Big data analysis method based on deep learning
By integrating multimodal data through deep learning methods and utilizing graph neural networks and dynamic data graphs, the problem of the inability of existing technologies to fully judge the production line operation status and predict production data has been solved, and accurate prediction and adaptive optimization of the production line operation status have been achieved.
Patent Information
- Application Number
- CN202510711132.3
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-05-29
- Publication Date
- 2025-09-12
AI Technical Summary
Existing technologies have difficulty integrating multimodal data, resulting in the inability to fully judge the operating status of production lines and predict production data in industrial production, and the lack of real-time graphical correlation analysis capabilities.
A deep learning-based approach is used to perform real-time processing and analysis of multimodal data through real-time data collection, distributed computing framework, graph neural network and dynamic data graph, combined with convolutional neural network, recurrent neural network and graph attention network, to generate decision support information and form a closed-loop control mechanism.
It achieves a comprehensive judgment of the production line's operating status and accurate prediction of production data, improves the timeliness of decision-making and the system's adaptability, and ensures the normal operation and optimization of the production line.
Smart Images
Figure CN120632414A_ABST
Abstract
Description
Technical Field
[0001] The present application relates to the field of data analysis technology, and specifically to a big data analysis method based on deep learning. Background Art
[0002] With the popularization of industrial Internet of Things and intelligent manufacturing technologies, sensors, equipment and control systems in modern factories continue to generate massive amounts of heterogeneous data, including equipment vibration timing signals, quality inspection images, process parameter logs, energy consumption records, etc.
[0003] Traditional industrial data analysis methods rely on threshold alarms and statistical process control, such as monitoring equipment temperature exceeding the limit through fixed thresholds, or determining product quality abnormalities based on historical averages. Such methods are difficult to deal with the problem of multi-factor coupling under complex working conditions: a single sensor data cannot reflect the overall health status of the equipment (for example, it is difficult to predict mechanical wear based on current data alone), and the rule engine cannot dynamically adapt to fluctuations in production parameters (such as process deviations caused by differences in raw material batches). In the existing technology, there are single equipment fault prediction models, but most of them do not integrate multimodal data such as images and text, resulting in missed detection of complex faults; there are also some solutions that can centrally collect data, but lack real-time graphical correlation analysis capabilities, making it difficult to predict production data on the production line. Summary of the Invention
[0004] To this end, this application provides a big data analysis method based on deep learning to solve the problem of the existing technology that multimodal data such as images and texts are not integrated.
[0005] In order to achieve the above objectives, this application provides the following technical solutions:
[0006] In the first aspect, a big data analysis method based on deep learning comprises the following steps:
[0007] Step 1: Data collection: Using a real-time data collection system, a real-time data stream containing structured and unstructured data is acquired from the production line or user terminal. The data stream is then cleaned and integrated to generate a unified data stream set.
[0008] Step 2: Obtaining behavioral pattern information, using a distributed computing framework to perform real-time pattern recognition processing on the data stream set to extract behavioral pattern information from the data. The pattern recognition processing includes feature extraction based on convolutional neural networks and time series analysis based on recurrent neural networks.
[0009] Step 3: Construct a data graph. Based on the behavior pattern information, a dynamic data graph is constructed. The graph represents data entities through nodes and relationships between entities through edges. The weights of nodes and edges in the graph are dynamically updated to reflect changes in data flow.
[0010] Step 4: Analyze the data graph using a graph neural network model to analyze the dynamic data graph, calculate the influencing factors between nodes, and predict future trends based on the historical production data of the production line to obtain the predicted results of the production line operation. The analysis includes node embedding representation learning, edge weight optimization, and global feature fusion.
[0011] Step 5: Generate decision support information based on the influencing factors and prediction results. The decision support information includes risk warnings for the production line, trend predictions for the operating status of each component of the production line, and optimization suggestions for the production line.
[0012] Step 6: Form a closed-loop control mechanism to feed back the decision support information to the data source or user terminal, and adjust the production line equipment operating parameters or production line operation strategy in real time;
[0013] Step 7: Visualization display: The visualization module displays the forecast results of production line operation in the form of multi-dimensional charts, including dynamic heat maps, relationship network diagrams, and trend curve charts;
[0014] Step 8: Data compression and storage: Use data compression algorithm to encode and compress the processed data, and store the compressed data in the cloud or local database through encrypted transmission protocol.
[0015] Preferably, the structured data and unstructured data in step 1 include sensor data, log files, images, videos and text data, and the data cleaning process includes missing value filling, outlier removal and noise filtering.
[0016] Preferably, the distributed computing framework in step 2 adopts Apache Spark or Flink, and the pattern recognition process includes the following sub-steps:
[0017] Use CNN to extract spatial features of unstructured data;
[0018] Use RNN or long short-term memory network to analyze the periodicity of time series data;
[0019] Classify and label structured data through clustering algorithms.
[0020] Preferably, the construction of the dynamic data graph in step 3 is based on an association rule mining algorithm, specifically including:
[0021] Calculate the initial edge weights based on the co-occurrence frequency and correlation between data entities;
[0022] The graph is updated in real time through a sliding window mechanism, and the window size is dynamically adjusted according to the data flow rate.
[0023] Preferably, the graph neural network model in step 4 adopts a graph attention network or a graph convolutional network, and the impact factor calculation is implemented through the following steps:
[0024] Aggregate features of neighboring nodes of each node;
[0025] Use the attention mechanism to assign weights to different neighbor nodes;
[0026] Combined with the global pooling layer to generate comprehensive influence factors between nodes.
[0027] Preferably, the closed-loop control mechanism in step six includes:
[0028] Automatically trigger adjustments to production line equipment operation strategies based on decision support information;
[0029] The prediction results are pushed to the visualization module through the API interface to achieve real-time response.
[0030] Preferably, the visualization module in step seven supports interactive operations, and the user can dynamically filter the prediction results of the production line operation by adjusting parameters and export them as editable reports.
[0031] Preferably, the data compression algorithm in step eight adopts Huffman coding or LZ77 algorithm, and the encryption transmission protocol is TLS1.3 or AES-256.
[0032] Compared with the prior art, this application has at least the following beneficial effects:
[0033] 1. When implemented, the present invention supports parallel processing of structured and unstructured data through real-time data acquisition and a distributed computing framework, solving the problem of being unable to fully judge the operating status of production lines and unable to predict production data in industrial production;
[0034] 2. By combining graph neural networks and dynamic data graphs, it can capture data correlation changes and predict trends in real time, improving the timeliness and accuracy of decision-making and providing a strong guarantee for the normal operation of the production line;
[0035] 3. Through real-time feedback of decision support information, a dynamic adjustment mechanism for production strategies is formed, enhancing the adaptability of the system. BRIEF DESCRIPTION OF THE DRAWINGS
[0036] To more intuitively illustrate the prior art and the present application, exemplary drawings are provided below. It should be understood that the specific shapes and structures shown in the drawings should not generally be considered as limiting conditions for implementing the present application; for example, based on the technical concepts disclosed in this application and the exemplary drawings, those skilled in the art are capable of easily making routine adjustments or further optimizations to the addition / reduction / attribution division, specific shapes, positional relationships, connection methods, dimensional ratios, etc. of certain units (components).
[0037] Figure 1 This is a flowchart of a big data analysis method based on deep learning in this application. DETAILED DESCRIPTION
[0038] The present application will be further described below in detail through specific embodiments in conjunction with the accompanying drawings.
[0039] like Figure 1 As shown, a big data analysis method based on deep learning includes the following steps:
[0040] Step 1: Data collection: Using a real-time data collection system, real-time data streams containing structured and unstructured data are acquired from production lines or user terminals. These data streams are then cleaned and integrated to generate a unified data stream set. This unified cleaning and distributed processing overcomes the bottleneck of traditional methods that make parallel analysis of multi-format data difficult, thereby improving data processing efficiency.
[0041] Step 2: Obtaining behavioral pattern information, performing real-time pattern recognition processing on the data stream set using a distributed computing framework to extract behavioral pattern information from the data. The pattern recognition processing includes feature extraction based on convolutional neural networks (CNNs) and time series analysis based on recurrent neural networks (RNNs);
[0042] Step 3: Construct a data graph. Based on the behavior pattern information, a dynamic data graph is constructed. The graph represents data entities through nodes and relationships between entities through edges. The weights of nodes and edges in the graph are dynamically updated to reflect changes in data flow.
[0043] Step 4: Analyze the data graph using a graph neural network (GNN) model to analyze the dynamic data graph, calculate the influencing factors between nodes, and predict future trends based on the historical production data of the production line and obtain the predicted results of the production line operation. The analysis includes node embedding representation learning, edge weight optimization, and global feature fusion. Based on the coupling of dynamic graphs and GNNs, it can capture the implicit causal relationships in the data stream. For example, through the energy consumption data of the equipment, the wear of the equipment can be judged, and the operating status of each device on the production line can be judged and predicted. The prediction accuracy is significantly improved compared with traditional statistical models.
[0044] Step 5: Generate decision support information based on the influencing factors and prediction results. The decision support information includes risk warnings for the production line, trend predictions for the operating status of each component of the production line, and optimization suggestions for the production line.
[0045] Step 6: Form a closed-loop control mechanism to feed back the decision support information to the data source or user terminal, and adjust the production line equipment operating parameters or production line operation strategy in real time. The feedback mechanism enables the system to adjust the strategy in real time based on the prediction results;
[0046] Step 7: Visualization. The visualization module displays the predicted results of the production line operation in the form of multi-dimensional charts, including dynamic heat maps, relationship network diagrams, and trend curve charts. The visualization and encryption design takes into account both the interpretability of the results and the protection of data privacy, providing a safe and easy-to-understand display for the production line operation.
[0047] Step 8: Data compression and storage: Use data compression algorithm to encode and compress the processed data, and store the compressed data in the cloud or local database through encrypted transmission protocol.
[0048] First, a real-time data acquisition system (individual sensors collect data) integrates data collected on the production line (such as sensor data, text, images, etc.). Data cleaning techniques are used to eliminate noise and redundancy, forming a standardized data stream. Subsequently, a hybrid model consisting of CNN and RNN is used for feature extraction: CNN is responsible for capturing spatial features from unstructured data (such as images and videos), while RNN analyzes periodicity and trend changes in time series data (such as logs and sensor streams). Based on feature extraction, the dynamic data graph dynamically updates the weights of nodes and edges (e.g., based on a sliding window mechanism) to reflect the changes in the associations between various devices on the production line in real time.
[0049] After the graph is constructed, the graph neural network (GNN) aggregates node neighbor information and global features to quantify the influencing factors between nodes (such as the correlation strength of user behavior on system load). It then combines historical data to train a predictive model, outputting future trends and providing decision makers with operational recommendations for production lines. Once decision-support information is generated, the system uses a closed-loop feedback mechanism to push optimization suggestions to the data source or visualization module, such as adjusting the operating parameters of production line equipment. The visualization module also converts the predicted results of complex production line operations into multidimensional charts for intuitive user understanding. The data compression and encryption module ensures the efficient storage and secure transmission of processed results.
[0050] The structured data and unstructured data described in step one include sensor data, log files, images, videos, and text data. The data cleaning process includes missing value filling, outlier removal, and noise filtering, which ensures the accuracy of production line data collection and improves the accuracy of prediction results.
[0051] By specifying that production line data includes sensor data, log files, and other data, this method ensures its applicability to typical industrial production scenarios. During the cleaning process, missing values are filled using averaging or interpolation, outliers are removed using box plots or the 3σ principle, and noise is filtered using wavelet transforms or sliding averages. This design eliminates interference during data collection and improves the reliability of subsequent analysis. For example, in smart factories, sensor data often generates noise due to equipment jitter. This method effectively filters out abnormal signals, ensuring accurate equipment status assessment.
[0052] The distributed computing framework described in step 2 uses Apache Spark or Flink, and the pattern recognition process includes the following sub-steps:
[0053] Use CNN to extract spatial features of unstructured data;
[0054] Use RNN or long short-term memory (LSTM) networks to analyze the periodicity of time series data;
[0055] Classify and label structured data through clustering algorithms.
[0056] Using the Spark or Flink framework enables low-latency batch and stream processing of data streams. CNNs are used to extract features from images (such as the location of equipment on a production line), RNNs / LSTMs analyze time series data, and clustering algorithms (such as K-means) classify structured data.
[0057] Structured data includes equipment operating parameters (such as temperature, pressure, vibration amplitude, current / voltage values and equipment status), production process data (such as product batch number, process parameters, raw material input, energy consumption records, hourly output quantity, qualified / defective product counts, etc.); unstructured data includes visual inspection data (such as infrared thermal imaging images of the equipment), text and log record data (such as equipment maintenance reports, operating manual updates, emails or instant messaging records of production schedule changes), audio and video data (such as equipment operation noise recordings, production line monitoring videos, and remote maintenance guidance images recorded by AR / VR devices, etc.), equipment configuration file data, etc.
[0058] The construction of the dynamic data graph described in step 3 is based on the association rule mining algorithm, which specifically includes:
[0059] Calculate the initial edge weights based on the co-occurrence frequency and correlation between data entities;
[0060] The graph is updated in real time through a sliding window mechanism, and the window size is dynamically adjusted according to the data flow rate.
[0061] Association rule mining uses the Apriori or FP-Growth algorithm, initializing edge weights based on data co-occurrence frequency (e.g., the probability that a failure in one device on a production line will cause failures in other devices). A sliding window mechanism dynamically adjusts the window length (e.g., 5-10 seconds) based on the data flow rate (e.g., thousands of records per second) to ensure real-time graph updates.
[0062] The graph neural network model described in step 4 uses a graph attention network (GAT) or a graph convolutional network (GCN), and calculates the impact factor through the following steps:
[0063] Aggregate features of neighboring nodes of each node;
[0064] Use the attention mechanism to assign weights to different neighbor nodes;
[0065] Combined with the global pooling layer to generate comprehensive influence factors between nodes.
[0066] GAT uses an attention mechanism to assign weights to different neighboring nodes (e.g., key equipment in a production line has a higher weight), while GCN aggregates local features through spectral convolution. Global pooling layers (such as SortPooling) convert node features into graph-level representations for calculating comprehensive impact factors.
[0067] The closed-loop control mechanism described in step six includes:
[0068] Automatically trigger adjustments to production line equipment operation strategies based on decision support information;
[0069] The prediction results are pushed to the visualization module through the API interface to achieve real-time response.
[0070] Automatic policy adjustments include: when it is predicted that the equipment on the production line exceeds the limit, triggering the equipment's operating power reduction instruction (or other instructions to control the production line equipment based on the prediction results); the API interface uses RESTful or gRPC protocol to push the prediction results to third-party systems (such as ERP, CRM).
[0071] The visualization module described in step 7 supports interactive operations. Users can dynamically filter the prediction results of production line operation by adjusting parameters and export them as editable reports.
[0072] Users can adjust parameters (such as time range and data dimensions) by dragging and dropping sliders, dynamically filtering results and generating heat maps. Export reports in PDF and Excel formats include links to the original data for further analysis.
[0073] The data compression algorithm described in step 8 adopts Huffman coding or LZ77 algorithm, and the encryption transmission protocol is TLS1.3 or AES-256.
[0074] Huffman coding can achieve compression rates of 50%-70% for text data, while LZ77 is suitable for log data with many repetitive patterns. Regarding encryption protocols, TLS 1.3 ensures transmission channel security, and AES-256 is used for static data encryption.
[0075] The technical features of the above embodiments can be combined arbitrarily (as long as there is no contradiction in the combination of these technical features). In order to make the description concise, not all possible combinations of the technical features in the above embodiments are described; these embodiments that are not explicitly written should also be considered to be within the scope of this specification.
Claims
1. A big data analysis method based on deep learning, characterized in that: The following steps are included: Step 1: Data collection: Using a real-time data collection system, a real-time data stream containing structured and unstructured data is acquired from the production line or user terminal. The data stream is then cleaned and integrated to generate a unified data stream set. Step 2: Obtaining behavioral pattern information, using a distributed computing framework to perform real-time pattern recognition processing on the data stream set to extract behavioral pattern information from the data. The pattern recognition processing includes feature extraction based on convolutional neural networks and time series analysis based on recurrent neural networks. Step 3: Construct a data graph. Based on the behavior pattern information, a dynamic data graph is constructed. The graph represents data entities through nodes and relationships between entities through edges. The weights of nodes and edges in the graph are dynamically updated to reflect changes in data flow. Step 4: Analyze the data graph using a graph neural network model to analyze the dynamic data graph, calculate the influencing factors between nodes, and predict future trends based on the historical production data of the production line to obtain the predicted results of the production line operation. The analysis includes node embedding representation learning, edge weight optimization, and global feature fusion. Step 5: Generate decision support information based on the influencing factors and prediction results. The decision support information includes risk warnings for the production line, trend predictions for the operating status of each component of the production line, and optimization suggestions for the production line. Step 6: Form a closed-loop control mechanism to feed back the decision support information to the data source or user terminal, and adjust the production line equipment operating parameters or production line operation strategy in real time; Step 7: Visualization display: The visualization module displays the forecast results of production line operation in the form of multi-dimensional charts, including dynamic heat maps, relationship network diagrams, and trend curve charts; Step 8: Data compression and storage: use a data compression algorithm to encode and compress the processed data, and store the compressed data in the cloud or local database through an encrypted transmission protocol; The construction of the dynamic data graph described in step 3 is based on the association rule mining algorithm, which specifically includes: Calculate the initial edge weights based on the co-occurrence frequency and correlation between data entities; The map is updated in real time through a sliding window mechanism, and the window size is dynamically adjusted according to the data flow rate; The graph neural network model described in step 4 uses a graph attention network or a graph convolutional network, and calculates the impact factor through the following steps: Aggregate features of neighboring nodes of each node; Use the attention mechanism to assign weights to different neighbor nodes; Combined with the global pooling layer to generate comprehensive influence factors between nodes.
2. The big data analysis method based on deep learning according to claim 1, characterized in that: The structured data and unstructured data mentioned in step 1 include sensor data, log files, images, videos and text data. The data cleaning process includes missing value filling, outlier removal and noise filtering.
3. The big data analysis method based on deep learning according to claim 1, characterized in that: The distributed computing framework described in step 2 uses Apache Spark or Flink, and the pattern recognition process includes the following sub-steps: Use CNN to extract spatial features of unstructured data; Use RNN or long short-term memory network to analyze the periodicity of time series data; Classify and label structured data through clustering algorithms.
4. The big data analysis method based on deep learning according to claim 1, characterized in that: The closed-loop control mechanism described in step six includes: Automatically trigger adjustments to production line equipment operation strategies based on decision support information; The prediction results are pushed to the visualization module through the API interface to achieve real-time response.
5. The big data analysis method based on deep learning according to claim 1, characterized in that: The visualization module described in step 7 supports interactive operations. Users can dynamically filter the prediction results of production line operation by adjusting parameters and export them as editable reports.
6. The big data analysis method based on deep learning according to claim 1, characterized in that: The data compression algorithm described in step 8 adopts Huffman coding or LZ77 algorithm, and the encryption transmission protocol is TLS1.3 or AES-256.