A security monitoring data fusion system and method based on adaptive Kalman filtering

Abstract

This invention discloses a safety monitoring data fusion system and method based on adaptive Kalman filtering, belonging to the field of safety monitoring data fusion technology. The system includes: a data preprocessing module for handling outliers in safety monitoring data and aligning the cleaned safety monitoring data using a spatiotemporal alignment algorithm; a data fusion module for reconstructing the aligned safety monitoring data and performing adaptive weighted fusion processing on the reconstructed safety monitoring data based on an adaptive Kalman filtering algorithm; and a fusion optimization module for constructing a safety monitoring knowledge graph based on an equipment safety knowledge base, combining it with the preliminary fused data to perform quality assessment, and optimizing the fusion based on the assessment results to obtain fused safety monitoring data. This invention ensures the output of highly reliable, low-latency, and process-semantic-rich fused safety monitoring data, significantly reducing false alarms and false negatives, and enhancing the ability to detect equipment anomalies and process disturbances.

Description

Technical Field

[0001] This invention relates to the field of security monitoring data fusion technology, and more specifically, to a security monitoring data fusion system and method based on adaptive Kalman filtering. Background Technology

[0002] Safety monitoring data fusion refers to the process of generating highly consistent and reliable fused data from raw observation information from multiple heterogeneous sensors through key technologies such as spatiotemporal alignment, noise suppression, and redundancy elimination. This fusion is a crucial prerequisite for achieving accurate perception and intelligent decision-making. Especially in complex industrial environments, various sensors often suffer from fragmented, distorted, or even contradictory data due to differences in sampling frequencies, transmission delays, and susceptibility to electromagnetic interference or environmental disturbances. Therefore, providing a high-quality safety monitoring data fusion system is of paramount importance.

[0003] However, existing safety monitoring data fusion systems generally lack the ability to finely align multi-source heterogeneous data in time and space, which can easily misjudge mismatched signals in the process stage as abnormal states, resulting in a large number of false alarms. Moreover, traditional fusion often relies on fixed weights or simple statistical averaging, without combining equipment dynamics models and real-time noise characteristics for adaptive adjustment, which leads to a decrease in the accuracy of state estimation. In particular, it ignores the physical laws and engineering constraints in the process, and the output results lack interpretability, making them unreliable or difficult to use for root cause analysis.

[0004] No effective solutions have yet been proposed to address the problems in the relevant technologies. Summary of the Invention

[0005] To address the problems in related technologies, this invention proposes a security monitoring data fusion system and method based on adaptive Kalman filtering, in order to overcome the aforementioned technical problems existing in the existing related technologies.

[0006] Therefore, the specific technical solution adopted by the present invention is as follows:

[0007] According to one aspect of the present invention, a security monitoring data fusion system based on adaptive Kalman filtering is provided, comprising: a data preprocessing module, a data fusion module, and a fusion optimization module;

[0008] The data preprocessing module is used to acquire safety monitoring data, process outliers in the safety monitoring data to obtain cleaned safety monitoring data, and use a spatiotemporal alignment algorithm to align the cleaned safety monitoring data to obtain aligned safety monitoring data.

[0009] The data fusion module is used to perform reconstruction processing on the aligned security monitoring data to obtain reconstructed security monitoring data, and to perform adaptive weighted fusion processing on the reconstructed security monitoring data based on the adaptive Kalman filter algorithm to obtain preliminary fused data.

[0010] The fusion optimization module is used to build a safety monitoring knowledge graph based on the equipment safety knowledge base, and to combine the safety monitoring knowledge graph with the preliminary fusion data to carry out quality assessment. Based on the assessment results, the preliminary fusion data is fused and optimized to obtain the safety monitoring fused data.

[0011] Furthermore, the data preprocessing module includes: a data cleaning module, a data alignment module, and a data augmentation module;

[0012] The data cleaning module is used to acquire security monitoring data, identify outliers by comparing the security monitoring data with preset security statistical thresholds, and perform data correction and confidence calculation on the security monitoring data based on the outlier identification results to obtain cleaned security monitoring data.

[0013] The data alignment module is used to perform time-series alignment processing on the cleaned safety monitoring data based on the dynamic time warping algorithm to obtain time-series synchronized monitoring data, and to perform spatial semantic alignment processing on the time-series synchronized monitoring data to obtain spatiotemporally aligned monitoring data.

[0014] The data augmentation module is used to construct a slime mold excitation field based on the spatiotemporal alignment monitoring data, and to perform data augmentation processing on the spatiotemporal alignment monitoring data according to the slime mold excitation field to obtain aligned safety monitoring data.

[0015] Furthermore, the data alignment module includes: a time warp path acquisition module, a temporal synchronization module, a perceptual network construction module, and a spatial semantic alignment module;

[0016] The time warp path acquisition module is used to input the cleaned safety monitoring data into the dynamic time warping algorithm and calculate the nonlinear time warp path.

[0017] The timing synchronization module is used to map the cleaned safety monitoring data onto a unified reference time axis based on the time warp path and using the time axis resampling method to obtain timing synchronized monitoring data.

[0018] The sensing network construction module is used to construct the sensing network by taking time-series synchronous monitoring data as data node attributes and combining it with the pre-acquired physical connection relationship of monitoring devices to obtain a sensing topology map.

[0019] The spatial semantic alignment module is used to adaptively learn and perceive the spatial semantic dependencies between nodes in the topological graph by utilizing the multi-head attention mechanism of graph attention networks. Based on the spatial semantic dependencies, it dynamically aggregates the context information of neighboring nodes and optimizes the embedding representation of nodes to obtain spatiotemporal alignment monitoring data.

[0020] Furthermore, the data augmentation module includes: an excitation field construction module, an enhanced candidate trajectory generation module, and a weighted fusion module;

[0021] Among them, the excitation field construction module is used to obtain the node confidence of the spatiotemporal alignment monitoring data, use the node confidence as the nutrient source intensity, and construct the slime mold excitation field by combining the perception topology map of the spatiotemporal alignment monitoring data.

[0022] An enhanced candidate trajectory generation module is used to initialize a random slime mold distribution based on a slime mold excitation field, dynamically adjust the connection weights on the topological edges of the sensing topology graph according to the oscillation propagation mechanism, and generate enhanced candidate trajectories with consistent structure.

[0023] The weighted fusion module is used to perform weighted fusion of spatiotemporal aligned monitoring data and enhanced candidate trajectories according to connection weights and remove nodes that violate physical continuity constraints to obtain aligned safety monitoring data.

[0024] Furthermore, the data fusion module includes: an encoder construction module, a data reconstruction module, and a preliminary fusion module;

[0025] The encoder construction module is used to input the aligned safety monitoring data into the temporal autoencoder and embed domain prior knowledge into the loss function of the temporal autoencoder to construct a physical constraint encoder.

[0026] The data reconstruction module is used to introduce a smoothness regularization term during the encoding stage of the physical constraint encoder, generate a reconstruction-ready feature vector, and embed the reconstruction-ready feature vector and preset physical boundary conditions into the decoding process to perform constraint-guided data reconstruction, thereby obtaining the reconstructed safety monitoring data.

[0027] The preliminary fusion module is used to input the reconstructed safety monitoring data into the adaptive Kalman filter algorithm and combine it with the dynamic model of the monitoring equipment to estimate the noise covariance. Based on the noise covariance estimation results, the reconstructed safety monitoring data is weighted and fused to obtain the preliminary fused data.

[0028] Furthermore, the preliminary fusion module includes: a filter initialization module, a filter gain update module, and a recursive estimation module;

[0029] The filter initialization module is used to take the reconstructed safety monitoring data as the observation sequence of the adaptive Kalman filter algorithm, and construct the state space equation in combination with the dynamic model of the monitoring equipment. The state vector and covariance matrix are initialized according to the state space equation to obtain the initialized filter.

[0030] The filter gain update module is used to generate the observation residual sequence based on the initialized filter, calculate the process noise covariance and observation noise covariance based on the covariance matrix of the observation residual sequence, and dynamically update the filter gain using the covariance matching mechanism to obtain the updated filter gain.

[0031] The recursive estimation module is used to perform optimal recursive estimation based on the updated filter gain, and to perform adaptive weighted fusion of the reconstructed safety monitoring data based on the optimal recursive estimation result to obtain preliminary fused data.

[0032] Furthermore, the feature is that the fusion optimization module includes: a knowledge graph construction module, a quality assessment module, and a data fusion correction module;

[0033] Among them, the knowledge graph construction module is used to extract equipment security entities and semantic relationships based on the equipment security knowledge base, and to construct a security monitoring knowledge graph based on the equipment security entities and semantic relationships;

[0034] The quality assessment module is used to semantically align the preliminary fused data with the safety monitoring knowledge graph, quantify the multidimensional credibility of the data based on the semantic alignment results, and generate quality assessment results based on the multidimensional credibility and the equipment safety knowledge base.

[0035] The data fusion correction module is used to identify correction areas in the preliminary fused data based on the quality assessment results. The correction area identification results are combined with the safety monitoring knowledge graph to generate a knowledge-guided correction strategy. The preliminary fused data is then weighted and adjusted according to the correction strategy to obtain the safety monitoring fused data.

[0036] Furthermore, the quality assessment module is characterized by including: a semantic association alignment module, a multidimensional credibility quantification module, a comprehensive score calculation module, and a quality grading determination module;

[0037] The semantic association alignment module is used to identify the devices and status nodes corresponding to the monitoring parameters in the preliminary fusion data in the security monitoring knowledge graph through entity links, and to establish semantic associations with security rules in the security monitoring knowledge graph by combining relationship matching, so as to obtain semantic alignment results.

[0038] The multidimensional credibility metric module is used to perform multidimensional credibility metric on the semantic alignment results from the dimensions of physical rationality, logical consistency and temporal integrity, and obtain multidimensional credibility.

[0039] The comprehensive score calculation module is used to determine the initial dimension weights based on the multidimensional credibility and the priority of different risk scenarios in the device security knowledge base, and to calculate the comprehensive credibility score by using the dynamic weighted fusion method on the multidimensional credibility and the initial dimension weights.

[0040] The quality grading determination module is used to determine the quality grading based on the comprehensive credibility score and a preset credibility threshold, and obtain the quality assessment result.

[0041] Furthermore, the feature is that the semantic alignment results are subjected to multi-dimensional credibility quantification from the dimensions of physical rationality, logical consistency, and temporal integrity, including:

[0042] In the dimension of physical rationality, the deviation of the theoretical feasible region is calculated based on the physical constraints of the monitoring parameters associated in the semantic alignment results, and the deviation is mapped to the standard interval to obtain the credibility of the physical rationality dimension.

[0043] In terms of logical consistency, domain logical consistency is detected based on the fault causal chain and security rules in the semantic alignment results, and a logical consistency dimension credibility is generated.

[0044] In the temporal integrity dimension, the continuity, smoothness, and sampling regularity of the data on the time axis are analyzed based on the temporal context in the semantic alignment results. The reliability of the data temporal structure is evaluated based on the degree analysis results, and the credibility of the temporal integrity dimension is obtained.

[0045] According to another aspect of the present invention, a security monitoring data fusion method based on adaptive Kalman filtering is provided, the method comprising the following steps:

[0046] S1. Obtain safety monitoring data and perform outlier processing on the safety monitoring data to obtain cleaned safety monitoring data. Then, use a spatiotemporal alignment algorithm to align the cleaned safety monitoring data to obtain aligned safety monitoring data.

[0047] S2. Perform reconstruction processing on the aligned security monitoring data to obtain reconstructed security monitoring data, and perform adaptive weighted fusion processing on the reconstructed security monitoring data based on the adaptive Kalman filter algorithm to obtain preliminary fused data.

[0048] S3. Construct a safety monitoring knowledge graph based on the equipment safety knowledge base, and combine the safety monitoring knowledge graph with the preliminary fusion data to conduct a quality assessment. Based on the assessment results, optimize the preliminary fusion data to obtain the safety monitoring fusion data.

[0049] The beneficial effects of this invention are as follows:

[0050] 1. This invention, through the synergistic effect of the data preprocessing module, data fusion module, and fusion optimization module, can accurately capture the correspondence between various sensor signals in the process stage, eliminate false correlations caused by time misalignment or black-box misjudgments caused by pure data-driven approaches; ensure the output of highly reliable, low-latency, and process-semantic-rich safety monitoring fusion data, significantly reduce false alarm and false negative rates, and enhance the ability to perceive equipment anomalies and process disturbances.

[0051] 2. This invention, through dynamic time warping algorithm and graph attention network, can effectively address the nonlinear asynchronous problem in the time dimension and the complex physical coupling relationship in the spatial dimension of multi-source security monitoring data. It not only ensures strict synchronization of multi-source data in time, but also explicitly models the physical associations across devices in space, thereby outputting spatiotemporally aligned embeddings rich in global semantic information. This significantly improves the accuracy and interpretability of subsequent data reconstruction, fusion and quality assessment, laying a solid data foundation for high-reliability security monitoring in advanced processes.

[0052] 3. This invention constructs and assesses the quality of knowledge graphs, structuring equipment failure modes, process constraints, and causal rules into a semantic network. This network is then used to perform multi-dimensional credible quantification with the initially fused data, enabling refined and interpretable quality assessment of the monitoring data. Based on this, the system can automatically identify logical conflicts or physical anomalies and guide correction strategies based on the knowledge graph, significantly improving the accuracy, consistency, and confidence of the final fused data. This effectively reduces false alarms and missed alarms, providing highly reliable and semantically rich decision support for predictive maintenance and process optimization. Attached Figure Description

[0053] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0054] Figure 1 This is a structural block diagram of a security monitoring data fusion system based on adaptive Kalman filtering according to an embodiment of the present invention;

[0055] Figure 2 This is a flowchart of a security monitoring data fusion method based on adaptive Kalman filtering according to an embodiment of the present invention.

[0056] In the picture:

[0057] 1. Data preprocessing module; 2. Data fusion module; 3. Fusion optimization module. Detailed Implementation

[0058] To further illustrate the various embodiments, the present invention provides accompanying drawings, which are part of the disclosure of the present invention. These drawings are mainly used to illustrate the embodiments and can be used in conjunction with the relevant descriptions in the specification to explain the operating principles of the embodiments. With reference to these drawings, those skilled in the art should be able to understand other possible implementation methods and the advantages of the present invention.

[0059] According to an embodiment of the present invention, a security monitoring data fusion system and method based on adaptive Kalman filtering is provided.

[0060] The present invention will now be further described in conjunction with the accompanying drawings and specific embodiments, such as... Figure 1 As shown, according to an embodiment of the present invention, a security monitoring data fusion system based on adaptive Kalman filtering is provided, including: a data preprocessing module 1, a data fusion module 2, and a fusion optimization module 3;

[0061] Among them, the data preprocessing module 1 is used to acquire safety monitoring data, process outliers in the safety monitoring data to obtain cleaned safety monitoring data, and use a spatiotemporal alignment algorithm to perform alignment processing on the cleaned safety monitoring data to obtain aligned safety monitoring data.

[0062] Specifically, data preprocessing module 1 includes: a data cleaning module, a data alignment module, and a data enhancement module;

[0063] The data cleaning module is used to acquire security monitoring data, identify outliers by comparing the security monitoring data with preset security statistical thresholds, and perform data correction and confidence calculation on the security monitoring data based on the outlier identification results to obtain cleaned security monitoring data.

[0064] The data alignment module is used to perform time-series alignment processing on the cleaned safety monitoring data based on the dynamic time warping algorithm to obtain time-series synchronized monitoring data, and to perform spatial semantic alignment processing on the time-series synchronized monitoring data to obtain spatiotemporally aligned monitoring data.

[0065] Specifically, the data alignment module includes: a time warp path acquisition module, a time sequence synchronization module, a perceptual network construction module, and a spatial semantic alignment module;

[0066] The time warp path acquisition module is used to input the cleaned safety monitoring data into the dynamic time warping algorithm and calculate the nonlinear time warp path.

[0067] The timing synchronization module is used to map the cleaned safety monitoring data onto a unified reference time axis based on the time warp path and using the time axis resampling method to obtain timing synchronized monitoring data.

[0068] The sensing network construction module is used to construct the sensing network by taking time-series synchronous monitoring data as data node attributes and combining it with the pre-acquired physical connection relationship of monitoring devices to obtain a sensing topology map.

[0069] The spatial semantic alignment module is used to adaptively learn and perceive the spatial semantic dependencies between nodes in the topological graph by utilizing the multi-head attention mechanism of graph attention networks. Based on the spatial semantic dependencies, it dynamically aggregates the context information of neighboring nodes and optimizes the embedding representation of nodes to obtain spatiotemporal alignment monitoring data.

[0070] The data augmentation module is used to construct a slime mold excitation field based on the spatiotemporal alignment monitoring data, and to perform data augmentation processing on the spatiotemporal alignment monitoring data according to the slime mold excitation field to obtain aligned safety monitoring data.

[0071] Specifically, the data augmentation module includes: an excitation field construction module, an enhanced candidate trajectory generation module, and a weighted fusion module;

[0072] Among them, the excitation field construction module is used to obtain the node confidence of the spatiotemporal alignment monitoring data, use the node confidence as the nutrient source intensity, and construct the slime mold excitation field by combining the perception topology map of the spatiotemporal alignment monitoring data.

[0073] An enhanced candidate trajectory generation module is used to initialize a random slime mold distribution based on a slime mold excitation field, dynamically adjust the connection weights on the topological edges of the sensing topology graph according to the oscillation propagation mechanism, and generate enhanced candidate trajectories with consistent structure.

[0074] The weighted fusion module is used to perform weighted fusion of spatiotemporal aligned monitoring data and enhanced candidate trajectories according to connection weights and remove nodes that violate physical continuity constraints to obtain aligned safety monitoring data.

[0075] Specifically, in integrated circuit manufacturing and high-reliability operation scenarios, the safety monitoring data fusion system can achieve reliable end-to-end processing of data generated by high-density heterogeneous sensor networks in key wafer fab equipment, such as lithography machines, etching machines, CVD / PVD chambers, ion implanters, probe stations, and cleanroom environments, through collaborative sensing of chip-level physical security, process stability, and equipment health status. Safety monitoring data inputs encompass multi-dimensional high-frequency time-series signals, such as temperature (including chamber wall temperature and wafer back temperature), with a sampling frequency of 10-100Hz and measurement accuracy up to ±0.1°C; voltage or current signals, such as RF matching converter output voltage and electrostatic chuck (ESC) bias voltage, with a sampling frequency of 1-10kHz; and pressure and gas flow parameters, i.e., process chamber vacuum and Ar, etc. The system monitors the flow rate of process gases, with response delays typically between 50-200 ms; vibration and acoustic emission signals, such as the vibration acceleration of robotic arm motors or the spectrum of pump noise, with sampling frequencies no lower than 10 kHz; particulate matter concentration data, i.e., real-time counting of airborne molecular pollutants (AMC), with detectable particle sizes ranging from 0.1 to 5 μm; and various process log information, such as key process indicators like exposure dose, etching rate, and film thickness uniformity. The data preprocessing module 1 employs an anomaly detection mechanism. It first performs preliminary screening based on the 3σ principle or interquartile range (IQR) rule, then combines this with a sliding window filtering method to model the residual sequence, dynamically identifying abrupt changes or drift phenomena in the data. For detected anomalies, such as a sudden temperature rise from 350°C to 600°C in a CVD chamber thermocouple due to plasma interference, the system utilizes the temporal and spatial correlation of neighboring sensors, employs Gaussian process regression (GPR) for interpolation correction, and assigns a corresponding confidence label, such as 0.4, to the corrected data points, obtaining the safety monitoring data after cleaning. The data alignment module addresses multi-source asynchronous issues. Using the timestamp of the main control PLC or the processing cycle of a reference wafer as a benchmark, it calculates nonlinear time warping paths using a Dynamic Time Warping (DTW) algorithm. For example, in the resist removal process, a certain etching cavity may have a 120-millisecond delay in heating compared to the main cavity. A cubic spline resampling method is employed to uniformly map all sensor data to a common time grid with a 1-millisecond granularity. Simultaneously, based on the physical topology of the equipment—material flow paths, gas path connections, and heat conduction channels—a perception topology map is constructed, and a lightweight graph attention network (GAT) is deployed to learn implicit spatial dependencies between nodes. For instance, the temperature changes of two process cavities that, while not directly connected, share a cooling loop, exhibit a strong correlation, corresponding to an attention weight of up to 0.87. The output is a spatiotemporally aligned embedding representation that fuses local measurements and global semantic information, yielding spatiotemporally aligned monitoring data. For low-confidence regions, such as beam saturation due to high-voltage arcing during ion implantation, the data confidence level may be only 0.2. To address this issue, the data augmentation module constructs a so-called slime mold excitation field. This field uses the confidence level of nodes as the nutrient source intensity to guide the slime mold algorithm (SMA) to oscillate and propagate on the physical topology of the device, thereby dynamically strengthening physically meaningful coupling paths, such as the relationship between beam current, bias voltage, and vacuum degree. This method generates augmentation candidate trajectories that conform to process rules, such as smoothing the beam recovery process. Based on dynamic weights, the original spatiotemporally aligned data and the augmented trajectories are convexly combined and fused. A physical continuity constraint checker is introduced, checking for things like exceeding the second derivative of temperature limits and whether the current satisfies KCL's law. Nodes violating fundamental physical laws are eliminated, resulting in highly complete and consistent aligned safety monitoring data.For example, limiting the beam current change rate to no more than 5 mA / ms and outputting highly complete data allows the data confidence level to be increased to 0.85. This not only improves the reliability of the data but also ensures the physical rationality of the data evolution process.

[0076] Data fusion module 2 is used to perform reconstruction processing on the aligned security monitoring data to obtain reconstructed security monitoring data, and to perform adaptive weighted fusion processing on the reconstructed security monitoring data based on the adaptive Kalman filter algorithm to obtain preliminary fused data.

[0077] Specifically, data fusion module 2 includes: encoder construction module, data reconstruction module, and preliminary fusion module;

[0078] The encoder construction module is used to input the aligned safety monitoring data into the temporal autoencoder and embed domain prior knowledge into the loss function of the temporal autoencoder to construct a physical constraint encoder.

[0079] The data reconstruction module is used to introduce a smoothness regularization term during the encoding stage of the physical constraint encoder, generate a reconstruction-ready feature vector, and embed the reconstruction-ready feature vector and preset physical boundary conditions into the decoding process to perform constraint-guided data reconstruction, thereby obtaining the reconstructed safety monitoring data.

[0080] The preliminary fusion module is used to input the reconstructed safety monitoring data into the adaptive Kalman filter algorithm and combine it with the dynamic model of the monitoring equipment to estimate the noise covariance. Based on the noise covariance estimation results, the reconstructed safety monitoring data is weighted and fused to obtain the preliminary fused data.

[0081] Specifically, the preliminary fusion module includes: a filter initialization module, a filter gain update module, and a recursive estimation module;

[0082] The filter initialization module is used to take the reconstructed safety monitoring data as the observation sequence of the adaptive Kalman filter algorithm, and construct the state space equation in combination with the dynamic model of the monitoring equipment. The state vector and covariance matrix are initialized according to the state space equation to obtain the initialized filter.

[0083] The filter gain update module is used to generate the observation residual sequence based on the initialized filter, calculate the process noise covariance and observation noise covariance based on the covariance matrix of the observation residual sequence, and dynamically update the filter gain using the covariance matching mechanism to obtain the updated filter gain.

[0084] The recursive estimation module is used to perform optimal recursive estimation based on the updated filter gain, and to perform adaptive weighted fusion of the reconstructed safety monitoring data based on the optimal recursive estimation result to obtain preliminary fused data.

[0085] Specifically, the aligned safety monitoring data input to data fusion module 2 originates from the high-density sensor network of critical equipment in the wafer fab. For example, in a multi-cavity plasma etching machine used for advanced logic chip production, the collected safety monitoring data includes: the inner wall temperature of cavities one through six, with a sampling rate of 50Hz; the output power of the RF generator, with a sampling rate of 1kHz; the bias voltage, with a sampling rate of 2kHz; the inlet and outlet temperatures of the cooling water, with a measurement accuracy of ±0.1°C; and the process gases such as... and The data includes the flow rate and the vibration acceleration of the robotic arm in the X, Y, and Z directions, for example, in m / s², with a sampling rate of 10 kHz. This data has been spatiotemporally aligned in the preceding module and accompanied by confidence level labels. For example, at a certain moment, the wall temperature reading of cavity 3 is 382.5°C with a confidence level of 0.63, while the confidence level of cavity 5 is 0.94 due to good sensor calibration.

[0086] The encoder building module inputs the aligned safety monitoring data into the timing autoencoder and embeds domain prior knowledge, namely the physical laws of semiconductor processes, into the loss function. For example, the RF power input should be equal to the sum of cavity heat loss and plasma energy consumption, or the cooling water temperature rise should not exceed 10°C. This results in the construction of a physically constrained encoder, ensuring that the model learning process strictly follows the basic laws of thermodynamics and electrodynamics. The data reconstruction module introduces smoothness regularization terms, such as total variation penalty, during the encoding stage to suppress temperature signal jitter caused by high-frequency electromagnetic interference, generating a reconstruction-ready feature vector. During the decoding stage, the reconstruction-ready feature vector is co-embedded with preset physical boundary conditions to constrain and guide data reconstruction. For example, it forces that the cavity wall temperature at any given time must not exceed the material's tolerance limit of 450°C, and that the temperature change rate between adjacent milliseconds must not exceed 50°C / s. This outputs reconstructed safety monitoring data that is both smooth and strictly conforms to the physically feasible domain. At this point, the original abrupt changes in cavity three, such as a sudden increase from 370°C to 410°C, are corrected to a smooth curve conforming to thermal inertia, and the confidence level is simultaneously increased to 0.82. Based on this, the preliminary fusion module performs adaptive weighted fusion on the reconstructed safety monitoring data. Taking the temperature sequence of the six cavities as an example, the filter initialization module uses the temperature estimate every millisecond as the observation sequence of the adaptive Kalman filter algorithm, and combines it with the dynamic model of the monitoring equipment, namely the first-order heat conduction state equation based on the heat capacity of the aluminum material of the cavity of 120J / K and the thermal resistance of 0.05K / W, to construct a state space model; the state vector is initialized as a two-dimensional vector composed of temperature and its rate of change, and the covariance matrix is ​​initially set as a diagonal matrix diag(1,0.1), thus obtaining the initialized filter. The filter gain update module performs prediction and update loops based on the filter to generate the observation residual sequence. It calculates the residual covariance matrix through a sliding window and estimates the process noise covariance and observation noise covariance online based on the covariance matching mechanism. For example, the mean residual of cavity 2 is +3.2°C and the standard deviation is 2.1 within 100ms, indicating that it has a systematic shift. Its observation noise covariance increases from 0.04 to 0.36, while that of cavity 5 remains at 0.02. The system dynamically adjusts the weights of each channel accordingly and dynamically updates the filter gain using the covariance matching mechanism to obtain the updated filter gain, so that the high-confidence channel dominates the fusion result. The recursive estimation module performs optimal recursive estimation based on the updated filter gain. At each time step, it fuses the system's dynamic prediction and multi-source observations, and outputs the optimal state estimate in the sense of minimum mean square error. For example, at time t = 1523ms, the original temperatures of the six chambers are 378.1°C, 381.5°C, 379.0°C, 377.8°C, 378.3°C, and 378.6°C, respectively. After adaptive weighted fusion, the preliminary fused data output is 378.25°C, with a standard deviation of only ±0.07°C.This process is also applicable to other critical scenarios. In EUV lithography machines, the reconstructed data from three sets of infrared thermography points—the mirror center, edge, and support frame—are fused using adaptive Kalman filtering. The updated filter gain effectively suppresses edge channel drift caused by carbon buildup on the mirror surface. In ion implanters, the reconstructed sequences of beam current, bias voltage, and vacuum level are jointly modeled to achieve high-precision fusion monitoring of beam stability. The entire data fusion module 2 is deployed on an edge AI server, utilizing parallel computing to accelerate temporal autoencoder inference and Kalman recursion. The output preliminary fusion data provides highly reliable and consistent core input for subsequent knowledge graph-based quality assessment and optimization, effectively ensuring the yield, equipment lifespan, and production safety of advanced processes.

[0087] The fusion optimization module 3 is used to construct a safety monitoring knowledge graph based on the equipment safety knowledge base, and to combine the safety monitoring knowledge graph with the preliminary fusion data to conduct a quality assessment. Based on the assessment results, the preliminary fusion data is fused and optimized to obtain the safety monitoring fused data.

[0088] Specifically, the fusion optimization module 3 includes: a knowledge graph construction module, a quality assessment module, and a data fusion correction module;

[0089] Among them, the knowledge graph construction module is used to extract equipment security entities and semantic relationships based on the equipment security knowledge base, and to construct a security monitoring knowledge graph based on the equipment security entities and semantic relationships;

[0090] The quality assessment module is used to semantically align the preliminary fused data with the safety monitoring knowledge graph, quantify the multidimensional credibility of the data based on the semantic alignment results, and generate quality assessment results based on the multidimensional credibility and the equipment safety knowledge base.

[0091] Specifically, the quality assessment module includes: a semantic association alignment module, a multi-dimensional credibility quantification module, a comprehensive score calculation module, and a quality grading determination module;

[0092] The semantic association alignment module is used to identify the devices and status nodes corresponding to the monitoring parameters in the preliminary fusion data in the security monitoring knowledge graph through entity links, and to establish semantic associations with security rules in the security monitoring knowledge graph by combining relationship matching, so as to obtain semantic alignment results.

[0093] The multidimensional trust metric module is used to perform multidimensional trust metric on the semantic alignment results from the dimensions of physical rationality, logical consistency and temporal integrity, so as to obtain multidimensional trustworthiness.

[0094] Specifically, the semantic alignment results are subjected to multi-dimensional credibility metric measurement from the dimensions of physical rationality, logical consistency, and temporal integrity, including:

[0095] In the dimension of physical rationality, the deviation of the theoretical feasible region is calculated based on the physical constraints of the monitoring parameters associated in the semantic alignment results, and the deviation is mapped to the standard interval to obtain the credibility of the physical rationality dimension.

[0096] In terms of logical consistency, domain logical consistency is detected based on the fault causal chain and security rules in the semantic alignment results, and a logical consistency dimension credibility is generated.

[0097] In the temporal integrity dimension, the continuity, smoothness, and sampling regularity of the data on the time axis are analyzed based on the temporal context in the semantic alignment results. The reliability of the data temporal structure is evaluated based on the degree analysis results, and the credibility of the temporal integrity dimension is obtained.

[0098] The comprehensive score calculation module is used to determine the initial dimension weights based on the multidimensional credibility and the priority of different risk scenarios in the device security knowledge base, and to calculate the comprehensive credibility score by using the dynamic weighted fusion method on the multidimensional credibility and the initial dimension weights.

[0099] The quality grading determination module is used to determine the quality grading based on the comprehensive credibility score and a preset credibility threshold, and obtain the quality assessment result.

[0100] The data fusion correction module is used to identify correction areas in the preliminary fused data based on the quality assessment results. The correction area identification results are combined with the safety monitoring knowledge graph to generate a knowledge-guided correction strategy. The preliminary fused data is then weighted and adjusted according to the correction strategy to obtain the safety monitoring fused data.

[0101] Specifically, the fusion optimization module 3 undertakes the crucial task of elevating the initial fused data into domain-semantic, interpretable, and highly reliable safety monitoring fused data. Through a three-tiered architecture of knowledge graph construction, quality assessment, and data fusion correction, it achieves a closed-loop process from data to knowledge to optimization decisions. The knowledge graph construction module extracts equipment safety entities and semantic relationships based on an equipment safety knowledge base. For example, it extracts equipment entities such as EUV mirrors, etching cavities, and cooling circuits from SEMI E10 standards, manufacturer maintenance manuals, historical fault reports, and process specifications, as well as causal rules or constraints such as temperature exceeding limits, thermal deformation, overlay errors, cooling water flow reduction, cavity temperature rise, and plasma extinction. Based on this, it constructs a structured safety monitoring knowledge graph, where nodes represent physical equipment or state variables, and edges represent physical dependencies, control logic, or fault propagation paths. The quality assessment module performs multi-dimensional reliability quantification and hierarchical judgment on the initial fused data. Internally, it includes a semantic association alignment module, a multi-dimensional reliability quantification module, a comprehensive score calculation module, and a quality grading judgment module. The semantic association alignment module identifies the corresponding equipment and status nodes in the safety monitoring knowledge graph of the preliminary fused data using entity linking technology. For example, it maps the temperature of cavity three (382.5°C) to the cavity wall temperature attribute node of plasma cavity three in the graph, and establishes a semantic association with safety rules by combining relation matching. That is, if the temperature exceeds 450°C, the thermal stress risk rule is triggered, thus obtaining the semantic alignment result. The multidimensional trust quantification module quantifies and evaluates the semantic alignment result from three dimensions: physical rationality, logical consistency, and temporal integrity. In the dimension of physical rationality, the system calculates the deviation of the theoretical feasible region based on the physical constraints of the monitoring parameters associated in the semantic alignment results. For example, if the upper limit of the EUV mirror temperature is 55°C, and the fused value is 58°C, then the deviation is (58−55) / 55≈0.055. This is then mapped to the [0,1] interval using the Sigmoid function, resulting in a physical rationality confidence score of 0.72. In the dimension of logical consistency, the system performs domain logical consistency detection based on the fault causal chain and safety rules in the semantic alignment results. For example, when the cavity... When the temperature reaches 460°C but no downstream effects such as plasma extinction or RF power drop are observed, a logical conflict is determined, and the reliability of the logical consistency dimension is 0.65. In terms of temporal integrity, the system analyzes the continuity, smoothness, and sampling regularity of the data on the time axis. For example, if a beam data segment experiences a non-physical jump from 10mA to 0 and then rises back within 10ms, and its second-order difference exceeds the equipment's allowable change rate of 5mA / ms, then the temporal structure is determined to be unreliable, and the reliability of the temporal integrity dimension is 0.58.The comprehensive score calculation module determines the initial dimension weights based on the multi-dimensional credibility and the priority of different risk scenarios in the equipment safety knowledge base. For example, in the lithography process, the weight of physical rationality is set to 0.5, logical consistency to 0.3, and timing integrity to 0.2; while in the etching process, due to the emphasis on process stability, the weight of timing integrity is increased to 0.4. The system uses a dynamic weighted fusion method to calculate the comprehensive credibility score based on the multi-dimensional credibility and the initial dimension weights. For example, physical rationality 0.72×0.5 + logical consistency 0.65×0.3 + timing integrity 0.58×0.2 = 0.673, resulting in a comprehensive credibility score of 0.673. The quality grading judgment module performs quality grading judgment based on the comprehensive credibility score and a preset credibility threshold. For example, it sets high credibility ≥ 0.8, medium credibility 0.6–0.8, and low credibility < 0.6, thus outputting a quality assessment result of medium credibility, indicating a potential logical conflict. Based on this assessment, the data fusion correction module initiates an optimization process, identifying correction areas in the preliminary fused data. For example, it locates inconsistencies between the cavity temperature and logical rules within the t=1523ms to 1550ms interval. Subsequently, the correction area identification results are combined with the safety monitoring knowledge graph to generate a knowledge-guided correction strategy. Specifically, if high temperature fails to trigger plasma extinguishing and is deemed abnormal, the system reverts to redundant infrared sensor data or interpolates based on historical similar operating conditions with the same gas ratio, RF power, and cooling flow rate. The preliminary fused data is then weighted and adjusted according to the correction strategy. For instance, the original fused value of 378.25°C is replaced with 376.8°C based on historical operating condition regression, and the confidence level is updated to 0.89, resulting in the safety monitoring fused data. This entire process manifests as a precise response to specific process events on the actual production line. For example, during an EUV exposure, preliminary fusion data showed a mirror temperature of 56.3°C, with a physical plausibility confidence score of only 0.68. Semantic alignment revealed a violation of the rule that mirror temperature ≤ 55°C. Logical consistency checks showed no significant increase in concurrent overlay error, indicating a contradiction. Time-series analysis revealed a 3°C temperature jump within 2ms, exceeding the thermal inertia limit. A comprehensive score of 0.65 triggered a moderately reliable alarm. The system then invoked the mirror carbonization fault mode from the knowledge graph, using an uncontaminated support frame temperature measurement point as the primary reference. Combined with thermal conduction model interpolation to correct the center temperature, the optimized safety monitoring fusion data was output as 54.1°C, with a comprehensive confidence score improved to 0.91. Similar mechanisms are also used in scenarios such as ion implantation beam recovery and CVD film thickness uniformity correction. The fusion optimization module 3, deployed on an edge AI platform, works in pipeline with data preprocessing module 1 and data fusion module 2, significantly reducing false alarm rates and improving early warning accuracy, providing a highly semantic and reliable data foundation for digital twins, predictive maintenance, and real-time process control.

[0102] According to another embodiment of the invention, such as Figure 2As shown, a security monitoring data fusion method based on adaptive Kalman filtering is provided, which includes the following steps:

[0103] S1. Obtain safety monitoring data and perform outlier processing on the safety monitoring data to obtain cleaned safety monitoring data. Then, use a spatiotemporal alignment algorithm to align the cleaned safety monitoring data to obtain aligned safety monitoring data.

[0104] S2. Perform reconstruction processing on the aligned security monitoring data to obtain reconstructed security monitoring data, and perform adaptive weighted fusion processing on the reconstructed security monitoring data based on the adaptive Kalman filter algorithm to obtain preliminary fused data.

[0105] S3. Construct a safety monitoring knowledge graph based on the equipment safety knowledge base, and combine the safety monitoring knowledge graph with the preliminary fusion data to conduct a quality assessment. Based on the assessment results, optimize the preliminary fusion data to obtain the safety monitoring fusion data.

[0106] In summary, this invention designs a dual-engine fusion framework driven by physical mechanism data, overcoming the challenge of inconsistent spatiotemporal benchmarks for multimodal data. Simultaneously, based on a spatiotemporal alignment algorithm using Dynamic Time Warping (DTW) and Graph Attention Network (GAT), it achieves efficient correlation of multi-source data such as vibration, infrared, and gas concentration. A domain knowledge graph is constructed, embedding physical field coupling rules and device operating principles. Through dynamic allocation of fusion weights and spatiotemporal alignment of multi-source heterogeneous data with knowledge graph-driven fusion, redundant information interference is eliminated, resulting in a lower false alarm rate for anomaly detection, improved data utilization, and effectively solving the temporal asynchrony problem caused by differences in sampling frequency, transmission delay, and device response characteristics among multi-source data.

[0107] In summary, by utilizing the above-mentioned technical solutions of this invention, the present invention, through the synergistic effect of the data preprocessing module, data fusion module, and fusion optimization module, can accurately capture the correspondence between various sensor signals in the process stage, eliminating false correlations caused by time misalignment or black-box misjudgments generated by pure data-driven approaches; ensuring the output of highly reliable, low-latency, and process-semantic-rich safety monitoring fusion data, significantly reducing false alarm and false negative rates, and enhancing the perception of equipment anomalies and process disturbances; the present invention, through dynamic time warping algorithm and graph attention network, can effectively address the nonlinear asynchronous problem in the time dimension and the complex physical coupling relationship in the spatial dimension of multi-source safety monitoring data, ensuring not only strict temporal synchronization of multi-source data, but also explicit spatial modeling of cross-equipment data. Physical associations, resulting in spatiotemporally aligned embeddings rich in global semantic information, significantly improve the accuracy and interpretability of subsequent data reconstruction, fusion, and quality assessment, laying a solid data foundation for high-reliability safety monitoring in advanced processes. This invention, through knowledge graph construction and quality assessment, structures equipment failure modes, process constraints, and causal rules into a semantic network, and performs multi-dimensional reliable quantification with the initially fused data, achieving refined and interpretable quality assessment of monitoring data. Based on this, the system can automatically identify logical conflicts or physical anomalies and guide correction strategies based on the knowledge graph, significantly improving the accuracy, consistency, and confidence of the final fused data, effectively reducing false alarms and false negatives, and providing highly reliable and semantically rich decision support for predictive maintenance and process optimization.

[0108] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A security monitoring data fusion system based on adaptive Kalman filtering, characterized in that, include: Data preprocessing module, data fusion module, and fusion optimization module; The data preprocessing module is used to acquire security monitoring data, process outliers in the security monitoring data to obtain cleaned security monitoring data, and use a spatiotemporal alignment algorithm to align the cleaned security monitoring data to obtain aligned security monitoring data. The data fusion module is used to perform reconstruction processing on the aligned security monitoring data to obtain reconstructed security monitoring data, and to perform adaptive weighted fusion processing on the reconstructed security monitoring data based on the adaptive Kalman filter algorithm to obtain preliminary fused data. The fusion optimization module is used to construct a safety monitoring knowledge graph based on the equipment safety knowledge base, combine the safety monitoring knowledge graph with the preliminary fusion data to conduct a quality assessment, and optimize the preliminary fusion data according to the assessment results to obtain safety monitoring fusion data. The data preprocessing module includes: a data cleaning module, a data alignment module, and a data enhancement module; The data cleaning module is used to acquire security monitoring data, identify outliers by comparing the security monitoring data with preset security statistical thresholds, and perform data correction and confidence calculation on the security monitoring data based on the outlier identification results to obtain cleaned security monitoring data. The data alignment module is used to perform time-series alignment processing on the cleaned safety monitoring data based on the dynamic time warping algorithm to obtain time-series synchronized monitoring data, and to perform spatial semantic alignment processing on the time-series synchronized monitoring data to obtain spatiotemporally aligned monitoring data. The data augmentation module is used to construct a slime mold excitation field based on the spatiotemporal alignment monitoring data, and to perform data augmentation processing on the spatiotemporal alignment monitoring data according to the slime mold excitation field to obtain aligned safety monitoring data. The data alignment module includes: a time warp path acquisition module, a time synchronization module, a perceptual network construction module, and a spatial semantic alignment module; The time warp path acquisition module is used to input the cleaned safety monitoring data into the dynamic time warping algorithm and calculate the nonlinear time warp path. The time synchronization module is used to map the cleaned safety monitoring data onto a unified reference time axis based on the time warp path and using the time axis resampling method to obtain time synchronization monitoring data. The sensing network construction module is used to construct a sensing network by taking time-series synchronous monitoring data as data node attributes and combining it with the pre-acquired physical connection relationship of monitoring devices to obtain a sensing topology map. The spatial semantic alignment module is used to adaptively learn and perceive the spatial semantic dependencies between nodes in the topological graph using the multi-head attention mechanism of the graph attention network, dynamically aggregate the context information of neighboring nodes according to the spatial semantic dependencies and optimize the embedding representation of the nodes to obtain spatiotemporal alignment monitoring data. The data augmentation module includes: an excitation field construction module, an enhanced candidate trajectory generation module, and a weighted fusion module; The excitation field construction module is used to acquire the node confidence of the spatiotemporal alignment monitoring data, use the node confidence as the nutrient source intensity, and construct the slime mold excitation field by combining the perception topology map of the spatiotemporal alignment monitoring data. The enhanced candidate trajectory generation module is used to initialize a random slime mold distribution based on a slime mold excitation field, dynamically adjust the connection weights on the topological edges of the sensing topology graph according to the oscillation propagation mechanism, and generate enhanced candidate trajectories with consistent structure. The weighted fusion module is used to perform weighted fusion of spatiotemporal aligned monitoring data and enhanced candidate trajectories according to connection weights and remove nodes that violate physical continuity constraints to obtain aligned safety monitoring data.

2. The security monitoring data fusion system based on adaptive Kalman filtering according to claim 1, characterized in that, The data fusion module includes: an encoder construction module, a data reconstruction module, and a preliminary fusion module; The encoder construction module is used to input the aligned safety monitoring data into the temporal autoencoder and embed domain prior knowledge into the loss function of the temporal autoencoder to construct a physical constraint encoder. The data reconstruction module is used to introduce a smoothness regularization term during the encoding stage of the physical constraint encoder, generate a reconstruction-ready feature vector, and embed the reconstruction-ready feature vector and preset physical boundary conditions into the decoding process to perform constraint-guided data reconstruction, thereby obtaining reconstructed safety monitoring data. The preliminary fusion module is used to input the reconstructed safety monitoring data into the adaptive Kalman filter algorithm, and combine it with the dynamic model of the monitoring equipment to estimate the noise covariance. Based on the noise covariance estimation result, the reconstructed safety monitoring data is weighted and fused to obtain preliminary fused data.

3. The safety monitoring data fusion system based on adaptive Kalman filtering according to claim 2, characterized in that, The preliminary fusion module includes: a filter initialization module, a filter gain update module, and a recursive estimation module; The filter initialization module is used to take the reconstructed safety monitoring data as the observation sequence of the adaptive Kalman filter algorithm, and construct the state space equation in combination with the dynamic model of the monitoring equipment. The state vector and covariance matrix are initialized according to the state space equation to obtain the initialized filter. The filter gain update module is used to generate an observation residual sequence based on the initialized filter, calculate the process noise covariance and observation noise covariance based on the covariance matrix of the observation residual sequence, and dynamically update the filter gain using a covariance matching mechanism to obtain the updated filter gain. The recursive estimation module is used to perform optimal recursive estimation based on the updated filter gain, and to perform adaptive weighted fusion of the reconstructed safety monitoring data based on the optimal recursive estimation result to obtain preliminary fused data.

4. The security monitoring data fusion system based on adaptive Kalman filtering according to claim 1, characterized in that, The fusion optimization module includes: a knowledge graph construction module, a quality assessment module, and a data fusion correction module; The knowledge graph construction module is used to extract device security entities and semantic relationships based on the device security knowledge base, and to construct a security monitoring knowledge graph based on the device security entities and semantic relationships. The quality assessment module is used to semantically align the preliminary fused data with the safety monitoring knowledge graph, quantify the multidimensional credibility of the data based on the semantic alignment results, and generate quality assessment results based on the multidimensional credibility and the equipment safety knowledge base. The data fusion correction module is used to identify correction areas in the preliminary fusion data based on the quality assessment results, combine the correction area identification results with the security monitoring knowledge graph to generate a knowledge-guided correction strategy, and adjust the preliminary fusion data according to the correction strategy to obtain the security monitoring fusion data.

5. The safety monitoring data fusion system based on adaptive Kalman filtering according to claim 4, characterized in that, The quality assessment module includes: a semantic association alignment module, a multidimensional credibility quantification module, a comprehensive score calculation module, and a quality grading determination module; The semantic association alignment module is used to identify the devices and status nodes corresponding to the monitoring parameters in the preliminary fused data in the security monitoring knowledge graph through entity links, and to establish semantic associations with security rules in the security monitoring knowledge graph by combining relationship matching, so as to obtain semantic alignment results. The multidimensional credibility metric module is used to perform multidimensional credibility metric on the semantic alignment result from the dimensions of physical rationality, logical consistency and temporal integrity, so as to obtain multidimensional credibility. The comprehensive score calculation module is used to determine the initial dimension weights based on the multidimensional credibility and the priority of different risk scenarios in the device security knowledge base, and to calculate the comprehensive credibility score by using the dynamic weighted fusion method on the multidimensional credibility and the initial dimension weights. The quality grading determination module is used to determine the quality grading based on the comprehensive credibility score and in combination with a preset credibility threshold, so as to obtain the quality assessment result.

6. The safety monitoring data fusion system based on adaptive Kalman filtering according to claim 5, characterized in that, The multi-dimensional credibility measurement of semantic alignment results from the dimensions of physical rationality, logical consistency, and temporal integrity includes: In the dimension of physical rationality, the deviation of the theoretical feasible region is calculated based on the physical constraints of the monitoring parameters associated in the semantic alignment results, and the deviation is mapped to the standard interval to obtain the credibility of the physical rationality dimension. In terms of logical consistency, domain logical consistency is detected based on the fault causal chain and security rules in the semantic alignment results, and a logical consistency dimension credibility is generated. In the temporal integrity dimension, the continuity, smoothness, and sampling regularity of the data on the time axis are analyzed based on the temporal context in the semantic alignment results. The reliability of the data temporal structure is evaluated based on the degree analysis results, and the credibility of the temporal integrity dimension is obtained.

7. A security monitoring data fusion method based on adaptive Kalman filtering, implemented based on the security monitoring data fusion system based on adaptive Kalman filtering as described in any one of claims 1-6, characterized in that, The method includes: S1. Obtain safety monitoring data and perform outlier processing on the safety monitoring data to obtain cleaned safety monitoring data. Then, use a spatiotemporal alignment algorithm to align the cleaned safety monitoring data to obtain aligned safety monitoring data. S2. Perform reconstruction processing on the aligned security monitoring data to obtain reconstructed security monitoring data, and perform adaptive weighted fusion processing on the reconstructed security monitoring data based on the adaptive Kalman filter algorithm to obtain preliminary fused data. S3. Construct a safety monitoring knowledge graph based on the equipment safety knowledge base, and combine the safety monitoring knowledge graph with the preliminary fusion data to conduct a quality assessment. Based on the assessment results, optimize the preliminary fusion data to obtain the safety monitoring fusion data.

Images