An edge-computing-based production data real-time analysis method and system

Abstract

The application belongs to the field of production control, and specifically discloses a production data real-time analysis method and system based on edge computing. The method first collects operation data under different processes of production equipment, trains sub-models by process after pre-processing, and constructs a data analysis model adapted to multiple processes. Process context labels are generated by real-time acquisition of process switching information at edge nodes and are bound with real-time operation data. When responding to a data analysis request, the corresponding sub-models are automatically switched according to the labels to analyze data. When the process switching frequency exceeds a preset threshold, a meta-learning mechanism is started, model parameters are optimized based on historical process switching sample data, and the optimized model is used to analyze real-time data. The application can solve the model adaptation problem caused by process switching and improve the real-time analysis accuracy and efficiency of production data.

Description

Technical Field

[0001] This invention belongs to the field of production control, specifically relating to a method and system for real-time analysis of production data based on edge computing. Background Technology

[0002] In the development of Industry 4.0 and intelligent manufacturing, the demand for flexible scheduling of equipment on production lines is increasing. Some equipment needs to temporarily undertake multiple processes, causing the distribution of production data collected by the same edge node to change dynamically with process switching. Currently, although edge computing has been widely used in real-time analysis of production data and can reduce data transmission latency, traditional data analysis models have significant limitations: they are trained based on data distribution of fixed processes and can only adapt to single or a small number of fixed working conditions. When equipment processes switch, the model cannot quickly adapt to the new data distribution, easily confuses the characteristic differences of different processes, and leads to a significant decrease in analysis accuracy (such as mistakenly judging the vibration data of the milling process as the abnormality of the drilling process). Moreover, manual intervention is required to supplement a large amount of new data to retrain the model, which usually takes several hours or even days, violating the core requirements of industrial production for real-time performance and high accuracy.

[0003] Meanwhile, the lack of a dynamic adaptation mechanism for frequent process switching in existing technologies makes it difficult to meet the actual application needs of flexible production lines such as automotive parts processing and electronic component assembly, thus restricting the further promotion of edge computing in the field of real-time production data analysis. Summary of the Invention

[0004] To address this, the present invention provides a method and system for real-time analysis of production data based on edge computing, thereby solving the aforementioned technical problems.

[0005] This invention provides a real-time analysis method for production data based on edge computing, comprising the following steps: The operation data of production equipment under different processes are collected to form a sample dataset. After preprocessing, the sample dataset is used to train sub-models for each process to build a multi-process adapted data analysis model. The operation data includes equipment sensor data, operation parameters and process identification information. At the edge node, process switching information is acquired in real time and process context tags are generated. The process context tags are then bound to the real-time running data collected by the edge node. In response to data analysis requests from edge nodes, the system reads the process context tags bound to real-time running data, automatically switches to the corresponding process sub-model based on the tag information, and analyzes the real-time running data. When the process switching frequency exceeds the preset threshold, the meta-learning mechanism is activated to optimize the data analysis model parameters based on historical process switching sample data, and to analyze the real-time operation data based on the optimized data analysis model.

[0006] Furthermore, the equipment sensor data includes at least temperature, pressure, rotational speed, and vibration frequency; the operating parameters include equipment start / stop status, load rate, and processing accuracy setpoint; the process identification information includes process name, process number, process execution duration, and corresponding quality standard parameters.

[0007] Furthermore, the step of automatically switching to the corresponding sub-model of the process based on the label information and analyzing the real-time running data includes: edge nodes reading the process context labels of the real-time data, extracting the process ID and matching it with the pre-stored process-model mapping table; if the match is successful, the corresponding sub-model is called; if the match fails, the default model is triggered and a label exception prompt is issued. The default model uses transfer learning to initialize parameters.

[0008] Furthermore, the meta-learning mechanism includes: constructing a meta-learning dataset and collecting sample data from historical process switching; training the meta-learning dataset using a model-independent meta-learning algorithm to learn a fast-adaptive parameter update strategy; and when a new process switching occurs, iteratively optimizing the model parameters based on a small amount of initial data from the new process using the parameter update strategy to reduce the model prediction error to below a preset threshold, wherein the sample data includes running data before and after the switching, model parameter adjustment records, and analysis error values.

[0009] Furthermore, the multi-process adapted data analysis model adopts a modular architecture, with each process corresponding to an independent sub-model module. The sub-model adopts a hybrid structure of convolutional neural network and long short-term memory network. The number of neurons in the input layer matches the number of filtered feature parameters. There are 3 hidden layers with Leaky ReLU activation function. The output layer is the production quality prediction value or abnormal risk level.

[0010] Furthermore, the process context label is encoded in JSON format and includes the following fields: process ID, process name, switch timestamp, associated device ID, and standard parameter range.

[0011] Furthermore, the sample dataset preprocessing includes data cleaning, standardization, and feature selection. Data cleaning removes outliers caused by sensor malfunctions and fluctuations during equipment start-up and shutdown. Standardization uses the Z-score standardization method to convert operational data of different dimensions into values ​​within a uniform range. Feature selection uses the Pearson correlation coefficient method to retain feature parameters that are more correlated with production quality than a preset threshold.

[0012] In another aspect, this application also provides a real-time production data analysis system based on edge computing, comprising: The data acquisition and preprocessing module is used to collect the operating data of the production equipment under different processes to form a sample dataset. After preprocessing, the sample dataset is used to train sub-models for each process to build a multi-process adapted data analysis model. The operating data includes equipment sensor data, operating parameters and process identification information. The process context label binding module is used to acquire process switching information in real time at the edge node and generate process context labels, and bind the process context labels to the real-time running data collected by the edge node; The real-time running data analysis module is used to respond to data analysis requests from edge nodes, read the process context tags bound to the real-time running data, and automatically switch to the corresponding process sub-model based on the tag information to analyze the real-time running data; The optimization module is used to activate the meta-learning mechanism when the process switching frequency exceeds a preset threshold. It optimizes the data analysis model parameters based on historical process switching sample data and analyzes the real-time running data based on the optimized data analysis model.

[0013] Furthermore, it also includes a transfer learning module, which is used by edge nodes to read the process context labels of real-time data, extract the process ID and match it with the pre-stored process-model mapping table. If the match is successful, the corresponding sub-model is called; if the match fails, the default model is triggered and a label exception prompt is issued. The default model uses transfer learning to initialize parameters.

[0014] Furthermore, it also includes a meta-learning mechanism module, which is used to construct a meta-learning dataset and collect sample data from historical process switching. The meta-learning dataset is trained using a model-independent meta-learning algorithm to learn a fast-adaptive parameter update strategy. When a new process switching occurs, the model parameters are iteratively optimized based on a small amount of initial data from the new process through the parameter update strategy, so that the model prediction error is reduced to below a preset threshold. The sample data includes running data before and after the switching, model parameter adjustment records, and analysis error values.

[0015] This invention effectively solves the problems of data distribution changes and model confusion caused by equipment process switching on the production line through a collaborative design that includes training sub-models for different processes, binding process context labels, automatic sub-model switching, and meta-learning optimization. It utilizes dedicated sub-models to ensure the accuracy of data analysis for regular processes, constructs default models through transfer learning to handle cold start scenarios for new processes, and leverages a meta-learning mechanism to rapidly optimize model parameters under frequent switching, significantly improving model adaptation efficiency and analytical accuracy while reducing manual intervention costs. Simultaneously, it achieves data processing locally based on edge computing, reducing transmission latency and meeting the real-time and high-precision requirements of industrial production, adapting to the dynamic adjustment scenarios of multiple processes on flexible production lines. Attached Figure Description

[0016] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the embodiments will be briefly introduced below. It should be understood that the following drawings only show some embodiments of the present invention and should not be regarded as a limitation on the scope. For those skilled in the art, other related drawings can be obtained based on these drawings without creative effort.

[0017] Other features, objects, and advantages of this application will become more apparent from the following detailed description of non-limiting embodiments with reference to the accompanying drawings: Figure 1 The flowchart illustrates a real-time production data analysis method based on edge computing, as provided in this embodiment of the invention.

[0018] Figure 2 This is a schematic diagram of the sub-model architecture provided in an embodiment of the present invention.

[0019] Figure 3 This is a schematic diagram illustrating the specific process of the meta-learning mechanism provided in an embodiment of the present invention.

[0020] Figure 4 This is a schematic diagram of a real-time production data analysis system based on edge computing, provided as an embodiment of the present invention. Detailed Implementation

[0021] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0022] like Figure 1 As shown, this embodiment of the invention discloses a real-time analysis method 100 for production data based on edge computing, including the following method steps: S101, collect the operating data of the production equipment under different processes to form a sample dataset. After preprocessing, the sample dataset is used to train sub-models for each process to build a multi-process adapted data analysis model. The operating data includes equipment sensor data, operating parameters and process identification information. S102, at the edge node, process switching information is acquired in real time and process context labels are generated, and the process context labels are bound to the real-time running data collected by the edge node; S103, in response to the data analysis request from the edge node, reads the process context label bound to the real-time running data, and automatically switches to the sub-model of the corresponding process according to the label information to analyze the real-time running data; S104, when the process switching frequency exceeds the preset threshold, the meta-learning mechanism is activated to optimize the data analysis model parameters based on historical process switching sample data, and to analyze the real-time operation data based on the optimized data analysis model.

[0023] In one embodiment, the equipment sensor data includes at least temperature, pressure, rotational speed, and vibration frequency; the operating parameters include equipment start / stop status, load rate, and processing accuracy setting; the process identification information includes process name, process number, process execution time, and corresponding quality standard parameters.

[0024] Specifically, by deploying sensor arrays at key locations on production equipment, such as temperature sensors, pressure transmitters, speed encoders, and vibration sensors, real-time operating data of the equipment under different processes can be collected synchronously.

[0025] For example, the device sensor data covers physical quantities such as temperature (sampling accuracy ±0.5℃), pressure (range 0-10MPa), rotational speed (measurement range 0-3000r / min), and vibration frequency (detection frequency band 10-5000Hz); For example, operating parameters are obtained through the equipment control system interface, including equipment start / stop status (Boolean value), load rate (0-100%), and machining accuracy setting value (e.g., ±0.01mm). For example, the process identification information is synchronously pushed by the production execution system (MES), which includes the process name (such as "drilling" or "milling"), process number (such as "PROC-001"), process execution time and corresponding quality standard parameters (such as dimensional tolerance range).

[0026] In one embodiment, sample dataset preprocessing includes data cleaning, standardization, and feature selection. Data cleaning removes outliers caused by sensor malfunctions and fluctuations during equipment start-up and shutdown. Standardization uses the Z-score standardization method to convert operational data of different dimensions into values ​​within a uniform range. Feature selection uses the Pearson correlation coefficient method to retain feature parameters that are more correlated with production quality than a preset threshold.

[0027] Specifically, the collected raw data undergoes three levels of preprocessing, which, for example, includes: For data cleaning, for example, the 3σ criterion is used to remove outliers caused by sensor malfunctions (data that exceed the mean ± 3 times the standard deviation), while filtering out fluctuation data within 10 seconds of the device starting and stopping. For standardization, for example, the Z-score algorithm (x_norm=(x-μ) / σ, where μ is the mean and σ is the standard deviation) is used to transform parameters of different dimensions to the standard normal distribution interval; For feature selection, for example, the Pearson correlation coefficient is used to calculate (|r|>0.6 is the selection threshold), retaining feature parameters that are strongly correlated with production quality indicators (such as dimensional accuracy and surface roughness) and eliminating redundant information.

[0028] In one embodiment, based on the preprocessed sample dataset, a modular architecture is adopted to build a dedicated analysis model for each process. For example, a one-process-one-module design principle is used, with each process corresponding to an independent sub-model. The overall architecture is a hybrid structure of Convolutional Neural Network (CNN) and Long Short-Term Memory Network (LSTM), such as... Figure 2 As shown, the model structure specifically includes: A CNN layer, for example, consists of two convolutional blocks, each containing 32 3×3 convolutional kernels and one 2×2 max pooling layer, used to extract local features from the data; The LSTM layer, for example, has 64 hidden cells and uses a gating mechanism to capture timing dependencies, such as the impact of rotational speed changes on subsequent processing accuracy, and the correlation between temperature accumulation and equipment wear. Fully connected layers, for example, contain three hidden layers, exemplarily with 64, 32, and 16 neurons respectively, and the activation function is Leaky ReLU (slope 0.01) to solve the gradient vanishing problem; The output layer, for example, outputs production quality predictions (such as dimensional error values) or abnormal risk levels (levels 1-5, where level 1 is normal and level 5 is an emergency fault) based on process requirements.

[0029] In one embodiment, the dataset is divided into process categories, and a training set (80%), a validation set (10%), and a test set (10%) are allocated separately for each process. The Adam optimizer is used (initial learning rate 0.001, decaying by 10% every 100 rounds), and the mean squared error (MSE) loss function is selected. The parameters are updated through backpropagation. Implement an early stopping strategy, that is, stop training when the validation set loss does not decrease for 10 consecutive rounds to avoid overfitting; After training, the model performance is verified using a test set. The prediction accuracy of the single-process model is required to be ≥95% (e.g., the surface roughness prediction error of the milling process is ≤0.1μm). Otherwise, the model is returned to the data acquisition stage to supplement samples.

[0030] Optionally, a model library can be built to store sub-models and weight parameters for each process, using version control, such as "milling sub-model v2.3" to record the iteration process. The model library supports dynamic expansion; when a new process is added, a new sub-model can be quickly trained based on the same architecture and added to the library, achieving flexible adaptation to production line adjustments.

[0031] In one embodiment, the process context label is encoded in JSON format and includes the following fields: process ID, process name, switch timestamp, associated device ID, and standard parameter range.

[0032] Specifically, the edge node generates a process context label in a standardized format based on the received process switching information. For example, the specific specification is as follows: First, the tag structure design uses JSON format for structured encoding, and for example, includes the following core fields: Process ID (process_id), for example, a unique identifier consisting of 8 alphanumeric characters, such as "GRD-0012", is used to accurately distinguish different processes; The process name (process_name) is used, for example, in full Chinese, such as "high-precision bearing inner ring grinding process", which is easy for manual identification; The switch timestamp, for example, is the process switch time accurate to the millisecond level, generated based on the local high-precision clock of the edge node (synchronized to the NTP server, with a time error ≤5ms); Associate the device ID (device_id), for example, a unique code corresponding to the production equipment, such as "EQP-2023-045", to ensure the binding relationship between the tag and the device; The standard parameter range (standard_params) is, for example, the allowable fluctuation range of key process parameters in this process, such as spindle speed 1500-1800 r / min and feed rate 0.05-0.08 mm / r, which serves as a benchmark reference for data analysis.

[0033] In one embodiment, edge nodes bind process context labels to real-time acquired runtime data using a time alignment and logical association algorithm. For example, the specific process is as follows: Real-time data acquisition and tag generation at edge nodes share the same clock source, ensuring that the synchronization accuracy of the timestamps between operational data and process tags is controlled within ±10ms. For each set of acquired operational data (including sensor data and operational parameters), its acquisition timestamp is extracted and compared with the process switching timestamp in the tag buffer to determine the process cycle to which the data belongs. For example, if the acquisition time of a certain data is 15:30:46.123 and the most recent process switching time is 15:30:45.789, then the data is determined to belong to the new process after the switch.

[0034] A one-to-one binding model is adopted, where each set of running data is matched with a unique process context label. The binding process is implemented through memory mapping technology, which allocates a special field in the data storage structure to embed label information, forming a complete data unit containing the original data and the process label.

[0035] Through the above process, the edge nodes realize the dynamic association between production data and process information, so that each set of operation data has a clear process attribute identifier, providing a basis for judgment for subsequent model switching based on process tags.

[0036] In one embodiment, for S103, the analysis request includes metadata such as the storage address of the data to be analyzed, the data acquisition timestamp, and the associated device ID. Upon receiving the request, the process context tag is first extracted from the real-time running data unit to be analyzed. For example, the parsing process can be as follows: Tag extraction: The tag field is located using the structured index of the data unit, and the complete JSON-formatted tag content is read. For example, extracting tags from the operational data of a bearing grinding process: {"process_id":"GRD-0012","process_name":"High-precision bearing inner ring grinding process","switch_timestamp":"2024-09-02T15:30:45.789","device_id":"EQP-2023-045","standard_params":{"spindle_speed":[1500,1800],"feed_rate":[0.05,0.08]}}.

[0037] The extraction tags extract the `process_id` (unique identifier for the process) and `standard_params` (standard parameter range). `process_id` is used to determine the appropriate sub-model, and `standard_params` serves as the baseline threshold for subsequent analysis (e.g., determining if the current spindle speed is within a reasonable range of 1500-1800 r / min). If an error in the tag format (e.g., missing `process_id` field) or abnormal content (e.g., an empty standard parameter range) is found during parsing, the default processing mechanism is triggered, calling the general backup model for analysis and recording the error in the log.

[0038] In one embodiment, the edge node reads the process context label of the real-time data, extracts the process ID and matches it with the pre-stored process-model mapping table. If the match is successful, the corresponding sub-model is called; if the match fails, the default model is triggered and a label exception prompt is issued. The default model uses transfer learning to initialize parameters.

[0039] Specifically, based on the parsed process_id, the matching and switching operation of the sub-model is performed. For example, the specific process includes: First, query the model mapping table, specifically the pre-stored process-model mapping table for edge nodes. This table uses `process_id` as the key and the storage path, version number, and input / output parameter specifications of the corresponding sub-model as values. For example, GRD-0012 corresponds to "Bearing Inner Ring Grinding Sub-model v3.2", with the storage path ` / model / grinding / inner_ring_v3.2`. Locate the target sub-model through key-value matching.

[0040] Next, the sub-model is loaded and initialized. Based on the storage path found in the mapping table, the target sub-model is loaded from the model library of the edge nodes (which uses distributed storage to support fast reading of model files). The loading process includes model structure verification (checking whether the number of network layers and neuron configuration match the current process), parameter integrity verification (ensuring that the weight matrix is ​​not damaged), and initialization of the input layer (adjusting the number of neurons according to the number of feature parameters of the current process; for example, if the grinding process contains 5 feature parameters, then the input layer is set to 5 neurons).

[0041] Optionally, if no sub-model matching process_id is found, such as when a new, unpreset process is added, the default model switching mechanism is automatically triggered. The default model is built based on transfer learning and is initialized using the parameters of the historical process model most similar to the current process (determined by calculating the cosine similarity of the process feature vectors; a similarity ≥ 0.8 is considered similar). This ensures that basic analytical capabilities can still be provided even when there is no dedicated sub-model. At the same time, the edge node sends a "model missing alarm" to the production management system, prompting technicians to supplement and train the corresponding sub-model.

[0042] Specifically, the default model's basic architecture is inherited from a widely applicable sub-model of a production line process, prioritizing mature process models with ample historical data and similar feature dimensions as transfer sources. For example, in the machining scenario, a general metal cutting sub-model is selected as the foundation. This model has been trained on more than 10 similar processes such as turning and milling, and includes a CNN-LSTM hybrid network structure (the input layer supports 8-12 feature parameters, and the hidden layer configuration is consistent with the multi-process sub-model). It also possesses strong generalization ability in feature extraction layers (such as vibration frequency features and temperature change trend capture modules). The parameters of the basic model have been validated through long-term production, and its accuracy in the analysis of similar processes remains stable at over 90%, providing a reliable knowledge foundation for transfer learning.

[0043] In this embodiment, to ensure the effectiveness of knowledge transfer, the default model calculates the similarity between the new process and the historical process before initialization and selects the most suitable transfer source parameters. For example, this includes the following steps: S201, construct feature vectors by extracting key feature parameters (such as spindle speed range, material hardness, equipment load characteristics, etc.) from the initial operation data (first 50 samples) of the new process to form process feature vectors; at the same time, retrieve the standard feature vectors of each process from the historical process library; S202, the cosine similarity algorithm is used to quantify the degree of matching between the new process and the historical process. For example, the calculation formula is as follows: Similarity = cosθ = (A・B) / (|A|・|B|) Here, A represents the feature vector of the new process, and B represents the feature vector of the historical process. The closer the ratio of the vector dot product (A・B) to the modulus product (|A|・|B|) is to 1, the more similar the features are. For example, the feature vector similarity between the newly added "precision boring process" and the historical "reaming process" is 0.87, which is considered highly similar.

[0044] S203, Select the transfer source parameters, set the similarity threshold to 0.75, and select the sub-model parameters of the process with the highest similarity that exceeds the threshold from the historical processes as the transfer source parameters. If there are multiple transfer source parameters that meet the conditions (such as the top 3 similarity parameters all exceeding 0.75), then the knowledge of multiple models is integrated through a weighted fusion strategy (allocating the parameter proportion according to the similarity weight) to improve the transfer effect.

[0045] In this embodiment, the default model adapts to the data characteristics of the new process through parameter migration and local fine-tuning. For example, this specifically includes: Feature extraction layer parameter transfer, for example, completely transfers the parameters of the CNN layer (convolution kernel weights, pooling layer parameters) and LSTM layer (gating unit weights, memory cell initialization parameters) in the source model to the default model, retaining its ability to extract general industrial features (such as the periodicity of mechanical vibration and the distribution pattern of temperature field).

[0046] Output layer parameter reconstruction involves, for example, reconstructing the number of neurons and weight matrix of the output layer based on the analysis objectives of the new process (e.g., a quality prediction indicator of "aperture roundness" rather than "surface roughness"). For instance, for the quality standard parameters of the new process, the number of neurons in the output layer is adjusted to match the indicator dimension (e.g., one neuron for a single quality indicator, five neurons for risk level assessment), and the weights are reset using the Xavier initialization method to ensure the output conforms to the physical meaning of the new process.

[0047] Local fine-tuning optimization, for example, involves rapidly fine-tuning the model using a small amount of initial data (50-100 sets) of data from the new process. The parameters of the feature extraction layer are frozen (to avoid disrupting general knowledge transferred from the previous step), and only the parameters of the output layer and adjacent fully connected layers are updated. For instance, the fine-tuning process uses mini-batch gradient descent (batch size 16), with a learning rate set to 0.0005 (lower than the learning rate used in new training to prevent drastic parameter fluctuations), iterating for 10-20 rounds until the loss function (mean squared error) converges to a stable range. Through fine-tuning, the model can capture the unique features of the new process, reducing initial analysis errors.

[0048] In one embodiment, after the target sub-model is loaded, real-time running data is analyzed and calculated to output results that can be used for production decisions. For example, this includes the following process: First, the real-time running data is standardized according to the requirements of the sub-model (using the same Z-score parameters as during model training to ensure data distribution matching), and the key feature parameters of the process are selected (such as retaining 5 features for the grinding process: "grinding wheel speed, feed rate, grinding depth, workpiece temperature, and vibration frequency"), and irrelevant parameters are removed to reduce the amount of calculation.

[0049] The sub-model performs forward propagation calculations on the preprocessed data, extracts local features through a CNN layer, captures temporal trends through an LSTM layer, and finally generates analysis results through the output layer. For example, the analysis result for the inner ring grinding process of a bearing might be "Current grinding depth deviation is 0.002mm, within the allowable range (±0.005mm), surface roughness prediction value Ra 0.4μm, quality grade A".

[0050] The analysis results are compared with the standard_params in the label to verify whether there are any anomalies (e.g., if the rotation speed exceeds the range of 1500-1800 r / min, it is marked as "parameter anomaly"). Results that pass verification are stored in a standardized format (including analysis timestamp, process ID, quality indicators, anomaly identifier, and recommended measures) in the result database of the edge node and are simultaneously pushed to the equipment control panel (for real-time viewing by operators) and the cloud management platform (for overall production monitoring). Results that detect anomalies trigger tiered alarms (minor anomalies are indicated by indicator lights, while severe anomalies trigger audible and visual alarms and automatically pause the equipment) to ensure timely handling.

[0051] Through the above process, the edge nodes achieve closed-loop linkage of data, tags, and models, ensuring that the operational data of each process can be analyzed by a dedicated sub-model, effectively avoiding confusion of characteristics of different processes, and providing a reliable basis for real-time optimization of the production process.

[0052] In one embodiment, for S104, in industrial production, some equipment may undertake multiple processes. When process switching is frequent, such as more than three times per hour, traditional models, trained based on a fixed data distribution, require a large amount of new data to retrain and adjust parameters to adapt to changes when facing new processes. This is time-consuming, and the analysis results have large errors before training is complete, failing to meet the real-time and high-precision requirements of industrial production. The meta-learning mechanism, on the other hand, can utilize historical process switching experience to quickly optimize model parameters, enabling rapid adaptation to new processes and ensuring the accuracy and timeliness of real-time analysis of production data.

[0053] Specifically, such as Figure 3 As shown, the specific process of the meta-learning mechanism includes: S301, constructing a meta-learning dataset and collecting sample data from historical process switching times; S302, training the meta-learning dataset using a model-independent meta-learning algorithm to learn a fast-adaptive parameter update strategy; S303, when a new process switching occurs, based on a small amount of initial data from the new process, iteratively optimizing the model parameters through the parameter update strategy to reduce the model prediction error to below a preset threshold, wherein the sample data includes running data before and after the switching, model parameter adjustment records, and analysis error values.

[0054] Specifically, for the construction of meta-learning datasets, for example, during the production process, continuous collection of operational data, model parameter adjustment records, and analysis error values ​​before and after each process changeover is performed. The operational data includes equipment sensor data, operational parameters, and process identification information, maintaining consistency with the data type and format collected when constructing multi-process adaptive data analysis models. The model parameter adjustment log records the changes in parameters of each layer during each model training or fine-tuning session; the error analysis includes the deviation data between the model's prediction results and the actual production quality detection values. For example, this data is indexed by timestamps and uniformly stored in the meta-learning data repository of the edge nodes.

[0055] Meanwhile, based on the different characteristics of process switching scenarios, the data is labeled, such as the similarity of processes before and after the switch, and the load status of the equipment during the switch, forming a structured meta-learning dataset. The dataset is divided into a meta-training set and a meta-validation set in an 8:2 ratio for subsequent training and validation of the meta-learning model.

[0056] For model-agnostic meta-learning algorithms trained on meta-learning datasets, for example, during the meta-training phase, the process switching sample data in the meta-training set are grouped, with each group containing running data before and after the switch. The algorithm first uses this data to rapidly update the parameters of the current data analysis model to simulate the model's adaptation process when switching to a new process. Then, by calculating the loss function of the updated model on the meta-validation set, the effectiveness of the parameter update strategy is evaluated. Finally, based on the gradient of the loss function, the initial parameters of the model are adjusted in reverse, enabling the model to quickly adapt to new processes through a small number of gradient updates. After multiple iterations of training (e.g., a preset 5000 iterations), a general parameter update strategy applicable to various process switching scenarios is learned.

[0057] Optionally, when a new process switching event occurs, the meta-learning mechanism collects 50 sets of initial running data from the new process as adaptation samples. Based on the parameter update strategy learned in the meta-training phase, the parameters of the current data analysis model are rapidly iterated and optimized for 10-15 rounds. In each iteration, the prediction error of the model is calculated based on the adaptation samples, and the model parameters are adjusted according to the error gradient to rapidly reduce the model prediction error. During the optimization process, the model's performance on the meta-validation set is monitored in real time. When the prediction error drops below a preset threshold (e.g., 0.04), the iteration stops, and the optimized model parameters are determined.

[0058] Next, the optimized data analysis model is applied to the analysis of real-time running data. Following the same process as conventional analysis, the process context labels bound to the real-time running data are read, and the corresponding sub-model of the process is switched to (if it is a new process, the optimized default model is used) for inference calculation, and the analysis results are output to support production decisions. This will not be elaborated further here.

[0059] Figure 4 An edge computing-based real-time production data analysis system 400 is illustrated. This device embodiment is similar to... Figure 1 Corresponding to the illustrated method embodiments, the specific methods include: The data acquisition and preprocessing module 401 is used to collect the operating data of the production equipment under different processes to form a sample dataset. After preprocessing, the sample dataset is used to train sub-models for each process to build a multi-process adapted data analysis model. The operating data includes equipment sensor data, operating parameters and process identification information. The process context label binding module 402 is used to acquire process switching information in real time at the edge node and generate process context labels, and bind the process context labels to the real-time running data collected by the edge node; The real-time running data analysis module 403 is used to respond to data analysis requests from edge nodes, read the process context tags bound to the real-time running data, and automatically switch to the corresponding process sub-model based on the tag information to analyze the real-time running data. The optimization module 404 is used to activate the meta-learning mechanism when the process switching frequency exceeds the preset threshold, optimize the data analysis model parameters based on historical process switching sample data, and analyze the real-time running data based on the optimized data analysis model.

[0060] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A method for real-time analysis of production data based on edge computing, characterized in that, Includes the following steps: The sample dataset is formed by collecting the operating data of the production equipment under different processes. After preprocessing, the sample dataset is used to train sub-models for each process to build a multi-process adapted data analysis model. The operating data includes equipment sensor data, operating parameters and process identification information. At the edge node, process switching information is acquired in real time and process context tags are generated. The process context tags are then bound to the real-time running data collected by the edge node. In response to data analysis requests from edge nodes, the process context tags bound to real-time running data are read. Based on the tag information, the process is automatically switched to the corresponding sub-model to analyze the real-time running data. Specifically, the edge node reads the process context tags of the real-time data, extracts the process ID and matches it with the pre-stored process-model mapping table. If the match is successful, the corresponding sub-model is called. If a match fails, the default model is triggered and a label error message is issued. The default model uses transfer learning to initialize parameters. When the process switching frequency exceeds the preset threshold, the meta-learning mechanism is activated to optimize the data analysis model parameters based on historical process switching sample data, and to analyze the real-time operation data based on the optimized data analysis model. The meta-learning mechanism includes: constructing a meta-learning dataset and collecting sample data from historical process switching; training the meta-learning dataset with a model-independent meta-learning algorithm to learn a fast-adaptive parameter update strategy; when a new process switching occurs, iteratively optimizing the model parameters based on a small amount of initial data from the new process through the parameter update strategy to reduce the model prediction error to below a preset threshold. The sample data includes running data before and after the switching, model parameter adjustment records, and analysis error values.

2. The method for real-time analysis of production data based on edge computing according to claim 1, characterized in that, The equipment sensor data includes at least temperature, pressure, rotational speed, and vibration frequency; the operating parameters include equipment start / stop status, load rate, and processing accuracy settings; the process identification information includes process name, process number, process execution time, and corresponding quality standard parameters.

3. The method for real-time analysis of production data based on edge computing according to claim 1, characterized in that, The multi-process adaptive data analysis model adopts a modular architecture, with each process corresponding to an independent sub-model module. The sub-model adopts a hybrid structure of convolutional neural network and long short-term memory network. The number of neurons in the input layer matches the number of filtered feature parameters. There are 3 hidden layers with Leaky ReLU activation function. The output layer is the production quality prediction value or abnormal risk level.

4. The method for real-time analysis of production data based on edge computing according to claim 1, characterized in that, The process context tag is encoded in JSON format and includes the following fields: process ID, process name, switch timestamp, associated device ID, and standard parameter range.

5. The method for real-time analysis of production data based on edge computing according to claim 4, characterized in that, The sample dataset preprocessing includes data cleaning, standardization, and feature filtering. Data cleaning involves removing outliers caused by sensor malfunctions and fluctuations in data during device start-up and shutdown. The standardization process uses the Z-score standardization method to convert operational data of different dimensions into values ​​within a uniform range; feature selection uses the Pearson correlation coefficient method to retain feature parameters that are more correlated with production quality than a preset threshold.

6. A real-time production data analysis system based on edge computing, characterized in that, include: The data acquisition and preprocessing module is used to collect the operating data of the production equipment under different processes to form a sample dataset. After preprocessing, the sample dataset is used to train sub-models for each process to build a multi-process adapted data analysis model. The operating data includes equipment sensor data, operating parameters and process identification information. The process context label binding module is used to acquire process switching information in real time at the edge node and generate process context labels, and bind the process context labels to the real-time running data collected by the edge node; The real-time running data analysis module is used to respond to data analysis requests from edge nodes, read the process context tags bound to the real-time running data, and automatically switch to the corresponding process sub-model based on the tag information to analyze the real-time running data. Specifically, the edge node reads the process context tags of the real-time data, extracts the process ID and matches it with the pre-stored process-model mapping table. If the match is successful, the corresponding sub-model is called. If a match fails, the default model is triggered and a label error message is issued. The default model uses transfer learning to initialize parameters. The optimization module is used to activate the meta-learning mechanism when the process switching frequency exceeds a preset threshold. It optimizes the data analysis model parameters based on historical process switching sample data and analyzes the real-time running data based on the optimized data analysis model. The meta-learning mechanism includes: constructing a meta-learning dataset and collecting sample data from historical process switching; training the meta-learning dataset with a model-independent meta-learning algorithm to learn a fast-adaptive parameter update strategy; when a new process switching occurs, iteratively optimizing the model parameters based on a small amount of initial data from the new process through the parameter update strategy to reduce the model prediction error to below a preset threshold. The sample data includes running data before and after the switching, model parameter adjustment records, and analysis error values.

Images