A state recognition method for an airborne inertial measurement unit

Abstract

This invention relates to a method for state identification of an airborne inertial measurement unit (INS). The method includes: acquiring operational data of the INS and converting it into a joint dataset according to a unified standard mapping relationship; performing feature extraction and coordinated modulation on the joint dataset, and outputting health status classification results under different operating conditions through a full-condition health status identification model; optimizing the full-condition health status identification model to generate an INS health status identification model oriented towards all operating conditions; and deploying locally to output the INS health status classification results in real time. This invention establishes a standardized benchmark with unified data modalities and aligned physical features, utilizes a health status feature encoding network to map multi-source time-series operational data into high-dimensional failure feature vectors, and implements an embedded operating condition adaptive verification and feature collaborative optimization strategy to ensure the accuracy and stability of health status identification results under different flight conditions.

Description

Technical Field

[0001] This invention relates to the technical field of intelligent detection of avionics equipment, and more particularly to a state recognition method for an airborne inertial measurement unit. Background Technology

[0002] As the core sensing device for aircraft navigation, attitude control, and motion perception, the airborne inertial measurement unit (IMU) directly determines the flight accuracy, handling stability, and flight safety of an aircraft. It is widely used in various aviation platforms, including civil airliners, military aircraft, unmanned aerial vehicles, and space launch vehicles. This device mainly consists of a gyroscope, accelerometer, inertial measurement module, and signal processing unit. Its health status is primarily reflected in measurement accuracy stability, zero-bias drift, scale factor error, dynamic response characteristics, and performance retention under complex flight environments. These macroscopic health performance characteristics are not determined by the state of a single component but are constrained by the collaborative working state of multiple components, including the internal mechanical structure, electronic components, sensing units, and signal transmission links. Its microscopic failure characteristics specifically involve multi-dimensional and multi-scale operational features such as gyroscope rotor wear, accelerometer sensing element fatigue state, circuit contact impedance changes, signal noise amplitude and frequency distribution, and data synchronization deviations among multiple sensing units.

[0003] However, in actual aviation operation and equipment support, the health status identification of inertial measurement units (IMUs) still heavily relies on traditional manual inspection and threshold judgment paradigms. Technicians typically pre-set judgment thresholds for key indicators such as zero-bias drift and scale factor error based on factory calibration parameters, empirical thresholds, and periodic offline testing data. They then collect unit operating data using ground-based testing equipment and use oscilloscopes, data acquisition instruments, and manual analysis to determine the equipment's operating parameters and health status. In this forward inspection mode, the identification of equipment health status and fault warnings often depend on manual experience and fixed thresholds, making it difficult to form an efficient, accurate, and real-time intelligent identification path, and thus unable to adapt to the complex and ever-changing flight environment and dynamic operating conditions of aircraft.

[0004] Existing technologies have the following limitations: First, the operating parameters of airborne inertial measurement units (IMUs) are highly dimensional and susceptible to numerous environmental interference factors. Environmental variables such as temperature, vibration, and air pressure during flight are coupled with the degradation of the equipment's own performance, making it difficult for a single fixed threshold to cover the health assessment requirements under all operating conditions. Furthermore, offline detection cannot capture instantaneous failure characteristics during dynamic operation, resulting in low identification accuracy and a high rate of missed and false positives. Second, there is a highly nonlinear and complex mapping relationship between the equipment's microscopic failure characteristics and macroscopic health performance. Different failure modes exhibit cross-coupling effects. For example, gyroscope rotor wear can lead to increased zero-bias drift and also cause an increase in signal noise amplitude. Relying solely on human experience or qualitative thresholds cannot accurately characterize the quantitative contribution of each failure characteristic to the health status, let alone achieve accurate identification and early warning of early, minor faults. Furthermore, in actual aviation flight scenarios, real-time online health status identification of IMUs is often required to meet the dynamic decision-making needs of the aircraft flight control system. However, existing methods struggle to extract effective failure characteristics from multi-source dynamic operating data and achieve accurate health status classification under constraints of strong real-time performance and high reliability.

[0005] Furthermore, the vast amounts of operational data, fault records, and environmental monitoring data accumulated by different aviation research and development institutions, aircraft manufacturers, and maintenance units during long-term operations are typically stored separately in their respective maintenance systems. Moreover, some of this data involves core aviation flight safety information and commercial technical secrets, making centralized sharing and unified modeling difficult without compromising privacy. This reality objectively restricts the construction of large-scale, high-quality aviation equipment health datasets, further limiting the widespread application of data-driven health status identification models in actual aviation maintenance scenarios. Therefore, how to achieve full-condition, real-time, and high-precision intelligent health status identification of aviation inertial measurement units while ensuring data privacy and aviation information security has become a critical technical problem that urgently needs to be solved in this field.

[0006] From the perspective of existing technologies, the field of aviation equipment health identification mainly employs traditional machine learning classification methods and physical model-based fault diagnosis methods. While physical model-based fault diagnosis methods can analyze failure mechanisms from the perspective of equipment operating principles, their modeling process is complex, highly dependent on equipment structural parameters, difficult to adapt to the personalized characteristics of different inertial measurement units (IMUs), and unable to effectively handle random environmental interference during flight. Traditional machine learning methods typically focus on establishing simple classification relationships between operating parameters and health status, lacking the ability to model the dynamic operating characteristics of equipment over time, struggling to extract deep correlation features from multi-source heterogeneous data, and having insufficient ability to perceive early, subtle fault characteristics. Therefore, the identification results fail to meet the high reliability and high accuracy requirements of the aviation field. In summary, the existing technological system cannot yet meet the health status identification needs of aviation IMUs in terms of full operating conditions, real-time operation, high precision, and early fault warning.

[0007] Therefore, it is necessary to improve one or more of the problems existing in the above-mentioned related technical solutions.

[0008] It should be noted that this section is intended to provide background or context for the technical solutions of the invention as set forth in the claims. The description herein does not imply acceptance as prior art simply because it is included in this section. Summary of the Invention

[0009] The purpose of this invention is to provide a state identification method for an airborne inertial measurement unit, thereby at least partially solving one or more of the problems caused by the limitations and defects of related technologies.

[0010] This invention provides a state identification method for an airborne inertial measurement unit, comprising: The operational data of the airborne inertial measurement unit is acquired and converted into a joint dataset according to a unified standard mapping relationship; wherein the standard mapping relationship includes four dimensions: operating parameters, failure characteristics, environmental conditions and health status. Feature extraction and coordinated modulation are performed on the joint dataset, and the health status classification results under different working conditions are output through the full-condition health status recognition model; wherein, the health status classification results include a working condition feature verification header; The operating condition feature verification head is used to perform feature validity discrimination and operating condition consistency verification. The composite objective function of adversarial loss, operating condition regression loss and feature consistency loss is used to optimize the full operating condition health status identification model, thereby generating an airborne inertial measurement unit health status identification model for all operating conditions. The trained health status recognition model for airborne inertial measurement units (INS) under all operating conditions is deployed locally, and the health status classification results of INS are output in real time.

[0011] Optionally, the step of acquiring the operational data of the airborne inertial measurement unit and converting it into a joint dataset according to a unified standard mapping relationship includes: The continuous and discrete operating parameters of the airborne inertial measurement unit are acquired and normalized to generate a set of operating parameters. Multi-source data acquisition is performed on the operation of the airborne inertial measurement unit, and standardized time-series operation data sequences are generated through time-series synchronization and standardization processing. The standardized time-series operational data is subjected to feature analysis to construct a failure feature vector; Construct a health status label that includes sub-labels for fault location and fault severity; The joint dataset is generated by associating the set of operating parameters, the standardized time-series operating data sequence, and the failure feature vector with the health status label in a one-to-one correspondence.

[0012] Optionally, the step of performing feature extraction and coordinated modulation on the joint dataset, and outputting health status classification results under different working conditions through the full-condition health status identification model, includes: The full-condition health status recognition model includes an encoding network, a classification network, and a discrimination network.

[0013] Optionally, the step of performing feature extraction and coordinated modulation on the joint dataset, and outputting health status classification results under different working conditions through the full-condition health status identification model, includes: The full-condition health status identification model maps the standardized time-series operation data sequence into a failure feature embedding vector through the coding network.

[0014] Optionally, the step of performing feature extraction and coordinated modulation on the joint dataset, and outputting health status classification results under different working conditions through the full-condition health status identification model, includes: The full-condition health status identification model uses the classification network to perform condition-adaptive correction on the failure feature embedding vector and outputs the probability distribution of the health status of the airborne inertial measurement unit.

[0015] Optionally, the step of performing feature extraction and coordinated modulation on the joint dataset, and outputting health status classification results under different working conditions through the full-condition health status identification model, includes: The full-condition health status recognition model performs comprehensive condition consistency supervision through the discrimination network, generates condition feature verification headers, and adds them to the health status classification results.

[0016] Optionally, the step of optimizing the full-condition health status identification model to generate a full-condition airborne inertial measurement unit health status identification model includes: The loss function of the discrimination network is obtained by summing the loss of the true feature discrimination, the loss of the extracted feature discrimination, and the loss of the working condition regression supervision according to the preset weights; The anti-deception loss, working condition consistency loss, and classification accuracy loss are summed according to preset weights to obtain the common loss function of the encoding network and the classification network. The full-condition health status recognition model is optimized and trained using a dynamic game strategy that alternately updates the encoding network, classification network, and discrimination network.

[0017] Optionally, the step of locally deploying the trained, full-condition-oriented airborne inertial measurement unit (INS) health status recognition model and outputting the INS health status classification results in real time includes: Real-time acquisition of multi-source operational data from airborne inertial measurement units through local deployment; The trained, full-condition-oriented airborne inertial measurement unit (INS) health status identification model outputs the real-time collected multi-source operational data as the INS health status classification results. The health status classification results are decoupled and output as independent single failure features.

[0018] Optionally, the step of locally deploying the trained, full-condition-oriented airborne inertial measurement unit (INS) health status recognition model and outputting the INS health status classification results in real time includes: By analyzing the single failure characteristics after decoupling, the system outputs a report on the cause of failure and the prediction of its development trend.

[0019] The technical solution provided by this invention may include the following beneficial effects: In this invention, a standardized benchmark with unified data modalities and aligned physical features is established by collaboratively acquiring multi-source heterogeneous operational data of the airborne inertial measurement unit (INS) and constructing a joint dataset. Based on a deep temporal feature learning architecture, a reverse identification model for the health status of the INS is built for all operating conditions. A health status feature encoding network maps multi-source temporal operational data into high-dimensional failure feature vectors, and the adaptive attention mechanism in the health status classification network enables the targeted extraction of key failure features. Furthermore, an embedded adaptive verification and feature collaborative optimization strategy is implemented, reusing the embedded condition feature verification head in the discrimination network to construct a self-supervised feedback loop for the agentless model, ensuring the accuracy and stability of the health status identification results under different flight conditions. Attached Figure Description

[0020] The accompanying drawings, which are incorporated in and constitute a part of this specification, illustrate embodiments consistent with the invention and, together with the description, serve to explain the principles of the invention. It is obvious that the drawings described below are merely some embodiments of the invention, and those skilled in the art can obtain other drawings based on these drawings without any inventive effort.

[0021] Figure 1 A flowchart illustrating the state recognition method for an airborne inertial measurement unit in an exemplary embodiment of the present invention is shown. Figure 2 A more specific logical diagram of the state recognition method for an airborne inertial measurement unit in an exemplary embodiment of the present invention is shown; Figure 3 This diagram illustrates the logic of collaborative acquisition of multi-source heterogeneous operational data and construction of a joint dataset in an exemplary embodiment of the present invention. Figure 4 This diagram illustrates the logic of constructing a health status identification model for an airborne inertial measurement unit oriented to all operating conditions in an exemplary embodiment of the present invention. Figure 5 This diagram illustrates a comparison of health status recognition accuracy in an exemplary embodiment of the present invention. Figure 6 This diagram illustrates a comparison of recognition stability under different operating conditions in an exemplary embodiment of the present invention. Figure 7 This diagram illustrates an experimental comparison of early fault identification capabilities in an exemplary embodiment of the present invention. Detailed Implementation

[0022] Exemplary embodiments will now be described more fully with reference to the accompanying drawings. However, these exemplary embodiments can be implemented in many forms and should not be construed as limited to the examples set forth herein; rather, they are provided so that the invention will be more comprehensive and complete, and will fully convey the concept of the exemplary embodiments to those skilled in the art. The described features, structures, or characteristics may be combined in any suitable manner in one or more embodiments.

[0023] Furthermore, the accompanying drawings are merely illustrative diagrams of embodiments of the present invention and are not necessarily drawn to scale. The same reference numerals in the drawings denote the same or similar parts, and therefore repeated descriptions of them will be omitted. Some block diagrams shown in the drawings are functional entities and do not necessarily correspond to physically or logically independent entities.

[0024] This invention provides a state identification method for airborne inertial measurement units, with reference to... Figure 1 As shown, it includes the following steps: Step S101: Acquire the operational data of the airborne inertial measurement unit and convert it into a joint dataset according to a unified standard mapping relationship.

[0025] Step S102: Perform feature extraction and coordinated modulation on the joint dataset, and output the health status classification results under different working conditions through the full-condition health status recognition model.

[0026] Step S103: Use the operating condition feature verification head to perform feature validity discrimination and operating condition consistency verification, and optimize the full-operating condition health status identification model with a composite objective function of adversarial loss, operating condition regression loss and feature consistency loss, to generate an airborne inertial measurement unit health status identification model for all operating conditions.

[0027] Step S104: Deploy the trained airborne inertial measurement unit health status recognition model for all operating conditions locally, and output the health status classification results of the airborne inertial measurement unit in real time.

[0028] It is important to understand that the standard mapping relationship includes four dimensions: operating parameters, failure characteristics, environmental conditions, and health status. The health status classification result includes an operating condition characteristic verification header.

[0029] It is also important to understand that, adopting an airborne inertial measurement unit health status identification model architecture design oriented towards full operating condition constraints, a deep identification model architecture based on temporal adaptive normalization and multi-scale feature fusion identification is proposed. The architecture first constructs a health status feature encoding network for airborne inertial measurement units (INS). Through cascaded temporal feature extraction layers and multi-source feature fusion layers, it maps multi-source heterogeneous temporal operational data, including gyroscope measurements, accelerometer outputs, zero-bias drift, scaling factor errors, and environmental monitoring data, to a high-dimensional semantic feature space, forming a failure feature embedding vector with strong operational condition adaptability. Subsequently, an INS health status classification network is constructed. Abandoning the traditional direct feature concatenation input method, it adopts a temporal feature decoding mechanism based on adaptive attention. It uses the failure feature embedding vector to dynamically generate attention weight parameters and performs dual feature modulation of the temporal feature map in both channel and time dimensions, thereby deeply integrating key failure features under different flight conditions during feature extraction. Simultaneously, a multi-scale feature fusion and discrimination network including an operational condition feature verification head is constructed. This network not only judges the authenticity and effectiveness of the extracted failure features at coarse, medium, and fine time scales, but also extracts abstract operational condition features and regresses to predict flight environment parameters through a global temporal pooling layer, achieving dual constraints on the operational condition adaptability and physical consistency of the feature extraction results.

[0030] It is also important to understand that an end-to-end adversarial training strategy based on an embedded condition-adaptive verification and feature-guided alternating collaborative optimization strategy is proposed, which abandons the independent frozen proxy model and is based on an embedded condition feature verification head. During training, the airborne inertial measurement unit health status discrimination network is directly reused as a differentiable condition environment simulator. The condition feature verification head embedded in its architecture is used to measure the condition adaptability of the extracted features in real time. By constructing a composite objective function that includes adversarial discrimination loss and condition consistency loss, a dynamic game is executed in which the feature encoding network, classification network, and discrimination network are updated alternately. Under this mechanism, the condition feature verification error generated by the discrimination network is transformed into a gradient flow that can guide feature adjustment through backpropagation. This gradient flow is directly fed back to the airborne inertial measurement unit health status feature encoding network and classification network, driving them to directionally correct key parameters such as feature extraction weights and attention allocation coefficients. This avoids the error accumulation caused by the accuracy deviation of the proxy model in the traditional step-by-step method, ensuring that the finally extracted failure features accurately anchor the actual health status of the equipment while meeting the distribution law of different flight conditions.

[0031] It is also important to understand that by adopting a multi-source heterogeneous data standardization construction and a health status classification and fault tracing system based on feature decoupling, a full-link data standardization and health status analysis system is established to address the problems of complex data modes, strong operating condition coupling, and difficulty in tracing the causes of faults in the research and development and operation and maintenance of aviation inertial measurement units. First, a joint dataset aligned with physical features is constructed. The heterogeneity of multi-source operating data is eliminated through time synchronization, dimension normalization, and operating condition labeling. Strict indexing and association are established for operating parameters, failure features, environmental conditions, and health status labels. Second, a feature effectiveness optimization mechanism based on dual discrimination criteria is designed. Combined with a feature importance quantification algorithm, the key features that contribute the most to health status identification are selected from the extracted high-dimensional failure features. Most importantly, a dedicated fault tracing network is constructed. Using the key failure feature vector as input, a deep convolutional regression and attention feature parsing architecture is used to decouple the nonlinear mapping relationship between micro-failure features and macro-failure modes. The abstract high-dimensional failure features are restored to standardized analysis results that include health status level, specific fault location, and fault development trend. This realizes a complete intelligent identification closed loop from multi-source operating data to key failure features, and then to interpretable health status determination and fault tracing.

[0032] It is also necessary to understand that, such as Figure 2As shown, the collaborative acquisition of multi-source heterogeneous operational data from inertial measurement units (IMUs) and the construction of a joint dataset aim to build a standardized benchmark dataset with aligned physical features and unified data modes. This dataset establishes a strict mathematical mapping relationship between IMUs of different models and flight conditions across four dimensions: operational parameters, failure characteristics, environmental conditions, and health status. This provides a learnable sample foundation for subsequent health status identification models. The specific process is as follows: First, operational parameter vectorization and sample index binding are performed, transforming continuous and discrete operational parameters into a unified operational parameter vector and establishing a unique index. Second, the collaborative acquisition of multi-source heterogeneous operational data from inertial measurement units (IMUs) and the construction of a joint dataset aim to build a standardized benchmark dataset with aligned physical features and unified data modes. This enables IMUs of different models and flight conditions to establish strict mathematical mapping relationships across four dimensions: operational parameters, failure characteristics, environmental conditions, and health status. This provides a learnable sample foundation for subsequent health status identification models. The system acquires and standardizes multi-source time-series operational data of the inertial measurement unit (IMU). Through time synchronization, dimension normalization, and noise suppression, data heterogeneity is eliminated to obtain a standardized time-series operational data sequence. Then, the system performs parallel extraction of equipment failure features and characterization of environmental operating conditions. Through various data analysis methods, it obtains multi-dimensional failure feature labels, including zero-bias drift, scaling factor error, and signal-noise characteristics. Finally, the system constructs a multi-source heterogeneous joint dataset, linking the above operational, feature, operating condition, and health status data with sample indices as keys to create a health status identification dataset for model training of the IMU.

[0033] It is also important to understand that the health status classification and fault cause tracing based on multi-dimensional feature decoupling, through feature importance quantification analysis and fault tracing network, maps the extracted deep failure features into interpretable health status levels and specific fault causes, thereby significantly improving the real-time performance, accuracy and interpretability of health status identification of airborne inertial measurement units, and enhancing the operational safety and maintenance efficiency of aircraft.

[0034] The aforementioned state recognition method for airborne inertial measurement units (INS) establishes a standardized benchmark with unified data modalities and aligned physical features by collaboratively acquiring multi-source heterogeneous operational data and constructing a joint dataset. Based on a deep temporal feature learning architecture, a reverse health status recognition model for INS under all operating conditions is built. A health status feature encoding network maps multi-source temporal operational data into high-dimensional failure feature vectors, and an adaptive attention mechanism in the health status classification network enables the targeted extraction of key failure features. Furthermore, an embedded adaptive verification and feature collaborative optimization strategy is implemented, reusing the embedded condition feature verification head in the discrimination network to construct a self-supervised feedback loop for the agentless model, ensuring the accuracy and stability of health status recognition results under different flight conditions. The following will describe in more detail the various steps of the state recognition method for an airborne inertial measurement unit described in this example embodiment.

[0035] In some embodiments, reference Figure 3 As shown, step S101 includes: The continuous and discrete operating parameters of the airborne inertial measurement unit are obtained and a set of operating parameters is generated through normalization.

[0036] Multi-source data acquisition is performed on the operation of the airborne inertial measurement unit, and standardized time-series operation data sequences are generated through time-series synchronization and standardization processing.

[0037] The standardized time-series running data is subjected to feature analysis to construct a failure feature vector.

[0038] Construct a health status label that includes sub-labels for the location and severity of the fault.

[0039] The joint dataset is generated by associating the set of operating parameters, the standardized time-series operating data sequence, and the failure feature vector with the health status label in a one-to-one correspondence.

[0040] It is important to understand that the vectorization of operating parameters is bound to the sample index. For each measured sample of an airborne inertial measurement unit, key operating parameters during its flight process are recorded in detail to establish an operating parameter set. This operating parameter set specifically includes five continuous operating parameters: gyroscope three-axis measurement values, accelerometer three-axis output values, zero-bias drift real-time values, scaling factor error values, and signal sampling frequency; and two discrete operating parameters: equipment model and flight stage. To eliminate the interference of different physical dimensions on the model weights, the five continuous operating parameters are linearly normalized according to preset upper and lower bounds of their dimensions, mapping them to a numerical range of zero to one. At the same time, the two discrete operating parameters are one-hot encoded and converted into sparse vector format. Finally, the normalized continuous parameters and the encoded discrete parameters are concatenated to form a unique operating parameter vector for that sample, and a globally unique sample index number is assigned to it.

[0041] Multi-source time-series operational data acquisition and standardized preprocessing: Using an aircraft flight control system data acquisition module and external high-precision sensing equipment, multi-source data acquisition is performed on the operation of the airborne inertial measurement unit (INS), obtaining raw time-series operational data from gyroscopes, accelerometers, signal processing units, and environmental monitoring. Given the issues of asynchronous timing, inconsistent dimensions, and noise interference in data acquired from different acquisition devices and under different flight conditions, this application establishes a unified data preprocessing workflow. First, time-series synchronization standardization is performed, aligning timestamps and interpolating the raw multi-source time-series operational data to obtain time-series operational data with a unified time dimension. Second, dimension and amplitude normalization is performed, linearly scaling and truncation of the time-series operational data to eliminate dimensional differences and suppress outlier interference. Finally, adaptive noise suppression is performed, using a combination of wavelet transform and Kalman filtering to remove Gaussian and impulse noise from the time-series operational data, obtaining a standardized airborne inertial measurement unit time-series operational data sequence, thereby ensuring the comparability of operational characteristics between different samples in both numerical and temporal dimensions.

[0042] Equipment Failure Feature Extraction and Environmental Condition Information Characterization: To reveal deeper failure characteristics and operational constraints beyond the operating parameters of the airborne inertial measurement unit (IMU), multi-dimensional feature and operational condition characterization was performed. For failure feature extraction, time series analysis and signal processing methods were employed to analyze standardized time-series operating data, calculating five core failure features: zero-bias drift, scaling factor error, signal noise amplitude and frequency distribution, data synchronization deviation, and dynamic response delay. A failure feature vector was constructed, and this data quantitatively describes the microscopic failure state of the equipment. For environmental condition information characterization, environmental monitoring data was collected during flight, and four key environmental parameters—temperature, vibration amplitude, air pressure, and flight speed—were analyzed. An operational condition feature vector was constructed, and flight phase labels were simultaneously added to the flight logs. This data quantitatively describes the external environmental constraints of the equipment operation.

[0043] Health status index calibration and label construction: Standardized health status calibration is performed on samples of airborne inertial measurement units (IMUs) to obtain five levels of health status labels, including healthy, slightly degraded, moderately faulty, severely faulty, and failed, as well as corresponding sub-labels for fault location and fault severity. Specifically, multi-dimensional health status judgment criteria are formulated by combining aviation industry standards and equipment manufacturer technical specifications, comprehensively considering the correlation between failure characteristics such as zero-bias drift and scaling factor error and the actual measurement accuracy of the equipment. Offline full-performance testing experiments are designed to conduct comprehensive performance tests on the samples, and the health status level of the samples is determined by combining human expert judgment. For samples with faults, specific fault locations (gyroscopes, accelerometers, circuit units, etc.) and fault development levels (early, middle, late) are labeled through fault disassembly and component inspection. Finally, the health status levels are digitally encoded to construct standardized health status label vectors, which serve as classification targets for subsequent model training.

[0044] A multi-source heterogeneous joint dataset is constructed. Using the sample index number as the unique primary key, the operating parameter vector, standardized time-series operating data sequence, failure feature vector, operating condition feature vector, and health status label vector are associated one-to-one to construct an airborne inertial measurement unit health status identification dataset containing a four-element relationship of operation, feature, operating condition, and health status. Among them, the failure feature vector and the environmental operating condition feature vector together form the airborne inertial measurement unit operating condition-failure composite feature label. The dataset is randomly divided into training set, validation set, and test set according to a preset ratio (7:2:1) as the standard data input for the end-to-end health status identification model and the operating condition consistency identification network in subsequent steps.

[0045] In some embodiments, reference Figure 2 As shown, step S102 includes: The full-condition health status recognition model includes an encoding network, a classification network, and a discrimination network.

[0046] It is important to understand that the construction of a health status identification model for airborne inertial measurement units (INS) under all operating conditions is crucial. This step aims to establish a precise mapping relationship between multi-source time-series operational data and equipment health status. First, an INS health status feature encoding network is built to map standardized time-series operational data sequences into implicit high-dimensional failure feature embedding vectors. Second, an INS health status classification network is constructed, utilizing the collaborative modulation of failure feature embedding vectors and operating condition feature vectors to achieve accurate classification of equipment health status under different operating conditions. Finally, an INS health status identification network is constructed, employing a multi-scale feature fusion architecture to determine the validity of extracted failure features and verify operating condition consistency. Adversarial training is used to improve the accuracy and stability of feature extraction and health status classification results under all operating conditions. In some embodiments, reference Figure 4 As shown, step S102 includes: The full-condition health status identification model maps the standardized time-series operation data sequence into a failure feature embedding vector through the coding network.

[0047] It is important to understand that the health status feature encoding network is constructed. This network consists of a cascaded temporal feature extraction layer, a multi-source feature fusion layer, and a semantic mapping layer. First, a standardized temporal operational data sequence is acquired, containing temporal operational data from six channels, including gyroscopes and accelerometers. This sequence is then input into the temporal feature extraction layer, which comprises stacked gated recurrent units (GRUs) and one-dimensional convolutional layers. By jointly extracting the temporal and spatial features of the temporal data, the inherent correlations and temporal evolution patterns between operational data from different channels are explored, outputting preliminary temporal fusion features. Subsequently, the temporal fusion features and operational condition feature vectors are input into the multi-source feature fusion layer. Through cross-dimensional feature concatenation and attention-weighted fusion, the interference of environmental conditions on the operational features is eliminated, outputting operational condition-adapted fusion features. Finally, the fusion features are input into the semantic mapping layer, projecting the low-dimensional fusion features onto a high-dimensional feature space, outputting a semantically rich airborne inertial measurement unit failure feature embedding vector. This vector serves as the core input to the subsequent classification network, determining the feature basis for health status recognition.

[0048] In some embodiments, reference Figure 4 As shown, step S102 includes: The full-condition health status identification model uses the classification network to perform condition-adaptive correction on the failure feature embedding vector and outputs the probability distribution of the health status of the airborne inertial measurement unit.

[0049] It is important to understand that the health status classification network is constructed using an adaptive attention-based temporal feature decoding architecture. This architecture comprises a feature modulation layer, three cascaded adaptive attention decoding modules, a global feature pooling layer, and a classification output layer. First, the failure feature embedding vector is input to the feature modulation layer. Combined with the operating condition feature vector, feature modulation parameters are dynamically generated to adaptively correct the failure feature embedding vector, resulting in the operating condition-modulated failure feature vector. Second, the operating condition-modulated failure feature vector is input to the cascaded adaptive attention decoding modules, each containing attention weights. The system consists of a generation layer and a feature decoding layer. During feature transfer, the attention weight generation layer dynamically generates attention weights in the time and channel dimensions based on the classification requirements of the device's health status. This allows for targeted extraction of failure features, highlighting key failure features and suppressing irrelevant interference features. After three rounds of attention decoding and feature optimization, a high-dimensional feature map containing core failure information is obtained. Finally, a global feature pooling layer reduces the dimensionality of the high-dimensional feature map, extracting the global core feature vector. This vector is then input into the classification output layer and processed by a Softmax classifier to output the probability distribution of the health status of the inertial measurement unit, achieving a preliminary classification of the health status.

[0050] In some embodiments, reference Figure 4 As shown, step S102 includes: The full-condition health status recognition model performs comprehensive condition consistency supervision through the discrimination network, generates condition feature verification headers, and adds them to the health status classification results.

[0051] It is important to understand that the health status identification network is constructed. The health status identification network of the airborne inertial measurement unit adopts a multi-scale feature fusion identification architecture, which includes three identification sub-networks at different scales, a global temporal pooling layer, and a condition feature verification head. The failure feature embedding vector and the real failure feature vector are simultaneously decomposed into three feature versions with time scales of the original sequence, 1 / 2 sequence, and 1 / 4 sequence, which are respectively input into the three identification sub-networks. Each identification sub-network consists of stacked one-dimensional convolutional layers, spectral normalization layers, and activation functions, which are responsible for capturing the temporal evolution law, channel correlation features, and abnormal mutation features of failure features at different temporal receptive field scales, and independently output the validity discrimination results corresponding to the extracted failure feature vector and the real failure feature vector. The discrimination results output by the three identification sub-networks are not merged into a single score, but are weighted and summed after calculating the loss with the real label and the fake label, thereby applying adversarial constraints at three levels: fine temporal features, local correlation features, and global evolution features.

[0052] Meanwhile, in order to integrate multi-scale features and achieve comprehensive monitoring of operational conditions, the feature maps of different scales output from the three discriminator sub-networks are respectively input into a global temporal pooling layer for pooling processing to obtain three pooled feature vectors of different scales. The three pooled feature vectors are concatenated along the channel dimension to form a global abstract feature vector representing the operating status of the equipment. This global abstract feature vector is input into the operational condition feature verification head, and the corresponding environmental operational parameters are predicted through regression by a fully connected layer. This forces the health status feature encoding network to not only follow the failure mechanism logic of the equipment when extracting failure features, but also to conform to the distribution law of operating features under different flight conditions, ensuring that the extracted failure features are physically credible and have strong adaptability at the operational condition level.

[0053] In some embodiments, reference Figure 2 As shown, step S103 includes: The loss function of the discrimination network is obtained by summing the loss of the true feature discrimination, the loss of the extracted feature discrimination, and the loss of the working condition regression supervision according to the preset weights. The loss function for both the encoding and classification networks is obtained by summing the anti-spoofing loss, the condition consistency loss, and the classification accuracy loss according to preset weights.

[0054] The full-condition health status recognition model is optimized and trained using a dynamic game strategy that alternately updates the encoding network, classification network, and discrimination network.

[0055] It is important to understand that this step aims to establish an end-to-end operational condition constraint mechanism without the need for an independent proxy model. It reuses the health status identification network of the airborne inertial measurement unit (IMU) and utilizes the operational condition feature verification head embedded in its architecture to achieve the dual functions of feature validity judgment and operational condition consistency verification. This allows for simultaneous feature extraction optimization and health status classification accuracy improvement within a single adversarial training cycle. First, a composite objective function is constructed, comprising adversarial loss, operational condition regression loss, and feature consistency loss. Second, adversarial training based on alternating optimization is executed. The operational condition regression gradient returned by the health status identification network directly guides the health status feature encoding network and classification network to adjust feature extraction and classification parameters, ensuring the accuracy and stability of health status identification results under different flight conditions.

[0056] A composite objective function based on operational condition verification is constructed. To quantify model performance and provide optimization gradients during training, a composite objective function based on operational condition verification is constructed, specifically including the following calculation process: For the health status identification network of the airborne inertial measurement unit, a discriminator loss function is defined and calculated. This loss function is composed of a weighted sum of feature validity discrimination error and operational condition regression error. First, the true failure feature vector is input into the identification network to obtain its corresponding validity discrimination result. The cross-entropy loss between this result and the true label is calculated to obtain the true feature discrimination loss. Simultaneously, the failure feature embedding vector is input into the identification network to obtain its corresponding validity discrimination result. The cross-entropy loss between this result and the forged label is calculated to obtain the extracted feature discrimination loss. Second, the operational condition feature verification head in the identification network is used to predict the true failure feature vector, outputting the predicted operational condition feature vector. The mean square error between the predicted vector and the true operational condition feature vector is calculated to obtain the operational condition regression supervision loss. Finally, the true feature discrimination loss, the extracted feature discrimination loss, and the operational condition regression supervision loss are summed according to preset weights to obtain the identification network loss function.

[0057] For the health status feature encoding network and classification network of an airborne inertial measurement unit (INS), a recognition loss function is defined and calculated. This loss function consists of adversarial deception loss, operational condition consistency loss, and classification accuracy loss. First, the failure feature embedding vector is input into the health status identification network to obtain its corresponding validity judgment result. The cross-entropy loss between this result and the true label is calculated and used as the adversarial deception loss. This loss forces the extracted failure features to approximate the distribution pattern of the true failure features. Second, the operational condition feature verification head is used to predict the extracted failure feature embedding vector, outputting a predicted operational condition feature vector. The mean square error between this predicted vector and the true operational condition feature vector is calculated and used as the operational condition consistency loss, ensuring that the extracted failure features match the actual flight operational conditions. Finally, the cross-entropy loss between the output of the classification network and the health status label vector is used as the classification accuracy loss, ensuring the model's health status classification ability. Finally, the adversarial deception loss, operational condition consistency loss, and classification accuracy loss are summed according to preset weights to obtain the recognition loss function. The system employs condition-oriented alternating adversarial training. It abandons the pre-training step of freezing parameters and adopts a dynamic game strategy where the health status feature encoding network, classification network, and discrimination network are updated alternately. In each training iteration, the parameters of the feature encoding and classification networks are first fixed, while the parameters of the discrimination network are updated to enable it to accurately identify non-physical artifacts and condition mismatches in the extracted features and calibrate its condition prediction benchmark. Subsequently, the parameters of the discrimination network are fixed, while the parameters of the feature encoding and classification networks are updated. During this stage, the health status discrimination network acts as a differentiable condition environment simulator. The condition regression error calculated by the condition feature verification head is transformed into a gradient flow that guides feature adjustment through backpropagation. This gradient flow is directly fed back to the feature encoding and classification networks, driving them to directionally correct key parameters such as temporal feature extraction weights, attention allocation coefficients, and classifier weights. As training progresses, the feature encoding network gradually learns to extract real and effective equipment failure features under different flight conditions, while the classification network learns to accurately classify the health status based on the extracted key failure features. Ultimately, this achieves high-precision health status identification of the airborne inertial measurement unit under all conditions without agent-assisted model support.

[0058] In addition, the gradient flow generated by the health status identification network's condition feature verification head will also propagate back to the multi-source feature fusion layer of the feature encoding network, driving its weight parameters to be updated synchronously. This enables the network to learn to dynamically adjust the fusion ratio of operational features and condition features according to different flight conditions, further improving the model's condition adaptability.

[0059] In some embodiments, reference Figure 2 As shown, step S104 includes: Real-time acquisition of multi-source operational data from airborne inertial measurement units through local deployment.

[0060] The trained, full-condition-oriented airborne inertial measurement unit (INS) health status identification model outputs the real-time collected multi-source operational data as the INS health status classification results.

[0061] The health status classification results are decoupled and output as independent single failure features.

[0062] It is important to understand that this step involves deploying the trained, full-condition-oriented airborne inertial measurement unit (INS) health status identification model to the aircraft flight control system or ground maintenance platform. This enables end-to-end intelligent analysis, from multi-source real-time operational data to health status classification and fault cause tracing. First, real-time data acquisition and online feature extraction are performed to obtain key failure features during the dynamic operation of the INS. Second, accurate health status determination based on dual discrimination criteria is performed, using a trained health status identification network and classification network to comprehensively analyze the extracted failure features and output high-precision health status classification results. Third, key failure feature decoupling and importance quantification are performed, decoupling and analyzing the extracted high-dimensional failure features to identify the core failure features with the greatest impact on health status. Finally, fault cause tracing and development trend prediction are performed, constructing a fault tracing network to map the core failure features to explainable fault causes and development trends, outputting a complete analysis report including health status level, fault location, fault cause, and early warning suggestions.

[0063] The system performs real-time operational data acquisition and online feature extraction. Users can acquire multi-source operational data from the aircraft's flight control system or ground maintenance platform in real time. The system first calls a standardized preprocessing procedure to perform time synchronization, dimension normalization, and noise suppression on the real-time acquired multi-source operational data from gyroscopes, accelerometers, etc., to obtain a standardized real-time time-series operational data sequence. Then, it calls a trained health status feature encoding network to map the standardized real-time time-series operational data sequence into a real-time failure feature embedding vector rich in semantic information, completing the online extraction from real-time operational data to deep failure features, meeting the real-time requirements of the aviation field.

[0064] The system performs precise health status determination based on a dual-criteria approach. To achieve high-precision health status determination of the airborne inertial measurement unit (INS), the system utilizes a pre-trained health status discrimination network and a classification network as a comprehensive decision maker. The system sequentially inputs the real-time failure feature embedding vectors into the discrimination and classification networks, performing parallel analysis in two dimensions: First, the discrimination network judges the validity and consistency of the real-time failure features with the operating conditions, outputting a feature validity score to eliminate abnormal features caused by sudden environmental interference, ensuring the reliability of the identification results. Second, the classification network classifies the real-time failure features into health status categories, outputting a probability distribution of five health status levels: healthy, slightly degraded, moderately faulty, severely faulty, and failed. The system selects the level corresponding to the highest probability as the preliminary health status determination result. The system then weights and fuses the feature validity score with the probability value of the preliminary determination result according to preset weights, outputting the final INS health status classification result, ensuring the accuracy and reliability of the determination result.

[0065] The system performs key failure feature decoupling and importance quantification. To give the health status assessment results an interpretable physical meaning, the system uses feature decoupling and importance quantification algorithms to perform post-processing analysis on the real-time failure feature embedding vector. First, the high-dimensional failure feature embedding vector is decoupled using a variational autoencoder, decomposing it into mutually independent single failure feature dimensions, including gyroscope zero-bias drift, accelerometer scaling factor error, signal noise, and data synchronization deviation. Then, the SHAP value feature importance analysis algorithm is used to calculate the contribution of each decoupled single failure feature to the health status assessment result, generating a feature importance ranking. The system automatically selects the top N single failure features with the highest contribution as core failure features. These core failure features directly reflect the key reasons for changes in equipment health status, providing a clear analytical direction for subsequent fault tracing.

[0066] In some embodiments, step S104 includes: By analyzing the single failure characteristics after decoupling, the system outputs a report on the cause of failure and the prediction of its development trend.

[0067] It is important to understand that the process of tracing the cause of failures, predicting their development trends, and outputting reports is crucial. This step aims to address the problem of locating and warning of failures following abnormal health conditions in airborne inertial measurement units (IMUs). A lightweight fault tracing and trend prediction network is constructed and utilized. This network takes core failure feature vectors as input and includes a feature analysis layer, a fault tracing layer, and a trend prediction layer. Through supervised training on the aforementioned dataset, it has mastered the nonlinear correspondence between core failure features and failure modes and development trends. The feature analysis layer performs in-depth analysis of the core failure features to uncover the underlying equipment operating mechanisms. Based on the analysis results, the fault tracing layer outputs the specific fault location (gyroscope, accelerometer, circuit unit, signal transmission). The system analyzes the health status of an inertial measurement unit (IMU) based on various factors, including links, specific causes of failure (component wear, fatigue aging, poor contact, environmental interference, etc.), and the stage of failure development (early, middle, and late stages). The trend prediction layer, based on the time evolution of core failure characteristics, uses a time series prediction algorithm to predict the future trend of equipment health status and provides corresponding fault warning levels and maintenance recommendations. Finally, the system integrates the health status classification results, core failure characteristics, fault cause tracing results, trend predictions, and maintenance recommendations into a single health status analysis report that directly guides aviation maintenance work. This report is pushed in real-time to the flight control system or ground maintenance platform, providing decision support for the safe operation and precise maintenance of aircraft.

[0068] Before deployment, the fault tracing and trend prediction network needs to undergo independent supervised training: using the health status recognition dataset of airborne inertial measurement units, the failure feature vector is extracted as input data, and the health status label, fault location, and fault development stage are extracted as label data to construct a multi-task prediction model; during the training process, a combination loss function of cross-entropy loss and mean squared error loss is used to calculate the deviation between the prediction result and the true label, and the backpropagation algorithm is used to update the network weights until the model can accurately achieve fault tracing and trend prediction.

[0069] In conjunction with the above-mentioned state recognition method for airborne inertial measurement units, and in order to fully verify the effectiveness and reliability of the airborne inertial measurement unit health status recognition method proposed in this invention for all operating conditions, an intelligent health status recognition simulation experimental platform for airborne inertial measurement units was constructed.

[0070] The platform fully replicates the data acquisition process, including a multi-source operational data acquisition module, a failure feature extraction module, an environmental condition simulation module, and a health status calibration module. The experiment collected 8,000 sets of sample data from different models and flight conditions of airborne inertial measurement units, focusing on evaluating three core indicators: the model's health status recognition accuracy, recognition stability under different conditions, and early fault recognition capability.

[0071] I. Experimental Analysis of Health Status Recognition Accuracy This experiment aims to quantitatively evaluate the accuracy of the constructed health status recognition model for airborne inertial measurement units (IMUs) in identifying the actual health status of the equipment. In the experiment, multi-source time-series operational data from the test set were randomly selected as input. The health status recognition model outputs health status judgment results, and the matching degree between the judgment results and the actual health status labels is calculated. Simultaneously, the recognition accuracy, precision, and recall rates for different health status levels are statistically analyzed.

[0072] refer to Figure 5 As shown, experimental results demonstrate that the method of this invention achieves an overall accuracy of 98.2% in identifying the health status of airborne inertial measurement units (INS). The accuracy rates for identifying both healthy and failed states exceed 99%, while the accuracy rates for identifying mild attenuation, moderate faults, and severe faults reach 97.5%, 98.1%, and 98.7%, respectively. The precision and recall rates for each level remain above 97%. Compared to traditional threshold-based methods and single machine learning classification methods, the accuracy of this invention is improved by 23.5% and 12.8%, respectively, fully demonstrating that the deep temporal feature learning architecture constructed in this invention can effectively extract deep failure features from multi-source heterogeneous operational data, achieving high-precision identification of health status.

[0073] II. Comparative Analysis of Recognition Stability under Different Working Conditions This experiment focuses on examining the stability of the model's health status recognition under different flight conditions. A condition complexity coefficient was defined; a larger coefficient indicates more complex interference factors in the flight environment, such as temperature and vibration, and stronger dynamics during the flight phase. The experiment compared the coefficient of variation of the recognition accuracy of the proposed method with that of traditional methods under different condition complexity coefficients; a smaller coefficient of variation indicates stronger stability.

[0074] refer to Figure 6 As shown in the experimental results, under low operating condition complexity, the recognition stability of the proposed method is not significantly different from that of the traditional method. However, with the increase of the operating condition complexity, the coefficient of variation of the recognition accuracy of the traditional method increases exponentially, and the recognition stability drops sharply. This is because the traditional method lacks operating condition adaptability and struggles to handle the feature coupling problem caused by environmental interference and changes in operating conditions. In contrast, the coefficient of variation of the recognition accuracy of the proposed method remains below 3%, with a gradual downward trend, maintaining extremely high recognition stability even under high operating condition complexity. This is attributed to the adaptive attention decoding architecture and multi-scale feature fusion identification mechanism adopted. The former ensures that the model can dynamically focus on key failure features and suppress environmental interference during feature extraction, while the latter guarantees the physical consistency and effectiveness of the extracted failure features under different operating conditions. This demonstrates that the proposed method has excellent adaptability to all operating conditions and can meet the health status recognition requirements of aircraft in complex flight environments.

[0075] III. Experimental Analysis of Early Fault Identification Capability This experiment aims to verify the ability of the method of this invention to identify early, minor faults in airborne inertial measurement units (INS), which is one of the core requirements of airborne equipment health management. The experiment selected 100 INS samples with early, minor faults and used both the method of this invention and traditional methods to identify their health status. The identification rate of early faults was statistically analyzed, and the fault warning lead time was calculated.

[0076] refer to Figure 7 As shown in the experimental results, the method of this invention achieves a 96% recognition rate for early minor faults in airborne inertial measurement units (INS), accurately identifying mild degradation states in the early stages of equipment performance decline, with an average fault warning lead time of 210 hours, providing ample time for fault investigation and equipment maintenance in aviation operations. In contrast, traditional methods only achieve a 35% recognition rate for early minor faults, with most early faults being classified as healthy states, failing to provide effective fault warnings. This fully demonstrates that the high-dimensional failure feature extraction and analysis system constructed by the method of this invention can effectively capture minor failure features during the early operation of equipment, achieving accurate identification and early warning of early faults, significantly improving the foresight and effectiveness of health management of airborne INS, and providing important protection for the flight safety of aircraft.

[0077] In the description of this specification, the references to terms such as "one embodiment," "some embodiments," "example," "specific example," or "some examples," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of the invention. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples. In addition, those skilled in the art can combine different embodiments or examples described in this specification.

[0078] Other embodiments of the invention will readily occur to those skilled in the art upon consideration of the specification and practice of the invention disclosed herein. This invention is intended to cover any variations, uses, or adaptations of the invention that follow the general principles of the invention and include common knowledge or customary techniques in the art not disclosed herein. The specification and examples are to be considered exemplary only, and the true scope and spirit of the invention are indicated by the appended claims.

Claims

1. A state identification method for an airborne inertial measurement unit, characterized in that, include: The operational data of the airborne inertial measurement unit is acquired and converted into a joint dataset according to a unified standard mapping relationship; wherein the standard mapping relationship includes four dimensions: operating parameters, failure characteristics, environmental conditions and health status. Feature extraction and coordinated modulation are performed on the joint dataset, and the health status classification results under different working conditions are output through the full-condition health status recognition model; wherein, the health status classification results include a working condition feature verification header; The operating condition feature verification head is used to perform feature validity discrimination and operating condition consistency verification. The composite objective function of adversarial loss, operating condition regression loss and feature consistency loss is used to optimize the full operating condition health status identification model, thereby generating an airborne inertial measurement unit health status identification model for all operating conditions. The trained health status recognition model for airborne inertial measurement units (INS) under all operating conditions is deployed locally, and the health status classification results of INS are output in real time.

2. The state identification method for an airborne inertial measurement unit according to claim 1, characterized in that, The steps of acquiring operational data from the airborne inertial measurement unit and converting it into a joint dataset according to a unified standard mapping relationship include: The continuous and discrete operating parameters of the airborne inertial measurement unit are acquired and normalized to generate a set of operating parameters. Multi-source data acquisition is performed on the operation of the airborne inertial measurement unit, and standardized time-series operation data sequences are generated through time-series synchronization and standardization processing. The standardized time-series operational data is subjected to feature analysis to construct a failure feature vector; Construct a health status label that includes sub-labels for fault location and fault severity; The joint dataset is generated by associating the set of operating parameters, the standardized time-series operating data sequence, and the failure feature vector with the health status label in a one-to-one correspondence.

3. The state identification method for an airborne inertial measurement unit according to claim 2, characterized in that, The steps of performing feature extraction and coordinated modulation on the joint dataset, and outputting health status classification results under different working conditions through the full-condition health status identification model, include: The full-condition health status recognition model includes an encoding network, a classification network, and a discrimination network.

4. The state identification method for an airborne inertial measurement unit according to claim 3, characterized in that, The steps of performing feature extraction and coordinated modulation on the joint dataset, and outputting health status classification results under different working conditions through the full-condition health status identification model, include: The full-condition health status identification model maps the standardized time-series operation data sequence into a failure feature embedding vector through the coding network.

5. The state identification method for an airborne inertial measurement unit according to claim 4, characterized in that, The steps of performing feature extraction and coordinated modulation on the joint dataset, and outputting health status classification results under different working conditions through the full-condition health status identification model, include: The full-condition health status identification model uses the classification network to perform condition-adaptive correction on the failure feature embedding vector and outputs the probability distribution of the health status of the airborne inertial measurement unit.

6. The state identification method for an airborne inertial measurement unit according to claim 5, characterized in that, The steps of performing feature extraction and coordinated modulation on the joint dataset, and outputting health status classification results under different working conditions through the full-condition health status identification model, include: The full-condition health status recognition model performs comprehensive condition consistency supervision through the discrimination network, generates condition feature verification headers, and adds them to the health status classification results.

7. The state identification method for an airborne inertial measurement unit according to claim 6, characterized in that, The step of optimizing the full-condition health status identification model to generate a full-condition airborne inertial measurement unit health status identification model includes: The loss function of the discrimination network is obtained by summing the loss of the true feature discrimination, the loss of the extracted feature discrimination, and the loss of the working condition regression supervision according to the preset weights; The anti-deception loss, working condition consistency loss, and classification accuracy loss are summed according to preset weights to obtain the common loss function of the encoding network and the classification network. The full-condition health status recognition model is optimized and trained using a dynamic game strategy that alternately updates the encoding network, classification network, and discrimination network.

8. The state identification method for an airborne inertial measurement unit according to any one of claims 1-7, characterized in that, The steps of locally deploying the trained, full-condition-oriented airborne inertial measurement unit (INS) health status recognition model and outputting the INS health status classification results in real time include: Real-time acquisition of multi-source operational data from airborne inertial measurement units through local deployment; The trained, full-condition-oriented airborne inertial measurement unit (INS) health status identification model outputs the real-time collected multi-source operational data as the INS health status classification results. The health status classification results are decoupled and output as independent single failure features.

9. The state identification method for an airborne inertial measurement unit according to claim 8, characterized in that, The steps of locally deploying the trained, full-condition-oriented airborne inertial measurement unit (INS) health status recognition model and outputting the INS health status classification results in real time include: By analyzing the single failure characteristics after decoupling, the system outputs a report on the cause of failure and the prediction of its development trend.

Images