Data processing method and system for monitoring data based on big data analysis

Abstract

The application discloses a kind of based on big data analysis's monitoring data data processing method and system, the present application relates to the technical field of big data analysis, obtains the monitoring data of industrial equipment monitoring end collection, based on edge computing node carries out data analysis to monitoring data, in the analysis process, the corresponding local space-time feature is extracted, and corresponding standard monitoring data flow is generated in combination with the space-time alignment mechanism of cross mode, and the standard monitoring data flow generated eliminates space-time asynchrony and scale difference, provides uniform, high-fidelity data basis for subsequent big data analysis, significantly improves the real-time performance and accuracy of data processing.Meanwhile, using space-time joint encoder to extract features, and combining time decay factor to construct a continuously evolving knowledge graph, can accurately present the multi-dimensional state of industrial equipment in action details and trajectory trends, and the introduction of time decay factor synchronously eliminates historical noise interference.

Description

Technical Field

[0001] This invention relates to the technical field of big data analysis, and in particular to a data processing method and system for monitoring data based on big data analysis. Background Technology

[0002] With the rapid development of the Industrial Internet of Things (IIoT) and intelligent manufacturing technologies, real-time monitoring and intelligent analysis of the operating status of industrial equipment have become crucial for ensuring production safety and product quality. Existing industrial equipment monitoring systems typically employ various sensors, industrial cameras, and PLC controllers to collect equipment operating data, which is then processed using big data analytics.

[0003] In industrial settings, monitoring data often originates from different modalities of acquisition terminals, such as video streams acquired by vision sensors, command streams fed back by controllers, and analog data acquired by various sensors. Existing technologies typically employ simple data splicing or hard alignment methods based on timestamps, which makes it difficult to eliminate spatiotemporal misalignments between data, resulting in low-quality monitoring data streams. This affects the accuracy of data processing and consequently the dynamic perception of the operational status of industrial equipment. Summary of the Invention

[0004] The purpose of this invention is to overcome the shortcomings of the prior art. This invention provides a data processing method and system for monitoring data based on big data analysis.

[0005] This invention provides a data processing method for monitoring data based on big data analytics, comprising: The system acquires monitoring data collected from industrial equipment monitoring terminals, performs data analysis on the monitoring data based on edge computing nodes, extracts corresponding local spatiotemporal features during the analysis process, and generates corresponding standard monitoring data streams by combining cross-modal spatiotemporal alignment mechanisms. The standard data stream is input into the big data analysis space, and the spatiotemporal joint features are extracted through the spatiotemporal joint encoder. Industrial equipment is used as the entity node. The temporal interaction similarity between nodes is determined based on the spatiotemporal joint features, and a continuously evolving industrial equipment knowledge graph is constructed by combining the time decay factor. The target nodes of the industrial equipment knowledge graph are labeled, the target nodes are feature-masked, and the distribution difference of the event evolution state transition network before and after the intervention is determined by combining the topological intervention mechanism. The abnormal state of the industrial equipment is quantified by the distribution difference, and then the corresponding abnormal prediction result is output. Based on the identification of anomaly prediction results, the corresponding early warning response instructions are determined, the execution feedback data of the early warning response instructions is captured, and the execution feedback data is jointly aligned with the standard monitoring data stream to perform closed-loop updates of data processing events of industrial equipment.

[0006] This invention provides a data processing system for monitoring data based on big data analysis. This system is applied to the aforementioned data processing method for monitoring data based on big data analysis. The data processing system for monitoring data based on big data analysis includes: The monitoring data stream module is used to acquire monitoring data collected by the monitoring terminal of industrial equipment, perform data analysis on the monitoring data based on edge computing nodes, extract corresponding local spatiotemporal features during the analysis process, and generate corresponding standard monitoring data streams by combining cross-modal spatiotemporal alignment mechanisms. The big data analysis module is used to input standard data streams into the big data analysis space, extract spatiotemporal joint features through a spatiotemporal joint encoder, use industrial equipment as entity nodes, determine the temporal interaction similarity between nodes based on spatiotemporal joint features, and construct a continuously evolving industrial equipment knowledge graph by combining time decay factors. The prediction module is used to label the target nodes of the industrial equipment knowledge graph, perform feature masking on the target nodes, and combine the topological intervention mechanism to determine the distribution difference of the event evolution state transition network before and after the intervention. The distribution difference is used to quantify the abnormal state of the industrial equipment, and then output the corresponding abnormal prediction results. The data processing event module is used to determine the corresponding early warning response command based on the identification of anomaly prediction results, capture the execution feedback data of the early warning response command, and jointly align the execution feedback data with the standard monitoring data stream to perform closed-loop updates of the data processing events of industrial equipment.

[0007] Compared with the prior art, the beneficial effects of the present invention are: (1) Acquire monitoring data collected by industrial equipment monitoring terminal, perform data analysis on monitoring data based on edge computing node, extract corresponding local spatiotemporal features during the analysis process, and generate corresponding standard monitoring data stream by combining cross-modal spatiotemporal alignment mechanism. The generated standard monitoring data stream eliminates spatiotemporal asynchrony and scale differences, providing a unified and high-fidelity data foundation for subsequent big data analysis, and significantly improving the real-time performance and accuracy of data processing.

[0008] (2) By using a spatiotemporal joint encoder to extract features and combining them with a time decay factor to construct a continuously evolving knowledge graph, the multidimensional state of industrial equipment in terms of action details and trajectory trends can be accurately presented. The introduction of a time decay factor at the same time eliminates historical noise interference, ensuring that the knowledge graph can update its topology in real time with changes in working conditions, thereby improving the dynamic perception and accurate description of the operating status of industrial equipment.

[0009] (3) By performing feature masking and topological intervention on the target node, the abnormal state is quantified based on the distribution difference of the state transition network before and after the intervention. Causal inference logic is introduced, which effectively identifies the real causal relationship between variables and avoids false alarms and false alarms under complex interference in traditional correlation analysis. The quantitative evaluation index based on distribution difference can accurately capture potential abnormal intentions and significantly improve the robustness of industrial equipment anomaly prediction. Attached Figure Description

[0010] Figure 1 This is a flowchart illustrating the data processing method for monitoring data based on big data analysis in an embodiment of the present invention. Figure 2 This is a simplified flowchart of a specific application scenario of the data processing method for monitoring data based on big data analysis in this embodiment of the invention. Figure 3 This is a schematic diagram of the structural composition of a data processing system for monitoring data based on big data analysis in an embodiment of the present invention. Detailed Implementation

[0011] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present invention.

[0012] Please see Figures 1 to 3 A data processing method for monitoring data based on big data analytics, applied to big data analytics scenarios; the data processing method for monitoring data based on big data analytics includes: Step S11: Obtain monitoring data collected by the industrial equipment monitoring terminal, perform data analysis on the monitoring data based on the edge computing node, extract the corresponding local spatiotemporal features during the analysis process, and generate the corresponding standard monitoring data stream by combining the cross-modal spatiotemporal alignment mechanism. Step S12: Input the standard data stream into the big data analysis space, extract the spatiotemporal joint features through the spatiotemporal joint encoder, take industrial equipment as the entity node, determine the temporal interaction similarity between nodes based on the spatiotemporal joint features, and construct a continuously evolving industrial equipment knowledge graph by combining the time decay factor. Step S13: Mark the target nodes of the industrial equipment knowledge graph, perform feature masking on the target nodes, and combine the topology intervention mechanism to determine the distribution difference of the event evolution state transition network before and after intervention. Quantify the abnormal state of industrial equipment with the distribution difference, and then output the corresponding abnormal prediction results. Step S14: Determine the corresponding early warning response command based on the identification of the anomaly prediction result, capture the execution feedback data of the early warning response command, and jointly align the execution feedback data with the standard monitoring data stream to perform closed-loop updates of the data processing events of industrial equipment.

[0013] In step S11, the specific steps are as follows: S111: Real-time acquisition of monitoring data collected by industrial equipment monitoring terminals. The monitoring data includes action videos, processing procedure information and yield rate. The feature extraction network deployed on edge computing nodes performs data analysis on the monitoring data and extracts local spatiotemporal features of different modalities during the data analysis process. S112: Perform cross-modal spatiotemporal alignment of local spatiotemporal features of different modalities. During the alignment process, determine the corresponding equipment processing content and perform dynamic timestamp normalization alignment and spatial projection mapping to eliminate the spatiotemporal asynchrony and scale differences between monitoring data, thereby generating a standard monitoring data stream with a unified time reference and spatial coordinate system.

[0014] In step S111, the specific steps are as follows: At this time, the edge gateway deployed around the industrial equipment acquires the raw data stream collected by the monitoring terminal in real time; among them, the motion video data is acquired by a high frame rate industrial camera installed in the robotic arm's working area, and the data format is RGB image sequence, which is used to capture the motion trajectory of the robotic arm's end effector and the shape of the rubber strip accumulation; the processing information is read from the PLC controller in real time through the OPCUA protocol, which includes the encoder angle values ​​of the six joints of the robotic arm, the speed vector of the end effector, and the currently executed program segment number, reflecting the control command status of the equipment.

[0015] The yield rate data comes from the downstream visual quality inspection unit and includes real-time measurements of the tape width and defect labels. During the acquisition process, the edge computing node performs noise reduction and region of interest (ROI) cropping on the motion video, normalizes the processing information, and performs sliding window smoothing on the yield rate data to eliminate high-frequency noise and outliers in the original data, providing a clean data source for subsequent feature extraction.

[0016] Lightweight 3D convolutional neural networks (3D-CNNs) are deployed within edge computing nodes to mine spatiotemporal features from motion video data. The network takes a continuous sequence of video frames as input and performs sliding calculations in both the temporal and spatial domains using 3D convolutional kernels to extract local spatiotemporal features during the robotic arm's movement. Specifically, the spatial convolutional operation is used to extract the geometric features of the glue gun head, the texture features of the glue strip edges, and the spatial position features of the window surface. The temporal convolutional operation is used to capture the dynamic trend features of the robotic arm's movement, such as the changes in motion acceleration of the glue gun head in different stages such as "accelerated start-up," "uniform glue movement," and "corner deceleration," outputting a local spatiotemporal feature vector of the motion video modality. This vector not only contains the spatial structure information of a single frame but also encodes the motion coherence and stability information within a short time window.

[0017] For machining process information with strong temporal dependencies, a temporal encoder based on Long Short-Term Memory (LSTM) or Gated Recurrent Unit (GRU) is used for feature extraction. Edge computing nodes input the normalized six-joint angle data and velocity vectors into the encoder in the order of time steps, and use a gating mechanism to selectively remember and forget historical states. In this process, not only are the configuration features of the robotic arm under a single time step extracted, i.e., the instantaneous position of each joint, but the coupling features and temporal dependencies between the movements of each joint are also explored. For example, when performing the process of applying glue to the R-corner of a car window, the network extracts the torque change features and trajectory tracking error features when the 5th and 6th axes of the robotic arm are linked, thereby forming a local spatiotemporal feature vector that can characterize the dynamic characteristics and control logic state of the robotic arm.

[0018] To address the dual numerical and semantic characteristics of yield rate data, a processing module combining statistical feature extraction and semantic embedding is constructed. For real-time adhesive strip width measurements, statistical functions within a sliding time window are used to extract statistical features such as mean, variance, and extreme values ​​to quantify the stability of the adhesive application process. For defect labels and process type labels, a word embedding layer maps them into low-dimensional dense vectors to preserve their semantic relevance. Furthermore, a fully connected layer fuses the statistical feature vectors and semantic embedding vectors to generate a local spatiotemporal feature vector for the yield rate modality. This vector reflects the trend of processing quality evolution with process stages in the time dimension and maps the correspondence between quality status and specific process segments in the spatial dimension.

[0019] The edge computing nodes perform preliminary channel concatenation or feature cascading on the extracted motion video feature vectors, processing procedure feature vectors, and yield rate feature vectors. During this process, an attention mechanism is introduced to initially evaluate the contribution of each modality feature in the current time slice. For example, motion video features are given higher weight in the corner glue application stage, and processing procedure information is given higher weight in the stable glue feeding stage. A set of local spatiotemporal feature matrices containing multimodal information is output. This matrix retains the original spatiotemporal resolution of each modality and eliminates some redundant information.

[0020] In step S112, the specific steps are as follows: At this point, the edge computing nodes construct a multimodal feature association model and perform semantic-level fusion analysis using the local spatiotemporal features of each modality extracted in step S111 to determine the specific processing content of the industrial equipment at the current moment. Specifically, the motion trajectory features in the motion video features and the semantic features of the program segments in the processing procedure information are used to perform similarity calculation and semantic matching to identify the correspondence between the action mode of the robotic arm end effector in the physical space and the instruction mode in the control logic space. Through this matching process, the current working status label of the equipment is determined, such as "straight-line uniform speed glue application segment", "corner deceleration glue application segment", or "glue lifting / retracting segment", thereby providing a unified semantic context constraint for subsequent cross-modal alignment and ensuring the consistency of video data stream and process data stream in terms of logical content.

[0021] To address the asynchronous issues caused by the difference in sampling frequency and transmission delay between motion video data streams and processing procedure information data streams, a dynamic time warping algorithm is employed for nonlinear alignment on the time axis. Specifically, using the high-frequency motion video feature sequence as the reference time axis, the cumulative distance matrix between it and the low-frequency or processing procedure feature sequence with transmission jitter is calculated, and the optimal curved path is searched to minimize the feature distance between the two in the time dimension.

[0022] In this process, the focus is on correcting the discrepancy between the arrival time of video frames and the timestamp of PLC commands caused by network transmission jitter. For example, the timestamp of the video frame when the robotic arm performs a "corner action" is precisely aligned with the timestamp of the PLC issuing the "corner command" to eliminate time offset and ensure the synchronization of different modal data with millisecond-level precision.

[0023] To address the scale differences in physical dimensions and feature space distribution among different modal data, a cross-modal shared feature space is constructed, and spatial projection mapping is performed using a deep metric learning network. Specifically, the pixel coordinate system containing motion video features, the joint angle coordinate system containing processing procedure information, and the tolerance coordinate system containing yield rate data are projected onto the same high-dimensional hyperspherical feature space through modality-specific nonlinear mapping functions, such as multilayer perceptrons. Within this space, a triplet loss function is used to constrain the feature distribution, ensuring that feature vectors describing the same processing content, such as "normal gluing," are spatially closer to each other, while feature vectors describing different processing states are further apart. This eliminates feature scale differences and achieves homogenized representation of heterogeneous data.

[0024] The multimodal feature vectors, after time alignment and spatial mapping, are concatenated or weighted and fused to generate a standard monitoring data stream with a unified time base and spatial coordinate system. Specifically, a unified metadata header is encapsulated for the fused feature vectors. The metadata header includes a globally aligned timestamp, a unique device identifier, a defined processing content tag, such as "process segment ID_03", and a data quality check bit. This standard data stream is encapsulated using a unified communication protocol, such as DataQueue or Protobuf format, to ensure that the data maintains temporal integrity, spatial consistency, and semantic clarity during the transmission from edge computing nodes to the big data analysis space, providing standardized input for the subsequent construction of an industrial equipment knowledge graph.

[0025] In step S12, the specific steps are as follows: S121: The standard monitoring data stream is input in parallel to the big data analysis space in the cloud. In this big data analysis space, the encoder performs downsampling and feature aggregation along the time and space dimensions through stacked spatiotemporal graph convolutional networks, and extracts multi-scale spatiotemporal joint features, which include the processing details and processing trajectory trends of industrial equipment. S122: Industrial equipment identified in the monitoring screen is treated as entity nodes. Its corresponding spatiotemporal joint features are projected into a high-dimensional latent space. In the high-dimensional latent space, multi-factor fusion is performed by combining the past abnormal information of the industrial equipment to determine the temporal interaction similarity between nodes. Furthermore, a time decay factor that decreases exponentially over time is introduced. The relationship between nodes and edges is continuously updated based on a dynamic graph reasoning mechanism, thereby constructing an industrial equipment knowledge graph whose topology evolves continuously over time.

[0026] In step S121, the specific steps are as follows: At this point, the standard monitoring data stream generated by the edge computing nodes is input in parallel to the big data analysis space in the cloud through a high-bandwidth backbone network. The big data analysis space performs tensor reconstruction on the received multi-source heterogeneous data to construct a unified high-dimensional data tensor. In this tensor, the standard monitoring data stream at each moment is regarded as a feature vector, and combined with the time series length and feature dimension, a three-dimensional input tensor X∈RT×N×C is constructed, where T represents the time step, N represents the number of nodes of the industrial equipment and its "key components, such as robotic arm joints and end effectors", and C represents the unified feature dimension generated in step S112. During the construction process, the adjacency matrix is ​​initialized according to the physical topology of the industrial equipment, and the physical connection relationship between the robotic arm base, each joint node and the end effector is defined, providing structural prior knowledge for subsequent graph convolution operations.

[0027] The spatiotemporal joint encoder deployed in the big data analysis space utilizes stacked graph convolutional layers to aggregate features along the spatial dimension. Specifically, for each time step, the graph convolutional network performs weighted summation and nonlinear transformation on the features of each node of the robotic arm according to the adjacency matrix, realizing the propagation and aggregation of features between nodes. In this process, the network captures the coupling relationship between the joint nodes of the robotic arm. For example, when the end effector performs the "window R-angle glue application" action, the graph convolutional layer can aggregate the features of the base, upper arm, forearm and wrist joints, and extract the spatial structural features of "multi-joint linkage to maintain constant end effector posture", thereby eliminating the limitations of single node data and forming a spatial feature map that can characterize the configuration of the entire robotic arm system.

[0028] The spatiotemporal co-encoder utilizes temporal convolution kernels to perform sliding window calculations and downsampling operations along the time dimension. Specifically, dilated convolution or stride convolution operations are employed to reduce the dimensionality of the input time-series features, gradually expanding the temporal receptive field. During downsampling, the network not only reduces the temporal resolution of the data to decrease computational complexity but also focuses on extracting long-term temporal dependent features. For example, it downsamples millisecond-level raw motion features to second-level or process-level semantic features, capturing the complete motion evolution trend of the robotic arm from "starting and accelerating" to "uniform speed applying glue" and then to "decelerating and stopping," thereby extracting robust trend features in the time dimension and filtering out high-frequency noise interference.

[0029] Iterative computation is performed using a multi-layered, stacked spatiotemporal graph convolutional network to extract spatiotemporal joint features encompassing different time spans and spatial ranges. Specifically, shallow network layers focus on extracting microscopic, instantaneous processing details, such as the momentary vibration of the nozzle and the microscopic morphology of the adhesive strip cross-section. Deep network layers, on the other hand, focus on extracting macroscopic, global processing trajectory trend features, such as the envelope shape of the entire adhesive application trajectory and the overall execution continuity of the process. Shallow and deep features are then connected in leaps or fused with weights to output a set of multi-scale spatiotemporal joint feature vectors containing rich semantic information. These feature vectors contain both precise processing details and implicit long-term trend information of equipment operation, comprehensively representing the real-time operating status of industrial equipment.

[0030] For the spatiotemporal co-encoder, it consists of three cascaded spatiotemporal graph convolutional modules. Each module contains a spatial graph convolutional layer, a temporal convolutional layer, and a residual connection unit. The input dimension of the first module is consistent with the feature dimension of the standard monitoring data stream. The input dimension of subsequent modules is the number of output channels of the previous module. The number of output channels of the three modules are set to 64, 128, and 256 respectively.

[0031] In the spatial dimension, each module employs a graph convolutional layer based on Chebyshev polynomial approximation for spatial feature aggregation. This graph convolutional layer utilizes a predefined adjacency matrix based on the physical topology of industrial equipment, and scales and expands the adjacency matrix using a second-order Chebyshev polynomial to achieve a weighted summation of the features of each node's first-order and second-order neighbors. After each graph convolutional layer, batch normalization is used to standardize the aggregated features, and a random deactivation operation with a dropout rate of 0.2 is used to suppress overfitting.

[0032] In the temporal dimension, each module employs a temporal convolutional layer to extract temporal features from the spatially aggregated features. The temporal convolutional layer consists of two sets of operations cascaded together: the first set is a one-dimensional convolutional kernel with a width of 3 and a stride of 1, used to extract local temporal dependencies; the second set is a one-dimensional convolutional kernel with a width of 5 and a stride of 2, used to downsample the temporal dimension, reducing the temporal length to half its original value. Both sets of convolutional operations are followed by a modified linear unit activation function. Through these downsampling operations, the three modules sequentially reduce the temporal resolution of the input from the original millisecond level to the second level, and the temporal receptive field expands from the initial 10 time steps to 30 and then 70 time steps.

[0033] The output of each spatiotemporal graph convolutional module is added to its input via residual connections to mitigate the vanishing gradient problem during deep network training. The residual connections use 1×1 convolutions to adapt the number of channels in the input features to match the number of channels in the module's output. After the three modules are cascaded, the deep features output from the third module, the mid-level features output from the second module, and the shallow features output from the first module are fused at multiple scales. The fusion method involves applying global average pooling to each of the three feature layers to compress the temporal dimension into a single time step, and then concatenating the pooled features along the channel dimension to obtain a multi-scale spatiotemporal joint feature vector containing micro-processing details, mid-range dynamic trends, and macro-trajectory evolution. This feature vector is then input to a fully connected layer for dimensional projection, with the output dimension set to 128, serving as the input representation for downstream knowledge graph construction.

[0034] In step S122, the specific steps are as follows: At this point, the big data analysis space performs entity recognition on the monitoring screen, locking onto the industrial equipment and its key components in the screen, defining them as entity nodes in the knowledge graph. Specifically, the entire industrial equipment robotic arm, its six joint units (J1-J6), and the "end effector: glue gun head" are mapped as independent nodes in the graph. Using the spatiotemporal joint features extracted in step S121, the feature vectors of each node are projected onto a unified high-dimensional latent space through a nonlinear mapping function constructed by a fully connected layer. In this high-dimensional latent space, the originally discrete multimodal features are transformed into dense vector representations, and the geometric distance between vectors can reflect the semantic similarity of the equipment's operating status. For example, the node vectors of nodes with good operating status will cluster in the healthy cluster, while the node vectors of nodes with potential anomalies will shift towards the abnormal cluster.

[0035] For node vectors in the high-dimensional latent space, the temporal interaction similarity between different nodes is calculated to quantify the degree of cooperation between equipment components and the correlation with fault propagation. Specifically, an attention mechanism is introduced to calculate the number of cross-relationships between nodes, and multi-factor fusion is performed by combining the past anomaly information of industrial equipment. The past anomaly information comes from the historical fault database, which contains records of high temperature, vibration or trajectory deviation that specific nodes have experienced in historical operation. During the fusion process, if "two nodes, such as the 5th and 6th joints of a robotic arm" frequently and simultaneously exhibit anomalies in historical data, or if the Euclidean distance between their current spatiotemporal joint feature vectors in the high-dimensional latent space is less than a preset threshold, they are given a higher interaction similarity weight, indicating that there is a strong coupling relationship or a potential fault propagation path between the two, thus forming a dynamically changing adjacency matrix weight.

[0036] To reflect the aging trend and changes in operating conditions of industrial equipment over time, a time decay factor is introduced to correct the interaction similarity between nodes. The time decay factor is defined as α(t) = e−λΔt, where λ is the decay coefficient and Δt is the time interval since the last data update. This mechanism emphasizes the importance of recent monitoring data and weakens the influence of long-term historical data on the current graph structure. For example, the influence weight of a momentary vibration signal of a robotic arm that occurred ten minutes ago will decrease exponentially over time, while the weight of a trajectory deviation that is currently occurring will remain at its maximum value. Through this correction, the graph can avoid non-continuous random disturbances in a timely manner and retain state characteristics with long-term trends.

[0037] Optionally, if a complete adhesive application cycle is sixty seconds, then the λ value is set to 1.5 times one-sixtieth, i.e., 0.025. This setting ensures that after three cycles, the influence weight of historical features decays to less than five percent of the initial value, thus preserving recent key state information while promptly eliminating interference.

[0038] Based on the corrected node similarity matrix, the topology of the knowledge graph is continuously updated using a dynamic graph reasoning mechanism. Specifically, a temporal graph neural network (DynamicGNN) is used to iteratively update the graph on a time-slice basis. If the interaction similarity between nodes exceeds a preset threshold in multiple consecutive time slices, a connection edge is built or maintained between the corresponding nodes, and corresponding attribute labels are assigned, such as "mechanical coupling", "heat conduction" or "fault association". Conversely, if the similarity is lower than the threshold, the connection edge is disconnected. In this way, an industrial equipment knowledge graph with a continuously evolving topology is constructed. This graph not only records the current static structural information of the industrial equipment, but also dynamically represents the real-time interaction status and evolution trend of each component during the glue application process.

[0039] In step S13, the specific steps are as follows: S131: Mark target nodes with potential risk characteristics in the industrial equipment knowledge graph, use a self-supervised masking strategy to randomly mask the key behavioral characteristics of the target nodes and their neighboring nodes, output the masked industrial equipment knowledge graph, and combine the topological intervention mechanism of causal inference to trigger virtual intervention of the masked industrial equipment knowledge graph. S132: Reconstruct the event evolution state transition network before and after intervention using a variational graph autoencoder, and determine the distribution difference of the state transition probability distribution before and after intervention; map the distribution difference into an abnormal intent quantitative evaluation index, and perform nonlinear mapping on the abnormal intent evaluation index in combination with a preset intent classifier, thereby outputting the abnormal prediction results of industrial equipment in the future time window.

[0040] In step S131, the specific steps are as follows: At this point, the big data analysis engine performs a full scan of the continuously evolving industrial equipment knowledge graph constructed in step S122. By comparing the real-time state feature vector with the preset abnormal pattern feature library, it identifies and marks target nodes with potential risk characteristics. Specifically, it calculates the Mahalanobis distance between the feature vector of each entity node in the current time slice and its historical steady-state feature vector. If the distance exceeds the preset risk threshold, it determines that the node has potential risk.

[0041] To set a preset risk threshold, the system collects feature vectors from each node of the industrial equipment during 72 consecutive hours of normal operation. It calculates the Mahalanobis distance sequence of each node's feature vector relative to its steady-state mean during this period, and uses the 95th percentile of this sequence as the initial risk threshold. During subsequent operation, the system continuously updates the Mahalanobis distance distribution under normal operating conditions using a sliding time window (with a window width set to 12 hours). Every 60 minutes, it recalculates the current 95th percentile and replaces the original threshold with this updated value, thus achieving adaptive tracking of the threshold to slow equipment aging or operational condition drift.

[0042] At the same time, based on the topological location attributes of the nodes, nodes located in "critical process paths, such as the core path of the adhesive application trajectory" or in high connectivity center positions are prioritized for marking. The marked target nodes will be assigned specific risk identifier bits for subsequent masking processing and intervention analysis, thereby locking in the key sources that may cause system-level failures.

[0043] A masking strategy from self-supervised learning is employed to randomly mask key behavioral features of labeled target nodes and their first-order neighbor nodes. Specifically, a binary mask matrix with the same dimension as the node features is constructed, and mask positions are randomly generated according to a preset masking ratio, such as 15%-30%. For the target node, core features reflecting its operational state, such as joint torque and end-effector velocity, are masked. For neighbor nodes, closely related features interacting with the target node are masked. The mask matrix is ​​then multiplied by the original feature vector, and the feature values ​​at the masked positions are either set to zero or replaced with learnable mask tokens. This operation aims to simulate scenarios of missing node data or functional failure, forcing subsequent models to reconstruct missing features using contextual information, thereby assessing the impact of node failure on the overall system.

[0044] Combining intervention theory in causal inference, a topological intervention mechanism is constructed to virtually intervene in the masked industrial equipment knowledge graph. Specifically, the intervention operation is defined as cutting off a specific information flow channel between the target node and its neighboring nodes, that is, forcibly resetting the weight of the corresponding edge in the graph adjacency matrix to zero, or modifying the attribute state of the edge to "blocking state". This operation is logically equivalent to cutting off the physical connection or signal transmission path between equipment components, simulating the extreme working condition of "assuming the component is completely failed or isolated". Through the message passing mechanism of graph neural network, the graph structure disturbance and feature propagation blocking effect caused by the intervention operation are observed, thus providing a virtual experimental environment for subsequent analysis of abnormal propagation paths and quantification of system vulnerability.

[0045] For the topology intervention mechanism, the system traverses the set of first-order neighbor nodes of the target node, checking one by one whether there is a connection edge between the target node and each first-order neighbor node in the knowledge graph. If a connection edge exists, it is added to the set of edges to be intervened. At the same time, the system calculates the weighted degree centrality index of the target node, which is obtained by summing the weights of all connection edges of the target node. If the weighted degree centrality value exceeds twice the average weighted degree centrality value of all nodes, the connection edges of the target node's second-order neighbors are also added to the set of edges to be intervened, in order to cover the multi-cascade failure propagation paths that may be caused by high-influence nodes.

[0046] The weight values ​​of the corresponding edges in the knowledge graph adjacency matrix are forcibly modified from the originally calculated interaction similarity values ​​to zero. Simultaneously, the attribute status labels of these edges are uniformly changed from the original connection status categories (including "mechanical coupling," "heat conduction," or "fault association") to the preset blocking status labels. When performing the blocking operation, the system does not delete the topological record of the edge; instead, it retains the node numbers at both ends of the edge and the original weight values, and adds an intervention timestamp field and an intervention flag field. The intervention flag field is set to true to record the original connection state of the edge before intervention, facilitating subsequent comparative analysis of the states before and after intervention.

[0047] The system replaces the original adjacency matrix with a modified one and re-executes the forward propagation calculation through the graph convolutional layer in the constructed spatiotemporal joint encoder. During this recalculation, feature information between the nodes at the two ends of the blocked edge is no longer transmitted along that edge. The aggregated features of the target node consist only of its own features and the features of its unblocked neighboring nodes, thus simulating the equipment state changes after the connection path completely fails in a real industrial scenario. The system records the changes in the feature vectors of each node before and after the recalculation. Nodes with changes exceeding a preset threshold are marked as affected nodes, and the number and distribution range of affected nodes are counted to quantify the degree of disturbance to the overall system state caused by the target node's anomaly.

[0048] The system sequentially performs the aforementioned intervention operations on different target nodes, recording the corresponding changes in node characteristics and the distribution data of affected nodes after each intervention, forming an intervention effect record table. This record table uses the target node number as an index and stores the intervention timestamp, the number of blocked edges, the number of affected nodes, and the spatial distribution characteristics of the affected nodes for each intervention operation. Based on this record table, the system ranks and compares the intervention effects of multiple target nodes, identifying the target node with the largest number of affected nodes and the widest distribution range after intervention as the most critical vulnerable node in the system, and its corresponding abnormal propagation path as the highest-risk propagation path.

[0049] In step S132, the specific steps are as follows: At this point, a variational graph autoencoder is deployed within the big data analysis space to perform low-dimensional embedding reconstruction of the industrial equipment knowledge graph before and after intervention generated in step S131. The encoder part of the variational graph autoencoder maps the node features and topological structure in the graph to the mean and variance vectors of the latent space, and obtains the latent representation through sampling operations. The decoder part uses the latent representation to reconstruct the adjacency relationships and feature distributions between nodes. During this process, reconstruction operations are performed on the original graph before intervention and the virtual graph after intervention, generating two corresponding event evolution state transition networks. This reconstruction process not only recovers the feature information obscured by the mask but also predicts the transition path of the equipment state in future time steps under the current topological constraints, forming a dynamic state manifold.

[0050] The state transition probability distributions of the evolutionary state transition networks before and after the intervention are calculated separately, and the differences between the two distributions are quantified. Specifically, the reconstructed graph feature space is divided into several discrete state intervals, such as "normal operation", "accuracy degradation" and "fault shutdown", and the transition probability density of node features in these intervals is statistically analyzed. KL divergence is used as a metric to calculate the information loss of the state distribution after intervention relative to the state distribution before intervention. If the KL divergence value is large, it indicates that the "intervention operation: i.e., simulated node failure or edge disconnection" has a significant disturbance to the state evolution of the system, which means that the target node has a key influence in the system, and its abnormality will cause a drastic deviation in the overall operating trajectory of the system.

[0051] The calculated "distribution difference value: KL divergence" is mapped to an abnormal intent quantitative assessment index. Here, the abnormal intent quantitative index Ia is defined as a nonlinear transformation function of the distribution difference, Ia=f(DKL), where the function f is designed as a monotonically increasing function to amplify the weight of significant differences and suppress fluctuations caused by small noise. This index not only reflects the degree to which the equipment deviates from normal operating conditions, but also implies the "intent" of the equipment's operating state to evolve towards a specific failure mode. For example, when the index value exceeds a preset threshold, it indicates that the equipment not only has a current abnormality, but also has a strong tendency to evolve into a serious failure.

[0052] A pre-defined intent classifier is constructed, taking the quantitative evaluation index of abnormal intent as input, and outputting specific anomaly prediction results through nonlinear mapping. The intent classifier adopts a multilayer perceptron or support vector machine structure, and its output layer is mapped to several predefined anomaly category probabilities, such as "mechanical wear anomaly", "control misalignment anomaly", and "process deviation anomaly". The classifier outputs the confidence score of each anomaly category through the Softmax function, and selects the category with the highest confidence as the final prediction result. At the same time, combined with a time series prediction model, the predicted value of the time window for the anomaly to occur is output, thereby realizing the leap from "current status monitoring" to "future risk prediction".

[0053] In step S14, the specific steps are as follows: S141: Based on the danger level and intent type of the anomaly prediction results, retrieve the preset emergency response strategy library to automatically match and issue the corresponding early warning response instructions to the execution terminal; capture the execution feedback data returned by the execution terminal after completing the early warning response instructions in real time, including action execution logs and monitoring screens; S142: A cross-modal contrastive learning mechanism is adopted to jointly align and feature-concatenate the execution feedback data with the standard monitoring data stream on the same time axis, thereby generating a closed-loop feedback feature vector; the closed-loop feedback feature vector is input into the reinforcement learning reward model to calculate the comprehensive reward value of the current data processing event, and the comprehensive reward value is mechanically propagated back to the source, so as to realize the adaptive closed-loop update of the data processing event of the industrial equipment during the propagation process.

[0054] In step S141, the specific steps are as follows: At this point, the big data analysis center has built a pre-defined emergency response strategy library. This strategy library is stored in a multi-dimensional index structure. The index key values ​​include hazard levels, such as "general warning", "serious fault", and "emergency shutdown", and intent types, such as "mechanical jam", "trajectory deviation", and "process deviation". The system reads the anomaly prediction results output in step S132, extracts the hazard level labels and intent type labels, and uses them as query keys to search and match in the emergency response strategy library. The matching process uses a weighted similarity calculation algorithm to select the strategy entries with the highest matching degree to the current abnormal situation. If there are multiple matching strategies, a unique optimal execution strategy is determined according to the pre-defined "priority rules, such as safety taking precedence over continuity". This strategy includes a specific control instruction sequence, target parameter set, and execution logic.

[0055] The retrieved emergency response strategies are transformed into machine-executable early warning response instructions and sent to the execution terminals on the edge. The control center encapsulates the strategy parameters into standard industrial control protocol messages, such as Modbus TCP messages or OPCUA service call instructions. These messages contain instruction codes, such as "speed degradation," "emergency stop trigger," target device address, and execution parameters. Through a highly reliable transmission channel of the Industrial Internet of Things (IIoT), the instructions are pushed in real-time to the PLC control unit of the industrial equipment. During the distribution process, a handshake mechanism and heartbeat detection are introduced to ensure that the instructions arrive at the execution terminal accurately within milliseconds, preventing control failures due to network packet loss.

[0056] While executing the warning response command, the "Execution Terminal: PLC Controller" generates an action execution log in real time. This log records the command reception timestamp, command parsing result, execution action type (e.g., "modify servo gain", "switch air path valve island"), and execution completion status ("success", "failure", "timeout"). The edge computing node captures this action execution log in real time through data bypass technology using high-frequency sampling and adds a globally unified timestamp. This log data reflects the physical response process of the equipment to the abnormal warning and is a key objective basis for evaluating the effectiveness of the response.

[0057] To verify the actual physical effect after the early warning response command is executed, the system synchronously triggers the special capture or recording function of the industrial camera to obtain the monitoring image after the execution terminal action. Specifically, when the "decelerate" command is executed, the camera immediately captures the motion trajectory image of the robotic arm end effector. When the "reset" command is executed, the camera captures the state image of the robotic arm returning to zero. The edge node packages the captured image data with the action execution log to generate a structured execution feedback data packet, which is then transmitted back to the big data analysis space in real time, forming a complete closed-loop evidence chain of "command-action-image".

[0058] In step S142, the specific steps are as follows: At this point, the big data analysis space employs a cross-modal contrastive learning mechanism to deeply align and fuse the execution feedback data returned in step S141 with the standard monitoring data stream on the same time axis. Specifically, a dual-branch feature extraction network is constructed to encode both the "execution feedback data (including action logs and monitoring screens)" and the standard monitoring data stream. A contrastive loss function is used to maximize the mutual information between the feedback data features and the standard data stream features, eliminating distribution shifts caused by "response actions, such as speed reduction." During the joint alignment process, the system performs feature-level concatenation of the "actual state after the warning response" and the "theoretical standard state" to generate a high-dimensional closed-loop feedback feature vector. This vector not only encodes the device's response result but also implicitly contains the actual corrective effect of the warning strategy on the device's operating state.

[0059] A deep learning-based reinforcement learning reward model is introduced, using closed-loop feedback feature vectors as input to quantify the execution effect of the current data processing event. The reward model design includes a multi-dimensional scoring mechanism, calculating "safety reward" (whether the device safely shuts down or recovers), "quality reward" (whether the adhesive strip returns to tolerance range), and "efficiency reward" (whether the response is timely and without excessive intervention). Through weighted summation, the comprehensive reward value Rtotal for the current event is output. This reward value reflects the quality of the warning response command in resolving potential anomalies: a positive value indicates the strategy is effective and the anomaly has been successfully eliminated; a negative value indicates the strategy is ineffective or has side effects, such as ineffective shutdown due to false alarms.

[0060] With the goal of maximizing long-term cumulative reward, the gradient direction of each network layer parameter is calculated. If the reward value is positive, the connection weight of the neuron that triggers the warning is enhanced. If the reward value is negative, the expression weight of related features is suppressed and the decision boundary is adjusted. In this process, the spatiotemporal joint encoder parameters in step S121, the knowledge graph construction weights in step S122, and the anomaly threshold judgment parameters in step S131 are updated in detail, so that the model can learn autonomously from successful interventions or lessons learned from failures and achieve adaptive optimization of parameters.

[0061] The optimized model parameters are synchronized to edge computing nodes and cloud analytics engines to update the anomaly detection criteria and the topology of the knowledge graph. For example, if a warning is confirmed as a false alarm by feedback data, the system will automatically raise the trigger threshold for that type of anomaly to reduce the false alarm rate. Conversely, if a missed alarm is detected, the threshold will be lowered to improve sensitivity. In this way, the system realizes a complete closed loop from "data perception" to "decision execution" and then to "feedback evolution", ensuring that the data processing method can continuously adapt to the changing operating conditions of industrial equipment in a flexible manufacturing environment.

[0062] Compared with the prior art, the beneficial effects of the present invention are: (1) Acquire monitoring data collected by industrial equipment monitoring terminal, perform data analysis on monitoring data based on edge computing node, extract corresponding local spatiotemporal features during the analysis process, and generate corresponding standard monitoring data stream by combining cross-modal spatiotemporal alignment mechanism. The generated standard monitoring data stream eliminates spatiotemporal asynchrony and scale differences, providing a unified and high-fidelity data foundation for subsequent big data analysis, and significantly improving the real-time performance and accuracy of data processing.

[0063] (2) By using a spatiotemporal joint encoder to extract features and combining them with a time decay factor to construct a continuously evolving knowledge graph, the multidimensional state of industrial equipment in terms of action details and trajectory trends can be accurately presented. The introduction of a time decay factor at the same time eliminates historical noise interference, ensuring that the knowledge graph can update its topology in real time with changes in working conditions, thereby improving the dynamic perception and accurate description of the operating status of industrial equipment.

[0064] (3) By performing feature masking and topological intervention on the target node, the abnormal state is quantified based on the distribution difference of the state transition network before and after the intervention. Causal inference logic is introduced, which effectively identifies the real causal relationship between variables and avoids false alarms and false alarms under complex interference in traditional correlation analysis. The quantitative evaluation index based on distribution difference can accurately capture potential abnormal intentions and significantly improve the robustness of industrial equipment anomaly prediction.

[0065] Please see Figure 2 The specific application scenarios and corresponding embodiments of this application are as follows: In step S11, on the production line of Flexible Manufacturing 4.0, the industrial equipment is a 6-DOF robotic arm that performs precision adhesive application on car windows. Its task is to apply sealant evenly to the edge of the window, requiring the width error of the adhesive strip to be controlled within ±0.1mm, and there to be no accumulation or breakage at the corners.

[0066] Edge computing nodes acquire monitoring data in real time: industrial cameras capture video of the movement of the end effector of the robotic arm at the corner of the car window at a frame rate of 60fps, the PLC controller synchronously feeds back the rotation angle and trajectory tracking instructions of each joint of the robotic arm, and the visual quality inspection sensor provides real-time feedback that the width of the rubber strip at the corner is 1.5mm.

[0067] For motion videos, a 3D convolutional network extracts the spatial feature of "the movement trajectory of the gun head at the R-angle is arc-shaped" and the temporal feature of "the gun head speed decreases from 200mm / s to 50mm / s to make a corner", forming a local spatiotemporal feature vector for the video. For processing procedure information, a timing encoder extracts the control feature of "the 5th and 6th axes perform linkage compensation" and the state feature of "currently in the 'R-angle glue application subroutine segment'", forming a local spatiotemporal feature vector for the process. For yield rate data, the statistics module calculates that the variance of the glue strip width fluctuation within the current time window is extremely small, and the semantic module identifies the current quality status as "good", forming a local spatiotemporal feature vector for quality.

[0068] The edge nodes fuse and output the above three sets of feature vectors, accurately depicting the action details, control state and quality performance of industrial equipment in the local time period of "R-corner gluing", providing high-fidelity underlying data support for the construction of a continuously evolving knowledge graph in the subsequent big data analysis space.

[0069] After the edge node obtains the features extracted in step S111, it identifies that the end effector of the robotic arm in the motion video is performing the action of "decelerating and changing the direction of motion". At the same time, the processing information shows that the "PNT_105_R50_Corner" program segment is currently being executed. The feature association model performs semantic matching between the two to determine that the processing content at the current moment is "corner gluing operation of the R50mm arc segment at the upper left corner of the window".

[0070] Since the video capture frequency is 60fps, while the PLC instruction cycle is 20ms, and network transmission causes the process information to arrive about 15ms later than the video data, the optimal matching path between the starting frame of the nozzle speed decrease in the video feature sequence and the starting time of the "deceleration instruction" in the process feature sequence is calculated by using a dynamic time warping algorithm. The late-arriving process data is shifted forward on the time axis, correcting the transmission delay and achieving precise synchronization between the time of nozzle contact inflection point in the video and the inflection point time recorded in the instruction.

[0071] The gun head position extracted from video features is represented by pixel coordinates, the joint angle in process information is represented by radians, and the adhesive strip width in yield data is represented by millimeters. These three have different dimensions. Through a deep metric learning network, the above three heterogeneous feature vectors are mapped to the same latent space. During the mapping process, the three feature vectors "gun head is located at the inflection point", "joint angle combination is θ1-θ6", and "adhesive strip width is 1.2mm" are brought closer in the latent space, so that their Euclidean distance is less than a preset threshold, thus eliminating the scale difference of physical dimensions.

[0072] The aligned feature vectors are encapsulated into a standard data package, along with a global timestamp "T_global_20231027103005.120" and a processing content tag "Process:Corner_R50", and output to the big data analysis space. This data stream accurately records the real state of the industrial equipment at the moment of corner gluing, providing high-quality input data for subsequent judgment of whether there is a risk of "corner accumulation" or "glue breakage".

[0073] In step S12, the edge node uploads the standard monitoring data stream of the industrial equipment over the past minute to the cloud. The cloud constructs it into a three-dimensional tensor with shape [6000,7,128], assuming a sampling rate of 100Hz, containing 6 joint nodes of the robotic arm and 1 end effector node, with a feature dimension of 128.

[0074] When applying adhesive to the R50 arc at the upper left corner of the car window, the spatial convolution kernel of the spatiotemporal graph convolutional network aggregates the angular change features of the 4th, 5th, and 6th joints. The network identifies that in order to keep the tangent direction of the gun head perpendicular to the edge of the car window in the arc segment, the 5th and 6th axes are finely coordinated and compensated. This spatial coupling feature is effectively extracted, characterizing the robot arm's posture adjustment capability on complex curved surfaces.

[0075] The encoder downsamples the time-series data within 60 seconds with a step size of 4, compressing 6000 time steps into 1500. During the dimensionality reduction process, the network captures the speed change trend of industrial equipment when passing through the corner of the car window: that is, "the speed first decreases linearly to 50mm / s to prevent the rubber strip from accumulating, and then linearly increases back to 200mm / s after passing the inflection point". This long-term time-series trend feature is preserved and enhanced, while instantaneous high-frequency voltage fluctuation interference is filtered out.

[0076] In the multi-scale feature vector, the processing detail feature shows that at time T=30.5s, the actual position of the gun head deviates from the theoretical trajectory by 0.05mm; while the processing trajectory trend feature shows that the overall glue coating trajectory is smooth and there is no obvious drift. The features of these two dimensions together constitute an accurate description of the current state of the industrial equipment, providing a high-fidelity spatiotemporal information carrier for building a knowledge graph in the subsequent step S12.

[0077] The big data analysis space identifies the key components of the industrial equipment robotic arm, defining the "5th axis reducer", "6th axis servo motor" and "end-effector glue gun head" as graph nodes; using the features of step S121, the current fluctuation characteristics and speed characteristics of the "6th axis servo motor" when performing circular interpolation are projected into a high-dimensional latent space to form the node feature vector VJ6.

[0078] When applying adhesive to the upper left corner of the car window (R50 arc), the system detected an abnormal current spike in the "6th axis servo motor" and a slight deviation in the trajectory of the "end-point adhesive gun". Calculating the vector distance between the two in the high-dimensional latent space revealed a significant reduction in distance. Combined with historical anomaly information, database records showed that when this type of robotic arm makes high-speed corners, overload of the J6 axis motor often causes gun head vibration. The multi-factor fusion model determined that there is a strong correlation between the two. The temporal interaction similarity between the nodes "J6 axis motor" and "gun head" was set to 0.92, close to 1.0 indicating a strong correlation, suggesting that motor anomalies directly affect the adhesive application quality.

[0079] Since the current spike occurred at the current time (T=10:30:05), the attenuation factor α(t) is close to 1, and this abnormal information is given the highest weight, significantly enhancing the "fault-related edge" in the graph. However, a transient voltage fluctuation that occurred ten minutes ago (T=10:20:00) has its influence weight reduced to 0.05 since the equipment has returned to normal, and it no longer affects the current graph structure.

[0080] Based on dynamic graph reasoning, the graph generated a dynamic connection edge from "J6 axis motor" to "gun head", labeled as "corner overload association". The graph topology update shows that during the straight glue application stage, the weight of this connection edge is 0.1, indicating a weak connection; once entering the corner glue application stage, the weight of this connection edge instantly jumps to 0.92, indicating a strong connection. This evolution graph clearly reveals the weak links and dynamic fault propagation paths of industrial equipment under specific processes, providing accurate topological basis for subsequent anomaly detection.

[0081] In step S13, the big data analysis engine scans the knowledge graph and finds that when the "6th axis servo motor node" of the industrial equipment robot arm performs R50 corner glue application, the Mahalanobis distance between its temperature characteristic value and vibration characteristic value exceeds the risk threshold, indicating that the node has a potential risk of "dynamic accuracy reduction due to overheating". At the same time, topology analysis shows that the node is directly connected to the "end glue gun head" and is on a critical process link. The system then marks the "6th axis servo motor node" as the target node and marks its downstream node "end glue gun head" as the affected neighbor node.

[0082] The system generates a mask matrix to mask the two key features of the "6th axis servo motor node"—"angle encoder reading" and "torque output value"—setting the feature value to [MASK]. At the same time, it masks the "actual position coordinates" feature of the neighboring node, the "end glue gun head." This simulates the working scenario where the end glue gun head cannot be accurately positioned when the control signal of the 6th axis motor is disordered due to a fault. This forces the model to infer the possible position of the gun head based on the state information of "other joints, such as the 4th and 5th axes."

[0083] To quantify the specific impact of the 6th axis failure on the adhesive coating quality, the system implemented virtual intervention: severing the causal edge from the "6th axis servo motor node" to the "end adhesive gun head" in the knowledge graph, that is, forcibly changing the corresponding weight in the adjacency matrix from "0.85, strong coupling" to "0, blocking"; at this time, the graph formed a virtual "broken chain" structure. By observing the loss of information propagation in the overall graph after severing this edge, the system constructed a virtual experimental field to "assume the 6th axis completely fails and how the adhesive coating trajectory will deviate".

[0084] The variational graph autoencoder reconstructs the knowledge graph of industrial equipment. For the graph before intervention: normal state, the feature vector sequence of the gun head trajectory reconstructed by the decoder presents a smooth arc shape. However, for the graph after intervention: simulated failure of the 6th axis motor, due to the masking of the 6th axis features and the cutting of the topological edges, the feature vector sequence of the gun head trajectory reconstructed by the decoder shows obvious sawtooth oscillation and position drift at the R50 corner.

[0085] The system calculates the state transition probability distribution before and after intervention. Under normal conditions, the nozzle position error is concentrated in the ±0.05mm range, and the transition probability distribution is Gaussian. After intervention, the reconstructed state distribution shifts to the "large error range, such as ±0.2mm", and the distribution variance increases significantly. Through KL divergence calculation, the distribution difference value is 3.5, which is significantly greater than the baseline value of 1.0, indicating that the anomaly of the 6th axis has a catastrophic impact on the coating accuracy.

[0086] Inputting the KL divergence value of 3.5 into the nonlinear mapping function, the abnormal intent quantification index Ia is calculated to be 0.92, ranging from 0 to 1. The larger the value, the higher the risk. This high index value reveals that the industrial equipment currently has a very strong intention to "deviate from process standards". Moreover, this intention is not a random fluctuation, but a deterministic trend caused by the failure of core driving components.

[0087] The intent classifier receives an input of Ia=0.92 and outputs a prediction result after nonlinear mapping: within the next 30-second glue application window, the probability of "corner glue accumulation" anomaly occurring in industrial equipment is 95%, and the probability of "glue strip breakage" anomaly occurring is 30%. The system determines the highest risk category as "corner glue accumulation" and immediately generates a prediction report containing the anomaly type, risk level, and expected occurrence time, providing accurate decision-making basis for subsequent early warning response.

[0088] In step S14, it has been predicted that the probability of "corner adhesive accumulation" anomaly occurring in the industrial equipment within the next 30 seconds is 95%, and the danger level is determined to be "serious warning" with the intent type as "process deviation". The system uses this tag to search the strategy library and matches the strategy number "RS-050: corner overload speed reduction compensation strategy". This strategy aims to eliminate the risk of adhesive strip accumulation by reducing the corner speed.

[0089] The control center encapsulates the strategy RS-050 into an OPCUA instruction, with the target address pointing to the PLC of the industrial equipment's robotic arm. The instruction parameters are set to "reduce the corner feed speed of the 5th and 6th axes from 200mm / s to 80mm / s, and simultaneously reduce the glue output by 20%". This instruction is sent to the control unit of the industrial equipment in real time via industrial Ethernet to ensure that it takes effect before the next corner action begins.

[0090] After receiving the instruction, the PLC of the industrial equipment modifies the speed loop parameters of the servo drive and records the action execution log as follows: "10:30:06.500: Received speed reduction instruction > 10:30:06.600: Parameter modification completed > 10:30:06.700: Speed ​​has been reduced to 80mm / s". The edge node captures this log in real time, verifying that the instruction has been successfully executed.

[0091] When the robotic arm passed through the R50 corner of the car window at a new speed, the industrial camera triggered a special shooting, capturing the glue gun head smoothly passing through the corner with uniform glue strip shape and no accumulation. This monitoring image and action execution log were packaged into execution feedback data and transmitted back to the big data center in real time, confirming that the early warning response command effectively eliminated potential quality anomalies and provided successful case data for subsequent closed-loop updates.

[0092] The system compares the "monitoring image of the robot arm corner after deceleration" captured in step S141 with the "theoretical trajectory of the corner in the standard monitoring data stream". Through cross-modal comparative learning, it finds that the cosine similarity between the actual trajectory feature vector and the theoretical trajectory feature vector of the robot arm after deceleration reaches 0.98, which is highly overlapping, and the action log shows that "the speed is stable at 80mm / s". The system concatenates these two sets of feature vectors to generate a closed-loop feedback feature vector Vfeedback, which clearly indicates that "the deceleration strategy has successfully corrected the trajectory deviation".

[0093] Inputting Vfeedback into the reinforcement learning reward model, the model calculates the following: quality reward is +0.9, the rubber strip width error is corrected to 0.05mm, meeting the ±0.1mm requirement; safety reward is +1.0, the equipment experienced no impact or alarm; efficiency reward is -0.2, the cycle time of this process was extended by 0.5 seconds due to the speed reduction; after weighted calculation, the comprehensive reward value Rtotal=0.8, this high positive value indicates that the "corner overload speed reduction compensation strategy" is successful.

[0094] The system performs backpropagation based on a reward value Rtotal=0.8. The gradient update process strengthens the causal connection weight between "high temperature of the 6th axis motor" and "corner deceleration command" in the knowledge graph. At the same time, the threshold of the anomaly detection model is fine-tuned, so that the model learns a flexible strategy of "prioritizing deceleration compensation rather than emergency stop when the temperature of the 6th axis rises slightly but the trajectory is normal".

[0095] The updated parameters are synchronized to the edge gateway; in the next window adhesive application operation, when the industrial equipment approaches the R50 corner again and detects similar vibration characteristics, the system can more accurately predict the risk and smoothly execute the deceleration action in advance. The adaptive adjustment of process parameters can be achieved without manual intervention, which meets the requirements of flexible manufacturing 4.0 for intelligent operation and maintenance.

[0096] Please see Figure 3 The data processing system for monitoring data based on big data analysis is applied to the aforementioned data processing method for monitoring data based on big data analysis; the data processing system for monitoring data based on big data analysis includes: The monitoring data stream module 21 is used to acquire monitoring data collected by the monitoring terminal of industrial equipment, perform data analysis on the monitoring data based on the edge computing node, extract the corresponding local spatiotemporal features during the analysis process, and generate the corresponding standard monitoring data stream by combining the cross-modal spatiotemporal alignment mechanism. The big data analysis module 22 is used to input standard data streams into the big data analysis space, extract spatiotemporal joint features through a spatiotemporal joint encoder, use industrial equipment as entity nodes, determine the temporal interaction similarity between nodes based on spatiotemporal joint features, and construct a continuously evolving industrial equipment knowledge graph by combining time decay factors. The prediction module 23 is used to label the target nodes of the industrial equipment knowledge graph, perform feature masking on the target nodes, and combine the topological intervention mechanism to determine the distribution difference of the event evolution state transition network before and after the intervention. The distribution difference is used to quantify the abnormal state of the industrial equipment, and then output the corresponding abnormal prediction result. The data processing event module 24 is used to determine the corresponding early warning response command based on the identification of the anomaly prediction result, capture the execution feedback data of the early warning response command, and jointly align the execution feedback data with the standard monitoring data stream to perform closed-loop updates of the data processing events of industrial equipment.

[0097] It should be noted that although multiple modules are mentioned in the detailed description above, this division is not mandatory; in fact, according to the embodiments of this disclosure, the features and functions of two or more modules or described above can be embodied in one module; conversely, the features and functions of one module described above can be further divided into multiple modules to be embodied.

[0098] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A data processing method for monitoring data based on big data analysis, characterized by, include: The system acquires monitoring data collected from industrial equipment monitoring terminals, performs data analysis on the monitoring data based on edge computing nodes, extracts corresponding local spatiotemporal features during the analysis process, and generates corresponding standard monitoring data streams by combining cross-modal spatiotemporal alignment mechanisms. The standard data stream is input into the big data analysis space, and the spatiotemporal joint features are extracted through the spatiotemporal joint encoder. Industrial equipment is used as the entity node. The temporal interaction similarity between nodes is determined based on the spatiotemporal joint features, and a continuously evolving industrial equipment knowledge graph is constructed by combining the time decay factor. The target nodes of the industrial equipment knowledge graph are labeled, the target nodes are feature-masked, and the distribution difference of the event evolution state transition network before and after the intervention is determined by combining the topological intervention mechanism. The abnormal state of the industrial equipment is quantified by the distribution difference, and then the corresponding abnormal prediction result is output. Based on the identification of anomaly prediction results, the corresponding early warning response instructions are determined, the execution feedback data of the early warning response instructions is captured, and the execution feedback data is jointly aligned with the standard monitoring data stream to perform closed-loop updates of data processing events of industrial equipment.

2. The data processing method for monitoring data based on big data analysis according to claim 1, characterized in that, The process involves acquiring monitoring data collected by the industrial equipment monitoring terminal, performing data analysis on the monitoring data based on edge computing nodes, extracting corresponding local spatiotemporal features during the analysis, and generating corresponding standard monitoring data streams by combining cross-modal spatiotemporal alignment mechanisms, including: The system acquires real-time monitoring data from industrial equipment monitoring terminals, including action videos, processing procedure information, and yield rates. A feature extraction network deployed on edge computing nodes performs data analysis on the monitoring data, extracting local spatiotemporal features of different modalities during the data analysis process.

3. The data processing method for monitoring data based on big data analysis according to claim 2, characterized in that, The process of acquiring monitoring data collected by the industrial equipment monitoring terminal, performing data analysis on the monitoring data based on edge computing nodes, extracting corresponding local spatiotemporal features during the analysis, and generating corresponding standard monitoring data streams by combining cross-modal spatiotemporal alignment mechanisms also includes: Cross-modal spatiotemporal alignment is performed on local spatiotemporal features of different modalities. During the alignment process, the corresponding equipment processing content is determined, and timestamp dynamic normalization alignment and spatial projection mapping are performed to eliminate the spatiotemporal asynchrony and scale differences between monitoring data, thereby generating a standard monitoring data stream with a unified time reference and spatial coordinate system.

4. The data processing method for monitoring data based on big data analysis according to claim 1, characterized in that, The process involves inputting standard data streams into a big data analytics space, extracting spatiotemporal joint features through a spatiotemporal joint encoder, using industrial equipment as entity nodes, determining the temporal interaction similarity between nodes based on the spatiotemporal joint features, and constructing a continuously evolving industrial equipment knowledge graph by combining a time decay factor. This includes: The standard monitoring data stream is input in parallel to the big data analysis space in the cloud. In this big data analysis space, the encoder performs downsampling and feature aggregation along the time and space dimensions through stacked spatiotemporal graph convolutional networks, and extracts multi-scale spatiotemporal joint features, which include the processing details and processing trajectory trends of industrial equipment.

5. The data processing method for monitoring data based on big data analysis according to claim 4, characterized in that, The process of inputting standard data streams into a big data analysis space, extracting spatiotemporal joint features through a spatiotemporal joint encoder, determining the temporal interaction similarity between nodes based on spatiotemporal joint features using industrial equipment as entity nodes, and constructing a continuously evolving industrial equipment knowledge graph by combining a time decay factor, also includes: Industrial equipment identified in the monitoring footage is treated as entity nodes. Its corresponding spatiotemporal joint features are projected into a high-dimensional latent space. In the high-dimensional latent space, multiple factors are fused by combining the past abnormal information of the industrial equipment to determine the temporal interaction similarity between nodes. Furthermore, a time decay factor that decreases exponentially over time is introduced. The relationship between nodes and edges is continuously updated based on a dynamic graph reasoning mechanism, thereby constructing an industrial equipment knowledge graph whose topology evolves continuously over time.

6. The data processing method for monitoring data based on big data analysis according to claim 1, characterized in that, The target nodes of the labeled industrial equipment knowledge graph are feature-masked, and the distribution differences of the event evolution state transition network before and after intervention are determined by combining a topological intervention mechanism. The abnormal state of the industrial equipment is quantified by the distribution differences, and the corresponding anomaly prediction results are output, including: In the industrial equipment knowledge graph, target nodes with potential risk characteristics are marked. A self-supervised masking strategy is used to randomly mask the key behavioral characteristics of the target nodes and their neighboring nodes to output the masked industrial equipment knowledge graph. A topological intervention mechanism based on causal inference is combined to trigger virtual intervention in the masked industrial equipment knowledge graph.

7. The data processing method for monitoring data based on big data analysis according to claim 6, characterized in that, The target nodes of the labeled industrial equipment knowledge graph are feature-masked, and the distribution differences of the event evolution state transition network before and after intervention are determined by combining a topological intervention mechanism. The abnormal state of the industrial equipment is quantified by the distribution differences, and the corresponding anomaly prediction results are output. The method also includes: The event evolution state transition network before and after intervention is reconstructed using a variational graph autoencoder, and the distribution difference of the state transition probability distribution before and after intervention is determined. The distribution difference is mapped to an abnormal intent quantitative evaluation index, and the abnormal intent evaluation index is nonlinearly mapped by a preset intent classifier, thereby outputting the abnormal prediction results of industrial equipment in the future time window.

8. The data processing method for monitoring data based on big data analysis according to claim 1, characterized in that, The process of determining the corresponding early warning response command based on the anomaly prediction results, capturing the execution feedback data of the early warning response command, and jointly aligning the execution feedback data with the standard monitoring data stream to perform closed-loop updates of data processing events for industrial equipment includes: Based on the risk level and intent type of the anomaly prediction results, a preset emergency response strategy library is retrieved to automatically match and issue corresponding early warning response instructions to the execution terminal; the execution feedback data returned by the execution terminal after completing the early warning response instructions is captured in real time, including action execution logs and monitoring screens.

9. The data processing method for monitoring data based on big data analysis according to claim 8, characterized in that, The process of determining the corresponding early warning response command based on the identification of anomaly prediction results, capturing the execution feedback data of the early warning response command, and jointly aligning the execution feedback data with the standard monitoring data stream to perform closed-loop updates of data processing events for industrial equipment also includes: A cross-modal contrastive learning mechanism is adopted to jointly align and concatenate the execution feedback data with the standard monitoring data stream on the same time axis, thereby generating a closed-loop feedback feature vector. The closed-loop feedback feature vector is input into the reinforcement learning reward model to calculate the comprehensive reward value of the current data processing event. The comprehensive reward value is mechanically propagated back to the model, and adaptive closed-loop updates of the data processing events of industrial equipment are realized during the propagation process.

10. A data processing system for monitoring data based on big data analytics, characterized in that, The data processing system for monitoring data based on big data analysis is applied to the data processing method for monitoring data based on big data analysis as described in any one of claims 1-9; The data processing system for monitoring data based on big data analytics includes: The monitoring data stream module is used to acquire monitoring data collected by the monitoring terminal of industrial equipment, perform data analysis on the monitoring data based on edge computing nodes, extract corresponding local spatiotemporal features during the analysis process, and generate corresponding standard monitoring data streams by combining cross-modal spatiotemporal alignment mechanisms. The big data analysis module is used to input standard data streams into the big data analysis space, extract spatiotemporal joint features through a spatiotemporal joint encoder, use industrial equipment as entity nodes, determine the temporal interaction similarity between nodes based on spatiotemporal joint features, and construct a continuously evolving industrial equipment knowledge graph by combining time decay factors. The prediction module is used to label the target nodes of the industrial equipment knowledge graph, perform feature masking on the target nodes, and combine the topological intervention mechanism to determine the distribution difference of the event evolution state transition network before and after the intervention. The distribution difference is used to quantify the abnormal state of the industrial equipment, and then output the corresponding abnormal prediction results. The data processing event module is used to determine the corresponding early warning response command based on the identification of anomaly prediction results, capture the execution feedback data of the early warning response command, and jointly align the execution feedback data with the standard monitoring data stream to perform closed-loop updates of the data processing events of industrial equipment.

Images