A physically-inlined sparse recognition architecture system, and an anomaly recognition method and device

Abstract

This invention discloses a physically embedded sparse identification architecture system and anomaly identification method and device. By embedding a physical anomaly response dictionary into the identification architecture, the observed differential response is mapped to the component pattern atomic space. A structured physical semantic activation is generated in this atomic space through a dictionary structure interaction layer, ultimately achieving the identification of anomaly components, anomaly patterns, anomaly support sets, anomaly scores, and optional effective sparse coefficients. This invention can embed the physical anomaly response dictionary into the anomaly identification process, ensuring consistency between anomaly semantic output, component pattern support, effective sparse coefficients, and physical response reconstruction, thereby improving the physical interpretability of anomaly identification, the stability of support set recognition, and adaptability to multiple anomaly scenarios.

Description

Technical Field

[0001] This invention belongs to the field of physical system anomaly identification and intelligent diagnosis technology, specifically involving a physically embedded sparse identification architecture system and anomaly identification method and device. Background Technology

[0002] Modular physical systems typically consist of multiple physical components, and different components may exhibit different types of anomalous modes. For such systems, anomaly identification tasks not only need to determine whether an anomaly exists in the system, but also need to further determine which component the anomaly occurs in, which anomaly mode it belongs to, and the corresponding anomaly score or confidence level for the diagnostic result.

[0003] For example, in wave systems, electromagnetic systems, acoustic systems, heat conduction systems, structural dynamics systems, and other multi-component coupled physical systems, changes in local source terms, material parameters, structural parameters, boundary conditions, or local physical properties will produce corresponding response changes at the observation point. This type of anomaly identification problem can essentially be understood as: identifying the activated combination of component modes based on changes in the observed system response.

[0004] Existing anomaly identification methods mainly include rule-based diagnostic methods, manual feature-based classification methods, sparse identification methods based on optimization inversion, and data-driven methods based on neural networks.

[0005] Rule-based or feature-based methods typically rely on expert experience and require manual design of diagnostic features. As the number of system components increases, abnormal patterns multiply, or physical response coupling strengthens, these methods often struggle to maintain stable recognition performance.

[0006] Sparse identification methods based on optimization inversion typically transform the anomaly identification problem into a sparse coefficient solving problem. For example, the anomaly response can be represented as a combination of a physical dictionary and sparse coefficients, and the anomaly support locations can be solved using LASSO, OMP, or other sparse optimization methods. These methods offer some physical interpretability, but typically rely on regularization parameters, iterative solutions, and dictionary properties. Their solution stability and identification performance are easily affected when measurement noise is strong, dictionary atom correlations are high, or the physical model exhibits mismatch.

[0007] Data-driven methods based on neural networks can learn a nonlinear mapping from observed responses to anomaly labels through training data, exhibiting strong fitting capabilities. However, ordinary neural networks typically input observed responses directly into a black-box feature extraction module and then output classification results. The intermediate features lack clear physical meaning, making it difficult to explain why the model judges a component or pattern as abnormal. Furthermore, the generalization ability of purely data-driven models is easily limited when training samples are limited, system operating conditions change, or the physical structure alters.

[0008] For modular physical systems, each component and its candidate anomaly modes can typically generate corresponding typical anomaly responses through a physical model. For example, the observed changes caused by a component under a certain anomaly mode can be obtained through analytical physical models, numerical simulation models, semi-analytical models, equivalent physical models, experimentally calibrated or data-corrected physical models, or combinations of the above models. If these anomaly responses are constructed into a physical anomaly response dictionary, and this dictionary is used as the basis for physical coordinates in the identification architecture, the observed differential responses can first be transformed into the component mode atomic space with clear physical semantics, and then the trainable modules can learn complex structural relationships.

[0009] The following shortcomings still exist in the existing technology.

[0010] First, pure data-driven anomaly diagnosis models typically do not explicitly introduce a physical anomaly response dictionary, resulting in the lack of physical coordinate meaning for intermediate network features and insufficient interpretability of the diagnosis process.

[0011] Second, although traditional sparse optimization methods can be used for anomaly support identification, they often limit the problem to the solution of explicit sparse coefficients, which is easily affected by regularization parameters, optimization algorithms, noise levels and dictionary correlations.

[0012] Third, existing physical-guided neural networks focus more on physical equation residuals, physical constraint losses, or data fitting accuracy, and rarely establish a complete anomaly identification chain from the architectural level, namely "observational differential response - physical dictionary coordinates - structured anomaly representation - physical semantic output".

[0013] Fourth, existing anomaly diagnosis models often treat component localization, pattern recognition, sparse coefficient estimation, and physical response reconstruction in a fragmented manner, lacking a unified architecture that can keep anomaly semantics, component pattern support, effective sparse coefficients, and dictionary reconstruction response consistent.

[0014] Fifth, existing methods rarely explicitly utilize same-component pattern competition, same-pattern cross-component sharing, dictionary coherence, and physical adjacency relationships at the physical dictionary atom level, resulting in insufficient stability of anomalous support set identification under conditions of high-coherence dictionary atoms, weak cross-response, or noise perturbation.

[0015] Therefore, it is necessary to propose a physically embedded sparse identification architecture that can embed a physical anomaly response dictionary into the anomaly identification process, learn structured anomaly representations in the component pattern atomic space, generate structured physical semantic activations through dictionary structure interaction, and obtain anomaly components, anomaly patterns, and anomaly scores through physical semantic decoding. Summary of the Invention

[0016] The purpose of this invention is to address the problems existing in current physical system anomaly identification methods, such as insufficient physical interpretability, over-reliance on explicit sparse optimization, insufficient utilization of component pattern structure relationships, and insufficient consistency between anomaly semantic output, sparse coefficient estimation, and physical response reconstruction. This invention provides a physically embedded sparse identification architecture system and anomaly identification method and device. The system takes the observed differential response as input and a physical anomaly response dictionary as the embedded physical coordinate basis. Through a physical dictionary projection layer, a dictionary structure interaction layer, and a physical semantic decoding layer, a complete identification chain from observed response changes to anomaly diagnosis results is established.

[0017] To achieve the above objectives, in one aspect, the present invention provides a physically embedded sparse identification architecture system. This sparse identification architecture system is used to identify anomalous components and anomalous patterns in a modular physical system. The modular physical system includes multiple physical components, each physical component corresponding to one or more candidate anomalous patterns. The sparse identification architecture system includes:

[0018] The differential response acquisition unit is used to obtain the observation differential response based on the reference state observation response and the state observation response to be identified of the modular physical system.

[0019] The physical anomaly response dictionary unit is used to generate anomaly response atoms for component pattern combinations consisting of physical components and candidate anomaly patterns, and multiple anomaly response atoms constitute the physical anomaly response dictionary.

[0020] And a sparse identification architecture unit, which includes a physical dictionary projection layer, a dictionary structure interaction layer, and a physical semantic decoding layer connected in sequence:

[0021] The physical dictionary projection layer is used to map the observed differential response to the component pattern atomic space based on the physical anomaly response dictionary, so as to obtain the projection features with component pattern physical semantics.

[0022] The dictionary structure interaction layer is used to interactively model the structural relationships between component pattern atoms based on the projection features, so as to obtain a structured physical semantic representation.

[0023] The physical semantic decoding layer is used to output anomaly diagnosis results based on the projection features and the structured physical semantic representation;

[0024] The anomaly diagnosis result includes at least anomaly components and anomaly patterns, and the physical semantic decoding layer is also used to output effective sparse coefficients corresponding to the atomic space of the component patterns. The effective sparse coefficients are used to reconstruct the observed differential response in the dictionary subspace corresponding to the physical anomaly response dictionary.

[0025] In the aforementioned sparse identification architecture system embedded in physics, the modular physical system includes wave systems, electromagnetic systems, acoustic systems, heat conduction systems, fluid systems, structural dynamic systems, and complex networked physical systems; the candidate anomaly modes include one or more of the following: source term perturbation, material parameter perturbation, structural parameter perturbation, boundary condition perturbation, local property change, connection relationship change, load change, damping change, scattering characteristic change, or coupling relationship change.

[0026] In the aforementioned physically embedded sparse identification architecture system, in the differential response acquisition unit, the reference state observation response can be a health state observation response, a baseline operating condition observation response, a calibration state observation response, or a historical reference observation response; the state observation response to be identified can be a current observation response, an abnormal candidate state observation response, or a condition observation response to be diagnosed.

[0027] The observation differential response includes one or more of the following: time domain response differential, frequency domain response differential, field response sampling differential, sensor response differential, boundary response differential, or multi-condition response differential; when the observation response is a complex response, the observation differential response is expanded into a real number form and then input into the physical dictionary projection unit.

[0028] The anomalous response atom is used to characterize the typical differential response morphology generated by the corresponding component mode combination in the observation space; the anomalous response atom is generated by an analytical physical model, a numerical simulation model, a semi-analytical model, an equivalent physical model, an experimental calibration model, a data correction physical model, or a combination of the above models.

[0029] In the aforementioned physically embedded sparse identification architecture system, the physical anomaly response dictionary includes a dictionary matrix formed by the arrangement of anomaly response atoms corresponding to different component pattern combinations; the anomaly support set in the component pattern atom space is used to represent the activated component pattern combination.

[0030] Specifically, in an optional embodiment, the dictionary of observed differential response and physical anomaly response is constructed as follows:

[0031] This invention pertains to modular physical systems. Suppose the modular physical system comprises C physical components, and each physical component contains M candidate anomaly modes. Then, the c-th physical component and the m-th candidate anomaly mode constitute a component mode combination, denoted as:

[0032]

[0033] As mentioned above, the input to the architecture is not the absolute observed response of the modular object system. The reference state observed response can be the health state observed response, the baseline operating condition observed response, the calibration state observed response, or the historical reference observed response, as long as it can serve as a differential reference for the observed response of the state to be identified.Figure 2 As shown. Let the reference state observed response be... The observed response of the state to be identified is The observed difference response is:

[0034]

[0035] in, This indicates the change in the observed response caused by the anomaly.

[0036] For complex observation responses, they can be expanded into real number form:

[0037]

[0038] In one implementation, the model input That is rather than absolute observation response .

[0039] For each component pattern combination Generate corresponding anomalous response atoms based on the physical model. This anomalous response atom represents the typical differential response pattern generated in the observation space when the c-th physical component experiences the m-th candidate anomalous mode.

[0040] like Figure 3 As shown, the physical exception response dictionary is formed by arranging the corresponding exception response atoms of all component patterns:

[0041]

[0042] If the total number of atoms in the component pattern is ,but Can be represented as containing A dictionary of physical differential responses for each anomalous response atom.

[0043] Under the aforementioned physical anomaly response dictionary, the observed differential response can be expressed as:

[0044]

[0045] Where x represents the anomalous sparsity coefficients in the component pattern atomic space, and r represents the structural residuals, model mismatch residuals, or unmodeled perturbations outside the dictionary subspace. This indicates measurement noise.

[0046] In its real number expansion form, the above equation can be written as:

[0047]

[0048] in, This represents the real number expansion of the dictionary of physical anomaly responses.

[0049] The exception support set in the atomic space of the component pattern is defined as follows:

[0050]

[0051] in, This represents the anomaly sparsity coefficient corresponding to the c-th component and the m-th candidate anomaly mode.

[0052] It should be noted that the sparsity in this invention refers to the support sparsity of anomalies in the component pattern atomic space, that is, usually only a small number of component pattern atoms are activated, rather than limiting this architecture to use an explicit sparse optimization algorithm for solving.

[0053] In the aforementioned physically embedded sparse identification architecture system, the physically embedded sparse identification architecture unit provided by this invention comprises three sequentially connected parts: a Physical Dictionary Projection (PDP) layer, a Dictionary Structure Interaction (DSI) layer, and a Physical Semantic Decoding (PSD) layer. Its overall chain is as follows:

[0054]

[0055] Wherein, PDP represents the physical dictionary projection layer, DSI represents the dictionary structure interaction layer, and PSD represents the physical semantic decoding layer; Represents the physical semantic activation of relevance. This represents the component pattern block features organized according to component and pattern dimensions. This represents the structured physical semantic activation output by the dictionary structure interaction layer. This represents the set of predicted anomaly diagnoses.

[0056] In the aforementioned sparse identification architecture system with embedded physics, the physical dictionary projection layer is used to calculate one or more of the following: similarity, correlation, projection coefficient, regularized projection feature, or amplitude semantic feature between the observed differential response and the anomalous response atom in the physical anomalous response dictionary, and organize them into component pattern block features.

[0057] Preferably, the physical dictionary projection layer normalizes the anomalous response atoms and calculates the correlation physical semantic activation between the observed differential response and the normalized anomalous response atoms. The correlation physical semantic activation is used to characterize the signed similarity between the observed differential response and each component mode atom.

[0058] Specifically, in an optional embodiment, the Physical Dictionary Projection (PDP) layer maps the observed differential responses to the component pattern atomic coordinate space defined by the physical anomaly response dictionary, as follows: Figure 4 As shown.

[0059] In one implementation, each anomalous response atom in the real number expansion dictionary is first normalized:

[0060]

[0061] in, To prevent division by zero of extremely small positive numbers, a normalized physical anomaly response dictionary is constructed from all normalized anomaly response atoms.

[0062]

[0063] PDP calculates the correlation activation between the observed differential response and each normalized anomalous response atom:

[0064]

[0065] in, This represents the correlation physical semantic activation, reflecting the signed similarity between the observed differential response and the mode atoms of each component.

[0066] In an alternative form, the PDP can also construct regularized projection features:

[0067]

[0068] in, Let I be the regularization parameter and I be the identity matrix. This regularized projection feature can be used as an auxiliary projection feature, but not as the final anomalous sparsity coefficient.

[0069] Furthermore, amplitude semantic features can be constructed:

[0070]

[0071] Reorganize one or more of the above projection features according to component and pattern dimensions to obtain component pattern block features:

[0072]

[0073] in, This represents the projection feature corresponding to each component mode atom. For example, when including correlation projection, regularized projection, and amplitude features, F=3; when including correlation projection and amplitude features, F=2. The specific value of F can be determined according to the type of projection feature used, and is not limited to the values ​​mentioned above.

[0074] The core function of PDP is that it is not a regular black-box feature extraction layer, but rather uses a physical anomaly response dictionary to transform the observed differential response into a physical coordinate space with explicit component pattern semantics.

[0075] In the above-mentioned physically embedded sparse identification architecture system, the structural relationships modeled by the dictionary structure interaction layer include at least one of the following: the relationship between different abnormal modes within the same component, the relationship between the same or related abnormal modes between different components, the atomic response correlation determined by the coherence of the physical abnormal response dictionary, the spatial proximity relationship determined by the geometric position of the component, the topological relationship determined by the physical connection relationship of the component, and the co-occurrence relationship or mutual exclusion relationship obtained by statistical analysis of training data.

[0076] Preferably, the dictionary structure interaction layer performs a structured transformation on the projected features through a trainable mapping, wherein the trainable mapping includes threshold mapping, multilayer perceptron, convolutional mapping, attention mechanism, graph neural network, gating network or a combination thereof.

[0077] More preferably, the threshold mapping includes a learnable signed soft threshold transformation for suppressing low-amplitude or weakly cross-component mode atomic responses and preserving component mode atomic activations with anomalous physical significance; the threshold is set at the global granularity, component granularity, mode granularity, or atomic granularity.

[0078] Specifically, in an optional embodiment, the Dictionary Structure Interaction (DSI) layer is used to model the structural relationships between component pattern atoms based on the physical dictionary projection features, and to generate structured physical semantic activations, as shown in the flowchart below. Figure 5 As shown. Each dictionary atom corresponds to a component pattern composition:

[0079]

[0080] In one implementation, the input to the DSI is the relevance physical semantic activation of the PDP output. After rearranging according to component and pattern dimensions, the projected features corresponding to the c-th component and the m-th candidate anomaly pattern can be obtained.

[0081] DSI Apply a learnable signed soft thresholding transformation:

[0082]

[0083] in, Indicates the learnable threshold. This represents the structured physical semantic activation of the DSI output. To ensure the threshold is non-negative, we can set:

[0084]

[0085] in, These are trainable parameters.

[0086] The learnable threshold can be set at the global granularity, component granularity, pattern granularity, or atomic granularity. Global granularity means all component pattern atoms share the same threshold; pattern granularity means the same abnormal pattern shares a threshold; atomic granularity means each component pattern atom has an independent threshold.

[0087] In an extended implementation, DSI can further incorporate global observation features u and dictionary atom embedding features. and dictionary structure relationship set This enhances the modeling of structural interactions between atoms in the component pattern.

[0088] in, This represents the global observation features extracted from the observed differential response. This represents the global observation feature extraction function, which can be a statistical feature extraction function, a trainable coding network, or other feature mapping function.

[0089] , represents the embedding feature of the (c,m)th dictionary atom. This represents a dictionary atomic embedding function, which can be a fixed feature map, a trainable embedding layer, or an encoding function based on dictionary atomic metadata;

[0090] A dictionary-structured set of relations can include one or more relations: The relationships can include mutual exclusion, competition, co-occurrence, or correlation. This indicates the intra-pattern relationship between different exception patterns within the same component; Indicates cross-component relationships between the same or related exception patterns among different components; This represents the atomic response correlation determined by the coherence of the physical anomaly response dictionary; This represents spatial relationships determined by the geometric location, spatial proximity, or physical connection of components. In the absence of geometric location or physical proximity data, this is useful when such data is unavailable. It can be ignored.

[0091] In this extended implementation, DSI can be represented as:

[0092]

[0093] in, This represents a trainable dictionary-structured interaction function. This function can employ soft-thresholding mapping, multilayer perceptrons, convolutional mapping, gating networks, attention mechanisms, graph neural networks, or other trainable mapping functions that can utilize component pattern structural relationships. The above extended implementation does not change the basic function of DSI, namely, generating structured physical semantic activation based on physical dictionary projection features. .

[0094] In the aforementioned physically embedded sparse identification architecture system, the physical semantic decoding layer includes a component-level decoding subunit, a component pattern atomic-level decoding subunit, a coefficient candidate output subunit, and an effective coefficient generation subunit.

[0095] The component-level decoding subunit is used to output component-level exception support information;

[0096] The component mode atomic-level decoding subunit is used to output component mode atomic-level exception support information;

[0097] The coefficient candidate output subunit is used to output continuous coefficient candidates corresponding to the component mode atom;

[0098] The effective coefficient generation subunit is used to gate or weight the continuous coefficient candidates based on the component-level anomaly support information, the component pattern atomic-level anomaly support information, the structured physical semantic representation, or a combination thereof, to obtain the effective sparse coefficients.

[0099] Preferably, the effective sparsity coefficients are jointly determined by the hierarchical effective gating value and the candidate continuous coefficients; the hierarchical effective gating value is obtained by combining component-level anomaly support information and component pattern atomic-level anomaly support information.

[0100] In the aforementioned physically embedded sparse identification architecture system, the anomaly diagnosis results also include one or more of the following: anomaly support set, number of anomalies, anomaly score, component confidence, component pattern atomic confidence, effective sparsity coefficient, dictionary subspace differential response reconstruction result, or anomaly severity.

[0101] Preferably, the anomaly score, component confidence, or pattern confidence is used to characterize the degree of support of the model for predicting anomaly entries. It is determined by component-level anomaly support information, component pattern atomic-level anomaly support information, structured physical semantic representation, hierarchical effective gating value, effective sparse coefficient magnitude, or a combination thereof, and is not limited to statistical probabilities that have been probability-calibrated.

[0102] In the aforementioned sparse identification architecture system with embedded physics, the physical semantic decoding layer reconstructs the response components of the observed differential response in the dictionary subspace based on the effective sparse coefficients and the physical anomaly response dictionary; the reconstructed response components are used to constrain or verify the consistency between the anomaly diagnosis results and the physical response changes.

[0103] Specifically, in an optional embodiment, the Physical Semantic Decoding (PSD) layer is used to perform component pattern block features based on the PDP output and the structured physical semantic activation based on the DSI output. Its flowchart and output are as follows: Figure 6 As shown. Decoding exception components, exception patterns, component pattern atomic support, effective sparse coefficients, and dictionary reconstruction responses.

[0104] The component pattern block characteristics output by PDP are:

[0105]

[0106] Where C represents the number of components, M represents the number of candidate anomaly modes corresponding to each component, and F represents the projection feature dimension of each component mode atom.

[0107] For the c-th component and the m-th candidate anomaly mode, its corresponding component mode block feature vector is denoted as:

[0108]

[0109] The structured physical semantic activation output by DSI is denoted as: , representing the structured physical semantic activation corresponding to the c-th component and the m-th candidate exception mode.

[0110] PSD first extracts the component pattern block feature vector With structured physical semantic activation Combined, atomic-level input features are formed:

[0111]

[0112] Generate atomic embeddings of component patterns using atomic-level semantic encoding functions:

[0113]

[0114] in, This represents a trainable atomic-level semantic encoding function. This represents the sample-related atom embedding corresponding to the c-th component and the m-th candidate anomaly pattern. This encoding process preserves the component pattern structure and is used to assist PSD decoding.

[0115] Component-level physical evidence can be obtained by aggregating structured physical semantic activations corresponding to various anomalous patterns within the same component. For example:

[0116]

[0117] in, This represents the physical evidence strength of the c-th component. This quantity is the evidence strength obtained by aggregating structured physical semantic activations, and is not necessarily limited to a probability-calibrated statistical probability.

[0118] Component-level atomic embeddings can be obtained by aggregating the atomic embeddings of candidate exception patterns under the same component:

[0119]

[0120] PSD outputs component-level logits based on component-level embedding and component-level physical evidence:

[0121]

[0122] Component-level anomaly probabilities can be obtained from DSI gating, atomic embedding, or their aggregated features within the same component:

[0123]

[0124] in, This represents the probability that the c-th component will experience an anomaly. This represents the sigmoid function.

[0125] The atomic logits of the component pattern can be represented as:

[0126]

[0127] in Represents a trainable component pattern atomic classification function used to embed the semantics of the (c,m)th component pattern atom. and structured physical semantic activation The mapping is converted to atomic logits in a component-mode framework. This function can be implemented using linear mappings, multilayer perceptrons, or other trainable classification mappings.

[0128] The probability of atomic support for component pattern is:

[0129]

[0130] This represents the probability that the c-th component will exhibit the m-th candidate anomaly pattern.

[0131] In a preferred embodiment, component pattern atomic support can be directly gated by DSI. Characterization can also be derived from and The combined representation of.

[0132] PSD also outputs ungated continuous coefficient candidates:

[0133]

[0134] in, This represents the candidate ungated continuous coefficients corresponding to the c-th component and the m-th candidate anomaly mode. This represents a trainable coefficient prediction function.

[0135] To maintain consistency in component-level support, atomic support for component patterns, and effective sparsity coefficients, PSD introduces hierarchical effective gating values:

[0136]

[0137] in, This represents the effective gating value of the hierarchy corresponding to the c-th component and the m-th candidate exception mode.

[0138] The final effective sparsity coefficients used in the physical dictionary reconstruction are:

[0139]

[0140] in, This represents the effective sparse coefficient corresponding to the c-th component and the m-th candidate anomaly mode. This design suppresses the continuous coefficient candidates of atoms with low component-level support probability or low component mode atomic-level support probability; only component mode atoms with high support evidence can form effective sparse coefficients and participate in physical reconstruction.

[0141] All A vector is formed according to the atomic arrangement order in the physical anomaly response dictionary. Based on the effective sparsity coefficients and the normalized physical anomaly response dictionary It is possible to reconstruct the observed difference response in the dictionary subspace:

[0142]

[0143] The output is defined as follows.

[0144] The basic output of this invention is a set of abnormal diagnostics:

[0145]

[0146] in, Indicates the first One predictive anomaly component, Indicates the first One predicted abnormal pattern, This indicates the predicted number of anomalies.

[0147] Predicting anomaly support set Probabilistic support at the atomic level can be based on component patterns. Hierarchical effective gate value Or a combination of both. For example, in one implementation, the following can be taken:

[0148] ,or:

[0149] in, This indicates an anomaly support threshold. In a preferred embodiment, the threshold... The support set determination can be based on the validation set, target false positive rate, target recall rate, or task requirements to improve the stability of support set determination under different data scales, different noise conditions, and different dictionary coherence conditions.

[0150] In an optional implementation, the anomaly diagnosis result may further include an anomaly score or confidence level for each predicted anomaly entry, expressed as:

[0151]

[0152] Abnormal scoring The probability used to characterize the model's support for the j-th predicted outlier should not be narrowly limited to probability-calibrated statistical probabilities. Probabilities can be atomically supported by component patterns. Hierarchical effective gate value Component-level physical evidence Determined by or in combination thereof. In a preferred form, the following may be taken:

[0153]

[0154] In another implementation, the following can also be adopted:

[0155]

[0156] In another implementation, a scoring fusion function can be used:

[0157]

[0158] in, This represents the scoring fusion function, which can be a product, weighted sum, maximum value, learnable mapping, or other scoring functions.

[0159] Except for abnormal diagnostic collection In addition, this invention can also output an effective sparse coefficient vector. Differential response reconstructed from dictionary subspace This is used to enhance the physical interpretability and consistency verification of diagnostic results.

[0160] In addition to the above outputs, the severity of the anomaly It can also be used as an extension of the basic architecture of this invention.

[0161] When the training data contains labeled anomaly severity tags, an anomaly severity estimation output header can be added to the physical semantic decoding layer to expand the output as follows:

[0162]

[0163] in, This represents the predicted severity of the j-th anomaly entry.

[0164] In the above-mentioned physically embedded sparse identification architecture system, the architecture unit adopts a training objective function that includes at least one of the following objectives during the training phase: component-level semantic support objective, component pattern atomic-level semantic support objective, effective sparse coefficient supervision objective, differential response reconstruction objective, hierarchical sparse prior objective, support probability and continuity coefficient consistency objective, confidence calibration objective, class imbalance weighting objective, or parameter regularization objective.

[0165] The differential response reconstruction objective is used to constrain the consistency between the reconstructed differential response obtained from the effective sparse coefficients and the physical anomaly response dictionary and the input observed differential response; the hierarchical sparse prior objective is used to encourage the activation of minority components and minority component mode atoms.

[0166] Specifically, in an optional embodiment, the training objective and loss function are designed as follows.

[0167] To ensure that the output of this invention remains consistent across anomaly diagnosis semantics, component pattern atomic support, effective sparsity coefficients, and physical dictionary reconstruction response, this invention may employ a physical semantic sparsity consistency training objective.

[0168] The training objective can be organized into three composite parts: a semantic support consistency objective, a physical coefficient reconstruction consistency objective, and a hierarchical sparse prior. The overall training objective can be expressed as:

[0169]

[0170] in, The semantics indicate that the consistency goal is supported. This indicates the consistency target for physical coefficient reconstruction. This indicates sparse prior knowledge at different levels. , and These are the weighting coefficients for the corresponding loss terms.

[0171] In a preferred embodiment of the present invention, the physical coefficient reconstruction consistency target Effective sparsity coefficients based on dictionary-structured interactive gating The observed differential response is constructed and further reconstructed using a physical anomaly response dictionary. .therefore, Used to constrain the consistency between DSI sparse gating, effective sparse coefficients, and dictionary subspace reconstruction response.

[0172] In a preferred implementation of the present invention, a hierarchical sparse prior is used. The structured physical semantic activation applied to the dictionary structure interaction layer output is used to encourage the activation of a few components and a few component pattern atoms, and can further include consistency constraints of component-level gating and atom-level gating, as well as competition or inhibition constraints between highly coherent dictionary atoms.

[0173] It should be noted that the above three composite objectives can be used in combination according to the training data label format and task requirements. In different implementations, one, two, or all of the loss terms can be enabled, and confidence calibration loss, class imbalance weighting terms, or other regularization terms can be further added.

[0174] The tag is defined as follows.

[0175] For the i-th training sample, if its true anomaly set is:

[0176]

[0177] in, This represents the number of real anomalies in the i-th sample.

[0178] Component-level real labels can be constructed from the set of real exceptions:

[0179]

[0180] The atomic-level real tags of the component pattern are:

[0181]

[0182] If the training data provides true sparse coefficients, then it is denoted as... If no real sparse coefficients are provided, the coefficient supervision term can be omitted, and only the semantic support loss, reconstruction loss, or sparse prior can be retained.

[0183] Furthermore, the semantic support consistency objective is used to constrain component-level support prediction and component pattern atomic-level support prediction, ensuring that the model output has explicit anomaly diagnostic semantics. This can be expressed as:

[0184]

[0185] in, This indicates a loss of component-level support. This indicates that the component pattern provides atomic-level support for loss. and These are the weighting coefficients.

[0186] Component-level support loss can be achieved using binary cross-entropy loss:

[0187]

[0188] The atomic support loss of the component pattern can be expressed as:

[0189]

[0190] in, This represents the prediction support score corresponding to the c-th component and the m-th candidate anomaly pattern.

[0191] In multi-label anomaly detection scenarios, the aforementioned BCE loss can be averaged or weighted over all components and all component pattern atoms. In cases of class imbalance, positive sample weighted BCE, focus loss, or other multi-label classification losses can also be used.

[0192] It should be noted that this invention does not require the introduction of an additional, independent explicit hierarchical consistency penalty term. Component-level support probability. Gating has been effectively supported at different levels. It participates in the generation of effective sparse coefficients, enabling component-level support and component pattern atomic-level support to form a closed-loop consistency in the coefficient generation and physical reconstruction stages.

[0193] Furthermore, the physical coefficient reconstruction consistency objective is used to constrain the consistency between the component pattern support probability, the effective sparse coefficients, and the dictionary subspace reconstruction response. It can be expressed as:

[0194]

[0195] in, Indicates the coefficient of supervised loss. Represents the dictionary subspace reconstruction loss. This represents the consistency loss between the support probability and the continuous coefficients; , and These are the weighting coefficients.

[0196] When the training data provides the true sparsity coefficients At that time, the effective sparsity coefficient can be adjusted. Supervision is performed. Considering that positive samples are usually few in the atomic space of component patterns, a positive sample-weighted SmoothL1 loss can be used:

[0197]

[0198] in, Indicates the total number of atoms in the component pattern. This represents the sample weight. The sample weight can be determined based on whether a component pattern atom is a genuine anomalous atom, for example:

[0199]

[0200] in, Increase the weights for positive samples.

[0201] To ensure that the predicted effective sparse coefficients can explain the observed difference response, a dictionary subspace reconstruction loss is introduced:

[0202]

[0203] in:

[0204]

[0205] In different implementations, the reconstruction loss can also be in a normalized form:

[0206]

[0207] This reconstruction focuses on the difference response, not the absolute response. This term ensures that the predicted effective sparse coefficients can explain the observed response variations within the physical dictionary subspace.

[0208] To avoid large hidden continuous coefficients on atoms with low support probability, a consistency constraint between support probability and continuous coefficients can be added:

[0209]

[0210] This item is related to gating. Together, they ensure that the effective coefficient of low-probability anomalous atoms is suppressed.

[0211] Furthermore, hierarchical sparse priors are used to express the physical priors of "few component anomalies" and "few component mode atomic activations" in this invention. They can be represented as:

[0212]

[0213] in, This indicates sparse terms in a component-level group. Represents atomic-level sparse terms; and These are the weighting coefficients.

[0214] The sparse term of a component-level group can be defined as:

[0215]

[0216] Only a few components were activated under this incentive.

[0217] Atomic-level sparsity terms can be defined as:

[0218]

[0219] This encourages sparse activation in the atomic space of the component pattern. Component-level sparse terms and atomic-level sparse terms together constitute the hierarchical sparse prior.

[0220] Furthermore, there are broader combinations of loss functions:

[0221]

[0222] in, Indicates the confidence level calibration loss. This indicates parameter regularization or other auxiliary regularization terms. The above loss terms do not necessarily need to be enabled simultaneously; one or more can be selected based on the label format, task type, and physical constraints.

[0223] The key to this invention is not to limit a specific loss function form, but to organize semantic support consistency, effective sparse coefficients, physical dictionary difference response reconstruction, and hierarchical sparse prior as a unified training objective, so that the anomaly diagnosis results, component pattern support, effective sparse coefficients, and physical response reconstruction of this invention are consistent.

[0224] On the other hand, the present invention also provides a method for identifying anomalies in a physical system. This method is based on the aforementioned sparse identification architecture system embedded in physical systems and includes the following steps:

[0225] Obtain the reference state observation response and the state observation response to be identified of the modular physical system, and construct the observation differential response;

[0226] Component pattern combinations are constructed based on physical components and candidate anomaly patterns, and anomaly response atoms are generated for each component pattern combination to form a physical anomaly response dictionary;

[0227] Based on the physical anomaly response dictionary, the observed differential response is mapped to the component pattern atomic space to obtain projection features with component pattern physical semantics.

[0228] The projected features are subjected to dictionary structure interaction modeling to obtain a structured physical semantic representation;

[0229] Based on the projection features and the structured physical semantic representation, physical semantic decoding is performed to obtain anomaly diagnosis results;

[0230] Generate effective sparse coefficients corresponding to the component pattern atomic space, and reconstruct the response components of the observed differential response in the dictionary subspace based on the effective sparse coefficients and the physical anomaly response dictionary.

[0231] In the above-mentioned physical system anomaly identification method, the anomaly diagnosis results include one or more of the following: anomaly component, anomaly pattern, anomaly support set, anomaly score, anomaly number, component confidence, pattern confidence, or anomaly severity.

[0232] In addition, the present invention also provides a computer device, including a processor and a memory, wherein the memory stores a computer program, and the computer program, when executed by the processor, implements the aforementioned anomaly identification method.

[0233] The present invention also provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the aforementioned anomaly identification method.

[0234] Compared with the prior art, the physically embedded sparse identification architecture system and anomaly identification method and device provided by the present invention have at least the following beneficial effects:

[0235] (1) This invention embeds a physical anomaly response dictionary into the anomaly identification architecture, so that the observed differential response is first mapped to the component pattern atomic coordinate space defined by the physical model. Compared with directly using ordinary neural networks for black box classification, the intermediate features of this invention have clear physical semantics.

[0236] (2) The sparsity in this invention represents the support for anomaly identification in the atomic space of component patterns, rather than being limited to explicit sparse optimization algorithms. Therefore, this invention retains the essential nature of sparse anomaly identification while avoiding the strong dependence of traditional sparse optimization methods on regularization parameters, optimization algorithms, and dictionary coherence.

[0237] (3) The present invention performs learnable sparse structure transformation on the physical dictionary projection features through the dictionary structure interaction layer, which can suppress weak cross-atom responses and retain sparse anomalous activations with physical meaning.

[0238] (4) This invention links the component-level support probability, the component pattern atomic-level support probability and the effective sparse coefficients through hierarchical effective support gating, so as to form a closed-loop consistency between abnormal semantic output and physical coefficient reconstruction.

[0239] (5) The present invention directly outputs abnormal components, abnormal patterns and abnormal scores through the physical semantic decoding layer, so that the architecture output is consistent with the engineering diagnostic requirements, which facilitates subsequent abnormal location, pattern judgment and maintenance decision-making.

[0240] (6) This invention is not a dedicated model for a specific verification object, but a general architecture for anomaly support and identification in modular physical systems. As long as the target system can construct component pattern anomaly response atoms and form a physical anomaly response dictionary, the architecture described in this invention can be used for anomaly identification.

[0241] (7) This invention is compatible with complex observations and real expansion observations, and can also be applied to single anomaly identification and multiple anomaly identification scenarios. It can also extend the severity estimation output when there is a severity label, and has good scalability. Attached Figure Description

[0242] Figure 1 This is a flowchart of a physically embedded sparse identification architecture system according to the present invention;

[0243] Figure 2 This is a schematic diagram of differential response acquisition in this invention;

[0244] Figure 3 This is the method for constructing the physical anomaly response dictionary in this invention;

[0245] Figure 4 This is a flowchart of the physical dictionary projection layer in this invention;

[0246] Figure 5 This is a flowchart of the dictionary structure interaction layer in this invention;

[0247] Figure 6 This is a flowchart of the physical semantic decoding layer in this invention;

[0248] Figure 7 This is a schematic diagram of the two-dimensional Helmholtz wave field system, local components, and boundary sensor arrangement in Example 2;

[0249] Figure 8 It is the F1 value identified by the component in Example 2;

[0250] Figure 9 It is the F1 score of pattern recognition in Example 2;

[0251] Figure 10 It is the F1 value of the anomaly support set identification in Example 2;

[0252] Figure 11 This refers to the normalized mean square error of the differential response reconstruction in Example 2.

[0253] Figure 12 This is a schematic diagram of the component confidence distribution results under different numbers of abnormal components in Example 2. Detailed Implementation

[0254] The present invention will be further described below with reference to embodiments. It should be understood that the following embodiments are only used to illustrate the implementation of the present invention and are not intended to limit the scope of protection of the present invention. Any equivalent substitutions or modifications made based on the physical anomaly response dictionary, component pattern atomic space, physical dictionary projection, dictionary structure interaction, and physical semantic decoding ideas disclosed in the present invention should be included within the scope of protection of the present invention.

[0255] The physically embedded sparse identification architecture system described in this invention can be implemented by a computer program and deployed in computer devices, servers, edge computing devices, embedded computing platforms, or other electronic devices with data processing capabilities. The computer program is stored in a computer-readable storage medium and, when executed by a processor, implements the anomaly identification method flow described in this invention.

[0256] Example 1

[0257] This embodiment provides a method for anomaly identification by implementing a physically embedded sparse identification architecture system in a general modular physical system. This embodiment is used to illustrate the general implementation process of the architecture of the present invention and does not limit the specific physical system type, physical equation form, number of components, number of anomaly modes, sensor type, observation data type, or specific combination of training loss.

[0258] In this embodiment, the object to be identified is a modular physical system composed of multiple physical components. The physical system can be a wave system, electromagnetic system, acoustic system, heat conduction system, fluid system, structural dynamics system, complex networked physical system, or other multi-component system capable of establishing a physical response model. Each physical component can correspond to one or more candidate anomaly modes. Anomaly modes can be source term perturbations, material parameter perturbations, structural parameter perturbations, boundary condition perturbations, local property changes, connectivity changes, or other anomalies that can cause changes in the observed response.

[0259] The physically embedded sparse identification architecture system provided in this embodiment includes: a differential response acquisition unit, used to obtain the observed differential response based on the reference state observation response and the observation response of the state to be identified of the modular physical system; a physical anomaly response dictionary unit, used to generate anomaly response atoms for component pattern combinations composed of physical components and candidate anomaly patterns, and to construct a physical anomaly response dictionary from multiple anomaly response atoms; and a sparse identification architecture unit.

[0260] The sparse identification architecture unit comprises a physical dictionary projection layer, a dictionary structure interaction layer, and a physical semantic decoding layer connected in sequence. The physical dictionary projection layer maps observed differential responses to the component pattern atomic space based on the physical anomaly response dictionary, obtaining projection features with component pattern physical semantics. The dictionary structure interaction layer interactively models the structural relationships between component pattern atoms based on the projection features, obtaining a structured physical semantic representation. The physical semantic decoding layer outputs anomaly diagnosis results based on the projection features and the structured physical semantic representation. The physical semantic decoding layer includes a component-level decoding subunit, a component pattern atomic-level decoding subunit, a coefficient candidate output subunit, and an effective coefficient generation subunit. The component-level decoding subunit outputs component-level anomaly support information; the component pattern atomic-level decoding subunit outputs component pattern atomic-level anomaly support information; the coefficient candidate output subunit outputs continuous coefficient candidates corresponding to component pattern atoms; and the effective coefficient generation subunit gates or weights the continuous coefficient candidates based on the component-level anomaly support information, the component pattern atomic-level anomaly support information, the structured physical semantic representation, or a combination thereof, to obtain effective sparse coefficients.

[0261] Suppose a modular physical system contains C physical components, and each physical component contains M candidate anomaly modes. Then the c-th physical component and the m-th candidate anomaly mode constitute a component mode combination:

[0262]

[0263] C and M can be set according to the target physical system. In different physical systems, components can represent spatial regions, functional modules, structural units, circuit units, material regions, or other physical units with independently definable abnormal modes.

[0264] First, obtain the observed response of the physical system in a healthy state. The observed response y of the system under the state to be identified can be a sensor measurement, boundary response, field response sample, frequency domain response, time domain response, complex response, or a combination of multiple operating conditions.

[0265] This embodiment does not directly input the absolute response y of the state to be identified into the invention, but instead constructs the observed difference response between the healthy state and the state to be identified:

[0266]

[0267] in, This represents the change in the observed response caused by the anomaly. The purpose of using the observed differential response is to reduce the influence of the healthy background response, so that the subsequent identification process can focus on the changes in the physical response caused by the anomalous disturbance.

[0268] When the observed response is a complex response, it can be expanded into real number form:

[0269]

[0270] When the observed response is itself a real response, it can be directly set as follows: In this embodiment, the input to the architecture is the real-number representation of the differential response. Instead of an absolute response.

[0271] Then, for each component pattern combination (c, m), the corresponding abnormal response atom is constructed based on the physical model of the target physical system. The physical model can be an analytical physical model, a numerical simulation model, a semi-analytical model, an equivalent physical model, an experimental calibration model, or a combination of the above. Anomaly response atoms. This represents the typical differential response pattern in the observation space caused by the occurrence of the m-th candidate anomaly mode in the c-th component.

[0272] Arrange the corresponding exception response atoms of all component patterns to form a physical exception response dictionary:

[0273]

[0274] If the total number of atoms in the component pattern is ,but:

[0275] Under the physical anomaly response dictionary, the observed differential response can be expressed as:

[0276]

[0277] in, This represents the dictionary of physical anomaly responses after real number expansion, where x represents the anomaly sparsity coefficients in the component pattern's atomic space. This represents structural residuals, model mismatch residuals, or unmodeled perturbations outside the dictionary subspace. This indicates measurement noise.

[0278] The exception support set in the atomic space of the component pattern is defined as follows:

[0279]

[0280] in, This represents the anomalous sparsity coefficient corresponding to the c-th component and the m-th candidate anomalous mode. In this embodiment, "sparse" means that the anomalous mode typically activates only a few component mode atoms, rather than requiring an explicit sparse optimization algorithm to solve for x.

[0281] After constructing the physical anomaly response dictionary, the PNSIA architecture is used to identify the observed differential responses. The PNSIA architecture consists of a physical dictionary projection layer, a dictionary structure interaction layer, and a physical semantic decoding layer connected in sequence. Its overall link is represented as follows:

[0282]

[0283] Wherein, PDP represents the physical dictionary projection layer, DSI represents the dictionary structure interaction layer, PSD represents the physical semantic decoding layer, z represents the physical dictionary projection feature, and h represents the structured physical semantic activation. This represents the set of predicted anomaly diagnoses.

[0284] In the physical dictionary projection layer, the anomalous response atoms in the physical anomaly response dictionary are normalized:

[0285]

[0286] in, To prevent division by zero of extremely small positive numbers, a normalized physical anomaly response dictionary is formed from the normalized anomaly response atoms. .

[0287] The physical dictionary projection layer calculates the correlation between the observed differential response and each normalized anomalous response atom, activating the physical semantics.

[0288]

[0289] in, This represents the signed similarity between the observed differential response and the mode atoms of each component.

[0290] In an optional approach, the physical dictionary projection layer can also compute regularized projection features:

[0291]

[0292] in, Let I be the regularization parameter and I be the identity matrix. This regularized projection feature can be used as an auxiliary physical projection feature, but not as the final anomalous sparsity coefficient.

[0293] We can also construct amplitude semantic features based on correlation physical semantic activation:

[0294]

[0295] Reorganize one or more of the correlation projection features, regularized projection features, and magnitude semantic features according to the component and mode dimensions to obtain component mode block features:

[0296]

[0297] Where F represents the projection feature dimension of each component mode atom. The specific value of F can be determined based on whether regularized projection features, magnitude features, or other auxiliary features are enabled.

[0298] In the dictionary-structured interaction layer, a learnable structured transformation is applied to the relevance physical semantic activations to suppress weak cross-atomic responses and preserve activations with physical anomalous meaning. In a preferred embodiment, the dictionary-structured interaction layer employs a signed soft-thresholding transformation.

[0299]

[0300] in, This represents the relevance physical semantic activation corresponding to the c-th component and the m-th candidate anomaly pattern. Indicates the learnable threshold. This indicates the activation of structured physical semantics.

[0301] Learnable thresholds can be implemented at the global, component, pattern, or atomic level. Global granularity means all component pattern atoms share the same threshold; pattern granularity means the same abnormal pattern shares a threshold; atomic granularity means each component pattern atom has an independent threshold.

[0302] It should be noted that, It is neither the final exception support label nor the final exception sparse coefficient, but the structured physical semantic activation output by the dictionary structure interaction layer, which is used for subsequent physical semantic decoding.

[0303] In the physical semantic decoding layer, the output component-level support, component pattern atomic-level support, ungated continuous coefficients, hierarchical effective gating, effective sparse coefficients, and dictionary subspace reconstruction response are activated based on structured physical semantics.

[0304] Component-level support probability is represented as:

[0305]

[0306] in, This indicates that the component level supports logits. This represents the sigmoid function.

[0307] The atomic-level support probability representation of the component pattern is as follows:

[0308]

[0309] in, This indicates that the component pattern supports logits at the atomic level.

[0310] The physical semantic decoding layer also outputs ungated continuous coefficient candidates. To maintain consistency between component-level support, atomic support for component patterns, and effective sparsity coefficients, a hierarchical effective support gating system is introduced:

[0311]

[0312] in, This represents the effective gating value of the hierarchy corresponding to the c-th component and the m-th candidate exception mode.

[0313] The final effective sparsity coefficients used in the physical dictionary reconstruction are:

[0314]

[0315] in, Indicates ungated continuous coefficient candidates. This represents the effective sparsity coefficients after hierarchical effective support gating. This design suppresses continuous coefficients corresponding to low component support probabilities or low component pattern support probabilities, allowing only component pattern atoms with high support evidence to form effective sparsity coefficients.

[0316] Based on the effective sparsity coefficients and the normalized physical anomaly response dictionary, the observed differential responses in the dictionary subspace can be reconstructed:

[0317]

[0318] in, This represents the difference response reconstruction result interpreted from the physical anomaly response dictionary. This reconstruction result is not the system absolute response and does not explicitly predict residual terms. Or measure noise item .

[0319] For the i-th training sample, its true anomaly set can be represented as:

[0320]

[0321] in, This represents the number of real anomalies in the i-th sample.

[0322] Component-level real labels can be constructed from the set of real exceptions:

[0323]

[0324]

[0325]

[0326] During the training phase, semantic support consistency objectives, physical coefficient reconstruction consistency objectives, hierarchical sparse priors, confidence calibration objectives, parameter regularization objectives, or combinations thereof, are employed. The overall training objective can be expressed as:

[0327]

[0328] in, The semantics indicate that the consistency goal is supported. This indicates the consistency target for physical coefficient reconstruction. This indicates a sparse prior hierarchy. Indicates the confidence level calibration target. This indicates parameter regularization or other auxiliary regularization terms; each These are the weighting coefficients for the corresponding loss terms.

[0329] The semantic support consistency objective can be used to constrain component-level support prediction and component pattern atom-level support prediction; the physical coefficient reconstruction consistency objective can be used to constrain the consistency between effective sparse coefficients and observed differential response reconstruction; and the hierarchical sparse prior can be used to encourage the activation of a minority of components and a minority of component pattern atoms. These loss terms do not need to be enabled simultaneously and can be selected based on whether the training data contains component labels, pattern labels, sparse coefficient labels, reconstruction constraints, or confidence calibration requirements.

[0330] During the inference phase, atomic support probability is provided based on the component pattern. Hierarchical effective gate value Component-level physical evidence or a combination thereof, to obtain a set of support for predicting anomalies:

[0331]

[0332] in, This indicates the threshold for determining whether an anomaly is supported. Threshold The threshold can be a fixed threshold or can be set based on validation data, target false positive rate, target recall rate, or task requirements.

[0333] Final output: Anomaly diagnosis set

[0334]

[0335] in, This indicates the j-th predicted anomaly component. This represents the j-th predicted abnormal pattern. This represents the anomaly score or confidence level of the j-th anomaly entry. This indicates the predicted number of anomalies.

[0336] Abnormal scoring It can be obtained from component-level support probability, component pattern atomic-level support probability, hierarchical effective gating value, component-level physical evidence, or a combination of the above quantities. For example, in one implementation:

[0337]

[0338] In another implementation, anomaly scoring can also be performed by... , The anomaly score is derived from component-level physical evidence or a weighted combination thereof. The anomaly score characterizes the model's support for anomaly entries and is not limited to probability-calibrated statistical probabilities.

[0339] Through the above process, this invention achieves integrated identification of observation differential responses, physical dictionary coordinates, structured physical semantic activation, component pattern anomaly support, and dictionary subspace reconstruction in a general modular physical system. This embodiment illustrates that as long as the target system can construct dictionaries of healthy observations, observations to be identified, and component pattern anomaly responses, the architecture of this invention can be used for anomaly component and anomaly pattern identification.

[0340] Example 2

[0341] This embodiment uses a two-dimensional Helmholtz wave field system as an example to illustrate the verification method of the physically embedded sparse identification architecture system of the present invention in a specific physical scenario. This embodiment is only used to verify the feasibility and technical effects of the present invention and does not constitute a limitation on the types of physical systems to which the present invention is applicable.

[0342] The physically embedded sparse identification architecture system provided in this embodiment includes: a differential response acquisition unit, used to obtain the observed differential response based on the reference state observation response and the observation response of the state to be identified of the modular physical system; a physical anomaly response dictionary unit, used to generate anomaly response atoms for component pattern combinations composed of physical components and candidate anomaly patterns, and to construct a physical anomaly response dictionary from multiple anomaly response atoms; and a sparse identification architecture unit.

[0343] The sparse identification architecture unit comprises a physical dictionary projection layer, a dictionary structure interaction layer, and a physical semantic decoding layer connected in sequence. The physical dictionary projection layer maps observed differential responses to the component pattern atomic space based on the physical anomaly response dictionary, obtaining projection features with component pattern physical semantics. The dictionary structure interaction layer interactively models the structural relationships between component pattern atoms based on the projection features, obtaining a structured physical semantic representation. The physical semantic decoding layer outputs anomaly diagnosis results based on the projection features and the structured physical semantic representation. The physical semantic decoding layer includes a component-level decoding subunit, a component pattern atomic-level decoding subunit, a coefficient candidate output subunit, and an effective coefficient generation subunit. The component-level decoding subunit outputs component-level anomaly support information; the component pattern atomic-level decoding subunit outputs component pattern atomic-level anomaly support information; the coefficient candidate output subunit outputs continuous coefficient candidates corresponding to component pattern atoms; and the effective coefficient generation subunit gates or weights the continuous coefficient candidates based on the component-level anomaly support information, the component pattern atomic-level anomaly support information, the structured physical semantic representation, or a combination thereof, to obtain effective sparse coefficients.

[0344] In this embodiment, the system to be identified is a two-dimensional Helmholtz wave field system. For example... Figure 7 As shown, the system generates a wavefield response within a two-dimensional physical region, and boundary sensors acquire the complex responses of the system under incident waves from different directions. The system incorporates multiple local physical components, each corresponding to several candidate anomaly modes. These anomaly modes can include local source term perturbations, local material parameter perturbations, local boundary or scattering characteristic perturbations, or other physical perturbations capable of causing changes in Helmholtz wavefield observations.

[0345] In one specific implementation, the two-dimensional physical region is divided into C=16 local components, and each component is configured with M=3 candidate anomaly modes: local source term perturbation, local material parameter perturbation, and local boundary perturbation. In this case, the total number of mode atoms in the component is: .

[0346] It should be noted that C=16 and M=3 are specific settings only for this embodiment. In other physical systems or other Helmholtz verification scenarios, the number of components and the number of abnormal modes can be adjusted according to the system architecture and diagnostic requirements.

[0347] For a two-dimensional Helmholtz wavefield system, the boundary complex observation response in a healthy state is denoted as . Let y denote the boundary complex observation response under abnormal conditions. First, construct the observed difference response:

[0348]

[0349] Since the Helmholtz wave field response is usually a complex response, it is expanded into real form:

[0350]

[0351] In this embodiment, the input is , rather than the absolute response y of an abnormal state.

[0352] For each local component and its candidate anomaly patterns, corresponding anomaly response atoms are generated based on a two-dimensional Helmholtz physical model. This atom represents the typical differential response generated at the boundary sensor observation point when the c-th local component experiences the m-th anomalous mode. Arranging all anomalous response atoms forms a physical anomalous response dictionary:

[0353]

[0354] Under this dictionary representation, the observed differential response of a two-dimensional Helmholtz wavefield system can be expressed as:

[0355]

[0356] Where x represents the anomalous sparsity coefficient in the atomic space of the component pattern. This represents structural residuals or model mismatch residuals outside the dictionary subspace. This indicates measurement noise.

[0357]

[0358]

[0359] in, This represents the set of atomic patterns of the exception components that are actually activated.

[0360] In this embodiment, a three-layer structure is still adopted: a physical dictionary projection layer, a dictionary structure interaction layer, and a physical semantic decoding layer. Its input / output path is as follows:

[0361]

[0362] First, in the physical dictionary projection layer, the two-dimensional Helmholtz physical anomaly response dictionary is normalized to obtain... And calculate the relevant physical semantic activation:

[0363]

[0364] in, This represents the signed similarity between the observed differential response and the mode atoms of each Helmholtz component. This quantity is not the final anomalous sparsity coefficient, but rather a physical dictionary projection feature. Reorganizing the projection features according to the component and mode dimensions yields the component-mode block features:

[0365]

[0366] Then, in the dictionary structure interaction layer, for Perform learnable signed soft thresholding transformation:

[0367]

[0368] in, This represents the physicosine activations after dictionary-structured interactions. This layer is used to suppress weak cross-atom responses and preserve physicosine activations with anomalous structural meaning.

[0369] In the physical semantic decoding layer, the output component-level support probability and component pattern atomic-level support probability are:

[0370] ,

[0371] in, This represents the support probability of the c-th local component being abnormal. This represents the support probability of the c-th local component exhibiting the m-th abnormal mode.

[0372] Furthermore, it effectively supports atomic-level effective gating values ​​in the gated computing component pattern through hierarchical support:

[0373]

[0374] Physical semantic decoding layer outputs ungated continuous coefficient candidates And by effectively supporting gating at different levels, the final effective sparsity coefficients are obtained:

[0375]

[0376] Reconstructing the dictionary subspace difference response of a two-dimensional Helmholtz wavefield system based on effective sparse coefficients:

[0377]

[0378] in, This represents the differential response reconstruction result interpreted from the physical anomaly response dictionary. This result is not an absolute wavefield response, nor does it explicitly predict residuals or noise terms.

[0379] During the training phase, supervisory labels are constructed based on the true anomaly components and true anomaly patterns of the two-dimensional Helmholtz data samples. For the i-th sample, its true anomaly set is represented as:

[0380]

[0381] in, This represents the number of true anomalies in the i-th sample. Component-level labels are constructed from the set of true anomalies.

[0382]

[0383] And construct atomic tags for the component pattern:

[0384]

[0385] Training objectives may include semantic support consistency, physical coefficient reconstruction consistency, and hierarchical sparse priors. The total loss can be expressed as:

[0386]

[0387] in, Used to constrain component-level support and atomic-level support for component patterns. Used to constrain the consistency of effective sparse coefficients and dictionary subspace reconstruction response. Sparse priors used to express the atomic activation of minority components and minority component modes.

[0388] During the inference phase, the predicted anomaly support set is obtained based on the atomic support probability of the component pattern or the effective gating value at the hierarchical level:

[0389]

[0390] in, This indicates the threshold for anomaly detection. The final output is a set of anomaly diagnoses.

[0391]

[0392] in, This indicates the j-th predicted anomaly component. This represents the j-th predicted abnormal pattern. This represents the support set for predicting anomalies. This represents the confidence level of the j-th anomaly entry. This indicates the predicted number of anomalies.

[0393] In this embodiment, to verify the technical effectiveness of the physical embedded sparse identification architecture system of the present invention, the method of the present invention is compared with several typical comparative models under the same dataset partitioning, the same training conditions, and the same evaluation process. The comparative models include convolutional neural network models, multilayer perceptron models, Transformer models, models containing only physical dictionary projection layers, models combining physical dictionary projection with convolutional neural networks, and models combining physical dictionary projection with Transformer.

[0394] This embodiment evaluates the identification performance of the present invention from four aspects: abnormal component identification, abnormal pattern identification, abnormal support set identification, and differential response reconstruction. Specifically, abnormal component identification evaluates whether the model can correctly locate the physical component experiencing the abnormality; abnormal pattern identification evaluates whether the model can correctly distinguish different types of candidate abnormal patterns; abnormal support set identification evaluates whether the model can simultaneously identify the dictionary atom set formed by the abnormal component and its corresponding abnormal pattern; and the normalized mean square error of the differential response reconstruction evaluates the interpretability of the physical semantic coefficients output by the model for the observed differential response, expressed as:

[0395]

[0396] Specifically, component recognition performance is evaluated using the component recognition F1 score, and the results are as follows: Figure 8 As shown; pattern recognition performance is evaluated using the pattern recognition F1 score, and the results are as follows. Figure 9 As shown; the performance of anomaly support set identification is evaluated using the support set F1 score, and the results are as follows. Figure 10 As shown; the performance of the differential response reconstruction is evaluated using the normalized mean square error, and the results are as follows. Figure 11 As shown in the figure. Among them, the larger the F1 value, the better the recognition performance, and the smaller the normalized mean square error, the lower the response reconstruction error.

[0397] Depend on Figures 8 to 10 As can be seen, compared to the comparative model that does not introduce physical dictionary projection or does not model dictionary structure relationships, the method of this invention achieves higher F1 scores in component identification, pattern recognition, and anomaly support set identification tasks. This indicates that the present invention can more accurately locate anomalous components, identify anomalous patterns, and maintain the consistency of component-pattern combination relationships. Figure 11 As can be seen, the method of this invention has a lower normalized mean square error in the differential response reconstruction task, indicating that its output physical semantic coefficients not only have better diagnostic and discriminative capabilities, but also can more effectively reconstruct the observed differential physical response. Therefore, compared with existing pure data-driven models and models that only introduce physical projection but lack structural interaction modeling, this invention can improve the overall performance of abnormal component localization, abnormal pattern discrimination, abnormal support set identification, and differential response reconstruction while maintaining physical interpretability.

[0398] In addition to the quantitative evaluation indicators mentioned above, this embodiment further provides the component confidence distribution results under different numbers of abnormal components, such as... Figure 12 As shown, the horizontal axis represents the component number, the vertical axis represents the component confidence score or component anomaly score, the bar chart represents the anomaly support level for each component, and the circle markers indicate components judged as anomalies. It can be seen that in the presence of 1, 2, and 3 anomaly components, this invention can form a relatively obvious confidence peak at the location of the anomaly component, while the confidence scores for non-anomaly components are generally lower. This indicates that this invention can maintain good anomaly component discrimination ability and output interpretability under different anomaly numbers. Here, the component confidence score is the component anomaly support score output by the model, and is not limited to statistical probabilities after probability calibration.

[0399] In summary, the architecture of this invention is designed for modular physical systems composed of multiple physical components and various candidate anomaly patterns. First, it obtains the observational difference response between the observed response of the healthy state and the observed response of the state to be identified. Then, based on the physical model, it generates anomaly response atoms for each component pattern combination and constructs a physical anomaly response dictionary. Next, it maps the observational difference response to the component pattern atom space through a physical dictionary projection layer, obtaining projection features with physical semantics. A dictionary structure interaction layer models the structural relationships between component pattern atoms, generating structured physical semantic activation. Finally, a physical semantic decoding layer outputs anomaly components, anomaly patterns, anomaly support sets, and anomaly scores, and can generate effective sparse coefficients and dictionary subspace difference response reconstruction results. This invention embeds the physical anomaly response dictionary into the anomaly identification process, ensuring consistency between anomaly semantic output, component pattern support, effective sparse coefficients, and physical response reconstruction, thereby improving the physical interpretability of anomaly identification, the stability of support set recognition, and adaptability to multiple anomaly scenarios.

[0400] Those skilled in the art should understand that the above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit the scope of protection of the present invention. Without departing from the spirit and principles of the present invention, those skilled in the art can make various modifications, alterations, equivalent substitutions, or improvements to the above embodiments, all of which should fall within the scope of protection defined by the claims of the present invention.

Claims

1. A physically embedded sparse identification architecture system, characterized in that, This sparse identification architecture system is used to identify anomalous components and anomalous patterns in a modular physical system, which includes multiple physical components, each of which corresponds to one or more candidate anomalous patterns. The sparse identification architecture system includes: The differential response acquisition unit is used to obtain the observation differential response based on the reference state observation response and the state observation response to be identified of the modular physical system. The physical anomaly response dictionary unit is used to generate anomaly response atoms for component pattern combinations consisting of physical components and candidate anomaly patterns, and multiple anomaly response atoms constitute the physical anomaly response dictionary. And a sparse identification architecture unit, which includes a physical dictionary projection layer, a dictionary structure interaction layer, and a physical semantic decoding layer connected in sequence: The physical dictionary projection layer is used to map the observed differential response to the component pattern atomic space based on the physical anomaly response dictionary, so as to obtain the projection features with component pattern physical semantics. The dictionary structure interaction layer is used to interactively model the structural relationships between component pattern atoms based on the projection features, so as to obtain a structured physical semantic representation. The physical semantic decoding layer is used to output anomaly diagnosis results based on the projection features and the structured physical semantic representation; The anomaly diagnosis result includes at least anomaly components and anomaly patterns, and the physical semantic decoding layer is also used to output effective sparse coefficients corresponding to the atomic space of the component patterns. The effective sparse coefficients are used to reconstruct the observed differential response in the dictionary subspace corresponding to the physical anomaly response dictionary.

2. The physically embedded sparse identification architecture system according to claim 1, characterized in that, The physical dictionary projection layer calculates one or more of the following: similarity, correlation, projection coefficient, regularized projection feature, or amplitude semantic feature between the observed differential response and the anomalous response atom in the physical anomaly response dictionary, and organizes them into component pattern block features.

3. The physically embedded sparse identification architecture system according to claim 1, characterized in that, The dictionary structure interaction layer performs structured transformation on the projected features through trainable mappings, which include threshold mapping, multilayer perceptron, convolutional mapping, attention mechanism, graph neural network, gating network or a combination thereof. The structural relationships modeled by the dictionary structure interaction layer include at least one of the following: the relationship between different abnormal modes within the same component, the relationship between the same or related abnormal modes between different components, the atomic response correlation determined by the coherence of the physical abnormal response dictionary, the spatial proximity relationship determined by the geometric position of the component, the topological relationship determined by the physical connection relationship of the component, and the co-occurrence relationship or mutual exclusion relationship obtained by statistical analysis of training data.

4. The physically embedded sparse identification architecture system according to claim 1, characterized in that, The physical semantic decoding layer includes a component-level decoding subunit, a component-mode atomic-level decoding subunit, a coefficient candidate output subunit, and an effective coefficient generation subunit; The component-level decoding subunit is used to output component-level exception support information; The component mode atomic-level decoding subunit is used to output component mode atomic-level exception support information; The coefficient candidate output subunit is used to output continuous coefficient candidates corresponding to the component mode atoms; The effective coefficient generation subunit is used to gate or weight the continuous coefficient candidates based on the component-level anomaly support information, the component pattern atomic-level anomaly support information, the structured physical semantic representation, or a combination thereof, to obtain the effective sparse coefficients.

5. The physically embedded sparse identification architecture system according to claim 4, characterized in that, The effective sparsity coefficients are jointly determined by the hierarchical effective gating value and the candidate continuous coefficients; the hierarchical effective gating value is obtained by combining component-level anomaly support information and component pattern atomic-level anomaly support information.

6. The physically embedded sparse identification architecture system according to claim 1, characterized in that, The anomaly diagnosis results also include one or more of the following: anomaly support set, number of anomalies, anomaly score, component confidence, component pattern atomic confidence, effective sparsity coefficient, dictionary subspace difference response reconstruction results, or anomaly severity.

7. The physically embedded sparse identification architecture system according to claim 1, characterized in that, The architecture employs a training objective function during the training phase that includes at least one of the following objectives: component-level semantic support objective, component pattern atomic-level semantic support objective, effective sparse coefficient supervision objective, differential response reconstruction objective, hierarchical sparse prior objective, support probability and continuous coefficient consistency objective, confidence calibration objective, class imbalance weighting objective, or parameter regularization objective.

8. A method for identifying anomalies in a physical system, characterized in that, This method, based on the physically embedded sparse identification architecture system described in any one of claims 1 to 7, includes the following steps: Obtain the reference state observation response and the state observation response to be identified of the modular physical system, and construct the observation differential response; Component pattern combinations are constructed based on physical components and candidate anomaly patterns, and anomaly response atoms are generated for each component pattern combination to form a physical anomaly response dictionary; Based on the physical anomaly response dictionary, the observed differential response is mapped to the component pattern atomic space to obtain projection features with component pattern physical semantics. The projected features are subjected to dictionary structure interaction modeling to obtain a structured physical semantic representation; Based on the projection features and the structured physical semantic representation, physical semantic decoding is performed to obtain anomaly diagnosis results; Generate effective sparse coefficients corresponding to the component pattern atomic space, and reconstruct the response components of the observed differential response in the dictionary subspace based on the effective sparse coefficients and the physical anomaly response dictionary.

9. A computer device, characterized in that, It includes a processor and a memory, wherein the memory stores a computer program, which, when executed by the processor, implements the anomaly identification method of claim 8.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program that, when executed by a processor, implements the anomaly identification method of claim 8.

Images