Intelligent manufacturing-oriented real-time data analysis method and system based on edge calculation

Abstract

The invention discloses an intelligent manufacturing-oriented real-time data analysis method and system based on edge calculation. The method comprises the following steps: collecting original time sequence data at a basic sampling frequency on an edge computing node; inputting the data into a dynamic data governance kernel, extracting high-order statistical features and lightweight convolutional neural network features in parallel by the kernel based on a sliding window, and generating a structured feature vector sequence through adaptive weighted fusion of a dynamic gating fusion unit; inputting the sequence into a time sequence prediction and anomaly detection joint model based on a bidirectional long-short-term memory network and an attention mechanism, and generating a real-time analysis result containing state prediction and classification; and the result and the key summary information are compressed and uploaded to a cloud server through an MQTT protocol in an event-driven mode. According to the method, deep coupling of data governance and analysis decision is realized, and while low delay and low bandwidth occupation are ensured, the accuracy of real-time analysis and the side cloud cooperation efficiency of the system are remarkably improved.

Description

Technical Field

[0001] This invention relates to the field of intelligent manufacturing technology, and in particular to a real-time data analysis method and system based on edge computing for intelligent manufacturing. Background Technology

[0002] With the deepening of the intelligent manufacturing strategy, modern manufacturing production lines are connecting to massive amounts of intelligent sensors, industrial robots, and programmable logic controllers at an unprecedented speed. These devices generate terabytes (TB) or even petabytes (PB) of real-time time-series data during operation, providing a valuable data foundation for production process monitoring, equipment health management, and product quality optimization. However, how to extract effective information that can guide production decisions in real time and accurately from this massive, high-dimensional, and noisy data has become the core bottleneck of the current intelligent transformation of the manufacturing industry.

[0003] Traditional manufacturing data analysis primarily relies on centralized cloud computing architectures. This architecture transmits raw data collected from all underlying devices via the network to a remote cloud center for unified storage and computation. However, this "data to the cloud" processing model faces three major technical hurdles in smart manufacturing scenarios: First, insufficient real-time performance. Industrial control loops require millisecond-level response times, such as suppressing cutting chatter or recovering from material blockages. However, data transmission, queuing, and cloud processing scheduling delays often exceed seconds, causing intelligent decision-making to lag behind changes in the physical world and failing to meet the needs of fault early warning and real-time control. Second, significant waste of network and computing resources. Industrial data often exhibits high temporal correlation and periodicity. During steady-state production, many data points do not carry new information. Uploading all "redundant" data indiscriminately not only consumes expensive industrial network bandwidth but also results in a huge waste of cloud computing and storage resources. Third, a disconnect between data governance and model adaptation. Existing technologies typically separate data preprocessing (such as cleaning and noise reduction) from advanced analysis (such as deep learning prediction). The former performs coarse filtering at the edge, while the latter invokes complex neural network models in the cloud or at the edge. This fragmented processing flow leads to the preprocessing stage often losing weak transient features used for early fault diagnosis in order to reduce the amount of data, while advanced analysis models have to spend a lot of computing resources to adapt to the low-quality data after "crude" processing, resulting in a decline in both the efficiency and accuracy of the entire data analysis system.

[0004] Although existing edge computing solutions attempt to push some computing down to industrial gateways closer to the data source, most of them only achieve simple data aggregation or lightweight filtering, failing to fundamentally solve the problem of dynamic adaptation between data governance and analysis models. In particular, when faced with non-stationary signals caused by operating condition drift and equipment aging during the manufacturing process, traditional fixed threshold preprocessing and static neural network models are inadequate. Summary of the Invention

[0005] To achieve the above objectives, this invention provides a real-time data analysis method based on edge computing for intelligent manufacturing, executed on an edge computing node close to the manufacturing equipment, wherein the edge computing node is communicatively connected to a cloud server, and the method includes the following steps:

[0006] Step 1: On the edge computing node, continuously collect raw time-series datasets from multiple industrial sensors at a preset basic sampling frequency;

[0007] Step 2: Input the original time series dataset into a pre-built dynamic data governance kernel for real-time processing to obtain a sequence of structured feature vectors after governance; wherein, the dynamic data governance kernel is a composite computational model based on a lightweight convolutional neural network and sliding window statistics, which is used to adaptively fuse statistical features and data-driven features through a dynamic gating fusion unit;

[0008] Step 3: Input the structured feature vector sequence into a pre-trained joint model for time-series prediction and anomaly detection to generate real-time analysis results R on equipment health status or process quality;

[0009] Step 4: Compress and upload the real-time analysis result R and the key summary information in the treated structured feature vector sequence to the cloud server in an event-driven manner via the MQTT protocol.

[0010] Preferably, step 2 specifically includes:

[0011] Step 2.1: Construct a sliding data window of length L. The sliding data window slides on the time axis with a step size s. For each window position, extract data segments from all sensors within the window to form a multivariate data matrix, where n is the number of sensors and s is the preset window time length.

[0012] Step 2.2: Input the multivariate data matrix X into a statistical feature extraction branch and a lightweight convolutional neural network feature extraction branch in parallel;

[0013] In the statistical feature extraction branch, for each sensor channel in the multivariate data matrix X, multiple higher-order statistics of each sensor channel in the one-dimensional time direction are calculated to form a statistical feature vector.

[0014] In the feature extraction branch of the lightweight convolutional neural network, the multivariate data matrix X is regarded as a two-dimensional image with n channels. The two-dimensional image is input into a lightweight convolutional neural network containing multiple depth separable convolutional layers, and a data-driven feature vector is output.

[0015] Step 2.3: Design a dynamic gating fusion unit. The dynamic gating fusion unit first concatenates the statistical feature vector and the data-driven feature vector into a joint feature vector. Then, the joint feature vector is input into a gating mechanism composed of a fully connected network. The gating mechanism outputs a weight scalar. Finally, the statistical feature vector and the data-driven feature vector are weighted and summed according to the weight scalar to obtain the fused structured feature vector v. As the sliding data window moves, all the fused structured feature vectors v are arranged in chronological order, thus forming the structured feature vector sequence.

[0016] Preferably, in step 2.2, the plurality of higher-order statistics calculated in the statistical feature extraction branch include at least the mean, standard deviation, skewness, and kurtosis; for the sensor channel that collects vibration data, the plurality of higher-order statistics also include the frequency domain energy centroid obtained based on the fast Fourier transform.

[0017] Preferably, in step 2.2, the lightweight convolutional neural network includes three depthwise separable convolutional layers: the first convolutional layer uses a convolutional kernel of size and a stride of 1 to capture short-term temporal patterns within each sensor channel; the second convolutional layer uses a depthwise convolutional kernel of size to fuse the instantaneous correlation features of different sensors at the same time; the third convolutional layer is a pointwise convolutional layer, using a convolutional kernel to perform nonlinear transformation and dimensionality reduction on the combined features extracted by the first two layers.

[0018] Preferably, in step 2.3, the gating mechanism consists of two fully connected layers, the two fully connected layers use the Sigmoid activation function as the last layer, the weight scalar is used to characterize the confidence of the statistical features under the current data window, and the weighted summation calculation formula is as follows.

[0019] Preferably, in step 3, the joint model for temporal prediction and anomaly detection is a deep learning model based on a bidirectional long short-term memory network and an attention mechanism. The joint model receives the structured feature vector sequence as input, where m is the current accumulated window size. First, the joint model captures the forward and backward dependencies of the structured feature vector sequence in the time dimension through a bidirectional long short-term memory network layer, generating a set of hidden state sequences. Then, a temporal attention mechanism layer performs a weighted summation of the hidden state sequences H, generating a context vector c. Finally, the context vector c is fed into a linear output layer for regression and a softmax output layer for classification. The linear output layer outputs the predicted value of the key performance indicator for the next time step. The softmax output layer outputs the classification probability of the current device in different state categories. Residual analysis is performed between the predicted value and the actual collected value, and a logical AND operation is performed between the residual and the classification probability to obtain the comprehensive real-time analysis result R.

[0020] Preferably, in step 4, the key summary information refers to the statistical and network features calculated in the dynamic data governance kernel and confirmed by the dynamic gating fusion unit to be representative, as well as the timestamp that triggers the state change; the event-driven method is as follows: when the state classification probability in the real-time analysis result R shows a state change, or when the residual between the predicted value and the true value exceeds a preset dynamic threshold, the edge computing node immediately triggers an upload task to package and upload several structured feature vectors v before and after the event and their corresponding timestamps.

[0021] Preferably, the preset basic sampling frequency setting satisfies at least the Nyquist sampling theorem to ensure that the highest frequency component of the key characteristic frequency of the target device can be collected without distortion; each data in the original time series dataset is accompanied by a high-precision timestamp, which is synchronized with the production line master clock by the edge computing node through the industrial network time protocol.

[0022] Preferably, the window time length of the sliding data window is set according to the dynamic response time constant of the monitored device to ensure that at least one complete device operation cycle or transient event can be captured within a window; the step size s is set according to the required real-time analysis resolution, and the step size s is less than the window length L to achieve partial overlap between data windows.

[0023] Accordingly, embodiments of the present invention also provide a real-time data analysis system based on edge computing for intelligent manufacturing, the system being used to execute the method described in any one of the embodiments of the present invention, the system comprising:

[0024] At least one edge computing node is located near the manufacturing equipment. The edge computing node includes a memory and one or more processors. The memory stores executable code. When the one or more processors execute the executable code, they are used to implement a real-time data analysis method based on edge computing for intelligent manufacturing, as described in any one of the embodiments of the present invention.

[0025] The cloud server is communicatively connected to the edge computing node. The cloud server is used to receive and store the real-time analysis result R and the key summary information uploaded by the edge computing node, and to perform global retraining on the dynamic data governance kernel and the joint model of time series prediction and anomaly detection.

[0026] The beneficial effects of this invention are:

[0027] 1. This invention addresses the disconnect between data preprocessing and advanced analysis in traditional solutions by constructing a "dynamic data governance kernel" that includes a statistical feature extraction branch, a lightweight convolutional neural network feature extraction branch, and a dynamic gating fusion unit. This kernel adaptively fuses high-order statistics with clear physical meaning and data-driven features highly sensitive to weak transient characteristics based on data quality, generating high-quality structured feature vectors. This design enables subsequent joint models for time-series prediction and anomaly detection to perform efficient inference in a cleaner, more information-rich feature space, achieving synergistic optimization of feature extraction and analytical decision-making.

[0028] 2. This invention deploys core computing tasks on edge computing nodes close to the data source, avoiding network latency caused by long-distance transmission of raw data. Simultaneously, a dynamic data governance kernel transforms high-dimensional raw data into low-dimensional structured feature vectors, significantly reducing the computational complexity of subsequent deep learning models. This enables the system to complete the entire process from data acquisition to analysis result output within milliseconds, meeting the stringent requirements of real-time control and early warning in industrial settings. Furthermore, high-quality structured feature input significantly improves the accuracy of time-series prediction and the precision of abnormal state classification, achieving a leap forward in analytical accuracy while maintaining real-time performance.

[0029] 3. This invention employs an event-driven upload mechanism, compressing and uploading key structured feature vectors and analysis results to the cloud only when a state transition occurs or the prediction residual exceeds limits; under steady-state conditions, simplified health indicators are uploaded only at a low frequency. This fundamentally solves the problems of high bandwidth consumption and high cloud storage and computing costs in traditional cloud architectures, focusing network and cloud resources on critical events. The high-quality feature data received in the cloud can be directly used for global retraining of the dynamic data governance kernel and the joint model for time series prediction. The optimized model can then be distributed to edge nodes, forming a virtuous cycle of "real-time edge decision-making and global cloud optimization," enabling the system to continuously learn and adaptively evolve. Attached Figure Description

[0030] To more clearly illustrate the technical solutions in this invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, those skilled in the art can obtain other drawings based on these drawings without creative effort.

[0031] Figure 1 This is a flowchart of the steps of the method of the present invention. Detailed Implementation

[0032] The present invention will now be described in detail with reference to the accompanying drawings and specific embodiments. It should also be noted that, to make the embodiments more comprehensive, the following embodiments are the best and preferred embodiments, and those skilled in the art can use other alternative methods to implement some well-known technologies; moreover, the accompanying drawings are only for more specific description of the embodiments and are not intended to specifically limit the present invention.

[0033] Please see Figure 1 This invention provides a real-time data analysis method and system based on edge computing for intelligent manufacturing. The method is executed on an edge computing node close to the manufacturing equipment, which communicates with a cloud server. The following describes the technical solution of this invention in detail using the spindle health monitoring of a precision CNC machine tool as an example. This CNC machine tool is a key piece of equipment in the manufacturing of aero-engine blades, with a spindle speed reaching up to 24,000 rpm. The health status of the spindle directly affects machining accuracy; therefore, real-time monitoring of vibration and temperature data is necessary to achieve early fault warning.

[0034] To achieve this goal, four industrial sensors are installed near the spindle: three piezoelectric vibration acceleration sensors, installed in the X, Y, and Z orthogonal directions of the spindle housing; and one platinum resistance temperature sensor, embedded near the spindle bearing housing. These sensors are connected to an edge computing node deployed next to the machine tool via the EtherCAT industrial Ethernet protocol. This edge computing node, based on an embedded industrial computer platform, has computing, storage, and communication capabilities and can exchange data with a remote cloud server via an industrial wireless network.

[0035] The method of the present invention includes the following four main steps:

[0036] Step 1: On the edge computing node, at a preset base sampling frequency Continuously collect raw time-series datasets from multiple industrial sensors. ;

[0037] Step 2: Transfer the original time series dataset The input is fed into a pre-built dynamic data governance kernel for real-time processing, resulting in a governed sequence of structured feature vectors. The dynamic data governance kernel is a composite computational model based on a lightweight convolutional neural network and sliding window statistics, which is used to adaptively fuse statistical features and data-driven features through a dynamic gating fusion unit.

[0038] Step 3: Convert the structured feature vector sequence Input into a pre-trained joint model for time series prediction and anomaly detection to generate real-time analysis results R on equipment health status or process quality;

[0039] Step 4: Combine the real-time analysis result R with the structured feature vector sequence after treatment. The key summary information is compressed and uploaded to the cloud server via the MQTT protocol in an event-driven manner.

[0040] The following section will elaborate on each step with specific examples.

[0041] Step 1: Acquisition of raw time series data

[0042] On edge computing nodes, a basic sampling frequency is pre-set. The frequency is set at 12.8 kHz. This frequency setting is based on the Nyquist sampling theorem: the highest frequency component of the target equipment's key characteristic frequencies needs to be acquired without distortion. Through analysis of the spindle's historical fault data, the highest characteristic frequency of bearing faults is approximately 6 kHz. Therefore, the sampling frequency needs to be at least 12 kHz. Choosing 12.8 kHz provides sufficient margin to ensure that all key fault frequency components can be captured.

[0043] A data acquisition service program running on the edge computing node continuously polls four sensors using a circular buffer. Whenever a sensor reading is acquired, the program immediately adds a high-precision timestamp to that data point. This timestamp is synchronized with the production line's master clock by the edge computing node via the IEEE 1588 precision time protocol, ensuring that the timestamp error between different sensors is less than 1 microsecond. This forms a continuous, multi-channel raw time-series dataset. The dataset consists of four time series: X-axis vibration. Y-axis vibration Z-axis vibration ,temperature Each time series corresponds to a physical quantity value at time t, and the unit of vibration value is . The unit of temperature value is .

[0044] Step 2: Real-time processing of the dynamic data governance kernel

[0045] A pre-built dynamic data governance kernel is loaded into the memory of the edge computing node. This kernel processes the raw data stream using a sliding window approach, specifically including the following sub-steps.

[0046] Step 2.1: Constructing the sliding data window

[0047] Set window duration The duration is 0.1 seconds. This duration is determined based on the dynamic response time constant of the monitored device (spindle): the spindle's maximum speed is 24,000 rpm, or 400 rpm. 0.1 seconds contains 40 complete rotational cycles, sufficient to capture fault characteristics related to the rotational frequency, while also capturing most transient events. This is based on the base sampling frequency. Hz, number of sampling points within the window One point.

[0048] The sliding step size s is set to 64 sampling points. The step size is smaller than the window length (1280 points), resulting in an overlap of 1216 points between adjacent windows, achieving an overlap rate of 95%. Although this high overlap sliding method increases the computational load, it greatly improves the temporal resolution, ensuring that any tiny transient event can be captured by multiple windows, avoiding missed detections.

[0049] For each window position, the program extracts four sensor readings from the circular buffer in the past. All data points within a given time period are arranged in order of sensor channels to form a multivariate data matrix. The matrix has rows corresponding to sensor channels (4 rows) and columns corresponding to sampling time points (1280 columns). Each column represents the synchronous readings of 4 sensors at a sampling time.

[0050] Step 2.2: Parallel dual-branch feature extraction

[0051] Matrix X is fed into two independent computational branches simultaneously: a statistical feature extraction branch and a lightweight convolutional neural network feature extraction branch.

[0052] In the statistical feature extraction branch, for each row vector of matrix X (i.e., each sensor channel), multiple higher-order statistics in the one-dimensional time direction are calculated. Taking the X-axis vibration channel as an example, its 1280 data points are denoted as... The calculation process is as follows:

[0053] mean It reflects the average energy level of vibration.

[0054] Standard deviation This reflects the amplitude of vibration.

[0055] Skewness It reflects the symmetry of the vibration signal distribution. A deviation of 0 from the skewness value may indicate scratching or impact.

[0056] Kudo It is extremely sensitive to impact pulses and is an important indicator of early bearing failure.

[0057] For vibration data (X, Y, Z channels), the frequency domain energy centroid is also calculated. The specific procedure is as follows: Perform a Fast Fourier Transform on 1280 data points to obtain the spectrum, and then calculate the energy centroid frequency in the 0-6kHz frequency range.

[0058]

[0059] Where P(f) is the power spectral density at frequency f. This index can reflect whether the frequency band where vibration energy is concentrated has drifted, for example, from the normal rotation frequency to the bearing failure frequency.

[0060] The above calculations are repeated for the Y and Z vibration channels and the temperature channel (the frequency domain energy centroid is not calculated for the temperature channel; only the mean, standard deviation, skewness, and kurtosis are calculated). Ultimately, five statistics are obtained for each vibration channel, and four statistics are obtained for the temperature channel; therefore, the statistical eigenvectors are... The dimension is dimension.

[0061] In the feature extraction branch of the lightweight convolutional neural network, matrix X is reshaped into an "image" with a height of 1, a width of 1280, and 4 channels. This "image" is then fed into a lightweight convolutional neural network designed for edge computing, which contains three depthwise separable convolutional layers:

[0062] The first layer is a depthwise convolutional layer, using four layers of size [missing information]. The convolutional kernels (one kernel per input channel) have a stride of 1 and are padded with "SAME". This layer slides independently over the time series of each sensor channel, capturing short-term local waveform patterns, such as rising edges, falling edges, and peaks, with a length of 3 samples. The output remains 4 channels with a width of 1280.

[0063] The second layer is a depthwise convolutional layer, using a matrix of size [missing information]. A deep convolutional kernel with a stride of 1 and "VALID" padding is used. This kernel slides between four channels, fusing instantaneous correlation features between different sensors at the same time, such as the co-variation relationship between X, Y, and Z vibrations and temperature. The number of output channels becomes 1, and the width becomes 1278.

[0064] The third layer is a pointwise convolutional layer, using 32 integers of size 1. The convolutional kernel is used. This layer performs linear combination and non-linear transformation (followed by ReLU activation) on the combined feature map output from the previous layer, increasing the number of channels of the feature map from 1 to 32 while maintaining the width of 1278. Then, global average pooling is performed in the width direction, finally obtaining a 32-dimensional data-driven feature vector. This vector encapsulates the most significant time-series and cross-channel patterns mined from data within the entire window.

[0065] Step 2.3: Dynamic Gating Fusion

[0066] To adaptively combine statistical features with explicit physical meaning and sensitive data-driven features, a dynamic gating fusion unit is designed. This unit first combines a 19-dimensional statistical feature vector... and 32-dimensional data-driven feature vectors Concatenate them into a 51-dimensional joint feature vector. .

[0067] Then, The data is fed into a gating mechanism consisting of two fully connected layers. The first fully connected layer contains 16 neurons with the ReLU activation function, used to learn higher-order nonlinear relationships between features. The second fully connected layer has only one neuron with the Sigmoid activation function, outputting a weight scalar between 0 and 1. This weight scalar Used to characterize the confidence or dominance of statistical features within the current data window. When the signal is stable and the signal-to-noise ratio is high, gating networks tend to assign higher weights to statistical features. (Approaching 1); when unknown transient disturbances appear in the signal and statistical characteristics are difficult to describe, the gating network assigns higher weights to data-driven features. (Close to 0).

[0068] Finally, according to the formula Weighted summation yields the final fused structured feature vector v, a 51-dimensional vector. As the sliding window moves forward every 64 sampling points, a new structured feature vector is generated. (j represents the j-th window). These vectors are stored in chronological order, thus forming a structured feature vector sequence. Thus, the original high-dimensional, high-noise time series data has been transformed into a high-quality, low-dimensional structured feature stream.

[0069] Step 3: Inference of the joint model for time series prediction and anomaly detection

[0070] The accumulated structured feature vector sequence The input is fed into a pre-trained joint model for temporal prediction and anomaly detection. This model is a deep learning model based on a bidirectional long short-term memory network and an attention mechanism, with an input length set to 10, meaning that the 10 most recent window feature vectors are used each time. To reason.

[0071] The internal processing flow of the model is as follows:

[0072] Bidirectional Long Short-Term Memory (LSTM) network layer: This layer contains 64 hidden units (32 forward and 32 backward). It processes the 10 input 51-dimensional vectors sequentially and outputs a sequence of hidden states. Each of them It is a 64-dimensional vector that incorporates information from the current moment and the surrounding context.

[0073] Temporal attention mechanism layer: The attention mechanism calculates each hidden state. Weights for the current prediction task The weights are learned through a small fully connected network and applied to all... Perform Softmax normalization. Then, sum all hidden states according to their weights to generate a context vector. The vector c focuses on the most important information from the past 10 windows.

[0074] Parallel output layer: The context vector c is fed into two different output layers simultaneously.

[0075] Regression output layer: Consists of a linear layer without an activation function, outputting a scalar value. This refers to the predicted value of the spindle temperature at the next time step (the (j+1)th window).

[0076] Classification output layer: Consists of a linear layer with a softmax activation function, outputting a 3-dimensional probability vector. , representing the probability that the current main axis is in one of the three states: "normal", "attention", and "abnormal".

[0077] Obtain the predicted value Then, the system waits until the actual temperature data for the (j+1)th window is available. After data collection is complete, the residuals are calculated. The predicted residuals and classification probabilities are logically ANDed to obtain a comprehensive real-time analysis result R. Specifically, if the residuals... The temperature exceeds the preset dynamic threshold of 5°C three times consecutively (this threshold is derived from statistical analysis of temperature fluctuations under historical normal operating conditions), and the probability of the "abnormal" category in the classification probability is also high. If the probability is greater than 85%, R is defined as "spindle overheating, failure is predicted, immediate shutdown and inspection recommended"; if only the classification probability is "caution" or the residual occasionally exceeds the limit, a warning or normal state mark of the corresponding level is given.

[0078] Step 4: Event-driven upload for edge-cloud collaboration

[0079] To maximize the use of network bandwidth and cloud computing resources, an event-driven approach is adopted to upload data to the cloud server. The uploaded content includes real-time analysis results (R) and key summary information. Key summary information refers to statistical features (such as mean and kurtosis) and network features (such as the principal components in the data-driven feature vector) that are calculated in the dynamic data governance kernel and confirmed as representative by the dynamic gating fusion unit, as well as the timestamps that triggered the state changes.

[0080] The upload trigger conditions are as follows:

[0081] Event-triggered upload: An emergency upload task is triggered when any of the following conditions are met:

[0082] The probability of state classification changes, for example, from the "normal" state ( ) becomes "attention" state ( ), or change from "attention" to "abnormal" ( ).

[0083] Temperature prediction residual The temperature exceeded 5°C for three consecutive times.

[0084] Once triggered, the edge node immediately packages the 10 structured feature vectors v before and after the event (20 in total), along with their respective timestamps, and the final analysis result R. The data packet is compressed using the MQTT protocol and immediately uploaded to the cloud server.

[0085] Steady-state heartbeat upload: In a "normal" state where the device is running smoothly and no events are triggered, to maintain cloud awareness of long-term device trends, edge nodes upload a set of highly concise health metrics at a low heartbeat frequency (e.g., once per minute). These metrics include: the average temperature over the past minute, the temperature standard deviation, the effective value of X-axis vibration, and the average weight output by the gating fusion unit. This data was also packaged and uploaded.

[0086] After receiving the data, the cloud server stores it in a historical database. This data (especially the feature vectors containing rich contextual information uploaded when the event is triggered) is valuable material for subsequent model retraining and optimization. The cloud can use this high-quality data to periodically optimize the dynamic data governance kernel in step 2 and the joint model in step 3, and then distribute the updated model parameters to edge nodes, forming a continuously evolving intelligent closed loop.

[0087] System Composition

[0088] A real-time data analysis system based on edge computing for intelligent manufacturing, which implements the above methods, includes:

[0089] At least one edge computing node is located near the manufacturing equipment. The edge computing node includes a memory and one or more processors. The memory stores executable code. When the one or more processors execute the executable code, they are used to implement the above-mentioned real-time data analysis method based on edge computing for intelligent manufacturing.

[0090] The cloud server is communicatively connected to the edge computing node. The cloud server is used to receive and store the real-time analysis result R and the key summary information uploaded by the edge computing node, and to perform global retraining on the dynamic data governance kernel and the joint model of time series prediction and anomaly detection.

[0091] Example;

[0092] This invention provides a real-time data analysis method and system based on edge computing for intelligent manufacturing, which can be widely applied to online monitoring and intelligent operation and maintenance of various manufacturing equipment. The following section uses the spindle health status monitoring of a precision CNC machine tool as a specific application scenario to elaborate on the technical solution of this invention.

[0093] This precision CNC machine tool is a key piece of equipment in a certain aero-engine blade manufacturing workshop. Its spindle speed can reach up to 24,000 rpm, and the health of the spindle directly determines the machining accuracy and surface quality of the blades. During high-speed rotation, the spindle may experience faults such as bearing wear, imbalance, and loosening, which will manifest through vibration and temperature signals. Therefore, real-time monitoring of the spindle status and early fault warning are crucial for ensuring production safety and product quality.

[0094] In this embodiment, four industrial sensors are installed near the spindle, including three piezoelectric vibration acceleration sensors, installed in the X, Y, and Z orthogonal directions of the spindle housing respectively; and one platinum resistance temperature sensor, embedded near the spindle bearing housing. These sensors are connected to an edge computing node deployed next to the machine tool via the EtherCAT industrial Ethernet protocol. This edge computing node is based on an embedded industrial computer platform, possesses certain computing power and storage space, and supports communication with a remote cloud server via an industrial wireless network.

[0095] Step 1: Acquisition of raw time series data

[0096] On edge computing nodes, a basic sampling frequency is pre-set. The frequency is set at 12.8kHz. This frequency setting is based on the fact that, through analysis of historical fault data, the fault characteristic frequency of the spindle bearing can reach up to about 6kHz. According to the Nyquist sampling theorem, the sampling frequency must be at least twice the highest frequency of the signal. The setting of 12.8kHz provides sufficient margin to ensure that all key fault frequency components can be captured without distortion.

[0097] A data acquisition service program running on the edge computing node continuously polls the four sensors using a circular buffer. Whenever a sensor reading is acquired, the program immediately adds a high-precision timestamp to that data point. This timestamp is synchronized between the edge computing node and the production line's master clock via the IEEE 1588 precision time protocol, ensuring that the timestamp error from different sensors is less than 1 microsecond. This results in a continuous, multi-channel raw time-series dataset. The dataset consists of four time series, denoted as follows: (X-axis vibration) (Y-axis vibration) (Z-axis vibration) (Temperature), each time series corresponds to a specific physical quantity value at time t, and the unit of vibration value is... The unit of temperature value is .

[0098] Step 2: Real-time processing of the dynamic data governance kernel

[0099] A pre-built dynamic data governance kernel is loaded into the memory of the edge computing node. This kernel processes the raw data stream using a sliding window approach.

[0100] Step 2.1: Constructing the sliding data window

[0101] Set window duration The duration is 0.1 seconds. This duration is chosen by considering both the spindle rotation cycle and the duration of the fault transient event. When the spindle operates at its highest speed, rotating at 400 revolutions per second, 0.1 seconds contains 40 complete rotation cycles, sufficient to capture fault characteristics related to the rotation frequency. The number of sampling points within the window is determined based on the base sampling frequency. The window slides forward 64 points at a time, with an overlap of 1280-64=1216 points between adjacent windows, resulting in an overlap rate of up to 95%. This high overlap sliding method, while increasing computational load, significantly improves temporal resolution, ensuring that any tiny transient event can be captured by multiple windows, thus avoiding missed detections.

[0102] For each window position, the program extracts four sensor readings from the circular buffer in the past. All data points within the time period are arranged in a 4-row, 1280-column multivariate data matrix X according to the sensor channel order. Each row of the matrix corresponds to the time series of one sensor, and each column represents the synchronous readings of the four sensors at a sampling time point.

[0103] Step 2.2: Parallel dual-branch feature extraction

[0104] Matrix X is fed into two independent computational branches simultaneously: a statistical feature extraction branch and a lightweight convolutional neural network feature extraction branch.

[0105] In the statistical feature extraction branch, for each row vector of matrix X (i.e., each sensor channel), five higher-order statistics are calculated along the one-dimensional time direction. Taking the X-axis vibration channel as an example, its 1280 data points are denoted as... The calculation process is as follows:

[0106] mean It reflects the average energy level of vibration.

[0107] Standard deviation This reflects the amplitude of vibration.

[0108] Skewness It reflects the symmetry of the vibration signal distribution. A deviation of 0 from the skewness value may indicate scratching or impact.

[0109] Kudo It is extremely sensitive to impact pulses and is an important indicator of early bearing failure.

[0110] For vibration data (X, Y, Z channels), the frequency domain energy centroid is also calculated. The specific method involves performing a Fast Fourier Transform on 1280 data points to obtain the spectrum, and then calculating the energy centroid frequency in the 0-6kHz frequency range. , where P(f) is the power spectral density at frequency f. This index can reflect whether the frequency band where vibration energy is concentrated has drifted, such as from the normal rotation frequency to the bearing failure frequency.

[0111] The above calculations were repeated for the other three channels, resulting in 5 statistics for each channel (5 for the vibration channel and 4 for the temperature channel, since the frequency domain energy centroid is not calculated for temperature). Therefore, the statistical eigenvectors... The total dimension is dimension.

[0112] In the feature extraction branch of the lightweight convolutional neural network, matrix X is reshaped into an "image" with a height of 1, a width of 1280, and 4 channels. This "image" is then fed into a lightweight convolutional neural network designed specifically for edge computing. This network contains three depthwise separable convolutional layers, designed to extract deep features with a minimal number of parameters.

[0113] The first layer is a depthwise convolutional layer, using four layers of size [missing information]. The convolutional kernels (one kernel per input channel) have a stride of 1 and are padded with "SAME" to maintain a constant width. This layer independently slides across the time series of each sensor channel, capturing short-term local waveform patterns (e.g., rising edges, falling edges, peaks) with a length of 3 samples. The output remains 4 channels, but each position within each channel contains local temporal information.

[0114] The second layer is a depthwise convolutional layer, using a matrix of size [missing information]. The depthwise convolution kernel (cross-channel depthwise convolution) has a stride of 1 and padding of "VALID". This kernel slides between the four channels to fuse instantaneous correlation features between different sensors at the same time, such as the co-variation relationship between X, Y, and Z vibrations and temperature. The output channel number of this layer becomes 1, and the size is [size missing]. .

[0115] The third layer is a pointwise convolutional layer, using 32 integers of size [missing information]. The convolutional kernel is used. This layer performs linear combination and non-linear transformation (followed by ReLU activation) on the combined feature map output from the previous layer, increasing the number of channels of the feature map from 1 to 32 while keeping the width unchanged at 1278. Then, global average pooling is performed in the width direction (1278 points), finally obtaining a 32-dimensional data-driven feature vector. This vector encapsulates the most significant time-series and cross-channel patterns mined from data within the entire window.

[0116] Step 2.3: Dynamic Gating Fusion

[0117] To adaptively combine statistical features with explicit physical meaning and sensitive data-driven features, a dynamic gating fusion unit was designed. This unit first combines 19-dimensional statistical features with sensitive data-driven features. And 32-dimensional Concatenate them into a 51-dimensional joint feature vector. .

[0118] Subsequently, The data is fed into a gating mechanism consisting of two fully connected layers. The first fully connected layer contains 16 neurons with the ReLU activation function, used to learn higher-order nonlinear relationships between features. The second fully connected layer has only one neuron with the Sigmoid activation function, outputting a weight scalar between 0 and 1. .this This can be understood as the reliability or dominance of statistical features in the final judgment within the current data window. When the signal is stable and the signal-to-noise ratio is high, gating networks tend to give higher weights to statistical features. (Approaching 1); when unknown transient disturbances appear in the signal and statistical characteristics are difficult to describe, the gating network assigns higher weights to data-driven features. (Close to 0).

[0119] Finally, according to the formula Weighted summation yields the final fused structured feature vector v, a 51-dimensional vector. A new feature vector is generated every 64 sampling points as the sliding window moves forward. (j represents the j-th window). These vectors are stored in chronological order, thus forming a structured feature vector sequence. Thus, the original high-dimensional, high-noise time series data has been transformed into a high-quality, low-dimensional structured feature stream.

[0120] Step 3: Inference of the joint model for time series prediction and anomaly detection

[0121] The accumulated structured feature vector sequence The input is fed into a pre-trained joint model for temporal prediction and anomaly detection. This model is a deep learning model based on a bidirectional long short-term memory network and an attention mechanism, with an input length set to 10, meaning that the 10 most recent window feature vectors are used each time. To reason.

[0122] The internal processing flow of the model is as follows:

[0123] 1. Bidirectional Long Short-Term Memory (LSTM) network layer: This layer contains 64 hidden units (32 forward and 32 backward). It processes the 10 input 51-dimensional vectors sequentially and outputs a sequence of hidden states. Each h is a 64-dimensional vector that incorporates information from the current moment and the surrounding context.

[0124] 2. Temporal Attention Mechanism Layer: The attention mechanism calculates each hidden state. Weights for the current prediction task The weights are learned through a small fully connected network and applied to all... Perform Softmax normalization. Then, sum all hidden states according to their weights to generate a context vector. The vector c focuses on the most important information from the past 10 windows.

[0125] 3. Parallel output layer: The context vector c is fed into two different output layers simultaneously.

[0126] Regression output layer: Consists of a linear layer without an activation function, outputting a scalar value. This refers to the predicted value of the spindle temperature at the next time step (the (j+1)th window).

[0127] Classification output layer: Consists of a linear layer with a softmax activation function, outputting a 3-dimensional probability vector. , representing the probability that the current main axis is in one of the three states: "normal", "attention", and "abnormal".

[0128] Obtain the predicted value Then, the system waits until the actual temperature data for the (j+1)th window is available. After data collection is complete, the residuals are calculated. .like If the temperature exceeds the preset dynamic threshold of 5°C three times consecutively (this threshold is derived from statistical analysis of temperature fluctuations under historical normal operating conditions), it is considered a predicted anomaly. Simultaneously, the classification probability... The category with the highest probability is taken as the current state. The final comprehensive real-time analysis result R is a combination of the two. For example, if the probability of the state being "abnormal" exceeds 85% and the prediction residual continuously exceeds the limit, then R is defined as "spindle overheating, prediction will fail, immediate shutdown and inspection recommended".

[0129] Step 4: Event-driven upload for edge-cloud collaboration

[0130] To maximize the use of network bandwidth and cloud computing resources, the upload mechanism adopts an event-driven approach. Edge computing nodes continuously monitor the output R from step 3 and intermediate data during the computation process.

[0131] Event-triggered upload: An emergency upload task is triggered when any of the following conditions are met:

[0132] The probability of state classification changes, for example, from the "normal" state ( ) becomes "attention" state ( ), or change from "attention" to "abnormal" ( ).

[0133] Temperature prediction residual The temperature exceeded 5°C for three consecutive times.

[0134] Once triggered, the edge node immediately packages the 10 structured feature vectors v before and after the event (20 in total), along with their respective timestamps, and the final analysis result R. The data packet is compressed using the MQTT protocol and immediately uploaded to the cloud server.

[0135] Steady-state heartbeat upload: In a "normal" state where the device is running smoothly and no events are triggered, to maintain cloud awareness of long-term device trends, edge nodes upload a set of highly concise health metrics at a low heartbeat frequency (e.g., once per minute). These metrics include: the average temperature over the past minute, the temperature standard deviation, the effective value of X-axis vibration, and the average weight output by the gating fusion unit. This data was also packaged and uploaded.

[0136] After receiving the data, the cloud server stores it in a historical database. This data (especially the feature vectors containing rich contextual information uploaded when the event is triggered) is valuable material for subsequent model retraining and optimization. The cloud can use this high-quality data to periodically optimize the dynamic data governance kernel in step 2 and the joint model in step 3, and then distribute the updated model parameters to edge nodes, forming a continuously evolving intelligent closed loop.

[0137] To highlight the technical effects of the present invention, three comparative examples were set up for comparative analysis. All comparative examples used the same hardware sensor configuration and the same experimental conditions as this embodiment (i.e., subjecting the spindle to three stages: normal operation, slight wear, and serious failure).

[0138] Comparative Example 1 (Pure Cloud-Based Centralized): The edge node acts only as a data forwarding gateway, uploading all 12.8kHz raw data collected by the four sensors to the cloud server in real time via the industrial network. The cloud server is equipped with a high-performance computer that first performs unified filtering and normalization preprocessing on the raw data, and then inputs it into a large deep residual network for fault classification and prediction.

[0139] Comparative Example 2 (Simple Edge Filtering + Cloud Analysis): Edge nodes perform simple threshold filtering on the raw data; for example, only data segments with valid vibration values ​​exceeding a preset threshold are uploaded, while the rest are discarded. The uploaded data undergoes the same analysis and processing in the cloud as in Comparative Example 1.

[0140] Comparative Example 3 (Traditional Statistical Features for Edges + Classifier): The edge nodes undergo the same calculations as the statistical feature extraction branches in steps 2.1 and 2.2 of this embodiment, but lightweight convolutional neural network feature extraction and dynamic gating fusion are not performed. The 19-dimensional statistical feature vector is directly used as input, and a traditional random forest classifier is run on the edge nodes to determine the state. The determination results and statistical features are then uploaded.

[0141] The main metrics compared include: average end-to-end response delay (time from sensor data generation to analysis result output), average daily data upload volume, fault detection accuracy, and false alarm rate. The experimental results are shown in Table 1.

[0142] Table 1 Comparison of the effects of different technical solutions

[0143] Technical solution Average response latency (milliseconds) Average daily data upload volume (megabytes) Fault detection accuracy (%) False alarm rate (%) Comparative Example 1 (Pure Cloud-Based Centralized System) 850 Approximately 42,000 94.5 5.2 Comparative Example 2 (Simple Edge Filtering) 820 Approximately 8,500 82.3 18.7 Comparative Example 3 (Marginal Traditional Statistics) 120 Approximately 0.8 88.1 11.5 Embodiments of the present invention 95 Approximately 1.2 97.8 2.4

[0144] As can be seen from Table 1:

[0145] Comparative Example 1 suffers from a massive data volume due to uploading all raw data, and the long response time (850 milliseconds) caused by network transmission and cloud queuing latency fails to meet real-time control requirements. Although the cloud model is powerful and has acceptable accuracy, its high latency is a fatal flaw.

[0146] Comparative Example 2 significantly reduced the amount of data through simple filtering, but also lost a large amount of weak but critical early fault information, resulting in a sharp drop in accuracy (82.3%) and a surge in false alarm rate (18.7%), making it the worst performing example.

[0147] Comparative Example 3 extracts statistical features at the edges, which significantly reduces latency (120 milliseconds) and data volume. However, relying solely on fixed statistical features cannot fully capture nonlinear fault modes under complex operating conditions, and the accuracy (88.1%) and false alarm rate (11.5%) are still unsatisfactory.

[0148] This invention achieves the highest detection accuracy (97.8%) and the lowest false alarm rate (2.4%) while maintaining extremely low latency (95 milliseconds, even lower than Comparative Example 3 due to higher quality subsequent model input and faster computation) and extremely low data volume (1.2 megabytes, only slightly higher than Comparative Example 3). This fully demonstrates that this invention achieves deep adaptive fusion of statistical features and data-driven features through a "dynamic data governance kernel," and completes high-quality inference at the edge, ultimately achieving optimal performance in terms of real-time performance, accuracy, and resource consumption.

[0149] In summary, this invention provides a real-time data analysis method and system based on edge computing for intelligent manufacturing. By constructing a dynamic data governance kernel on edge computing nodes, it achieves intelligent cleaning and feature enhancement of raw data. It then utilizes a joint model of time-series prediction and anomaly detection for accurate analysis, and finally achieves efficient edge-cloud collaboration through an event-driven mechanism. This solution effectively solves the challenge of balancing latency, bandwidth, and accuracy in real-time data analysis in intelligent manufacturing, and has extremely high industrial application value.

[0150] This invention encompasses any substitutions, modifications, equivalent methods, and solutions made within the spirit and scope of this invention. To provide the public with a thorough understanding of this invention, specific details are described in detail in the following preferred embodiments; however, those skilled in the art will fully understand the invention even without these details. Furthermore, to avoid unnecessary misunderstanding of the essence of this invention, well-known methods, processes, procedures, components, and circuits are not described in detail.

[0151] The above description is only a preferred embodiment of the present invention. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the principle of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. A real-time data analysis method based on edge computing for intelligent manufacturing, characterized in that, The method, executed on an edge computing node located close to manufacturing equipment and communicatively connected to a cloud server, includes the following steps: Step 1: On the edge computing node, at a preset base sampling frequency Continuously collect raw time-series datasets from multiple industrial sensors. ; Step 2: Transfer the original time series dataset The input is fed into a pre-built dynamic data governance kernel for real-time processing, resulting in a governed sequence of structured feature vectors. The dynamic data governance kernel is a composite computational model based on a lightweight convolutional neural network and sliding window statistics, which is used to adaptively fuse statistical features and data-driven features through a dynamic gating fusion unit. Step 3: Convert the structured feature vector sequence Input into a pre-trained joint model for time series prediction and anomaly detection to generate real-time analysis results R on equipment health status or process quality; Step 4: Combine the real-time analysis result R with the structured feature vector sequence after treatment. The key summary information is compressed and uploaded to the cloud server via the MQTT protocol in an event-driven manner.

2. The real-time data analysis method based on edge computing for intelligent manufacturing according to claim 1, characterized in that, Step 2 specifically includes: Step 2.1: Construct a sliding data window of length L, which slides along the time axis in steps s. For each window position, extract data segments from all sensors within the window to form a multivariate data matrix. Where n is the number of sensors. , This is the preset window duration. Step 2.2: Input the multivariate data matrix X into a statistical feature extraction branch and a lightweight convolutional neural network feature extraction branch in parallel; In the statistical feature extraction branch, for each sensor channel in the multivariate data matrix X, multiple higher-order statistics of each sensor channel in the one-dimensional time direction are calculated to form a statistical feature vector. ; In the feature extraction branch of the lightweight convolutional neural network, the multivariate data matrix X is treated as a two-dimensional image with n channels. The two-dimensional image is input into a lightweight convolutional neural network containing multiple depthwise separable convolutional layers, and a data-driven feature vector is output. ; Step 2.3: Design a dynamic gating fusion unit, which first processes the statistical feature vector... and the data-driven feature vector Concatenate them into a joint feature vector Then the joint feature vector The input is fed into a gating mechanism consisting of a fully connected network, and the gating mechanism outputs a weight scalar. Finally, based on the aforementioned weight scalar For the statistical feature vector and the data-driven feature vector A weighted sum is performed to obtain the fused structured feature vector v. As the sliding data window moves, all the fused structured feature vectors v are arranged in chronological order, thus forming the structured feature vector sequence. .

3. The real-time data analysis method based on edge computing for intelligent manufacturing according to claim 2, characterized in that, In step 2.2, the plurality of higher-order statistics calculated in the statistical feature extraction branch include at least the mean. Standard deviation skewness kurtosis For sensor channels that acquire vibration data, the multiple higher-order statistics also include the frequency domain energy centroid obtained based on the fast Fourier transform. .

4. A real-time data analysis method based on edge computing for intelligent manufacturing according to claim 2, characterized in that, In step 2.2, the lightweight convolutional neural network comprises three depthwise separable convolutional layers: the first convolutional layer uses a size of... The first convolutional kernel, with a stride of 1, is used to capture short-term temporal patterns within each sensor channel; the second convolutional layer uses a kernel of size 1. The deep convolutional kernel is used to fuse the instantaneous correlation features of different sensors at the same time. The third convolutional layer is a pointwise convolutional layer, using... The convolutional kernel is used to perform nonlinear transformation and dimensionality reduction on the combined features extracted from the first two layers.

5. A real-time data analysis method based on edge computing for intelligent manufacturing according to claim 2, characterized in that, In step 2.3, the gating mechanism consists of two fully connected layers, with the sigmoid activation function used as the last layer. The weight scalar... The weighted summation is used to characterize the confidence level of statistical features within the current data window. .

6. A real-time data analysis method based on edge computing for intelligent manufacturing according to claim 1, characterized in that, In step 3, the joint model for temporal prediction and anomaly detection is a deep learning model based on a bidirectional long short-term memory network and an attention mechanism; the joint model for temporal prediction and anomaly detection receives the structured feature vector sequence. As input, where m is the current accumulated number of windows; the joint model of temporal prediction and anomaly detection first captures the structured feature vector sequence through a bidirectional long short-term memory network layer. The forward and backward dependencies in the time dimension generate a set of hidden state sequences. Next, a temporal attention mechanism layer performs a weighted summation on the hidden state sequence H to generate a context vector c. Finally, the context vector c is fed into a linear output layer for regression and a softmax output layer for classification. The linear output layer outputs the predicted values ​​of the key performance indicators for the next time step. The Softmax output layer outputs the classification probability of the current device being in different state categories. ; the predicted value Perform residual analysis with the actual collected values, and then correlate the residuals with the classification probabilities. A logical AND operation is performed to obtain the comprehensive real-time analysis result R.

7. A real-time data analysis method based on edge computing for intelligent manufacturing according to claim 1, characterized in that, In step 4, the key summary information refers to the statistical and network features that are calculated in the dynamic data governance kernel and confirmed by the dynamic gating fusion unit to be representative, as well as the timestamp that triggers the state change; the event-driven approach is as follows: when the state classification probability in the real-time analysis result R shows a state change, or when the residual between the predicted value and the true value exceeds a preset dynamic threshold, the edge computing node immediately triggers an upload task to package and upload several structured feature vectors v before and after the event and their corresponding timestamps.

8. A real-time data analysis method based on edge computing for intelligent manufacturing according to claim 1, characterized in that, The preset basic sampling frequency The settings must at least satisfy the Nyquist sampling theorem to ensure that the highest frequency components of the key feature frequencies of the target device can be acquired without distortion; all acquired original time series datasets Each piece of data in the dataset is accompanied by a high-precision timestamp, which is synchronized with the production line master clock by the edge computing node via an industrial network time protocol.

9. A real-time data analysis method based on edge computing for intelligent manufacturing according to claim 2, characterized in that, The window time length of the sliding data window The step size s is set according to the dynamic response time constant of the monitored device to ensure that at least one complete device operation cycle or transient event can be captured within a window; the step size s is set according to the required real-time analysis resolution, and the step size s is less than the window length L to achieve partial overlap between data windows.

10. A real-time data analysis system based on edge computing for intelligent manufacturing, characterized in that, The system is used to perform the method according to any one of claims 1 to 9, the system comprising: At least one edge computing node located near a manufacturing device, the edge computing node comprising a memory and one or more processors, the memory storing executable code, the one or more processors executing the executable code to implement a real-time data analysis method based on edge computing for intelligent manufacturing as described in any one of claims 1 to 9; The cloud server is also connected to the edge computing node. The cloud server is used to receive and store the real-time analysis result R and the key summary information uploaded by the edge computing node, and to perform global retraining on the dynamic data governance kernel and the joint model of time series prediction and anomaly detection.

Images