Incremental data fusion method for industrial internet based on granulation attribute reduction and sparse autoencoder
By employing granular attribute reduction and sparse autoencoders, the problems of high computational cost and long response latency in incremental data fusion in the Industrial Internet are solved, achieving efficient and accurate data fusion, which is applicable to scenarios such as equipment status monitoring and production scheduling.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-04-02
- Publication Date
- 2026-07-03
AI Technical Summary
Existing industrial internet data fusion methods suffer from high computational costs and long response times when dealing with incremental data environments. They also struggle to balance the deep correlation between temporal characteristics and textual descriptions of project information. Furthermore, traditional attribute reduction methods are ill-suited to adapting to dynamic shifts in data distribution, leading to a decline in model stability and accuracy.
We adopt a method based on granular attribute reduction and sparse autoencoders, which achieves efficient fusion of incremental data from the industrial Internet through dynamic attribute reduction, feature extraction from multi-layer fully connected networks, and unsupervised representation learning of sparse autoencoders, combined with an incremental update mechanism for knowledge rules.
It reduces computational burden, improves fusion accuracy and response latency, and maintains the real-time performance and stability of the model, making it suitable for applications such as equipment status monitoring, anomaly warning, and production scheduling.
Smart Images
Figure CN122333344A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of Internet of Things (IoT) technology, and more particularly to an incremental data fusion method for the Industrial Internet based on granular attribute reduction and sparse autoencoders. Background Technology
[0002] Data in industrial internet scenarios typically originates from multiple sources, including equipment sensors, PLC controllers, MES / ERP business systems, alarm logs, maintenance records, and work order texts. It is characterized by diverse sources, heterogeneous structures, inconsistent sampling frequencies, and significant quality fluctuations.
[0003] Existing industrial data fusion methods are mostly designed for single-batch, static, or homogeneous data. A common approach is to first perform simple feature concatenation, then use LSTM, matrix factorization, or traditional dimensionality reduction algorithms to complete the fusion representation learning. While these methods can achieve some success with small sample sizes or relatively stable scenarios, in the incremental data environment of the Industrial Internet, as new batches are continuously added, the number of data attributes, time span, and redundancy increase simultaneously. Models often need to repeatedly process historical data, leading to high computational costs, increased response latency, and difficulty in simultaneously addressing the deep correlation between temporal features and the descriptive features of textual descriptions. Furthermore, traditional attribute reduction methods mostly rely on fixed thresholds or static rules, making it difficult to adapt to the dynamic shifts in industrial field data distribution as operating conditions change. This can easily result in insufficient retention of key attributes or incomplete removal of redundant attributes, thus affecting the stability, accuracy, and real-time performance of subsequent fusion models.
[0004] Therefore, there is an urgent need for an incremental data fusion method that can address multi-source heterogeneous incremental data in the Industrial Internet, perform dynamic attribute reduction in the preprocessing stage, reduce temporal modeling overhead in the feature extraction stage, and achieve deep complementary representation learning in the fusion stage. Summary of the Invention
[0005] To address the shortcomings of the aforementioned background technologies, this invention proposes an improvement scheme to solve the core problems in incremental data fusion scenarios of the Industrial Internet, thereby overcoming the deficiencies of existing technologies. The technical solution adopted by this invention is as follows: A method for incremental data fusion in the industrial internet based on granular attribute reduction and sparse autoencoders is provided, including the following steps: S1. Dynamically reduce the attributes of incremental industrial internet data within the current time window based on granular and extensional decision-making; S2. The numerical sequence after dynamic attribute reduction is normalized and detrended and then input into a multi-layer fully connected network to extract text features and temporal features; S3. The text features and temporal features are concatenated to obtain the joint input, which is then input into the sparse autoencoder for unsupervised representation learning. Finally, the fused feature matrix is output and directly supplied to the upper-level industrial internet intelligent analysis platform. S4. When a new batch of incremental data is detected, the knowledge rule incremental update mechanism is triggered to update the affected information particles and local model parameters in the new data.
[0006] Preferably, the dynamic attribute reduction process includes, for any two industrial internet incremental data, and Its absolute correlation can be expressed as ,in The association function is constructed based on the differences in sample distribution and decision consistency. Greater than the preset threshold If the two attributes are highly redundant or not correlated enough in the current time window, the comprehensive retention score is calculated by combining the attribute weight and data quality coefficient, and the attribute with the higher score is retained in the reduction result set.
[0007] A better method for calculating the overall retained score is as follows: ; in, This indicates the overall retained score. Indicates attribute weight. This represents the data quality coefficient.
[0008] Preferably, the attribute reduction process also records the mapping relationship between deleted attributes and retained attributes, which is used to quickly locate the affected information particles when new data arrives, thereby reducing the search range during subsequent local updates.
[0009] Preferably, the detrending process involves the following steps: Let the original time series of a certain attribute be... The trend term is estimated using a sliding window or local regression. The detrended sequence is then... .
[0010] As a preferred approach, in a multi-layer fully connected network, for text description data, word segmentation, stop word filtering, and terminology normalization are performed first, then the processed text sequence is mapped into a word vector matrix, and semantic features are extracted through convolution and pooling operations.
[0011] Preferably, in the training process of unsupervised representation learning, the following objective function is used for training: ; The first term represents the reconstruction error, the second term represents the sparsity penalty term, the third term represents the weight regularization term, and N represents the sample weight. This represents the joint feature of the i-th sample input to the autoencoder. This represents the i-th reconstructed feature output after network decoding. Let KL represent the hyperparameter controlling the weights of the sparse penalty term, and let KL represent the divergence, the Kullback-Leibler divergence. This represents the target activation level of neurons in the hidden layer, as set manually. Let represent the actual average activation response of the j-th neuron in the hidden layer on the training set. This is represented by the regularization coefficient that controls weight decay. This represents the weight matrix parameters in the encoder and decoder.
[0012] Even better, after training, the hidden layer output That is, as the final fusion feature SAE uses an optimization algorithm to fuse multiple features, extracting complementary information between features while significantly reducing feature redundancy.
[0013] Preferably, the feature is that a fusion adjustment coefficient is constructed based on data quality indicators and data volume indicators to reduce data loss during incremental fusion, as follows: ; in, Indicates the fusion adjustment coefficient. Indicates adjustable weights. Indicates data quality indicators, Indicates the amount of data. This represents the total number of samples in the current time window.
[0014] Preferably, steps S1 to S4 are repeated at preset time intervals, and the fusion features output in each round are written into the feature cache or knowledge base to support subsequent model iteration, visualization analysis and cross-system calls.
[0015] Compared with the prior art, the present invention has the following significant advantages: Compared to existing fusion methods that rely on static feature concatenation or direct LSTM retraining, this invention first utilizes granular attribute reduction to eliminate weakly correlated and redundant attributes before data enters the fusion model, reducing input dimensionality and noise propagation risks. Then, it extracts temporal features through detrending processing and a fully connected network, reducing the computational burden caused by medium- to long-term dependencies. Furthermore, it employs a knowledge-granular incremental update mechanism, only updating locally in changed data regions, avoiding repeated calculations of all historical samples each time new data arrives. Finally, it uses a sparse autoencoder to complete deep feature fusion, ensuring sufficient preservation of complementary information between different modalities. Therefore, this invention demonstrates superior overall performance in terms of fusion accuracy, response latency, resource utilization, and continuous online update capabilities, making it particularly suitable for applications in the Industrial Internet context that require continuous processing of incremental data, such as equipment status monitoring, anomaly warning, quality analysis, and production scheduling. Attached Figure Description
[0016] Figure 1 This is a general framework diagram of the DIDF-GRSAE method in this embodiment of the invention; Figure 2 This is a flowchart illustrating the overall process of incremental data fusion in the industrial internet, as described in this embodiment of the invention. Figure 3 This is a flowchart of the deep feature fusion process based on SAE in an embodiment of the present invention; Figure 4 This is a flowchart of the multi-granularity data knowledge incremental dynamic update process in an embodiment of the present invention. Detailed Implementation
[0017] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.
[0018] To more clearly illustrate the specific implementation method of the present invention, the system modules, data processing flow, and key calculation processes involved in the present invention will be further described below with reference to the accompanying drawings. (Reference) Figure 1 As shown, in this embodiment, industrial internet field data can originate from equipment sensors, controller operation logs, work order texts, alarm records, and production process parameters. Due to the inconsistent sampling periods, significant differences in data dimensions, and varying semantic expressions among different data sources, the system first organizes the multi-source data according to a unified time window and establishes an object-attribute-decision ternary relationship within the window. Then, steps S1 to S4 are executed sequentially to complete incremental data fusion.
[0019] refer to Figure 2 As shown, this invention provides an incremental data fusion method for the industrial internet based on granular attribute reduction and sparse autoencoders. The specific implementation steps are as follows: S1: Based on granularity and extension decision-making, dynamic attribute reduction is performed on incremental industrial internet data within the current time window, specifically including: In the dynamic attribute reduction phase, let the original sample set within the current time window be... The attribute set is Each sample simultaneously contains equipment runtime sequence attributes, operating condition attributes, and textual description attributes. The system first divides the attribute space into multiple information granules according to a preset granulation criterion. Furthermore, by combining the extension decision-making concept, the contribution of different attributes to the decision results can be calculated to avoid the problem of insufficient identification of key attributes in complex working conditions by traditional static screening methods.
[0020] For any attribute and Its absolute correlation can be expressed as ,in The association function is constructed based on differences in sample distribution and consistency of decision-making; when This indicates that the two attributes have strong redundancy or insufficient correlation within the current time window. Further analysis using attribute weights can help address this. and data quality coefficient Calculate the overall retained score The attributes with higher scores are retained in the reduced result set. .
[0021] In practical implementation, threshold The system adaptively adjusts based on the volatility, attribute density, and anomaly ratio of the current batch of data, ensuring stable preservation of key attributes even when equipment operating conditions change, noise increases, or new data types are added. After this step, the reduced feature matrix is output. This provides low-redundancy input for subsequent feature extraction and incremental updates.
[0022] The attribute reduction process also records the mapping relationship between deleted attributes and retained attributes, which is used to quickly locate the affected information particles when new data arrives, thereby reducing the search range during subsequent local updates.
[0023] S2: The numerical sequence after dynamic attribute reduction is normalized and detrended before being input into a multi-layer fully connected network to extract textual and temporal features, specifically including: In the time series modeling and multidimensional feature extraction stages, the reduced numerical sequences are first normalized and detrended. Let the original time series of a certain attribute be... Estimating the trend term using a sliding window or local regression Then the detrended sequence can be written as This process can reduce the interference of long-term trends on model training, allowing the network to pay more attention to changes in device state, abnormal fluctuations, and short-term dynamic characteristics.
[0024] The detrended sequence is input into a multilayer fully connected network to extract temporal adjoint features. Its forward computation can be represented as follows: Where σ(·) is a nonlinear activation function, These are the network weight matrices, This is the bias term. Compared to directly using LSTM for long sequence modeling, the above structure can complete feature extraction with lower computational complexity in industrial incremental scenarios.
[0025] For text description data, first perform word segmentation, stop word filtering, and terminology normalization, then process the text sequence. Mapped to word vector matrix Then, semantic features are extracted through convolution and pooling operations, denoted as... The textual features and temporal features complement each other in semantic space, providing a more complete representation of industrial scenarios for subsequent deep fusion.
[0026] S3: Concatenate text features and temporal features to obtain a joint input, which is then fed into a sparse autoencoder for unsupervised representation learning. The final output fused feature matrix is directly supplied to the upper-layer industrial internet intelligent analysis platform, specifically including: Please see Figure 3 In the deep feature fusion stage, time features are... Text features By concatenating the data, we obtain the combined input. The SAE encoding process can be represented as follows: [The SAE encoding process is then fed into a sparse autoencoder for unsupervised representation learning.] The decoding process can be represented as ,in These are the encoder and decoder parameters, respectively.
[0027] During training, a sparsity constraint penalty term (such as KL divergence) is introduced to force hidden layer neurons to maintain a sparse activation state. This mechanism prompts SAE to automatically learn the underlying patterns within the data, transforming the original feature data into a more abstract representation. To ensure that the fusion result retains the main information while possessing strong sparse expressive power, the following objective function is used for training: ; The first term represents the reconstruction error, the second is the sparsity penalty term, and the third is the weight regularization term. Specifically, N represents the total number of samples within the time window, and z... i z represents the joint feature of the i-th sample input to the autoencoder. i r Let represent the i-th reconstructed feature output after network decoding; β is a hyperparameter controlling the weight of the sparse penalty term; KL represents the Kullback-Leibler divergence; ρ represents the manually set target activation of the hidden layer neurons (usually a small constant close to 0), ρ j λ represents the actual average activation response of the j-th neuron in the hidden layer on the training set; λ is the regularization coefficient that controls weight decay; and W represents the weight matrix parameters in the encoder and decoder.
[0028] By minimizing this objective function, the model can actively suppress ineffective feature responses and highlight discriminative shared representations among different modalities. After training, the hidden layer outputs... That is, as the final fusion feature SAE uses an optimization algorithm to fuse multiple features, extracting complementary information between features while significantly reducing feature redundancy. The final output of the network hidden layer features is a highly available and accurate fusion feature matrix that can be directly supplied to the upper-layer industrial internet intelligent analysis platform. This feature not only retains the temporal evolution information of the equipment's operating status but also absorbs contextual information from alarm texts, work order semantics, and process descriptions. Therefore, it is more suitable as input for status recognition, anomaly detection, quality prediction, and decision support in the industrial internet intelligent analysis platform.
[0029] S4: When a new batch of incremental data is detected, the knowledge rule incremental update mechanism is triggered to update the affected information particles and local model parameters in the new data. The specific operation is as follows: Please see Figure 4 During the dynamic knowledge update phase, this invention does not repeatedly train on all historical data, but instead focuses on newly added data. Update the affected information particles and local model parameters. Let historical knowledge be represented as... The overlap constraint between the new batch and the historical batch is: Then the knowledge granule update result at the current moment can be represented as ,in The coefficients are preserved for historical knowledge, and Ψ(·) is the local solution function performed for newly added samples.
[0030] By prioritizing the identification of granular units that overlap with new data in terms of attributes, time, or decision-making, and recalculating attribute reduction results, sample weights, and local statistics only within these units, the update scope is ensured to match the range of data changes. This approach significantly reduces response latency caused by full recalculation and lowers resource consumption during continuous online operation in industrial settings.
[0031] Furthermore, to reduce data loss during incremental fusion, based on data quality indicators... and data volume metrics Constructing the fusion adjustment coefficient ,in For adjustable weights, This represents the total number of samples in the current time window. The adjustment coefficient is used to dynamically correct the local update intensity, so that new batches can influence the fusion model more quickly when the quality is high, and maintain smooth updates when the quality fluctuates greatly.
[0032] In a preferred embodiment, steps S1 to S4 are repeated at preset time intervals, and the fused features output in each round are written into a feature cache or knowledge base to support subsequent model iteration, visualization analysis, and cross-system calls. This implementation achieves efficient processing of multi-source heterogeneous incremental data from the Industrial Internet through a sequential process of "dynamic attribute reduction—de-trending temporal modeling—knowledge granularity incremental update—SAE deep fusion."
[0033] In summary, this invention provides the DIDF-GRSAE method, which efficiently removes irrelevant attributes in the early stages of data inflow through granular extension decisions, avoids the high computational load of LSTM through power fading processing of FC networks during feature extraction, and achieves deep complementarity by leveraging the powerful unsupervised abstraction capabilities of SAE during the fusion phase. Compared with existing methods, this invention reduces the computational burden caused by redundant attributes while ensuring fusion accuracy, and maintains good real-time performance and stability even with continuous inflow of new data. This method requires only a small number of labeled training samples to efficiently adapt to the rapid identification and strong versatility requirements of industrial internet big data.
[0034] The above embodiments are merely descriptions of preferred embodiments of the present invention and are not intended to limit the scope of the present invention. Various modifications and improvements made by those skilled in the art to the technical solutions of the present invention without departing from the spirit of the present invention should fall within the protection scope defined by the claims of the present invention.
Claims
1. An incremental data fusion method for the Industrial Internet based on granular attribute reduction and sparse autoencoders, characterized in that, Includes the following steps: S1. Dynamically reduce the attributes of incremental industrial internet data within the current time window based on granular and extensional decision-making; S2. The numerical sequence after dynamic attribute reduction is normalized and detrended and then input into a multi-layer fully connected network to extract text features and temporal features; S3. The text features and temporal features are concatenated to obtain the joint input, which is then input into the sparse autoencoder for unsupervised representation learning. Finally, the fused feature matrix is output and directly supplied to the upper-level industrial internet intelligent analysis platform. S4. When a new batch of incremental data is detected, the knowledge rule incremental update mechanism is triggered to update the affected information particles and local model parameters in the new data.
2. The industrial internet incremental data fusion method based on granular attribute reduction and sparse autoencoder as described in claim 1, characterized in that, The dynamic attribute reduction process includes, for any two industrial internet incremental data, and Its absolute correlation can be expressed as ,in The association function is constructed based on the differences in sample distribution and decision consistency. Greater than the preset threshold If the two attributes are highly redundant or not correlated enough in the current time window, the comprehensive retention score is calculated by combining the attribute weight and data quality coefficient, and the attribute with the higher score is retained in the reduction result set.
3. The industrial internet incremental data fusion method based on granular attribute reduction and sparse autoencoder as described in claim 2, characterized in that, The overall retained score is calculated as follows: ； in, This indicates the overall retained score. Indicates attribute weight. This represents the data quality coefficient.
4. The industrial internet incremental data fusion method based on granular attribute reduction and sparse autoencoder as described in claim 1, characterized in that, The attribute reduction process also records the mapping relationship between deleted attributes and retained attributes, which is used to quickly locate the affected information particles when new data arrives, thereby reducing the search range during subsequent local updates.
5. The industrial internet incremental data fusion method based on granular attribute reduction and sparse autoencoder as described in claim 1, characterized in that, The specific steps of the detrending process are as follows: Suppose the original time series of a certain attribute is The trend term is estimated using a sliding window or local regression. The detrended sequence is then... .
6. The industrial internet incremental data fusion method based on granular attribute reduction and sparse autoencoder as described in claim 1, characterized in that, In a multi-layer fully connected network, for text description data, word segmentation, stop word filtering and term normalization are performed first, then the processed text sequence is mapped into a word vector matrix, and then semantic features are extracted through convolution and pooling operations.
7. The industrial internet incremental data fusion method based on granular attribute reduction and sparse autoencoder as described in claim 1, characterized in that, In the training process of unsupervised representation learning, the following objective function is used for training: ； The first term represents the reconstruction error, the second term represents the sparsity penalty term, the third term represents the weight regularization term, and N represents the sample weight. This represents the joint feature of the i-th sample input to the autoencoder. This represents the i-th reconstructed feature output after network decoding. Let KL represent the hyperparameter controlling the weights of the sparse penalty term, and let KL represent the divergence, the Kullback-Leibler divergence. This represents the target activation level of neurons in the hidden layer, as set manually. Let represent the actual average activation response of the j-th neuron in the hidden layer on the training set. This is represented by the regularization coefficient that controls weight decay. This represents the weight matrix parameters in the encoder and decoder.
8. The industrial internet incremental data fusion method based on granular attribute reduction and sparse autoencoder as described in claim 7, characterized in that, After training, the hidden layer outputs That is, as the final fusion feature SAE uses an optimization algorithm to fuse multiple features, extracting complementary information between features while significantly reducing feature redundancy.
9. The industrial internet incremental data fusion method based on granular attribute reduction and sparse autoencoder as described in claim 1, characterized in that, A fusion adjustment coefficient is constructed based on data quality and data volume indicators to reduce data loss during incremental fusion, as follows: ； in, Indicates the fusion adjustment coefficient. Indicates adjustable weights. Indicates data quality indicators, Indicates the amount of data. This represents the total number of samples in the current time window.
10. The industrial internet incremental data fusion method based on granular attribute reduction and sparse autoencoder as described in claim 1, characterized in that, Steps S1 to S4 are repeated at preset time intervals, and the fused features output in each round are written into the feature cache or knowledge base to support subsequent model iteration, visualization analysis and cross-system calls.