A Meta-learning-based Intelligent Fault Diagnosis Method and System for Cross-Domain Few-Shot Processes

Abstract

This invention discloses a cross-domain few-sample intelligent fault diagnosis method and system based on meta-learning, belonging to the field of intelligent fault diagnosis technology. The proposed scheme includes acquiring source domain fault samples, a small number of support samples in the target domain, and target domain samples to be diagnosed; uniformly extracting acoustic and vibration fragments to obtain source domain fragment representations, support fragment representations, and samples to be diagnosed; extracting the fault category of the source domain fault samples; and performing intra-class consistency aggregation on the source domain fragment representations according to the fault category to obtain the source domain class prototype. This application addresses the technical deficiency of existing meta-learning and domain adaptation methods, which mix scene-shared offsets and fault discrimination information for feature alignment or fine-tuning, causing the class relationships formed by a small number of target samples to be influenced by the shared background.

Description

Technical Field

[0001] This invention relates to the field of intelligent fault diagnosis technology, and more specifically, this application relates to a cross-domain few-sample intelligent fault diagnosis method and system based on meta-learning. Background Technology

[0002] In intelligent fault diagnosis of industrial equipment, acoustic and vibration signals are often used to reflect the operating status of components such as bearings, transmission mechanisms, motors, and pumps and valves. In actual deployment, the model usually needs to be migrated from existing equipment or historical operating conditions to new equipment, installation location, load conditions, or acquisition environment. However, there are often few labeled fault samples available in the target scenario. Therefore, meta-learning, transfer learning, or domain adaptation methods are usually adopted in this field to enable existing models to quickly adapt using a small number of target samples.

[0003] Existing solutions generally use a small number of samples in the target scene as the basis for adaptation. Through feature alignment, prototype correction, or fine-tuning of a small number of samples, the target samples are made closer to the existing category distribution in the feature space. The premise for the applicability of this type of method is that the difference between the target samples and historical samples mainly reflects the change in fault performance under the new scene. Therefore, absorbing this difference helps to improve the recognition ability of the target scene.

[0004] However, in conventional industrial acoustic and vibration acquisition, different fault samples within the same target scene are often affected by the equipment structure transmission path, sensor installation status, background load, and acquisition link. These effects will manifest as similar scene-based shifts in multiple fault categories, but they do not necessarily correspond to specific fault types. If existing methods use this type of shift along with the fault information that is actually used to distinguish the categories for feature alignment or fine-tuning, it may weaken the separability between different fault categories while reducing scene differences. This will cause the category relationship formed by a small number of target samples to be pulled by the common background, which will lead to misjudgment or decreased recognition stability of subsequent samples to be diagnosed. Therefore, a cross-domain few-sample intelligent fault diagnosis method and system based on meta-learning is proposed to solve this problem. Summary of the Invention

[0005] To address the aforementioned technical problems, this paper provides a cross-domain few-sample intelligent fault diagnosis method and system based on meta-learning. This technical solution solves the problems mentioned in the background section.

[0006] To achieve the above objectives, the technical solution of the present invention is as follows:

[0007] Firstly, this application provides a cross-domain few-shot intelligent fault diagnosis method based on meta-learning, the method comprising:

[0008] We acquire source domain fault samples, a small number of target domain support samples, and target domain samples to be diagnosed, and extract acoustic and vibration segments in a unified manner to obtain source domain segment representations, support segment representations, and samples to be diagnosed.

[0009] Extract the fault categories from the source domain fault samples, and perform intra-class consistency aggregation of the source domain fragment representations according to the fault categories to obtain the source domain class prototype;

[0010] The supporting fragment representation is compared with the prototype of the same source domain class, and the deviation part relative to the prototype of the source domain class is extracted as the target domain residual representation. It is then cross-compared according to the fault category. The common deviation part that appears repeatedly in multiple fault categories and whose difference between categories is lower than the preset residual retention condition is selected as the cross-class co-occurrence domain residual.

[0011] The cross-class co-occurrence domain residues are removed from the supporting fragment representations, and the parts whose distance relationship with the prototype of the same source domain class satisfies the preset same-class preservation condition are retained. Based on this, the target domain class reference is generated.

[0012] For each supporting fragment representation, calculate its distance to the class reference of the same type of target domain and the nearest class reference of the different type of target domain, and form a boundary reference set;

[0013] Based on the common deviation direction of the cross-class co-occurrence domains remaining in the support fragment representation, a candidate domain alignment processing amount is generated. This amount is then subtracted from the support fragment representation to obtain the candidate support fragment representation. Finally, the distance to the class reference of the same class target domain is subtracted from the distance to the candidate support fragment representation to obtain the candidate class boundary interval.

[0014] If the candidate category boundary interval satisfies the preservation condition corresponding to the boundary reference set, then the common deviation part in the support fragment representation is deducted according to the candidate domain alignment processing amount and the target domain class reference is regenerated; otherwise, the target domain class reference remains unchanged.

[0015] The common deviation portion in the representation of the segment to be diagnosed is subtracted by the candidate domain alignment processing amount, and the fault diagnosis result is obtained based on the category boundary interval between the subtracted representation of the segment to be diagnosed and the target domain class reference.

[0016] Secondly, this application provides a cross-domain few-shot fault intelligent diagnosis system based on meta-learning, used to implement the aforementioned cross-domain few-shot fault intelligent diagnosis method based on meta-learning, including:

[0017] The segment representation extraction module is used to acquire source domain fault samples, a small number of target domain support samples, and target domain samples to be diagnosed, and to extract acoustic and vibration segments in a unified manner to obtain source domain segment representations, support segment representations, and samples to be diagnosed.

[0018] The class prototype aggregation module is used to extract the fault categories of source domain fault samples, perform intra-class consistency aggregation of source domain fragment representations according to fault categories, and obtain source domain class prototypes.

[0019] The common residue extraction module is used to compare the supporting fragment representation with the prototype of the same source domain class, extract the deviation part relative to the prototype of the source domain class as the target domain residue representation, and cross-compare it according to the fault category. The common deviation part that appears repeatedly in multiple fault categories and the difference between categories is lower than the preset residue retention condition is selected as the cross-class co-occurrence domain residue.

[0020] The target reference generation module is used to subtract cross-class co-occurrence domain residues from the supporting fragment representation, retain the part whose distance relationship with the prototype of the same source domain class satisfies the preset same-class preservation condition, and generate the target domain class reference accordingly.

[0021] The boundary set construction module is used to represent each supporting fragment, calculate its distance to the class reference of the same type of target domain and the nearest class reference of the opposite type of target domain, and form a boundary reference set;

[0022] The candidate processing evaluation module is used to generate candidate domain alignment processing amount based on the common deviation direction of cross-class co-occurrence domains remaining in the support fragment representation, subtract it from the support fragment representation to obtain the candidate support fragment representation, and subtract the distance to the same class target domain reference from the distance of the candidate support fragment representation to the nearest heterogeneous target domain class reference to obtain the candidate class boundary interval.

[0023] The reference update decision module is used to deduct the common deviation part in the support fragment representation and regenerate the target domain class reference if the candidate category boundary interval meets the preservation condition corresponding to the boundary reference set; otherwise, the target domain class reference remains unchanged.

[0024] The fault diagnosis module is used to subtract the common deviation part in the representation of the segment to be diagnosed by the candidate domain alignment processing amount, and obtain the fault diagnosis result based on the category boundary interval between the subtracted segment representation and the target domain class reference.

[0025] Thirdly, this application provides a computer device, the computer device including a memory and a processor, the memory storing code, and the processor being configured to acquire the code and execute the above-described meta-learning-based cross-domain few-sample intelligent fault diagnosis method.

[0026] Fourthly, this application provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the above-described cross-domain few-sample intelligent fault diagnosis method based on meta-learning.

[0027] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0028] This application addresses the technical shortcomings of existing meta-learning and domain adaptation methods, which mix scene common offsets and fault discrimination information and use them together for feature alignment or fine-tuning, resulting in the class relationship formed by a small number of target samples being pulled by the common background. This method first identifies the deviations that repeatedly occur in multiple fault categories as anchor points as cross-class co-occurrence domain residues, and then generates candidate domain alignment processing based on these anchor points.

[0029] This application solves the technical defect of existing methods that lack a class separability check link during the adaptation process, which weakens the separability between different fault categories while absorbing differences. By checking the execution result of each candidate domain alignment process one by one with the retention conditions corresponding to the candidate category boundary interval and the boundary reference set, and keeping the target domain class reference unchanged when the retention conditions are not met, this application solves the technical defect of existing methods that lack a class separability check link during the adaptation process.

[0030] This application addresses the technical shortcomings of existing methods that continuously and unconditionally update the reference when operating conditions drift or abnormal batches arrive, causing the identification stability to decrease with the accumulation of adaptation times. It constructs a closed-loop feedback optimization path for the candidate domain alignment process as "generation-trial deduction-condition verification-maintaining or updating the target domain class reference". This solves the problem of the existing methods continuously and unconditionally updating the reference when operating conditions drift or abnormal batches arrive. It realizes round-by-round self-verification and back-off fallback in the adaptation process, and achieves a balance between the rapid adaptation of a small number of samples under new equipment and new operating conditions and the stability of long-term continuous diagnosis. Attached Figure Description

[0031] The disclosure of this invention is illustrated with reference to the accompanying drawings. It should be understood that the drawings are for illustrative purposes only and are not intended to limit the scope of protection of this invention. Wherein:

[0032] Figure 1 This is a flowchart of the cross-domain few-shot intelligent fault diagnosis method based on meta-learning proposed in this invention;

[0033] Figure 2 This is a block diagram of the cross-domain few-sample intelligent fault diagnosis system based on meta-learning proposed in this invention. Detailed Implementation

[0034] It is readily understood that, based on the technical solution of this invention, those skilled in the art can propose various interchangeable structural methods and implementations without altering the essential spirit of the invention. Therefore, the following detailed embodiments and accompanying drawings are merely illustrative examples of the technical solution of this invention and should not be considered as the entirety of the invention or as limitations or restrictions on the technical solution of this invention.

[0035] Reference Figure 1 As shown, this application proposes a cross-domain few-shot intelligent fault diagnosis method based on meta-learning, including:

[0036] We acquire source domain fault samples, a small number of target domain support samples, and target domain samples to be diagnosed, and extract acoustic and vibration segments in a unified manner to obtain source domain segment representations, support segment representations, and samples to be diagnosed.

[0037] It should be noted that the source domain fault samples are acoustic and vibration signals with fault category labels collected from existing equipment or under historical operating conditions. The target domain support samples are acoustic and vibration signals with only a few labels for each type of fault collected from new equipment, new installation locations, or new operating conditions. The target domain samples to be diagnosed are acoustic and vibration signals without labels in the target domain. The same preprocessing and feature extraction chain is used for the source domain fault samples, the target domain support samples, and the target domain samples to be diagnosed. First, equal-length acoustic and vibration segments are obtained by slicing according to a uniform window length and overlap rate. Then, they are mapped into equal-dimensional vectors by a uniform feature extraction model, which are denoted as source domain segment representation, support segment representation, and sample to be diagnosed segment representation, respectively.

[0038] It should also be noted that the feature extraction model can use a feature extraction network based on convolution or time-frequency joint structure. Its input is an equal-length sound and vibration segment, and its output is a vector of fixed dimensions. The model completes meta-learning training on source domain fault samples, enabling it to quickly adapt to a small number of samples in new scenarios.

[0039] Extract the fault categories from the source domain fault samples, and perform intra-class consistency aggregation of the source domain fragment representations according to the fault categories to obtain the source domain class prototype;

[0040] It should be noted that for all source domain fragment representations under each fault category, the intra-class center vector in the representation space is calculated, and the intra-class center vector is used as the source domain class prototype of the fault category. When the source domain fragment representation under a certain fault category is significantly discrete within the class, the source domain fragment representations that deviate from the intra-class center vector by more than a preset multiple of the intra-class standard deviation are removed, and the intra-class center vector is calculated again.

[0041] The supporting fragment representation is compared with the prototype of the same source domain class, and the deviation part relative to the prototype of the source domain class is extracted as the target domain residual representation. It is then cross-compared according to the fault category. The common deviation part that appears repeatedly in multiple fault categories and whose difference between categories is lower than the preset residual retention condition is selected as the cross-class co-occurrence domain residual.

[0042] It should be noted that for each supporting fragment representation, the source domain class prototype corresponding to its fault category is located, and the difference between the supporting fragment representation and the corresponding source domain class prototype is calculated to obtain the deviation of the supporting fragment representation relative to the source domain class prototype. This deviation is used as the target domain residual representation of the supporting fragment representation. The target domain residual representation has direction and amplitude in the representation space. Its direction reflects the offset direction of the target scene relative to the source domain, and its amplitude reflects the offset intensity.

[0043] The cross-class co-occurrence domain residues are removed from the supporting fragment representations, and the parts whose distance relationship with the prototype of the same source domain class satisfies the preset same-class preservation condition are retained. Based on this, the target domain class reference is generated.

[0044] It should be noted that the generation process of the target domain class reference is as follows:

[0045] Sub-step 1.1.1: For each support fragment representation, subtract the projection of each common deviation direction component in the cross-class co-occurrence domain residue onto the support fragment representation from its vector in terms of direction, to obtain the support fragment representation after deducting the co-occurrence offset;

[0046] Sub-step 1.1.2: For the support fragment representation after deducting co-occurrence offset, recalculate its distance to the prototype of the same source domain class and compare it with the distance before deduction;

[0047] Sub-step 1.1.3: When the distance from the support fragment representation after deducting the co-occurrence offset to the prototype of the same source domain class is not greater than the product of the distance before deduction and the preset class preservation coefficient, it is determined that the class preservation condition is met, and the support fragment representation is retained in the generation set of the target domain class reference; otherwise, the support fragment representation is discarded.

[0048] Sub-step 1.1.4: For the support fragment representations of the same fault category within the generated set after deducting co-occurrence offsets, calculate their intra-class center vectors and use these intra-class center vectors as the target domain class references for that fault category;

[0049] Since the class retention condition is used to maintain the class affiliation relationship between the supporting fragment representation and the source domain class prototype after deducting the co-occurrence offset, a class retention coefficient is set, with a value ranging from 1.05 to 1.30, and a typical value of 1.15.

[0050] For each supporting fragment representation, calculate its distance to the class reference of the same type of target domain and the nearest class reference of the different type of target domain, and form a boundary reference set;

[0051] Based on the common deviation direction of the cross-class co-occurrence domains remaining in the support fragment representation, a candidate domain alignment processing amount is generated. This amount is then subtracted from the support fragment representation to obtain the candidate support fragment representation. Finally, the distance to the class reference of the same class target domain is subtracted from the distance to the candidate support fragment representation to obtain the candidate class boundary interval.

[0052] It should be noted that the process of generating the candidate domain alignment processing amount and calculating the candidate class boundary interval is as follows:

[0053] Sub-step 1.2.1: Read each common deviation direction component from the cross-class co-occurrence domain residue, configure an initial deduction intensity for each common deviation direction component, and use the sum of the products of each common deviation direction component and the initial deduction intensity as the candidate domain alignment processing amount;

[0054] Sub-step 1.2.2: For each support fragment representation, subtract the projection of the candidate domain alignment processing amount onto the support fragment representation from its vector in the direction to obtain the candidate support fragment representation;

[0055] Sub-step 1.2.3: For each candidate support fragment representation, calculate its distance to the class reference of the same type of target domain and its distance to the nearest class reference of the opposite type of target domain. Subtract the distance of the class reference of the same type of target domain from the distance of the nearest class reference of the opposite type of target domain to obtain the candidate category boundary interval corresponding to the candidate support fragment representation.

[0056] If the candidate category boundary interval satisfies the preservation condition corresponding to the boundary reference set, then the common deviation part in the support fragment representation is deducted according to the candidate domain alignment processing amount and the target domain class reference is regenerated; otherwise, the target domain class reference remains unchanged.

[0057] It should be noted that when subtracting the common deviation in the support fragment representation based on the candidate domain alignment processing amount, all support fragment representations are uniformly projected and subtracted based on the candidate domain alignment processing amount. The subtracted support fragment representations replace the original support fragment representations in the regeneration of the target domain class reference. When the candidate category boundary interval does not meet the preservation condition corresponding to the boundary reference set, the target domain class reference is not updated in this round of processing, and the original target domain class reference continues to be used as the reference for subsequent diagnosis.

[0058] The common deviation in the representation of the segment to be diagnosed is subtracted by the candidate domain alignment processing amount, and the fault diagnosis result is obtained based on the category boundary interval between the subtracted segment representation and the target domain class reference.

[0059] It should be noted that the process of obtaining the fault diagnosis results is as follows:

[0060] Sub-step 1.3.1: Subtract the projection amount of the final candidate domain alignment processing on the segment representation to be diagnosed from the segment representation to obtain the subtracted segment representation.

[0061] Sub-step 1.3.2: For the deducted segment to be diagnosed, calculate the distance to the target domain class reference corresponding to each fault category, and use the fault category corresponding to the closest distance as the preliminary diagnosis category;

[0062] Sub-step 1.3.3: Use the difference between the distance of the target domain class reference corresponding to the preliminary diagnosis category and the distance of the target domain class reference corresponding to the next nearest fault category as the category boundary interval, and output the fault diagnosis result together with the preliminary diagnosis category and the category boundary interval.

[0063] It should be noted that the fault diagnosis results include a fault category field, a category boundary interval field, a recent outlier fault category field, and a processing batch timestamp field. When the category boundary interval is positive and not lower than the preset confidence interval lower limit, the fault diagnosis results are given in a fixed output format, such as "fault category is outer ring pitting, category boundary interval is 0.42, recent outlier fault category is inner ring pitting, processing batch timestamp is April 27, 2026, 10:15". When the category boundary interval is positive but lower than the preset confidence interval lower limit, the fault diagnosis results are given in a candidate output format, including the primary fault category and the next closest candidate fault category. When the distance from the segment to be diagnosed to all target domain class references exceeds the preset anomaly upper limit, the fault diagnosis results are given in an unknown type format, including only the anomaly identifier and the processing batch timestamp. The confidence interval lower limit ranges from 0.05 to 0.30, with a typical value of 0.15.

[0064] It should be noted that the source domain fragment representation, support fragment representation, fragment representation to be diagnosed, source domain class prototype, target domain residual representation, cross-class co-occurrence domain residual, target domain class reference, candidate domain alignment processing amount, candidate support fragment representation, and the same class distance, nearest dissimilar class distance, original class boundary interval and candidate class boundary interval in the boundary reference set are all vectors in the same representation space or distances obtained based on the vectors. The distances are uniformly measured using Euclidean distance, with consistent units, and can be directly calculated using difference and ratio.

[0065] For non-ideal situations such as missing data, momentary interruptions, and jumps that may occur during the acquisition of acoustic and vibration signals, the following rules apply: When the proportion of effective sampling points of a certain acoustic and vibration signal within a slice window is not lower than the preset tolerance threshold, the missing part is filled in using linear interpolation of the effective sampling points before and after, and then the signal enters the feature extraction process; when the proportion of effective sampling points is lower than the preset tolerance threshold, the slice window is discarded and the signal is carried over to the next slice window; when multiple consecutive slice windows are discarded, the feature extraction of the current batch is interrupted and marked as an input anomaly, and the extraction process is restarted after a new batch of acoustic and vibration signals arrives. The tolerance threshold ranges from 0.80 to 0.95, with a typical value of 0.9.

[0066] Through the above technical solution, this embodiment identifies common background offsets by using "cross-class co-occurrence domain residues", constrains the execution of candidate domain alignment processing by using "preservation conditions corresponding to boundary reference sets", and obtains fault diagnosis results by subtracting the common deviation part represented by the segment to be diagnosed from the candidate domain alignment processing. This enables the model to quickly complete cross-domain adaptation in scenarios with a small number of target samples under new equipment and new working conditions. Moreover, the model will not weaken the separability between fault categories due to the absorption of common offsets during the adaptation process, providing a dual guarantee of cross-domain adaptation and category separability at the main chain level for the overall solution.

[0067] In an optional embodiment, the faults are cross-referenced by fault category, and common deviations that occur repeatedly in multiple fault categories and whose differences between categories are below a preset residue retention condition are selected as cross-category co-occurrence domain residues, specifically including:

[0068] The residual representation of each target domain is split into several residual direction components according to the deviation direction, and then grouped according to the fault category to form a set of residual direction components for each fault category;

[0069] It should be noted that for each residual representation of the target domain, its vector is decomposed into several residual direction components according to several preset base directions in the representation space. Each residual direction component retains the projection magnitude and direction identifier in its corresponding base direction. According to the fault category to which the residual representation of the target domain belongs (i.e. the fault category of the corresponding supporting fragment representation), the residual direction components obtained from its decomposition are collected into the set of residual direction components of that fault category.

[0070] Residual directional components with similar directions within the same fault category are merged to obtain representative residual directional components under each fault category.

[0071] It should be noted that for each fault category's set of residual directional components, the directional angle between any two residual directional components within the set is calculated. Residual directional components whose directional angle does not exceed the upper limit of the preset merging angle are assigned to the same merging group. The residual directional components in each merging group are weighted by amplitude to obtain the composite direction of the merging group. This composite direction is used as a representative residual directional component under the fault category. The upper limit of the merging angle ranges from 5 degrees to 15 degrees, with a typical value of 10 degrees.

[0072] The representative residual directional components under different fault categories are matched across categories according to directional similarity to obtain the number of occurrence categories and the degree of directional consistency for each representative residual directional component.

[0073] It should be noted that the calculation process for cross-category matching and consistency is as follows:

[0074] Sub-step 2.1.1: For any representative residual direction component, traverse the representative residual direction components under other fault categories, and include the representative residual direction components whose angle with its direction does not exceed the upper limit of the preset cross-class matching angle into the same cross-class matching family;

[0075] Sub-step 2.1.2: For each cross-class matching family, count the number of different fault categories covered by the component representing the residual direction, and use this as the number of occurrence categories corresponding to that cross-class matching family;

[0076] Sub-step 2.1.3: For each cross-class matching family, calculate the average value of the cosine of the direction angle between each pair of all representative residual direction components in the family, and use it as the direction consistency degree corresponding to the cross-class matching family.

[0077] For example, the formula for calculating the degree of directional consistency is:

[0078] ;

[0079] in, This represents the degree of directional consistency corresponding to the cross-class matching family. This represents the number of residual directional components in this cross-class matching family. Let S be the angle between the i-th representative residual direction component and the j-th representative residual direction component in the cross-class matching family; for example, the upper bound of the cross-class matching angle ranges from 8 degrees to 20 degrees, with a typical value of 12 degrees; the value range of S is [-1, 1].

[0080] The representative residual directional components that have a number of categories not less than the preset lower limit of the number of categories and a degree of directional consistency not less than the preset lower limit of directional consistency are identified as common deviation directional components, and all common deviation directional components are summarized to form cross-class co-occurrence domain residues;

[0081] It should be noted that for each cross-class matching family, the number of its corresponding occurrence categories is compared with the preset lower limit of the number of categories, and the degree of directional consistency is compared with the preset lower limit of directional consistency. Only when both conditions are met, the representative residual directional components in the cross-class matching family are weighted by amplitude and combined into a common deviation directional component, and added to the cross-class co-occurrence domain residue. The entire cross-class matching family is traversed to obtain the complete cross-class co-occurrence domain residue.

[0082] Since the lower limit of the number of categories is used to limit the minimum coverage width of co-occurrence, a lower limit of the number of categories is set, and its value ranges from 50% to 70% of the total number of fault categories, with a typical value of 60% of the total number of fault categories; since the lower limit of directional consistency is used to limit the directional stability of co-occurrence, a lower limit of directional consistency is set, and its value ranges from 0.85 to 0.97, with a typical value of 0.92.

[0083] Additionally, the upper bound of the merging angle, the upper bound of the cross-class matching angle, and the lower bound of the direction consistency are all parameters for measuring directional similarity. The units of the upper bound of the merging angle and the upper bound of the cross-class matching angle are both degrees, and they can be directly subtracted and compared. The lower bound of the direction consistency is given in the form of a cosine value, which is a different expression from the upper bound of the cross-class matching angle, but they can be converted one-to-one through the cosine function. In this embodiment, the cosine form is used uniformly in the calculation of the determination of the degree of direction consistency.

[0084] Through the above technical solution, this embodiment determines the common deviation direction component by splitting the target domain residual representation into residual direction components, merging them according to fault categories, cross-category matching, and using a dual filtering method of the number of occurrence categories and the degree of direction consistency. This makes the extraction of cross-class co-occurrence domain residuals from "conceptual common deviation" to an executable direction-level co-occurrence screening, providing a refined and reproducible cross-class co-occurrence domain residual generation path.

[0085] In an optional embodiment, for each supporting fragment representation, the distances to the nearest class reference of the same type of target domain and the nearest class reference of the opposite type of target domain are calculated to form a boundary reference set, specifically including:

[0086] For each supporting fragment representation, calculate its class distance to the class reference of the same target domain and its nearest dissimilar distance to the class reference of the nearest dissimilar target domain;

[0087] It should be noted that for each supporting fragment representation, the target domain class reference corresponding to its fault category is located, and the Euclidean distance between the two is calculated as the class distance of the supporting fragment representation; at the same time, all target domain class references other than the fault category to which the supporting fragment representation belongs are traversed, the Euclidean distance between the two is calculated, and the minimum value is taken as the nearest dissimilar distance of the supporting fragment representation.

[0088] Subtracting the nearest out-of-class distance from the same-class distance yields the original class boundary interval corresponding to each support fragment representation, and the boundary reference set is composed of the same-class distance, the nearest out-of-class distance, and the original class boundary interval;

[0089] It should be noted that for each supporting fragment representation, the distance to the nearest out-of-class class is subtracted from the distance to the same class class to obtain the original class boundary interval corresponding to the supporting fragment representation; the distance to the same class class class, the distance to the nearest out-of-class class, and the original class boundary interval corresponding to the supporting fragment representation are used as a boundary reference set record; the boundary reference set records corresponding to all supporting fragment representations are summarized to form a boundary reference set.

[0090] Support fragments with an original class boundary interval greater than zero are marked as valid support fragments, and the corresponding boundary preservation lower limit is obtained based on the original class boundary interval of the valid support fragments and the preset preservation ratio.

[0091] It should be noted that for each supporting fragment representation, it is determined whether its original class boundary interval is greater than zero. If it is greater than zero, it is marked as a valid supporting fragment; otherwise, it does not participate in the subsequent preservation condition verification. For each valid supporting fragment, its original class boundary interval is multiplied by a preset preservation ratio to obtain the lower limit of boundary preservation corresponding to the valid supporting fragment. Since the preservation ratio is used to prevent significant class boundary shrinkage while tolerating small separability loss, a preservation ratio is set, and its value ranges from 0.80 to 0.95, with a typical value of 0.85.

[0092] The candidate category boundary interval obtained after the effective supporting fragments are processed by deducting the amount of processing according to the candidate domain alignment is compared with the corresponding boundary lower limit.

[0093] When the candidate category boundary interval of each valid support fragment is not lower than the corresponding boundary preservation lower limit, and there is no relationship reversal where the nearest reference relationship of any valid support fragment changes from the reference of the same type of target domain to the reference of the nearest different type of target domain after trial deduction, it is determined that the preservation condition corresponding to the boundary reference set is satisfied.

[0094] It should be noted that the process for determining the hold condition is as follows:

[0095] Sub-step 3.1.1: For each valid support fragment, after trial deduction based on the candidate domain alignment processing amount, obtain the candidate category boundary interval corresponding to the valid support fragment;

[0096] Sub-step 3.1.2: Compare the candidate category boundary interval with the corresponding boundary preservation lower limit, and record whether it is not lower than the boundary preservation lower limit;

[0097] Sub-step 3.1.3: For each valid support segment, compare whether it is closer to the class reference of the same type of target domain or closer to the class reference of the nearest dissimilar target domain before the trial deduction, and the relative relationship between the two after the trial deduction. If it is closer to the class reference of the same type of target domain before the trial deduction and becomes closer to the class reference of the nearest dissimilar target domain after the trial deduction, then record it as a relationship reversal.

[0098] Sub-step 3.1.4: When the candidate category boundary intervals corresponding to all valid support fragments are not lower than their corresponding boundary preservation lower limit, and the nearest reference relationship reversal has not occurred before and after the trial deduction of all valid support fragments, it is determined that the candidate domain alignment processing amount meets the preservation condition corresponding to the boundary reference set; otherwise, it is determined that it does not meet the condition.

[0099] It should be noted that the same-class distance, nearest dissimilar distance, original class boundary interval, candidate class boundary interval, and boundary preservation lower limit are all distance quantities in the target domain representation space, with consistent units, and can be directly subtracted and compared; the preservation ratio is a dimensionless ratio quantity, and the boundary preservation lower limit obtained by multiplying it with the original class boundary interval has the same unit as the original class boundary interval.

[0100] It should be noted that the boundary reference set and boundary preservation lower limit determined in this specification are the basis for determining whether the "candidate category boundary interval meets the preservation condition". The multi-candidate trial deduction screening calls this judgment logic one by one for the alignment processing of multiple candidate domains based on this. Under this linkage, the selection process of candidate domain alignment processing has a unified and reproducible judgment caliber, avoiding inconsistencies in the judgment criteria between different candidate quantities.

[0101] Through the above technical solution, this embodiment defines the original category boundary interval by the difference between the distance of the same category and the distance of the nearest dissimilar category, calculates the boundary maintenance lower limit by maintaining the ratio, and then compares the candidate category boundary interval with the boundary maintenance lower limit for each effective support segment, plus the nearest reference relationship reversal detection, so that the "maintaining condition" is realized from a conceptual constraint into an executable dual-gate judgment, providing a mechanism for re-checking the separability of candidate domain alignment processing.

[0102] In an optional embodiment, before comparing the supporting fragment representation with the prototype of the same source domain class and extracting the deviation relative to the source domain class prototype as the residual representation of the target domain, a supporting fragment reliability screening process is further included, specifically including:

[0103] Based on the distance from each source domain segment representation to the prototype of the same source domain class, the intra-class distance distribution of the source domain segment representation under each fault category is statistically analyzed, and the upper bound of the deviation distribution of the corresponding fault category is set according to the intra-class distance distribution.

[0104] It should be noted that for each fault category, the distances of all source domain fragments represented under that fault category to the prototype of the same source domain are calculated to obtain the intra-class distance distribution of that fault category. Based on the mean and standard deviation of the intra-class distance distribution, the upper bound of the deviation distribution of that fault category is set according to the method of "mean plus a preset multiple of the standard deviation". The mean and standard deviation of the intra-class distance distribution are both in Euclidean distance, and the upper bound of the deviation distribution is in the same unit. The preset multiple ranges from 2 to 4, with a typical value of 3.

[0105] For each supporting fragment representation, calculate its distance to the prototype of the same source domain class and compare it with the upper bound of the deviation distribution of the corresponding fault category;

[0106] It should be noted that for each support fragment representation, the corresponding source domain class prototype and the upper bound of the deviation distribution are located according to the fault category to which it belongs. The Euclidean distance from the support fragment representation to the source domain class prototype is calculated. Then, this distance is compared with the upper bound of the deviation distribution of the corresponding fault category to determine whether the support fragment representation exceeds the distribution range within the source domain class.

[0107] Support segments whose distance does not exceed the upper bound of the deviation distribution are marked as reliable support segments, and support segments whose distance exceeds the upper bound of the deviation distribution are marked as support segments to be verified.

[0108] Reliable support fragments are used as effective fragments in the support fragment representation to extract residual representations of the target domain, and residual representations of the target domain are extracted based on the effective fragments.

[0109] The support fragments to be verified are excluded from the extraction range of the residual representation of the target domain, and after generating the candidate domain alignment processing amount, the support fragments to be verified and the reliable support fragments are used together to calculate the candidate class boundary interval.

[0110] It should be noted that for each supporting fragment representation, branching is performed based on its comparison with the upper bound of the deviation distribution of the corresponding fault category: when its distance to the source domain class prototype does not exceed the upper bound of the deviation distribution, it is marked as a reliable supporting fragment and included in the set of valid fragments for extracting residual representations of the target domain; when its distance to the source domain class prototype exceeds the upper bound of the deviation distribution, it is marked as a supporting fragment to be reviewed and is not included in the set of valid fragments for the time being, but is retained separately; after the candidate domain alignment processing quantity is generated, the supporting fragment to be reviewed is re-entered into the calculation stage of the candidate category boundary interval together with the reliable supporting fragments to participate in the construction of the boundary reference set and the verification of the preservation conditions.

[0111] It should also be noted that the construction criteria for the intra-class distance distribution of source domain fragment representations are as follows: the number of source domain fragment representations used to calculate the intra-class distance distribution under each fault category is no less than several hundred. Before statistics are performed, obviously abnormal samples whose distance from the prototype of the same source domain exceeds the upper limit of the intra-class distribution are removed. The mean and standard deviation of the intra-class distance distribution are calculated based on the samples after removal. When the source domain fault sample library partition is updated, the intra-class distance distribution and the corresponding upper bound of the deviation distribution are re-statistically calculated synchronously, and the statistical results of the old partition are not used.

[0112] Through the above technical solution, this embodiment sets an upper bound for the deviation distribution based on the intra-class distance distribution of the source domain, performs reliability grading and marking of the support fragment representation, and uses reliable support fragments and support fragments to be verified differently in the two stages of target domain residual representation extraction and boundary reference set construction. This significantly suppresses the contamination of cross-class co-occurrence domain residual extraction by occasional noise, acquisition anomalies or instantaneous shocks, provides purified input, and achieves dual utilization of a small number of target samples by retaining support fragments to be verified to participate in boundary verification.

[0113] In an optional embodiment, after extracting the residual characterization of the target domain using reliable support fragments, and before cross-referencing the residual characterization of the target domain by fault category, a residual direction stability verification process is further included, specifically including:

[0114] The target domain residual representations corresponding to the reliable support segments are grouped according to their respective fault categories to obtain the intra-class residual set for each fault category;

[0115] It should be noted that the residual representations of the target domain corresponding to the reliable support segments output by the support segment reliability screening process are aggregated according to their respective fault categories (i.e., the fault categories of the corresponding reliable support segments) to obtain the intra-class residual sets of each fault category.

[0116] Perform directional consistency statistics on the target domain residual representations in each class residual set to obtain the degree of residual directional concentration for the corresponding fault category;

[0117] It should be noted that the statistical process for determining the concentration of residual directions is as follows:

[0118] Sub-step 5.1.1: For each target domain residual representation in the intra-class residual set, normalize its vector to a unit direction vector;

[0119] Sub-step 5.1.2: For all unit direction vectors in the residual set within the class, calculate the cosine of the direction angle between each pair of vectors;

[0120] Sub-step 5.1.3: Take the average of the cosine values ​​of the included angles in all directions as the degree of concentration of residual directions for this fault category.

[0121] For example, the formula for calculating the degree of concentration of residual directions is:

[0122] ;

[0123] in, To determine the degree of concentration of residual directions corresponding to the fault category, This represents the number of target domain residual representations in the residual set within the fault category. is the directional angle between the p-th and q-th target domain residual representations in the residual set within this class; the output of this formula is used as a comparison measure of the lower limit of concentration in the subsequent division of single residual sets and residual subsets, and the directional consistency with the cross-class matching family adopts the same form of cosine average measure, both of which are cosine values ​​with consistent dimensions; the value range of C is [-1,1].

[0124] The set of residues within a class whose concentration of residue direction is not lower than the preset lower limit of concentration is taken as a single set of residues and entered into the subsequent cross-matching.

[0125] The intra-class residual sets whose residual direction concentration is lower than the preset concentration lower limit are clustered and split according to the similarity of residual directions to obtain multiple residual subsets;

[0126] Use the single residual set and residual subset as inputs for cross-matching by fault category, respectively;

[0127] It should be noted that for the residual set within each fault category, branching is performed based on the comparison result of the concentration degree of the residual direction with the preset lower limit of concentration degree: when the concentration degree of the residual direction is not lower than the lower limit of concentration degree, the residual set within the category is retained as a single residual set; when the concentration degree of the residual direction is lower than the lower limit of concentration degree, the target domain residual representations in the residual set within the category are clustered according to the similarity of directions, and the angle between the directions of any two target domain residual representations within each cluster does not exceed the upper limit of the preset clustering angle. After clustering, each cluster is treated as a residual subset; finally, the single residual set and residual subset obtained for each fault category are used as the input for the cross-comparison step by fault category.

[0128] Since the lower limit of concentration is used to distinguish whether the residual orientation within a category is a stable unimodal or multimodal distribution, a lower limit of concentration is set, with a value ranging from 0.75 to 0.92, a typical value of 0.85, and is dimensionless. Since the upper limit of cluster angle is used to control the directional compactness of each residual subset in the multimodal case, an upper limit of cluster angle is set, with a value ranging from 10 degrees to 25 degrees, a typical value of 15 degrees.

[0129] It should also be noted that the concentration of residual directions is given in the form of cosine average, and the lower limit of concentration is set in the form of cosine value. The two have the same dimension and can be directly compared. The upper limit of the clustering angle is in degrees, which has the same dimension as the upper limit of the merging angle and the upper limit of the cross-class matching angle. The three constitute different levels of thresholds under the same directional similarity measurement system.

[0130] Through the above technical solution, this embodiment further purifies the input from the dimension of directional stability in addition to sample reliability screening by statistically analyzing the directional concentration of the intra-class residual set, branching according to the lower limit of concentration, and clustering and splitting the multimodal intra-class residual set according to directional similarity. This forms a progressive preprocessing link of "sample level - directional level" with the support fragment reliability screening process, providing a more stable directional input for cross-class co-occurrence domain residual extraction.

[0131] In an optional embodiment, after generating candidate domain alignment processing quantities based on the common deviation direction of cross-class co-occurrence domains remaining in the supporting fragment representation, the process further includes a multi-candidate trial deduction screening process for the candidate domain alignment processing quantities, specifically including:

[0132] Based on the common deviation direction component in the cross-class co-occurrence domain residue and the preset deduction intensity level, multiple candidate domain alignment processing quantities are combined to form a candidate processing quantity set.

[0133] It should be noted that the process of generating the candidate processing volume set is as follows:

[0134] Sub-step 6.1.1: Read all common deviation direction components from the cross-class co-occurrence domain residues and preset several deduction intensity levels;

[0135] Sub-step 6.1.2: For each combination of common deviation direction component and each deduction intensity level, multiply the common deviation direction component by the deduction intensity level to obtain the deduction item corresponding to the combination;

[0136] Sub-step 6.1.3: Sum the deduction items of all common deviation direction components by direction to obtain a candidate domain alignment processing amount; by traversing the combination of deduction intensity levels of different common deviation direction components, obtain multiple candidate domain alignment processing amounts and form a candidate processing amount set.

[0137] Since the subtraction intensity level is used to balance the degree of removal of common background offset and the degree of preservation of class separability among multiple levels, the subtraction intensity level is set, and its value range is several discrete values ​​between 0.3 and 1.0, with typical levels being 0.4, 0.6, 0.8, and 1.0; the typical number of subtraction intensity levels is 3 to 5.

[0138] For each candidate domain alignment processing quantity in the candidate processing quantity set, the corresponding common deviation portion is subtracted from the support fragment representation to obtain the corresponding candidate support fragment representation group.

[0139] For each candidate support fragment representation group, calculate the candidate category boundary interval corresponding to each candidate support fragment representation to obtain the candidate category boundary interval distribution;

[0140] It should be noted that for each candidate domain alignment processing quantity in the candidate processing quantity set, the projection of the candidate domain alignment processing quantity onto the support fragment representation is subtracted from each support fragment representation in the direction to obtain the candidate support fragment representation corresponding to the candidate domain alignment processing quantity. The candidate support fragment representations corresponding to all support fragment representations constitute a candidate support fragment representation group. For each candidate support fragment representation in the candidate support fragment representation group, the difference between its distance to the nearest heterogeneous target domain class reference and its distance to the same type of target domain class reference is calculated according to the definition of "candidate category boundary interval". The candidate category boundary intervals corresponding to all candidate support fragment representations are summarized to obtain the candidate category boundary interval distribution corresponding to the candidate domain alignment processing quantity.

[0141] The candidate domain alignment processing amount that satisfies the boundary reference set correspondence preservation condition of the candidate category boundary spacing distribution is included in the qualified candidate set;

[0142] The candidate domain alignment processing amount with the highest degree of common deviation reduction and the best degree of preservation of the candidate category boundary interval distribution is selected from the qualified candidate set and used as the final candidate domain alignment processing amount.

[0143] It should be noted that the final selection process for candidate field alignment processing volume is as follows:

[0144] Sub-step 6.2.1: For each candidate domain alignment processing quantity in the candidate processing quantity set, according to the determination logic of the preservation condition as described in claim 3, verify the distribution of the corresponding candidate category boundary interval, and include the candidate domain alignment processing quantities that pass the verification into the qualified candidate set;

[0145] Sub-step 6.2.2: For each candidate domain alignment processing amount in the qualified candidate set, calculate its common deviation reduction degree and candidate category boundary interval distribution preservation degree. The former is represented by the composite amplitude of the candidate domain alignment processing amount in the common deviation direction, and the latter is represented by the ratio of the mean of the candidate category boundary interval distribution to the mean of the original category boundary interval corresponding to the effective support fragment.

[0146] Sub-step 6.2.3: Sort the candidate domain alignment processing quantities in the qualified candidate set from high to low according to the degree of common deviation reduction. Further compare the candidate domain alignment processing quantities at the top of the deduction degree ranking to see the degree of preservation of the candidate category boundary interval distribution. Select the candidate domain alignment processing quantity with the highest deduction degree and a preservation degree no lower than the other top candidate domain alignment processing quantities as the final candidate domain alignment processing quantity.

[0147] For example, the formula for calculating the degree of preservation of the candidate category boundary spacing distribution is as follows:

[0148] ;

[0149] in, The degree to which the candidate category boundary spacing distribution is preserved corresponding to the amount of alignment processing for this candidate domain. The candidate category boundary interval is the total number of valid supporting fragments corresponding to the alignment processing amount for this candidate domain. This is the original category boundary interval corresponding to the valid supporting fragment.

[0150] When the qualified candidate set is empty, select the candidate domain alignment processing amount with the lowest degree of decrease in the candidate category boundary interval, and process it according to the path that keeps the target domain class reference unchanged;

[0151] It should be noted that when no candidate domain alignment processing quantity in the candidate processing quantity set passes the preservation condition verification, the difference between the mean of the candidate category boundary interval distribution and the mean of the original category boundary interval of the corresponding valid support fragment is calculated for each candidate domain alignment processing quantity in the candidate processing quantity set. This difference is taken as the degree of decrease in the candidate category boundary interval corresponding to that candidate domain alignment processing quantity. The candidate domain alignment processing quantity with the lowest degree of decrease in the candidate category boundary interval is selected as the candidate domain alignment processing quantity record value for this batch. It is processed according to the path of "keeping the target domain class reference unchanged if the candidate category boundary interval does not meet the preservation condition". The target domain class reference is not updated in this batch, but the candidate domain alignment processing quantity is recorded in the cross-batch record for subsequent verification.

[0152] Through the above technical solution, this embodiment expands the single candidate domain alignment processing quantity into a candidate processing quantity set composed of common deviation direction components and deduction intensity levels. Combined with conditional filtering, dual index ranking of deduction degree and preservation degree, and candidate quantity downgrading processing when the qualified candidate set is empty, the generation of candidate domain alignment processing quantity is no longer a single point decision but a multi-point selection, providing dual protection of refined candidate quantity screening and boundary condition backoff.

[0153] In an optional embodiment, after obtaining the fault diagnosis result based on the category boundary interval between the deducted segment representation and the target domain class reference, the process further includes a cross-batch common deviation from stable records and a rejection update process, specifically including:

[0154] The candidate domain alignment processing amount and its corresponding common deviation direction component used in the diagnostic batch of each target domain are recorded in the cross-batch common deviation record in chronological order.

[0155] It should be noted that after each target domain diagnosis batch is completed, the candidate domain alignment processing amount and its common deviation direction components finally adopted by the selection process as described in claim 6 in that batch, together with the verification results of the timestamp and boundary reference set corresponding to the preservation conditions of that batch, are added to the cross-batch common deviation record in chronological order. The cross-batch common deviation record is stored using a circular buffer structure. The capacity of the circular buffer is set according to the observable batch duration, and its typical value is the most recent 20 to 50 batches. After the old batch records are filled, they are overwritten by the new batch records in chronological order.

[0156] For each common deviation direction component in the cross-batch common deviation record, count the number of times it appears repeatedly in the consecutive diagnostic batches, and verify whether the corresponding deduction process satisfies the preservation condition corresponding to the boundary reference set.

[0157] It should be noted that the statistical process for the number of repetitions and the maintenance condition verification is as follows:

[0158] Sub-step 7.1.1: For each common deviation direction component in the cross-batch common deviation record, determine the similarity of cross-class matching direction according to claim 2, and regard the common deviation direction components in different batches whose direction angle does not exceed the upper limit of the cross-class matching angle as the same stable candidate;

[0159] Sub-step 7.1.2: For each stable candidate, count the number of times it appears repeatedly in the continuous diagnosis batch. The continuous diagnosis batch is determined by the adjacency of the batch timestamps. If the intermediate batch lacks a component in the same direction, the continuous counting is interrupted.

[0160] Sub-step 7.1.3: For each stable candidate in each batch, during the deduction process, read the corresponding batch's boundary reference set and the corresponding maintenance condition verification results, and mark the stable candidate that passes each verification as a maintenance condition continuously satisfied item.

[0161] Common deviation direction components whose recurrence count is not less than the preset lower limit of consecutive batches and whose corresponding deduction process meets the retention conditions are registered as target domain stable common offsets.

[0162] It should be noted that for each stable candidate, its recurrence count is compared with the preset lower limit of consecutive batches, and it is verified whether the deduction process in the consecutive batches meets the retention conditions. Only when the recurrence count is not lower than the lower limit of consecutive batches and all corresponding deduction processes meet the retention conditions, the direction of the stable candidate (weighted synthesis according to the directions of each batch) is registered as a target domain stable common offset. The target domain stable common offset is stored in a separate stable offset record table, independent of the cross-batch common offset record.

[0163] Since the continuous batch lower limit is used to strike a balance between fast response and stability, a continuous batch lower limit is set, with a value ranging from 3 to 10 batches, and a typical value of 5 batches.

[0164] In subsequent diagnostic batches, the representation of the segment to be diagnosed is first pre-subtracted according to the stable common offset of the target domain, and then the generation and screening of candidate domain alignment processing volume is carried out.

[0165] It should be noted that the pre-deduction process for the stable common offset of the target domain is as follows:

[0166] Sub-step 7.2.1: After the new batch of segments to be diagnosed arrives, read all currently registered target domain stable common offsets from the stable offset record table;

[0167] Sub-step 7.2.2: Subtract the sum of the projections of all target domain stable common offsets onto the characterization of the segment to be diagnosed by subtracting them in the direction to obtain the pre-subtracted characterization of the segment to be diagnosed.

[0168] Sub-step 7.2.3: Using the pre-subtracted characterization of the segment to be diagnosed as input, process it according to the candidate domain alignment processing quantity generation and screening process to obtain the final candidate domain alignment processing quantity used in this batch.

[0169] If the target domain stable common offset causes the class boundary interval to fall below the lower limit allowed by the corresponding preservation condition in a subsequent diagnostic batch, the update of the target domain stable common offset for that diagnostic batch is rejected, and the corresponding common offset direction component is removed from the target domain stable common offset.

[0170] It should be noted that the process for rejecting and reversing updates of the target domain stable common offset is as follows:

[0171] Sub-step 7.3.1: For each new batch of processing results after pre-subtraction of the target domain stable common offset and then screening by the candidate domain alignment processing amount, call the preservation condition judgment logic to verify whether the candidate category boundary interval corresponding to the effective support fragment of the current batch is still not lower than the corresponding boundary preservation lower limit.

[0172] Sub-step 7.3.2: When the verification result shows that the candidate category boundary interval of the valid supporting fragment is lower than the corresponding boundary maintenance lower limit, locate the target domain stable common offset component that caused the decline. The method is to sequentially back down the pre-deduction of each target domain stable common offset component and re-screen the candidate domain alignment processing amount. The target domain stable common offset component that meets the boundary maintenance lower limit after backing down is identified as the component that caused the decline.

[0173] Sub-step 7.3.3: Remove the degrading component from the stable offset record table, and do not register the newly appearing common deviation direction component as a new target domain stable common offset in this batch. The characterization of the segment to be diagnosed in this batch is changed to be processed together with the target domain stable common offset after removing the degrading component and the candidate domain alignment processing amount finally adopted in this batch.

[0174] It should be noted that if the applicable conditions are no longer met, such as changes in the model of the data acquisition equipment, significant changes in the installation location, or changes in the load conditions, the entire stable offset record table should be cleared and the statistics of cross-batch common deviation stability records should be restarted. The reason is that the validity of the target domain stable common offset is based on the premise that the common background offset of the cross-batch scenario is continuously stable. The change of operating conditions will make the originally stable common deviation direction component no longer representative of the cross-batch, and continuing to use it will cause the pre-deduction direction to deviate from the actual scenario.

[0175] It should also be noted that the common deviation direction components in the target domain stable common offset, the degrading component, and the cross-batch common deviation record are all in the same representation space. Their directions are stored in the form of unit direction vectors, and their amplitudes are stored in the form of scalars with the same dimensions as the original segment representation. During pre-deduction, the amplitude is subtracted from the segment representation to be diagnosed in the form of "amplitude multiplied by direction". The dimensions are consistent with the segment representation to be diagnosed, so the subtraction operation can be performed directly.

[0176] Through the above technical solutions, this embodiment expands the single-batch deduction to a long-term working mode with cross-batch prior accumulation and boundary self-verification by cyclic storage of cross-batch common deviation records, continuous batch recurrence statistics of stable candidates, registration and pre-deduction of stable common offsets in the target domain, and rollback identification and cancellation of degrading components. This provides post-closed-loop support for the stable operation of the solution in real industrial scenarios such as long-term deployment, continuous diagnosis, and slow drift of operating conditions.

[0177] See Figure 2 As shown, this scheme proposes a cross-domain few-shot intelligent fault diagnosis system based on meta-learning, which is used to implement the above-mentioned cross-domain few-shot intelligent fault diagnosis method based on meta-learning, including:

[0178] The segment representation extraction module is used to acquire source domain fault samples, a small number of target domain support samples, and target domain samples to be diagnosed, and to extract acoustic and vibration segments in a unified manner to obtain source domain segment representations, support segment representations, and samples to be diagnosed.

[0179] The class prototype aggregation module is used to extract the fault categories of source domain fault samples, perform intra-class consistency aggregation of source domain fragment representations according to fault categories, and obtain source domain class prototypes.

[0180] The common residue extraction module is used to compare the supporting fragment representation with the prototype of the same source domain class, extract the deviation part relative to the prototype of the source domain class as the target domain residue representation, and cross-compare it according to the fault category. The common deviation part that appears repeatedly in multiple fault categories and the difference between categories is lower than the preset residue retention condition is selected as the cross-class co-occurrence domain residue.

[0181] The target reference generation module is used to subtract cross-class co-occurrence domain residues from the supporting fragment representation, retain the part whose distance relationship with the prototype of the same source domain class satisfies the preset same-class preservation condition, and generate the target domain class reference accordingly.

[0182] The boundary set construction module is used to represent each supporting fragment, calculate its distance to the class reference of the same type of target domain and the nearest class reference of the opposite type of target domain, and form a boundary reference set;

[0183] The candidate processing evaluation module is used to generate candidate domain alignment processing amount based on the common deviation direction of cross-class co-occurrence domains remaining in the support fragment representation, subtract it from the support fragment representation to obtain the candidate support fragment representation, and subtract the distance to the same class target domain reference from the distance of the candidate support fragment representation to the nearest heterogeneous target domain class reference to obtain the candidate class boundary interval.

[0184] The reference update decision module is used to deduct the common deviation part in the support fragment representation and regenerate the target domain class reference if the candidate category boundary interval meets the preservation condition corresponding to the boundary reference set; otherwise, the target domain class reference remains unchanged.

[0185] The fault diagnosis module is used to subtract the common deviation part in the representation of the segment to be diagnosed by the candidate domain alignment processing amount, and obtain the fault diagnosis result based on the category boundary interval between the subtracted segment representation and the target domain class reference.

[0186] In another embodiment, a computer device is provided, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the steps in the above embodiments.

[0187] In one embodiment, a computer-readable storage medium is provided storing a computer program that, when executed by a processor, implements the steps described above.

[0188] In one embodiment, a computer program product or computer program is provided, the computer program product or computer program including computer instructions stored in a computer-readable storage medium. A processor of a computer device reads the computer instructions from the computer-readable storage medium, and executes the computer instructions, causing the computer device to perform the steps described above.

[0189] The technical scope of this invention is not limited to the content described above. Those skilled in the art can make various modifications and variations to the above embodiments without departing from the technical concept of this invention, and all such modifications and variations should fall within the protection scope of this invention.

Claims

1. A cross-domain few-shot fault intelligent diagnosis method based on meta-learning, characterized in that, The method includes: We acquire source domain fault samples, a small number of target domain support samples, and target domain samples to be diagnosed, and extract acoustic and vibration segments in a unified manner to obtain source domain segment representations, support segment representations, and samples to be diagnosed. Extract the fault categories from the source domain fault samples, and perform intra-class consistency aggregation of the source domain fragment representations according to the fault categories to obtain the source domain class prototype; The supporting fragment representation is compared with the prototype of the same source domain class, and the deviation part relative to the prototype of the source domain class is extracted as the target domain residual representation. It is then cross-compared according to the fault category. The common deviation part that appears repeatedly in multiple fault categories and whose difference between categories is lower than the preset residual retention condition is selected as the cross-class co-occurrence domain residual. The cross-class co-occurrence domain residues are removed from the supporting fragment representations, and the parts whose distance relationship with the prototype of the same source domain class satisfies the preset same-class preservation condition are retained. Based on this, the target domain class reference is generated. For each supporting fragment representation, calculate its distance to the class reference of the same type of target domain and the nearest class reference of the different type of target domain, and form a boundary reference set; Based on the common deviation direction of the cross-class co-occurrence domains remaining in the support fragment representation, a candidate domain alignment processing amount is generated. This amount is then subtracted from the support fragment representation to obtain the candidate support fragment representation. Finally, the distance to the class reference of the same class target domain is subtracted from the distance to the candidate support fragment representation to obtain the candidate class boundary interval. If the candidate category boundary interval satisfies the preservation condition corresponding to the boundary reference set, then the common deviation part in the support fragment representation is deducted according to the candidate domain alignment processing amount and the target domain class reference is regenerated; otherwise, the target domain class reference remains unchanged. The common deviation portion in the representation of the segment to be diagnosed is subtracted by the candidate domain alignment processing amount, and the fault diagnosis result is obtained based on the category boundary interval between the subtracted representation of the segment to be diagnosed and the target domain class reference.

2. The method of claim 1, wherein, Cross-compare the faults by fault category, and select the common deviation portion that occurs repeatedly in multiple fault categories and whose inter-category differences are below the preset residue retention condition as the cross-category co-occurrence domain residue, specifically including: The residual representation of each target domain is split into several residual direction components according to the deviation direction, and then grouped according to the fault category to form a set of residual direction components for each fault category; Residual directional components with similar directions within the same fault category are merged to obtain representative residual directional components under each fault category. The representative residual directional components under different fault categories are matched across categories according to directional similarity to obtain the number of occurrence categories and the degree of directional consistency for each representative residual directional component. The representative residual directional components that have a number of categories not less than the preset lower limit of the number of categories and a degree of directional consistency not less than the preset lower limit of directional consistency are identified as common deviation directional components. All common deviation directional components are then aggregated to form cross-class co-occurrence domain residues.

3. The method of claim 1, wherein, For each supporting fragment representation, the distances to the nearest class reference of the same type of target domain and the nearest class reference of the opposite type of target domain are calculated to form a boundary reference set, which specifically includes: For each supporting fragment representation, calculate its class distance to the class reference of the same target domain and its nearest dissimilar distance to the class reference of the nearest dissimilar target domain; Subtracting the nearest out-of-class distance from the same-class distance yields the original class boundary interval corresponding to each support fragment representation, and the boundary reference set is composed of the same-class distance, the nearest out-of-class distance, and the original class boundary interval; Support fragments with an original class boundary interval greater than zero are marked as valid support fragments, and the corresponding boundary preservation lower limit is obtained based on the original class boundary interval of the valid support fragments and the preset preservation ratio. The candidate category boundary interval obtained after the effective supporting fragments are processed by deducting the amount of processing according to the candidate domain alignment is compared with the corresponding boundary lower limit. When the candidate category boundary interval of each valid support fragment is not lower than the corresponding boundary preservation lower limit, and there is no relationship reversal where the nearest reference relationship of any valid support fragment changes from the reference of the same type of target domain to the reference of the nearest different type of target domain after trial deduction, it is determined that the preservation condition corresponding to the boundary reference set is satisfied.

4. The method of claim 1, wherein, Before comparing the supporting fragment representation with the prototype of the same source domain class and extracting the deviation relative to the source domain class prototype as the residual representation of the target domain, the process also includes a supporting fragment reliability screening process, which specifically includes: Based on the distance from each source domain segment representation to the prototype of the same source domain class, the intra-class distance distribution of the source domain segment representation under each fault category is statistically analyzed, and the upper bound of the deviation distribution of the corresponding fault category is set according to the intra-class distance distribution. For each supporting fragment representation, calculate its distance to the prototype of the same source domain class and compare it with the upper bound of the deviation distribution of the corresponding fault category; Support segments whose distance does not exceed the upper bound of the deviation distribution are marked as reliable support segments, and support segments whose distance exceeds the upper bound of the deviation distribution are marked as support segments to be verified. Reliable support fragments are used as effective fragments in the support fragment representation to extract residual representations of the target domain, and residual representations of the target domain are extracted based on the effective fragments. The support fragments to be verified are excluded from the extraction range of the residual representation of the target domain, and after generating the candidate domain alignment processing amount, the support fragments to be verified and the reliable support fragments are used together to calculate the candidate class boundary interval.

5. The method of claim 4, wherein, After extracting the residual representation of the target domain using reliable support fragments, and before cross-referencing the residual representation of the target domain by fault category, a residual direction stability verification process is also included, specifically including: The target domain residual representations corresponding to the reliable support segments are grouped according to their respective fault categories to obtain the intra-class residual set for each fault category; Perform directional consistency statistics on the target domain residual representations in each class residual set to obtain the degree of residual directional concentration for the corresponding fault category; The set of residues within a class whose concentration of residue direction is not lower than the preset lower limit of concentration is taken as a single set of residues and entered into the subsequent cross-matching. The intra-class residual sets whose residual direction concentration is lower than the preset concentration lower limit are clustered and split according to the similarity of residual directions to obtain multiple residual subsets; The single residual set and residual subset are used as inputs for cross-matching by fault category, respectively.

6. The method of claim 2, wherein, Based on the common deviation direction of cross-class co-occurrence domains remaining in the supporting fragment representation, after generating candidate domain alignment processing quantities, a multi-candidate trial deduction screening process for candidate domain alignment processing quantities is also included, specifically including: Based on the common deviation direction component in the cross-class co-occurrence domain residue and the preset deduction intensity level, multiple candidate domain alignment processing quantities are combined to form a candidate processing quantity set. For each candidate domain alignment processing quantity in the candidate processing quantity set, the corresponding common deviation portion is subtracted from the support fragment representation to obtain the corresponding candidate support fragment representation group. For each candidate support fragment representation group, calculate the candidate category boundary interval corresponding to each candidate support fragment representation to obtain the candidate category boundary interval distribution; The candidate domain alignment processing amount that satisfies the boundary reference set correspondence preservation condition of the candidate category boundary spacing distribution is included in the qualified candidate set; The candidate domain alignment processing amount with the highest degree of common deviation reduction and the best degree of preservation of the candidate category boundary interval distribution is selected from the qualified candidate set and used as the final candidate domain alignment processing amount. When the qualified candidate set is empty, select the candidate domain alignment processing amount with the lowest degree of decrease in the candidate category boundary interval, and process it according to the path that keeps the target domain class reference unchanged.

7. The method according to claim 6, characterized in that, After obtaining the fault diagnosis result based on the category boundary interval between the deducted segment representation and the target domain class reference, the process also includes cross-batch common deviation stability record and update rejection procedure, specifically including: The candidate domain alignment processing amount and its corresponding common deviation direction component used in the diagnostic batch of each target domain are recorded in the cross-batch common deviation record in chronological order. For each common deviation direction component in the cross-batch common deviation record, count the number of times it appears repeatedly in the consecutive diagnostic batches, and verify whether the corresponding deduction process satisfies the preservation condition corresponding to the boundary reference set. Common deviation direction components whose recurrence count is not less than the preset lower limit of consecutive batches and whose corresponding deduction process meets the retention conditions are registered as target domain stable common offsets. In subsequent diagnostic batches, the characterization of the segment to be diagnosed is first pre-subtracted according to the stable common offset of the target domain, and then the generation and screening of candidate domain alignment processing volume is carried out. If the target domain stable common offset causes the class boundary interval to fall below the lower limit allowed by the corresponding preservation condition in a subsequent diagnostic batch, the update of the target domain stable common offset for that diagnostic batch is rejected, and the corresponding common deviation direction component is removed from the target domain stable common offset.

8. A cross-domain few-shot intelligent fault diagnosis system based on meta-learning, characterized in that, To implement the cross-domain few-sample intelligent fault diagnosis method based on meta-learning as described in any one of claims 1-7, comprising: The segment representation extraction module is used to acquire source domain fault samples, a small number of target domain support samples, and target domain samples to be diagnosed, and to extract acoustic and vibration segments in a unified manner to obtain source domain segment representations, support segment representations, and samples to be diagnosed. The class prototype aggregation module is used to extract the fault categories of source domain fault samples, perform intra-class consistency aggregation of source domain fragment representations according to fault categories, and obtain source domain class prototypes. The common residue extraction module is used to compare the supporting fragment representation with the prototype of the same source domain class, extract the deviation part relative to the prototype of the source domain class as the target domain residue representation, and cross-compare it according to the fault category. The common deviation part that appears repeatedly in multiple fault categories and the difference between categories is lower than the preset residue retention condition is selected as the cross-class co-occurrence domain residue. The target reference generation module is used to subtract cross-class co-occurrence domain residues from the supporting fragment representation, retain the part whose distance relationship with the prototype of the same source domain class satisfies the preset same-class preservation condition, and generate the target domain class reference accordingly. The boundary set construction module is used to represent each supporting fragment, calculate its distance to the class reference of the same type of target domain and the nearest class reference of the opposite type of target domain, and form a boundary reference set; The candidate processing evaluation module is used to generate candidate domain alignment processing amount based on the common deviation direction of cross-class co-occurrence domains remaining in the support fragment representation, subtract it from the support fragment representation to obtain the candidate support fragment representation, and subtract the distance to the same class target domain reference from the distance of the candidate support fragment representation to the nearest heterogeneous target domain class reference to obtain the candidate class boundary interval. The reference update decision module is used to deduct the common deviation part in the support fragment representation and regenerate the target domain class reference if the candidate category boundary interval meets the preservation condition corresponding to the boundary reference set; otherwise, the target domain class reference remains unchanged. The fault diagnosis module is used to subtract the common deviation part in the representation of the segment to be diagnosed by the candidate domain alignment processing amount, and obtain the fault diagnosis result based on the category boundary interval between the subtracted segment representation and the target domain class reference.

9. 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 method described in any one of claims 1-7.

10. A computer-readable storage medium storing a computer program, characterized in that, When the computer program is executed by a processor, it implements the method as described in any one of claims 1-7.

Images