Servo mechanism fault diagnosis method and system based on multi-source representation and mixed experts

Abstract

The application discloses a servo mechanism fault diagnosis method and system based on multi-source representation and mixed experts, and belongs to the technical field of intelligent fault diagnosis and servo system health monitoring. The method first carries out sliding window processing and independent branch coding on multi-source observation signals such as displacement, current, voltage and temperature rise of the servo mechanism, extracts local features and fuses to obtain unified fault representation; then a semantic prototype alignment and local time sequence context enhancement mechanism is introduced, and the representation is mapped to a diagnostic semantic space containing fault types, mechanisms and positions; then a mixed expert diagnosis model is constructed under the explicit constraint of the mechanism, a gating network is used for routing allocation and weighted fusion, and a hierarchical diagnosis result is output; finally, it is organized into a conditional vector and a structured result containing a conclusion and a disposal suggestion is generated. The application effectively improves the hierarchical fault recognition accuracy and the explainability of the diagnosis decision under complex working conditions.

Description

Technical Field

[0001] This invention belongs to the field of intelligent fault diagnosis and servo system health monitoring technology, and in particular to a servo mechanism fault diagnosis method and system based on multi-source characterization and hybrid experts. Background Technology

[0002] Servo mechanisms are widely used in aerospace, high-end manufacturing, and complex equipment control, and their operating status directly affects execution accuracy, dynamic response performance, and system safety. During long-term operation, factors such as servo valve wear, internal leakage, actuator degradation, sensor drift, and parameter deviation can all induce faults, which are manifested through multi-source observations such as displacement, current, voltage, and temperature rise. Due to the significant coupling between the execution, control, and measurement links, different faults may exhibit similar responses in local observations, while the same fault may show different temporal variation characteristics at different stages of evolution. Therefore, servo mechanism fault diagnosis not only requires fault identification but also further characterization of fault type, fault mechanism, and affected location.

[0003] Invention patent CN120030322A discloses a method for spatial hierarchical semantic construction and equipment fault diagnosis. This method does not perform in-depth modeling of multi-scale temporal evolution under complex dynamic working conditions. It only generates virtual samples by combining conditional variational autoencoders and generative adversarial networks with spatial hierarchical semantics. It relies on static semantic alignment and the feature distribution of known fault data. It has weak robustness to transient changes and nonlinear temporal fluctuations during equipment operation and is difficult to adapt to the complex operating environment of industrial sites where high-frequency sudden changes and gradual trends are intertwined. Moreover, the scenarios are mainly limited to static feature classification under generalized zero-shot learning, without considering the dynamic topological interaction requirements between multi-source sensor channels, and the temporal feature extraction dimension is single. Invention patent CN120162681B discloses a method and system for diagnosing complex equipment faults based on dynamic feature modeling. The core technology of this invention is direction-aware convolution combined with multi-scale modulation and trend-guided constraints. This method has weak adaptability to scenarios with unknown fault mechanisms and strong lack of prior knowledge. It only extracts features through data-driven local convolutional receptive fields and temporal difference assistance tasks. It assumes that multi-channel data is complete and ignores the logical causal relationship of fault propagation. Especially under strong noise and boundary conditions, feature extraction is easily affected by non-critical channel interference, which can lead to abrupt changes in diagnostic results. Moreover, its interpretability is limited to gradient response heatmaps. It does not integrate expert experience rules and structured mechanism knowledge, and its ability to interpret the deep semantics of complex equipment faults is insufficient. The invention patent CN121579935A invented a fault diagnosis method for power grid equipment based on temporal modeling and knowledge-enhanced reasoning. The core technology of this invention is a multi-path reasoning system that combines long short-term memory networks with graph neural networks, fuzzy logic, and dynamic Bayesian methods. This method has a large computational architecture and relies excessively on manually set prior expert knowledge. Although multi-path reasoning can produce collaborative output, it can significantly increase the system's reasoning latency and parameter tuning dimensionality. Especially when multimodal data is severely missing or asynchronous, the fault tolerance of the fusion decision is easily affected by the parameter drift of local models (such as fuzzy rule thresholds), resulting in a decrease in robustness. Furthermore, it only uses a multi-objective genetic algorithm for periodic hyperparameter optimization and does not integrate a lightweight end-to-end dynamic adaptive mechanism, which is not universally applicable to edge-side computing power-limited equipment and scenarios requiring high-frequency real-time response.

[0004] In addition, some methods have attempted to improve fault diagnosis performance through multi-source fusion, semantic embedding, prototype alignment, or expert division of labor. However, existing expert division of labor methods usually lack a clear correspondence with the fault mechanism, making it difficult to form a stable discrimination path consistent with the fault mechanism. Existing semantic alignment methods also mostly remain at the feature enhancement level and have not yet effectively solved the problem of differentiated modeling under the condition of "similar local responses but different mechanisms" in servo mechanism fault diagnosis. Summary of the Invention

[0005] The purpose of this invention is to address the shortcomings of the prior art. This invention provides a servo mechanism fault diagnosis method and system based on multi-source representation and hybrid experts, which has the advantages of deep collaboration of multi-source information, diagnostic decision-making supported by clear physical mechanisms, and high accuracy and robustness in hierarchical fault identification under complex working conditions.

[0006] The technical solution to achieve the purpose of this invention is as follows:

[0007] A fault diagnosis method for servo mechanisms based on multi-source characterization and hybrid experts includes the following steps:

[0008] Step S1, Observation Representation Learning: Collect multi-source observation signals of displacement, current, voltage and temperature rise during the operation of the servo mechanism, process the multi-source observation signals using a sliding window method to construct window-level diagnostic samples; perform representation learning on different types of observation signals in the window-level diagnostic samples, extract local features and fuse them to form a unified fault representation;

[0009] Step S2, Diagnostic semantic mapping: Based on the unified fault representation, a semantic prototype alignment mechanism is introduced to map the unified fault representation to the diagnostic semantic space corresponding to the fault type, fault mechanism and fault location, and obtain the semantically aligned sample representation.

[0010] Step S3, Local Temporal Context Enhancement: For the semantically aligned sample representation, neighborhood window information is introduced to perform context aggregation, resulting in a context-enhanced sample representation;

[0011] Step S4, Hybrid expert diagnosis with fault mechanism constraints: Input the context-enhanced sample representation into the hybrid expert diagnosis model, and use a gating network to perform expert route allocation to determine the route weight of each expert sub-network; Based on the route weight, perform weighted fusion of the outputs of each expert sub-network to obtain hierarchical diagnosis results of fault type, fault mechanism and fault location;

[0012] Step S5: Structured diagnostic output: Based on the hierarchical diagnostic results, generate a structured diagnostic result that includes diagnostic conclusions, mechanism explanations, and treatment suggestions.

[0013] Furthermore, in step S1, the window-level diagnostic sample construction process includes: using a sliding window mechanism to perform time-domain segmentation of the servo mechanism's operating state, and using a preset window length and sliding step size to transform continuous operating data into discrete diagnostic samples with local feature expression capabilities.

[0014] Specifically, the original observation sequence of the servo mechanism is represented as:

[0015]

[0016] In the formula, For sampling length, For the first Multi-source observation vectors at each time step;

[0017] Use a length of Step size is The discretized diagnostic samples are constructed using a sliding window, then the first... Each window sample is denoted as:

[0018]

[0019] In the formula, For the first A window of samples, For window length, This is the sliding step size;

[0020] Each window-level diagnostic sample is associated with a corresponding hierarchical diagnostic label; wherein, the hierarchical diagnostic label includes an abnormal state label, a fault type label, a fault mechanism label, and a fault location label.

[0021] Furthermore, the process of constructing a unified fault representation in step S1 includes, assuming the first... Each window sample contains For different types of input segments, separate encoders are set up to extract local features:

[0022]

[0023] In the formula, Indicates the first The encoder corresponding to the type of signal, Indicates the first In the nth window sample Input fragment.

[0024] Furthermore, the local features extracted from each branch are concatenated, and a shared mapping operator is introduced to reduce the dimensionality and semantically align the concatenated features, generating a shared fault representation for subsequent hierarchical diagnostic tasks.

[0025]

[0026] In the formula, Indicates a shared mapping function. Indicates the first A unified fault representation corresponding to each window sample;

[0027] By combining branch coding and sharing, a unified fault representation is obtained that characterizes the differences in statistical properties and physical meanings of different observations and corresponds to hierarchical diagnostic tasks.

[0028] Furthermore, the diagnostic semantic mapping includes:

[0029] To obtain a shared fault representation, a semantic prototype alignment mechanism is introduced, and the representation projection function is set as follows: Then the representation of the sample in the semantic space is:

[0030]

[0031] In the formula, Indicates the first The representation of each sample in the semantic space;

[0032] Semantic prototypes are constructed for each fault type, fault mechanism, and fault location, and hierarchical diagnosis prediction is completed by matching the similarity between the sample representation and each semantic prototype.

[0033] The semantic alignment stage employs a joint optimization approach, and the objective function of the joint optimization is written as:

[0034]

[0035] In the formula, This represents the diagnostic supervision loss based on semantic prototype matching. Indicates loss of auxiliary tasks. This represents a consistency constraint term between the fault type and the fault mechanism. This represents the distillation constraint term. , and These are non-negative weighting coefficients, used to characterize the weights of the auxiliary task loss, consistency constraint, and distillation constraint in the joint optimization objective, respectively.

[0036] The distillation constraint term uses the output distribution of the shared representation branch in the observation representation learning as a soft objective to constrain the discriminative structure shift in the semantic mapping process, thereby determining the discriminativeness and semantic consistency of the feature representation.

[0037] Furthermore, the local temporal context enhancement includes: assuming the semantically aligned sample representation is... Extract the neighborhood window information of the sample representation and perform context aggregation to obtain the context-enhanced representation:

[0038]

[0039] In the formula, This indicates a context aggregation function. This represents a local sequence centered on the current window. This represents context-enhanced representation;

[0040] The context-enhanced representation is further projected onto the diagnostic feature space using a shared mapping function to obtain shared diagnostic features, which serve as the input basis for subsequent gated route generation.

[0041]

[0042] In the formula, Indicates a shared mapping function;

[0043] Through the aforementioned context enhancement process, the current window representation is jointly modeled with the local temporal information of adjacent time points to obtain shared diagnostic features that represent the dynamic evolution of faults and cross-window correlation information.

[0044] Furthermore, the hybrid expert diagnosis constrained by the fault mechanism includes: inputting the shared diagnostic features into the hybrid expert diagnosis model, using a gating network to perform expert route allocation to generate route weights, and weighted fusing the outputs of each expert sub-network based on the route weights to obtain hierarchical diagnostic results. The shared diagnostic features specifically include:

[0045]

[0046] In the formula, Indicates the gating mapping parameters, Indicates the first The probability of each sample being assigned to each expert;

[0047] Let the set of activation experts be... The expert representation is obtained by weighting the corresponding routes according to their assigned weights. Used for fault type identification, fault mechanism determination, and fault location identification:

[0048]

[0049] In the formula, Indicates the first A network of experts Indicates the first The sample at the th The weighting of each expert.

[0050] Furthermore, the optimization objective of the hybrid expert diagnostic model is jointly composed of main task supervision, mechanism routing supervision, and hierarchical consistency constraints. The actual fault mechanism label of the sample is used as a supervision signal to constrain the gating learning process, ensuring that the expert routing allocation result corresponds to explicit diagnostic semantics. Let the actual fault mechanism label of the sample be... The gate monitoring loss is defined as:

[0051]

[0052] In the formula, Represents cross-entropy loss, This represents the expert allocation result output by the gating network.

[0053] Furthermore, the specific process of the structured diagnostic output in step S5 includes: outputting the hierarchical diagnostic results from the hybrid expert diagnostic model. and corresponding fusion diagnostic representation The organization is a unified condition vector:

[0054]

[0055] In the formula, Represents a conditional mapping function;

[0056] Let the target structured result sequence be Then the conditional generation probability is expressed as:

[0057]

[0058] Based on the aforementioned conditions, a structured diagnostic result containing diagnostic conclusions, mechanism explanations, and treatment recommendations is generated.

[0059] The servo mechanism fault diagnosis system based on multi-source characterization and hybrid experts specifically includes:

[0060] The observation characterization learning module is configured to perform branch coding and shared fusion of multi-source observation signals of displacement, current, voltage and temperature rise, and output a unified fault characterization.

[0061] The diagnostic semantic mapping module is configured to construct a semantic space mapping based on the unified fault representation and complete the matching with the semantic prototypes corresponding to the fault type, fault mechanism and fault location.

[0062] The local temporal context enhancement module is configured to perform neighborhood window aggregation on the semantically aligned sample representations and output context-enhanced representations.

[0063] The hybrid expert diagnostic module is configured to perform gated routing and expert assignment under fault mechanism constraints, and output hierarchical diagnostic results.

[0064] The structured output module is configured to generate structured diagnostic results based on hierarchical diagnostic results.

[0065] Compared with the prior art, the present invention has the following beneficial effects:

[0066] (1) This invention discloses a fault diagnosis method and system for servo mechanisms based on multi-source representation and hybrid experts. This method combines a hybrid expert diagnosis mechanism of unified representation of multi-source observation, semantic prototype alignment and fault mechanism constraints. Based on the unified fault representation, a stable correspondence between signal features and fault type, fault mechanism and fault location is established. Based on the fault mechanism constraints, a differentiated discrimination path for different fault modes is constructed. By combining the two, hierarchical fault diagnosis and structured diagnosis output can be achieved relatively stably when the servo mechanism faces similar local observation responses and fault mechanism differences hidden in multivariate correlation and dynamic evolution process.

[0067] (2) In the process of diagnostic modeling and result generation, this invention uses semantic prototype alignment to enhance the diagnostic orientation of observation representation, and uses local temporal context enhancement and gating routing mechanism to form expert division of labor constrained by fault mechanism, so as to improve the problem of insufficient ability to distinguish complex fault modes under the traditional single shared discrimination path; at the same time, the hierarchical diagnostic results are further organized into structured results containing diagnostic conclusions, mechanism explanations and treatment suggestions, thereby improving the fault identification performance and hierarchical consistency, while enhancing the interpretability, completeness and engineering application value of the diagnostic results;

[0068] (3) This invention introduces a local temporal context enhancement mechanism to deeply co-model the observation representation of the current sampling window with the local temporal information of adjacent time points, effectively overcoming the shortcomings of traditional single-time-window features that are susceptible to local transient disturbances and high-frequency noise. This mechanism not only accurately depicts the continuous dynamic evolution law of the fault state and the cross-window correlation features, but also provides a feature base with temporal smoothness for subsequent hybrid expert routing allocation, thereby significantly improving the model's anti-interference ability and diagnostic robustness under complex and non-stationary conditions. Attached Figure Description

[0069] Figure 1 This is an overall framework diagram of the servo mechanism fault diagnosis method based on multi-source characterization and hybrid experts proposed in this invention;

[0070] Figure 2 This is a schematic diagram of the observation representation learning process according to an embodiment of the present invention;

[0071] Figure 3 This is a schematic diagram of the diagnostic semantic mapping process according to an embodiment of the present invention;

[0072] Figure 4 This is a diagram of a hybrid expert architecture according to an embodiment of the present invention;

[0073] Figure 5 This is a fault mechanism expert routing distribution diagram according to an embodiment of the present invention;

[0074] Figure 6This is a fault type expert routing distribution diagram according to an embodiment of the present invention;

[0075] Figure 7 This is a structured output example diagram of an embodiment of the present invention. Detailed Implementation

[0076] To more clearly describe the ideas, technical solutions, and advantages of the present invention, specific embodiments are illustrated through examples and accompanying drawings. Obviously, the described embodiments are only a portion, not all, of the embodiments of the present invention. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without inventive effort are within the scope of protection of the present invention.

[0077] like Figure 1 As shown, the servo mechanism fault diagnosis method based on multi-source observation representation and mechanism constraint hybrid expert provided in this embodiment includes five steps: observation representation learning, diagnostic semantic mapping, local temporal context enhancement, hybrid expert diagnosis based on fault mechanism constraint, and structured diagnostic output.

[0078] This embodiment provides a servo mechanism fault diagnosis method based on a hybrid expert approach combining multi-source observation representation and mechanistic constraints. The method employs a phased modeling approach, comprising four stages: observation representation learning, diagnostic semantic mapping, hybrid expert diagnosis combining local temporal context enhancement and fault mechanism constraints, and structured diagnostic output. The representation results obtained in each stage serve as the input basis for the next stage, enabling the model to gradually transition from shared discriminative ability to semantic constraints, mechanistic division of labor, and diagnostic representation capabilities.

[0079] Step S1, Observation Representation Learning:

[0080] The purpose of this step is to uniformly characterize the multi-source observations during the operation of the servo mechanism, obtaining shared feature representations that can serve fault type identification, fault mechanism determination, and fault location identification. For example... Figure 2 As shown, this implementation first constructs window-level samples and extracts multi-branch representations from multi-source observation signals such as displacement, current, voltage and temperature rise, and then forms a unified fault representation through sharing and fusion.

[0081] S11, Window-level Sample Construction

[0082] First, the original observation sequence of the servo mechanism is represented as follows:

[0083]

[0084] in, For sampling length, For the first The multi-source observation vector at each time step uses a length of [length missing]. Step size is If a sliding window is used to construct samples, then the first... Each window sample is denoted as:

[0085]

[0086] Each window sample corresponds to a set of hierarchical diagnostic labels, used to describe the abnormal state, fault type, fault mechanism, and fault location. In this way, continuous operational observations can be transformed into a unified window-level input for hierarchical diagnostic tasks.

[0087] S12, Multi-branch local characterization extraction

[0088] Let the first Each window sample contains For different types of input segments, separate encoders are set up to extract local features:

[0089]

[0090] in, Indicates the first The encoder corresponding to the type of signal, Indicates the first In the nth window sample Input fragments, This represents the corresponding local feature representation. Through the aforementioned multi-branch encoding method, local discrimination information from observations of different sources, such as displacement, current, voltage, and temperature rise, can be extracted.

[0091] S13, Construction of Shared and Fusion Representations

[0092] After obtaining the local features of each branch, they are concatenated and mapped to a shared fault representation:

[0093]

[0094] in, Indicates a shared mapping function. Indicates the first A unified fault representation corresponding to each window sample.

[0095] In this embodiment, the local features of multiple branches are not simply added together during the sharing and fusion stage, but are further jointly modeled. Let the fused node be represented as... Its corresponding importance weight is Then, the unified fault characterization can also be expressed as:

[0096]

[0097] in, This indicates the number of nodes in the fusion phase. Indicates the first The degree of contribution of each node to the final unified representation.

[0098] This approach forms a shared fault characterization while preserving the differences in various observations, enabling it to simultaneously serve hierarchical diagnostic tasks such as fault type identification, fault mechanism determination, and fault location identification.

[0099] S14, Optimization of the Shared Representation Stage

[0100] To ensure that the shared representation possesses both diagnostic discriminative power and good hierarchical consistency, this implementation method employs a joint optimization approach for the shared representation stage. Let the main task loss for the fault type be... The failure mechanism branch loss is The hierarchical consistency constraint is The characterization of the comparison constraint term is The category boundary constraint term is Then the objective function of the shared representation stage can be written as:

[0101]

[0102] in, These are the weight coefficients for each constraint term. This joint optimization method enables the unified representation to form a relatively stable category structure and hierarchical semantic foundation in the early stages of training.

[0103] Step S2, Diagnose semantic mapping:

[0104] The purpose of this step is to introduce diagnostic semantic constraints based on shared fault representations, establishing a stable correspondence between sample representations and fault types, fault mechanisms, and fault locations. For example... Figure 3 As shown, after obtaining a unified fault representation, it is mapped to a diagnostic semantic space consisting of fault type, fault mechanism and fault location, and hierarchical semantic alignment is completed through the matching relationship between sample representation and corresponding semantic prototype.

[0105] S21, Semantic Space Projection

[0106] After obtaining the shared fault representation, a semantic prototype alignment mechanism is introduced, whereby the representation projection function is... Then the representation of the sample in the semantic space is:

[0107]

[0108] in, Indicates the first The representation of a sample in the semantic space.

[0109] S22, Semantic Prototype Construction

[0110] Semantic prototypes are constructed for each fault type, fault mechanism, and fault location. Let the first... The length of the descriptive text corresponding to each semantic prototype is . , No. Each word element is embedded as The mask is Then the first A semantic prototype vector can be represented as:

[0111]

[0112] in, This represents a normalization operation. Using the above method, semantic labels such as fault type, fault mechanism, and fault location can be mapped to a unified prototype representation.

[0113] S23, Prototype Matching and Alignment Supervision

[0114] Let the first The semantic representation of each sample is as follows , No. The semantic prototype is The temperature coefficient is The matching score between the sample and the semantic prototype can be expressed as:

[0115]

[0116] Hierarchical prediction of fault type, fault mechanism, and fault location is achieved by matching the similarity between sample representations and semantic prototypes. Compared with methods that rely solely on numerical spatial separability, semantic prototype alignment subjectes feature representations to the dual constraints of observation patterns and diagnostic semantics.

[0117] S24 Joint optimization of semantic alignment stage

[0118] To ensure that the semantic mapping process balances discriminative ability and training stability, a joint optimization approach is adopted in the semantic alignment stage, and its objective function can be written as:

[0119]

[0120] in, This represents the diagnostic supervision loss based on semantic prototype matching. Indicates loss of auxiliary tasks. This represents a consistency constraint term between the fault type and the fault mechanism. This represents the distillation constraint term. Through the aforementioned joint optimization method, the discriminative structure shift during the semantic mapping process can be suppressed while aligning diagnostic semantics.

[0121] Step S3, Local Temporal Context Enhancement:

[0122] The purpose of this step is to enhance the model's ability to express the continuous evolution of faults, neighborhood window association information, and local difference patterns. After completing the diagnostic semantic mapping, neighborhood window information is further introduced with the current window as the center, and context-enhanced representation is formed through context aggregation, thereby improving the model's ability to characterize the dynamic evolution of faults.

[0123] S31, Neighborhood Window Context Construction

[0124] Let the semantically aligned center window be represented as Select neighboring windows centered on this window to construct a local sequence:

[0125]

[0126] in, This represents the length of the local sequence. This method allows the semantic representations of the current window and its neighboring windows to be fed into the context aggregation module together.

[0127] S32, Context Aggregation

[0128] Let the context aggregation function be Then, context enhancement is represented as:

[0129]

[0130] In this embodiment, the context-enhanced representation is formed by a weighted average of the central window representation and the local sequence average representation. Let the central window encoding result be... The local sequence average is expressed as The context-enhanced representation can then be further written as:

[0131]

[0132] in, Let be the weight coefficients represented by the central window, and satisfy . In this way, while maintaining the diagnostic focus of the current window, the evolutionary information of neighboring windows is also introduced.

[0133] S33, Context-enhanced shared diagnostic features

[0134] To facilitate subsequent division of labor among experts, the context-enhanced representation is further mapped to shared diagnostic features:

[0135]

[0136] in, Indicates a shared mapping function. This represents the shared diagnostic features used for subsequent gated routing and expert fusion. Through this step, local temporal context information is uniformly compressed into a representation space suitable for subsequent hybrid expert discrimination.

[0137] Step S4: Hybrid expert diagnosis with fault mechanism constraints:

[0138] The purpose of this step is to utilize gated routing and expert division of labor to create differentiated discrimination paths for different fault mechanisms. For example... Figure 4 As shown, this implementation constructs a hybrid expert architecture based on context-enhanced representation, routes and allocates samples through a gating network, and completes differentiated discrimination and weighted fusion by multiple expert sub-networks.

[0139] S41, Gated Route Weight Generation

[0140] Build includes A hybrid expert structure with multiple expert subnetworks, where the gated network generates routing weights for each expert based on shared diagnostic features:

[0141]

[0142] in, Indicates the gating mapping parameters, Indicates the first The probability of each sample being assigned to each expert.

[0143] Figure 5 The distribution of expert routes corresponding to the fault mechanisms is presented to illustrate that, under the constraints of fault mechanisms, samples of different fault mechanism categories can form discriminative expert assignment results.

[0144] S42, Sparse Expert Activation and Fusion Representation

[0145] To enhance the effectiveness of expert division of labor, a sparse routing strategy is adopted, activating only the experts with the highest weights. Let the set of activated experts be . The experts after fusion expressed it as follows:

[0146]

[0147] in, Indicates the first A network of experts Indicates the first The sample at the th The weighting of each expert.

[0148] To simultaneously maintain the stable discrimination capability of shared paths and the local difference modeling capability of expert paths, let the global fault type output be... , No. The local fault type output by each expert is: The final fault type output can then be represented as:

[0149]

[0150] in, This represents the fusion coefficient of the local expert judgment results. Using the above method, the differentiated judgment capabilities resulting from expert division of labor can be further introduced on top of shared diagnostic results.

[0151] Figure 6 The distribution of expert routes corresponding to different fault types is presented to illustrate the differentiated routing distribution characteristics of samples of different fault types in the hybrid expert structure.

[0152] S43. Introduction of Fault Mechanism Constraints

[0153] To ensure that expert analysis tools have clear diagnostic semantic basis, fault mechanism labels are directly introduced into the gating learning process. Let the true fault mechanism label of the sample be... The output of the gating network is The gate monitoring loss is defined as:

[0154]

[0155] in, This represents the cross-entropy loss. By directly using fault mechanism information as a gating supervision signal, the division of labor among experts can be transformed from free allocation to differentiated discrimination paths constrained by the fault mechanism.

[0156] S44, Expert Load Balancing and Joint Optimization

[0157] To avoid excessive concentration of samples among a few experts, an expert load balancing constraint is introduced. Let the... The routing frequency of each expert is The average gate probability is Then the load balancing term can be written as:

[0158]

[0159] Based on this, the overall optimization objective of the hybrid expert diagnosis phase can be expressed as:

[0160]

[0161] in, This indicates the main task monitoring loss corresponding to the fault type, fault mechanism, and fault location. Indicates the loss of gate monitoring, This represents the expert load balancing constraint. This represents a hierarchical consistency constraint. Through the aforementioned joint optimization method, the division of labor among experts maintains differentiation while possessing a clear mechanistic orientation.

[0162] Step S5, Structured Diagnostic Output:

[0163] The purpose of this step is to transform the hierarchical diagnostic results into structured diagnostic text with fixed fields and engineering meaning. For example... Figure 7 As shown, after obtaining the hierarchical diagnostic results, this embodiment organizes the diagnostic information such as fault type, fault mechanism and fault location into a unified condition representation, and further generates structured diagnostic results.

[0164] S51, Conditional Representation Construction

[0165] After obtaining the hierarchical diagnostic results, the diagnostic information, such as fault type, fault mechanism, and fault location, is organized into a unified conditional representation. Let the hierarchical diagnostic results output by the hybrid expert diagnostic model be... The corresponding fusion diagnosis is represented as The conditional vector for generation is then written as:

[0166]

[0167] in, This represents a conditional mapping function.

[0168] S52, Conditional Output Generation

[0169] Let the target structured result sequence be Then the conditional generation probability is expressed as:

[0170]

[0171] Furthermore, if the target structured result sequence is written as Then, the process of generating its conditions can be expanded as follows:

[0172]

[0173] in, Indicates the first This method utilizes previously generated result fragments at each location. It transforms hierarchical diagnostic results into structured diagnostic expressions geared towards engineering applications.

[0174] S53. Structured Diagnostic Results Organization Format

[0175] In this embodiment, the structured diagnostic results are organized using a fixed-field format, including at least five parts: "Conclusion, Key Evidence, Mechanism Analysis, Risk Impact, and Recommended Measures." By using a fixed-field format, a consistent structure for output results from different samples can be ensured, while also enhancing the interpretability, completeness, and engineering usability of the diagnostic results. Figure 7 The example shown illustrates that the output of this implementation method is organized in a fixed field format, which can present the diagnostic conclusions, key evidence, mechanism analysis, risk impact, and recommended measures in a relatively complete manner.

[0176] In summary, this implementation method achieves the proposed method step by step through the sequence of "shared representation learning—semantic prototype alignment—local temporal context enhancement—hybrid expert diagnosis with fault mechanism constraints—structured diagnostic output." This implementation method retains the advantages of unified representation from multi-source observations, while improving the diagnostic capability under complex fault modes through semantic constraints, temporal enhancement, and expert division of labor, and ultimately achieving a unified output from fault discrimination to diagnostic expression.

[0177] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A servo mechanism fault diagnosis method based on multi-source representation and hybrid expert, characterized in that, Includes the following steps: Step S1, Observation Representation Learning: Collect multi-source observation signals of displacement, current, voltage and temperature rise during the operation of the servo mechanism, process the multi-source observation signals using a sliding window method to construct window-level diagnostic samples; perform representation learning on different types of observation signals in the window-level diagnostic samples, extract local features and fuse them to form a unified fault representation; Step S2, Diagnostic semantic mapping: Based on the unified fault representation, a semantic prototype alignment mechanism is introduced to map the unified fault representation to the diagnostic semantic space corresponding to the fault type, fault mechanism and fault location, and obtain the semantically aligned sample representation. Step S3, Local Temporal Context Enhancement: For the semantically aligned sample representation, neighborhood window information is introduced to perform context aggregation, resulting in a context-enhanced sample representation; Step S4, Hybrid expert diagnosis with fault mechanism constraints: Input the context-enhanced sample representation into the hybrid expert diagnosis model, and use a gated network to perform expert route allocation to determine the routing weight of each expert sub-network; The outputs of each expert subnetwork are weighted and fused based on the routing weights to obtain hierarchical diagnostic results of fault type, fault mechanism and fault location; Step S5: Structured diagnostic output: Based on the hierarchical diagnostic results, generate a structured diagnostic result that includes diagnostic conclusions, mechanism explanations, and treatment suggestions.

2. The servo mechanism fault diagnosis method based on multi-source representation and hybrid expert as described in claim 1, characterized in that, In step S1, the window-level diagnostic sample construction process includes: using a sliding window mechanism to perform time-domain segmentation of the servo mechanism's operating state, and using a preset window length and sliding step size to transform continuous operating data into discrete diagnostic samples with local feature expression capabilities. Specifically, the original observation sequence of the servo mechanism is represented as: In the formula, For sampling length, For the first Multi-source observation vectors at each time step; Use a length of Step size is The discretized diagnostic samples are constructed using a sliding window, then the first... Each window sample is denoted as: In the formula, For the first A window of samples, For window length, This is the sliding step size; Each window-level diagnostic sample is associated with a corresponding hierarchical diagnostic label; wherein, the hierarchical diagnostic label includes an abnormal state label, a fault type label, a fault mechanism label, and a fault location label.

3. The servo mechanism fault diagnosis method based on multi-source representation and hybrid expert as described in claim 1, characterized in that, The process of constructing a unified fault representation in step S1 includes, assuming the first... Each window sample contains For different types of input segments, separate encoders are set up to extract local features: In the formula, Indicates the first The encoder corresponding to the type of signal, Indicates the first In the nth window sample Input fragment.

4. The servo mechanism fault diagnosis method based on multi-source representation and hybrid expert as described in claim 3, characterized in that, The local features extracted from each branch are concatenated, and a shared mapping operator is introduced to reduce the dimensionality and semantically align the concatenated features, generating a shared fault representation for subsequent hierarchical diagnostic tasks. In the formula, Indicates a shared mapping function. Indicates the first A unified fault representation corresponding to each window sample; By combining branch coding and sharing, a unified fault representation is obtained that characterizes the differences in statistical properties and physical meanings of different observations and corresponds to hierarchical diagnostic tasks.

5. The servo mechanism fault diagnosis method based on multi-source representation and hybrid expert as described in claim 1, characterized in that, The diagnostic semantic mapping includes: To obtain a shared fault representation, a semantic prototype alignment mechanism is introduced, and the representation projection function is set as follows: Then the representation of the sample in the semantic space is: In the formula, Indicates the first The representation of each sample in the semantic space; Semantic prototypes are constructed for each fault type, fault mechanism, and fault location, and hierarchical diagnosis prediction is completed by matching the similarity between the sample representation and each semantic prototype. The semantic alignment stage employs a joint optimization approach, and the objective function of the joint optimization is written as: In the formula, This represents the diagnostic supervision loss based on semantic prototype matching. Indicates loss of auxiliary tasks. This represents a consistency constraint term between the fault type and the fault mechanism. This represents the distillation constraint term. , and These are non-negative weighting coefficients, used to characterize the weights of the auxiliary task loss, consistency constraint, and distillation constraint in the joint optimization objective, respectively. The distillation constraint term uses the output distribution of the shared representation branch in the observation representation learning as a soft objective to constrain the discriminative structure shift in the semantic mapping process, thereby determining the discriminativeness and semantic consistency of the feature representation.

6. The servo mechanism fault diagnosis method based on multi-source representation and hybrid expert as described in claim 1, characterized in that, The local temporal context enhancement includes: assuming the semantically aligned sample representation is... Extract the neighborhood window information of the sample representation and perform context aggregation to obtain the context-enhanced representation: In the formula, This indicates a context aggregation function. This represents a local sequence centered on the current window. This represents context-enhanced representation; The context-enhanced representation is further projected onto the diagnostic feature space using a shared mapping function to obtain shared diagnostic features, which serve as the input basis for subsequent gated route generation. In the formula, Indicates a shared mapping function; Through the aforementioned context enhancement process, the current window representation is jointly modeled with the local temporal information of adjacent time points to obtain shared diagnostic features that represent the dynamic evolution of faults and cross-window correlation information.

7. The servo mechanism fault diagnosis method based on multi-source characterization and hybrid expert as described in claim 6, characterized in that, The hybrid expert diagnosis constrained by the fault mechanism includes: inputting the shared diagnostic features into the hybrid expert diagnosis model, using a gated network to perform expert route allocation to generate route weights, and weighting and fusing the outputs of each expert sub-network based on the route weights to obtain hierarchical diagnostic results. The shared diagnostic features specifically include: In the formula, Indicates the gating mapping parameters, Indicates the first The probability of each sample being assigned to each expert; Let the set of activation experts be... The expert representation is obtained by weighting the corresponding routes according to their assigned weights. Used for fault type identification, fault mechanism determination, and fault location identification: In the formula, Indicates the first A network of experts Indicates the first The sample at the th The weighting of each expert.

8. The servo mechanism fault diagnosis method based on multi-source representation and hybrid expert as described in claim 1, characterized in that, The optimization objective of the hybrid expert diagnostic model consists of main task supervision, mechanism routing supervision, and hierarchical consistency constraints. The actual fault mechanism labels of the samples are used as supervision signals to constrain the gating learning process, ensuring that the expert routing allocation results correspond to explicit diagnostic semantics. Let the actual fault mechanism label of the sample be... The gate monitoring loss is defined as: In the formula, Represents cross-entropy loss, This represents the expert allocation result output by the gating network.

9. The servo mechanism fault diagnosis method based on multi-source representation and hybrid expert as described in claim 1, characterized in that, The specific process of the structured diagnostic output in step S5 includes: outputting the hierarchical diagnostic results from the hybrid expert diagnostic model. and corresponding fusion diagnostic representation The organization is a unified condition vector: In the formula, Represents a conditional mapping function; Let the target structured result sequence be The conditional generation probability is expressed as: Based on the aforementioned conditions, a structured diagnostic result containing diagnostic conclusions, mechanism explanations, and treatment recommendations is generated.

10. A servo mechanism fault diagnosis system based on multi-source representation and hybrid expert, characterized in that, The servo mechanism fault diagnosis method based on multi-source characterization and hybrid expert as described in any one of claims 1 to 9, wherein the servo mechanism fault diagnosis system based on multi-source characterization and hybrid expert comprises: The observation characterization learning module is configured to perform branch coding and shared fusion of multi-source observation signals of displacement, current, voltage and temperature rise, and output a unified fault characterization. The diagnostic semantic mapping module is configured to construct a semantic space mapping based on the unified fault representation and complete the matching with the semantic prototypes corresponding to the fault type, fault mechanism and fault location. The local temporal context enhancement module is configured to perform neighborhood window aggregation on the semantically aligned sample representations and output context-enhanced representations. The hybrid expert diagnostic module is configured to perform gated routing and expert assignment under fault mechanism constraints, and output hierarchical diagnostic results. The structured output module is configured to generate structured diagnostic results based on hierarchical diagnostic results.

Images