connection constraint matrix with the initial connection matrix according to the neuron index mapping relationship; S54. After completing the position alignment, according to the structure encoding value in the network connection constraint matrix, perform element-by-element filtering on the corresponding matrix elements in the initial connection matrix to generate a structure-constrained connection matrix. S55. For the structure-restricted connection matrix, extract the set of allowed neuron index pairs according to the network layer order, and write the set of neuron index pairs into the network structure definition.

[0029] In this invention, the connections in the neural network are not dynamically adjusted during training, but determined through a matrix-based approach during network construction. Specifically, an initial connection matrix based on neuron indices is first generated for adjacent network layers. Then, the connection constraint matrix obtained from the homology structure mapping is aligned with this initial connection matrix at the index level, and the allowed cross-layer connections are determined through element-wise filtering. The filtered neuron index pairs are directly written into the network structure description, forming a fixed connection topology. This construction method allows the network's computational structure to be induced by the data structure, resulting in deterministic and repeatable connections, thus introducing homology information at the structural level.

[0030] In this embodiment, S6 includes: S61. Obtain the improved coherent neural network model and lock the network parameter set of the improved coherent neural network model; S62. Obtain multiple sets of network connection constraint matrices and assign a corresponding structural identifier to each network connection constraint matrix. S63. Without changing the set of network parameters, according to the order of the structure identifiers, write each network connection constraint matrix into the connection structure description of the improved homology neural network model in sequence to form multiple structurally constrained computational configurations. S64. For each structure-constrained computational configuration, perform forward propagation computation based on the same input sample vector to generate a modeling result that corresponds one-to-one with the structure identifier. S65. The modeling results are associated with and stored with the corresponding structural identifiers to form a set of structured modeling results.

[0031] In this invention, multiple sets of network connection constraint matrices are distinguished by structural identifiers and, while keeping the same set of network parameters unchanged, are sequentially loaded into the connection structure description of the coherent neural network, enabling the network to perform forward propagation computations under different connection topology configurations. Each forward propagation is based on the same input sample vector, only changing the allowed connection index pairs in the network, thereby forming multiple computational paths distinguished by structural differences within a unified parameter space. By associating and storing each modeling result with its corresponding structural identifier, the mapping relationship between the modeling results and their structural sources can be preserved at the data level, providing a consistent data foundation for subsequent result selection and structural analysis.

[0032] In this embodiment, S7 includes: S71. Obtain a set of structured modeling results, wherein the set of structured modeling results includes multiple modeling results obtained for the same input sample vector and a structure identifier corresponding to each modeling result. S72. For multiple modeling results, calculate the numerical difference between different modeling results one by one according to the position order of the corresponding elements in the output vector, and perform aggregation processing on the numerical difference to obtain the corresponding result difference value. S73. Sort the result difference values ​​corresponding to each modeling result and generate a result difference sorting list; S74. Based on the result difference sorting list, select the modeling results whose result difference values ​​are within the preset sorting range, and determine the corresponding structure identifiers as candidate structure identifiers. S75. Select a modeling result from the modeling results corresponding to the candidate structure identifier according to the preset selection rules as the target modeling result, and output the target modeling result.

[0033] In this invention, multiple sets of modeling results obtained for the same input sample vector are uniformly organized into structured data, and each modeling result is associated with its corresponding structural identifier. The calculation of result differences is performed at the output vector level. By comparing the numerical differences of corresponding elements in different modeling results item by item and performing aggregation processing, a difference value that can be directly sorted is formed. After the difference sorting is completed, the range of candidate results is determined according to the pre-set sorting interval. This makes the determination process of the target modeling result based on the numerical comparison between multiple structural calculation paths, thereby maintaining a consistent mapping relationship with the structural identifier at the result level and forming a stable and repeatable result determination process.

[0034] Example 1: To verify the feasibility and effectiveness of this invention in practical applications, the data modeling method based on deep neural networks proposed in this invention was applied to a scenario of modeling the operating status of industrial equipment. In this scenario, multiple devices on the production line continuously collect a large amount of operating data during long-term operation, including multi-dimensional features such as current, voltage, vibration amplitude, temperature, and rotational speed. Because the operating status of the equipment is affected by various factors such as load changes, environmental conditions, and component wear, the data collected at different time periods show significant differences in numerical distribution and structural relationships. Traditional modeling methods based on fixed network structures are prone to problems such as large fluctuations and insufficient stability in modeling results in this type of scenario.

[0035] In this embodiment, raw equipment operation data is first acquired continuously from the industrial control system. The raw data is recorded in time slices, with each record containing 18 numerical features and 6 discrete status indicators. The raw data is then cleaned, removing records with missing key fields or obviously abnormal values. The discrete status indicators are converted into corresponding numerical codes using a preset mapping table. Subsequently, a linear scaling transformation is performed on each feature component to ensure that features of different dimensions are expressed within a unified numerical range, ultimately forming a data modeling sample set for modeling purposes.

[0036] After the sample set is constructed, a correlation metric matrix is ​​calculated based on the feature differences between sample vectors. Specifically, the feature dimensions of the sample vectors are divided into multiple feature subsets according to electrical, mechanical, and environmental features. Within the same feature subset, the numerical differences between samples are calculated and subset aggregation is performed. Then, a weighted combination method is used to generate combined differences between sample vectors, thus obtaining the complete correlation metric matrix. Based on this correlation metric matrix, an ordered sequence of structure thresholds is generated. Corresponding structure graphs are constructed at different threshold levels, and a set of structure complexes containing higher-order simplexes is further generated.

[0037] For the constructed set of structural complexes, homology calculations are performed on each complex to obtain the number of connected components corresponding to the zero-dimensional simplex, the number of loop structures corresponding to the one-dimensional simplex, and the identifiers of higher-order connected structures corresponding to simplexes with dimensions greater than one. These homology structure descriptions are then divided into multiple levels according to the simplex dimension and converted into structural codes to generate a network connection constraint matrix. The generated network connection constraint matrix corresponds one-to-one with the indices of adjacent neurons in the neural network, limiting the allowed connections between neurons.

[0038] Based on this, an improved cohomological neural network model is constructed. In the initialization phase, an initial connection matrix in fully connected form is generated. Then, based on the network connection constraint matrix, the connections are filtered element-by-element, retaining only neuron connections that satisfy the structural constraints, thus forming a network topology with structural constraint characteristics. After the model parameters are determined, while keeping the network parameter set unchanged, multiple sets of network connection constraint matrices generated corresponding to different structural complexes are sequentially loaded. This drives the improved cohomological neural network model to perform forward propagation calculations under various structurally constrained configurations, obtaining multiple sets of modeling results.

[0039] For multiple sets of modeling results generated from the same input sample vector, the numerical differences between different modeling results are calculated according to the positional order of corresponding elements in the output vector. These numerical differences are then aggregated to obtain the result difference values. After sorting the result difference values, modeling results whose difference values ​​fall within a preset sorting interval are selected as candidate results. The target modeling result is then determined from the candidate results as the final output.

[0040] To verify the beneficial effects of this invention, a comparative experiment was conducted between the method of this invention and the traditional fixed-structure deep neural network method under the same dataset and network parameter scale. The experiment used equipment operation data from 30 consecutive time windows as test samples, and statistical analysis was performed on modeling error, result stability, and structural consistency indicators. The experimental results are shown in Table 1. Table 1. Comparison of modeling effects of different modeling methods in industrial equipment operation data scenarios.

[0041] As shown in Table 1, the improved coherent neural network model proposed in this invention exhibits significant advantages in several key indicators. Regarding the average modeling error, the method of this invention reduces it from 0.087 for traditional fixed-structure neural networks to 0.052, indicating a significant improvement in the model's fitting accuracy to the equipment's operating state under the same parameter scale. Simultaneously, the standard deviation of the modeling error decreases from 0.031 to 0.014, indicating that the model output is more stable across different time windows and less affected by data fluctuations and structural changes. In terms of the fluctuation amplitude of results within continuous time windows, the method of this invention has a fluctuation of only 7.9%, significantly lower than the 18.6% of the comparative method, reflecting higher consistency in the modeling results under multi-structure constraints. Furthermore, the consistency of multi-structure modeling results improves from 0.68 to 0.91, further verifying that introducing structural constraints and performing result screening under shared parameter conditions can effectively reduce the uncertainty caused by structure selection and improve overall modeling reliability.

[0042] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A data modeling method based on deep neural networks, characterized in that, Includes the following steps: S1. Obtain a data modeling sample set, which contains multiple sample vectors, and each sample vector is composed of multiple feature dimensions. S2. Calculate the correlation metric between sample vectors based on the data modeling sample set, and generate a correlation metric matrix; S3. Generate a structure threshold sequence based on the correlation metric matrix, and construct a corresponding structure complex set based on the structure threshold sequence; S4. Calculate the homology structure description results for the set of structural complexes respectively, and convert the homology structure description results into network connection constraint information; S5. Construct an improved homology neural network model, and limit the neuron connection set in the improved homology neural network model according to the network connection constraint information; S6. Under the condition of shared network parameters, the improved homology neural network model is driven to perform forward calculations based on different network connection constraint information to obtain multiple modeling results; S7. Filter the multiple modeling results, determine the target modeling result, and output the target modeling result.

2. The data modeling method based on deep neural networks according to claim 1, characterized in that, S1 includes: S11. Obtain the original data source, which contains multiple original sample records, and the original sample records consist of numerical data, discrete data, or a combination of both. S12. Perform data cleaning on the original sample records, remove sample records with missing key fields, compare the sample records with abnormal values ​​with the preset value range, and delete the sample records that do not meet the comparison conditions. S13. Perform feature parsing on the cleaned original sample records, and parse each original sample record into a sample vector composed of multiple feature components. Each feature component corresponds to a data dimension in the original sample record. S14. Perform numerical mapping processing on discrete feature components, wherein the numerical mapping processing converts discrete values ​​into numerical representations based on a preset mapping table. S15. Perform numerical scaling on each feature component in the sample vector and linearly transform the numerical range of each feature component. S16. Combine the sample vectors that have completed numerical scaling according to the sample index order to form a data modeling sample set.

3. The data modeling method based on deep neural networks according to claim 1, characterized in that, S2 includes: S21. Determine the number of feature dimensions for each sample vector in the data modeling sample set; S22. According to the predetermined feature grouping rules, the feature dimension of each sample vector is divided into multiple feature subsets, and each feature subset contains at least one feature dimension. S23. For any two sample vectors in the sample set, calculate the numerical difference between the corresponding feature components one by one in the same feature subset, and perform subset aggregation processing on the numerical difference in the same feature subset to obtain the subset difference. S24. Perform weighted combination processing on the subset differences corresponding to each feature subset according to the preset weights to generate the combined difference between sample vectors; S25. Calculate the correlation metric between sample vectors based on the combined difference; S26. Arrange the correlation metrics between the sample vectors in the data modeling sample set according to the index order of the sample vectors to form a correlation metric matrix.

4. The data modeling method based on deep neural networks according to claim 1, characterized in that, S3 includes: S31. Obtain the correlation metric matrix and determine the value range of each correlation metric value in the correlation metric matrix; S32. Generate an ordered structural threshold sequence based on the value range, wherein the structural threshold sequence contains multiple thresholds and the thresholds are arranged in numerical order. S33. For each threshold in the structural threshold sequence, select sample index pairs corresponding to the correlation metric values ​​that satisfy the preset threshold comparison rules from the correlation metric matrix; S34. Based on the sample index pairs, construct a node set and an edge set to form a corresponding structure graph; S35. Based on the structure diagram, construct a higher-order simplex according to the predetermined simplex generation rules to form the corresponding structural complex. S36. Collect the structural complexes generated from each threshold in the structural threshold sequence to form a set of structural complexes.

5. The data modeling method based on deep neural networks according to claim 1, characterized in that, S4 includes: S41. For the set of structural complexes, select each structural complex one by one, and perform homology calculation processing on the zero-dimensional simplex, one-dimensional simplex and simplex above one dimension in each structural complex to obtain the number of connected branches, the number of loop structures and the identifier of higher-order connected structures in the corresponding dimensions, and form a homology structure description result corresponding one-to-one with each structural complex. S42. According to the homology calculation of the corresponding simplex dimension, the homology structure description result is divided into multiple dimension levels, including the connected branch level based on the zero-dimensional simplex, the ring structure level based on the one-dimensional simplex, and the higher-order connected level based on the simplex above one dimension, and the structure descriptor result corresponding to each dimension level is generated respectively. S43. Based on the simplex connectivity reflected in the structural descriptor results, convert the structural descriptor results into binary or multi-valued structural codes; S44. Generate a network connection constraint matrix based on the structure encoding, wherein the row and column indices of the network connection constraint matrix correspond one-to-one with the indices of adjacent neurons in the neural network. S45. Determine the neuron index pairs that are allowed to establish connections in the neural network based on the network connection constraint matrix to form network connection constraint information.

6. The data modeling method based on deep neural networks according to claim 1, characterized in that, S5 includes: S51. Construct an improved coherent neural network model, wherein the improved coherent neural network model includes multiple sequentially arranged network layers, and each network layer contains multiple neurons with unique index identifiers; S52. For any two adjacent network layers in the improved coherent neural network model, generate an initial connection matrix in fully connected form according to the neuron index order. Each element of the initial connection matrix corresponds to a set of cross-layer neuron index pairs. S53. Obtain the network connection constraint matrix, and align the network connection constraint matrix with the initial connection matrix according to the neuron index mapping relationship; S54. After completing the position alignment, according to the structure encoding value in the network connection constraint matrix, perform element-by-element filtering on the corresponding matrix elements in the initial connection matrix to generate a structure-constrained connection matrix. S55. For the structure-restricted connection matrix, extract the set of allowed neuron index pairs according to the network layer order, and write the set of neuron index pairs into the network structure definition.

7. The data modeling method based on deep neural networks according to claim 1, characterized in that, S6 includes: S61. Obtain the improved coherent neural network model and lock the network parameter set of the improved coherent neural network model; S62. Obtain multiple sets of network connection constraint matrices and assign a corresponding structural identifier to each network connection constraint matrix. S63. Without changing the set of network parameters, according to the order of the structure identifiers, write each network connection constraint matrix into the connection structure description of the improved homology neural network model in sequence to form multiple structurally constrained computational configurations. S64. For each structure-constrained computational configuration, perform forward propagation computation based on the same input sample vector to generate a modeling result that corresponds one-to-one with the structure identifier. S65. The modeling results are associated with and stored with the corresponding structural identifiers to form a set of structured modeling results.

8. The data modeling method based on deep neural networks according to claim 1, characterized in that, S7 includes: S71. Obtain a set of structured modeling results, wherein the set of structured modeling results includes multiple modeling results obtained for the same input sample vector and a structure identifier corresponding to each modeling result. S72. For multiple modeling results, calculate the numerical difference between different modeling results one by one according to the position order of the corresponding elements in the output vector, and perform aggregation processing on the numerical difference to obtain the corresponding result difference value. S73. Sort the result difference values ​​corresponding to each modeling result and generate a result difference sorting list; S74. Based on the result difference sorting list, select the modeling results whose result difference values ​​are within the preset sorting range, and determine the corresponding structure identifiers as candidate structure identifiers. S75. Select a modeling result from the modeling results corresponding to the candidate structure identifier according to the preset selection rules as the target modeling result, and output the target modeling result.

Images