Intelligent factory data acquisition platform and method thereof

Abstract

The intelligent factory data acquisition platform comprises a data sensing layer, a multi-protocol transmission layer, a data fusion layer and an application service layer, the data sensing layer is used for collecting equipment state and environment parameter multi-source heterogeneous data through various sensors, preprocessing the collected data and transmitting the preprocessed data to the multi-protocol transmission layer; the multi-protocol transmission layer is used for receiving the preprocessed data transmitted by the data sensing layer, packaging the data through a multi-protocol data transmission unit, then carrying out multi-protocol transmission, and outputting the transmitted data to the data fusion layer; the data fusion layer is used for receiving the multi-source heterogeneous data transmitted by the multi-protocol transmission layer, performing feature extraction and fusion on the heterogeneous data by adopting a self-correcting convolutional recurrent neural network, and outputting fused feature representation to the application service layer; and the application service layer is used for receiving the fusion features output by the data fusion layer, and performing equipment health diagnosis, process parameter optimization and energy consumption statistical analysis intelligent application services based on the fusion features.

Description

Technical Field

[0001] The present invention relates to the technical field of data acquisition platforms, and more specifically, to a smart factory data acquisition platform and method thereof. Background Art

[0002] The various sensors deployed in industrial equipment continuously generate massive amounts of time-series data, containing crucial information about equipment health, production efficiency, and energy consumption. Effectively collecting, transmitting, integrating, and analyzing industrial big data has become a hot topic of concern in both academia and industry.

[0003] Currently, industrial data collection and analysis mainly have the following technical problems:

[0004] First, industrial sites are often populated with heterogeneous sensors from multiple sources. The data collected varies greatly in type, format, and frequency, making direct fusion and correlation analysis difficult. Existing data acquisition systems are mostly designed for a single sensor type and lack the ability to uniformly process and fuse heterogeneous data. Fusion of heterogeneous data requires extensive preprocessing, including data cleaning, conversion, and alignment, resulting in inefficient data analysis.

[0005] Secondly, time series data collected by industrial sensors is generally characterized by non-stationarity, high noise, and information redundancy, posing significant challenges to feature extraction and data modeling. Traditional feature engineering methods rely on manual design and domain knowledge, making it difficult to adaptively extract key features under different operating conditions. Existing time series analysis algorithms, such as ARIMA and Kalman filtering, mostly assume that the data follows a specific probability distribution, making it difficult to characterize the complex dynamic characteristics of industrial data.

[0006] Third, intelligent applications for equipment health diagnosis and process optimization mostly rely on shallow machine learning models, such as support vector machines and decision trees, whose performance relies heavily on the quality of manual feature engineering. Shallow models have limited feature representation capabilities, making it difficult to uncover deep correlations and time-varying patterns in time series data. Diagnostic and optimization algorithms developed for specific applications lack versatility and transferability, making them difficult to continuously learn and adapt in dynamic and changing industrial environments.

[0007] Taking mechanical fault diagnosis in equipment health diagnostics as an example, existing methods primarily extract features and classify faults based on vibration signals from mechanical equipment. While computationally simple, traditional time-domain statistical features (mean, variance, and kurtosis) and frequency-domain features (Fourier transform and wavelet transform) struggle to characterize the nonstationary and nonlinear characteristics of vibration signals. Some researchers have attempted to introduce shallow machine learning models, such as support vector machines and neural networks, but their diagnostic performance relies heavily on the quality of feature engineering, and their generalization capabilities are limited. In recent years, deep learning technology has achieved breakthroughs in image and speech processing, and some researchers have attempted to apply convolutional neural network models to mechanical fault diagnosis. However, existing deep learning diagnostic methods mostly rely on single-source sensor data and lack mechanisms for integrating heterogeneous multi-source data, making it difficult to fully exploit the complementarity and correlation between different sensor data. Furthermore, existing methods rarely consider the coordinated optimization of equipment fault diagnosis with process parameter optimization and energy consumption analysis, resulting in a lack of a fully closed-loop system architecture.

[0008] To sum up, industrial big data analysis urgently needs a new system architecture and data fusion method. Summary of the Invention

[0009] It is in this context that the present invention is proposed, aiming to address the limitations of existing technologies, provide a smart factory data acquisition platform and method, adaptively fuse multi-source heterogeneous sensor data, extract key features end-to-end, and simultaneously achieve a closed business loop at the intelligent application layer, conduct innovative design at the system and algorithm levels, achieve breakthroughs in key links such as data fusion, feature learning, and intelligent decision-making, and form a scalable, portable, and explainable overall solution for smart manufacturing.

[0010] The present invention provides a smart factory data acquisition platform, which includes: a data perception layer, a multi-protocol transmission layer, a data fusion layer and an application service layer;

[0011] The data perception layer is used to collect multi-source heterogeneous data on device status and environmental parameters through multiple sensors, and pre-process the collected data before transmitting it to the multi-protocol transmission layer;

[0012] The multi-protocol transmission layer is used to receive the pre-processed data transmitted by the data perception layer, and to package and encapsulate the data through the multi-protocol data transmission unit for multi-protocol transmission, and output the transmitted data to the data fusion layer;

[0013] The data fusion layer is used to receive multi-source heterogeneous data transmitted by the multi-protocol transport layer, use a self-correcting convolutional recurrent neural network to extract and fuse features of the heterogeneous data, and output the fused feature representation to the application service layer;

[0014] The application service layer is used to receive the fusion features output by the data fusion layer, perform equipment health diagnosis, process parameter optimization, and energy consumption statistical analysis intelligent application services based on the fusion features, and feed back the analysis results to the data perception layer and multi-protocol transmission layer to achieve closed-loop control based on the collected data.

[0015] Specifically, the data perception layer includes:

[0016] Vibration sensor, used to collect vibration signals from equipment and convert them into voltage signals for output;

[0017] Noise sensor, used to collect noise signals generated during equipment operation and convert the noise signals into voltage signals for output;

[0018] Temperature and humidity sensor, used to collect the temperature and humidity parameters of the environment where the equipment is located, and convert the temperature and humidity parameters into voltage signal output;

[0019] Visual sensors are used to capture images of key components of the equipment and extract image feature parameters;

[0020] The data preprocessing unit is used to receive the voltage signals of the vibration sensor, noise sensor, temperature and humidity sensor and the image feature parameters of the visual sensor, perform data screening, denoising, and normalization preprocessing, and transmit the preprocessed data to the multi-protocol transmission layer.

[0021] Specifically, the multi-protocol transport layer includes:

[0022] The multi-protocol data transmission unit is used to receive the pre-processed data transmitted by the data pre-processing unit, package and compress the data according to the preset packaging format, and support multiple industrial communication protocols to transmit data;

[0023] The industrial Ethernet communication unit is used to transmit the packaged and encapsulated data through the industrial Ethernet protocol under the control of the multi-protocol data transmission unit;

[0024] A wireless communication unit, configured to transmit the packaged and encapsulated data via a wireless communication protocol under the control of the multi-protocol data transmission unit;

[0025] The time-sensitive network synchronization unit is used to synchronize and schedule the data transmission of the industrial Ethernet communication unit and the wireless communication unit, ensure the real-time and reliability of key data, and output the final transmission data to the data fusion layer.

[0026] Specifically, the self-correcting convolutional recurrent neural network includes:

[0027] (1) Use 1×1, 3×3, and 5×5 convolution kernels to process the time series data X of N sensorsn Perform local feature extraction to obtain

[0028]

[0029] in Represents the time series data of the nth sensor, T is the number of time steps D n is the data dimension; k is the convolution kernel size; ReLU is the rectified linear unit activation function

[0030] ReLU(x)=max(0,x);

[0031] (2) Use the bidirectional long short-term memory network BiLSTM to extract the global temporal feature H n :

[0032]

[0033] in, Represent the hidden state outputs of the forward and backward LSTMs respectively, and ⊕ represents the tensor concatenation operation;

[0034] (3) Use attention mechanism to analyze different sensor features H n Assign dynamic weights and obtain the fusion feature Z:

[0035] A n =softmax(tanh(H n W a ))

[0036] Z=∑A n ·H n

[0037] Among them, A n is the attention weight of the nth sensor feature, W a is the attention parameter matrix, is the normalized exponential function;

[0038] 4) Calculate the confidence matrix Q based on the mutual information, perform self-correction on Z, and output the final fusion feature O:

[0039]

[0040] O=Q·Z

[0041] in is the longitudinal concatenation of all sensor features, H T is the transpose of H, d k is the hidden state dimension, mutual information It is used to measure the correlation between different sensor features, and Q is the normalized confidence matrix.

[0042] Specifically, the application service layer includes:

[0043] The equipment health diagnosis module is used to build an equipment health status model based on fusion features, and realize equipment fault diagnosis and predictive maintenance through anomaly detection and remaining life prediction;

[0044] The process parameter optimization module is used to mine and optimize process parameters based on fusion features. It combines the expert knowledge base and reinforcement learning algorithm to adjust equipment parameters in real time to improve production efficiency and product qualification rate.

[0045] Energy consumption statistics and analysis module, used to collect statistics and analyze equipment energy consumption based on fusion features, automatically generate energy consumption reports, explore energy-saving potential, and formulate energy-saving strategies;

[0046] The application service interface module is used to provide a standardized interface for equipment health diagnosis, process parameter optimization and energy consumption statistical analysis, realize data interaction and business collaboration with MES and ERP upper-level systems, and feed back the analysis results to the data perception layer and multi-protocol transmission layer through the data bus, forming a closed loop of data collection, analysis and control.

[0047] Specifically, the vibration sensor uses a PCB 356A16 three-axis accelerometer with a frequency response range of 0.5-5000Hz and a sensitivity of 50g; the noise sensor uses a GRAS46AE microphone with a frequency range of 3.15Hz-20kHz and a dynamic range of 17-146dB; the temperature and humidity sensor uses a Sensirion SHT35 digital temperature and humidity sensor with a temperature measurement range of -40-125°C, a humidity measurement range of 0-100%RH, and an accuracy of ±0.1°C and ±1.5%RH; the visual sensor uses a Keyence CV-X150 series smart camera with a resolution of 640×480, 200 frames / second, and is equipped with a LumiTrax light source.

[0048] Specifically, the industrial Ethernet communication unit supports Modbus TCP, PROFINET, and EtherCAT industrial Ethernet protocols; the wireless communication unit supports LoRaWAN and NB-IoT low-power wide area network protocols; the synchronization accuracy of the time-sensitive network synchronization unit is better than 1μs, and the clock drift is less than 10ns / s.

[0049] Specifically, the training process of the self-correcting convolutional recurrent neural network includes:

[0050] (1) Construct a loss function that includes a self-correcting recurrent neural network:

[0051]

[0052] Where Y is the softmax(O·W o )Calculate the predicted output, Y * is the true label, CrossEntropy(Y,Y * )=-∑Y * ·log(Y) is the cross entropy loss function, Q is the confidence matrix, I is the identity matrix, ||·|| F is the Frobenius norm of the matrix, λ is the regularization coefficient;

[0053] (2) Use the small batch stochastic gradient descent algorithm to adjust the neural network parameters W a 、W o And LSTM parameters are iteratively optimized to minimize the loss function L:

[0054]

[0055] Where W refers to all parameters to be optimized, η is the learning rate, is the gradient of the loss function with respect to the parameters;

[0056] (3) By setting an early stopping threshold, the validation set loss is used as a monitoring indicator, and the iteration is stopped when the threshold is reached to prevent overfitting.

[0057] The smart factory data collection method includes the following steps:

[0058] Step 1: Use vibration sensors, noise sensors, and temperature and humidity sensors to collect equipment status data and environmental parameter data, and use visual sensors to collect images of key equipment components. The collected data is filtered, denoised, and normalized to obtain preprocessed data.

[0059] Step 2: The pre-processed data is packaged, encapsulated and compressed through the multi-protocol data transmission unit, multi-protocol data transmission is performed through the industrial Ethernet communication unit and the wireless communication unit, and time synchronization is performed through the time-sensitive network synchronization unit, and the transmitted data is output to the data fusion layer;

[0060] Step 3: Extract local and global features of multi-source heterogeneous sensor data through self-correcting convolutional recurrent neural network, fuse multi-sensor data through attention mechanism and confidence matrix self-correction, and output the fused features to the application service layer;

[0061] Step 4: Through the equipment health diagnosis, process parameter optimization and energy consumption statistical analysis application service modules, intelligent analysis and optimization decisions are made based on the fusion characteristics, and the analysis results are fed back to the data perception layer and multi-protocol transmission layer through the application service interface to achieve closed-loop control.

[0062] Specifically, the training method of the self-correcting convolutional recurrent neural network in step 3 includes:

[0063] A loss function containing a self-correcting recurrent neural network is constructed, and the neural network parameters are iteratively optimized using a small-batch stochastic gradient descent algorithm. Overfitting is prevented by setting an early stopping threshold and using the validation set loss as a monitoring indicator, and the optimal model parameters after training are output.

[0064] The present invention has the following beneficial effects:

[0065] 1. First, at the system level, this invention adopts a modular and hierarchical design concept, decoupling data acquisition, transmission, fusion, and application, forming a closed-loop control data flow architecture. The data perception layer integrates multiple heterogeneous sensors to comprehensively perceive equipment status and process parameters. The multi-protocol transport layer is compatible with multiple industrial communication protocols, solving the problem of heterogeneous network interoperability. The data fusion layer and application service layer are decoupled through a data bus, enabling the independent evolution of data processing and business applications. This architecture is scalable, maintainable, and robust, and can flexibly adapt to the diversity and dynamic changes of industrial scenarios.

[0066] Secondly, at the algorithmic level, the core innovation of this invention lies in a self-correcting convolutional recurrent neural network. This network integrates multiple machine learning models, including CNN, RNN, attention, and confidence matrix, and can adaptively extract key features from multi-sensor data in both the temporal and spatial domains, overcoming the challenges of industrial data non-stationarity, high noise, and information redundancy. Specifically:

[0067] 1) Multi-scale one-dimensional convolution can adaptively extract local features at different time scales, taking into account both the instantaneous characteristics of the signal and the long-term trend, thus improving the robustness of feature representation;

[0068] 2) Bidirectional LSTM can mine the forward and backward dependencies of time series data and learn more comprehensive and in-depth time series features;

[0069] 3) The attention mechanism can adaptively assign different importance weights to different sensor data, highlighting the contribution of key sensors and suppressing the interference of redundant sensors;

[0070] 4) The confidence matrix characterizes the correlation and consistency between data from different sensors. By reweighting the fused features, the quality and reliability of data fusion are further improved. These four mechanisms form an organic synergy and complementarity, improving the accuracy and generalization performance of data fusion while also enhancing the model's robustness to anomalies and noise. Furthermore, compared to non-end-to-end manual feature engineering, the end-to-end feature learning paradigm significantly improves the automation and intelligence of feature extraction.

[0071] Thirdly, the intelligent applications of equipment health diagnosis, process parameter optimization, and energy consumption analysis provided by the present invention at the application service layer create synergistic benefits and complementary advantages with the data fusion layer. On the one hand, equipment health diagnosis relies on the cross-modal fusion features provided by the data fusion layer to accurately and comprehensively characterize the health status of the equipment. On the other hand, the prediction results and tuning instructions generated by process parameter optimization and energy consumption analysis are fed back to the data perception layer via the data bus, forming a real-time closed-loop control. The seamless coupling of data fusion and intelligent applications not only improves the accuracy of data analysis, but also enables the rapid implementation of analysis results, forming a benign interaction between data, algorithms, and business.

[0072] Finally, during its implementation, the present invention also properly resolved several technical contradictions and conflicts. For example, although multi-sensor data fusion can provide richer information, it also brings about data redundancy and coupling problems. To address this contradiction, the present invention innovatively introduces an attention mechanism and a confidence matrix. By adaptively weighting and resetting the confidence, it suppresses data redundancy while maximizing the complementarity of cross-modal data. For another example, although end-to-end neural networks simplify manual feature engineering, they also face the problems of high model complexity and large computational workload. To address this contradiction, the present invention introduces the L2 regularization term of the confidence matrix into the loss function. By limiting the rank and sparsity of the matrix, it reduces the model complexity while improving the fusion performance, and improves the training and inference efficiency.

[0073] In summary, this invention achieves global optimization and collaborative design across system architecture, core algorithms, and intelligent applications, forming an innovative mechanism of complementarity, superposition, and efficiency. The introduction of innovative modules such as multi-scale convolution, attention, and confidence matrices effectively balances and resolves common technical contradictions in industrial big data analysis. While improving data fusion and feature learning performance, it also addresses the non-functional requirements of system reliability, real-time performance, and interpretability. Overall, this method achieves technical results superior to conventional methods, providing new insights and solutions for the development of industrial artificial intelligence. BRIEF DESCRIPTION OF THE DRAWINGS

[0074] Figure 1 This is a schematic diagram of the framework principle of the smart factory data acquisition platform of the present invention. DETAILED DESCRIPTION

[0075] To further illustrate the technical means and effectiveness of the present invention in achieving its intended objectives, the following, in conjunction with the accompanying drawings and preferred embodiments, describes in detail the specific implementation, structure, features, and effectiveness of the smart factory data acquisition platform proposed in accordance with the present invention. In the following description, references to "one embodiment" or "another embodiment" do not necessarily refer to the same embodiment. Furthermore, specific features, structures, or characteristics of one or more embodiments may be combined in any suitable manner.

[0076] Unless defined otherwise, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention belongs.

[0077] The specific scheme of the smart factory data acquisition platform provided by the present invention is described in detail below with reference to the accompanying drawings.

[0078] Example 1

[0079] See also Figure 1 The present invention provides a smart factory data acquisition platform and data acquisition method. This platform collects multi-source heterogeneous data through a data perception layer, transmits data using a multi-protocol transport layer, extracts and fuses features from the heterogeneous data in a data fusion layer, and finally implements intelligent analysis and optimized decision-making through an application service layer. The entire system implements automated closed-loop control of data acquisition, transmission, analysis, and feedback. The smart factory data acquisition platform of the present invention comprises a data perception layer 1, a multi-protocol transport layer 2, a data fusion layer 3, and an application service layer 4.

[0080] Among them, the data perception layer 1 is used to collect multi-source heterogeneous data such as device status and environmental parameters through various sensors, and pre-process the collected data before transmitting it to the multi-protocol transmission layer.

[0081] In a preferred embodiment, the data perception layer includes a vibration sensor, a noise sensor, a temperature and humidity sensor, a visual sensor and a data preprocessing unit.

[0082] Vibration sensors are used to collect vibration signals from equipment and convert them into voltage output. A preferred vibration sensor is the PCB 356A16 triaxial accelerometer, which has a frequency response range of 0.5-5000Hz and a sensitivity of 50g. This sensor accurately captures equipment vibration waveforms, offers high sensitivity, and a wide frequency response, meeting vibration detection requirements in most industrial scenarios.

[0083] Noise sensors are used to collect noise signals generated by equipment during operation and convert them into voltage output signals. A GRAS46AE microphone is ideal for this noise sensor, with a frequency range of 3.15Hz-20kHz and a dynamic range of 17-146dB. This sensor has high sensitivity, a wide frequency response, and a large dynamic range, enabling it to accurately detect noise fluctuations in noisy industrial environments.

[0084] The temperature and humidity sensor collects the temperature and humidity parameters of the device's environment and converts them into voltage signals for output. The Sensirion SHT35 digital temperature and humidity sensor is a preferred choice, offering a temperature measurement range of -40-125°C and a humidity measurement range of 0-100% RH, with an accuracy of ±0.1°C and ±1.5% RH, respectively. This integrated temperature and humidity sensor offers a wide measurement range, high accuracy, and a digital output for convenient data processing.

[0085] Vision sensors are used to capture images of key equipment components and extract image feature parameters. Keyence's CV-X150 series smart camera is a preferred choice for this purpose. It features a resolution of 640×480, a sampling rate of 200 frames per second, and is equipped with a LumiTrax light source. This smart camera integrates image acquisition, processing, and analysis capabilities. Its resolution and sampling rate meet the requirements for equipment defect detection, and its accompanying light source ensures high-quality image acquisition.

[0086] The data preprocessing unit receives voltage signals from vibration sensors, noise sensors, and temperature and humidity sensors, as well as image feature parameters from visual sensors. It then performs preprocessing, including data screening, denoising, and normalization, before transmitting the processed data to the multi-protocol transport layer. Common algorithms such as Kalman filtering, wavelet transforms, and Z-score normalization can be used for data preprocessing to remove invalid or outliers, improve data quality, and standardize the data into a standard format for subsequent data fusion.

[0087] In a preferred embodiment, the multi-protocol transport layer 2 is used to receive pre-processed data transmitted by the data perception layer, and to package and encapsulate the data through the multi-protocol data transmission unit for multi-protocol transmission, and output the transmitted data to the data fusion layer.

[0088] The multi-protocol transport layer includes a multi-protocol data transmission unit, an industrial Ethernet communication unit, a wireless communication unit and a time-sensitive network synchronization unit.

[0089] The multi-protocol data transmission unit receives preprocessed data from the data preprocessing unit, packages and compresses the data according to a preset packaging format, and supports data transmission using a variety of industrial communication protocols. This packaging format can use compact data representations such as JSON and Protocol Buffers, and lossless compression algorithms such as GZIP and ZLIB can be used to reduce data size while ensuring data integrity. The multi-protocol data transmission unit supports industrial Ethernet protocols such as ModbusTCP, PROFINET, and EtherCAT, as well as low-power wide area network protocols such as LoRaWAN and NB-IoT, allowing flexible selection of transmission methods based on different application scenarios.

[0090] The Industrial Ethernet communication unit, under the control of the multi-protocol data transmission unit, transmits packaged and encapsulated data using the Industrial Ethernet protocol. The Industrial Ethernet protocol provides higher data transmission bandwidth and lower latency while ensuring real-time performance and reliability, making it suitable for transmitting large amounts of time-series data.

[0091] The wireless communication unit, under the control of the multi-protocol data transmission unit, transmits the packaged and encapsulated data via wireless communication protocols. Low-power wide-area network protocols such as LoRaWAN and NB-IoT use a star topology and long-distance communication, making them suitable for deployment in scenarios with dispersed sensor nodes.

[0092] The Time-Sensitive Network (TSN) synchronization unit synchronizes and schedules data transmission between the Industrial Ethernet and wireless communication units, ensuring the real-time and reliability of critical data. The final transmitted data is then output to the data fusion layer. Time synchronization utilizes the IEEE 1588 Precision Time Protocol (PTP), achieving nanosecond-level synchronization accuracy through a master-slave clock synchronization mechanism. The scheduling mechanism utilizes static scheduling based on time slicing, allocating fixed transmission slots to data streams of different priorities, ensuring deterministic transmission latency.

[0093] In a preferred embodiment, the data fusion layer 3 is used to receive multi-source heterogeneous data transmitted by the multi-protocol transport layer, use a self-correcting convolutional recurrent neural network to extract and fuse features of the heterogeneous data, and output the fused feature representation to the application service layer.

[0094] The self-correcting convolutional recurrent neural network first uses convolution kernels of different scales such as 1x1, 3x3, and 5x5 to analyze the time series data X of N sensors. n Perform local feature extraction to obtain feature maps of different receptive fields in Represents the time series data of the nth sensor, T is the number of time steps, D n is the data dimension; the convolution operation can be expressed as:

[0095]

[0096] in Represents the time series data of the nth sensor, T is the number of time steps D n is the data dimension; k is the convolution kernel size; ReLU is the rectified linear unit activation function

[0097] ReLU(x)=max(0,x);

[0098] Next, the self-correcting convolutional recurrent neural network uses a bidirectional long short-term memory network (BiLSTM) to extract forward and backward global temporal features of different sensor data. and Then the complete global feature H is obtained by splicing n By introducing LSTM units in both forward and backward directions, BiLSTM can simultaneously utilize the past and future contextual information of the data to enhance feature representation capabilities. The calculation process of BiLSTM is:

[0099]

[0100] in, Represent the hidden state outputs of the forward and backward LSTMs respectively, and ⊕ represents the tensor concatenation operation;

[0101] In order to adaptively fuse the features of different sensors, the self-correcting convolutional recurrent neural network uses the attention mechanism to n Assign dynamic weights to obtain fusion features Z. Attention weight A n By setting the sensor characteristic H n and the attention parameter matrix W a Multiply, and then calculate through tanh activation and softmax normalization to measure the importance of different sensor features. The softmax function can map the input vector x into a probability distribution, and the calculation formula of its i-th element is

[0102]

[0103] The sensor features are weighted and summed by the comprehensive attention weight to obtain the fusion feature Z. The calculation formula is:

[0104] A n =softmax(tanh(H n W a ))

[0105]

[0106] Finally, in order to further improve the representation ability and robustness of the fusion feature, the self-correcting convolutional recurrent neural network introduces the confidence matrix Q to self-correct the fusion feature Z. The confidence matrix Q is obtained by calculating the pairwise mutual information between different sensor features and then normalizing it with softmax. The mutual information I(H i ,H j ) is defined as two sensor characteristics H i and H jThe ratio of the inner product of the confidence matrix Q to the square root of the feature dimension is used to measure the correlation between different features. By multiplying the confidence matrix Q with the fusion feature Z, the fusion feature can be adaptively corrected to reduce the impact of abnormal or redundant features. The complete calculation process is:

[0107] H=[H1;...;H N ]

[0108]

[0109] O=Q·Z

[0110] where H∈R N T×d k represents the vertical concatenation of all sensor features, H T is the transpose of H, d k is the hidden state dimension, and O is the final corrected fusion feature.

[0111] In a preferred embodiment, the application service layer 4 is used to receive the fusion features output by the data fusion layer, perform intelligent application services such as equipment health diagnosis, process parameter optimization, and energy consumption statistical analysis based on the fusion features, and feed back the analysis results to the data perception layer and the multi-protocol transmission layer to achieve closed-loop control based on the collected data.

[0112] The application service layer includes equipment health diagnosis module, process parameter optimization module, energy consumption statistical analysis module and application service interface module.

[0113] The equipment health diagnosis module is used to construct an equipment health status model based on fused features, enabling equipment fault diagnosis and predictive maintenance through anomaly detection and remaining life prediction. The equipment health status model can be trained using historical operating data using machine learning algorithms such as support vector machines and random forests, with fused features as input and equipment health status as output. Anomaly detection constructs a baseline distribution of fused features under normal conditions and uses density estimation, distance measurement, and other methods to determine whether the current fused features deviate from the baseline distribution, thereby enabling real-time diagnosis of equipment faults. Remaining life prediction uses the regression relationship between fused features and lifespan, combined with the current equipment state and historical degradation trends, to estimate the remaining useful life of the equipment, thereby formulating a predictive maintenance strategy.

[0114] The process parameter optimization module is used to mine and optimize process parameters based on fused features. Combining an expert knowledge base with a reinforcement learning algorithm, it adjusts equipment parameters in real time to improve production efficiency and product qualification rates. Process parameter mining constructs a mapping relationship between fused features and process parameters, using statistical analysis and causal inference to identify key parameters that significantly impact product quality and energy consumption. Process parameter optimization utilizes a reinforcement learning framework, modeling the parameter adjustment problem as a Markov decision process. The optimal parameter combination strategy is gradually learned through trial and error and a reward-and-penalty mechanism. The expert knowledge base is used to constrain the parameter value range, improving the convergence speed and interpretability of the optimization.

[0115] The energy consumption statistics and analysis module is used to collect statistics and analyze equipment energy consumption based on fused features, automatically generate energy consumption reports, explore energy-saving potential, and formulate energy-saving strategies. Energy consumption statistics use the equipment operating status reflected by fused features, combined with energy consumption sensor data, to calculate the equipment's energy consumption level and energy efficiency indicators under different operating conditions. Energy consumption analysis uses data mining algorithms such as clustering and association rule mining to discover energy consumption patterns and influencing factors from massive amounts of historical data, and identify energy consumption anomalies and waste through comparative analysis. Energy-saving potential is explored by constructing energy consumption prediction models, simulating the effects of different energy-saving measures, and evaluating the return on investment and emission reduction benefits of energy-saving projects.

[0116] The application service interface module provides a standardized interface for equipment health diagnosis, process parameter optimization, and energy consumption statistical analysis. It enables data interaction and business collaboration with upper-layer systems such as MES and ERP. Analysis results are fed back to the data perception layer and multi-protocol transport layer via a data bus, forming a closed loop of data collection, analysis, and control. The application service interface adopts a RESTful architectural style, providing a Web API based on the HTTP protocol and supporting common data formats such as JSON and XML, facilitating integration and interaction between different systems. The interface design adheres to resource-oriented principles, using URLs to locate resources and HTTP verbs to describe operations. It supports security and quality assurance mechanisms such as authentication, permission management, and flow control.

[0117] The present invention also provides a smart factory data acquisition method, which includes four steps: data perception, multi-protocol transmission, data fusion and intelligent application.

[0118] In the data perception step, equipment status data and environmental parameter data are collected through vibration sensors, noise sensors, and temperature and humidity sensors, and images of key components of the equipment are collected through visual sensors. The collected data are filtered, denoised, and normalized to obtain preprocessed data.

[0119] In the multi-protocol transmission step, the pre-processed data is packaged, encapsulated and compressed through the multi-protocol data transmission unit, multi-protocol data transmission is carried out through the industrial Ethernet communication unit and the wireless communication unit, and time synchronization is carried out through the time-sensitive network synchronization unit, and the transmitted data is output to the data fusion layer.

[0120] In the data fusion step, the local and global features of multi-source heterogeneous sensor data are extracted through a self-correcting convolutional recurrent neural network, and the multi-sensor data is self-corrected and fused through the attention mechanism and confidence matrix, and the fused features are output to the application service layer.

[0121] In the intelligent application step, through application service modules such as equipment health diagnosis, process parameter optimization and energy consumption statistical analysis, intelligent analysis and optimization decisions are made based on fusion features. The analysis results are fed back to the data perception layer and multi-protocol transmission layer through the application service interface to achieve closed-loop control.

[0122] The training method of the self-correcting convolutional recurrent neural network includes: constructing a loss function containing the self-correcting recurrent neural network, iteratively optimizing the neural network parameters using a small-batch stochastic gradient descent algorithm, preventing overfitting by setting an early stopping threshold and using the validation set loss as a monitoring indicator, and outputting the optimal model parameters after training.

[0123] Among them, the loss function L is defined as the weighted sum of the cross entropy loss and the confidence matrix L2 regularization term:

[0124]

[0125] Where Y is the softmax(O·W o )Calculate the predicted output, Y * is the true label, CrossEntropy(Y,Y * ) represents Y and Y * The cross entropy of ||·|| F represents the Frobenius norm of the matrix, I is the identity matrix, and λ is the regularization coefficient. Mini-batch stochastic gradient descent randomly extracts a batch of samples from the training set, calculates the gradient of the loss function with respect to the neural network parameters, and updates the parameters in the opposite direction of the gradient until the loss function converges. The parameter update formula is:

[0126]

[0127] Where W represents all the parameters of the neural network, including the attention parameter matrix W a , output layer parameter matrix W o And the gate parameter matrix of LSTM, η is the learning rate, is the gradient of the loss function with respect to the parameters.

[0128] The early stopping strategy evaluates the model loss on the validation set after each training epoch. If the loss does not decrease over several consecutive epochs, training is terminated promptly to prevent overfitting of the model to the training set. Hyperparameter optimization methods such as grid search are used to select the optimal learning rate, regularization coefficient, and early stopping threshold, ultimately achieving the model parameters that achieve the best generalization performance.

[0129] In summary, the intelligent factory data acquisition platform proposed in this invention builds a complete industrial big data solution through the seamless integration and coordinated optimization of data perception, multi-protocol transmission, data fusion, and intelligent applications. It innovatively introduces a self-correcting convolutional recurrent neural network, which achieves adaptive fusion of multi-source heterogeneous sensor data through an attention mechanism and confidence matrix, overcoming the challenges of industrial data complexity, dynamics, and heterogeneity. The various intelligent applications at the application service layer further unlock the value of massive amounts of industrial data, addressing business needs such as equipment health management, production process optimization, and energy consumption analysis, and providing important support for improving the intelligence level of industrial enterprises. The data acquisition method provided by this invention also actively explores and beneficially implements the data-driven intelligent manufacturing paradigm.

[0130] It should be noted that the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, replacements, and improvements made within the principles of the present invention should be included in the scope of protection of the present invention.

Claims

1. Smart factory data acquisition platform, characterized by ,The platform includes: a data perception layer, a multi-protocol transmission layer, a data fusion layer and an application service layer; The data perception layer is used to collect multi-source heterogeneous data of equipment status and environmental parameters through multiple sensors, and transmit the collected data to the multi-protocol transmission layer after pre-processing; The multi-protocol transmission layer is used to receive the pre-processed data transmitted by the data perception layer, and to package and encapsulate the data through the multi-protocol data transmission unit for multi-protocol transmission, and output the transmitted data to the data fusion layer; The data fusion layer is used to receive multi-source heterogeneous data transmitted by the multi-protocol transmission layer, use a self-correcting convolutional recurrent neural network to extract and fuse features of the heterogeneous data, and output the fused feature representation to the application service layer; The application service layer is used to receive the fusion features output by the data fusion layer, perform equipment health diagnosis, process parameter optimization, and energy consumption statistical analysis intelligent application services based on the fusion features, and feed back the analysis results to the data perception layer and the multi-protocol transmission layer to achieve closed-loop control based on the collected data.

2. The intelligent factory data acquisition platform according to claim 1 is characterized in that , the data perception layer includes: Vibration sensor, used to collect vibration signals of equipment and convert the vibration signals into voltage signals for output; Noise sensor, used to collect noise signals generated when the equipment is running, and convert the noise signals into voltage signals for output; Temperature and humidity sensor, used to collect the temperature and humidity parameters of the environment where the equipment is located, and convert the temperature and humidity parameters into voltage signal output; Visual sensors are used to capture images of key components of equipment and extract image feature parameters; The data preprocessing unit is used to receive the voltage signals of the vibration sensor, noise sensor, temperature and humidity sensor and the image feature parameters of the visual sensor, perform data screening, denoising and normalization preprocessing, and transmit the preprocessed data to the multi-protocol transmission layer.

3. The intelligent factory data acquisition platform according to claim 1 is characterized in that , the multi-protocol transport layer includes: A multi-protocol data transmission unit is used to receive the pre-processed data transmitted by the data pre-processing unit, package and compress the data according to a preset packaging format, and support multiple industrial communication protocols to transmit the data; The industrial Ethernet communication unit is used to transmit the packaged and encapsulated data through the industrial Ethernet protocol under the control of the multi-protocol data transmission unit; A wireless communication unit, used to transmit the packaged and encapsulated data through a wireless communication protocol under the control of the multi-protocol data transmission unit; The time-sensitive network synchronization unit is used to synchronize and schedule the data transmission of the industrial Ethernet communication unit and the wireless communication unit, ensure the real-time and reliability of key data, and output the final transmission data to the data fusion layer.

4. The intelligent factory data acquisition platform according to claim 1 is characterized in that , the self-correcting convolutional recurrent neural network includes: (1) Use 1×1, 3×3, and 5×5 convolution kernels to process the time series data X of N sensors. n Perform local feature extraction and obtain in Represents the time series data of the nth sensor, T is the number of time steps D n is the data dimension; k is the convolution kernel size; ReLU is the rectified linear unit activation function ReLU(x)=max(0,x); (2) Use the bidirectional long short-term memory network BiLSTM to extract the global temporal features H n : in, Represent the hidden state outputs of the forward and backward LSTMs, respectively. Represents a tensor concatenation operation; (3) Use the attention mechanism to analyze different sensor features H n Assign dynamic weights and obtain the fusion feature Z: A n =softmax(tanh(H n ·W a )) Z=∑A n ·H n Among them, A n is the attention weight of the nth sensor feature, W a is the attention parameter matrix, is the normalized exponential function; 4) Calculate the confidence matrix Q based on mutual information, perform self-correction on Z, and output the final fusion feature O: O=Q·Z in is the longitudinal concatenation of all sensor features, H T is the transpose of H, d k is the hidden state dimension, mutual information It is used to measure the correlation between different sensor features, and Q is the normalized confidence matrix.

5. The intelligent factory data acquisition platform according to claim 1 is characterized in that , the application service layer includes: Equipment health diagnosis module, which is used to build equipment health status model based on fusion features, and realize equipment fault diagnosis and predictive maintenance through anomaly detection and remaining life prediction; Process parameter optimization module, which is used to mine and optimize process parameters based on fusion features, and adjust equipment parameters in real time by combining expert knowledge base and reinforcement learning algorithm to improve production efficiency and product qualification rate; Energy consumption statistics and analysis module, which is used to count and analyze the energy consumption of equipment based on fusion features, automatically generate energy consumption reports, explore energy-saving potential, and formulate energy-saving strategies; The application service interface module is used to provide a standardized interface for equipment health diagnosis, process parameter optimization and energy consumption statistical analysis, realize data interaction and business collaboration with MES and ERP upper-level systems, and feed back the analysis results to the data perception layer and multi-protocol transmission layer through the data bus, forming a closed loop of data collection, analysis and control.

6. The intelligent factory data acquisition platform according to claim 2 is characterized in that The vibration sensor uses a PCB 356A16 three-axis accelerometer with a frequency response range of 0.5-5000Hz and a sensitivity of 50g; the noise sensor uses a GRAS46AE microphone with a frequency range of 3.15Hz-20kHz and a dynamic range of 17-146dB; the temperature and humidity sensor uses a Sensirion SHT35 digital temperature and humidity sensor with a temperature measurement range of -40-125℃, a humidity measurement range of 0-100%RH, and an accuracy of ±0.1℃ and ±1.5%RH; the visual sensor uses a Keyence CV-X150 series smart camera with a resolution of 640×480, 200 frames / second, and is equipped with a LumiTrax light source.

7. The intelligent factory data acquisition platform according to claim 3 is characterized in that The industrial Ethernet communication unit supports Modbus TCP, PROFINET, and EtherCAT industrial Ethernet protocols; the wireless communication unit supports LoRaWAN and NB-IoT low-power wide area network protocols; the synchronization accuracy of the time-sensitive network synchronization unit is better than 1μs, and the clock drift is less than 10ns / s.

8. The intelligent factory data acquisition platform according to claim 4 is characterized in that ,The training process of the self-correcting convolutional recurrent neural network includes: (1) Construct a loss function containing a self-correcting recurrent neural network: Where Y is softmax(O·W o ) to calculate the predicted output, Y * is the true label, CrossEntropy(Y,Y * )=-∑Y * ·log(Y) is the cross entropy loss function, Q is the confidence matrix, I is the identity matrix, ||·|| F is the Frobenius norm of the matrix, λ is the regularization coefficient; (2) Use the mini-batch stochastic gradient descent algorithm to adjust the neural network parameters W a , W o And LSTM parameters are iteratively optimized to minimize the loss function L: Where W refers to all parameters to be optimized, η is the learning rate, is the gradient of the loss function with respect to the parameters; (3) By setting an early stopping threshold and using the validation set loss as a monitoring indicator, iteration is stopped when the threshold is reached to prevent overfitting.

9. The smart factory data collection method according to any one of claims 1 to 8, characterized in that , including the following steps: Step 1: Collect equipment status data and environmental parameter data through vibration sensors, noise sensors, and temperature and humidity sensors, collect images of key equipment components through visual sensors, and filter, denoise, and normalize the collected data to obtain preprocessed data; Step 2: Packaging, encapsulating and compressing the pre-processed data through the multi-protocol data transmission unit, performing multi-protocol data transmission through the industrial Ethernet communication unit and the wireless communication unit, and performing time synchronization through the time-sensitive network synchronization unit, and outputting the transmitted data to the data fusion layer; Step 3: Extract local and global features of multi-source heterogeneous sensor data through self-correcting convolutional recurrent neural network, fuse multi-sensor data through attention mechanism and confidence matrix self-correction, and output the fused features to the application service layer; Step 4: Through the equipment health diagnosis, process parameter optimization and energy consumption statistical analysis application service modules, intelligent analysis and optimization decisions are made based on fusion features, and the analysis results are fed back to the data perception layer and multi-protocol transmission layer through the application service interface to achieve closed-loop control.

10. The intelligent factory data collection method according to claim 9 is characterized in that The training method of the self-correcting convolutional recurrent neural network in step 3 includes: A loss function containing a self-correcting recurrent neural network is constructed, and the neural network parameters are iteratively optimized using a small batch stochastic gradient descent algorithm. Overfitting is prevented by setting an early stopping threshold and using the validation set loss as a monitoring indicator, and the optimal model parameters after training are output.

Images