A signal time-frequency analysis guided small sample fault intelligent identification method

Abstract

The application provides a signal time-frequency analysis guided small sample fault intelligent identification method, and relates to the field of machine learning, comprising: performing multi-gear time-frequency transformation on original samples and making structural level difference along the resolution axis, and respectively classifying the structural features which are invariant and change across gears into stable cores and variable responses to form time-frequency resolution response fingerprints; taking the intersection of stable cores and the union of variable responses of registered samples to form a frozen prototype, and registering, freezing and only allowing contraction; increment samples are checked through independent physical anchor checking, frozen consistency checking and historical explanation consistency checking, and single components are confirmed twice and reverse checked before execution to realize contraction evolution, and the non-contractible core constitutes the lower limit of the frozen prototype protection, and is only triggered to update by equipment physical configuration change; the time-frequency resolution response fingerprints of the to-be-judged samples are matched with the frozen prototypes of each fault category, and the fault category and the judgment basis are output.

Description

Technical Field

[0001] This invention relates to the field of machine learning, specifically to a method for intelligent identification of small-sample faults guided by signal time-frequency analysis. Background Technology

[0002] Intelligent equipment fault identification is a key technology to ensure the safe and stable operation of industrial equipment. It is widely used in the operation and maintenance of key equipment such as rotating machinery, motors, and pump sets. With the advancement of sensing and signal processing technologies, time-frequency analysis has become an important development direction for equipment fault identification, and the need for fault identification in complex working conditions is becoming increasingly urgent. Existing technologies employ deep learning-based fault identification schemes, which collect vibration and current signals during equipment operation, preprocess them, and input them into a neural network. The model automatically extracts features and learns the mapping relationship between features and fault categories. The model is trained using a large number of labeled samples. The trained model performs feature extraction and classification and outputs the fault category. Existing identification methods rely on a large number of labeled samples. It is difficult to obtain fault samples and the labeling cost is high. When the number of samples is insufficient, the identification performance drops significantly. After the equipment operating status changes, the existing identification model is difficult to absorb new information while maintaining the stability of the existing discrimination results. False judgments and false omissions are easy to occur during the model update process. Summary of the Invention

[0003] (a) Technical problems to be solved

[0004] To address the shortcomings of existing technologies, this invention provides a small-sample intelligent fault identification method guided by signal time-frequency analysis. This method solves the problems mentioned in the background technology, such as reliance on a large number of labeled samples, difficulty in obtaining fault samples and high labeling costs, significant performance degradation when the number of samples is insufficient, difficulty for existing identification models to absorb new information while maintaining the stability of existing discrimination results after changes in equipment operating status, and the tendency for misjudgments and missed judgments to occur during model updates.

[0005] (II) Technical Solution

[0006] To achieve the above objectives, the present invention provides the following technical solution: a small-sample intelligent fault identification method guided by signal time-frequency analysis, comprising: S1, Sample-aware fingerprint generation: Perform multi-level time-frequency transformation on the original sample according to the resolution step to form a resolution-time-frequency representation sequence. Perform structural level difference along the resolution axis, classify the structural features that are unchanged across levels into the stable kernel, and classify the structural features that change with the level into the variable response to form a time-frequency resolution response fingerprint. S2, Category Prototype Registration: The intersection of the stable kernels of all registered samples is taken as the stable kernel basis, and the union of the variable responses of all registered samples is taken as the variable response spectrum, forming a frozen prototype. Registration is freezing and only shrinkage is allowed. S3, Incremental Sample Admission: For incremental samples that are to enter the frozen prototype update, independent physical anchor verification, frozen consistency verification and historical interpretation consistency verification are performed sequentially. Independent physical anchor verification is based on the physical mechanism of the device under test. After all verifications are passed, the sample enters the prototype update. S4, Prototype shrinkage evolution, generates candidates to be confirmed with a single incremental sample, shrinks the frozen prototype and writes it back to the prototype storage area after double confirmation and reverse verification before execution. S5, Core Guard Update, consists of an innermost essential protection zone within a stable core base that forms an unshrinkable core and a lower limit of protection for a frozen prototype. The update of the unshrinkable core is triggered only by changes in the device's physical configuration and is closed to incremental data. S6, Fault Judgment Output: Match the time-frequency resolution response fingerprint of the sample to be judged with the frozen prototype of each fault category. When the match is successful, it is judged as the corresponding fault category and the judgment basis is output.

[0007] Preferably, the feature is that: S1 includes: S11, the resolution ladder is composed of multiple fixed gears arranged in order of increasing time resolution, the number of gears and the window length parameter are determined once during the initialization of the fault identification device, the window length of adjacent gears decreases sequentially and the time resolution increases sequentially; S12, the structural differential checks the time-frequency representation of adjacent gears according to the structural differential judgment rules, if all checked items do not exceed the deviation limit, it is determined to remain unchanged, if any checked item exceeds the deviation limit, it is determined to change; S13, the fault identification device shall not perform the optimal resolution selection operation and shall not discard the time-frequency representation of any gear.

[0008] Preferably, the feature is that: S2 includes: S21, the stable core base is obtained by taking the intersection of the stable cores of all registered samples, the intersection operation is to exclude stable core components that are not jointly confirmed by all registered samples from the stable core base, and only stable structures jointly confirmed by all registered samples are included in the stable core base; S22, the variable response spectrum is obtained by taking the union of the variable responses of all registered samples, and the variable response exhibited by any registered sample is recorded in the variable response spectrum; S23, freezing is performed so that the stable core base as a whole cannot be expanded, translated, or rewritten as a whole; S24, contraction is triggered by incremental samples that pass the admission verification to remove components that are no longer jointly confirmed by subsequent samples, and the fault identification device shall not expand the stable core base based on any single registered sample and shall not incorporate the variable response of the registered sample into the stable core base.

[0009] Preferably, the feature is that: S2 further includes the establishment of an unshrinkable core, which is divided into candidate nomination, continuous confirmation, physical endorsement and solidification stages: S25, candidate nomination stage, extracting stable core components that repeatedly appear and are consistent among multiple independent registered samples from the stable cores of registered samples that have passed the admission verification as candidate components; S26, continuous confirmation stage, when a candidate component overlaps with a component in the stable core of an incremental sample, it obtains a continuous confirmation mark, and the confirmation degree accumulates when it is repeatedly confirmed by multiple independent incremental samples. It cannot enter solidification if it does not reach the preset confirmation number lower limit; S27, physical endorsement stage, the candidate component must be traceable to a physical source corresponding to the physical mechanism of the device under test, and candidate components that have been marked as suspicious by independent physical anchor verification multiple times cannot enter the unshrinkable core; S28, solidification stage, the candidate component that has passed independent physical anchor verification is solidified into the unshrinkable core and enters a permanent frozen state.

[0010] Preferably, the characteristic of S3 is that S3 includes: S31, the benchmark for independent physical anchor verification does not come from statistical models, frozen prototypes, or training data distributions, and the time-frequency structure of incremental samples is checked according to the physical mechanism of the device under test. If all conditions are met, it passes; if any condition is not met, it is judged as a physically untrustworthy sample and rejected; S32, the frozen consistency verification compares the time-frequency resolution response fingerprint of the incremental sample with the interpretation range of the frozen prototype. The interpretation range is composed of the stable structure position covered by the stable core base and the essential structure position covered by the non-shrinkable core. If all stable cores fall within the range and the variable response falls within the dynamic morphology range, it passes; if there are samples that do not fall within the range, it is judged as a sample that rewrites existing conclusions and is blocked; S33, the historical interpretation consistency verification combines the incremental sample and the frozen prototype to form a proposed update. The proposed update is executed at the simulation level, and the updated frozen prototype is used to interpret the frozen historical registered sample group. If all samples can be completely interpreted, it is allowed; if any historical registered sample cannot be completely interpreted, the proposed update is rejected.

[0011] Preferably, the feature is that: S4 includes: S41, candidate nomination stage, generating candidates to be confirmed from a single incremental sample, with granularity limited to a single stable core component and batch nomination prohibited; S42, dual confirmation stage, when no candidates to be confirmed appear in the stable cores of multiple independent incremental samples that have passed the admission verification, or when the candidates to be confirmed are determined to be physically unreliable according to independent physical anchor verification, they are confirmed to be removed; if neither is true, they are determined to be temporarily unobserved; S43, pre-execution reverse verification stage, performing removal at the simulation level and verifying whether the frozen prototype after removal fully explains all historical evidence; if it cannot fully explain, it is rejected; S44, protection lower limit stage, removing is prohibited when removal touches an unshrinkable core, and shrinkage is performed when removal passes the pre-execution reverse verification and does not touch an unshrinkable core; S45, execution and audit stage, recording shrinkage events and supporting rollback, triggering overall integrity review.

[0012] Preferably, the feature is that: S5 includes: S51, freezing the prototype to form a structure from the outside to the inside, including an outermost variable response spectrum, an intermediate stable core substrate, and an innermost non-shrinkable core. The outermost layer allows the incorporation of new dynamic forms, the intermediate layer only allows shrinkage, and the innermost layer is permanently frozen; S52, variable response incorporation is an operation performed on incremental samples that have passed the admission verification. When the variable response of the incremental sample contains a new component whose similarity to any member in the variable response spectrum is lower than a preset threshold, it is incorporated into the variable response spectrum. If the similarity is not lower than the preset threshold, it cannot be repeatedly incorporated and cannot touch the stable core substrate and the non-shrinkable core; S53, the update of the non-shrinkable core is triggered only by changes in the physical configuration of the device and is divided into core removal and core addition. Core removal is the removal of solidified components that no longer meet the independent physical anchor verification. Core addition is the inclusion of components that have passed the independent physical anchor verification multiple times and have been confirmed by the frozen historical registered sample group. The update sequentially re-instantiates the independent physical anchor verification and updates the benchmark and triggers the overall consistency review.

[0013] Preferably, S6 includes: S61, matching whether the stable kernel of the sample to be judged completely covers the stable kernel base of the frozen prototype, where complete coverage means that the overlap between the stable kernel component and the stable kernel base is not less than a preset threshold; S62, performing matching one by one according to each fault category, and judging the corresponding fault category when a match is successful. When multiple frozen prototypes are matched simultaneously, a list of fault categories is output. If the stable kernel base of any frozen prototype is not completely covered, an undiscriminable result is output; S63, the review process is to transfer the sample to be judged into the review queue and perform manual review on the suspected contamination sample. If it is confirmed as a normal sample, it re-enters the incremental sample admission. If it is confirmed as contamination, it maintains the isolation state and prohibits it from entering the prototype update; S64, outputting the fault identification conclusion and the judgment basis. The judgment basis is the matching record of the time-frequency resolution response fingerprint of the sample to be judged and the frozen prototype. The matching record contains a list of stable kernel components and fault categories. Each fault identification conclusion can be traced back to the stable kernel component in the matching record.

[0014] (III) Beneficial Effects

[0015] This invention provides a small-sample intelligent fault identification method guided by signal time-frequency analysis. It has the following beneficial effects: 1. This invention generates time-frequency resolution response fingerprints from original samples through multi-level time-frequency transformation and structural level difference. The category frozen prototype consists only of a stable kernel base and variable response spectrum of a small number of registered samples. The establishment of the discrimination benchmark does not depend on the statistical training of large-scale labeled samples. Even when the number of fault samples is insufficient, the category prototype can still be registered and used for fault discrimination, reducing the dependence on fault sample acquisition and labeling.

[0016] 2. The incremental samples to be updated in this invention are sequentially verified by independent physical anchor, frozen consistency, and historical interpretation consistency. The update is performed at the granularity of a single stable kernel component after double confirmation and reverse verification before execution. The updated frozen prototype must fully interpret the frozen historical registered sample group at the simulation level, maintaining the stability of the existing discrimination conclusion. The non-shrinkable core is closed to the incremental data and is only triggered by changes in the physical configuration of the device. The discrimination conclusion can be traced back to the stable kernel component in the matching record, reducing the occurrence of misjudgment and missed judgment. Attached Figure Description

[0017] To make the content of this invention easier to understand, the invention will be further described in detail below with reference to specific embodiments and accompanying drawings, wherein: Figure 1 This is an overall flowchart of a small-sample intelligent fault identification method guided by signal time-frequency analysis as described in this invention. Detailed Implementation

[0018] The technical solutions in the embodiments of the present invention will be clearly and completely described below. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0019] Example 1: This invention provides a method for intelligent identification of small-sample faults guided by signal time-frequency analysis, comprising: In S1, the fault identification device is a signal processing device that performs intelligent identification of small-sample faults. The fault identification device uses a piezoelectric accelerometer and a Hall current sensor in conjunction with a data acquisition card to acquire raw vibration signal samples and raw current signal samples of the device under test. The data acquisition card is a multi-channel synchronous acquisition card, with each channel starting at the same time. The raw vibration signal samples and the raw current signal samples come from different channels of the data acquisition card. The data acquisition card performs analog-to-digital conversion on the charge signal output by the piezoelectric accelerometer or the voltage signal output by the Hall current sensor. The sampling rate of the data acquisition card is set to 25.6 kHz, and the quantization bit depth is set to 16 bits. The data acquisition card outputs digitized raw samples, which are fixed-length samples extracted from continuously acquired signals, with a sample length of 4096 sampling points. The fault identification device generates a time-frequency representation corresponding to each resolution level for the same raw sample according to a resolution ladder. The resolution ladder consists of five fixed levels arranged in ascending order of time resolution, with adjacent levels having a window length... Decreasing in powers of 2, the number of fixed gears and window length parameters are determined once during the initialization of the fault identification device and do not change with the sample content. The coarsest gear among the five fixed gears is the coarsest gear, and the gear adjacent to the coarsest gear is the second coarsest gear. The finest gear among the five fixed gears is the finest gear. The window length of the coarsest gear is set to 1024 sampling points, and the window length of the finest gear is set to 64 sampling points. The fault identification device sequentially processes the original samples starting from the coarsest gear in the resolution hierarchy. The system performs a short-time Fourier transform and outputs the time-frequency spectrum corresponding to each gear level. The short-time Fourier transform uses a Hanning window, with the window length decreasing sequentially and the time resolution increasing sequentially for adjacent gear levels. During the resolution ladder generation process, the fault identification device must not perform the optimal resolution selection operation at any gear level. The fault identification device arranges the time-frequency spectrum of each gear level in ascending order of gear level and stores it as a two-dimensional array, outputting a resolution-time-frequency representation sequence specific to the original sample. This resolution-time-frequency representation sequence serves as the input for subsequent processing in S1.

[0020] In step S1, the fault identification device receives the resolution-time-frequency representation sequence. The fault identification device performs structural level difference analysis on the time-frequency representations of adjacent levels in the resolution-time-frequency representation sequence along the resolution axis. The structural level difference analysis starts between the coarsest and second-coarsest levels in the resolution ladder, and compares the time-frequency representations of each pair of adjacent levels sequentially. Before the comparison, the time-frequency representations of the adjacent levels are interpolated to a unified frequency grid. The comparison checks the spectral peak position, energy distribution pattern, time-frequency ridge direction, concentration of time-frequency ridges, and breakage of time-frequency ridges one by one according to the structural level difference judgment rules. The direction of the time-frequency ridges is the migration direction of the spectral peak center frequency over time. When the migration direction is reversed, the direction is determined to change. The concentration of the time-frequency ridge is the frequency bandwidth corresponding to the ridge peak drop of 3dB. When the bandwidth change between adjacent gears exceeds 50%, the concentration is determined to change. The breakage of the time-frequency ridge is the continuous state of the ridge on the time-frequency spectrum. When the continuity between adjacent gears changes to breakage, the breakage is determined to change. The structural level difference determination rule is that the peak position deviation does not exceed 2 times the frequency resolution, the energy distribution morphology similarity is not less than 90%, the time-frequency ridge amplitude change exceeds 50%, or any check item is determined to change, it is determined to be a morphological change. Structural features that remain unchanged during gear switching are classified into stable kernels and constitute stable kernel components. "Remaining unchanged" means all checks are performed. None of the items exceeded the deviation limit. Structural features that appeared, disappeared, migrated, or changed their form with changes in gear were classified as volatile responses. The change was defined as any check item exceeding the deviation limit. The stable kernel and the volatile response together constituted the time-frequency resolution response fingerprint of the original sample. The stable kernel corresponds to the essential structural components of the device under test in a steady state, representing essential state information that remains unchanged across resolutions, serving as a robust criterion for judging the health of the device under test. The volatile response corresponds to the transient impact components or modulation sideband components caused by the fault, representing resolution-sensitive fault dynamic information, serving as a sensitive criterion for distinguishing different fault types. The time-frequency resolution response fingerprint is entirely composed of the cross-resolution relationships within the original sample. The system generates a complete set of discriminative features without relying on statistics between samples; a single original sample produces a complete set of discriminative features; the time-frequency resolution response fingerprint is insensitive to overall translation, scaling, and offset of the entire time-frequency representation, and these overall translations, scaling, and offsets cancel each other out in the structural-level difference; the fault identification device must not perform any optimal resolution selection operation and must not discard any time-frequency representation; the fault identification device records the gear configuration and difference action used in each structural-level difference as an audit record; the fault identification device outputs the time-frequency resolution response fingerprint, which serves as the input object for the S2 category prototype registration.

[0021] In step S2, the fault identification device receives time-frequency resolution response fingerprints of multiple registered samples that have passed admission verification. Admission verification is a verification performed by the fault identification device on samples participating in prototype registration or prototype update. The registered samples are samples used for category prototype registration, and the incremental samples are samples acquired during runtime and used for prototype update after the time-frequency resolution response fingerprints are generated in step S1. The fault category is a preset fault type to be identified by the fault identification device, and the number of registered samples for each fault category is set to 2 to 5. The fault identification device takes the intersection of the stable kernels of all registered samples under the fault category as the stable kernel basis of the fault category. The intersection operation excludes stable kernel components that are not jointly confirmed by all registered samples from the stable kernel basis, and only stable structures jointly confirmed by all registered samples are included in the stable kernel basis. The fault identification device takes the union of the variable responses of all registered samples under the fault category as the variable response spectrum of the fault category, and the variable response exhibited by any registered sample is recorded in the variable response spectrum. The stable kernel basis Together with the variable response spectrum, the frozen prototype of the fault category is formed. The frozen prototype is frozen upon registration. Freezing means that the stable core base as a whole cannot be expanded, translated, or rewritten. Rewriting involves replacing or rearranging the entire set of components of the stable core base. The only permitted change to the stable core base is contraction, triggered by incremental samples that pass the admission check, to remove components that are no longer jointly confirmed by subsequent samples passing the admission check. The fault identification device may not expand the stable core base based on the stable core of any single registered sample, nor may it incorporate the variable response of the registered sample into the stable core base. After the frozen prototype registration is completed, the fault identification device writes the frozen prototype into the prototype storage area, which is the storage area in the fault identification device that stores frozen prototypes for all fault categories. The fault identification device may not perform any prototype update operation based on the sample mean during the frozen prototype registration phase. The fault identification device outputs the frozen prototype, which serves as the criterion for the admission of the S3 incremental samples.

[0022] In step S2, the fault identification device establishes an inextensible core for each fault category. This inextensible core is the innermost essential protection zone within the stable core base, representing the essential definition of the fault category. The establishment of the inextensible core is divided into four stages: candidate nomination, continuous confirmation, physical endorsement, and solidification. These four stages are completed concurrently during the registration period and the initial operation phase. In the candidate nomination stage, the fault identification device extracts stable core components that repeatedly appear and are consistent among at least three independent samples from the stable cores of registered samples and incremental samples that have passed the access verification. These stable core components constitute an inextensible core candidate set, and the stable core components in the inextensible core candidate set are denoted as candidate components. The candidate nomination stage only nominates these candidate components and does not perform any solidification actions. In the continuous confirmation stage, the fault identification device receives incremental samples that have passed the access verification. Confirmation refers to the candidate component coinciding with the components in the stable core of the incremental sample according to the structural differential judgment rule. After confirmation, the candidate component receives a continuous confirmation mark, indicating it is a candidate component repeatedly confirmed by at least three independent incremental samples. The confirmation level is gradually accumulated, measured by the number of times it is confirmed by the incremental sample. When the confirmation level does not reach the preset minimum number of confirmations (3 times), the candidate component cannot enter the solidification stage. Candidate components that show discrepancies or are no longer confirmed during operation are removed from the non-shrinkable core candidate set. In the physical endorsement stage, the independent physical anchor verification is a verification based on the physical mechanism of the device under test. The candidate component must be traceable to a physical source that clearly corresponds to the physical mechanism of the device under test, and must pass the independent physical anchor verification at least 3 times without generating a suspicious mark. Candidate components with unknown physical sources or those repeatedly marked as suspicious by the independent physical anchor verification are not allowed to enter the non-shrinkable core. In the solidification stage, candidate components that simultaneously meet the aforementioned repeated joint confirmation conditions and pass the independent physical anchor verification at least 3 times are officially solidified into the non-shrinkable core. The solidification action is recorded as a core event, and the solidified component immediately enters a permanent frozen state. The non-shrinkable core serves as the protection lower limit for the frozen prototype. The fault identification device outputs the non-shrinkable core and stores the non-shrinkable core and the frozen prototype together in the prototype storage area.

[0023] In step S3, the fault identification device performs independent physical anchor verification on the incremental samples to be updated to the frozen prototype. The benchmark for the independent physical anchor verification does not come from the statistical model, the frozen prototype, or the training data distribution. The fault identification device acquires the time-frequency resolution response fingerprint of the incremental sample. The volatile response is divided into impulse fault components and modulation fault components. When the time-domain interval is uniformly periodically distributed, it is determined to be an impulse fault component. When the frequency domain shows symmetrical sidebands around the carrier frequency, it is determined to be a modulation fault component. The fault identification device verifies the stable structural position and the deterministic characteristic frequency position. The deterministic characteristic frequency is the inherent characteristic frequency of the device under test in a deterministic operating state. When the deviation does not exceed twice the frequency resolution, the position is determined to be consistent. The fault identification device verifies the time-domain interval of the impulse fault component. When the time-domain interval is uniformly periodically distributed and the deviation from the theoretical value determined by the structural parameters and operating parameters of the device under test does not exceed 5%, it is determined to satisfy periodicity. The fault identification device verifies the sidebands of the modulation fault component. When the symmetric deviation of the sidebands about the carrier frequency does not exceed twice the frequency resolution, it is determined to satisfy symmetry. The fault identification device checks the cross-gear evolution; if the energy amplitude change between adjacent gears exceeds 50%, or if it appears and disappears without intermediate transition in the structural differential, it determines it to be discontinuous. The fault identification device checks the energy source; if the energy is concentrated in a region outside the frequency band corresponding to the deterministic characteristic frequency, or is unrelated to the physical mechanism of the tested equipment, it determines it to be inexplicable. If any of the above five physical laws are not satisfied, the fault identification device determines the incremental sample to be a physically unreliable sample and performs a rejection operation, only writing it into the sample file, which stores all... The received samples and their recorded storage area; if all five physical laws mentioned above are satisfied, the incremental sample passes the independent physical anchor verification. The deterministic characteristic frequency is jointly determined by the nameplate parameters and field measurement parameters of the device under test. The nameplate parameters are the structural parameters marked by the device under test at the factory, and the field measurement parameters are the parameters obtained by on-site measurement of the device under test. Each device under test instantiates the independent physical anchor verification according to its own parameters, while the rules of the independent physical anchor verification remain unchanged. The independent physical anchor verification serves as the first gate for the S3 incremental sample admission.

[0024] In step S3, the incremental sample that passes the independent physical anchor verification enters the freeze consistency verification. The freeze consistency verification is a verification performed by the fault identification device comparing the time-frequency resolution response fingerprint of the incremental sample with the interpretation range of the frozen prototype. The interpretation range is the time-frequency structure range jointly defined by the stable core base and the non-shrinkable core of the frozen prototype. The interpretation range is composed of the stable structure position covered by the stable core base and the essential structure position covered by the non-shrinkable core. The fault identification device checks whether all stable cores of the incremental sample fall within the stable structure position covered by the stable core base. "All falling within" means that the overlap between the stable core component of the incremental sample and the stable structure position is not less than 90%. The fault identification device checks whether the variable response of the incremental sample falls within the dynamic morphological range covered by the variable response spectrum. "Falling within" means that the similarity between the variable response and any member of the variable response spectrum is not less than 90%. If all the above checks are true, the incremental sample passes the freeze consistency verification. If any stable core component does not fall within the stable structure covered by the stable core base, the verification fails. If the location or any of the variable responses does not fall within the dynamic morphological range covered by the variable response spectrum, the fault identification device determines that the incremental sample is a sample attempting to rewrite existing conclusions. The fault identification device performs an interception operation on the sample attempting to rewrite existing conclusions, and the interception operation prevents the incremental sample from entering the prototype update. The fault identification device writes the incremental sample into the sample file. Contamination refers to the shift in the judgment basis caused by the inclusion of a misjudged sample in the prototype update. The fault identification device marks the incremental sample as a suspected contamination sample and transfers the suspected contamination sample to the review queue. The review queue is a queue for isolating and storing suspected contamination samples. The review queue is stored in the sample file, and the suspected contamination samples in the review queue do not participate in any prototype update. The frozen consistency check relies on the frozen prototype and protects the integrity of the frozen prototype. The frozen boundary is the outer boundary of the interpretation range. The frozen consistency check prevents any incremental sample from crossing the frozen boundary and tampering with the category essence. The frozen consistency check serves as the second gate for the S3 incremental sample admission.

[0025] In step S3, the incremental sample that passes the frozen consistency check enters the historical interpretation consistency check. The historical interpretation consistency check verifies the interpretability of the frozen historical registration sample group using the updated frozen prototype. This check is performed on the proposed update action triggered by the incremental sample. The fault identification device combines the incremental sample with the frozen prototype to form a proposed update. This proposed update is a candidate update action that conforms to the stable kernel basis's rule of only allowing shrinkage. The proposed update only involves the shrinkage of the stable kernel basis and does not involve the volatile response spectrum. The incorporation of the volatile response spectrum does not weaken the interpretability of the frozen historical registration sample group. The fault identification device executes the proposed update at the simulation level. The simulation level is the process where the fault identification device copies a copy of the frozen prototype in memory, latches the prototype storage area, and then executes the proposed update. The fault identification device uses the updated frozen prototype to interpret the time-frequency resolution response fingerprint of each historical registration sample in the frozen historical registration sample group. The frozen historical registration sample group consists of historical samples that have completed registration and passed all the access checks. The set of historical samples refers to the samples received before the incremental sample. The frozen historical registration sample group is stored in the sample file. A complete interpretation means that the interpretation range of the updated frozen prototype covers the stable kernel and volatile response of the historical registration samples, and the coverage is an overlap of not less than 90%. If the updated frozen prototype can still completely interpret the time-frequency resolution response fingerprint of all the historical registration samples in the frozen historical registration sample group, the historical interpretation consistency check passes, the fault identification device allows the proposed update, and the approved proposed update is executed by S4. If the updated frozen prototype has any historical registration sample whose time-frequency resolution response fingerprint can no longer be completely interpreted, the historical interpretation consistency check fails, the fault identification device rejects the proposed update, and the rejection operation prohibits the proposed update from entering execution. The historical interpretation consistency check blocks the spread of pollution by rewriting historical conclusions. The independent physical anchor check, the frozen consistency check, and the historical interpretation consistency check together constitute the admission check, and the historical interpretation consistency check serves as the third gate for the incremental sample admission in S3.

[0026] In step S4, incremental samples that pass the admission verification enter the intersection shrinkage determination process. This intersection shrinkage determination consists of five stages: candidate nomination, double confirmation, pre-execution reverse verification, protection lower limit, execution, and auditing. In the candidate nomination stage, the fault identification device generates candidates to be confirmed using a single incremental sample. These candidates are single stable kernel components in the stable kernel base that are no longer confirmed by the incremental sample. The candidate nomination stage does not immediately change the stable kernel base and limits the granularity of the candidates to be confirmed to a single component, prohibiting batch nomination. In the double confirmation stage, the fault identification device performs multi-source co-disappearance confirmation and... The physical anchor reverse confirmation has two paths. The multi-source co-disappearance confirmation is that the component corresponding to the candidate to be confirmed no longer appears in the stable kernels of at least three independent incremental samples, each having passed the admission verification. The physical anchor reverse confirmation is that the component corresponding to the candidate to be confirmed is determined to be physically unreliable according to the independent physical anchor verification. If the multi-source co-disappearance confirmation or the physical anchor reverse confirmation is successful, the candidate to be confirmed is removed. If both the multi-source co-disappearance confirmation and the physical anchor reverse confirmation are unsuccessful, the candidate to be confirmed is determined to be temporarily unobserved, and the fault identification device does not... The process involves several steps: First, a pre-execution reverse verification step. The fault identification device performs the removal at the simulation level and checks whether the frozen prototype after removal can still fully explain all historical evidence. If it cannot, the removal is rejected, and the contraction action undergoes another independent physical anchor verification before execution. Second, a protection lower limit step. The fault identification device checks whether the removal touches the non-contractable core. If the removal touches the non-contractable core, the fault identification device prohibits the removal. If the removal passes the pre-execution reverse verification and does not touch the non-contractable core, the fault identification device performs the removal and contracts the frozen prototype. The prototype is written back to the prototype storage area; during the execution and auditing phase, the fault identification device records a shrinkage event with a complete context. The shrinkage event supports rollback and triggers an overall integrity review. If the overall integrity review finds inconsistencies, the fault identification device restores the frozen prototype before the update based on the shrinkage event. The intersection shrinkage determination prohibits batch operations, merge operations, and overall operations. The batch operation is the operation of nominating multiple candidates to be confirmed at once. The merge operation is the operation of merging multiple candidates to be confirmed into a single object to be removed. The overall operation is the operation of shrinking the entire stable core base.

[0027] In step S5, the frozen prototype forms a three-layer structure from the outside in. This three-layer structure includes an outermost variable response spectrum, a middle stable core substrate, and an innermost non-shrinkable core. The outermost variable response spectrum allows the incorporation of new dynamic morphologies, the middle stable core substrate only allows contraction and prohibits expansion, and the innermost non-shrinkable core is permanently frozen. Variable response incorporation is an operation performed by the fault identification device on incremental samples that have passed the access verification. If the variable response of the incremental sample contains a new component with a similarity of less than 90% to any member of the variable response spectrum, the fault identification device incorporates the new component into the variable response spectrum. If the similarity of the new component to any member of the variable response spectrum is not less than 90%, the fault identification device does not re-incorporate it. Variable response incorporation must not touch the stable core substrate or the non-shrinkable core. Updates to the non-shrinkable core are triggered only by changes in the device's physical configuration. These changes are entered and confirmed by maintenance personnel through the input interface of the fault identification device. These changes include component replacement or structural modification. The core update is divided into two controlled forms: core removal and core addition. Core removal involves removing at least three fixed components that no longer meet the independent physical anchor verification. Core addition involves incorporating new components that have passed the independent physical anchor verification at least three times and have been confirmed by the frozen historical registration sample group at least three times into the non-shrinkable core. The fault identification device executes a five-step mandatory process: Step 1, the fault identification device confirms that the only legitimate trigger source is the change in the device's physical configuration; Step 2, the fault identification device re-instantiates the independent physical anchor verification, which involves recalculating the deterministic feature frequency according to the changed parameters of the device under test and updating the benchmark of the independent physical anchor verification; Step 3, the fault identification device performs a global reverse verification of frozen historical evidence; Step 4, the fault identification device records the complete core change event; Step 5, the fault identification device triggers a global consistency review of the frozen prototypes for all fault categories. During normal operation, the non-shrinkable core is completely closed to incremental data, and the fault identification device must not trigger the update of the non-shrinkable core with data samples or ordinary operating events.

[0028] In step S6, the fault identification device uses the frozen prototype as a discrimination benchmark to perform fault discrimination on the sample to be judged. The sample to be judged is the sample whose time-frequency resolution response fingerprint was generated in step S1 and which did not participate in prototype updates. The fault identification device matches the time-frequency resolution response fingerprint of the sample to be judged with the frozen prototype of each fault category. The matching is to check whether the stable kernel of the sample to be judged completely covers the stable kernel base of the frozen prototype. Complete coverage means that the overlap between the stable kernel component of the sample to be judged and the stable kernel base is not less than 90%. The fault identification device performs the matching one by one according to each fault category and records the result of each matching. When the matching is successful, the fault identification device identifies the sample to be judged as the corresponding fault category and outputs the discrimination result. If the sample to be judged matches multiple frozen prototypes at the same time, the fault identification device outputs a list of matched fault categories. If the stable kernel of the sample to be judged fails to completely cover the stable kernel base of any frozen prototype, an undiscriminable result is output. The undiscriminable result triggers a review process. The review process is to check whether the stable kernel of the sample to be judged completely covers the stable kernel base of any frozen prototype. The process of transferring samples to the review queue is described. The fault identification device performs manual review on suspected contamination samples and unidentifiable samples in the review queue. The manual review is conducted independently by at least two operators. If the manual review confirms that the sample is normal, the suspected contamination sample can re-enter the incremental sample access in S3. If the manual review confirms that contamination exists, the fault identification device maintains the isolation state of the suspected contamination sample and prohibits it from entering any prototype update. The fault identification device outputs a fault identification conclusion and a judgment basis. The judgment basis is the matching record of the time-frequency resolution response fingerprint of the sample to be judged and the frozen prototype. The matching record includes a list of matched stable kernel components, the overlap value between the stable kernel component and the stable kernel base, and the fault category hit. The fault identification conclusion and the judgment basis together constitute a self-verifiable output. The self-verifiable output means that each fault identification conclusion can be traced back to the specific stable kernel component in the matching record. The fault identification device writes the judgment result and the judgment basis into the audit record and outputs them synchronously.

[0029] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A method for intelligent identification of small-sample faults guided by signal time-frequency analysis, characterized in that, include: S1, Sample-aware fingerprint generation: Perform multi-level time-frequency transformation on the original sample according to the resolution step to form a resolution-time-frequency representation sequence. Perform structural level difference along the resolution axis, classify the structural features that are unchanged across levels into the stable kernel, and classify the structural features that change with the level into the variable response to form a time-frequency resolution response fingerprint. S2, Category Prototype Registration: The intersection of the stable kernels of all registered samples is taken as the stable kernel basis, and the union of the variable responses of all registered samples is taken as the variable response spectrum, forming a frozen prototype. Registration is freezing and only shrinkage is allowed. S3, Incremental Sample Admission: For incremental samples that are to enter the frozen prototype update, independent physical anchor verification, frozen consistency verification and historical interpretation consistency verification are performed sequentially. Independent physical anchor verification is based on the physical mechanism of the device under test. After all verifications are passed, the sample enters the prototype update. S4, Prototype shrinkage evolution, generates candidates to be confirmed with a single incremental sample, shrinks the frozen prototype and writes it back to the prototype storage area after double confirmation and reverse verification before execution. S5, Core Guard Update, consists of an innermost essential protection zone within a stable core base that forms an unshrinkable core and a lower limit of protection for a frozen prototype. The update of the unshrinkable core is triggered only by changes in the device's physical configuration and is closed to incremental data. S6, Fault Judgment Output: Match the time-frequency resolution response fingerprint of the sample to be judged with the frozen prototype of each fault category. When the match is successful, it is judged as the corresponding fault category and the judgment basis is output.

2. The method for intelligent identification of small-sample faults guided by signal time-frequency analysis according to claim 1, characterized in that: S1 includes: S11, the resolution ladder consists of multiple fixed levels arranged in order of increasing time resolution, the number of levels and the window length parameter are determined once during the initialization of the fault identification device, the window length of adjacent levels decreases sequentially and the time resolution increases sequentially; S12, the structural differential checks the time-frequency representation of adjacent levels according to the structural differential judgment rules, if all checked items do not exceed the deviation limit, it is determined to remain unchanged, if any checked item exceeds the deviation limit, it is determined to change; S13, the fault identification device shall not perform the optimal resolution selection operation and shall not discard the time-frequency representation of any level.

3. The method for intelligent identification of small-sample faults guided by signal time-frequency analysis according to claim 2, characterized in that: S2 includes: S21, the stable core base is obtained by taking the intersection of the stable cores of all registered samples, and the intersection operation excludes stable core components that are not jointly confirmed by all registered samples from the stable core base, and only stable structures jointly confirmed by all registered samples are included in the stable core base; S22, the variable response spectrum is obtained by taking the union of the variable responses of all registered samples, and the variable response exhibited by any registered sample is recorded in the variable response spectrum; S23, freezing is performed so that the stable core base as a whole cannot be expanded, translated, or rewritten as a whole; S24, contraction is triggered by incremental samples that pass the admission verification to remove components that are no longer jointly confirmed by subsequent samples, and the fault identification device shall not expand the stable core base based on any single registered sample and shall not incorporate the variable response of the registered sample into the stable core base.

4. The method for intelligent identification of small-sample faults guided by signal time-frequency analysis according to claim 3, characterized in that: S2 further includes the establishment of an unshrinkable core, which is divided into candidate nomination, continuous confirmation, physical endorsement and solidification stages: S25, candidate nomination stage, extracting stable core components that repeatedly appear and are consistent among multiple independent registered samples from the stable cores of registered samples that have passed the admission verification as candidate components; S26, continuous confirmation stage, when a candidate component overlaps with a component in the stable core of an incremental sample, it obtains a continuous confirmation mark, and the confirmation degree accumulates when it is repeatedly confirmed by multiple independent incremental samples. It cannot enter solidification if it does not reach the preset confirmation number lower limit; S27, physical endorsement stage, candidate components must be traceable to a physical source corresponding to the physical mechanism of the device under test, and candidate components that have been marked as suspicious by independent physical anchor verification multiple times cannot enter the unshrinkable core. S28, the solidification stage, where candidate components verified by independent physical anchors are solidified into the non-shrinkable core and enter a permanently frozen state.

5. The method for intelligent identification of small-sample faults guided by signal time-frequency analysis according to claim 4, characterized in that... The S3 The process includes: S31, where the benchmark for independent physical anchor verification does not come from statistical models, frozen prototypes, or training data distributions. The incremental sample's time-frequency structure is checked against the physical mechanism of the device under test. If all criteria are met, the sample passes; if any criterion is not met, it is deemed a physically unreliable sample and rejected. S32, where frozen consistency verification compares the time-frequency resolution response fingerprint of the incremental sample with the interpretation range of the frozen prototype. The interpretation range consists of the stable structure position covered by the stable core base and the essential structure position covered by the non-shrinkable core. If all stable cores fall within the range and the variable response falls within the dynamic morphology range, the sample passes; if any of these do not fall within the range, the sample is deemed to be rewriting existing conclusions and is blocked. S33, where historical interpretation consistency verification combines the incremental sample with the frozen prototype to form a proposed update. The proposed update is executed at the simulation level, and the updated frozen prototype is used to interpret the frozen historical registered sample group. If all samples can be fully interpreted, the sample passes; if any historical registered sample cannot be fully interpreted, the proposed update is rejected.

6. The method for intelligent identification of small-sample faults guided by signal time-frequency analysis according to claim 5, characterized in that: S4 includes: S41, candidate nomination stage, generating candidates to be confirmed from a single incremental sample, with granularity limited to a single stable core component and batch nomination prohibited; S42, double confirmation stage, when no candidates to be confirmed appear in the stable cores of multiple independent incremental samples that have passed the admission verification, or when candidates to be confirmed are determined to be physically unreliable according to independent physical anchor verification, they are confirmed to be removed; if neither is true, they are determined to be temporarily unobserved; S43, pre-execution reverse verification stage, performing removal at the simulation level and verifying whether the frozen prototype after removal fully explains all historical evidence; if it cannot fully explain, it is rejected; S44, protection lower limit stage, removal is prohibited when it touches an unshrinkable core, and removal is performed when it passes the pre-execution reverse verification and does not touch an unshrinkable core; S45, execution and audit stage, recording the shrinkage event and supporting rollback, triggering an overall integrity review.

7. The method for intelligent identification of small-sample faults guided by signal time-frequency analysis according to claim 6, characterized in that: S5 includes: S51, freezing the prototype to form a structure from the outside in, including an outermost variable response spectrum, an intermediate stable core base and an innermost non-shrinkable core. The outermost layer allows the incorporation of new dynamic forms, the intermediate layer only allows shrinkage and the innermost layer is permanently frozen; S52, variable response incorporation is an operation performed on incremental samples that have passed the admission verification. When the variable response of an incremental sample contains a new component whose similarity to any member in the variable response spectrum is lower than a preset threshold, it is incorporated into the variable response spectrum. If the similarity is not lower than the preset threshold, it cannot be repeatedly incorporated and cannot touch the stable core base and the non-shrinkable core; S53, the update of the non-shrinkable core is triggered only by changes in the device's physical configuration and is divided into core removal and core addition. Core removal is the removal of solidified components that no longer meet the independent physical anchor verification. Core addition is the inclusion of components that have passed the independent physical anchor verification multiple times and have been confirmed by the frozen historical registered sample group. The update re-instantiates the independent physical anchor verification and updates the benchmark and triggers the overall consistency review.

8. The method for intelligent identification of small-sample faults guided by signal time-frequency analysis according to claim 7, characterized in that: S6 includes: S61, matching whether the stable kernel of the sample to be judged completely covers the stable kernel base of the frozen prototype, where complete coverage means that the overlap between the stable kernel component and the stable kernel base is not less than a preset threshold; S62, performing matching one by one according to each fault category, when a match is successful, it is judged as the corresponding fault category, and when multiple frozen prototypes are matched at the same time, a list of fault categories is output, and when the stable kernel base of any frozen prototype is not completely covered, an indiscriminate result is output; S63, the review process is to transfer the sample to be judged into the review queue and perform manual review on the suspected contamination sample, when it is confirmed as a normal sample, it re-enters the incremental sample admission, when it is confirmed as contamination, it maintains the isolation state and prohibits it from entering the prototype update; S64, outputting the fault identification conclusion and the judgment basis, the judgment basis is the matching record of the time-frequency resolution response fingerprint of the sample to be judged and the frozen prototype, the matching record contains a list of stable kernel components and fault categories, and each fault identification conclusion can be traced back to the stable kernel component in the matching record.

Images