Fault diagnosis method for aeronautical gear pump based on ground test system

Abstract

This invention relates to a fault diagnosis method for aircraft gear pumps based on a ground testing system, comprising the following steps: simulating aircraft flight attitude using a 6-DOF platform, and simulating different gear pump faults by replacing components of the gear pump with faulty parts; adding vibration sensors to the existing discharge pressure sensors in the airborne oil circuit system during ground testing to collect vibration acceleration signals and pressure pulsation signals; employing a cross-attention mechanism to perform soft projection between modes, extracting the shared consistency between the vibration and pressure domains; removing redundant information from the original features, and then introducing a variational information bottleneck to filter out sample-specific noise from the residuals, thereby obtaining purified unique features; applying Hadamard product to the purified unique features to explicitly construct high-order collaborative relationships; concatenating the decoupled semantic features and inputting them into a classifier for fault diagnosis; overcoming the dependence on dataset-level statistics in traditional methods.

Description

Technical Field

[0001] This invention relates to the field of fault diagnosis of aircraft gear oil pumps, and more particularly to a fault diagnosis method for aircraft gear pumps based on a ground testing system. Background Technology

[0002] The safety of aviation equipment, especially the safety of external gear pumps in aircraft engines, largely depends on accurate health monitoring capabilities. With the development of the Industrial Internet of Things, data-driven multimodal diagnostics has become a dominant mode, effectively replacing traditional single-sensor analysis.

[0003] In complex industrial environments, this approach is fundamentally based on the premise that heterogeneous modalities, such as vibration and pressure signals, capture mechanical dynamics from different physical perspectives. These modalities possess complementary advantages, enabling diagnostic systems to overcome the observational limitations inherent in any single sensor. Despite this theoretical potential, mainstream deep learning frameworks often rely on black-box fusion strategies, in which features are merely spliced ​​or weighted without recognizing their inherent information structure. However, reliable decision-making requires more than just simple data aggregation; it demands a clear understanding of diagnostic evidence. Specifically, it requires distinguishing between shared consistent patterns (redundancy) among sensors, fault characteristics specific to a single modality (uniqueness), and complex higher-order interactions (synergy) that only emerge when multiple modalities are jointly observed. To systematically quantify these semantic components, Partial Information Decomposition (PID) theory provides a principled mathematical framework. By decomposing total mutual information into these distinct atoms, PID offers a theoretical foundation for moving beyond black-box fusion and interpreting the complex interactions of multimodal features.

[0004] However, traditional PID estimators are fundamentally rooted in distributional statistical analysis, making them structurally incompatible with real-time, single-sample inference in constrained industrial environments, where obtaining the massive fault datasets required for such computations is practically impossible. Furthermore, mainstream methods treat synergy as a mathematical residual (total information minus redundancy and uniqueness), a rigid subtraction method heavily reliant on sample size. In practical computation, this often leads to severe numerical instability and illogical negative information values, fundamentally failing to capture the explicit high-order interactions required for diagnosing complex, compound faults. Therefore, a shift from statistical estimation to geometric construction is urgently needed to bridge the gap between theoretical rigor and industrial applicability. Moreover, a gap remains between theoretical ideals and industrial reality regarding noise interference. Standard PID theory strictly defines components based on their contribution to the target task, theoretically assigning zero mutual information to task-irrelevant noise. However, real-world sensor data is inevitably entangled with complex background interference. Without a dedicated cleanup mechanism, this entanglement prevents standard estimators from effectively separating valid diagnostic evidence from non-contributing noise, thus severely limiting the practical deployment of the PID framework in safety-critical scenarios. Summary of the Invention

[0005] The purpose of this invention is to provide a fault diagnosis method for aircraft gear pumps based on a ground testing system. To address the shortcomings of existing PID frameworks in aircraft gear pump fault diagnosis applications, a variational geometric PID (VG-PID) framework is proposed to overcome these shortcomings.

[0006] To achieve the above objectives, the technical solution adopted by this invention is: a fault diagnosis method for aviation gear oil pumps based on a ground testing system, comprising the following steps:

[0007] Step 1: During ground testing, a 6-DOF platform is used to simulate the aircraft's flight attitude. Different gear oil pump failures are simulated by replacing the components of the gear oil pump with faulty parts.

[0008] Step 2: Based on the existing oil pressure sensor in the airborne oil circuit system, a vibration sensor is added during ground testing to collect vibration acceleration signals and pressure pulsation signals;

[0009] Step 3: Redundancy extraction. A cross-attention mechanism is used to perform soft projection between modes to extract the shared consistency between the vibration domain and the pressure domain.

[0010] Step 4: Uniqueness purification, geometric subtraction, removes the extracted redundant information from the original features. Then, a variational information bottleneck is introduced to filter out sample-specific noise from the residuals, thereby obtaining the purified unique features.

[0011] Step 5: Collaborative Construction. By applying the Hadamard product to the purified unique features, higher-order collaborative relationships are clearly constructed.

[0012] Step six: Concatenate the decoupled semantic features and input them into the classifier for fault diagnosis.

[0013] Preferably, steps three to five employ partial information decomposition and variational information bottleneck. The partial information decomposition framework formally decomposes the total mutual information into three non-negative and disjoint components: redundancy, uniqueness, and synergy. The decomposition relationship is expressed as follows:

[0014]

[0015] in:

[0016] Redundancy: Overlapping information shared by two sources;

[0017] Uniqueness: Exclusive information provided by only one source;

[0018] Synergy: Information that only appears when X1 and X2 are observed simultaneously; it captures the complex cross-modal correlations necessary to identify compound faults.

[0019] Preferably, partial information decomposition approximates the 1B target by introducing a variational lower bound;

[0020] First, for the prediction item Using a parameterized decoder To approximate the difficult-to-handle posterior ;

[0021]

[0022] in, For expectation operators;

[0023] Secondly, regarding the compression item Minimize encoder distribution With fixed priors (usually a standard Gaussian distribution) The Kullback-Leibler (KL) divergence between ).

[0024]

[0025] Combining these terms, the minimization objective of VIB becomes:

[0026] .

[0027] Preferably, steps three through six utilize the variational geometry PID (VG-PID) framework for multimodal fault diagnosis of aviation fuel pumps, where vibration and pressure signals are encoded into latent feature vectors. and First, the cross-attention mechanism extracts redundant features. Next, by subtracting redundancy from the original representation, the unique features of a specific mode are obtained. and Then, through the high-order interactions of these unique features, collaborative features are explicitly constructed. Throughout the process, this module is used to filter out noise. Finally, all decoupled features are spliced ​​together and input into the classification head for fault diagnosis.

[0028] Preferably, PID theory is instantiated as a geometric neural framework by mapping information-theoretic concepts to vector geometry; it is assumed that if two eigenvectors are geometrically orthogonal, their mutual information is minimized; therefore, the statistical independence required by PID is reformulated as a vector orthogonality constraint:

[0029]

[0030] Based on this premise, the latent feature space is constructed as an orthogonal combination of semantic subspaces:

[0031]

[0032] in , and The eigenvectors corresponding to the theoretical PID components, and This represents the null space containing task-independent noise; This is a vector concatenation operator; this restatement decomposes abstract information into a verifiable feature projection task.

[0033] Within this geometric framework, the unique spaces of different modalities are orthogonal to each other, while task-irrelevant noise components are effectively incorporated into the null space of the diagnostic task; specifically, the latent feature decomposition of each modality is as follows:

[0034]

[0035] in, This represents the shared semantic information between the two modalities. and These respectively represent the unique semantics of vibration signals and pressure signals, while and This represents the specific noise component inherent in each sensor channel.

[0036] Preferred, redundant (denoted as This refers to consistent semantic information shared across heterogeneous physical channels. The design cascade process captures redundancy, including cross-modal soft projection and VIB-based redundancy cleanup, enabling... The vibration and pressure coding features are represented respectively. Since cross-modal fault features often exhibit local time shifts or nonlinear distortions, rigid element-wise alignment is not optimal. Instead, this application employs a cross-attention mechanism to perform soft projection, treating the feature space of one mode as a basis set to reconstruct another mode, thereby effectively aligning their geometric manifolds. Specifically, vibration features... Projected onto pressure base Above, through Activation function to obtain an aligned representation. :

[0037]

[0038] By utilizing symmetry, we obtain the back projection. To form a unified redundant candidate that is not biased towards any single modality, an average fusion strategy is used:

[0039]

[0040] Applying VIB constraints, the final cleanup redundancy characteristics It is sampled from the random distribution predicted by the VIB encoder:

[0041]

[0042] in Indicates VIB encoder, and These represent the mean and variance, respectively. This represents standard Gaussian white noise. This is the dot product operator;

[0043] In the reasoning process, the deterministic mean It is used as a feature vector.

[0044] Preferably, a "subtraction-purification" strategy is used to extract uniqueness and obtain global redundancy. Then, it is first projected back onto the local feature manifold to separate the patterns of specific modes. The geometric residuals are calculated by removing the projection of the common eigenvectors.

[0045]

[0046] in It is a learnable linear mapping used for dimension alignment;

[0047] Original residual Inevitably contains sensor-specific noise. To extract purely unique semantics, the VIB module is used again, similar to redundant branches, for unique features. and Sampling using reparameterization techniques:

[0048]

[0049] Extracted and Only information-rich, mode-specific fault characteristics are captured, and these are redundant. Strictly orthogonal.

[0050] Preferably, a constructive geometric approach is used to resolve the ambiguity of the subtraction residual; it is assumed that the true synergistic effect must be explicitly modeled as a nonlinear interaction between refined unique features, and the total feature interaction space is rigorously decomposed, assuming that the latent feature space of each mode is geometrically decomposed into... The total interaction space Q spanned by the Cartesian product of the two modes expands as follows:

[0051]

[0052] in This represents the tensor product operator. This is a placeholder;

[0053] Based on information theory principles, the true source of synergy is isolated by sequentially eliminating invalid candidates.

[0054] Group A exclusion: Involved The term represents random interference, since noise is statistically independent of fault labels. Their mutual information satisfies These interactions do not provide effective prediction gain;

[0055] Group B is excluded because it contains self-redundant and cross-terms, interactions inherent in the autocorrelation of a single mode. Since synergy is defined as information unavailable to any single mode, and Group B only obtains information through... or It can be fully acquired, therefore it cannot constitute synergy;

[0056] Therefore, after excluding groups A and B, group C becomes the only valid candidate group; this mathematical proof confirms that the synergistic effect must originate entirely from cross-modal interactions with unique features.

[0057] Based on the above derivation, the obtained purified unique characteristics are utilized. and This concretizes the concept of collaboration; and models higher-order interactions through the Hadamard product:

[0058]

[0059] This element-wise operation acts as a logic gate, activating only when a specific pattern appears simultaneously in both channels. However, direct multiplication can amplify random noise. To suppress these artifacts, VIB constraints are applied to cleanse the original interaction:

[0060]

[0061] A stable collaborative structure was generated. .

[0062] Preferably, task prediction, variational regularization, and geometric constraints are integrated into a comprehensive objective function, with a total loss of The formula is as follows:

[0063]

[0064] in , and It is a hyperparameter that balances the contributions of various factors;

[0065] Global task prediction: Instead of training discrete classifiers for each subspace, it focuses on the final fused representation. A uniform cross-entropy loss is used, which implicitly guides all upstream encoders to extract discriminative features through backpropagation:

[0066]

[0067] in, Represents the total number of samples. Indicates sample The true label, Indicates sample The predicted probability.

[0068] Unified variational regularization: To implement noise filtering across all semantic branches and minimize latent features. The posterior distribution and the standard Gaussian prior The sum of KL divergences between them:

[0069]

[0070] Apply a uniform weighting factor to all subspaces. ;

[0071] Geometric constraints: To strictly satisfy the geometric definition of PID, two structural penalties are imposed. First, Minimize the absolute cosine similarity between all pairs of different feature subspaces to prevent information leakage; secondly, Minimize the Euclidean distance between bidirectional projections in redundant branches to ensure semantic alignment:

[0072] .

[0073] The technical effects of this invention are as follows:

[0074] 1. A general variational geometric PID framework is proposed, which reformulates PID theory from a statistical estimation problem into a geometric construction task. By creating orthogonal subspaces for redundancy, uniqueness, and synergy, this framework achieves accurate sample-level inference, overcoming the inherent dependence on dataset-level statistics in traditional methods.

[0075] 2. To address the ambiguity of collaborative information, a mechanism is introduced to explicitly construct collaborativeity through high-order interactions of refined unique features. This method transforms collaborativeity from fuzzy mathematical residuals into perceptible diagnostic features, effectively capturing complex cross-modal correlations crucial for identifying compound faults.

[0076] 3. By implementing variational information bottleneck constraints in all semantic subspaces, a rigorous denoising strategy is designed. This branching adjustment suppresses noise while protecting weak but critical fault features, thus maintaining the semantic purity of the separated subspaces. Attached Figure Description

[0077] Figure 1 This diagram illustrates the signal processing and prediction methods for fault diagnosis of aircraft gear oil pumps based on a ground-based testing system. Detailed Implementation

[0078] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention. Furthermore, the technical features involved in the various embodiments of this invention described below can be combined with each other as long as they do not conflict with each other.

[0079] First, ground tests were conducted. During the ground tests, a 6-DOF platform was used to simulate the aircraft's flight attitude. Different failures of the gear oil pump were simulated by replacing the components of the gear oil pump with faulty parts.

[0080] Based on the existing oil pressure sensor in the airborne oil circuit system, a vibration sensor was added during ground testing to collect vibration acceleration signals and pressure pulsation signals.

[0081] The acquired diagnostic and pressure signals are processed as follows: a cross-attention mechanism is used to perform soft projection between modes to extract the shared consistency between the vibration and pressure domains;

[0082] Redundant information is removed from the original features. Then, a variational information bottleneck is introduced to filter out sample-specific noise from the residuals, thereby obtaining purified unique features. By applying the Hadamard product to the purified unique features, a high-order collaborative relationship is clearly constructed. The decoupled semantic features are concatenated and input into a classifier for fault diagnosis.

[0083] Specifically as follows:

[0084] Implicit decoupling in representation learning

[0085] From an information theory perspective, the predictive power of multi-source data is not a single, unified whole, but rather distributed across redundant, unique, and collaborative components. While standard deep learning models typically integrate these factors into a holistic representation, the field is increasingly shifting towards structured decoupling. The goal is to decompose the latent space into distinct semantic subspaces, adhering to the principle of partial information decomposition, to improve model interpretability and robustness to complex industrial noise.

[0086] While most existing frameworks have successfully separated redundant commonalities from unique specificities, the binary decomposition strategy employed ignores the potential direction of collaborative information. The greatest advantage of multimodal learning lies precisely in the simultaneous utilization of different modalities, which is a unique and significant feature of multimodal data. Current architectures limit feature representation to shared and private components, failing to fully utilize the intrinsic value of synchronous input from multiple sensors, thus leaving a fundamental gap in comprehensive multimodal representation learning.

[0087] Collaborative modeling and multimodal interaction

[0088] In multi-sensor fault diagnosis, the fusion of heterogeneous modes aims not only to accumulate information but also to reveal synergistic information—an emerging form of diagnostic evidence that can only be obtained by jointly observing multiple sensors. Theoretically, synergy differs from redundancy and uniqueness; it captures the complex nonlinear coupling mechanisms inherent in electromechanical systems, such as the interaction between fluid pressure dynamics and structural vibration. Therefore, modeling higher-order correlations is the key difference between advanced fusion and simple aggregation.

[0089] Recent literature has explored various mechanisms for capturing these cross-modal dependencies, utilizing hypergraph neural networks to construct hyperedges connecting multiple sensor nodes, implicitly modeling higher-order correlations beyond pairwise relationships. Similarly, contrastive learning is combined with hypergraph topology to align cross-modal features and enhance the model's sensitivity to cross-domain dependencies. To address the data scarcity problem, a collaborative feature fusion strategy based on deep convolutional generative adversarial networks is proposed, which improves the diversity of fused representations by synthesizing complementary pseudo-samples. Combining causal discovery with temporal graph convolutional networks shows great potential in modeling explicit directional dependencies between multivariate sensor signals. From a rigorous information theory perspective, a framework for de-entanglement of multimodal representation information bottlenecks is proposed. This method attempts to explicitly decompose representations into redundant, unique, and collaborative components by imposing specific mutual information constraints, marking a significant step forward in structural interpretability.

[0090] However, constructing synergistic effects remains challenging due to limitations in computational methods. Standard estimators typically employ a subtractive definition, calculating synergistic effects as the remainder after subtracting redundancy and uniqueness from the total information. This algebraic rigidity easily leads to numerical instability and often produces negative information values, contradicting physical interpretations and hindering stable gradient optimization. Furthermore, existing PID (Partial Information Decomposition)-based frameworks still suffer from structural limitations. Due to the lack of explicit geometric orthogonality, these methods often unintentionally confuse emerging synergistic patterns with shared redundant features. This confusion violates the strict condition that synergistic effects must originate entirely from a joint modal distribution, thus affecting the purity of diagnostic evidence.

[0091] Robustness and variational purification

[0092] In aviation safety-critical applications, such as health monitoring of external gear pumps, the operating environment is extremely complex. Unlike controlled laboratory environments, airborne equipment is subjected to severe multi-source heterogeneous interference, including high-frequency mechanical shocks and electromagnetic coupling. From an information theory perspective, this environmental noise poses a fundamental threat to the effectiveness of diagnostic models. The standard definition of mutual information is based on the assumption that input entropy contains meaningful changes related to the target state. However, in such harsh industrial scenarios, a large portion of the input entropy is dominated by task-irrelevant noise. If left unmitigated, this non-contributing variance propagates during PID decomposition, causing noise to be incorrectly attributed as "unique" information or diluting the synergistic effect of the estimation. Therefore, high-noise environments undermine the theoretical foundation of standard PID estimators, necessitating explicit denoising mechanisms.

[0093] To counteract the distortions caused by noise, recent research has incorporated uncertainty estimation and information-theoretic regularization into fusion architectures. A framework embracing arbitrary uncertainty (EAU) has been proposed, explicitly quantifying the arbitrary uncertainty inherent in unimodal data and weighting it based on feature reliability. EAU effectively suppresses the influence of noisy modalities during fusion. A multimodal information bottleneck (MIB) has been introduced, extending the fundamental deep variational information bottleneck principle to multimodal learning. Following the variational information bottleneck paradigm, MIB aims to learn a minimum sufficient representation by maximizing the mutual information between latent features and the target label while minimizing the mutual information between features and the original input. This variational objective acts as an information filter, theoretically causing the model to discard task-irrelevant noise while retaining predictive semantic information.

[0094] While MIB and its variants are theoretically sound, directly applying the information bottleneck principle to the entire feature space produces a key side effect: indiscriminate over-compression. Current MIB-based methods typically impose global compression constraints on the entire representation vector. Multimodal representations inherently contain complex features that contribute differently to the diagnostic target, ranging from dominant strong information to subtle weak information. When imposing global compression constraints, the optimization process in the early stages of training often overemphasizes strong information to minimize prediction loss. Consequently, weaker features are overcompressed and incorrectly discarded as task-irrelevant noise. This lack of subspace-specific tuning fails to protect these fragile weak features, leading to the loss of fine-grained fault features. Ultimately, this global approach makes the model robust to significant noise but extremely insensitive to subtle early anomalies. To address this critical limitation, a rigorous denoising strategy is designed, achieved by independently imposing variational information bottleneck constraints.

[0095] This application is detailed as follows:

[0096] Prerequisites

[0097] Before delving into the proposed framework, we first establish the theoretical foundations of partial information decomposition and variational information bottleneck, which are the mathematical cornerstones of feature decoupling and noise reduction, respectively.

[0098] Partial information decomposition

[0099] Standard information theory relies on mutual information (MI) to quantify the dependency between source variable X and target variable Y. Mathematically, mutual information is defined as the reduction in uncertainty of Y given X:

[0100]

[0101] in This represents Shannon entropy. However, in multi-source scenarios, standard mutual information only measures the total information jointly provided by the sources. It cannot distinguish how this information is distributed—specifically, it cannot distinguish whether X1 and X2 provide the same information or complementary details.

[0102] To resolve this ambiguity, the partial information decomposition framework formally decomposes the total mutual information into three non-negative and disjoint components: redundancy, uniqueness, and synergy. The decomposition relationship is expressed as:

[0103]

[0104] in:

[0105] Redundancy: Overlapping information shared by two sources.

[0106] Uniqueness: Exclusive information provided by only one source.

[0107] Synergy: Information that only appears when X1 and X2 are observed simultaneously; it captures the complex cross-modal correlations necessary to identify compound faults.

[0108] Relevance to this application: In the VG-PID framework proposed in this application, this theoretical decomposition is used to guide the geometric construction of the feature subspace, ensuring that the extracted features have structured semantics rather than being inherently entangled.

[0109] Variational information bottleneck

[0110] While PID provides a method for feature separation, it assumes that the variables are clean. In industrial scenarios, the input X is subject to noise. Interference (i.e.) Simply maximizing mutual information This leads to the encoding of interference noise. To solve this problem, the variational information bottleneck (IB) principle is adopted.

[0111] The goal of IB is to learn a compressed representation Z that best expresses the objective Y while maximally compressing the input X. The objective function is as follows:

[0112]

[0113] in It is a Lagrange multiplier that controls the trade-off between predictability and compression.

[0114] Since it is computationally difficult to calculate mutual information in high-dimensional space, this application adopts the VIB method, which approximates the 1B objective by introducing variational lower bounds.

[0115] First, for the prediction item Using a parameterized decoder To approximate the difficult-to-handle posterior ;

[0116]

[0117] Secondly, regarding the compression item Minimize encoder distribution With fixed priors (usually a standard Gaussian distribution) The Kullback-Leibler (KL) divergence between ).

[0118]

[0119] Combining these terms, the minimization objective of VIB becomes:

[0120]

[0121] This equation constitutes the core purification mechanism of this application. By specifically applying VIB constraints in the decomposed PID subspace, the network is forced to discard high-frequency, unstructured noise while retaining structured semantic features.

[0122] like Figure 1 As shown, the model framework proposed in this application aims to decouple high-dimensional features in a multi-source noise environment, thereby achieving robust fault diagnosis. This model receives multimodal input, denoted as... , representing vibration and pressure signals respectively; initially, these raw signals are processed by two parallel feature encoders, mapping the high-dimensional time series input to latent feature representations, denoted as . .

[0123] In this process, a variational information bottleneck (VIB) module (purple block) is applied to each semantic branch, such as Figure 1 As shown, the overall architecture of the proposed variational geometry PID (VG-PID) framework for multimodal fault diagnosis of aviation fuel pumps is presented, where vibration and pressure signals are encoded as latent feature vectors. and First, the cross-attention mechanism extracts redundant features. (Red block) Next, by subtracting redundancy from the original representation, we obtain the unique features of the specific modality. and (Blue and orange blocks), and then, through the high-order interactions of these unique features, explicitly construct collaborative features. (Green block); Throughout the process, this module is used to filter out noise. Finally, all decoupled features are concatenated and input into the classification head.

[0124] To achieve reliable sample-level inference, PID theory is instantiated as a geometric neural framework by mapping information-theoretic concepts to vector geometry. It is assumed that if two feature vectors are geometrically orthogonal, their mutual information is minimized; therefore, the statistical independence required by PID is reformulated as a vector orthogonality constraint.

[0125]

[0126] Based on this premise, the latent feature space is constructed as an orthogonal combination of semantic subspaces:

[0127]

[0128] in , and The eigenvectors corresponding to the theoretical PID components, and It represents a null space containing task-independent noise; this restatement decomposes abstract information into a verifiable feature projection task.

[0129] Within this geometric framework, the unique spaces of different modalities are mutually orthogonal, while task-irrelevant noise components are effectively incorporated into the null space of the diagnostic task. Specifically, the latent feature decomposition for each modality is as follows:

[0130]

[0131] in, This represents the shared semantic information between the two modalities. and These respectively represent the unique semantics of vibration signals and pressure signals, while and This represents the specific noise component inherent in each sensor channel.

[0132] The core motivation of this application is to construct multimodal representations as interpretable semantic components for fault diagnosis, while mitigating the impact of multi-source noise. To achieve this goal, the VG-PID module operates as a cascaded pipeline, comprising three consecutive stages;

[0133] Phase 1: Redundancy extraction. A cross-attention mechanism is used to perform soft projection between modes. This step extracts the shared consistency between the vibration domain and the pressure domain.

[0134] The second stage is uniqueness purification. Geometric subtraction then removes the extracted redundant information from the original features. Next, variational information bottleneck (VIB) is introduced to filter out sample-specific noise from the residuals, thereby obtaining the purified unique features.

[0135] The third stage: collaborative construction. By applying the Hadamard product to the purified unique features, high-order collaborative relationships are clearly constructed. This construction method captures emerging information that can only be obtained through multimodal joint observation, which is different from the information obtained from any single sensor perspective.

[0136] Then, the decoupled semantic features are concatenated and input into a classifier for fault diagnosis, ensuring that decisions are based only on clear, task-related information.

[0137] Redundancy extraction, soft projection, and variational purification

[0138] Redundancy (denoted as) This refers to consistent semantic information shared in heterogeneous physical channels. In order to robustly capture this intersection while adapting to the inherent temporal asynchrony of multi-sensor data, this application designs a cascaded process consisting of geometric soft projection and VIB-based purification.

[0139] Cross-modal soft projection

[0140] make The vibration and pressure coding features are represented respectively. Since cross-modal fault features often exhibit local time shifts or nonlinear distortions, rigid element-wise alignment is not optimal. Instead, this application employs a cross-attention mechanism to perform soft projection, treating the feature space of one mode as a basis set to reconstruct another mode, thereby effectively aligning their geometric manifolds. Specifically, vibration features... Projected onto pressure base Above, to obtain an aligned representation :

[0141]

[0142] By utilizing symmetry, we obtain the back projection. To form a unified redundant candidate that is not biased towards any single mode, this application adopts an average fusion strategy:

[0143]

[0144] VIB-based redundancy cleanup

[0145] Although While capturing common information, it may still retain shared noise. To extract rigorous diagnostic-related redundancy, this application applies VIB constraints, ultimately cleansing up the redundancy features. It is sampled from the random distribution predicted by the VIB encoder:

[0146]

[0147] In the reasoning process, the deterministic mean It is used as a feature vector.

[0148] Uniqueness extraction, geometric subtraction and purification

[0149] Uniqueness represents modality-specific information unique to a particular modality, and its geometric definition is the orthogonal components of the redundant subspace; this information is extracted through a "subtraction-purification" strategy.

[0150] Achieve global redundancy Then, it is first projected back onto the local feature manifold to separate the patterns of specific modes. The geometric residual (candidate for uniqueness) is calculated by removing the projection of the common feature vectors.

[0151]

[0152] in It is a learnable linear mapping used for dimension alignment.

[0153] Purification based on VIB residuals

[0154] Algebraic subtraction alone is insufficient because of the original residual. Inevitably contains sensor-specific noise. To extract purely unique semantics, the VIB module is used again, similar to redundant branches, for unique features. and Sampling using reparameterization techniques:

[0155]

[0156] Therefore, the extracted and Only information-rich, mode-specific fault characteristics are captured, and these are redundant. Strictly orthogonal.

[0157] Collaborative modeling, constructive high-order interaction

[0158] To address the ambiguity of subtractive residuals, a constructive geometric approach is adopted; it is assumed that true synergy must be explicitly modeled as nonlinear interactions between refined, unique features.

[0159] The theoretical basis of constructive collaboration

[0160] To verify this, the total feature interaction space was rigorously decomposed, assuming that the latent feature space of each modality is geometrically decomposed into... The total interaction space Q spanned by the Cartesian product of the two modes expands as follows:

[0161]

[0162] Based on information theory principles, the true source of synergy is isolated by sequentially eliminating invalid candidates.

[0163] Group A exclusion: Involved The term represents random interference, since noise is statistically independent of fault labels. Their mutual information satisfies Therefore, these interactions do not provide effective prediction gain.

[0164] Group B exclusion: This group contains self-redundant and cross-terms. These interactions are inherent in the autocorrelation of a single mode. Since synergy is defined as information that cannot be obtained by any single mode, and Group B only obtains information through... or It can be fully acquired. Therefore, it cannot constitute synergy.

[0165] Therefore, after excluding groups A and B, group C becomes the only valid candidate group; this mathematical proof confirms that the synergistic effect must originate entirely from cross-modal interactions of unique features.

[0166] Explicit crosstalk and variational filtering

[0167] Based on the above derivation, the obtained purified unique characteristics are utilized. and This concretizes the concept of collaboration; and models higher-order interactions through the Hadamard product:

[0168]

[0169] This element-wise operation acts as a logic gate, activating only when a specific pattern appears simultaneously in both channels. However, direct multiplication can amplify random noise. To suppress these artifacts, VIB constraints are applied to cleanse the original interaction:

[0170]

[0171] This resulted in a stable collaborative structure. This is crucial for identifying coupling failures.

[0172] Overall integration strategy optimization objectives

[0173] To train the VG-PID framework end-to-end, task prediction, variational regularization, and geometric constraints are integrated into a comprehensive objective function. Total loss The formula is as follows:

[0174]

[0175] in , and It is a hyperparameter that balances the contributions of various factors;

[0176] 1 / Global Task Prediction: Instead of training discrete classifiers for each subspace, the final fused representations are used. A uniform cross-entropy loss is used; this implicitly guides all upstream encoders to extract discriminative features through backpropagation.

[0177]

[0178] 2 / Unified Variational Regularization: To implement noise filtering on all semantic branches and minimize latent features The posterior distribution and the standard Gaussian prior The sum of KL divergences between them:

[0179]

[0180] Apply uniform weighting coefficients to all subspaces. This strategy prevents parameter explosion and ensures balanced denoising intensity, effectively suppressing unstructured noise without destroying the semantic structure of any particular branch.

[0181] (3) Geometric Constraints: To strictly satisfy the geometric definition of PID, two structural penalties are applied. First, Minimize the absolute cosine similarity between all pairs of different feature subspaces to prevent information leakage; secondly, Minimize the Euclidean distance between bidirectional projections in redundant branches to ensure semantic alignment:

[0182] .

[0183] It should be noted that, in this document, the terms “comprising,” “including,” or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0184] This article uses specific examples to illustrate the principles and implementation methods of the present invention. The above examples are only for the purpose of helping to understand the method and core ideas of the present invention. The above descriptions are only preferred embodiments of the present invention. It should be noted that due to the limitations of textual expression, while there are objectively infinite specific structures, those skilled in the art can make several improvements, modifications, or changes without departing from the principles of the present invention, and can also combine the above technical features in an appropriate manner. These improvements, modifications, changes, or combinations, or the direct application of the inventive concept and technical solution to other situations without modification, should all be considered within the scope of protection of the present invention.

Claims

1. A method for diagnosing faults in an aero gear pump based on a ground test system, characterized in that, Includes the following steps: Step 1: During ground testing, a 6-DOF platform is used to simulate the aircraft's flight attitude. Different gear oil pump failures are simulated by replacing the components of the gear oil pump with faulty parts. Step 2: Based on the existing oil pressure sensor in the airborne oil circuit system, a vibration sensor is added during ground testing to collect vibration acceleration signals and pressure pulsation signals; Step 3: Redundancy extraction. A cross-attention mechanism is used to perform soft projection between modes to extract the shared consistency between the vibration domain and the pressure domain. Step 4: Uniqueness purification, geometric subtraction, removes the extracted redundant information from the original features. Then, a variational information bottleneck is introduced to filter out sample-specific noise from the residuals, thereby obtaining the purified unique features. Step 5: Collaborative Construction. By applying the Hadamard product to the purified unique features, higher-order collaborative relationships are clearly constructed. Step six: Concatenate the decoupled semantic features and input them into the classifier for fault diagnosis.

2. The method of claim 1, wherein, Steps three through five employ partial information decomposition and variational information bottleneck. The partial information decomposition framework formally decomposes the total mutual information into three non-negative and disjoint components: redundancy, uniqueness, and synergy. The decomposition relationship is expressed as: in: Redundancy: Overlapping information shared by two sources; Uniqueness: Exclusive information provided by only one source; Synergy: Information that only appears when X1 and X2 are observed simultaneously; it captures the complex cross-modal correlations necessary to identify compound faults.

3. The method of claim 2, wherein, Partial information decomposition approximates the 1B target by introducing variational lower bounds; First, for the prediction term , a parametric decoder is used to approximate the intractable posterior ; in, For expectation operators; Secondly, regarding the compression item Minimize encoder distribution With fixed priors (usually a standard Gaussian distribution) The Kullback-Leibler (KL) divergence between ). Combining these terms, the minimization objective of VIB becomes: 。 4. The method for diagnosing aircraft gear pump faults based on a ground testing system according to claim 3, characterized in that, Steps three through six utilize the variational geometry PID (VG-PID) framework for multimodal fault diagnosis of aviation fuel pumps, where vibration and pressure signals are encoded into latent feature vectors. and First, the cross-attention mechanism extracts redundant features. Next, by subtracting redundancy from the original representation, the unique features of a specific mode are obtained. and Then, through the high-order interactions of these unique features, collaborative features are explicitly constructed. ; Throughout the process, this module is used to filter out noise. Finally, all decoupled features are spliced ​​together and input into the classification head for fault diagnosis.

5. The method for diagnosing aircraft gear pump faults based on a ground testing system according to claim 4, characterized in that, By mapping information theory concepts to vector geometry, PID theory is instantiated as a geometric neural framework; Assuming that mutual information is minimized if two eigenvectors are geometrically orthogonal, the statistical independence required for PID is reformulated as a vector orthogonality constraint: Based on this premise, the latent feature space is constructed as an orthogonal combination of semantic subspaces: in , and The eigenvectors corresponding to the theoretical PID components, and This represents the null space containing task-independent noise; This is a vector concatenation operator; this restatement decomposes abstract information into a verifiable feature projection task. Within this geometric framework, the unique spaces of different modalities are orthogonal to each other, while task-irrelevant noise components are effectively incorporated into the null space of the diagnostic task; specifically, the latent feature decomposition of each modality is as follows: in, This represents the shared semantic information between the two modalities. and These respectively represent the unique semantics of vibration signals and pressure signals, while and This represents the specific noise component inherent in each sensor channel.

6. The method for diagnosing aircraft gear pump faults based on a ground testing system according to claim 5, characterized in that, Redundancy (denoted as) This refers to consistent semantic information shared across heterogeneous physical channels. The design cascade process captures redundancy, including cross-modal soft projection and VIB-based redundancy cleanup, enabling... The vibration and pressure coding features are represented respectively. Since cross-modal fault features often exhibit local time shifts or nonlinear distortions, rigid element-wise alignment is not optimal. Instead, this application employs a cross-attention mechanism to perform soft projection, treating the feature space of one mode as a basis set to reconstruct another mode, thereby effectively aligning their geometric manifolds. Specifically, vibration features are encoded as follows: Projected onto pressure base Above, through Activation function to obtain an aligned representation. : By utilizing symmetry, we obtain the back projection. To form a unified redundant candidate that is not biased towards any single modality, an average fusion strategy is used: Applying VIB constraints, the final cleanup redundancy characteristics It is sampled from the random distribution predicted by the VIB encoder: in Indicates VIB encoder, and These represent the mean and variance, respectively. This represents standard Gaussian white noise. This is the dot product operator; In the reasoning process, the deterministic mean It is used as a feature vector.

7. The method for diagnosing aircraft gear pump faults based on a ground testing system according to claim 6, characterized in that, Uniqueness is extracted through a "subtraction-purification" strategy to obtain global redundancy. Then, it is first projected back onto the local feature manifold to separate the patterns of specific modes. The geometric residuals are calculated by removing the projection of the common eigenvectors. in It is a learnable linear mapping used for dimension alignment; Original residual Inevitably contains sensor-specific noise. To extract purely unique semantics, the VIB module is used again, similar to redundant branches, for unique features. and Sampling using reparameterization techniques: Extracted and Only information-rich, mode-specific fault characteristics are captured, and these are redundant. Strictly orthogonal.

8. The method for diagnosing aircraft gear pump faults based on a ground testing system according to claim 7, characterized in that, A constructive geometric approach is employed to resolve the ambiguity of subtractive residuals; it is assumed that true synergy must be explicitly modeled as nonlinear interactions between refined, unique features, and the total feature interaction space is rigorously decomposed, with the latent feature space of each mode geometrically decomposed as follows: The total interaction space Q spanned by the Cartesian product of the two modes expands as follows: in This represents the tensor product operator. This is a placeholder; Based on information theory principles, the true source of synergy is isolated by sequentially eliminating invalid candidates. Group A exclusion: Involved The term represents random interference, since noise is statistically independent of fault labels. Their mutual information satisfies These interactions do not provide effective prediction gain; Group B is excluded because it contains self-redundant and cross-terms, interactions inherent in the autocorrelation of a single mode. Since synergy is defined as information unavailable to any single mode, and Group B only obtains information through... or It can be fully acquired, therefore it cannot constitute synergy; Therefore, after excluding groups A and B, group C becomes the only valid candidate group; this mathematical proof confirms that the synergistic effect must originate entirely from cross-modal interactions with unique features. Based on the above derivation, utilizing the obtained purified unique characteristics and This concretizes the concept of collaboration; and models higher-order interactions through the Hadamard product: This element-wise operation acts as a logic gate, activating only when a specific pattern appears simultaneously in both channels. However, direct multiplication can amplify random noise. To suppress these artifacts, VIB constraints are applied to cleanse the original interaction: A stable collaborative structure was generated. .

9. The method for diagnosing aircraft gear pump faults based on a ground testing system according to claim 8, characterized in that, The task prediction, variational regularization, and geometric constraints are integrated into a comprehensive objective function, with a total loss. The formula is as follows: in , and It is a hyperparameter that balances the contributions of various factors; Global task prediction: Instead of training discrete classifiers for each subspace, it focuses on the final fused representation. A uniform cross-entropy loss is used, which implicitly guides all upstream encoders to extract discriminative features through backpropagation: in, Represents the total number of samples. Indicates sample The true label, Indicates sample The predicted probability. Unified variational regularization: To implement noise filtering across all semantic branches and minimize latent features. The posterior distribution and the standard Gaussian prior The sum of KL divergences between them: Apply a uniform weighting factor to all subspaces. ; Geometric constraints: To strictly satisfy the geometric definition of PID, two structural penalties are imposed. First, Minimize the absolute cosine similarity between all pairs of different feature subspaces to prevent information leakage; secondly, Minimize the Euclidean distance between bidirectional projections in redundant branches to ensure semantic alignment: 。

Images