A defect detection system and method for a pressure sensor chip based on big data

Abstract

The application relates to a big data-based defect detection system and method for a pressure sensor chip. The method comprises the following steps: constructing a graph of a multi-dimensional correlation relationship of a to-be-tested pressure sensor chip according to space-time process correlation data and test response data in a chip manufacturing detection data set, and obtaining a heterogeneous manufacturing graph; performing multi-dimensional correlation feature analysis according to the chip manufacturing detection data set and the heterogeneous manufacturing graph, and obtaining a chip comprehensive representation vector; performing group deviation analysis according to a reference group distribution model corresponding to the to-be-tested pressure sensor chip and the chip comprehensive representation vector, and obtaining a deviation representation result; performing multi-domain damage analysis according to the deviation representation result, and obtaining multi-domain damage data; and performing instability defect analysis according to the multi-domain damage data, and obtaining defect detection data. The method can improve the defect detection efficiency of the pressure sensor chip.

Description

Technical Field

[0001] This application relates to the field of big data processing technology, and in particular to a defect detection system and method for a pressure sensor chip based on big data. Background Technology

[0002] In traditional technologies, defect detection of pressure sensor chips typically involves acquiring chip images using industrial cameras or microscopes after chip manufacturing or packaging. This is followed by template matching and feature comparison to identify cracks, contamination, scratches, chipping, electrode abnormalities, and packaging defects. Electrical performance testing is then used to verify the chip's functional integrity, thus achieving a comprehensive screening of both appearance and performance defects in the pressure sensor chip. However, traditional technologies are poorly adaptable to complex, minute, and diverse defects, have limited robustness and automation, and rely heavily on human experience and pre-defined rules, making them prone to missed and false detections, resulting in low efficiency in pressure sensor chip defect detection. Summary of the Invention

[0003] Therefore, it is necessary to provide a big data-based defect detection system and method for pressure sensor chips that can improve the efficiency of defect detection for pressure sensor chips, addressing the aforementioned technical problems.

[0004] Firstly, this application provides a defect detection method for pressure sensor chips based on big data, including: Based on the spatiotemporal process correlation data and test response data in the chip manufacturing test dataset corresponding to the pressure sensor chip under test, a multidimensional correlation map of the pressure sensor chip under test in the manufacturing chain is constructed to obtain a heterogeneous manufacturing map. Based on the chip manufacturing test dataset and the heterogeneous manufacturing map, a multidimensional correlation feature analysis is performed on the pressure sensor chip under test to obtain a chip comprehensive characterization vector. Based on the reference population distribution model corresponding to the pressure sensor chip under test and the chip comprehensive characterization vector, a population deviation analysis is performed on the pressure sensor chip under test to obtain the deviation characterization results. Based on the deviation characterization results, multi-domain damage analysis is performed on the pressure sensor chip under test to obtain multi-domain damage data. Based on the multi-domain damage data, instability defect analysis is performed on the pressure sensor chip under test to obtain defect detection data.

[0005] Secondly, this application also provides a defect detection system for pressure sensor chips based on big data, the system including: a server and a terminal; The server is used to construct a graph of the multidimensional relationship between the pressure sensor chip under test in the manufacturing chain based on the spatiotemporal process correlation data and test response data in the chip manufacturing inspection dataset corresponding to the pressure sensor chip under test, thereby obtaining a heterogeneous manufacturing graph; the terminal is used to acquire the chip manufacturing inspection dataset. The server is used to perform multi-dimensional correlation feature analysis on the pressure sensor chip under test based on the chip manufacturing test dataset and the heterogeneous manufacturing map to obtain a chip comprehensive characterization vector. The server is used to perform population deviation analysis on the pressure sensor chip under test based on the reference population distribution model corresponding to the pressure sensor chip under test and the chip comprehensive characterization vector, and to obtain deviation characterization results. The server is used to perform multi-domain damage analysis on the pressure sensor chip under test based on the deviation characterization results, and obtain multi-domain damage data. The server is used to perform instability and defect analysis on the pressure sensor chip under test based on the multi-domain destruction data, and obtain defect detection data.

[0006] The aforementioned defect detection system and method for pressure sensor chips based on big data, by using the chip manufacturing inspection dataset corresponding to the pressure sensor chip under test as a basis, jointly models spatiotemporal process correlation data and test response data to construct a heterogeneous manufacturing map. This unified correlation characterizes the batch affiliation, wafer location, equipment experience, process flow, time evolution, and test response relationships of the pressure sensor chip under test in the manufacturing chain, thus overcoming the limitations of existing technologies that only target local characterization of a single chip for defect judgment. Furthermore, multidimensional correlation feature analysis is used to construct a comprehensive chip characterization vector, which is then combined with a reference population distribution model for further analysis. Population deviation analysis can identify abnormal deviations of the pressure sensor chip under test relative to the reference population from the perspective of population consistency, thereby improving the detection capability of hidden defects, early defects, and non-obvious defects. Further, multi-domain destructive analysis is performed on the deviation characterization results to reveal cross-domain propagation relationships and coupling imbalances between the process domain, equipment domain, time domain, spatial domain, and test response domain, thereby improving the identification accuracy of complex coupled defects and propagating defects. Finally, instability defect analysis is used to obtain structured defect detection data, which can not only output defect categories and defect severity, but also quantify defect risk, propagation activity, and instability sources. Overall, this approach provides more comprehensive detection dimensions, more refined detection results, and stronger foresight regarding potential defects in pressure sensor chips, thus improving the efficiency of defect detection for pressure sensor chips. Attached Figure Description

[0007] To more clearly illustrate the technical solutions in the embodiments or related technologies of this application, the accompanying drawings used in the description of the embodiments or related technologies will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0008] Figure 1 This is an application environment diagram of a defect detection method for a pressure sensor chip based on big data in one embodiment; Figure 2 This is a flowchart illustrating a defect detection method for a pressure sensor chip based on big data in one embodiment. Figure 3 This is an internal structural diagram of a computer device in one embodiment. Detailed Implementation

[0009] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0010] This application provides a defect detection method for pressure sensor chips based on big data, which can be applied to, for example... Figure 1 In the application environment shown, terminal 102 communicates with server 104 via a network. A data storage system can store the data that server 104 needs to process. The data storage system can be integrated onto server 104, or it can be located in the cloud or on other network servers. Server 104 can be implemented using a standalone server or a server cluster consisting of multiple servers.

[0011] In one exemplary embodiment, such as Figure 2 As shown, a defect detection method for pressure sensor chips based on big data is provided, which can be applied to... Figure 1 Taking the server in the example, the explanation includes the following steps 202 to 210. Wherein:

[0012] Step 202: Based on the spatiotemporal process correlation data and test response data in the chip manufacturing test dataset corresponding to the pressure sensor chip under test, construct a map of the multidimensional correlation of the pressure sensor chip under test in the manufacturing chain to obtain a heterogeneous manufacturing map.

[0013] Step 204: Based on the chip manufacturing inspection dataset and heterogeneous manufacturing map, perform multi-dimensional correlation feature analysis on the pressure sensor chip under test to obtain the chip comprehensive characterization vector.

[0014] Step 206: Based on the reference population distribution model and chip comprehensive characterization vector corresponding to the pressure sensor chip under test, perform population deviation analysis on the pressure sensor chip under test to obtain deviation characterization results.

[0015] Step 208: Based on the deviation characterization results, perform multi-domain damage analysis on the pressure sensor chip under test to obtain multi-domain damage data.

[0016] Step 210: Based on the multi-domain damage data, perform instability defect analysis on the pressure sensor chip under test to obtain defect detection data.

[0017] Among them, the pressure sensor chip under test refers to the pressure sensor chip that has entered the defect detection process and needs to be judged to have abnormal defects based on its manufacturing process information and test results.

[0018] Among them, the chip manufacturing test dataset refers to the complete set of data formed and organized in relation to the pressure sensor chip under test during the manufacturing, circulation, testing and environmental processes.

[0019] Among them, spatiotemporal process correlation data refers to data that can simultaneously reflect the process steps, equipment experiences, time sequence, and spatial location relationship of the pressure sensor chip under test in the manufacturing chain.

[0020] Among them, test response data refers to the electrical, timing, or dynamic response results output by the pressure sensor chip under test under different test conditions, environmental conditions, or excitation conditions.

[0021] Among them, multidimensional correlation refers to the various types of connection and coupling relationships formed between the pressure sensor chip under test and batch, equipment, process, time, location and test status.

[0022] Among them, graph construction refers to the process of establishing an associated network structure by organizing the multi-source data related to the pressure sensor chip under test according to the organization of nodes and edges.

[0023] Among them, heterogeneous manufacturing graph refers to a graph-structured manufacturing association network that includes different types of nodes, different types of connecting edges, and multi-layered manufacturing relationships.

[0024] Among them, multidimensional correlation feature analysis refers to the process of jointly analyzing the pressure sensor chip under test in multiple dimensions such as response, process, neighborhood and topology by combining manufacturing inspection datasets and heterogeneous manufacturing maps.

[0025] Among them, the chip comprehensive characterization vector refers to the vectorized result formed by uniformly encoding and fusing the associated features of the pressure sensor chip under test in multiple dimensions.

[0026] Among them, the reference group distribution model refers to a model that is pre-established based on the distribution pattern of historical chip groups in the characterization space and is used to compare the pressure sensor chips under test.

[0027] Among them, population deviation analysis refers to the analysis process of mapping the pressure sensor chip under test to a reference population distribution model and determining its positional offset relative to the reference population.

[0028] Among them, the deviation characterization result refers to the result obtained after the population deviation analysis, which describes the deviation direction, degree of deviation and deviation morphology of the pressure sensor chip under test relative to the reference population.

[0029] Among them, multi-domain destruction analysis refers to the process of jointly analyzing the propagation relationship, coupling relationship and imbalance relationship of the pressure sensor chip under test in multiple detection domains based on the deviation characterization results.

[0030] Among them, multi-domain destruction data refers to data obtained after multi-domain destruction analysis, which reflects the destruction intensity, propagation status and coupling imbalance between different detection domains.

[0031] Among them, instability defect analysis refers to the analytical process of comprehensively judging the defect mode, defect level and defect chain relationship of the pressure sensor chip under test based on multi-domain damage data.

[0032] Among them, defect detection data refers to the structured result data output after instability defect analysis, which is used to represent the defect status and detection conclusion of the pressure sensor chip under test.

[0033] Specifically, spatiotemporal process correlation data and test response data are extracted from the chip manufacturing inspection dataset. The spatiotemporal process correlation data includes the batch number, wafer number, wafer coordinate position, processing equipment number, process step identifier, process switch sequence, start and end times of each process step, process dwell time, and process correlation information of adjacent chips corresponding to the pressure sensor chip under test during the manufacturing process. The test response data includes the bridge output data, sensitivity data, zero-point offset data, temperature drift data, noise response data, dynamic response curve data, and environmental stimulus response data of the pressure sensor chip under test under different test conditions. Using the unique identifier of the pressure sensor chip under test as the main index, the spatiotemporal process correlation data and test response data are uniformly linked and expanded according to the correlation link of "chip—batch—wafer position—equipment—process step—time—test response" to map data from different sources and dimensions into uniformly computable node and edge data. Based on batch affiliation, wafer spatial proximity, equipment sharing, process sequence, temporal continuity, and test response similarity, we construct affiliation edges between chip nodes and batch nodes, spatial edges between chip nodes and location nodes, processing edges between chip nodes and equipment nodes, flow edges between chip nodes and process nodes, timing edges between chip nodes and time nodes, and response edges between chip nodes and test nodes. We assign corresponding association weights to each type of edge to characterize the degree of influence of different relationships on chip manufacturing behavior. We then heterogeneously integrate and organize these nodes and edges to form a heterogeneous manufacturing atlas that simultaneously includes chip entity relationships, process flow relationships, spatiotemporal distribution relationships, and test response relationships.

[0034] The test response data extracts fundamental response features characterizing the individual performance state of the pressure sensor chip under test, including at least static output features, dynamic response features, temperature drift features, noise fluctuation features, nonlinear variation features, and response stability features under multiple test conditions. Simultaneously, the spatiotemporal process correlation data extracts process correlation features characterizing the manufacturing process of the pressure sensor chip under test, including at least process path sequence features, equipment switching features, process dwell time features, wafer location distribution features, batch affiliation features, and adjacent chip co-change features. Combining the node relationships directly and indirectly connected to the pressure sensor chip under test in the heterogeneous manufacturing map, graph correlation mining is performed on the pressure sensor chip under test to obtain neighborhood consistency features, cross-node propagation features, relationship link features, and topological embedding features reflecting the strength of its correlation with chips in the same batch, chips on the same equipment, chips in adjacent locations, and chips on the same process path. Furthermore, the aforementioned basic response features, process correlation features, and graph correlation features are uniformly encoded, dimensionally normalized, weighted, and correlated and fused, so that data from different sources, scales, and structures are mapped to the same representation space. By strengthening the abnormal sensitivity information and group deviation information of the pressure sensor chip under test in the manufacturing chain through the coupling relationship between features, a chip comprehensive representation vector is obtained that comprehensively reflects the individual response state, manufacturing process behavior, neighborhood group relationship, and graph topological position of the pressure sensor chip under test.

[0035] A pre-built reference population distribution model is invoked, and the chip comprehensive representation vector is input into the reference population distribution model. During the operation of the reference population distribution model, based on the pre-divided stable clustering region, edge transition region, and abnormally sparse region, a back-transmission process is performed on the chip comprehensive representation vector. This process re-transmits the individual response features, process correlation features, and map topology features in the representation vector to the population representation space corresponding to the reference population distribution model according to a preset mapping rule, obtaining population mapping trigger points corresponding to different reference population sub-regions. Then, using the population mapping trigger points and the reference population distribution model as the starting point, within the population representation space, starting from the population mapping trigger points, the mapping trajectory of the pressure sensor chip under test is mirrored, unfolded, and sieved along multiple reference distribution directions to identify its deviation path relative to the stable clustering region, the degree of boundary crossing relative to the edge transition region, and the contact state relative to the abnormally sparse region, thereby obtaining a deviation candidate bundle formed by the combination of multiple deviation segments. Finally, the deviation direction, deviation magnitude, deviation level, deviation density, and deviation clustering pattern implicit in the population representation space of the deviation candidate bundle are converted into a quantifiable set of deviation parameters. The deviation parameter set is then combined and output to obtain the deviation representation result.

[0036] Deviation parameters corresponding to different detection domains are extracted from the deviation characterization results. Detection domains can include process domains, equipment domains, time domains, spatial domains, test response domains, and group relationship domains. The deviation direction, deviation magnitude, deviation density, deviation level, and deviation aggregation state in each detection domain are decomposed and encoded to form deviation data for each domain. Cross-domain reverse ordering processing is performed on the deviation data of each domain to break the unidirectional limitation of the original process order, test order, or time order on the judgment of anomaly propagation. Deviation domains that appear later but have high correlation are prioritized to obtain a domain perturbation evolution sequence that reflects the reverse traction relationship of multi-domain perturbations. According to the inter-domain correlation strength, node co-occurrence frequency, and perturbation similarity, multiple similar perturbation nodes in the domain perturbation evolution sequence are compressed into domain perturbation folds. Transition link search is performed on the domain perturbation folds in combination with heterogeneous manufacturing maps to obtain inter-domain contagion link data that characterizes the cross-domain propagation path of perturbations between different detection domains. Based on this, the propagation weights, backtracking influence weights, and cross-domain traction weights of subsequent link nodes in the inter-domain contagion link data are reassigned to preceding link nodes to obtain inter-domain coupling strength data. Then, the original heterogeneous manufacturing map is reconstructed based on this data to generate a domain destruction propagation map characterizing multi-domain disturbance diffusion relationships and inter-domain coupling destruction relationships. Finally, the propagation strength, propagation span, destruction depth, cross-domain coupling density, and local imbalance degree in the domain destruction propagation map are aggregated to output multi-domain destruction data.

[0037] High-damage components characterizing multi-domain abnormal linkage states are extracted from multi-domain damage data. These high-damage components can include cross-domain propagation intensity components, local imbalance density components, inter-domain coupling anomaly components, propagation span components, and damage depth components. These high-damage components are then folded and merged according to their co-occurrence, coupling, and evolution direction relationships to obtain instability source clusters corresponding to different instability sources. Using these instability source clusters as the guiding core, each cluster is correlated and mapped with pattern features in a pre-defined defect pattern library to identify candidate defect patterns in different defect directions of the pressure sensor chip under test, such as bridge imbalance, abnormal temperature drift, noise runaway, process disturbance propagation, packaging coupling anomalies, long-term drift risk, or response mismatch, resulting in a candidate defect pattern set. Furthermore, based on the instability propagation path length, peak coupling intensity, abnormal clustering degree, and cross-domain injection weight of each candidate defect pattern, the hazard levels of each candidate defect pattern are reordered and reversed to obtain a defect severity sequence reflecting the priority and risk sequence of defect evolution. Based on this, a defect association analysis is performed on the pressure sensor chip under test according to the defect severity sequence and candidate defect pattern set. This analysis associates different defect patterns with their corresponding damage domains, propagation chains, instability source clusters, and risk levels, yielding defect association results. Finally, the defect category, defect severity, defect risk value, defect propagation activity, and comprehensive detection conclusions from the defect association results are converted into structured numerical outputs to obtain defect detection data.

[0038] The aforementioned defect detection method for pressure sensor chips based on big data, by using the chip manufacturing inspection dataset corresponding to the pressure sensor chip under test as a basis, jointly models spatiotemporal process correlation data and test response data to construct a heterogeneous manufacturing map. This unified correlation characterizes the batch affiliation, wafer location, equipment experience, process flow, time evolution, and test response relationships of the pressure sensor chip under test in the manufacturing chain, thus overcoming the limitation of existing technologies that only target the local characterization of a single chip for defect judgment. Furthermore, multidimensional correlation feature analysis is used to construct a comprehensive chip characterization vector, which is then combined with a reference population distribution model for group analysis. Deviation analysis identifies anomalous offsets of the pressure sensor chip under test relative to a reference group from a population consistency perspective, thereby improving the detection capabilities for latent, early, and inconspicuous defects. Further, multi-domain destructive analysis of the deviation characterization results reveals cross-domain propagation and coupling imbalance relationships between the process domain, equipment domain, time domain, spatial domain, and test response domain, thus improving the identification accuracy of complex coupled and propagating defects. Finally, instability defect analysis yields structured defect detection data, outputting not only defect categories and severity but also quantifying defect risk, propagation activity, and instability sources. Overall, this approach provides more comprehensive detection dimensions, more refined results, and stronger foresight regarding potential defects in pressure sensor chips, thereby improving the efficiency of defect detection for pressure sensor chips.

[0039] In an exemplary embodiment, based on the deviation characterization results, a multi-domain damage analysis is performed on the pressure sensor chip under test to obtain multi-domain damage data, including steps 302 to 306. Wherein:

[0040] Step 302: The deviation data of each domain in the deviation characterization results are arranged in reverse order across domains to obtain the domain perturbation evolution sequence.

[0041] Step 304: Based on the domain perturbation evolution sequence and heterogeneous manufacturing map, perform inter-domain coupling contagion analysis on the pressure sensor chip under test to obtain the domain destruction propagation map.

[0042] Step 306: Perform imbalance aggregation on the domain destruction propagation graph to obtain multi-domain destruction data.

[0043] Among them, domain deviation data refers to the domain-specific data formed by splitting the deviation characterization results into different detection domains, each corresponding to the deviation state of each detection domain.

[0044] Cross-domain reverse order arrangement refers to the process of reordering the data deviating from each domain according to the backtracking traction relationship and cross-domain influence relationship between detection domains, rather than according to the original process order or time order.

[0045] Among them, the domain perturbation evolution sequence refers to the ordered sequence formed after cross-domain reverse arrangement, which can reflect the sequential connection and propagation direction of perturbations in different detection domains.

[0046] Inter-domain coupling contagion analysis refers to the process of jointly analyzing the perturbation transmission and coupling diffusion relationships between different detection domains by combining the domain perturbation evolution sequence and heterogeneous manufacturing map.

[0047] Among them, the domain destruction propagation graph refers to the graph structure result formed after inter-domain coupling infection analysis, which is used to describe the destruction propagation path, propagation direction and propagation intensity between different detection domains.

[0048] Among them, imbalance aggregation refers to the process of converging and calculating various propagation relationships and destruction relationships in the domain destruction propagation graph to form an overall multi-domain destruction result.

[0049] Specifically, the deviation data in the deviation characterization results are binned according to the process domain, equipment domain, time domain, spatial domain, test response domain, and group relationship domain. Within each detection domain, deviation direction value, deviation amplitude value, deviation level value, deviation density value, and deviation clustering location value are written according to a unified field format. Then, for any two detection domains, the number of times deviation events occur simultaneously within the same statistical window is counted to obtain the inter-domain co-occurrence frequency. Next, the number of times they occur synchronously within the same batch, the number of times they occur adjacently within the same equipment time window, the number of times they occur in conjunction at adjacent positions on the wafer, and the number of times the test response change direction is consistent are counted. The above count results are normalized and then weighted and summed according to preset weights to obtain the correlation strength between the two detection domains. After calculating the inter-domain correlation strength, following the rule of "prioritizing detection domains that appear later but have higher traction strength, and placing detection domains that appear earlier but have weaker diffusion capabilities later," for each detection domain, the number of effective connections that form reverse traction with other detection domains is first counted, and then weighted according to the reciprocal of the connection span to obtain the backtracking traction value. Next, the total number of detection domain switches in the effective propagation chain is counted to obtain the cross-domain jump count. Furthermore, backflow links containing the detection domain are identified, and abnormal backflow weights are accumulated according to the backflow edge weight and link length. Finally, the backtracking traction value, cross-domain jump count, and abnormal backflow weight are normalized and weighted summed to obtain the reverse priority of the detection domain. Using the reverse priority as the primary sorting key and the inter-domain correlation strength as the secondary sorting key, deviation items in multiple detection domains are sequentially concatenated, and cross-domain transfer markers, jump distance parameters, and transfer direction parameters are written between adjacent deviation items. Finally, the concatenated results are subjected to continuity correction and loop elimination. When repeated insertions occur in the same detection domain, deviation terms with higher reverse priority and stronger correlation are retained, and redundant transition segments are removed, thereby forming a domain perturbation evolution sequence with cross-domain order, transition path and perturbation sequence.

[0050] Each perturbation term in the domain perturbation evolution sequence is mapped to a corresponding node combination in the heterogeneous manufacturing graph, ensuring that each perturbation term is simultaneously associated with at least two nodes among chip nodes, device nodes, process nodes, time nodes, location nodes, and test nodes. Based on the adjacent transition relationships in the domain perturbation evolution sequence, a connectable path is searched in the heterogeneous manufacturing graph. Instead of using the shortest path rule, candidate links with more cross-layer jumps, more shared device windows, more concentrated spatial adjacency ranges, and shorter time backtracking spans are prioritized. For each candidate link, the number of times each node on the link co-occurs with its adjacent nodes in the same batch, the same device window, or the same wafer adjacency region is counted, and the co-occurrence density of each node is calculated. Then, the edge weights between adjacent nodes on the link are read and multiplied to obtain the edge weight product value. Furthermore, the number of times a node in the link is repeatedly traversed is counted and corrected according to a preset penalty coefficient. The average co-occurrence density of each node is multiplied by the edge weight product value, and combined with the correction result for the number of repeated traversals, to obtain the link coupling value. Similarly, for each candidate link, the order difference between its initial and final perturbation terms in the domain perturbation evolution sequence is first calculated to obtain the order distance, which is then converted into an order transfer factor. Next, the number of times the propagation direction reversals occur in the link is counted, and the link backflow value is obtained by combining the reversal edge weights. Furthermore, the proportion of consistent directions in each cross-domain transition of the link is calculated to obtain the cross-domain transition direction consistency. The order transfer factor, link backflow value, and cross-domain transition direction consistency are then normalized and weighted to obtain the link transfer value. Candidate links are screened according to their link coupling value and link transfer value, retaining only links that simultaneously meet the minimum coupling threshold and minimum transfer threshold. The retained links are then organized into propagation edges according to the initial perturbation term, transition node group, and final perturbation term. After forming propagation edges, propagation edges with common relay nodes are spliced ​​together, propagation edges with opposite propagation directions but sharing most nodes are merged backflow, and propagation edges with excessively short propagation spans and lacking cross-domain characteristics are pruned. Finally, using the perturbation term as the graph node and the spliced ​​and merged propagation edge as the graph edge, the edge weight of each propagation edge is set as a combination of the link coupling value, the link transmission value and the return weight, to obtain the domain destruction propagation graph.

[0051] In the domain disruption propagation graph, the total weight of incoming edges and the total weight of outgoing edges are calculated for each node according to the direction of the propagation edges. The edge weight difference between adjacent nodes, the weight dispersion between nodes in the same layer, and the backflow ratio between nodes across layers are also calculated. The difference between the total weight of incoming edges and the total weight of outgoing edges is used as the local imbalance base value, the cumulative result of the edge weight difference is used as the link imbalance base value, and the weight dispersion of nodes in the same layer and the backflow ratio of nodes across layers are used as extended imbalance correction values. Imbalance metrics are formed for each node, each propagation edge, and each propagation chain. Then, taking the detection domain as a unit, the node imbalance value, edge imbalance value, and link imbalance value belonging to the same detection domain are combined into an intra-domain imbalance value according to a preset ratio. The intra-domain imbalance value is then superimposed twice according to the inter-domain coupling relationship. A coupling gain is added to multiple detection domains with shared relay nodes, and a reduction coefficient is applied to detection domains with mutually canceling propagation directions. After completing cross-domain aggregation, all aggregation results are normalized and hierarchically expanded to obtain the total damage value, domain damage value, link damage value, propagation activity value, coupling imbalance value, and return flow imbalance value. Finally, the above values ​​are encapsulated according to a unified data structure to generate multi-domain damage data.

[0052] In this embodiment, by first reversing the cross-domain deviation data, then performing inter-domain coupling contagion analysis in conjunction with the heterogeneous manufacturing map, and finally aggregating the domain destruction propagation map, the method is no longer limited to linear judgment of anomalies according to the original process sequence, test sequence, or time sequence. Instead, it can incorporate the backtracking traction relationship, cross-domain jump relationship, and propagation diffusion relationship between different detection domains into the same processing link. This organizes the originally scattered anomalies within the domains into a destruction network with propagation direction and coupling hierarchy, making it easier to reveal the linkage destruction process between multiple detection domains and the evolution path of local anomalies spreading to global anomalies. At the same time, the obtained multi-domain destruction data not only reflects the anomaly degree of a single detection domain, but also reflects the imbalance aggregation state and propagation activity state between detection domains, thereby improving the ability to capture complex coupled defects and chain diffusion defects.

[0053] In an exemplary embodiment, based on the domain perturbation evolution sequence and heterogeneous manufacturing map, inter-domain coupling contagion analysis is performed on the pressure sensor chip under test to obtain a domain destruction propagation map, including steps 402 to 408. Wherein:

[0054] Step 402: Perform cross-layer compression mapping on the domain perturbation nodes in the domain perturbation evolution sequence to obtain the domain perturbation fold.

[0055] Step 404: Based on the domain perturbation fold body and heterogeneous manufacturing map, perform a transition link search on the pressure sensor chip under test to obtain inter-domain contagion link data.

[0056] Step 406: Perform reverse energy injection on the inter-domain infection link data to obtain the inter-domain coupling strength field.

[0057] Step 408: Based on the inter-domain coupling strength field, the graph structure of the pressure sensor chip under test is reconstructed to obtain the domain destruction propagation graph.

[0058] Among them, the domain perturbation node refers to the node unit in the domain perturbation evolution sequence that corresponds to the specific perturbation location and perturbation state within a certain detection domain.

[0059] Cross-layer compression mapping refers to the process of compressing and mapping multiple domain perturbation nodes that are distributed at different detection levels but have close relationships into a unified aggregation unit.

[0060] Among them, the domain perturbation fold body refers to the aggregate structure formed after cross-layer compression mapping, which contains multiple related domain perturbation nodes and their compression connection relationships.

[0061] Among them, the transition link search refers to the process of searching for connectable propagation paths between different detection domains in the heterogeneous manufacturing map, starting from the domain perturbation fold.

[0062] Inter-domain transmission link data refers to data formed after the jump link search, which records the transmission links and their connection relationships between different detection domains.

[0063] Among them, reverse energy reinjection refers to the process of redistributing and retrospectively transmitting the propagation intensity in the reverse direction of the inter-domain infection link.

[0064] Among them, the inter-domain coupling strength field refers to the field result formed after reverse energy back-injection, which reflects the distribution of coupling strength between different detection domains.

[0065] Among them, graph structure reconstruction refers to the process of reorganizing the original graph connection relationships, edge weight relationships, and node hierarchy relationships based on the inter-domain coupling strength field.

[0066] Specifically, each domain perturbation node in the domain perturbation evolution sequence is numbered and registered according to its detection domain, occurrence order, cross-domain transition direction, perturbation amplitude range, and connection relationship with chip nodes, device nodes, process nodes, time nodes, and location nodes in the heterogeneous manufacturing map. Then, using adjacent order windows as units, domain perturbation nodes in different detection domains but sharing the same device node, adjacent time node, or adjacent location node are paired side-by-side. The shortest connection step between two domain perturbation nodes in the heterogeneous manufacturing map is calculated and converted into a graph distance factor. Then, the order number difference between the two domain perturbation nodes in the domain perturbation evolution sequence is calculated and converted into an order proximity factor. The edge weight similarity factor is calculated using the edge weights of the connecting edges of the two domain perturbation nodes and the absolute value of the difference between their edge weights. The cross-domain transition directions of the two domain perturbation nodes are compared. If the transition directions are the same, the first direction value is used; if the transition directions are opposite, the second direction value is used; if the transition directions are partially consistent, the third direction value is used. This yields the transition direction factor. Finally, the graph distance factor, order proximity factor, edge weight similarity factor, and transition direction factor are normalized and weighted to obtain the cross-layer proximity. When the cross-layer proximity reaches a preset compression threshold, the corresponding domain perturbation nodes are merged into the same compression unit. Within the compression unit, the domain label, original order number, and edge connection index of each original node are retained. A center mapping is performed on the nodes within the same compression unit. The connecting edges of this compression unit in the heterogeneous manufacturing graph are rearranged according to their edge weights, with the connection group with the largest cumulative edge weight value becoming the primary connection group; the remaining connection groups become secondary connection groups. Simultaneously, the output order of the compression unit is reordered using a "first-order priority, strongest edge weight correction" method. Finally, each compression unit that has completed center mapping and order reordering is uniformly encapsulated to generate a domain perturbation fold that simultaneously retains the multi-detection domain co-connection relationship, the original perturbation order relationship, and the spectrum connection relationship.

[0067] For each domain disturbance fold, a search entry point is established according to the device node, process node, time node, and location node corresponding to its main connection group. A one-to-one initial correspondence is established between each search entry point and the current position of the pressure sensor chip under test in the heterogeneous manufacturing map. Starting from each search entry point, a multi-path parallel traversal is performed along the traversable edges between the five types of nodes in the heterogeneous manufacturing map: "device—process—time—location—test". During the traversal, a single shortest path rule is not used; instead, candidate paths with no more than a preset upper limit for cross-layer switching, an edge weight decay rate lower than a preset threshold, and no directional conflict between adjacent steps are prioritized. For each candidate path, the node sequence, edge weight sequence, number of cross-domain switching, number of back jumps, and the location of the closed loop are recorded step by step. Segments that continuously pass through shared relay nodes are spliced ​​together, and segments that repeatedly pass through the same nodes but include different detection domain nodes are retained in parallel. Then, all candidate paths are screened, with cumulative edge weight, cross-domain span, bounce control value, and loop integrity as the sorting criteria. Paths with cumulative edge weight below the threshold, insufficient cross-domain span, or broken loops are deleted. The remaining paths are organized into inter-domain infection links in the form of "starting fold body - relay node group - ending fold body". Finally, the start position, end position, path length, path edge weight, and cross-domain switching sequence of each inter-domain infection link are written into a unified link structure to form inter-domain infection link data.

[0068] The starting point for back-injection at the end of each inter-domain infection link is determined based on the terminating fold, path length, and cumulative edge weight. The basic back-injection value at the end of each link is set as a combination of the cumulative edge weight and the closed-loop integrity of that link. Back-injection is then distributed progressively from the end to the beginning of the link node sequence. During each back-injection distribution, the back-injection value received by the current node is split into a main back-injection amount and a bypass back-injection amount according to a preset ratio. The main back-injection amount is transmitted in reverse along the preceding edge of the original link, while the bypass back-injection amount is transmitted along the side edges in the heterogeneous manufacturing graph that share equipment, time windows, or location neighborhoods with the current node. For each reverse-transmission edge, a back-injection attenuation coefficient is calculated based on the edge weight, edge type priority, and the number of times the node is traversed repeatedly. This attenuation coefficient is then used to correct the main back-injection amount and the bypass back-injection amount. When the same node receives backfeed values ​​from multiple links or side edges simultaneously, the values ​​are first superimposed, and then truncated according to the backflow threshold and saturation upper limit of the detection domain where the node is located to prevent the backfeed values ​​of local nodes from being infinitely amplified. Then, the backfeed results of all nodes that have completed the reverse transmission of the entire link are summarized, and the backfeed values ​​of the equipment layer, process layer, time layer, location layer, and test layer are registered according to the hierarchical position of the node in the heterogeneous manufacturing map. Finally, with the node as the center and the backfeeded edges as the connection basis, a field distribution structure containing the backfeed intensity of each layer of nodes and the reverse transmission intensity between edges is constructed to obtain the inter-domain coupling strength field.

[0069] Based on the injection values ​​of nodes in each layer of the inter-domain coupling strength field, the original nodes in the heterogeneous manufacturing graph are reweighted. Nodes with injection values ​​higher than a first threshold are designated as reinforcement nodes, nodes with injection values ​​between the first and second thresholds are designated as transition nodes, and nodes with injection values ​​lower than the second threshold retain their original weights or undergo weakening. Then, the edge connecting two reinforcement nodes in the original graph is weighted and marked as a main propagation edge, the edge connecting a reinforcement node and a transition node is marked as a diffusion edge, and the edge connecting two low-weight nodes without effective injection transmission is pruned or downweighted. Furthermore, the local propagation subgraph is reorganized around the chip node containing the pressure sensor chip under test. Nodes in the first layer directly connected to this chip node via the main propagation edge are designated as first-level propagation circles, and nodes connected via the first-level propagation circles and further via the main propagation edge or diffusion edge are designated as second-level propagation circles. The cross-layer connections between propagation circles are reoriented according to the injection strength and path closure degree. Based on this, the repeated propagation segments in the graph are merged, the propagation segments with opposite directions but sharing relay nodes are backflowed and integrated, the broken propagation segments with spacing less than the preset connection threshold are reconnected, and finally the reconstructed graph structure is output as the domain destruction propagation graph.

[0070] In this embodiment, by first performing cross-layer compression mapping on the domain disturbance nodes, then conducting transition link search based on the heterogeneous manufacturing map, and performing reverse energy injection at the link level for graph structure reconstruction, the system no longer focuses on local judgment of a single disturbance node or a single propagation edge. Instead, it compresses the disturbance relationships scattered across different detection levels into a continuously traceable propagation skeleton, enabling indirect associations across devices, processes, time, and spatial locations to be connected to the same propagation network. On this basis, the inter-domain coupling strength field formed by reverse energy injection can redistribute the feedback influence of subsequent propagation segments on preceding propagation segments to the entire propagation link. As a result, the final domain destruction propagation map can not only reflect the forward diffusion trajectory of destruction propagation, but also reflect deep propagation characteristics such as propagation backflow, link coupling, and relay node amplification, which is conducive to more accurately distinguishing between isolated disturbances and chain-like contagion disturbances.

[0071] In an exemplary embodiment, based on the domain perturbation fold and heterogeneous manufacturing map, a transition link search is performed on the pressure sensor chip under test to obtain inter-domain contagion link data, including steps 502 to 508. Wherein:

[0072] Step 502: Prune the entry points of folded nodes in the domain perturbation fold that have perturbation strengths higher than a preset threshold to obtain a set of candidate transition sources.

[0073] Step 504: Based on the candidate transition source set and heterogeneous manufacturing map, perform a misaligned projection search on the pressure sensor chip under test to obtain cross-domain transition candidate segments.

[0074] Step 506: Perform a back-loop closure check on the cross-domain transition candidate links to obtain the set of valid transition links.

[0075] Step 508: Perform link convergence processing on the effective transition link set to obtain inter-domain infection link data.

[0076] Among them, disturbance intensity refers to the value used to measure the activity and influence of a certain folded node during the multi-domain disturbance propagation process.

[0077] Among them, the preset threshold refers to the pre-set judgment limit used to screen whether the disturbance intensity has reached the condition for entering the subsequent transition analysis.

[0078] Among them, folded nodes refer to node units that aggregate multiple associated domain perturbation nodes after cross-layer compression mapping.

[0079] Among them, entry pruning refers to the process of filtering and retaining the connection entry and exit of folded nodes in order to narrow down the subsequent transition search range.

[0080] Among them, the candidate transition source set refers to the set of folded nodes that are retained after entry pruning and can be used as the starting point for transition link search.

[0081] Among them, misaligned projection search refers to the process of searching for potential propagation paths by offsetting across layers instead of directly extending along the original same-layer connections.

[0082] Among them, cross-domain transition candidate chain segments refer to the set of chain segments obtained after misaligned projection search that connect potential propagation relationships between different detection domains.

[0083] Among them, the back-loop verification refers to the process of verifying whether there is a valid back-loop path and closed propagation structure in the candidate chain segments of cross-domain transitions.

[0084] Among them, the effective transition link set refers to the set of transition links that remain after the loopback verification and meet the requirements of connectivity and loop stability.

[0085] Link convergence processing refers to the process of merging, pruning, and unifying multiple links in the effective transition link set to form a stable link result.

[0086] Specifically, for each domain perturbation fold node in the domain perturbation fold body, the perturbation intensity value of all fold nodes is calculated by weighting the number of domain perturbation nodes contained within the fold node, the cumulative edge weight within the node, the number of cross-domain connections of the node, and the inversion priority of the node in the domain perturbation evolution sequence. Then, the perturbation intensity value of all fold nodes is compared with a preset threshold. Fold nodes with perturbation intensity values ​​greater than or equal to the preset threshold are retained, while fold nodes with perturbation intensity values ​​lower than the preset threshold and no direct graph connection with the pressure sensor chip under test are deleted. The retained fold nodes are then subject to entry pruning. That is, according to the direction of the main connection edge, the connection level, and the number of adjacent nodes of each fold node in the heterogeneous manufacturing graph, only the entry and exit edges that can reach the chip node, device node, process node, or time node within a preset step size are retained, while the entry and exit edges pointing to weakly associated nodes, low edge weight nodes, and redundant backflow branches are removed. After pruning, each retained folded node and its remaining inlet / outlet edges are combined into a candidate transition unit, and sorted according to the node perturbation intensity value, the total weight of the remaining inlet / outlet edges, and the graph distance from the node to the pressure sensor chip under test, thus obtaining the candidate transition source set.

[0087] Starting with each candidate transition unit in the candidate transition source set, a projection search coordinate system is established in the heterogeneous manufacturing map, where equipment nodes, process nodes, time nodes, location nodes, and test nodes are considered as different projection layers. Then, for each candidate transition unit, it first traverses one or more non-same-layer nodes before entering the target layer to form a cross-layer jump path. At each cross-layer jump, the jump direction, jump span, and edge weight decay value are recorded. Path searches are then performed in parallel within each projection layer. Path segments that meet the criteria of "two consecutive steps not falling into the same layer, two consecutive jump directions not being completely opposite, and cumulative edge weight decay not exceeding a preset upper limit" are retained. These path segments are registered according to the starting fold node, transition node sequence, and ending node position. When multiple path segments intersect at a relay node, the paths before and after the intersection are spliced ​​together to form a cross-domain transition candidate chain segment. Path segments that fail to intersect within the preset search depth, have continuously exceeded the edge weight decay limit, or have conflicting cross-layer sequences are deleted. Finally, all the remaining spliced ​​chain segments are written into the chain segment set according to the chain segment length, cumulative edge weight, number of cross-levels, and endpoint proximity to obtain cross-domain transition candidate chain segments.

[0088] For each cross-domain transition candidate chain segment, a chain segment sequence list is established according to the order of node appearance. The chain segment sequence list is then checked for any backtracking segments that return from a subsequent node to a preceding node. If a backtracking segment exists, a closed-loop detection unit is formed by the backtracking start point, backtracking end point, and backtracking relay node. The shortest connection step between the backtracking start point and backtracking end point in the heterogeneous manufacturing graph is calculated and converted into a graph distance factor. The edge weights of each edge in the backtracking segment are then summed to obtain a total edge weight value, which is converted into an edge weight factor. The proportion of recurring nodes within the closed loop to the total number of nodes in the closed loop is then calculated to obtain the node repetition rate, which is converted into a repetition rate factor. The consistency between the propagation direction and the backtracking direction of each segment within the closed loop is compared, and the direction consistency factor is obtained based on the proportion of edge segments with consistent directions to the total number of edge segments in the closed loop. Finally, the graph distance factor, edge weight factor, repetition rate factor, and direction consistency factor are normalized and weighted to obtain the closure degree. The closure degree is compared with a preset loop closure threshold. When the closure degree reaches the preset threshold, the jump segment is retained and recorded as a valid loop segment. When the closure degree is lower than the preset threshold, the candidate chain segment containing the corresponding jump segment is deleted or the invalid loop portion is removed. For the retained valid loop segments and their adjacent parts in the original chain segment, multiple consecutive valid loop segments in the same chain segment are merged into a complete transition link according to their order. For parts with local breaks but whose two ends can still be connected by high-weight edges in the graph, they are reconnected. Finally, all chain segments that have completed loop closure verification and reconnection are screened according to the number of loops, loop stability, chain segment connectivity, and cumulative edge weight. Chain segments with zero loop count and insufficient chain segment connectivity are deleted, and the remaining chain segments are retained to obtain the set of valid transition links.

[0089] All transition links in the effective transition link set are grouped according to the starting fold node, the position of the ending node, the cross-layer order, and the repetition of relay nodes. Transition links with the same starting node and adjacent ending regions are grouped into the same candidate convergence group. Within each candidate convergence group, the cumulative edge weight, link length, closed-loop stability, and relay node overlap rate of each transition link are compared. For multiple transition links with a relay node overlap rate exceeding a preset threshold, the main chain is retained and the branches are connected together. That is, the transition link with the largest cumulative edge weight and the highest closed-loop stability is retained as the main chain, and the remaining transition links are connected as branches to the corresponding relay nodes of the main chain. Then, the direction of the connected main chain and branches is unified, so that the propagation direction of all transition links in the same candidate convergence group is recalibrated according to the order from the starting fold node to the ending node, and branches with conflicting directions are reversed or partially reversed. Then, for each candidate convergence group, a converged link structure is output, and the starting node, ending node, relay node sequence, cumulative edge weight, cross-layer order, closed loop position, and branch connection relationship are recorded in the link structure. Finally, all converged link structures are summarized to generate inter-domain contagion link data.

[0090] In this embodiment, by first performing entry pruning on folded nodes with high disturbance intensity, and then combining heterogeneous manufacturing maps for misaligned projection search, loopback verification, and link convergence processing, it is possible to prioritize the retention of transition entry points more likely to form actual propagation in a large number of cross-layer connections. This organizes the originally scattered, misaligned, and intermittent propagation fragments into effective transition links with sequential connections, thereby reducing the interference of low-correlation links, redundant backflow links, and pseudo-loop links on subsequent analysis. Based on this, the resulting inter-domain contagion link data not only indicates whether nodes are connected, but also provides cross-domain jump paths, loop stability, and main-branch convergence relationships. Therefore, it is more conducive to locating the actual diffusion channels of disturbances between different detection domains and improving the adaptability of subsequent inter-domain coupling analysis to complex jump propagation and multi-branch propagation scenarios.

[0091] In an exemplary embodiment, based on the candidate transition source set and the heterogeneous manufacturing map, a misaligned projection search is performed on the pressure sensor chip under test to obtain cross-domain transition candidate segments, including steps 602 to 606. Wherein:

[0092] Step 602: Perform offset cross-layer mapping on the candidate transition source nodes in the candidate transition source set to obtain the heterogeneous projection contact point.

[0093] Step 604: Based on the heterogeneous projection contact and heterogeneous manufacturing pattern, perform discontinuous link splicing analysis on the pressure sensor chip under test to obtain candidate transition fragment clusters.

[0094] Step 606: Perform chain segment configuration processing on the candidate transition fragment cluster to obtain cross-domain transition candidate chains.

[0095] Among them, candidate transition source node: refers to the node unit located in the candidate transition source set that can serve as the starting position of the cross-domain transition path.

[0096] Among them, offset layer mapping refers to the process of making the candidate transition source node not extend directly along the original layer, but cross intermediate layers and map to the target layer according to the preset offset rules.

[0097] Heterogeneous projection contact: refers to the landing point node formed at different layer positions after the candidate transition source node is offset through layer mapping.

[0098] Intermittent link splicing analysis refers to the analysis process that allows multiple link segments with local gaps or bridging relationships to be connected and combined according to preset conditions.

[0099] Among them, candidate transition fragment clusters refer to a set of fragments composed of multiple related transition fragments formed after discontinuous link splicing analysis.

[0100] Among them, the chain segment configuration processing refers to the process of organizing the skeleton, unifying the orientation and refining the structure of the segments in the candidate transition segment cluster to form a complete transition chain segment.

[0101] Specifically, node positioning coordinates are established for each candidate transition source node in the candidate transition source set. The node positioning coordinates include at least its current layer position in the heterogeneous manufacturing map, the direction of the associated edge, the edge weight level, and the graph distance to the pressure sensor chip under test. According to the preset layer crossing order, a cross-layer connection edge is selected as the main layer crossing edge based on the edge type priority. Then, according to the preset offset step size, the candidate transition source node crosses at least one intermediate node before falling into the target layer. After falling into the target layer, the associated nodes within the preset step size range around the landing point are compared for adjacency. Only landing points that simultaneously meet the following conditions are retained: the edge weight is not lower than the preset threshold, the layer crossing direction does not conflict with the original transition direction, and the graph distance to the pressure sensor chip under test is shortened or remains unchanged. When there are multiple landing points that meet the conditions for the same candidate transition source node, they are sorted in the order of "fewer layers crossed, higher cumulative edge weight, and smaller direction offset". The top-ranked landing points are retained as the heterogeneous projection contacts corresponding to the candidate transition source node. Finally, each candidate transition source node and its corresponding heterogeneous projection contact are written into a unified mapping structure according to a one-to-one mapping relationship to obtain a heterogeneous projection contact set.

[0102] Starting from each heterogeneous projection contact point, the heterogeneous manufacturing map is expanded layer by layer along the equipment edge, process edge, time edge, location edge, and test edge connected to that contact point. Each local path formed by the expansion is recorded as a separate link segment. Instead of requiring the endpoint of the previous segment to completely coincide with the starting point of the next segment, a graph interval within a preset step size is allowed between two local paths. As long as this graph interval can be bridged through a high-weight connecting edge or a shared relay node, the two segments are considered mergeable. During the merging process, the hierarchical relationship, edge direction relationship, and bridging path length between the endpoint of one segment and the starting point of the next segment are compared. Segment combinations with excessively large hierarchical differences, completely opposite directions, or bridging path lengths exceeding a preset upper limit are eliminated. Then, sequential merging is performed on the selected segment combinations, connecting multiple local paths sequentially into longer chain segments. After each merging, the cumulative edge weight, number of layers crossed, and relay node repetition rate of the chain segment are recalculated. Merging results with excessively low cumulative edge weights, excessively dense layer jumps, or excessively high repetition rates are deleted. When multiple spliced ​​chain segments share the same starting contact or the same bridging relay node, they are grouped into the same candidate segment cluster and internally sorted according to chain segment length, bridging stability, and cumulative edge weight. Finally, the candidate segment clusters that have passed splicing and screening are output.

[0103] Within each candidate transition segment cluster, the segment with the largest cumulative edge weight and the highest bridging stability is selected as the main configuration skeleton. The remaining segments are then sequentially attached to the corresponding positions on the main configuration skeleton according to their node overlap rate, direction consistency rate, and bridging position adjacency. Next, local breaks in the overall chain segment after attachment are repaired, repeated parallel segments are merged, and conflicting branches are pruned or reversed to ensure the chain segment remains continuous in node order and consistent in propagation direction. The repaired chain segment is then re-labeled with its start point, end point, and relay node sequences. Based on its cross-layer order, the chain segment is divided into start segments, transition segments, and convergence segments. Simultaneously, the cumulative edge weight and bridging density of each segment are calculated, and chain segments with insufficient length or bridging density below a preset threshold are deleted. Among the retained chains, those whose endpoints are in the adjacent domain of the pressure sensor chip under test or have a high-weight connection relationship with the pressure sensor chip under test are retained first and sorted according to priority. Finally, all the chains that have passed the configuration modification and sorting are written into the chain result structure to obtain cross-domain transition candidate chains.

[0104] In this embodiment, by first performing offset cross-layer mapping on candidate transition source nodes, then conducting discontinuous link splicing analysis around the heterogeneous projection contact points, and performing chain segment configuration processing on the resulting candidate transition fragment clusters, cross-layer association relationships that were originally difficult to connect directly along continuous paths in the same layer can be transformed into splicable candidate propagation fragments. This allows discrete jump relationships distributed at different node levels, different connection depths, and different local regions to be organized into cross-domain transition candidate chain segments with stronger continuity. The chain segment results formed in this way no longer rely on the direct connection of a single path, but can retain the combination relationship between cross-layer landing points, bridging positions, and chain segment skeletons. Therefore, it is more conducive to covering potential propagation paths under conditions of multiple jumps, weak continuity, and local breaks, and reduces the situation of missing effective cross-domain diffusion channels due to insufficient direct connections in the same layer.

[0105] In an exemplary embodiment, based on the reference population distribution model corresponding to the pressure sensor chip under test and the chip comprehensive characterization vector, a population deviation analysis is performed on the pressure sensor chip under test to obtain the deviation characterization result, including steps 702 to 706. Wherein:

[0106] Step 702: Perform a back-transmission process on the chip comprehensive representation vector to obtain the group mapping trigger point.

[0107] Step 704: Based on the population mapping trigger point and the reference population distribution model, perform mirror sieving analysis on the pressure sensor chip under test to obtain the deviation candidate beam.

[0108] Step 706: Perform deviation explicit processing on the deviation candidate bundles to obtain the deviation characterization results.

[0109] Among them, the back-and-forth delivery process refers to the process of first sending the chip comprehensive representation vector to the target region boundary of the reference population distribution model according to the preset mapping rules, and then sending it back to the interior of the region in the opposite direction.

[0110] Among them, the group mapping trigger point refers to the landing point in the reference group distribution model formed after the chip comprehensive characterization vector has been processed by the return delivery, which is used to start the subsequent deviation analysis.

[0111] Among them, mirror sieving analysis refers to the process of unfolding, moving and sieving in the reference population distribution model in a mirror direction around the population mapping trigger point to form an analysis process of deviation trajectory.

[0112] Among them, the deviation candidate bundle refers to the bundled result formed after mirror sieving analysis, which consists of multiple interrelated deviation segments.

[0113] Among them, deviation explicit processing refers to the process of converting positional changes, directional changes, and continuity relationships in the deviation candidate bundle into output deviation results.

[0114] Specifically, the chip synthesis representation vector is divided into several representation segments according to a preset dimensional grouping rule, and a corresponding group mapping channel is assigned to each representation segment. The group mapping channel establishes a fixed correspondence with different distribution areas in the reference group distribution model. Starting from the current position of the chip synthesis representation vector, each representation segment is first sent to the boundary of the target distribution area in the reference group distribution model along the first mapping direction, and then sent back to the interior of the target distribution area in the opposite direction according to a preset backtracking rule, so that the same representation segment forms paired landing points at the boundary position and the interior position. Then, the spacing, backtracking angle offset, and landing point density between each pair of landing points are compared, and only landing point groups that simultaneously meet the following conditions are retained: the boundary fit reaches a preset threshold, the backtracking offset is less than a preset upper limit, and there is no conflict with the landing points of adjacent representation segments. When the same representation segment forms multiple landing point groups that meet the conditions, they are selected in the order of "priority given to internal landing points closer to high-density distribution areas, priority given to boundary landing points closer to partition transition surfaces, and priority given to landing point spacing smaller". Finally, all the retained landing point groups are written into the unified mapping structure according to their spatial location and channel number in the reference population distribution model to obtain the population mapping trigger point.

[0115] Centered on each population mapping trigger point, a local distribution region is selected within the reference population distribution model where the trigger point is located. Within this local distribution region, the direction of the most concentrated distribution of neighboring sample points is calculated as the local principal direction. Then, using the boundary position of the trigger point as a reference, the vertical direction pointing outwards from the boundary position is calculated as the boundary normal direction. Finally, the local principal direction and the boundary normal direction together form the planar direction of the mirror control surface, which is then made to pass through the trigger point to obtain the corresponding mirror control surface. Using the vertical projection line from the trigger point to the mirror control surface as a reference, the trigger point is mirrored, forming a pair of mirrored screening trajectories on both sides of the mirror control surface. The point then moves gradually along these two mirrored screening trajectories at a preset step size. During each movement, the distance between the current position and neighboring distribution units in the reference population distribution model, the directional angle, and the local density change rate are recorded. Different screening weights are assigned to trajectory segments that overlap with high-density distribution areas, cross boundary partitions, or adjoin sparse areas during the movement. All mirrored screening trajectory segments are segmented and compared. Segments with continuous direction, consistent density variation trend, and cumulative step length within a preset range are grouped into the same candidate deviation segment. Multiple candidate deviation segments are merged into a deviation candidate bundle according to the rule of having the same starting trigger point or adjacent screening ends. Finally, the deviation candidate bundles are sorted according to the total screening length, the number of boundary crossings, the number of sparse region adjacencies, and the cumulative screening weight. Bundles that are ranked lower and lack continuity are deleted to obtain the deviation candidate bundle.

[0116] For each candidate deviation segment in the deviation candidate bundle, a segment sequence is established according to the order of appearance. For each segment in the sequence, its starting position, ending position, screening direction, number of boundary crossings, cumulative movement step length, and adjacent distribution area type are recorded. Then, the connection relationship between adjacent segments within the same deviation candidate bundle is corrected for continuity. Segments with a connection spacing less than a preset threshold and a directional deviation not exceeding a preset angle are spliced ​​together, while segments with excessive connection spacing or conflicting directions are broken. The spliced ​​segment sequence is numerically expanded, and the total offset, average offset gradient, boundary crossing density, sparse region contact length, and number of screening backslides are calculated. These results are then written into a unified deviation parameter structure. After constructing the deviation parameter structure, the results of multiple deviation candidate bundles are aggregated. Candidate bundles with similar starting positions and consistent offset directions are merged into the same deviation cluster. Candidate bundles with opposite directions but adjacent ending points are designated as offset deviation bundles. The merged deviation clusters, offset deviation bundles, and corresponding deviation parameter structures are output uniformly to obtain the deviation characterization results.

[0117] In this embodiment, the chip comprehensive characterization vector is first processed by foldback delivery, and then mirror sieving analysis is performed in combination with the reference population distribution model. The deviation candidate bundles are then processed by deviation manifestation. Instead of using a single distance value or a single threshold to statically judge the difference between the pressure sensor chip under test and the reference population, the landing position of the pressure sensor chip under test in the reference population distribution space, the boundary approach process, the mirror wandering trajectory, and the continuous relationship between multiple deviation segments are all incorporated into the processing link. This allows the deviation results to reflect not only the magnitude of the deviation, but also the direction, path, and aggregation state of the deviation. This makes it easier to distinguish edge fluctuation type deviation, continuous outward movement type deviation, and hedging back swing type deviation, and improves the adaptability of multi-domain destruction analysis to different population deviation patterns.

[0118] In an exemplary embodiment, based on the population mapping trigger point and the reference population distribution model, a mirror sieving analysis is performed on the pressure sensor chip under test to obtain the deviation candidate bundle, including steps 802 to 806. Wherein:

[0119] Step 802: Perform mirror projection processing on the group mapping trigger points to obtain the mirror domain touch set.

[0120] Step 804: Based on the mirror domain contact set and the reference group distribution model, perform a wandering sieving analysis on the pressure sensor chip under test to obtain the deviation candidate chip bundle.

[0121] Step 806: Perform beam domain convergence on the off-target candidate bundles to obtain the off-target candidate bundles.

[0122] Among them, mirror-image projection processing refers to the process of unfolding and dividing the reference population distribution model along the mirror direction to form multiple local projection units, with the population mapping trigger point as the turning point.

[0123] Among them, the mirror field touch set refers to the set of multiple local touch units formed around the same group of mapping trigger points after mirror projection processing.

[0124] Among them, the wandering cutting and screening analysis refers to the analysis process in which the mirror field contact moves in a preset direction in the reference population distribution model and simultaneously performs cutting and screening.

[0125] Among them, the deviation candidate slice bundle refers to the slice bundle result formed after the wandering and screening analysis, which is a combination of multiple interrelated deviation slice columns.

[0126] Among them, beam convergence refers to the process of merging and integrating multiple deviating candidate slices according to their starting position, wandering direction and proximity relationship of the ending region.

[0127] Specifically, each group mapping trigger point is located to its corresponding distribution area in the reference group distribution model, and a local mirroring coordinate system is established based on the location of the group mapping trigger point. The local mirroring coordinate system includes at least a principal axis set along the local distribution extension direction and a return axis set along the outer boundary direction. Using the group mapping trigger point as the return point, mirroring operations are simultaneously performed along the principal axis and the return axis direction, so that the group mapping trigger point forms multiple discrete landing points on both sides after mirroring. Each discrete landing point is sliced ​​into pieces according to a preset slice width and preset slice length to generate multiple local film-laying units. Then, the positions of each local film-laying unit are adjusted, and only local film-laying units located within the allowable distribution area of ​​the reference group distribution model and without overlapping conflicts with adjacent film-laying units are retained. Local film-laying units located in the boundary transition area, dense distribution area, or sparse transition area are assigned different slice position labels. Based on this, local film-laying units with the same return source point and located on adjacent mirroring layers are organized sequentially to form a mirror domain touch set unfolding around the same group mapping trigger point.

[0128] Starting with each mirror touch in the mirror touch set as a unit, the current mirror touch is corrected along a preset wandering direction in the reference group distribution model based on the boundary relationship between the current mirror touch and the adjacent distribution area, ensuring that the mirror touch re-fits the current local distribution structure after each movement. Then, after each movement, the current mirror touch is divided into multiple parallel slice units according to a preset slice width. Each slice unit is then aligned with a local distribution unit in the reference group distribution model, retaining slice units tangent to high-density distribution units, intersecting with boundary partitions, or adjacent to sparse areas, while deleting slice units in invalid regions or those with continuously deviating directional conflicts. Furthermore, the slice units retained from multiple consecutive steps are concatenated according to temporal order and spatial adjacency. When the movement interval between adjacent slice units is less than a preset threshold and the directional deviation does not exceed a preset angle, they are merged into the same wandering slice column. When a local break occurs but the break spacing can still be repaired through a short bridging path, the corresponding slice column continues to be retained and bridging and merging are performed. After completion, multiple wandering slices with the same initial mirror field contact, consistent wandering direction, and ends falling into adjacent local distribution areas are merged into the same deviation candidate slice bundle to obtain all deviation candidate slice bundles.

[0129] For each candidate deviating slice bundle, a slice bundle index is established. All slice sequences within the candidate deviating slice bundle are registered according to their start position, end position, cumulative wander length, number of cut-and-screen retentions, and number of bridging and merging. For candidate deviating slice bundles with adjacent starting mirror region touch slices, consistent main wandering directions, and overlapping or adjacent ending regions, they are determined to belong to the same convergence region and are merged into the same candidate convergence group. Then, within each candidate convergence group, a main bundle is selected and auxiliary bundles are attached. That is, the candidate deviating slice bundle with the longest cumulative wander length, the most cut-and-screen retentions, and the ending region closest to the boundary turning area is selected as the main bundle. The remaining candidate deviating slice bundles are attached to the corresponding slice sequence positions of the main bundle according to node overlap and path adjacency. Then, auxiliary bundles with conflicting directions within the same candidate convergence group are pruned, auxiliary bundles with scattered ends but still within the same ending region are merged at the end, and locally duplicate slice sequences are merged to eliminate duplicates. Finally, the result after main bundle retention, auxiliary bundle merging, direction clipping and end merging is output as a complete deviation candidate bundle, and all complete deviation candidate bundles are summarized to obtain the deviation candidate bundle.

[0130] In this embodiment, by first performing mirroring and projection processing on the group mapping trigger point, then combining the reference group distribution model for wandering and screening analysis, and performing bundle domain convergence on the formed deviation candidate bundle, the group deviation relationship originally concentrated at a single point can be expanded into multiple continuously trackable local touch pieces and wandering piece columns. This allows the local expansion, boundary collision, and fragment merging relationships of the deviation process in the distribution space to be gradually preserved. On this basis, the deviation candidate bundle formed by bundle domain convergence is no longer an isolated set of deviation fragments, but a bundled structure with initial correlation, wandering continuity, and termination aggregation. Therefore, it is more conducive to reflecting the expansion contour and convergence direction of deviation in the local distribution area and reducing the instability of deviation judgment caused by the discrete appearance of a single fragment.

[0131] In an exemplary embodiment, based on the mirror domain contact set and the reference population distribution model, a wandering sieving analysis is performed on the pressure sensor chip under test to obtain the deviation candidate chip bundle, including steps 902 to 908. Wherein:

[0132] Step 902: Split-window projection is performed on the mirror field contacts in the mirror field contact set to obtain the wandering contact grid.

[0133] Step 904: Based on the wandering lattice and the reference group distribution model, perform staggered wandering analysis on the pressure sensor chip under test to obtain the deviation candidate patch clusters.

[0134] Step 906: Perform reverse pruning on the candidate patch clusters that deviate from the target to obtain the effective deviated patch clusters.

[0135] Step 908: Perform clustering processing on the effective deviated slice columns to obtain the deviated candidate slice columns.

[0136] Among them, split-window projection refers to the process of dividing the mirror field touch panel into multiple sub-windows according to preset window rules and then distributing them according to different layers or offset directions.

[0137] Among them, wandering lattice refers to a regular lattice structure formed after split-window projection, which can continue to wander and attach in the reference population distribution model.

[0138] Among them, the staggered layering analysis refers to the analysis process of moving the wandering lattice between different layers according to preset rules and continuously attaching and comparing it with the corresponding distribution area.

[0139] Among them, the off-candidate patch cluster refers to the cluster-like result formed after the cross-layered patch analysis, which is composed of multiple adjacent patches with related migration directions.

[0140] Among them, reverse cropping and retention processing refers to the process of going back from the end of the candidate patch cluster and deleting the part with insufficient continuity while retaining the stable patch part.

[0141] Among them, the effective deviation sequence refers to the deviation segment sequence that is retained after reverse pruning and meets the conditions of continuity and bridging.

[0142] Among them, the sheet clustering process refers to the process of merging and organizing multiple valid deviated sheets into sheet bundles according to their starting position, wandering direction and the proximity of their ending regions.

[0143] Specifically, each mirror touch piece in the mirror touch piece set is located to its corresponding local region in the reference group distribution model. Local window coordinates are established based on the center position of the mirror touch piece. Then, according to preset window width, preset window height, and preset overlap step size, each mirror touch piece is decomposed into multiple adjacent and partially overlapping sub-windows along the main extension direction and lateral offset direction. Based on this, according to the positional relationship of each sub-window relative to the local boundary, local dense area, and local sparse area, adjacent sub-windows are projected to different wandering layers according to preset offset rules. The original touch piece number, window sequence number, projection layer, offset step size, and boundary contact state are recorded for each projected sub-window. Sub-windows with excessive positional overlap, conflicting offset directions, or duplicate projection layers are deduplicated and merged, retaining only sub-window groups that meet preset spacing and layer difference requirements. Finally, all sub-windows retained after split-window projection of the same mirror touch piece are organized into a regular grid structure according to their arrangement order in the local distribution space, resulting in a wandering grid.

[0144] Using each wandering cell as the initial attachment unit, the cell moves gradually along a preset wandering direction within the reference population distribution model. After each move, it is determined whether the current cell crosses a local distribution boundary or enters an adjacent layer. If a layer change occurs, the cell is attached to the new layer region according to the staggered layer rule, instead of remaining in the original layer. Next, the attached cell is compared with adjacent distribution units in the current layer in terms of position, direction, and boundary contact. Only cell units that simultaneously meet the following conditions are retained: attachment angle deviation does not exceed a preset angle, attachment edge overlap reaches a preset ratio, and the wandering direction does not conflict with the previous time step. Cell units that still meet the above conditions after multiple consecutive wanders are then connected according to time sequence and spatial adjacency to form candidate attachment chains. Multiple candidate attachment chains with adjacent initial cell units, similar wandering directions, and converging termination regions are merged into the same deviation candidate attachment cluster. For lattice cells that experience long-distance jumps, sudden changes in attachment angles, or two consecutive failed boundary contacts during the migration process, the extension of the corresponding attachment chain is terminated, and they are no longer merged into the same deviation candidate cluster. Finally, all deviation candidate clusters formed after the staggered migration are output.

[0145] For each candidate offset patch cluster, a reverse traversal order is established, starting from the last attached unit of the cluster and proceeding back to the starting attached unit in its formation order. During the traversal, the spacing, directional deviation, and boundary continuity between each attached unit and the previous attached unit are checked. When the spacing between two adjacent units exceeds a preset threshold, the directional deviation exceeds a preset angle, or the boundary continuity is interrupted, the current attached unit is recorded as a trimming point. Using the trimming point as the boundary, the tail segment after the trimming point is deleted or shortened, retaining only the segment that maintains a continuous attachment relationship with the preceding attached unit. For candidate offset patches with multiple trimming points, the above trimming and retention operations are performed in reverse order, prioritizing the retention of the part that is closer to the starting segment and has higher continuity. The remaining attached units after trimming are reconnected in sequence. If there is still a break less than the preset bridging length between adjacent retained segments, the two segments are joined together to form the same valid patch array. Finally, all results after reverse trimming and bridging are uniformly registered to obtain the valid offset patch array.

[0146] All valid deviation fragments are grouped according to their starting position proximity, main wandering direction similarity, and termination region overlap. Valid deviation fragments whose starting positions are in the same mirror domain contact patch neighborhood, whose main wandering direction angle is less than a preset angle, and whose termination positions fall within adjacent local distribution areas are grouped into the same candidate clustering group. Within each candidate clustering group, the length, number of consecutive attached segments, number of bridging operations, and number of boundary collisions of each valid deviation fragment are sorted, and the valid deviation fragment with the longest length and the most consecutive attached segments is selected as the main fragment. Then, the remaining valid deviation fragments are attached to the sides or relay positions of the main fragment according to their node overlap rate and path proximity, forming a fragment bundle structure unfolding around the main fragment. Valid deviation fragments that significantly conflict with the main fragment direction, have excessively large termination region deviations, or have too many bridging operations are removed from the candidate clustering group and do not participate in fragment bundle construction. Next, repeated segments within the fragment bundle are merged, adjacent endpoints are closed, and the order of the starting, unfolding, and closing ends of the fragment bundle is unified. Finally, the clustered and bundled sheet structures are output to obtain the off-candidate sheet bundles.

[0147] In this embodiment, by first performing split-window projection on the mirror domain contact slices, then combining the reference population distribution model to perform staggered layer wandering analysis, and forming deviation candidate slice bundles through reverse trimming and slice bundle clustering, the deviation trajectory that was originally insufficient in continuity, scattered in layer and with obvious local jumps can be broken into wandering slice grids that can be advanced in layers. In subsequent processing, the combination of segments with higher attachment continuity and more stable boundary transition relationships is retained, so that the scattered deviation segments are tightened forward after reverse trimming, and after slice bundle clustering, a slice bundle result with primary and secondary levels and extension direction is formed. This is more conducive to suppressing the interference of local noisy segments and short-term jump segments on deviation judgment, while enhancing the continuous tracking ability of cross-layer wandering deviation and multi-slice convergence deviation.

[0148] In an exemplary embodiment, based on multi-domain damage data, an instability defect analysis is performed on the pressure sensor chip under test to obtain defect detection data, including steps 1002 to 1010. Wherein:

[0149] Step 1002: Perform instability and collapse processing on the active destruction units in the multi-domain destruction data to obtain unstable nucleus data.

[0150] Step 1004: Based on the unstable nucleus data, perform pattern pull-out analysis on the pressure sensor chip under test to obtain candidate defect clusters.

[0151] Step 1006: Arrange the candidate defect clusters in reverse order to obtain the defect level difference column.

[0152] Step 1008: Based on the defect hierarchy and candidate defect clusters, perform defect chain analysis on the pressure sensor chip under test to obtain the defect chain results.

[0153] Step 1010: Perform result shaping processing on the defect chain results to obtain defect detection data.

[0154] Among them, active destruction unit refers to a local unit in multi-domain destruction data that simultaneously satisfies high destruction intensity, strong propagation activity, or obvious coupling anomaly.

[0155] Among them, instability and collapse treatment refers to the process of compressing and synthesizing multiple adjacent and continuously active destructive units into a concentrated instability structure according to preset rules.

[0156] Among them, unstable core data refers to data that records the location, range, and connectivity of concentrated unstable clusters after instability and collapse processing.

[0157] Pattern elicitation analysis refers to the analysis process that starts with unstable nucleus data and organizes defect pattern chains along their connections and expansion directions.

[0158] Among them, candidate defect clusters refer to clusters formed after pattern extraction analysis, which are composed of multiple interrelated defect pattern fragments.

[0159] Among them, reverse order arrangement refers to the process of reordering candidate defective clusters not according to the original intensity order, but according to the activity level of the propagation terminal, the boundary expansion situation, or the return frequency.

[0160] Among them, the defect hierarchy sequence refers to the ordered sequence formed after inverse order arrangement, which reflects the hierarchical order and differences between different candidate defect clusters.

[0161] Among them, defect chain analysis refers to the analysis process of chaining and organizing multiple defect levels by combining the connection relationship between defect level difference columns and candidate defect type clusters.

[0162] Among them, defect chain results refer to the defect chain results formed after defect chain analysis, which include the main chain, branches and their connection relationships.

[0163] Among them, the result shaping process refers to the process of organizing, closing and structuring the defect chain results according to the preset output structure.

[0164] Specifically, multi-domain destruction data are uniformly numbered according to propagation chain number, detection domain number, coupling edge number, and imbalance location number. Adjacent destruction active units that appear consecutively within the same statistical window and whose destruction intensity exceeds a preset active threshold are included in the same set to be collapsed. Starting from each destruction active unit in the set to be collapsed, adjacent units are gradually merged according to the rule of "priority given to continuous propagation direction, priority given to larger coupling edge weight, and priority given to higher local imbalance density," so that multiple destruction active units that are spatially close, temporally continuous, and connected in the propagation chain converge into a single collapse cluster. Then, the destruction intensity, propagation span, local imbalance value, and cross-domain coupling value of each destruction active unit within each collapse cluster are accumulated according to a preset ratio, and the accumulated result is written to the cluster's central node; simultaneously, the cluster's outer boundary, the main propagation direction within the cluster, and the external propagation edges connecting the cluster are preserved. Multiple collapse clusters with overlapping boundaries or repeated external propagation edge connections are merged a second time, retaining only the cluster structure with the greater cumulative intensity and more stable external connections. Finally, the collapsed and merged clumps are output in a unified format, which records the clumping center location, clumping range, main propagation direction, cumulative damage intensity, and external connectivity, thus obtaining the unstable core clumping data.

[0165] Starting with each unstable core cluster data point, a core cluster expansion channel is established in a predefined pattern association space. The center position, outer boundary, main propagation direction, and external connectivity of the unstable core cluster data are written into this expansion channel. Other core clusters, propagation chains, and imbalanced edge groups connected to the unstable core cluster data are then sequentially expanded along this channel. Multiple closely connected, propagating, and overlapping outer boundaries are merged into the same pattern chain. Then, multiple pattern chain segments with short connection paths, high boundary proximity, and consistent propagation directions are synthesized into the same pattern cluster. The synthesized pattern clusters are then sorted according to their positional relationship in the multi-domain disruption network. For pattern chains with directional conflicts, excessively large connection spans, or frequent relay chain breaks, merging into the current pattern cluster is stopped, and the cluster is transferred to a new pattern cluster channel. For each pattern cluster, its starting core cluster, expansion range, internal cluster connection sequence, and inter-cluster boundary relationships are recorded, ultimately forming candidate defect clusters.

[0166] The initial ranking of each candidate defect cluster is calculated sequentially based on the cumulative destruction intensity, cross-domain coverage, propagation chain closure, and outward boundary length. Candidate defect clusters with more active propagation terminals, more pronounced boundary expansion, and more frequent cross-domain backflow are prioritized, while those with earlier propagation starts but lower current activity levels are placed later. Then, the difference in level is calculated for adjacent candidate defect clusters after the arrangement. The cumulative destruction intensity difference, cross-domain coverage difference, closure difference, and outward boundary difference between consecutive candidate defect clusters are converted into difference values ​​using a preset formula, and each difference value is written into the arrangement chain in sequence. Adjacent candidate defect clusters with excessively small difference values ​​and overlapping boundaries are merged to occupy the same arrangement level; adjacent candidate defect clusters with excessively large difference values ​​and obvious propagation breakpoints are skipped to create hierarchical transitions in the arrangement chain. Finally, all candidate defect clusters after inverse arrangement, merging, and skipping are output in chronological order to obtain a defect difference column containing hierarchical positions and difference relationships.

[0167] Each level position in the defect hierarchy is used as the primary attachment position, and the candidate defect clusters at that level are written into the attachment structure as main chain nodes. Hierarchical connection edges are established between adjacent main chain nodes according to the adjacent level order in the hierarchy. The connection direction and connection weight are set based on the boundary adjacency, propagation continuation, and shared core relationships between candidate defect clusters. For candidate defect clusters that are not in a main chain position but have a stable connection with a main chain node, they are connected to the lateral position of the corresponding main chain node, and the connection position, length, and direction between the branch and the main chain are recorded. For candidate defect clusters that have connections with multiple main chain nodes simultaneously, the connection weights and boundary overlap between them are compared, and only the attachment relationship with the largest connection weight and the highest boundary overlap is retained. After completing the main chain connection and branch chain attachment, the entire defect chain is subjected to continuity correction. For chain segments with breakpoints but spacing less than the preset reconnection threshold, short bridge reconnection is performed. For chain segments with conflicting directions, they are cut off or reconnected to obtain the defect chain attachment result.

[0168] The main chain nodes, branch chain nodes, hierarchical edges, and supplementary edges in the defect chain results are uniformly numbered, and a result output sequence is established according to the position on the chain. Using the main chain as the result skeleton, the cumulative damage intensity, hierarchical position, grade difference, and chain span corresponding to each main chain node are numerically written. Branch chains are used as supplementary data, with the attachment position, attachment direction, and attachment length corresponding to each branch chain node being added. Then, the start, end, and local turnaround points of the entire result output sequence are uniformly converted into fixed output fields, and duplicate node numbers are merged, while redundant branches are pruned. Next, the entire result output sequence is converted into a structured result set according to a preset output template, so that the primary and secondary relationships, hierarchical relationships, connection relationships, and span relationships of the same candidate defect chain are continuously presented in the same result structure. Finally, the defect detection data is output.

[0169] In this embodiment, by first performing instability and collapse processing on the active destructive units in the multi-domain destruction data, then conducting pattern extraction analysis around the unstable core, and forming defect detection data through inverse order arrangement, defect chain analysis, and result shaping processing, the destruction relationships that were originally scattered in different propagation positions, different coupling levels, and different active intervals can be compressed into a continuously organized defect chain structure. This allows defect judgment to no longer be limited to the isolated output of a single anomaly point or a single destruction value, but further provides the sequential hierarchy, chain connection relationship, and main branch development relationship between defects. This is more conducive to presenting the clustering trend, bifurcation trend, and level transition trend in the defect evolution process, and reduces the result dispersion and conclusion fragmentation caused by the coexistence of multiple similar defect patterns.

[0170] It should be understood that although the steps in the flowcharts of the embodiments described above are shown sequentially as indicated by the arrows, these steps are not necessarily executed in the order indicated by the arrows. Unless otherwise expressly stated herein, there is no strict order restriction on the execution of these steps, and they may be executed in other orders.

[0171] In one exemplary embodiment, a computer device is provided, which may be a server, and its internal structure diagram may be as follows: Figure 3 As shown. This computer device includes a processor, memory, input / output interfaces (I / O), and communication interfaces.

[0172] Those skilled in the art will understand that Figure 3 The structure shown is merely a block diagram of a portion of the structure related to the present application and does not constitute a limitation on the computer device to which the present application is applied. Specific computer devices may include more or fewer components than those shown in the figure, or combine certain components, or have different component arrangements.

[0173] In one embodiment, a computer device is also provided, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the steps in the above method embodiments.

[0174] In one embodiment, a computer-readable storage medium is provided storing a computer program that, when executed by a processor, implements the steps in the above method embodiments.

[0175] In one embodiment, a computer program product or computer program is provided, the computer program product or computer program including computer instructions stored in a computer-readable storage medium. A processor of a computer device reads the computer instructions from the computer-readable storage medium, and executes the computer instructions, causing the computer device to perform the steps in the above method embodiments.

[0176] It should be noted that the user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data used for analysis, data stored, data displayed, etc.) involved in this application are all information and data authorized by the user or fully authorized by all parties, and the collection, use and processing of the relevant data must comply with relevant regulations.

[0177] Those skilled in the art will understand that all or part of the processes in the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium. When the computer program is executed, it can include the processes of the embodiments of the above methods.

[0178] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0179] The embodiments described above are merely illustrative of several implementation methods of this application, and while the descriptions are specific and detailed, they should not be construed as limiting the scope of this patent application. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of this application, and these all fall within the protection scope of this application. Therefore, the protection scope of this application should be determined by the appended claims.

Claims

1. A defect detection method for pressure sensor chips based on big data, characterized in that, The method includes: Based on the spatiotemporal process correlation data and test response data in the chip manufacturing test dataset corresponding to the pressure sensor chip under test, a multidimensional correlation map of the pressure sensor chip under test in the manufacturing chain is constructed to obtain a heterogeneous manufacturing map. Based on the chip manufacturing test dataset and the heterogeneous manufacturing map, a multidimensional correlation feature analysis is performed on the pressure sensor chip under test to obtain a chip comprehensive characterization vector. Based on the reference population distribution model corresponding to the pressure sensor chip under test and the chip comprehensive characterization vector, a population deviation analysis is performed on the pressure sensor chip under test to obtain the deviation characterization results. Based on the deviation characterization results, multi-domain damage analysis is performed on the pressure sensor chip under test to obtain multi-domain damage data. Based on the multi-domain damage data, instability defect analysis is performed on the pressure sensor chip under test to obtain defect detection data.

2. The method according to claim 1, characterized in that, Based on the deviation characterization results, a multi-domain damage analysis is performed on the pressure sensor chip under test to obtain multi-domain damage data, including: The deviation data of each domain in the deviation characterization results are arranged in reverse order across domains to obtain the domain perturbation evolution sequence; Based on the domain perturbation evolution sequence and the heterogeneous manufacturing map, an inter-domain coupling infection analysis is performed on the pressure sensor chip under test to obtain a domain destruction propagation map. Imbalanced aggregation is performed on the domain destruction propagation graph to obtain the multi-domain destruction data.

3. The method according to claim 2, characterized in that, The step of performing inter-domain coupling contagion analysis on the pressure sensor chip under test based on the domain perturbation evolution sequence and the heterogeneous manufacturing map to obtain a domain destruction propagation map includes: A cross-layer compression mapping is performed on the domain perturbation nodes in the domain perturbation evolution sequence to obtain the domain perturbation fold body; Based on the domain perturbation fold and the heterogeneous manufacturing map, the transition link search is performed on the pressure sensor chip under test to obtain inter-domain infection link data. Reverse energy back-injection is performed on the inter-domain infection link data to obtain the inter-domain coupling strength field; Based on the inter-domain coupling strength field, the graph structure of the pressure sensor chip under test is reconstructed to obtain the domain destruction propagation graph.

4. The method according to claim 3, characterized in that, The step of performing a transition link search on the pressure sensor chip under test based on the domain perturbation fold and the heterogeneous manufacturing map to obtain inter-domain contagion link data includes: The entry points of the folded nodes in the domain perturbation folded volume with perturbation intensity higher than a preset threshold are pruned to obtain a candidate transition source set. Based on the candidate transition source set and the heterogeneous manufacturing map, a misaligned projection search is performed on the pressure sensor chip under test to obtain cross-domain transition candidate chain segments. Perform back-loop closure verification on the cross-domain transition candidate links to obtain the effective transition link set; The effective transition link set is subjected to link convergence processing to obtain the inter-domain infection link data.

5. The method according to claim 4, characterized in that, The step of performing a misaligned projection search on the pressure sensor chip under test based on the candidate transition source set and the heterogeneous manufacturing map to obtain cross-domain transition candidate segments includes: Offset cross-layer mapping is performed on the candidate transition source nodes in the candidate transition source set to obtain heterogeneous projection contacts; Based on the heterogeneous projection contact and the heterogeneous manufacturing pattern, the pressure sensor chip under test is subjected to discontinuous link splicing analysis to obtain a cluster of candidate transition segments. The candidate transition fragment cluster is subjected to chain segment configuration processing to obtain the cross-domain transition candidate chain segment.

6. The method according to claim 1, characterized in that, The step involves performing population deviation analysis on the pressure sensor chip under test based on the reference population distribution model corresponding to the chip and the chip's comprehensive characterization vector, to obtain deviation characterization results, including: The chip's comprehensive characterization vector is subjected to a back-transmission process to obtain the group mapping trigger point; Based on the population mapping trigger point and the reference population distribution model, a mirror sieving analysis is performed on the pressure sensor chip under test to obtain the deviation candidate bundle; The deviation candidate bundle is subjected to deviation explicit processing to obtain the deviation characterization result.

7. The method according to claim 6, characterized in that, The step of performing mirror sieving analysis on the pressure sensor chip under test based on the population mapping trigger point and the reference population distribution model to obtain the deviation candidate beam includes: The group mapping trigger points are subjected to mirror projection processing to obtain a set of mirror domain triggers; Based on the mirror domain contact set and the reference population distribution model, the pressure sensor chip under test is subjected to wandering sieving analysis to obtain the deviation candidate chip bundle; The deviation candidate sheet bundle is subjected to beam domain convergence to obtain the deviation candidate bundle.

8. The method according to claim 7, characterized in that, The step of performing a drift-and-screen analysis on the pressure sensor chip under test based on the mirror domain contact set and the reference population distribution model to obtain the deviation candidate chip bundle includes: The mirror field contacts in the mirror field contact set are split-windowed to obtain a floating contact grid; Based on the wandering lattice and the reference group distribution model, the chip of the pressure sensor under test is subjected to a staggered wandering analysis to obtain the deviation candidate patch clusters; The deviation candidate patch clusters are subjected to reverse pruning to obtain effective deviation patch columns; The effective deviation slices are subjected to slice bundle clustering processing to obtain the deviation candidate slice bundles.

9. The method according to claim 1, characterized in that, The step involves performing instability defect analysis on the pressure sensor chip under test based on the multi-domain damage data to obtain defect detection data, including: The active destructive units in the multi-domain destructive data are subjected to instability and collapse processing to obtain unstable nucleus data; Based on the unstable nucleus data, a pattern pull-out analysis is performed on the pressure sensor chip under test to obtain candidate defect clusters; The candidate defect type clusters are arranged in reverse order to obtain a defect level difference column; Based on the defect hierarchy and the candidate defect clusters, defect chain analysis is performed on the pressure sensor chip under test to obtain defect chain results. The defect chain result is processed to obtain the defect detection data.

10. A defect detection system for a pressure sensor chip based on big data, characterized in that, The system includes: a server and a terminal; The server is used to construct a graph of the multidimensional relationship between the pressure sensor chip under test in the manufacturing chain based on the spatiotemporal process correlation data and test response data in the chip manufacturing inspection dataset corresponding to the pressure sensor chip under test, thereby obtaining a heterogeneous manufacturing graph; the terminal is used to acquire the chip manufacturing inspection dataset. The server is used to perform multi-dimensional correlation feature analysis on the pressure sensor chip under test based on the chip manufacturing test dataset and the heterogeneous manufacturing map to obtain a chip comprehensive characterization vector. The server is used to perform population deviation analysis on the pressure sensor chip under test based on the reference population distribution model corresponding to the pressure sensor chip under test and the chip comprehensive characterization vector, and to obtain deviation characterization results. The server is used to perform multi-domain damage analysis on the pressure sensor chip under test based on the deviation characterization results, and obtain multi-domain damage data. The server is used to perform instability and defect analysis on the pressure sensor chip under test based on the multi-domain destruction data, and obtain defect detection data.

Images