A multi-institution quality data real-time comparison method based on big data analysis

Abstract

The application discloses a kind of based on big data analysis's multi-agency quality data real-time comparison method, it is related to quality data analysis technical field, including the following steps: acquisition multi-agency quality time series data, environmental covariate data and agency operating state data, and based on state transition inflection point carries out elastic time series alignment;Cross-agency quality state three-order tensor and environmental covariate matrix are constructed;Using environment orthogonal and sparse constraint's Tucker tensor decoupling algorithm peels off environmental covariant component, obtains agency specific residual tensor;Multi-agency residual flow hypergraph is constructed;It is input improved VGAE model, generates comparison score;When comparison score is over threshold, generate root cause agency mark, abnormal quality characteristic mark and abnormal hyperedge mark;And according to feedback online updates decomposition basis vector and hyperedge weight.The application improves the accuracy of multi-agency quality data real-time comparison, anti-environmental interference capability and abnormal root cause positioning efficiency.

Description

Technical Field

[0001] This invention relates to the field of quality data analysis and intelligent comparison technology, and in particular to a method for real-time comparison of quality data from multiple institutions based on big data analysis. Background Technology

[0002] With the increasing demand for collaborative quality data supervision across factories, hospitals, and regional centers, real-time comparison technology of multi-institutional quality time series data is gradually being used to detect differences in quality fluctuations between institutions. Existing technologies employ multivariate temporal anomaly detection, tensor decomposition, graph neural networks, or variational graph autoencoders to identify anomalies in multi-source data; for example, tensor decomposition has been publicly disclosed for multidimensional temporal anomaly detection, and variational graph autoencoders have been publicly disclosed for unsupervised latent variable representation learning of graph-structured data.

[0003] However, in multi-institution quality comparison scenarios, existing technologies still have the following shortcomings: different institutions have inconsistent sampling frequencies, equipment clocks, and process switching times, and direct interpolation or alignment by timestamp can easily smooth out real transient fluctuations and generate false anomalies; environmental covariates such as temperature, humidity, air pressure, and batch environment can cause synchronous shifts in quality data from multiple institutions, and static threshold or ordinary residual comparison is difficult to distinguish between environmental covariance and institution-specific quality anomalies; traditional matrix splicing or univariate comparison is difficult to retain the high-order coupling relationship between time, institution, and quality features, and is prone to missing hidden quality problems such as normal individual indicators but abnormal joint distribution; existing anomaly detection and root cause localization are usually performed separately, and after an anomaly is triggered, manual investigation of the institution, quality features, and related institution range is still required, resulting in a delayed response and making it difficult to meet the business needs of real-time comparison and rapid traceability of multi-institution quality data.

[0004] Therefore, how to provide a real-time comparison method for quality data from multiple institutions based on big data analysis is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0005] One objective of this invention is to propose a real-time comparison method for quality data from multiple institutions based on big data analysis. This invention fully utilizes flexible temporal alignment, Tucker tensor decoupling, streaming hypergraphs, and an improved VGAE model to perform environmental interference removal, residual correlation modeling, and anomaly comparison analysis on quality data from multiple institutions. It has the advantages of high comparison accuracy, strong resistance to environmental interference, and high root cause localization efficiency.

[0006] A method for real-time comparison of quality data from multiple institutions based on big data analysis, according to an embodiment of the present invention, includes the following steps: Step 1: Collect multi-mechanism mass time series data, environmental covariate data, and mechanism operation status data. Based on the mechanism operation status data, extract the state transition inflection point of the mass time series data from steady state to wave dynamics, and use the state transition inflection point as the alignment anchor point to perform elastic time series alignment to obtain the aligned mass data sequence. Step 2: Construct the aligned quality data sequence into a cross-institutional quality state third-order tensor according to the three dimensions of time, institution, and quality characteristics, and map the environmental covariate data into an environmental covariate matrix; Step 3: Using the environmental covariate matrix as a constraint, the Tucker tensor decoupling algorithm with environmental orthogonality and sparse constraints is used to remove the environmental covariant components from the third-order tensor of the cross-mechanism mass state to obtain the mechanism-specific residual tensor. Step 4: Using the mechanism residual vector in the mechanism-specific residual tensor as the residual node feature, construct a multi-mechanism residual streaming hypergraph through a streaming hypergraph topology mapping mechanism; Step 5: Input the multi-mechanism residual streaming hypergraph into the improved VGAE model, and process it through the differential high-pass graph attention encoder, compact class cohesion loss function and hyperedge feature and topology joint decoder to obtain the joint reconstruction error as the comparison score; Step 6: When the comparison score is not less than the preset comparison threshold, calculate the Mahalanobis distance of the current residual node latent variable from the center of the compact Gaussian kernel, and generate root cause mechanism markers, anomalous quality feature markers, and anomalous hyperedge markers. Step 7: Based on the root cause mechanism label, abnormal quality feature label, abnormal hyperedge label and quality inspection results, filter the non-abnormal residual feature feedback, and update the decomposed basis vectors of the Tucker tensor decoupling algorithm and the hyperedge weights of the multi-mechanism residual flow hypergraph online.

[0007] Optionally, step one specifically includes: Based on the multi-institution access configuration table, a binding relationship is established between institution identifier, acquisition terminal identifier, quality feature identifier, environmental covariate identifier and institution operation status identifier. Binding matching, clock correction and period identifier mapping are performed on the multi-institution quality time series data, environmental covariate data and institution operation status data to obtain the quality time series data to be aligned. Based on the stable operation marker, load switching marker, process switching marker, and equipment start-up and shutdown marker, the steady-state candidate interval and the wave dynamic candidate interval are determined, and the first-order absolute difference of the quality measurement value is calculated at the boundary between the steady-state candidate interval and the wave dynamic candidate interval. The sliding window mean and standard deviation are generated based on the absolute value of the first difference within a preset period before the boundary position, and the sliding window mean, standard deviation and running state weights are weighted to obtain the adaptive fluctuation threshold. The first period in which the absolute value of the first-order difference is not less than the adaptive fluctuation threshold and the fluctuation condition is met for a continuous preset period is determined as the state transition inflection point, and an alignment anchor point group is generated according to the order of the mechanism identifier, quality characteristic identifier and the occurrence of the state transition inflection point. Using adjacent alignment anchor point groups as boundaries, the median number of cycles between two adjacent state transition inflection points for each mechanism is determined as the reference number of cycles. The ratio of the corresponding number of cycles for each mechanism to the reference number of cycles is used as the elastic scaling factor. The alignment quality time series data is stretched or compressed according to the elastic scaling factor to obtain the alignment quality data sequence.

[0008] Optionally, step two specifically includes: Based on the period identifier, mechanism identifier, and quality feature identifier in the aligned quality data sequence, time dimension index, mechanism dimension index, and quality feature dimension index are generated respectively, and the quality measurement values ​​are assigned to the corresponding tensor units of the third-order tensor of the cross-mechanism quality state. A missing mask is generated based on the missing marker, an anomaly mask is generated based on out-of-bounds records and invalid binding records, a valid marker is generated for tensor units that do not correspond to the missing mask and anomaly mask, and the quality measurement values ​​are unified to the standard dimension according to the dimension conversion coefficients corresponding to the quality characteristics. An environmental covariate matrix is ​​constructed based on the period identifier, environmental covariate identifier, and environmental covariate value in the environmental covariate data. The time dimension index of the environmental covariate matrix is ​​consistent with the time dimension index of the third-order tensor of cross-institutional quality state. When the same period identifier and the same environmental covariate identifier correspond to multiple environmental covariate values, a weighted average is performed using the proportion of the effective operating time of the corresponding institution as the weight to obtain the elements of the environmental covariate matrix. For missing environmental covariate matrix elements, the elements of the previous effective period are used to generate environmental missing markers. Establish a mapping relationship between the cross-institutional quality state third-order tensor, missing mask, anomaly mask, valid label, environmental covariate matrix, and environmental missing label under the same period identifier.

[0009] Optionally, step three specifically includes: Based on the missing mask and the outlier mask, valid and invalid computational units are determined. Invalid computational units are excluded from the calculation range of Tucker decomposition error. The mean is removed and the standard deviation is normalized on the environmental covariate matrix to obtain the standardized environmental covariate matrix. After temporarily filling the invalid computational units, a third-order tensor to be decomposed is formed. The third-order tensor to be decomposed is expanded along the time dimension, mechanism dimension and quality feature dimension, respectively. Principal component decomposition is performed on each expanded matrix to obtain the time decomposition basis vector set, mechanism decomposition basis vector set and quality feature decomposition basis vector set. The third-order tensor to be decomposed is projected onto the subspaces corresponding to the time decomposition basis vector set, the mechanism decomposition basis vector set, and the quality feature decomposition basis vector set, respectively, to obtain the Tucker core tensor. The Tucker core tensor is then restored to the original tensor dimension to form a common reconstruction tensor. Within the effective computational unit, the difference between the third-order tensor of cross-mechanism quality state and the common reconstruction tensor is calculated to obtain the candidate residual tensor. For the residual time series corresponding to the same mechanism identifier and the same quality feature identifier, the environmental covariance residual series is calculated based on the standardized environmental covariance matrix. By subtracting the environmental covariant residual sequence from the residual time series, an environmental orthogonal residual sequence is obtained. Based on the median absolute value of the residuals of the environmental orthogonal residual sequence and the preset sparsity coefficient, mechanism-specific residual units are screened to generate mechanism-specific residual tensors. At the same time, the decomposition basis vector set of the Tucker tensor decoupling algorithm is recorded.

[0010] Optionally, step four specifically includes: Based on the mechanism-specific residual tensor, the mechanism residual vector corresponding to the same period and the same mechanism is extracted, and the mechanism residual vector, mechanism identifier, period identifier and quality feature identifier set are combined into residual node features; The residual vector of the mechanism is normalized to generate a residual direction sequence and a residual magnitude sequence. Candidate associated node pairs are determined based on the cosine similarity, residual direction consistency rate and residual magnitude difference rate between residual nodes. Candidate associated node pairs that have common residual nodes and the same set of quality feature identifiers participating in similarity calculation are merged into candidate superedges, and the weights of candidate superedges are calculated based on average cosine similarity, average residual direction consistency rate, and average residual magnitude proximity. Candidate hyperedges with weights not less than a preset hyperedge weight threshold are identified as valid hyperedges. The residual nodes, valid hyperedges, and hyperedge weights are then updated in a streaming manner according to the periodic order to generate a multi-mechanism residual streaming hypergraph.

[0011] Optionally, the improved VGAE model includes a differential high-pass graph attention encoder, a variational coding layer, a compact class cohesion loss function, and a joint decoder of hyperedge features and topology. Based on the residual node features, effective hyperedges, and hyperedge weights in the multi-mechanism residual flow hypergraph, construct the node feature matrix, hypergraph association matrix, and hyperedge weight diagonal matrix; The node feature matrix is ​​input into the differential high-pass graph attention encoder. The first-order neighborhood node set is determined based on the hypergraph association matrix and the hyperedge weight diagonal matrix. The neighborhood aggregation feature is obtained through the graph attention network. Residual high-pass filtering is performed on the target residual node features and neighborhood aggregation features to obtain high-pass residual features. The high-pass residual features are then input into the variational coding layer, and the latent variables of the residual nodes are obtained through the mean mapping layer, the log-variance mapping layer, and reparameterized sampling. The compact Gaussian kernel center and compact Gaussian kernel covariance matrix are calculated based on the latent variables of normal residual nodes, and the compact class cohesion loss function is constructed by KL divergence loss, center loss and margin loss. The latent variables of residual nodes within the same effective hyperedge are subjected to bilinear decoding and covariance decoding by the joint decoder of hyperedge features and topology, resulting in the hyperedge topology reconstruction probability and the hyperedge feature covariance reconstruction matrix. The binary cross-entropy loss is calculated based on the hyperedge topological reconstruction probability, the Frobenius norm loss is calculated based on the reconstruction matrix of the feature covariance within the hyperedge, and the joint reconstruction error is obtained by weighting. The joint reconstruction error is used as the alignment score.

[0012] Optionally, step six specifically includes: Based on the historical periodic statistical comparison score distribution confirmed as normal by the quality inspection results during the calibration phase, and based on the quantile value corresponding to the preset upper bound quantile in the normal comparison score distribution, the preset comparison threshold is determined. When the alignment score of the current period is less than the preset alignment threshold, a normal alignment mark is generated. When the alignment score of the current period is not less than the preset alignment threshold, the Mahalanobis distance calculation of the latent variables of the current residual node is triggered. Based on the current residual node latent variable, the compact Gaussian kernel center, and the compact Gaussian kernel covariance matrix, calculate the offset vector of the current residual node latent variable relative to the compact Gaussian kernel center, and use the inverse matrix of the compact Gaussian kernel covariance matrix to perform covariance normalization on the offset vector to obtain the Mahalanobis distance. Based on the distribution of Mahalanobis distance of latent variables of normal residual nodes, a preset Mahalanobis distance threshold is determined. Residual nodes with Mahalanobis distance not less than the preset Mahalanobis distance threshold are identified as candidate root cause nodes, and the institution identifiers corresponding to the candidate root cause nodes are identified as root cause institution markers. Abnormal quality feature labels are determined based on the institutional residual vector and the distribution of the absolute value of normal residuals corresponding to candidate root cause nodes. Abnormal hyperedge labels are determined based on the hyperedge topology reconstruction error, hyperedge feature reconstruction error and hyperedge weight of the effective hyperedge to which the candidate root cause node belongs.

[0013] Optionally, step seven specifically includes: Based on the root cause mechanism marker, abnormal quality feature marker, abnormal over-edge marker and quality inspection results of the current cycle, the cycle in which the quality inspection result is qualified and there is a root cause mechanism marker, abnormal quality feature marker or abnormal over-edge marker is determined as a non-abnormal drift cycle. The residual node features corresponding to the root cause mechanism marker, abnormal quality feature marker, or abnormal super-edge marker in the non-abnormal drift cycle are determined as non-abnormal residual feature feedback, and the absorbable non-abnormal drift is determined based on the consistency of the residual direction and the change in the residual amplitude of the non-abnormal residual feature feedback. The third-order tensor fragments of cross-mechanism mass states corresponding to absorbable non-abnormal drifts are incorporated into the incremental decomposition samples. Incremental principal component decomposition is performed on the incremental decomposition samples to obtain the updated candidate decomposition basis vector set. The common reconstruction tensor and mechanism-specific residual tensor of non-abnormal drift cycles are recalculated using the updated candidate decomposition basis vector set. When the updated reconstruction error and environmental projection correlation quantities do not increase, the updated candidate decomposition basis vector set is determined as the decomposition basis vector set updated by the Tucker tensor decoupling algorithm. Hyperedge stability is generated based on the co-occurrence relationship of residual nodes that recur in the feedback of non-abnormal residual features. The hyperedge weights of the multi-mechanism residual streaming hypergraph are updated according to the hyperedge stability and the preset decay coefficient. The updated set of decomposed basis vectors and hyperedge weights are used for the environmental covariance component stripping and multi-mechanism residual streaming hypergraph construction in the next cycle.

[0014] The beneficial effects of this invention are: This invention extracts state transition inflection points from the operational status data of the mechanism and uses these inflection points as alignment anchors for flexible temporal alignment. This solves the interpolation distortion problem caused by inconsistencies in sampling frequencies, equipment clocks, and process switching times among different mechanisms, ensuring that the aligned quality data sequence retains the true dynamic changes of the wave. By preserving the orthogonal structure between time, mechanism, and quality features through a third-order tensor of cross-mechanism quality states, and by using an environmental covariate matrix-constrained Tucker tensor decoupling algorithm to remove environmental covariant components, it solves the problem of false alarms due to synchronization offsets caused by temperature, humidity, air pressure, and batch environment, making the mechanism-specific residual tensor more concentrated in reflecting the mechanism's own quality anomalies. Finally, by using a streaming hypergraph topology mapping mechanism to construct residual vectors into residual nodes and form effective hyperedges, it addresses the difficulty of ordinary pairwise graphs in representing multiple mechanisms and multiple... The problem of joint coupling of quality features is addressed. By employing a differential high-pass graph attention encoder, a compact class cohesive loss function, and a joint decoder of hyperedge features and topology in the improved VGAE model, the issues of smoothing small anomalies in standard graph encoding, loose distribution of latent variables, and inability to reconstruct hyperedge features are resolved, enabling the joint reconstruction error to characterize high-order residual coupling anomalies. Furthermore, by generating root cause mechanism markers, anomalous quality feature markers, and anomalous hyperedge markers using the Mahalanobis distance between the current residual node's latent variable and the center of the compact Gaussian kernel, the problem of the disconnect between anomaly comparison and root cause localization is solved. Finally, by using feedback from non-anomaly residual features to update the decomposition basis vectors and hyperedge weights online, the comparison benchmark is adaptively adjusted with normal process drift, which is of great significance for reducing environmental false alarms, improving the detection capability of hidden joint anomalies, and shortening the quality traceability link for multiple mechanisms. Attached Figure Description

[0015] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings: Figure 1 This is a flowchart of a method for real-time comparison of quality data from multiple institutions based on big data analysis, as proposed in this invention. Figure 2 This is a schematic diagram of a real-time comparison method for quality data from multiple institutions based on big data analysis proposed in this invention. Figure 3 This is a data flow diagram of the improved VGAE model in a real-time comparison method for quality data from multiple institutions based on big data analysis proposed in this invention. Detailed Implementation

[0016] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0017] refer to Figures 1-3 A real-time comparison method for quality data from multiple institutions based on big data analysis includes the following steps: Step 1: Collect multi-mechanism mass time series data, environmental covariate data, and mechanism operation status data. Based on the mechanism operation status data, extract the state transition inflection point of the mass time series data from steady state to wave dynamics, and use the state transition inflection point as the alignment anchor point to perform elastic time series alignment to obtain the aligned mass data sequence. Step 2: Construct the aligned quality data sequence into a cross-institutional quality state third-order tensor according to the three dimensions of time, institution, and quality characteristics, and map the environmental covariate data into an environmental covariate matrix; Step 3: Using the environmental covariate matrix as a constraint, the Tucker tensor decoupling algorithm with environmental orthogonality and sparse constraints is used to remove the environmental covariant components from the third-order tensor of the cross-mechanism mass state to obtain the mechanism-specific residual tensor. Step 4: Using the mechanism residual vector in the mechanism-specific residual tensor as the residual node feature, construct a multi-mechanism residual streaming hypergraph through a streaming hypergraph topology mapping mechanism; Step 5: Input the multi-mechanism residual streaming hypergraph into the improved VGAE model, and process it through the differential high-pass graph attention encoder, compact class cohesion loss function and hyperedge feature and topology joint decoder to obtain the joint reconstruction error as the comparison score; Step 6: When the comparison score is not less than the preset comparison threshold, calculate the Mahalanobis distance of the current residual node latent variable from the center of the compact Gaussian kernel, and generate root cause mechanism markers, anomalous quality feature markers, and anomalous hyperedge markers. Step 7: Based on the root cause mechanism label, abnormal quality feature label, abnormal hyperedge label and quality inspection results, filter the non-abnormal residual feature feedback, and update the decomposed basis vectors of the Tucker tensor decoupling algorithm and the hyperedge weights of the multi-mechanism residual flow hypergraph online.

[0018] In this embodiment, step one specifically includes: Read the multi-institution access configuration table, establish the binding relationship between institution identifier, acquisition terminal identifier, quality characteristic identifier, environmental covariate identifier and institution operation status identifier, receive the quality time series data, environmental covariate data and institution operation status data uploaded by each institution, extract the institution identifier, quality characteristic identifier, acquisition time, quality measurement value, environmental covariate value, stable operation flag, load switching flag, process switching flag and equipment start and stop flag, correct the acquisition time to the corrected acquisition time according to the edge gateway clock, and map the corrected acquisition time to the period identifier according to the preset comparison period; Records that do not match a binding relationship are marked as invalid binding records and removed. Records whose quality measurement values ​​exceed the upper or lower limit of the corresponding quality feature range are marked as out-of-bounds records and removed. For duplicate records under the same institution identifier, the same quality feature identifier, and the same period identifier, the record whose calibration acquisition time is closest to the period center is retained. For periods that have not received quality measurement values, a missing mark is generated to obtain the quality time series data to be aligned. In one specific implementation, the lower limit of the product dimensional deviation range is set to -0.50mm, and the upper limit is set to 0.50mm; the lower limit of the assembly gap range is set to 0.10mm, and the upper limit is set to 1.20mm; the lower limit of the pressure resistance test value range is set to 0MPa, and the upper limit is set to 5MPa; the lower limit of the appearance defect rate range is set to 0%, and the upper limit is set to 10%. If mechanism A uploads two dimensional deviation data points within cycle T002, with acquisition times of 10:06:00 and 10:07:40 respectively, and the cycle center is 10:07:30, then the data corresponding to 10:07:40 is retained. If mechanism B does not upload assembly gap data in cycle T003, then a missing marker is generated for that cycle. For each institutional identifier and quality feature identifier corresponding to the quality time series data to be aligned, if the stable operation markers are all valid and the load switching marker, process switching marker, and equipment start-stop marker are all invalid within M consecutive cycles, then the M consecutive cycles are determined as a steady-state candidate interval. If any valid marker among the load switching marker, process switching marker, or equipment start-stop marker appears within N cycles after the steady-state candidate interval, then the N cycles are determined as a wave dynamic candidate interval, where M and N are both preset positive integers. In one specific implementation, M can be 6 and N can be 4. Since the preset comparison period is 5 minutes, 6 consecutive periods correspond to 30 minutes of stable operation time, and 4 consecutive periods correspond to 20 minutes of fluctuation observation time. If the dimensional deviation data of mechanism A shows that all stable operation markers are valid within periods T010 to T015, and the load switching marker, process switching marker, and equipment start / stop marker are all invalid, then T010 to T015 is determined as a steady-state candidate interval; if a process switching marker is valid within periods T016 to T019, then T016 to T019 is determined as a wave dynamic candidate interval. At the boundary between the steady-state candidate interval and the wave dynamic candidate interval, the first-order absolute value of the quality measurement value is calculated, and the sliding window mean and sliding window standard deviation are calculated using the first-order absolute value of the difference of the M periods before the boundary. The sliding window mean, sliding window standard deviation and operating state weight are weighted and summed to generate an adaptive fluctuation threshold. The operating state weight is obtained by adding the preset weights of the load switching flag, process switching flag and equipment start-stop flag. The preset weights for load switching markers, process switching markers, and equipment start / stop markers can be set to 0.4, 0.4, and 0.2, respectively. The weight of the sliding window mean is set to 1, the weight of the sliding window standard deviation is set to 2, and the operating status weight is used as an additional adjustment quantity in the adaptive fluctuation threshold calculation. If the mean absolute value of the first-order difference of the six cycles before the boundary position is 0.012 and the standard deviation is 0.004, and the current process switching marker is valid, then the operating status weight is 0.4. The adaptive fluctuation threshold can be obtained by combining the mean, standard deviation, and operating status weight terms, and is used to determine whether the cycle has entered the wave dynamic. The first period in which the absolute value of the first-order difference is not less than the adaptive fluctuation threshold and the absolute value of the first-order difference for N consecutive periods from that period is not less than the average value of the sliding window is determined as the state transition inflection point. According to the mechanism identifier, quality feature identifier and the order of occurrence of the state transition inflection point, the state transition inflection points in different mechanisms that correspond to the same quality feature and have the same occurrence order are grouped into alignment anchor points. For example, if mechanisms A, B, C, and D detect the first state transition inflection point at T016, T018, T017, and T019 respectively, based on the dimensional deviation characteristic, then these four state transition inflection points form the first group of alignment anchor points because they correspond to the same quality characteristic and are all the first inflection points to occur. If a second state transition inflection point is subsequently detected at T060, T062, T061, and T063 respectively, then a second group of alignment anchor points is formed. Using adjacent alignment anchor point groups as boundaries, calculate the number of cycles for each mechanism between two adjacent state transition inflection points. Determine the median number of cycles for each mechanism as the reference number of cycles. Use the ratio of the number of cycles for each mechanism to the reference number of cycles as the elastic scaling factor. Then, stretch or compress the time-series data of the alignment quality according to the elastic scaling factor. During the stretching or compression process, maintain the temporal order of quality measurement values ​​under the same mechanism identifier and the same quality feature identifier to obtain the alignment quality data sequence. For example, the number of cycles between the first and second alignment anchor groups for mechanisms A, B, C, and D are 44, 40, 42, and 46 respectively, with a median of 43. Therefore, the base number of cycles is 43. The elastic scaling factor for mechanism A is 44 / 43, and for mechanism B it is 40 / 43. The corresponding segment of mechanism A is compressed by a ratio of 44 to 43, and the corresponding segment of mechanism B is stretched by a ratio of 40 to 43. Stretching or compressing the time axis only changes the aligned cycle position, not the order of mass measurements within the original mechanism.

[0019] In this embodiment, step two specifically includes: Read the period identifier, organization identifier, quality feature identifier and quality measurement value from the aligned quality data sequence, generate a time dimension index according to the time sequence of the period identifier, generate an organization dimension index according to the fixed coding order of the organization identifier, and generate a quality feature dimension index according to the fixed coding order of the quality feature identifier. Assign the mass measurement values ​​corresponding to the same period identifier, the same mechanism identifier, and the same mass characteristic identifier to the corresponding tensor units of the third-order tensor of the cross-mechanism mass state, and unify the mass measurement values ​​to the standard dimensions according to the dimension conversion coefficient of the corresponding mass characteristic; In one specific implementation, product dimensional deviations are measured in millimeters, pressure resistance test values ​​in MPa, and appearance defect rates in percentage. If an organization uploads dimensional deviations in micrometers, these are converted to millimeters using a dimensional conversion factor of 1mm = 1000μm before being assigned to the tensor unit. This avoids comparison distortion caused by different organizations using different data acquisition units.

[0020] Generate a missing mask for tensor cells with missing labels, generate an anomaly mask for tensor cells corresponding to out-of-bounds records or invalid binding records, and generate a valid label for tensor cells that do not have corresponding missing or anomaly masks. Read the period identifier, environmental covariate identifier, and environmental covariate value from the environmental covariate data, arrange the environmental covariate values ​​according to the same time dimension index as the third-order tensor of cross-institutional quality status, and generate an environmental covariate column index according to the fixed coding order of the environmental covariate identifier; If the same period identifier and the same environmental covariate identifier correspond to multiple environmental covariate values, then the weight of the multiple environmental covariate values ​​is taken as the proportion of the effective operating time of the corresponding institution to the sum of the effective operating times of all effective operating institutions, and the elements of the environmental covariate matrix for that period are obtained. If there is no corresponding environmental covariate value for that period, then the elements of the environmental covariate matrix of the previous effective period are used and an environmental missing marker is generated. In one specific implementation, for example, if the effective running times of mechanisms A, B, C, and D within period T020 are 5 minutes, 5 minutes, 3 minutes, and 2 minutes respectively, then the total effective running time is 15 minutes, with corresponding weights of 5 / 15, 5 / 15, 3 / 15, and 2 / 15 respectively. If the workshop temperatures uploaded by the four mechanisms are 24.0℃, 24.4℃, 23.8℃, and 25.0℃ respectively, then the workshop temperature matrix elements for that period are obtained by weighting according to the above weights. If no workshop humidity is received in period T021, then the workshop humidity matrix elements from period T020 are used, and an environment missing marker is generated.

[0021] A mapping relationship is established under the same period identifier for the cross-institutional quality state third-order tensor, missing mask, anomaly mask, valid label, environmental covariate matrix, and environmental missing label, so that the environmental covariate matrix and the cross-institutional quality state third-order tensor are aligned periodically in the time dimension.

[0022] In this embodiment, step three specifically includes: Read the cross-institutional quality state third-order tensor, environmental covariate matrix, missing mask and anomalous mask, determine the tensor cells that do not correspond to missing mask and anomalous mask as valid calculation cells, determine the tensor cells that correspond to missing mask or anomalous mask as invalid calculation cells, and exclude invalid calculation cells from the Tucker decomposition error calculation range. Read each environmental covariate column in the environmental covariate matrix according to the time dimension index, calculate the mean and standard deviation of each environmental covariate column within the effective time period, subtract the corresponding mean from each environmental covariate value and divide it by the corresponding standard deviation to form a standardized environmental covariate matrix, identify the environmental covariate columns with a standard deviation of zero as invalid environmental covariate columns, and remove invalid environmental covariate columns from the environmental orthogonal calculation range; For invalid computational units in the cross-mechanism quality state third-order tensor, the quality measurement values ​​of adjacent effective time periods under the same mechanism identifier and the same quality feature identifier are temporarily filled to form a third-order tensor to be decomposed. The third-order tensor to be decomposed is expanded along the time dimension, mechanism dimension and quality feature dimension to obtain the time expansion matrix, mechanism expansion matrix and quality feature expansion matrix. Principal component decomposition is performed on the time expansion matrix, mechanism expansion matrix and quality feature expansion matrix respectively, and a preset number of principal component vectors are extracted to obtain the time decomposition basis vector set, mechanism decomposition basis vector set and quality feature decomposition basis vector set. In one specific implementation, the temporary padding of invalid computational cells is only used to ensure that Tucker tensor decomposition can be performed and is not used as the final residual output. For example, if mechanism C lacks an assembly gap measurement value in period T050, and the corresponding values ​​for T049 and T051 are 0.42mm and 0.46mm respectively, then the temporary padding value can be 0.44mm. During principal component decomposition, the number of time decomposition basis vectors can be 12, the number of mechanism decomposition basis vectors can be 3, and the number of mass feature decomposition basis vectors can be 3, used to retain the time-varying principal mode, the mechanism difference principal mode, and the mass feature coupled principal mode, respectively.

[0023] The third-order tensor to be decomposed is projected along the time dimension to the subspace corresponding to the time decomposition basis vector set, along the mechanism dimension to the subspace corresponding to the mechanism decomposition basis vector set, and along the quality feature dimension to the subspace corresponding to the quality feature decomposition basis vector set, thus obtaining the Tucker core tensor. The Tucker core tensor is then restored to the original tensor dimension according to the time decomposition basis vector set, the mechanism decomposition basis vector set, and the quality feature decomposition basis vector set, forming a common reconstruction tensor. Within the effective computational unit, the mass measurement value of the cross-mechanism mass state third-order tensor is subtracted from the corresponding reconstruction value of the common reconstruction tensor to obtain the candidate residual tensor. For the residual time series corresponding to the same mechanism identifier and the same mass feature identifier in the candidate residual tensor, the environmental linear projection coefficient is calculated with the standardized environmental covariate value of the same time period as the independent variable and the residual time series as the dependent variable. The environmental covariate residual series is obtained by multiplying the standardized environmental covariate matrix with the environmental linear projection coefficient. After subtracting the environmental covariant residual series from the residual time series, an environmental orthogonal residual series is obtained. The median of the absolute value of the residuals in the environmental orthogonal residual series is calculated. Residual cells whose absolute residual value is less than the product of the median of the absolute residual value and the preset sparse coefficients are assigned a value of zero. The remaining residual cells are identified as mechanism-specific residual cells. All mechanism-specific residual cells are arranged according to the time dimension, mechanism dimension, and quality feature dimension to generate a mechanism-specific residual tensor. At the same time, the current time decomposition basis vector set, mechanism decomposition basis vector set, and quality feature decomposition basis vector set are recorded as the decomposition basis vector set of the Tucker tensor decoupling algorithm.

[0024] In one specific implementation, the preset sparsity coefficient can be set to 1.2. If the median absolute value of the residuals in a certain residual time series is 0.015, then the residual retention boundary is 0.018. Residual cells with an absolute value less than 0.018 of environmental orthogonal residuals are assigned a value of zero, indicating that they are suppressed as insignificant residuals; residual cells with an absolute value not less than 0.018 are retained as mechanism-specific residual cells. This processing can reduce the impact of small random fluctuations on the subsequent construction of the streaming hypergraph.

[0025] In this embodiment, step four specifically includes: Read the mechanism-specific residual tensor, extract all quality feature residual values ​​corresponding to the same period and the same mechanism according to the time dimension index and the mechanism dimension index, arrange them in order of quality feature dimension to form the mechanism residual vector, and combine the mechanism residual vector, mechanism identifier, period identifier and quality feature identifier set into residual node features; The residual vector of the mechanism is normalized, and a residual direction sequence is generated according to the positive and negative directions of the residual values ​​of each quality feature. A residual amplitude sequence is generated according to the absolute value of the residual values ​​of each quality feature. The mechanism residual vector, residual direction sequence and residual amplitude sequence after residual amplitude normalization are determined as the topology mapping input of the residual node. Within the current period, the cosine similarity, residual direction consistency rate, and residual amplitude difference rate are calculated pairwise for the residual nodes corresponding to different institutions. The residual direction consistency rate is determined by the proportion of the number of quality features with the same sign to the total number of quality features. The residual amplitude difference rate is determined by the ratio between the mean of the residual amplitude difference between two residual nodes and the mean of the residual amplitude. Residual node pairs with a cosine similarity not less than a preset similarity threshold, a residual direction consistency rate not less than a preset direction consistency threshold, and a residual amplitude difference rate not greater than a preset amplitude difference threshold are identified as candidate associated node pairs. Candidate associated node pairs with common residual nodes and the same quality feature identifier set participating in similarity calculation are merged into candidate superedges, and the number of residual nodes contained in each candidate superedge is not less than the preset minimum number of superedge nodes. In one specific implementation, the preset similarity threshold can be 0.82, the preset direction consistency threshold can be 0.75, the preset amplitude difference threshold can be 0.30, and the preset minimum number of superedge nodes can be 3. If the residual direction consistency rate of the three residual nodes A, B, and C in terms of product size deviation, assembly clearance, and pressure resistance test value all reach 0.75 or higher, and the pairwise cosine similarity is not lower than 0.82, while the residual amplitude difference rate is not greater than 0.30, then these three residual nodes form a candidate superedge.

[0026] For all residual nodes within the candidate superedge, calculate the average cosine similarity, average residual direction consistency rate, and average residual amplitude difference rate. Subtract the average residual amplitude difference rate from one to obtain the average residual amplitude proximity. Weight the average cosine similarity, average residual direction consistency rate, and average residual amplitude proximity rate according to a preset weight coefficient to obtain the candidate superedge weight. Candidate superedges with a weight not less than the preset superedge weight threshold are determined as valid superedges. In one specific implementation, the weights of the average cosine similarity, average residual direction consistency rate, and average residual magnitude closeness can be set to 0.4, 0.35, and 0.25, respectively, and the preset hyperedge weight threshold can be set to 0.70. If the average cosine similarity of a candidate hyperedge is 0.86, the average residual direction consistency rate is 0.80, and the average residual magnitude difference rate is 0.20, then the average residual magnitude closeness is 0.80, and the candidate hyperedge weight is the sum of 0.4×0.86, 0.35×0.80, and 0.25×0.80. When this weight is not less than 0.70, the candidate hyperedge is determined to be a valid hyperedge.

[0027] The residual nodes and effective superedges of the current period are merged into the multi-mechanism residual streaming hypergraph formed in the previous period according to the periodic order. Residual nodes that exceed the preset streaming retention window and effective superedges composed only of the residual nodes are deleted. The weights of effective superedges that recur across consecutive periods are updated according to the preset attenuation coefficient to generate a multi-mechanism residual streaming hypergraph containing residual nodes, effective superedges and superedge weights.

[0028] In one specific implementation, the preset streaming retention window can be set to 24 comparison periods, meaning that residual nodes and valid hyperedges from the most recent 2 hours are retained in a 5-minute comparison period. The preset decay coefficient can be set to 0.90. If a valid hyperedge appears repeatedly in consecutive periods, its hyperedge relationship is retained and its hyperedge weight is updated; if a valid hyperedge does not appear repeatedly for more than 24 comparison periods, it is removed from the multi-organization residual streaming hypergraph or downgraded to a failed state.

[0029] In this embodiment, step five specifically includes: Read the residual node features, effective hyperedges, hyperedge weights and periodic identifiers in the multi-organization residual streaming hypergraph, construct the residual node features into a node feature matrix, construct the membership relationship between residual nodes and effective hyperedges into a hypergraph association matrix, and construct the hyperedge weights into a hyperedge weight diagonal matrix. The first-order neighborhood node set of the residual node is determined based on the hypergraph association matrix and the hyperedge weight diagonal matrix. The node feature matrix is ​​input into the differential high-pass graph attention encoder. The attention coefficient between the target residual node and the first-order neighborhood node is calculated through the graph attention network. The attention coefficient is weighted and normalized according to the hyperedge weight to obtain the neighborhood aggregation feature. Residual high-pass filtering is performed on the target residual node features and neighborhood aggregation features. The difference between the two is used as the high-pass residual features. The high-pass residual features are input into the variational coding layer. The mean mapping layer outputs the latent variable mean vector, and the log-variance mapping layer outputs the latent variable log-variance vector. The latent variables of the residual node are obtained through reparameterized sampling. During the training phase of the improved VGAE model, the compact Gaussian kernel center and compact Gaussian kernel covariance matrix are calculated based on the latent variables of normal residual nodes. The distribution of latent variables of residual nodes is constrained by KL divergence loss to approximate the compact Gaussian kernel represented by the compact Gaussian kernel center and compact Gaussian kernel covariance matrix. The normal residual node latent variables are constrained to converge to the compact Gaussian kernel center by center loss. The distance between the latent variables of abnormal residual nodes and the compact Gaussian kernel center is constrained to be no less than a preset distance threshold by interval loss. In one specific implementation, the node feature dimension of the improved VGAE model can be set to 4, the hidden dimension of the graph attention network can be set to 32, the number of attention heads can be set to 4, and the dimension of residual node latent variables can be set to 16. During training, historical periods with qualified quality detection results are used as normal samples. The weight of KL divergence loss is set to 1.0, the weight of center loss is set to 0.5, the weight of interval loss is set to 0.3, and the preset interval threshold is set to 2.0. Through the above settings, the latent variables of normal residual nodes are concentrated near the compact Gaussian kernel, while the latent variables of abnormal residual nodes maintain a separable distance from the center of the compact Gaussian kernel.

[0030] The joint decoder of hyperedge features and topology reads the latent variables of residual nodes and the hypergraph correlation matrix. It performs bilinear decoding on the latent variables of residual nodes within the same valid hyperedge to obtain the hyperedge topology reconstruction probability, and performs covariance decoding on the latent variables of residual nodes within the same valid hyperedge to obtain the feature covariance reconstruction matrix within the hyperedge. The binary cross-entropy loss is calculated based on the actual topological label of the effective hyperedge and the topological reconstruction probability of the hyperedge. The Frobenius norm loss is calculated based on the feature covariance matrix inside the actual hyperedge and the feature covariance reconstruction matrix inside the hyperedge. The joint reconstruction error is obtained by weighting the loss according to the preset topological loss weight and the preset feature loss weight. The joint reconstruction error is used as the comparison score for the current period.

[0031] In one specific implementation, the preset topology loss weight and the preset feature loss weight can be set to 0.5 and 0.5, respectively. If the hyperedge topology reconstruction error for a certain period is 0.18 and the hyperedge feature covariance reconstruction error is 0.22, then the joint reconstruction error is obtained by weighting the two by 0.5 and 0.5, and is used as the comparison score for that period. The larger the comparison score, the more significant the difference between the current multi-mechanism residual flow hypergraph and the higher-order coupling relationship under normal conditions.

[0032] In this embodiment, step six specifically includes: Read the historical cycles that are confirmed to be normal by the quality inspection results during the calibration phase, statistically analyze the comparison score distribution corresponding to the normal historical cycles, and determine the preset comparison threshold based on the quantile value corresponding to the preset upper bound quantile in the normal comparison score distribution. In one specific implementation, during the calibration phase, a historical period of 30 consecutive days with all quality inspection results meeting the required standards can be selected as the normal sample. The distribution of comparison scores for these historical periods is then statistically analyzed. The preset upper bound quantile can be the 95th percentile, meaning the value corresponding to the 95th percentile in the normal comparison score distribution is used as the preset comparison threshold. This threshold is used to exclude the vast majority of normal fluctuations from the abnormal trigger range.

[0033] Read the alignment score of the current period, compare the alignment score with the preset alignment threshold, generate a normal alignment mark when the alignment score is less than the preset alignment threshold, and trigger the Mahalanobis distance calculation of the latent variables of the current residual node when the alignment score is not less than the preset alignment threshold. Read the current residual node latent variables, compact Gaussian kernel center, and compact Gaussian kernel covariance matrix output by the improved VGAE model, and establish the correspondence between the current residual node latent variables and the residual node, mechanism identifier, and quality feature identifier set; Read the Mahalanobis distance distribution of the latent variables of normal residual nodes relative to the center of the compact Gaussian kernel during the calibration phase, and determine the preset Mahalanobis distance threshold according to the quantile value corresponding to the preset upper bound quantile in the normal Mahalanobis distance distribution. For each current latent variable of residual node, calculate its offset vector with the center of the compact Gaussian kernel, and use the inverse matrix of the compact Gaussian kernel covariance matrix to perform covariance normalization on the offset vector to obtain the Mahalanobis distance of the corresponding residual node. In one specific implementation, the preset upper bound quantile can be the 97th percentile. That is, during the calibration phase, the Mahalanobis distances of all normal residual node latent variables relative to the center of the compact Gaussian kernel are first calculated, and then the 97th percentile of this distribution is taken as the preset Mahalanobis distance threshold. When the Mahalanobis distance of the current residual node latent variable is not less than this threshold, it indicates that the residual node deviates significantly from the normal state in the compact latent space and can be used as a candidate root cause node.

[0034] Residual nodes with a Mahalanobis distance not less than a preset Mahalanobis distance threshold are identified as candidate root cause nodes, and the candidate root cause nodes are sorted in descending order of Mahalanobis distance. The institution identifiers corresponding to the candidate root cause nodes are identified as root cause institution markers. Read the institutional residual vector corresponding to the candidate root cause node, statistically analyze the distribution of the absolute value of the residual in the normal sample during the calibration stage according to the quality feature identifier, and determine the normal residual benchmark value of the corresponding quality feature based on the quantile value corresponding to the preset upper bound quantile in the distribution of the absolute value of the normal residual. The difference between the absolute value of the current quality feature residual and the normal residual benchmark value is determined as the deviation score, and the quality feature identifier with a deviation score not less than zero is determined as the abnormal quality feature marker. In one specific implementation, for product dimensional deviations, assembly clearances, pressure resistance test values, and appearance defect rates, the distribution of the absolute values ​​of the residuals of normal samples during the calibration phase is statistically analyzed, and the 95th percentile is used as the normal residual benchmark value for the corresponding quality characteristic. For example, if the normal residual benchmark value for product dimensional deviation is 0.018 mm, and the absolute value of the product dimensional deviation residual for the current candidate root cause node is 0.026 mm, then the deviation score is 0.008 mm, and this quality characteristic is identified as an abnormal quality characteristic marker.

[0035] Read the topology reconstruction error, feature reconstruction error, and weight of normal and valid hyperedges during the calibration phase. Calculate the abnormal score distribution of normal hyperedges according to the preset error weights. Determine the preset hyperedge abnormality threshold based on the quantile value corresponding to the preset upper bound quantile in the normal hyperedge abnormality score distribution. Read the topology reconstruction error, feature reconstruction error, and weight of the valid hyperedge to which the candidate root cause node belongs. Calculate the current hyperedge abnormal score according to the same preset error weights. Mark valid hyperedges with a current hyperedge abnormality score not less than the preset hyperedge abnormality threshold and containing candidate root cause nodes as abnormal hyperedges.

[0036] In one specific implementation, the preset error weights include a hyperedge topology reconstruction error weight of 0.5 and a hyperedge feature reconstruction error weight of 0.5, with the hyperedge weight serving as an adjustment factor for the anomaly score. During the calibration phase, the anomaly score distribution of normal valid hyperedges is statistically analyzed, and the 95th percentile is used as the preset hyperedge anomaly threshold. If a current valid hyperedge contains a candidate root cause node, and the current hyperedge anomaly score of that valid hyperedge reaches the preset hyperedge anomaly threshold, then that valid hyperedge is identified as an anomaly hyperedge.

[0037] In this embodiment, step seven specifically includes: Read the root cause mechanism marker, abnormal quality feature marker, abnormal over-edge marker and quality inspection results of the current cycle. Determine the cycle in which the quality inspection results are confirmed as unqualified as the abnormal feedback cycle. Determine the cycle in which the quality inspection results are confirmed as qualified and there are root cause mechanism markers, abnormal quality feature markers or abnormal over-edge markers as the non-abnormal drift cycle. Extract the mechanism-specific residual tensor, residual node features, effective superedge and superedge weight corresponding to the non-abnormal drift cycle, and determine the residual node features corresponding to the root cause mechanism label, abnormal quality feature label or abnormal superedge label in the non-abnormal drift cycle as non-abnormal residual feature feedback; The residual mean, residual direction consistency, and residual amplitude change are statistically analyzed based on the time dimension, institutional dimension, and quality characteristic dimension of non-abnormal residual characteristic feedback. If the residual direction remains consistent within a continuous preset number of periods and the corresponding quality test results are all qualified, then the corresponding residual change is determined as an absorbable non-abnormal drift. In one specific implementation, the consecutive preset number of cycles can be 3. If the assembly clearance residual of mechanism B is positively offset in all three consecutive comparison cycles, the residual amplitude change does not exceed the preset drift range, and the corresponding quality inspection results are all qualified, then the residual change is determined to be an absorbable non-abnormal drift. This type of drift indicates that there is a stable long-term change in the production process that does not affect the quality qualification, and it can be used to update the benchmark.

[0038] Read the current time decomposition basis vector set, mechanism decomposition basis vector set, and mass feature decomposition basis vector set of the Tucker tensor decoupling algorithm. Incorporate the third-order tensor fragment of cross-mechanism mass state corresponding to absorbable non-abnormal drift into the incremental decomposition sample. Perform incremental principal component decomposition on the incremental decomposition sample to obtain the updated candidate decomposition basis vector set. The common reconstruction tensor and mechanism-specific residual tensor of non-abnormal drift cycles are recalculated using the updated candidate decomposition basis vector set. If the reconstruction error does not increase compared to the previous version and the projection correlation between the mechanism-specific residual tensor and the environmental covariate matrix does not increase compared to the previous version, then the updated candidate decomposition basis vector set is determined as the decomposition basis vector set updated by the Tucker tensor decoupling algorithm. The co-occurrence relationship of recurring residual nodes is read from the feedback of non-abnormal residual features. The number of occurrences of the same effective hyperedge in consecutive periods, the consistency of residual direction, and the closeness of residual magnitude are combined to form the hyperedge stability. The product of the hyperedge stability and the preset weight gain coefficient is used as the weight increment. The original hyperedge weight is added to the weight increment to obtain the updated hyperedge weight of the recurring effective hyperedge. For effective hyperedges that do not recur in the preset streaming retention window, the original hyperedge weight is multiplied by the preset attenuation coefficient to obtain the updated hyperedge weight. The updated decomposition basis vector set and hyperedge weight are used for the environmental covariance component stripping and multi-mechanism residual streaming hypergraph construction in the next period.

[0039] In one specific implementation, the preset weight gain coefficient can be set to 0.05, the preset attenuation coefficient can be set to 0.90, and the preset streaming retention window can be set to 24 comparison periods. If the stability of a valid hyperedge is 0.8, the weight increment is 0.04, and the updated hyperedge weight is the original hyperedge weight plus 0.04. If a valid hyperedge does not reappear within 24 comparison periods, its updated hyperedge weight is the original hyperedge weight multiplied by 0.90. In this way, long-term stable non-abnormal residual coupling relationships are preserved, while occasional residual correlations are gradually weakened.

[0040] Example 1: To verify the feasibility of this invention in practice, it was applied to a real-time product quality comparison scenario across four regional quality inspection centers. The four institutions, designated A, B, C, and D, all periodically inspect the same type of precision assemblies. Quality characteristics included product dimensional deviations, assembly gaps, pressure resistance test values, and appearance defect rates. Environmental covariates included workshop temperature, workshop humidity, regional air pressure, and production batch environmental codes. Institutional operating status data included stable operation markers, load switching markers, process switching markers, and equipment start / stop markers. Each institution uploaded data via an edge gateway. The preset comparison period was set to 5 minutes. 30 days of historical qualified data were continuously collected as calibration samples, and another 7 days of continuous operating data were used as verification samples.

[0041] In this scenario, the edge gateway first binds the institution identifier, acquisition terminal identifier, quality feature identifier, environmental covariate identifier, and institution operating status identifier according to the multi-institution access configuration table. The product size deviation range is set to -0.50mm to 0.50mm, the assembly gap range to 0.10mm to 1.20mm, the pressure resistance test value range to 0MPa to 5MPa, and the appearance defect rate range to 0% to 10%. When duplicate data for the same quality feature exists within the same period, the data closest to the period center at the acquisition time is retained, and missing periods generate missing markers. The M value for the steady-state candidate interval is 6, and the N value for the wave dynamic candidate interval is 4. The weights for load switching markers, process switching markers, and equipment start / stop markers are 0.4, 0.4, and 0.2, respectively, used to generate adaptive fluctuation thresholds and determine state transition inflection points. Between adjacent alignment anchor point groups, the median number of periods for each institution is used as the baseline number of periods. The time-axis stretching or compression of the time-series data of the quality to be aligned is then performed to obtain the alignment quality data sequence.

[0042] After alignment, the data is used to construct a cross-institutional quality state third-order tensor based on three dimensions: time, institution, and quality features. A validation day contains 288 alignment periods, with the four institutions and four quality features forming 288×4×4 tensor units. The environmental covariate matrix is ​​arranged along the same time dimension; when multiple environmental covariate values ​​exist in the same period, they are weighted averaged based on the proportion of effective runtime of each institution. In the Tucker tensor decoupling algorithm, the number of time decomposition basis vectors is 12, the number of institution decomposition basis vectors is 3, the number of quality feature decomposition basis vectors is 3, and the preset sparsity coefficient is 1.2. In the streaming hypergraph topology mapping, the cosine similarity threshold is 0.82, the residual direction consistency rate threshold is 0.75, the residual magnitude difference rate threshold is 0.30, the minimum number of hyperedge nodes is 3, and the streaming retention window is 24 alignment periods.

[0043] In the improved VGAE model, the graph attention hidden dimension is 32, the number of attention heads is 4, and the dimension of the residual node latent variables is 16. The weights of KL divergence loss, center loss, and margin loss are 1.0, 0.5, and 0.3, respectively, and the preset margin threshold is 2.0. The weights of the binary cross-entropy loss and Frobenius norm loss in the joint reconstruction error are both 0.5. During the calibration phase, the preset alignment threshold is the 95th percentile of the normal alignment score distribution, the preset Mahalanobis distance threshold is the 97th percentile of the Mahalanobis distance distribution of the normal residual node latent variables, the normal residual baseline value is the 95th percentile of the distribution of the absolute value of the corresponding quality feature normal residuals, and the preset hyperedge anomaly threshold is the 95th percentile of the distribution of the normal hyperedge anomaly scores.

[0044] For comparison, Method A is set to "timestamp linear interpolation + static threshold comparison", Method B is set to "timestamp linear interpolation + ordinary Tucker decomposition + standard VGAE model", and Method C is set to "state transition anchor alignment + Tucker tensor decoupling algorithm with environment orthogonality and sparse constraints + unmodified VGAE model".

[0045] Table 1. Comparison of the Implementation Effects of Real-time Comparison of Quality Data from Multiple Institutions

[0046] As shown in Table 1, Method A had an anomaly detection rate of 82.6% and an environmental false alarm rate of 13.8%. This indicates that relying solely on linear interpolation of timestamps and static threshold comparisons can easily lead to misjudgments of synchronous fluctuations in temperature, humidity, and air pressure as quality anomalies. Furthermore, inconsistent sampling times can smooth out some real wave dynamic changes. Method B, by adding ordinary Tucker decomposition and a standard VGAE model to Method A, reduced the environmental false alarm rate to 9.6% and increased the anomaly detection rate to 87.9%, demonstrating that tensor decomposition and graph latent variable modeling can improve the representation of multivariate quality states. However, because it did not use state transition inflection points as alignment anchors, misalignment issues before and after process switching still exist, resulting in a hidden joint anomaly detection rate of only 70.8%.

[0047] Method C, by employing state transition anchor alignment and the Tucker tensor decoupling algorithm with environmental orthogonality and sparse constraints, further reduced the false alarm rate to 5.4% and increased the anomaly detection rate to 91.7%, indicating that aligning the quality data sequence and the mechanism-specific residual tensor can reduce false anomalies and isolate the covariant effects of the environment. However, Method C still uses the unimproved VGAE model, resulting in an insufficiently compact distribution of latent variables in the residual nodes, and it does not jointly reconstruct the hyperedge topology and the feature covariance within the hyperedge. Therefore, the hidden joint anomaly detection rate is 78.6%, and the average source tracing time remains at 48.0 seconds.

[0048] The present invention achieves an anomaly detection rate of 96.3%, reduces the environmental false alarm rate to 2.7%, and achieves a hidden joint anomaly detection rate of 90.4%, while reducing the average source tracing time to 12.5 seconds. Compared to Method C, this invention, while also employing the Tucker tensor decoupling algorithm with state transition anchor alignment and environmental orthogonality and sparse constraints, further amplifies the deviation of residual nodes from neighborhood aggregation features through a differential high-pass graph attention encoder in the improved VGAE model. It compresses the latent variable distribution of normal residual nodes through a compact class cohesion loss function and simultaneously reconstructs the hyperedge topology and hyperedge-internal feature covariance through a hyperedge feature and topology joint decoder. Therefore, it can more accurately identify hidden joint anomalies where "single quality features do not significantly exceed limits but multi-mechanism residual coupling structure is abnormal". The accuracy of root cause location improved from 83.5% in method C to 92.8%, and the accuracy of abnormal quality feature location improved from 80.6% to 89.7%, indicating that the Mahalanobis distance of the current residual node latent variable from the center of the compact Gaussian kernel can directly serve the generation of root cause labeling and abnormal quality feature labeling.

[0049] As can be seen from this embodiment, the present invention reduces false anomalies caused by inconsistencies in sampling frequencies and process switching times across multiple mechanisms by aligning state transition inflection points; it removes covariant environmental components from the third-order tensor of cross-mechanism quality states using the Tucker tensor decoupling algorithm with environmental orthogonality and sparse constraints; it preserves high-order coupling relationships across mechanisms and across quality features using a multi-mechanism residual streaming hypergraph; it enhances the comparison sensitivity of hidden joint anomalies using an improved VGAE model; and it directly generates root cause mechanism markers, anomalous quality feature markers, and anomalous hyperedge markers using Mahalanobis distance. As shown in Table 1, the present invention has significant effects in reducing false environmental alarms, improving anomaly detection, and shortening tracing time, and can meet the engineering application requirements for real-time comparison and rapid location of multi-mechanism quality data.

[0050] The above are merely preferred embodiments of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A method for real-time comparison of quality data from multiple institutions based on big data analysis, characterized in that, Includes the following steps: Step 1: Collect multi-mechanism mass time series data, environmental covariate data, and mechanism operation status data. Based on the mechanism operation status data, extract the state transition inflection point of the mass time series data from steady state to wave dynamics, and use the state transition inflection point as the alignment anchor point to perform elastic time series alignment to obtain the aligned mass data sequence. Step 2: Construct the aligned quality data sequence into a cross-institutional quality state third-order tensor according to the three dimensions of time, institution, and quality characteristics, and map the environmental covariate data into an environmental covariate matrix; Step 3: Using the environmental covariate matrix as a constraint, the Tucker tensor decoupling algorithm with environmental orthogonality and sparse constraints is used to remove the environmental covariant components from the third-order tensor of the cross-mechanism mass state to obtain the mechanism-specific residual tensor. Step 4: Using the mechanism residual vector in the mechanism-specific residual tensor as the residual node feature, construct a multi-mechanism residual streaming hypergraph through a streaming hypergraph topology mapping mechanism; Step 5: Input the multi-mechanism residual streaming hypergraph into the improved VGAE model, and process it through the differential high-pass graph attention encoder, compact class cohesion loss function and hyperedge feature and topology joint decoder to obtain the joint reconstruction error as the comparison score; Step 6: When the comparison score is not less than the preset comparison threshold, calculate the Mahalanobis distance of the current residual node latent variable from the center of the compact Gaussian kernel, and generate root cause mechanism markers, anomalous quality feature markers, and anomalous hyperedge markers. Step 7: Based on the root cause mechanism label, abnormal quality feature label, abnormal hyperedge label and quality inspection results, filter the non-abnormal residual feature feedback, and update the decomposed basis vectors of the Tucker tensor decoupling algorithm and the hyperedge weights of the multi-mechanism residual flow hypergraph online.

2. The method for real-time comparison of quality data from multiple institutions based on big data analysis according to claim 1, characterized in that, Step one specifically includes: Based on the multi-institution access configuration table, a binding relationship is established between institution identifier, acquisition terminal identifier, quality feature identifier, environmental covariate identifier and institution operation status identifier. Binding matching, clock correction and period identifier mapping are performed on the multi-institution quality time series data, environmental covariate data and institution operation status data to obtain the quality time series data to be aligned. Based on the stable operation marker, load switching marker, process switching marker, and equipment start-up and shutdown marker, the steady-state candidate interval and the wave dynamic candidate interval are determined, and the first-order absolute difference of the quality measurement value is calculated at the boundary between the steady-state candidate interval and the wave dynamic candidate interval. The sliding window mean and standard deviation are generated based on the absolute value of the first difference within a preset period before the boundary position, and the sliding window mean, standard deviation and running state weights are weighted to obtain the adaptive fluctuation threshold. The first period in which the absolute value of the first-order difference is not less than the adaptive fluctuation threshold and the fluctuation condition is met for a continuous preset period is determined as the state transition inflection point, and an alignment anchor point group is generated according to the order of the mechanism identifier, quality characteristic identifier and the occurrence of the state transition inflection point. Using adjacent alignment anchor point groups as boundaries, the median number of cycles between two adjacent state transition inflection points for each mechanism is determined as the reference number of cycles. The ratio of the corresponding number of cycles for each mechanism to the reference number of cycles is used as the elastic scaling factor. The alignment quality time series data is stretched or compressed according to the elastic scaling factor to obtain the alignment quality data sequence.

3. The method for real-time comparison of quality data from multiple institutions based on big data analysis according to claim 1, characterized in that, Step two specifically includes: Based on the period identifier, mechanism identifier, and quality feature identifier in the aligned quality data sequence, time dimension index, mechanism dimension index, and quality feature dimension index are generated respectively, and the quality measurement values ​​are assigned to the corresponding tensor units of the third-order tensor of the cross-mechanism quality state. A missing mask is generated based on the missing marker, an anomaly mask is generated based on out-of-bounds records and invalid binding records, a valid marker is generated for tensor units that do not correspond to the missing mask and anomaly mask, and the quality measurement values ​​are unified to the standard dimension according to the dimension conversion coefficients corresponding to the quality characteristics. An environmental covariate matrix is ​​constructed based on the period identifier, environmental covariate identifier, and environmental covariate value in the environmental covariate data. The time dimension index of the environmental covariate matrix is ​​consistent with the time dimension index of the third-order tensor of cross-institutional quality state. When the same period identifier and the same environmental covariate identifier correspond to multiple environmental covariate values, a weighted average is performed using the proportion of the effective operating time of the corresponding institution as the weight to obtain the elements of the environmental covariate matrix. For missing environmental covariate matrix elements, the elements of the previous effective period are used to generate environmental missing markers. Establish a mapping relationship between the cross-institutional quality state third-order tensor, missing mask, anomaly mask, valid label, environmental covariate matrix, and environmental missing label under the same period identifier.

4. The method for real-time comparison of quality data from multiple institutions based on big data analysis according to claim 1, characterized in that, Step three specifically includes: Based on the missing mask and the outlier mask, valid and invalid computational units are determined. Invalid computational units are excluded from the calculation range of Tucker decomposition error. The mean is removed and the standard deviation is normalized on the environmental covariate matrix to obtain the standardized environmental covariate matrix. After temporarily filling the invalid computational units, a third-order tensor to be decomposed is formed. The third-order tensor to be decomposed is expanded along the time dimension, mechanism dimension and quality feature dimension, respectively. Principal component decomposition is performed on each expanded matrix to obtain the time decomposition basis vector set, mechanism decomposition basis vector set and quality feature decomposition basis vector set. The third-order tensor to be decomposed is projected onto the subspaces corresponding to the time decomposition basis vector set, the mechanism decomposition basis vector set, and the quality feature decomposition basis vector set, respectively, to obtain the Tucker core tensor. The Tucker core tensor is then restored to the original tensor dimension to form a common reconstruction tensor. Within the effective computational unit, the difference between the third-order tensor of cross-mechanism quality state and the common reconstruction tensor is calculated to obtain the candidate residual tensor. For the residual time series corresponding to the same mechanism identifier and the same quality feature identifier, the environmental covariance residual series is calculated based on the standardized environmental covariance matrix. By subtracting the environmental covariant residual sequence from the residual time series, an environmental orthogonal residual sequence is obtained. Based on the median absolute value of the residuals of the environmental orthogonal residual sequence and the preset sparsity coefficient, mechanism-specific residual units are screened to generate mechanism-specific residual tensors. At the same time, the decomposition basis vector set of the Tucker tensor decoupling algorithm is recorded.

5. The method for real-time comparison of quality data from multiple institutions based on big data analysis according to claim 1, characterized in that, Step four specifically includes: Based on the mechanism-specific residual tensor, the mechanism residual vector corresponding to the same period and the same mechanism is extracted, and the mechanism residual vector, mechanism identifier, period identifier and quality feature identifier set are combined into residual node features; The residual vector of the mechanism is normalized to generate a residual direction sequence and a residual magnitude sequence. Candidate associated node pairs are determined based on the cosine similarity, residual direction consistency rate and residual magnitude difference rate between residual nodes. Candidate associated node pairs that have common residual nodes and the same set of quality feature identifiers participating in similarity calculation are merged into candidate superedges, and the weights of candidate superedges are calculated based on average cosine similarity, average residual direction consistency rate, and average residual magnitude proximity. Candidate hyperedges with weights not less than a preset hyperedge weight threshold are identified as valid hyperedges. The residual nodes, valid hyperedges, and hyperedge weights are then updated in a streaming manner according to the periodic order to generate a multi-mechanism residual streaming hypergraph.

6. The method for real-time comparison of quality data from multiple institutions based on big data analysis according to claim 1, characterized in that, The improved VGAE model includes a differential high-pass graph attention encoder, a variational coding layer, a compact class cohesion loss function, and a joint decoder of hyperedge features and topology. Based on the residual node features, effective hyperedges, and hyperedge weights in the multi-mechanism residual flow hypergraph, construct the node feature matrix, hypergraph association matrix, and hyperedge weight diagonal matrix; The node feature matrix is ​​input into the differential high-pass graph attention encoder. The first-order neighborhood node set is determined based on the hypergraph association matrix and the hyperedge weight diagonal matrix. The neighborhood aggregation feature is obtained through the graph attention network. Residual high-pass filtering is performed on the target residual node features and neighborhood aggregation features to obtain high-pass residual features. The high-pass residual features are then input into the variational coding layer, and the latent variables of the residual nodes are obtained through the mean mapping layer, the log-variance mapping layer, and reparameterized sampling. The compact Gaussian kernel center and compact Gaussian kernel covariance matrix are calculated based on the latent variables of normal residual nodes, and the compact class cohesion loss function is constructed by KL divergence loss, center loss and margin loss. The latent variables of residual nodes within the same effective hyperedge are subjected to bilinear decoding and covariance decoding by the joint decoder of hyperedge features and topology, resulting in the hyperedge topology reconstruction probability and the hyperedge feature covariance reconstruction matrix. The binary cross-entropy loss is calculated based on the hyperedge topological reconstruction probability, the Frobenius norm loss is calculated based on the reconstruction matrix of the feature covariance within the hyperedge, and the joint reconstruction error is obtained by weighting. The joint reconstruction error is used as the alignment score.

7. The method for real-time comparison of quality data from multiple institutions based on big data analysis according to claim 1, characterized in that, Step six specifically includes: Based on the historical periodic statistical comparison score distribution confirmed as normal by the quality inspection results during the calibration phase, and based on the quantile value corresponding to the preset upper bound quantile in the normal comparison score distribution, the preset comparison threshold is determined. When the alignment score of the current period is less than the preset alignment threshold, a normal alignment mark is generated. When the alignment score of the current period is not less than the preset alignment threshold, the Mahalanobis distance calculation of the latent variables of the current residual node is triggered. Based on the current residual node latent variable, the compact Gaussian kernel center, and the compact Gaussian kernel covariance matrix, calculate the offset vector of the current residual node latent variable relative to the compact Gaussian kernel center, and use the inverse matrix of the compact Gaussian kernel covariance matrix to perform covariance normalization on the offset vector to obtain the Mahalanobis distance. Based on the distribution of Mahalanobis distance of latent variables of normal residual nodes, a preset Mahalanobis distance threshold is determined. Residual nodes with Mahalanobis distance not less than the preset Mahalanobis distance threshold are identified as candidate root cause nodes, and the institution identifiers corresponding to the candidate root cause nodes are identified as root cause institution markers. Abnormal quality feature labels are determined based on the institutional residual vector and the distribution of the absolute value of normal residuals corresponding to candidate root cause nodes. Abnormal hyperedge labels are determined based on the hyperedge topology reconstruction error, hyperedge feature reconstruction error and hyperedge weight of the effective hyperedge to which the candidate root cause node belongs.

8. The method for real-time comparison of quality data from multiple institutions based on big data analysis according to claim 1, characterized in that, Step seven specifically includes: Based on the root cause mechanism marker, abnormal quality feature marker, abnormal over-edge marker and quality inspection results of the current cycle, the cycle in which the quality inspection result is qualified and there is a root cause mechanism marker, abnormal quality feature marker or abnormal over-edge marker is determined as a non-abnormal drift cycle. The residual node features corresponding to the root cause mechanism marker, abnormal quality feature marker, or abnormal super-edge marker in the non-abnormal drift cycle are determined as non-abnormal residual feature feedback, and the absorbable non-abnormal drift is determined based on the consistency of the residual direction and the change in the residual amplitude of the non-abnormal residual feature feedback. The third-order tensor fragments of cross-mechanism mass states corresponding to absorbable non-abnormal drifts are incorporated into the incremental decomposition samples. Incremental principal component decomposition is performed on the incremental decomposition samples to obtain the updated candidate decomposition basis vector set. The common reconstruction tensor and mechanism-specific residual tensor of non-abnormal drift cycles are recalculated using the updated candidate decomposition basis vector set. When the updated reconstruction error and environmental projection correlation quantities do not increase, the updated candidate decomposition basis vector set is determined as the decomposition basis vector set updated by the Tucker tensor decoupling algorithm. Hyperedge stability is generated based on the co-occurrence relationship of residual nodes that recur in the feedback of non-abnormal residual features. The hyperedge weights of the multi-mechanism residual streaming hypergraph are updated according to the hyperedge stability and the preset decay coefficient. The updated set of decomposed basis vectors and hyperedge weights are used for the environmental covariance component stripping and multi-mechanism residual streaming hypergraph construction in the next cycle.

Images