A chip data self-testing method, device, and medium based on artificial intelligence.

Abstract

This invention discloses a chip data self-testing method, device, and medium based on artificial intelligence, relating to the field of data processing technology. The method includes: performing a unified caliber mapping on consistency gating, and combining this with a gating weight modulation neural network to perform gating weight modulation, obtaining a gating fusion score, and simultaneously extracting the unstable window sequence number in the threshold neighborhood to generate a retest candidate list; based on the retest candidate list, performing sampling window retesting with window length expansion and sliding step size reduction to obtain the retest self-test acquisition window, and performing residual entry reconstruction and gating mark registration backfilling to generate a retest gating score; comparing the retest gating score with a data self-test evaluation threshold to obtain a judgment mark, and encapsulating an anomaly fingerprint packet to generate a chip data self-test output set. This invention achieves cross-parameter anomaly collaborative verification, constructs a physically interpretable evidence chain, and improves the confidence and anti-interference capability of composite fault detection.

Description

Technical Field

[0001] This invention relates to the field of data processing technology, and in particular to a chip data self-testing method, device and medium based on artificial intelligence. Background Technology

[0002] With the continuous miniaturization of integrated circuit processes and the widespread application of heterogeneous integrated architectures, ensuring the reliability of chips under complex operating conditions faces severe challenges. Traditional chip testing technologies are mainly geared towards post-manufacturing static verification, relying on preset test vectors and hardware redundancy structures. In recent years, the popularization of on-chip sensor networks has driven the development of a data-driven self-testing paradigm: by real-time acquisition of multi-dimensional operational data such as output signals, local temperature, power supply parameters, and clock edges, anomaly detection is performed in combination with statistical process control, isolated forest, or time-series deep learning models.

[0003] A thorough analysis of existing technologies reveals two core limitations: First, the collaborative analysis mechanism for multi-source heterogeneous data is weak. Current methods often employ independent threshold judgments or static weighted fusion strategies, leading to broken evidence chains in complex fault scenarios and insufficient confidence and anti-interference capabilities in anomaly detection. Second, the self-checking decision-making process lacks dynamic adaptability and result traceability. Mainstream AI models are mostly end-to-end classifiers with fixed thresholds, unable to trigger targeted retesting for ambiguous samples at the judgment boundary, easily causing false positives and false negatives. Summary of the Invention

[0004] In view of the aforementioned existing problems, the present invention is proposed.

[0005] Therefore, this invention provides an artificial intelligence-based chip data self-testing method to solve the problems of low confidence in anomaly detection and high false alarm / false negative rates.

[0006] To solve the above-mentioned technical problems, the present invention provides the following technical solution:

[0007] In a first aspect, the present invention provides a chip data self-testing method based on artificial intelligence, comprising: acquiring multi-view sampling data of the chip, and performing out-of-order callback, missing test completion, and jump dual-track recording to generate a self-testing acquisition window; constructing output residual entries, temperature residual entries, power supply residual entries, and time sequence residual entries based on the self-testing acquisition window, and performing peak time position alignment and supporting evidence counting registration to generate a consistency gating; performing unified caliber mapping on the consistency gating, and combining the gating weight modulation neural network to perform gating weight modulation to obtain a gating fusion score, while extracting the threshold neighborhood unstable window sequence number to generate a retest candidate list; performing sampling window retesting with window length expansion and sliding step size reduction based on the retest candidate list to obtain a retesting self-testing acquisition window, and performing residual entry reconstruction and gating mark registration backfilling to generate a retesting gating score; performing data self-testing evaluation threshold comparison on the retesting gating score to obtain a judgment mark, and performing abnormal fingerprint packet encapsulation to generate a chip data self-testing output set.

[0008] As a preferred embodiment of the AI-based chip data self-testing method of the present invention, the steps for generating the self-test acquisition window are as follows:

[0009] The chip's multi-view sampling data is mapped to a unified sampling scale, and out-of-order rollback and duplicate arrival merging are performed. At the same time, window numbers are added to generate unified sampling scale alignment entries.

[0010] For entries aligned to the uniform sampling scale, perform missing measurement completion according to the sampling scale, and write the missing measurement mark and the completion source index to generate missing measurement completion alignment entries;

[0011] For missing and aligned items, perform jump detection and retain the jump records of the original value track and the corrected value track. At the same time, perform sliding window slicing to generate a self-test acquisition window.

[0012] As a preferred embodiment of the AI-based chip data self-testing method of the present invention, the steps for generating a consistency gating are as follows:

[0013] Based on the unified sampling scale of the self-test acquisition window, the chip multi-view sampling data are aligned to the same time axis within the same window and collected into a set of aligned segments within the window.

[0014] Based on the in-window aligned fragment set, output residual entries, temperature residual entries, power supply residual entries, and time series residual entries are constructed window by window and aggregated into a multi-view residual entry set;

[0015] In the multi-view residual entry set, locate the peak time position of the output residual entry, and perform peak neighborhood hit check to generate supporting evidence to support the count;

[0016] Based on supporting evidence, the evidence consistency screening and grouping of the multi-perspective residual item set are performed to generate a consistency gating.

[0017] As a preferred embodiment of the AI-based chip data self-testing method of the present invention, the steps for generating the retest candidate list are as follows:

[0018] Extract the cumulative residual values ​​of the multi-perspective residual entry set from the consistency gating to form a peer residual value set, and set the corroborating evidence support count as peer gating information;

[0019] Perform a unified mapping on the set of residual values ​​from the same window, and perform cumulative intensity per unit time and amplitude scale alignment to generate unified residuals;

[0020] Based on peer gating information, gating weight modulation is performed on the uniform residual in the gating weight modulation neural network, and the results are aggregated to generate a gating fusion score.

[0021] Extract the window numbers of those whose gating fusion scores fall within the neighborhood of the data self-assessment threshold and exhibit up-and-down swing reversal, and attach retest configuration entries to generate a retest candidate list.

[0022] As a preferred embodiment of the AI-based chip data self-testing method of the present invention, the steps for generating the retest gating score are as follows:

[0023] The retest time range corresponding to the unstable window number in the threshold neighborhood of the candidate list for retesting is located, and the retesting sampling window parameter set is generated by combining the window length expansion factor and the sliding step size reduction factor.

[0024] Based on the sampling window parameter set, the sampling window is retested by expanding the window length and reducing the sliding step size, and multi-view data is collected simultaneously to generate a retest self-check acquisition window;

[0025] Residual entries are reconstructed for the retest self-inspection acquisition window, and peak time position alignment and supporting evidence count verification are performed to generate retest consistency gating.

[0026] A unified caliber mapping and gating weight modulation are performed on the retest consistency gating to obtain the retest gating fusion score. The retest gating score is then generated by replacing and backfilling the unstable window number in the threshold neighborhood.

[0027] As a preferred embodiment of the AI-based chip data self-testing method of the present invention, the steps for generating the chip data self-test output set are as follows:

[0028] Based on the retested gating score, the gating fusion score is compared with the data self-inspection evaluation threshold to obtain the abnormal judgment mark, the health judgment mark and the judgment mark to be reviewed, and then integrated into the judgment mark sequence;

[0029] Based on the judgment marker sequence, the abnormal window is located by backtracking, and the peak time position of the multi-view residual entries, the source summary of the supporting evidence for the count, and the retest configuration entries are extracted to generate an abnormal fingerprint element set;

[0030] The abnormal fingerprint element set is encapsulated with the retest gating score, the comparison result of the data self-test evaluation threshold, and the retest configuration items to generate the chip data self-test output set.

[0031] As a preferred embodiment of the chip data self-testing method based on artificial intelligence described in this invention, the gated weight modulation neural network includes an input normalization layer, a feature projection fully connected layer, a cross-window context extraction one-dimensional convolutional layer, a gated weight generation fully connected layer, and a weight normalization layer.

[0032] As a preferred embodiment of the chip data self-testing method based on artificial intelligence described in this invention, wherein: the input normalization layer performs scale normalization on the uniform caliber residual and the same window gate control information;

[0033] The feature projection fully connected layer maps the normalized features to a unified latent space.

[0034] The cross-window context extraction one-dimensional convolutional layer extracts local change patterns related to gated fusion scoring along the window number direction;

[0035] The gating weights generate a gating weight sequence corresponding to the fully connected layer output and the uniform caliber residual;

[0036] The weight normalization layer performs nonnegation and total constraint on the gating weight sequence, and performs weighted aggregation on the uniform residuals to generate a gating fusion score.

[0037] In a second aspect, the present invention provides a computer device, including a memory and a processor, wherein the memory stores a computer program, wherein when the computer program is executed by the processor, it implements any step of the chip data self-testing method based on artificial intelligence as described in the first aspect of the present invention.

[0038] Thirdly, the present invention provides a computer-readable storage medium having a computer program stored thereon, wherein: when the computer program is executed by a processor, it implements any step of the chip data self-testing method based on artificial intelligence as described in the first aspect of the present invention.

[0039] The beneficial effects of this invention are as follows: by aligning the peak time of residuals from multiple perspectives and counting corroborating evidence, cross-parameter anomaly collaborative verification is achieved, a physically interpretable evidence chain is constructed, and the confidence and anti-interference ability of composite fault detection are improved; by adaptive retesting through threshold neighborhood windows, fuzzy sample closed-loop enhanced verification is achieved, the sampling strategy is dynamically optimized and embedded in the retesting context, the false alarm and false negative rates are reduced, and a structured and traceable anomaly fingerprint is generated. Attached Figure Description

[0040] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the following description of the embodiments will be briefly introduced. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0041] Figure 1 This is a flowchart of an AI-based chip data self-testing method.

[0042] Figure 2 A schematic diagram of adaptive retesting of the neighborhood window for gating fusion scoring threshold.

[0043] Figure 3 This is a comparison chart of the gate control fusion scoring calibration before and after the retest.

[0044] Figure 4 This is a distribution chart showing the percentage of markers used in the chip data self-test. Detailed Implementation

[0045] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

[0046] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and those skilled in the art can make similar extensions without departing from the spirit of the invention. Therefore, the invention is not limited to the specific embodiments disclosed below.

[0047] Secondly, the term "one embodiment" or "embodiment" as used herein refers to a specific feature, structure, or characteristic that may be included in at least one implementation of the present invention. The phrase "in one embodiment" appearing in different places in this specification does not necessarily refer to the same embodiment, nor is it a single or selective embodiment that is mutually exclusive with other embodiments.

[0048] Reference Figures 1-4 This is one embodiment of the present invention, which provides a chip data self-testing method based on artificial intelligence, including the following steps:

[0049] S1: Collect multi-view sampling data from the chip, and perform out-of-order callback, missing test completion and jump dual-track recording to generate a self-test acquisition window;

[0050] S1.1: The chip's multi-view sampling data includes output interface data, interface area temperature data, power supply voltage sampling data, power supply current sampling data, and clock edge arrival time sampling data;

[0051] Furthermore, the system collects output data from the output interface and adds sampling time stamps and output interface identifiers; it also collects temperature data from the interface area and adds sampling time stamps and area identifiers; it collects power supply voltage and current sampling data and adds sampling time stamps and power supply channel identifiers; it collects clock edge arrival time sampling data and adds sampling time stamps and timing channel identifiers; it performs format standardization and time zone verification on the sampling time stamps; and it performs parallel aggregation and missing item registration on various types of sampling data with the same sampling time stamp to generate multi-view sampling data for the chip.

[0052] S1.2: Map the chip's multi-view sampling data to a unified sampling scale, perform out-of-order rollback and duplicate arrival merging, and add window sequence numbers to generate unified sampling scale alignment entries.

[0053] Furthermore, a unified sampling scale is established, and the sampling time stamps of the chip's multi-view sampling data are rounded and mapped. The mapping result is used as the unified sampling scale key to perform aggregation and merging on the chip's multi-view sampling data. Records with out-of-order arrival of the unified sampling scale key are sorted according to the unified sampling scale key, and the out-of-order segments are reordered to restore the time order. Records with duplicate arrival of the same unified sampling scale key are preferentially merged according to the coverage integrity of output interface output data, interface area temperature data, power supply voltage sampling data, power supply current sampling data, and clock edge arrival time sampling data, and conflict summaries are retained. The unified sampling scale key sequence is numbered by window boundary according to fixed window length and fixed sliding step size, and the window number is appended to the corresponding record to generate unified sampling scale alignment entries.

[0054] S1.3: Perform missing measurement completion on the unified sampling scale alignment entries according to the sampling scale, and write the missing measurement mark and the completion source index to generate missing measurement completion alignment entries;

[0055] Furthermore, the unified sampling scale alignment entries are traversed in the unified sampling scale order, and the chip multi-view sampling data coverage integrity of the unified sampling scale alignment entries within the same unified sampling scale is checked. For unified sampling scale alignment entries with missing items, a missing test mark is written, and the nearest valid unified sampling scale alignment entry is located as the completion source index. The chip multi-view sampling data corresponding to the completion source index is filled with nearest neighbor value or linear interpolation, while keeping the unified sampling scale key and window number unchanged. For unified sampling scale alignment entries with consecutive missing test spans exceeding the completion upper limit, an uncompleted mark is written and the missing test mark is retained, generating a missing test completion alignment entry.

[0056] The upper limit for padding is set by using the time span corresponding to the unified sampling scale as the constraint boundary and according to the maximum continuous missing test duration allowed by the chip's multi-view sampling data. For example, the upper limit for padding can be set to no more than 5 consecutive missing tests of the unified sampling scale (e.g., 5ms when the unified sampling scale is 1ms).

[0057] S1.4: Perform jump detection on missing test completion and alignment entries, retain the jump dual-track records of the original value track and the corrected value track, and perform sliding window slicing to generate a self-test acquisition window.

[0058] Furthermore, the missing measurement completion and alignment entries are traversed in a unified sampling scale order. The change amplitude of the missing measurement completion and alignment entries of adjacent unified sampling scales is calculated and checked against the jump limit value. Records with change amplitudes exceeding the jump limit value are judged as jumps. The original value track of the jump record is retained and the sampling segments before and after the jump are extracted to form the original value track segment. The corrected value track is generated for the jump record and interpolated and smoothed by the adjacent non-jump sampling segments to form the corrected value track segment. The original value track segment and the corrected value track segment are aggregated in parallel into a jump dual-track record. The jump dual-track record is sliced ​​by sliding window according to a fixed window length and a fixed sliding step size, and the in-window sequence of each window is encapsulated to generate a self-check acquisition window.

[0059] The formula for calculating the magnitude of change is:

[0060] ;

[0061] in, Indicates the range of change. Indicates the first The values ​​of the missing measurement padding alignment entries corresponding to the sampling scale. Indicates the first The values ​​of the missing measurement padding alignment entries corresponding to the sampling scale. Indicates a time interval.

[0062] It should be noted that the fixed window length is defined as the continuous time span (e.g., 2s) covered by the self-test acquisition window on the uniform sampling scale, and the fixed sliding step is defined as the time interval between the starting uniform sampling scale of adjacent self-test acquisition windows (e.g., 0.5s). The fixed window length and fixed sliding step are converted into the corresponding number of sampling points according to the sampling frequency of the uniform sampling scale and remain unchanged within the self-test acquisition window sequence.

[0063] Jump detection calculates the change amplitude by filling in missing measurements and aligning entries for adjacent uniform sampling scales, and compares it with the jump amplitude limit. Sampling scales with change amplitudes exceeding the jump amplitude limit are marked as jumps, and the original value track and the corrected value track are retained simultaneously to form a jump dual-track record.

[0064] The jump limit value is an upper limit threshold used to determine whether the change amplitude of missing measurement completion alignment items of adjacent uniform sampling scales belongs to abnormal abrupt change. For example, the jump limit value of power supply voltage sampling data can be taken as the difference between adjacent scales exceeding 0.2V.

[0065] S2: Based on the self-test acquisition window, construct output residual entries, temperature residual entries, power supply residual entries, and time sequence residual entries, and perform peak time position alignment and corroborating evidence support count registration to generate consistency gating;

[0066] S2.1: Based on the unified sampling scale of the self-test acquisition window, align the chip multi-view sampling data to the same time axis within the same window and collect them into a set of aligned segments within the window;

[0067] Furthermore, the self-test acquisition windows are traversed according to their window numbers, and the unified sampling scale at the beginning and end of each window is determined. The unified sampling scale at the beginning of the window is used as the reference for the time axis within the window. Continuous segments within the window are extracted from the chip's multi-view sampling data according to the unified sampling scale, and the unified sampling scale is converted into an offset within the window. The missing measurement fill-in alignment results are used for the sampling scale hole positions. The original value track segment and the corrected value track segment are extracted from the jump dual-track recording positions respectively, and the time axis length within the window is unified. The output data segments of the output interface, the temperature data segments of the interface area, the power supply voltage sampling data segments, the power supply current sampling data segments, and the clock edge arrival time sampling data segments are collected in parallel according to the time axis within the window to generate a set of aligned segments within the window.

[0068] S2.2: Based on the in-window aligned fragment set, construct output residual entries, temperature residual entries, power supply residual entries and time series residual entries window by window, and aggregate them into a multi-view residual entry set;

[0069] Furthermore, the system iterates through the set of aligned segments within the window. Within each self-test acquisition window, it extracts the smoothing benchmark for the output data segment from the output interface and accumulates the deviation amplitude within the window to generate output residual entries. It checks the temperature data segment in the interface area for out-of-bounds intervals and accumulates the out-of-bounds duration scale within the window to generate temperature residual entries. It locates the drop segment and surge segment for the power supply voltage sampling data segment and the power supply current sampling data segment and accumulates the amplitude and duration scale within the window to generate power supply residual entries. It checks the adjacent arrival intervals for the clock edge arrival time sampling data segment and accumulates the abnormal interval duration scale within the window to generate timing residual entries. Finally, it aggregates the output residual entries, temperature residual entries, power supply residual entries, and timing residual entries according to the window number to generate a multi-view residual entry set.

[0070] S2.3: Locate the peak time position of the output residual in the multi-view residual entry set, and perform peak neighborhood hit check to generate supporting evidence for counting;

[0071] Furthermore, the multi-view residual entry set is traversed, and the deviation of the output residual entry is scanned along the time axis within each self-test acquisition window. The window offset with the largest deviation is selected as the peak time position of the output residual entry. The peak neighborhood time range is intercepted on both sides of the peak time position of the output residual entry according to a fixed neighborhood width. Within the peak neighborhood time range, it is checked whether the temperature residual entry hits the out-of-bounds scale, whether the power supply residual entry hits the drop segment or the sudden increase segment, and whether the timing residual entry hits the abnormal interval. Each type of hit result is counted as one support and the non-hit result is counted as zero support. The number of hits of each type is summarized to generate the supporting evidence count.

[0072] S2.4: Based on the supporting evidence count, perform evidence consistency screening and grouping merging on the multi-perspective residual item set to generate consistency gating.

[0073] Furthermore, the system traverses the set of multi-perspective residual entries and reads the corresponding supporting evidence counts. For self-test acquisition windows where the supporting evidence counts meet the consistency conditions, evidence consistency screening is performed. The output residual entries are retained, and temperature residual entries, power supply residual entries, and time series residual entries that hit the peak neighborhood time range are grouped into a consistent group. For self-test acquisition windows where the supporting evidence counts do not meet the consistency conditions, evidence consistency screening is performed. The output residual entries are retained, and temperature residual entries, power supply residual entries, and time series residual entries that do not hit the peak neighborhood time range are grouped into a non-consistent group. The consistent group and the non-consistent group are respectively fixed into the same window entry set while maintaining the window number order. The consistent group and the non-consistent group of each window are aggregated and encapsulated to generate a consistency gating.

[0074] S3: Perform a unified caliber mapping on the consistency gating, and combine the gating weight modulation neural network to perform gating weight modulation, obtain the gating fusion score, and extract the threshold neighborhood unstable window number to generate a retest candidate list.

[0075] S3.1: Extract the cumulative value of residual entries from the multi-perspective residual entry set from the consistency gating to form a set of residual values ​​in the same window, and set the supporting evidence count as the gating information in the same window;

[0076] Furthermore, the system iterates through the consistency gating and locates the multi-view residual entry set corresponding to the window number within the consistency gating. It reads and outputs the cumulative values ​​of residual entries, temperature residual entries, power supply residual entries, and time series residual entries from the multi-view residual entry set and performs null value verification. For cumulative values ​​with missing items, it selects the cumulative value of the consistent group or the cumulative value of the non-consistent group according to the merging result of the same window group within the consistency gating to fill in the missing items. It aggregates the cumulative values ​​of each residual entry under the same window number to form a set of residual values ​​for the same window. At the same time, it reads the supporting evidence count corresponding to the same window number and binds the supporting evidence count to the same window gating information, generating the set of residual values ​​for the same window and the same window gating information.

[0077] It should be noted that the window gating information is a description of the intra-window consistency strength based on the number of supporting evidence corresponding to the same window number. It is used to characterize how much cross-view hit support the cumulative value of multi-view residual entries obtains within the same window for gating weight modulation.

[0078] S3.2: Perform a unified mapping on the set of residual values ​​in the same window, and perform cumulative intensity and amplitude scale alignment over unit time to generate unified residuals;

[0079] Furthermore, the system iterates through the set of residual values ​​within the same window and reads the window length time span of the self-test acquisition window. The cumulative value of each residual entry in the set of residual values ​​within the same window is divided by the window length time span to obtain the cumulative intensity per unit time while keeping the window number of the set of residual values ​​within the same window unchanged. Amplitude scale alignment is performed on the cumulative intensity per unit time. Amplitude scale alignment eliminates magnitude differences and preserves relative size patterns by truncating, limiting, and mapping the cumulative intensity per unit time within the set of residual values ​​within the same window through interval mapping. Missing measurement completion alignment entries are marked as uncompletable window numbers, and the confidence weight of the corresponding cumulative intensity per unit time is reduced and written to the scale alignment bias simultaneously. The results of completing the cumulative intensity per unit time and amplitude scale alignment are aggregated to generate a unified caliber residual.

[0080] S3.3: Based on peer gating information, gating weight modulation is performed on the uniform residual in the gating weight modulation neural network, and the results are aggregated to generate a gating fusion score;

[0081] Furthermore, the unified caliber residuals are paired with the gating information of the same window in order of window number. The unified caliber residuals are scaled and normalized and concatenated with the gating information of the same window to generate the input sequence of the gating weight modulation neural network. The input sequence of the gating weight modulation neural network is input into the gating weight modulation neural network for forward inference to obtain the gating weight sequence. Boundary clipping and adjacent window smoothing constraints are performed on the gating weight sequence. The unified caliber residuals are weighted and aggregated using the gating weight sequence to obtain the window-level weighted residual summary value, which is then encapsulated as a gating fusion score.

[0082] The formula for calculating the gating fusion score is:

[0083] ;

[0084] in, Indicates the first Window door control fusion scoring, Indicates the first Window uniform diameter residuals Indicates the first Window access control information, This represents a composite function of normalization, pruning, and smoothing operations, used to smooth and constrain gated weights, ensuring the stability of the gated weight sequence during training. This represents the output of the gated weighted modulation neural network, which is weighted based on uniform residuals and windowed gating information. This indicates the transpose operation.

[0085] S3.4: The gated weight modulation neural network includes an input normalization layer, a feature projection fully connected layer, a cross-window context extraction one-dimensional convolutional layer, a gated weight generation fully connected layer, and a weight normalization layer;

[0086] Furthermore, the gated weight modulation neural network is connected in sequence as follows: input normalization layer, feature projection fully connected layer, cross-window context extraction one-dimensional convolutional layer, gated weight generation fully connected layer, and weight normalization layer to form an end-to-end weight generation link. The input normalization layer, feature projection fully connected layer, and cross-window context extraction one-dimensional convolutional layer together constitute the gated evidence extraction link, which first compresses the multi-source residuals into comparable low-dimensional representations under the same numerical caliber, and then converges the cross-window correlation along the time window direction. The gated weight generation fully connected layer and weight normalization layer constitute the weight landing link, which maps the cross-window correlation representation to the initial weight values ​​of each gated channel and performs constraint normalization, so that the output gated weights can be directly used for gated weight modulation without weight drift and channel imbalance.

[0087] It should be noted that the training of the gated weight modulation neural network adopts a pseudo-supervised training driven by consistency gating and a joint training of smoothing constraints. The training samples are generated from multiple batches of self-check acquisition windows and sequentially construct the in-window aligned fragment set, multi-view residual entry set, corroborating evidence support count, and consistency gating. The consistent group and non-consistent group in the consistency gating are used as the source of pseudo-supervisory labels, so that the gated fusion score output by the gated weight modulation neural network shows a higher degree of aggregation in the consistent group window and a lower degree of aggregation in the non-consistent group window. At the same time, the corroborating evidence support count is used to construct ranking supervision, so that the gated fusion score corresponding to the self-check acquisition window with a higher corroborating evidence support count maintains a monotonically increasing trend. During the training process, the adjacent window smoothing penalty is applied to the gated weight sequence and non-negativity constraints and total amount constraints are applied to the gated weight sequence. After the training is completed, the network parameters are fixed for online gated weight modulation of uniform caliber residuals and output of gated fusion scores.

[0088] The gating evidence extraction chain compresses the residuals and gating information of the same window into comparable low-dimensional representations through an input normalization layer, a feature projection fully connected layer, and a cross-window context extraction one-dimensional convolutional layer, and converges the cross-window associations along the time window direction to form evidence representations that can directly drive the generation of gating weights.

[0089] S3.5: The input normalization layer performs scale normalization on the uniform residual and the same window gating information, and the feature projection fully connected layer maps the normalized features to the uniform latent space.

[0090] Furthermore, the input normalization layer receives the unified caliber residuals and the window gating information. The unified caliber residuals are centered and pruned according to the statistical range within the window, and the scale is unified. The window gating information is linearly scaled according to the gating value range, and missing items are filled in by interpolation of adjacent windows. The unified caliber residual normalization result and the window gating information normalization result are concatenated in channel order to form a normalized feature sequence. The feature projection fully connected layer performs weight mapping and bias translation on the normalized feature sequence, outputs a dimension-aligned latent space vector, and applies a fixed dimension constraint on the vector length to obtain the feature representation mapped to the unified latent space.

[0091] S3.6: Cross-window context extraction: The one-dimensional convolutional layer extracts local change patterns related to the gated fusion score along the window number direction, and the gated weights generate a gated weight sequence corresponding to the output of the fully connected layer and the uniform residual.

[0092] Furthermore, the cross-window context extraction one-dimensional convolutional layer receives the feature representation of the unified latent space, sorts the feature representations according to the window number to form a window sequence and fills in the gaps at the beginning and end, sets the convolutional receptive field and sliding stride, performs sliding convolution on the window sequence along the window number direction and outputs the local change response sequence, and performs amplitude compression and abnormal peak suppression on the local change response sequence according to the numerical range of the gated fusion score; the gated weight generation fully connected layer receives the local change response sequence, expands it window by window according to the window number as the input for weight generation and performs window-by-window mapping, outputs the gated weight vector with the same number of residual channels as the unified aperture, and concatenates the gated weight vectors of each window according to the window number to form the gated weight sequence.

[0093] S3.7: Weight normalization performs nonnegation and total constraint on the gated weight sequence, and performs weighted aggregation on the uniform residuals to generate the gated fusion score;

[0094] Furthermore, the weights are normalized by traversing the gated weight sequence and extracting the gated weight vector of each window according to the window number. The negative components of the gated weight vectors are truncated to obtain non-negative gated weight vectors. The components of the non-negative gated weight vectors are summed to obtain the total. The gated weight vectors with a total of zero are subjected to equal distribution backoff. The gated weight vectors with a total of non-zero are scaled proportionally to fix the total to a preset total. The scaling results of each window are concatenated according to the window number to form a normalized gated weight sequence. The normalized gated weight sequence is aligned with the unified caliber residual according to the window number. Channel multiplication is performed on the unified caliber residual window by window and the results are aggregated into a single value to generate a gated fusion score.

[0095] It should be noted that zero lower bound pruning directly sets the components in the gated weight vector that are less than zero to 0 to ensure that the gated weight sequence is non-negative and to avoid the negative weights canceling out the weighted convergence of the uniform residuals.

[0096] S3.8: Extract the window number that falls into the neighborhood of the data self-inspection evaluation threshold and shows up-down swing reversal, and attach the retest configuration entry to generate a retest candidate list.

[0097] Furthermore, the gated fusion scoring sequence is traversed, and the gated fusion score of each window is compared with the data self-assessment threshold for neighborhood fall-in. Neighborhood fall-in verification is achieved by comparing whether the absolute value of the difference between the gated fusion score and the data self-assessment threshold falls within the neighborhood range. The window number whose absolute value of the difference falls within the neighborhood range is recorded as the threshold neighborhood unstable window number. The gated fusion scores of several adjacent windows are extracted from the threshold neighborhood unstable window number, and the adjacent difference sign sequence is obtained. The threshold neighborhood unstable window number whose difference sign sequence changes from positive to negative or from negative to positive is determined as the up-and-down swing inversion window number. The up-and-down swing inversion window number is combined with the window length expansion factor and sliding step size reduction factor of the corresponding window number to generate a retest configuration entry and appended to the up-and-down swing inversion window number, and the retest candidate list is generated.

[0098] It should be noted that the data self-assessment threshold (example range: 0.60–0.80) and its neighborhood range can be set as the tolerance bandwidth range around the data self-assessment threshold (e.g., ±0.03–±0.08). The data self-assessment threshold is determined by the minimum overlap point of the gating fusion score distribution of healthy windows and abnormal windows in the training samples or by the quantile under the specified false alarm rate constraint. The neighborhood range is obtained by superimposing the natural fluctuation amplitude of the gating fusion score within the stable healthy window with the upper bound of the quantization noise and is used to screen out suspicious windows close to the data self-assessment threshold.

[0099] The window number where the swings and reversals occur: This refers to the unstable window number in the threshold neighborhood when the gated fusion score changes from rising to falling or from falling to rising between adjacent windows when the gated fusion score is within the data self-assessment threshold neighborhood range. It is used to locate the fluctuating window near the boundary.

[0100] Retest configuration entry: Pointer to the set of retest sampling window adjustment instructions given for each up-and-down swing inversion window number, used to determine the window length expansion factor and sliding step size reduction factor corresponding to the retest time range and drive the sampling window retest.

[0101] Figure 2 The relationship between the gating fusion scoring sequence and the data self-assessment threshold is shown with window number as the horizontal axis and score as the vertical axis. Figure 2 The full-window sequence display area is marked with a red dashed rectangle to indicate the threshold neighborhood magnification window and the interval is marked with a double arrow. This intuitively corresponds to the extraction of the unstable window number in the threshold neighborhood and generates a retest candidate list. The retest candidate list window (the landing point before retesting) is marked with scatter points, which reflects the convergence of the sampling strategy that only performs sampling on unstable windows. Figure 2The local magnified display area in the image refines the display of the defined interval, marking several feature peaks and points of maximum difference, which is used to illustrate the local correction effect of the retest gated score (after backfilling) relative to the gated fusion score (before retesting), and to demonstrate the role of "threshold neighborhood window adaptive retesting + backfilling" in combating interference and improving confidence from the perspective of the evidence chain.

[0102] S4: Based on the retest candidate list, perform sampling window retest with window length expansion and sliding step size reduction, obtain the retest self-test acquisition window, and perform residual item reconstruction and gate mark registration backfilling to generate retest gate score;

[0103] S4.1: Locate the retest time range corresponding to the unstable window number in the threshold neighborhood of the candidate list for retesting, and generate the retesting sampling window parameter set by combining the window length expansion factor and the sliding step size reduction factor;

[0104] Furthermore, the candidate list for retesting is sorted by window number and the unstable window numbers in the threshold neighborhood are traversed. Based on the unified sampling scale at the start and end of the self-test acquisition window, the retesting time range corresponding to the unstable window number in the threshold neighborhood is calculated. The retesting time range is then extended forward and backward by several unified sampling scales to cover the swing-up and reversed neighborhood. The retesting configuration entries attached to the unstable window numbers in the threshold neighborhood in the candidate list for retesting are read and the window length expansion factor and sliding step size reduction factor are parsed. The window length expansion factor is applied to the fixed window length of the self-test acquisition window to obtain the retesting window length, and the sliding step size reduction factor is applied to the fixed sliding step size of the self-test acquisition window to obtain the retesting sliding step size. The retesting time range is generated according to the retesting sliding step size to form a retesting window start scale sequence and according to the retesting window length to form a retesting window end scale sequence, which are then aggregated into a retesting sampling window parameter set.

[0105] S4.2: Perform sampling window retesting based on the retest sampling window parameter set, expanding the window length and reducing the sliding step size, and simultaneously collect multi-view data to generate a retest self-check acquisition window;

[0106] Furthermore, the parameter set of the retest sampling window is sorted by window number and the start and end scales of the retest window are expanded one by one. Within the time range from the start to the end of each retest window, the output data of the output interface, the temperature data of the interface area, the power supply voltage sampling data, the power supply current sampling data, and the clock edge arrival time sampling data are collected synchronously and the retest window number is attached. The synchronously collected chip multi-view sampling data is mapped to a unified sampling scale and the disordered reordering and duplicate arrival merging are performed. Missing test fill is performed at the hole positions of the unified sampling scale and the missing test mark and fill source index of the missing test fill alignment entry are retained. For the jump position, the jump dual track record of the original value track and the corrected value track is retained. The alignment result of each retest window is encapsulated into an in-window sequence according to the retest window number and aggregated to generate the retest self-test acquisition window.

[0107] S4.3: Residual entries are reconstructed for the retest self-test acquisition window, and peak time position alignment and supporting evidence count verification are performed to generate retest consistency gating;

[0108] Furthermore, the process iterates through the retest self-check acquisition windows and constructs an in-window alignment fragment set for each retest window. The in-window alignment fragment set is then reconstructed window by window to output residual entries, temperature residual entries, power supply residual entries, and time series residual entries, which are then aggregated into a retest multi-view residual entry set. The peak time position of the output residual entries is located within the retest multi-view residual entry set, and the time axis within the retest window is mapped to a unified sampling scale to complete peak time position alignment. The peak time position alignment result is used to extract the peak neighborhood time range, and the hit rate of temperature residual entries, power supply residual entries, and time series residual entries within the peak neighborhood time range is verified to generate a retest supporting evidence count. Evidence consistency screening and grouping are performed on the retest supporting evidence count and the retest multi-view residual entry set, and consistent and inconsistent groups are encapsulated to generate a retest consistency gating.

[0109] S4.4: Perform unified caliber mapping and gating weight modulation on the retest consistency gating, obtain the retest gating fusion score, and generate the retest gating score by replacing and backfilling the unstable window number in the threshold neighborhood.

[0110] Furthermore, the retest consistency gating is traversed by window number and a unified caliber mapping is performed. The multi-view residual entries in the retest consistency gating are aligned in magnitude according to standardization rules to generate unified caliber residual entries. The gating weight modulation is then input into the gating weight modulation neural network for forward inference to obtain the retest gating weight sequence. The retest gating weight sequence is smoothed and restricted to the range of gating weights. The unified caliber residuals are weighted according to the retest gating weight sequence and aggregated to generate the retest gating fusion score. In the retest gating fusion score sequence, the unstable window number in the neighborhood of the threshold is extracted. The retest gating fusion score sequence is then replaced and backfilled with the unstable window number in the neighborhood of the threshold to generate the retest gating score.

[0111] It should be noted that the gating weight range (example range: 0-1) is adjusted based on the target distribution and optimization objective in the training data, and the boundaries are usually set according to actual needs and error tolerance.

[0112] The retest gating score reconstructs the residuals and gating weights by using a retest self-test acquisition window with expanded window length and reduced sliding step size within the time range corresponding to the unstable window number in the threshold neighborhood. This replaces the "coarse-grained fluctuation value near the boundary" in the gating fusion score with a "robust weighted value with denser sampling and longer coverage".

[0113] Figure 3The interpretability and stability of "gated fusion score (before retesting)" and "retested gated score (after backfilling)" were compared using a binning approach. The horizontal axis represents the average score within each bin, and the vertical axis represents the actual failure rate within each bin. An ideal calibration line was also provided as a reference. Before retesting, the curve was more likely to deviate from the ideal calibration line in the middle bins, indicating that fuzzy samples representing the threshold neighborhood were more prone to generating unstable scores. After retesting and backfilling, the curve was closer to the ideal calibration line, demonstrating that retesting of the sampling window and reconstruction of residual entries could strengthen the supporting evidence for counting, making the evidence chain of consistency gating more closed. The gated fusion score's expression of the failure probability was more reliable, thereby improving the confidence of composite fault detection and reducing the false alarm and false negative rates.

[0114] S5: Perform a self-test evaluation threshold comparison on the retest gate control score, obtain the judgment mark, encapsulate the abnormal fingerprint packet, and generate a chip data self-test output set.

[0115] S5.1: Based on the retest gating score, perform a data self-inspection and evaluation threshold comparison on the gating fusion score, obtain anomaly judgment marks, health judgment marks and judgment marks to be reviewed, and integrate them into a judgment mark sequence;

[0116] Furthermore, the retest gating score sequence is traversed, and the retest gating score of each window is compared with the data self-assessment threshold. Window numbers whose retest gating scores are greater than the data self-assessment threshold and exceed the upper bound of the neighborhood range are marked as anomaly judgment marks. Window numbers whose retest gating scores are less than the data self-assessment threshold and exceed the upper bound of the neighborhood range are marked as health judgment marks. Window numbers whose retest gating scores fall within the threshold bandwidth range corresponding to the neighborhood range or whose window numbers belong to the unstable window numbers in the threshold neighborhood are marked as pending review judgment marks. The anomaly judgment marks, health judgment marks, and pending review judgment marks are organized into a judgment mark sequence according to the window number and the corresponding retest sampling window parameter set information is attached to generate the integrated judgment mark sequence.

[0117] Figure 4 The results of the self-inspection acquisition window after comparing the data self-inspection evaluation thresholds are summarized by proportion according to "healthy, fuzzy samples, interference disturbances, compound faults, and single output anomalies". The proportion of healthy samples is the majority, indicating that the consistency gating and gating fusion score maintain stable low alarms within the normal window; the proportion of fuzzy samples and interference disturbances can be used to indicate that the unstable window sequence in the threshold neighborhood can be identified and enter the retest closed loop, avoiding direct misjudgment; the proportion of compound faults and single output anomalies reflects the ability of cross-parameter residual collaborative verification to distinguish compound failures. This distribution can intuitively reflect the beneficial effects of improved false alarm and false alarm control and anti-interference capabilities.

[0118] S5.2: Based on the judgment marker sequence, backtrack to locate the abnormal window, and extract the peak time position of the multi-view residual entries, the source summary of the supporting evidence for counting, and the retest configuration entries to generate an abnormal fingerprint element set;

[0119] Furthermore, the system traverses the judgment mark sequence and filters the window numbers corresponding to the anomaly judgment marks. It maps the anomaly judgment mark window numbers to the time range of the anomaly window for retest gating score and retest consistency gating location. It associates the anomaly window time range with the retest multi-view residual item set and reads the peak time position of the output residual items. It reads the supporting evidence count and extracts the hit source summary corresponding to the supporting evidence count. The hit source summary is limited to the hit type and hit neighborhood range of temperature residual item hit, power supply residual item hit, and time series residual item hit. It associates the anomaly judgment mark window number with the retest candidate list and extracts the corresponding retest configuration item. It aggregates and encapsulates the peak time position of the output residual item, the source summary of the supporting evidence count, and the retest configuration item to generate an anomaly fingerprint element set.

[0120] S5.3: Encapsulate the abnormal fingerprint element set with the retest gating score, the comparison result of the data self-test evaluation threshold, and the retest configuration items to generate the chip data self-test output set.

[0121] Furthermore, the abnormal fingerprint element set, along with the retest gating score, the data self-inspection evaluation threshold comparison results, and the retest configuration entries, are encapsulated according to window number. The peak time position of the output residual entries in the abnormal fingerprint element set, the source summary of the supporting evidence count, and the retest configuration entries are mapped one by one to the abnormal window corresponding to the retest gating score. The abnormal fingerprint elements are marked according to the data self-inspection evaluation threshold comparison results. The markings include abnormal judgment markings, health judgment markings, and pending review judgment markings. These are aggregated into an abnormal fingerprint package and encapsulated into a chip data self-inspection output set.

[0122] It should be noted that the anomaly fingerprint packet encapsulation aggregates the anomaly fingerprint element set, the retest gating score, the data self-inspection evaluation threshold comparison results, and the retest configuration entries into a traceable unified payload by window number. This allows each anomaly window to simultaneously carry the peak time location, the summary of the source of supporting evidence counts, and the retest configuration entries for subsequent verification and location.

[0123] This embodiment also provides a computer device applicable to the chip data self-testing method based on artificial intelligence, including: a memory and a processor; the memory is used to store computer-executable instructions, and the processor is used to execute the computer-executable instructions to implement the chip data self-testing method based on artificial intelligence as proposed in the above embodiment.

[0124] The computer device can be a terminal, comprising a processor, memory, communication interface, display screen, and input devices connected via a system bus. The processor provides computing and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system and computer programs. The internal memory provides an environment for the operation of the operating system and computer programs stored in the non-volatile storage media. The communication interface is used for wired or wireless communication with external terminals; wireless communication can be achieved through Wi-Fi, carrier networks, NFC (Near Field Communication), or other technologies. The display screen can be an LCD screen or an e-ink screen. The input devices can be a touch layer covering the display screen, buttons, a trackball, or a touchpad on the computer device's casing, or an external keyboard, touchpad, or mouse.

[0125] This embodiment also provides a storage medium storing a computer program, which, when executed by a processor, implements the chip data self-testing method based on artificial intelligence as proposed in the above embodiments. The storage medium can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as Static Random Access Memory (SRAM), Electrically Erasable Programmable Read-Only Memory (EEPROM), Erasable Programmable Read Only Memory (EPROM), Programmable Red-Only Memory (PROM), Read-Only Memory (ROM), magnetic storage, flash memory, magnetic disk, or optical disk.

[0126] In summary, this invention achieves cross-parameter anomaly collaborative verification by: multi-view residual peak time alignment and corroborating evidence counting, constructing a physically interpretable evidence chain, and improving the confidence and anti-interference capability of composite fault detection; and achieves fuzzy sample closed-loop enhanced verification by adaptive retesting through threshold neighborhood window, dynamically optimizing the sampling strategy and embedding the retesting context, reducing false alarm and false negative rates and generating structured traceable anomaly fingerprints.

[0127] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A chip data self-testing method based on artificial intelligence, characterized in that: include, Collect multi-view sampling data from the chip, and perform out-of-order callback, missing test completion and jump dual-track recording to generate a self-test acquisition window; Based on the self-test acquisition window, output residual entries, temperature residual entries, power supply residual entries, and time sequence residual entries are constructed, and peak time position alignment and supporting evidence counting are performed to generate consistency gating. A unified caliber mapping is performed on the consistency gating, and the gating weight modulation is combined with the gating weight modulation neural network to obtain the gating fusion score. At the same time, the threshold neighborhood unstable window number is extracted to generate a retest candidate list. Based on the candidate list for retesting, perform sampling window retesting with window length expansion and sliding step size reduction, obtain the retest self-test acquisition window, and perform residual item reconstruction and gate mark registration backfilling to generate retest gate score; The retest gate control score performs a data self-test evaluation threshold comparison, obtains a judgment mark, encapsulates an abnormal fingerprint packet, and generates a chip data self-test output set; The steps for generating the self-test acquisition window are as follows: The chip's multi-view sampling data is mapped to a unified sampling scale, and out-of-order rollback and duplicate arrival merging are performed. At the same time, window numbers are added to generate unified sampling scale alignment entries. For entries aligned to the uniform sampling scale, perform missing measurement completion according to the sampling scale, and write the missing measurement mark and the completion source index to generate missing measurement completion alignment entries; Perform jump detection on missing and aligned items, and retain the jump dual-track record of the original value track and the corrected value track. At the same time, perform sliding window slicing to generate a self-test acquisition window. The steps for generating the consistency gating are as follows: Based on the unified sampling scale of the self-test acquisition window, the chip multi-view sampling data are aligned to the same time axis within the same window and collected into a set of aligned segments within the window. Based on the in-window aligned fragment set, output residual entries, temperature residual entries, power supply residual entries, and time series residual entries are constructed window by window and aggregated into a multi-view residual entry set; In the multi-view residual entry set, locate the peak time position of the output residual entry, and perform peak neighborhood hit check to generate supporting evidence to support the count; Based on supporting evidence, the evidence consistency screening and grouping of the multi-perspective residual item set are performed to generate a consistency gating.

2. The chip data self-testing method based on artificial intelligence as described in claim 1, characterized in that: The steps for generating the candidate list for retesting are as follows: Extract the cumulative residual values ​​of the multi-perspective residual entry set from the consistency gating to form a peer residual value set, and set the corroborating evidence support count as peer gating information; Perform a unified mapping on the set of residual values ​​from the same window, and perform cumulative intensity per unit time and amplitude scale alignment to generate unified residuals; Based on peer gating information, gating weight modulation is performed on the uniform residual in the gating weight modulation neural network, and the results are aggregated to generate a gating fusion score. Extract the window numbers of those whose gating fusion scores fall within the neighborhood of the data self-assessment threshold and exhibit up-and-down swing reversal, and attach retest configuration entries to generate a retest candidate list.

3. The chip data self-testing method based on artificial intelligence as described in claim 2, characterized in that: The steps for generating the retest gating score are as follows: The retest time range corresponding to the unstable window number in the threshold neighborhood of the candidate list for retesting is located, and the retesting sampling window parameter set is generated by combining the window length expansion factor and the sliding step size reduction factor. Based on the sampling window parameter set, the sampling window is retested by expanding the window length and reducing the sliding step size, and multi-view data is collected simultaneously to generate a retest self-check acquisition window; Residual entries are reconstructed for the retest self-inspection acquisition window, and peak time position alignment and supporting evidence count verification are performed to generate retest consistency gating. A unified caliber mapping and gating weight modulation are performed on the retest consistency gating to obtain the retest gating fusion score. The retest gating score is then generated by replacing and backfilling the unstable window number in the threshold neighborhood.

4. The chip data self-testing method based on artificial intelligence as described in claim 3, characterized in that: The steps for generating the chip data self-test output set are as follows: Based on the retested gating score, the gating fusion score is compared with the data self-inspection evaluation threshold to obtain the abnormal judgment mark, the health judgment mark and the judgment mark to be reviewed, and then integrated into the judgment mark sequence; Based on the judgment marker sequence, the abnormal window is located by backtracking, and the peak time position of the multi-view residual entries, the source summary of the supporting evidence for the count, and the retest configuration entries are extracted to generate an abnormal fingerprint element set; The abnormal fingerprint element set is encapsulated with the retest gating score, the comparison result of the data self-test evaluation threshold, and the retest configuration items to generate the chip data self-test output set.

5. The chip data self-testing method based on artificial intelligence as described in claim 2, characterized in that: The gated weight modulation neural network includes an input normalization layer, a feature projection fully connected layer, a cross-window context extraction one-dimensional convolutional layer, a gated weight generation fully connected layer, and a weight normalization layer.

6. The chip data self-testing method based on artificial intelligence as described in claim 5, characterized in that: The input is normalized to a single layer to standardize the residuals and gate control information of the same window; The feature projection fully connected layer maps the normalized features to a unified latent space. The cross-window context extraction one-dimensional convolutional layer extracts local change patterns related to gated fusion scoring along the window number direction; The gating weights generate a gating weight sequence corresponding to the fully connected layer output and the uniform caliber residual; The weight normalization layer performs nonnegation and total constraint on the gating weight sequence, and performs weighted aggregation on the uniform residuals to generate a gating fusion score.

7. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that: When the processor executes the computer program, it implements the steps of the chip data self-testing method based on artificial intelligence as described in any one of claims 1 to 6.

8. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by the processor, it implements the steps of the chip data self-testing method based on artificial intelligence as described in any one of claims 1 to 6.

Images