Deep learning micro-fault diagnosis method and system considering fault time location

Abstract

The application provides a deep learning micro-fault diagnosis method and system considering fault time positioning, comprising the following steps: acquiring time series observation of a target sensor of a hypersonic aircraft to be diagnosed within T w segments of time before the current time; determining a target state variable set according to the residual between the time series observation and the corresponding observation of the nominal system of the hypersonic aircraft to be diagnosed; refining sensitive features of the target state variable set through an encoding module in a Transformer model to obtain refined time series features; inputting the refined time series features into a trained time convolution network as input data, and using the trained time convolution network to diagnose the fault of the hypersonic aircraft to be diagnosed to obtain a fault type sequence output; and determining a fault occurrence time based on the fault type sequence output. The application alleviates the technical problems of low precision and inability to locate the fault occurrence time in the prior art.

Description

Technical Field

[0001] This invention relates to the field of micro-fault diagnosis technology, and in particular to a deep learning-based micro-fault diagnosis method and system that considers fault time localization. Background Technology

[0002] High-dynamic near-space vehicles (HLPs) refer to aircraft that fly at speeds exceeding Mach 5 in the near-space region of 20–100 km. Due to their high speed and rapid maneuverability, they serve as platforms for various advanced technologies, possessing extremely strong penetration capabilities, making them among the most advanced and effective penetration weapons in the world today. With the development of modern science and technology, complex and high-precision equipment is widely used in the airborne systems of HLPs. Due to the highly disruptive aerodynamic environment and complex structural dynamic parameters, HLPs exhibit strong nonlinearity, and their airborne component failures exhibit diversity, coupling, secondaryity, and uncertainty. Once a significant failure occurs within the limits of redundancy or passive fault tolerance, the battlefield survivability, operational lifespan, and strike capability of the HLP will be greatly weakened. Significant failures often evolve from slowly changing minor faults accumulated over wartime. Therefore, accurate and real-time identification of minor faults, and designing active fault-tolerant control rates using the fault information obtained from diagnosis, can reduce or avoid the occurrence of significant failures and improve the flight reliability of HLPs.

[0003] A preferred fault diagnosis system can provide complete fault information to a fault-tolerant control system. Therefore, real-time fault diagnosis for the purpose of active fault tolerance needs to achieve: (1) fault isolation; (2) fault time location; and (3) fault identification. Existing technologies focus on fault identification but neglect the important influence of fault occurrence time. It is generally believed that the moment a fault is detected is the fault occurrence time, and the size of the fault at the instant of occurrence is used instead of the actual size at that time as the fault identification result. However, for complex nonlinear systems such as highly dynamic near-space vehicles, the diagnostic delay caused by the time it takes for the fault to manifest cannot be completely eliminated. In particular, small faults have the characteristics of indistinct fault features and are easily masked by noise and disturbances, further increasing the diagnostic delay. The development of fault size within the delay leads to significant differences. Therefore, the existing technology for diagnosing small faults in aircraft has the technical problems of low accuracy and inability to locate the fault occurrence time. Summary of the Invention

[0004] In view of this, the purpose of the present invention is to provide a deep learning-based method and system for diagnosing minor faults that takes into account fault time localization, so as to alleviate the technical problems of low accuracy and inability to locate the fault occurrence time in the prior art.

[0005] In a first aspect, embodiments of the present invention provide a deep learning-based method for diagnosing minor faults that considers fault time localization, including: obtaining the time T of the hypersonic vehicle to be diagnosed before the current moment. w The process involves: collecting temporal observations of the target sensor over a given period; determining the target state variable set by analyzing the residuals between the temporal observations and the corresponding observations of the nominal system of the hypersonic vehicle to be diagnosed; refining the target state variable set using the encoding module in the Transformer model to obtain refined temporal features; using the refined temporal features as input data for a trained temporal convolutional network to perform fault diagnosis on the hypersonic vehicle to be diagnosed, resulting in a fault type sequence output; and determining the fault occurrence time based on the fault type sequence output.

[0006] Further, the target state variable set is refined using sensitive features to obtain refined temporal features, including: mapping the target state variable set to a high-dimensional feature space through linear projection to obtain spatial channel refined features; adding corresponding timestamps to the spatial channel refined features and performing multi-granularity scanning and filtering operations to obtain elite features; and performing multi-head self-attention processing on the elite features to obtain the refined temporal features.

[0007] Furthermore, the target sensor includes: an inertial navigation system sensor and an embedded atmospheric data sensing system sensor; the timing observation includes: the angle-of-attack timing observation of the inertial navigation system sensor, the angle-of-attack timing observation of the embedded atmospheric data sensing system sensor, and the complementary filter joint timing observation.

[0008] Furthermore, it also includes: establishing a nominal system model, a sensor model, and a minor fault model for the hypersonic vehicle to be diagnosed.

[0009] Secondly, embodiments of the present invention also provide a deep learning-based micro-fault diagnosis system considering fault time localization, comprising: an acquisition module, a determination module, a feature refinement module, and a fault diagnosis module; wherein, the acquisition module is used to acquire the time T of the hypersonic vehicle to be diagnosed before the current moment. wThe system comprises: a time-series observation of the target sensor over a period of time; a determination module, used to determine the target state variable set by the residual between the time-series observation and the corresponding observation of the nominal system of the hypersonic vehicle to be diagnosed; a feature refinement module, used to refine the target state variable set by using the encoding module in the Transformer model to obtain refined time-series features; a fault diagnosis module, used to use the refined time-series features as input data of a trained temporal convolutional network, and to use the trained temporal convolutional network to perform fault diagnosis on the hypersonic vehicle to be diagnosed, obtaining a fault type sequence output; and a fault diagnosis module, also used to determine the fault occurrence time based on the fault type sequence output.

[0010] Furthermore, the feature refinement module is also used to: map the target state variable set to a high-dimensional feature space through a linear projection operation to obtain spatial channel refinement features; add a corresponding timestamp to the spatial channel refinement features and perform multi-granularity scanning and filtering operations to obtain elite features; and perform multi-head self-attention processing on the elite features to obtain the refined temporal features.

[0011] Furthermore, the target sensor includes: an inertial navigation system sensor and an embedded atmospheric data sensing system sensor; the timing observation includes: the angle-of-attack timing observation of the inertial navigation system sensor, the angle-of-attack timing observation of the embedded atmospheric data sensing system sensor, and the timing observation combined with a complementary filter.

[0012] Furthermore, it also includes a training module for training a preset temporal convolutional network to obtain the trained temporal convolutional network.

[0013] Thirdly, embodiments of the present invention also provide an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the steps of the method described in the first aspect above.

[0014] Fourthly, embodiments of the present invention also provide a computer-readable medium having processor-executable non-volatile program code, the program code causing the processor to perform the method described in the first aspect above.

[0015] This invention provides a deep learning-based method and system for diagnosing minor faults that considers fault timing. By fusing the encoder structure of the Transformer model with Temporal Convolutional Networks (TCNs), a fundamental shift from sequential distributed processing to an attention-based memory mechanism is achieved, constructing an end-to-end deep learning network. This invention utilizes the parallel processing capabilities of TCNs to obtain a sequence output containing high-precision fault type and fault occurrence time information, alleviating the technical problems of low accuracy and inability to pinpoint fault occurrence time in existing technologies. Attached Figure Description

[0016] To more clearly illustrate the specific embodiments of the present invention or the technical solutions in the prior art, the drawings used in the description of the specific embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0017] Figure 1 A flowchart of a deep learning-based micro-fault diagnosis method considering fault time localization provided in an embodiment of the present invention;

[0018] Figure 2 This is a schematic diagram of a deep learning fault diagnosis model structure provided in an embodiment of the present invention;

[0019] Figure 3 A schematic diagram of a deep learning micro-fault diagnosis system that considers fault time localization provided in an embodiment of the present invention;

[0020] Figure 4 This is a schematic diagram of another deep learning-based micro-fault diagnosis system that considers fault time localization, provided as an embodiment of the present invention. Detailed Implementation

[0021] The technical solution of the present invention will now be clearly and completely described with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0022] Example 1:

[0023] Figure 1 This is a flowchart illustrating a deep learning-based micro-fault diagnosis method considering fault time localization, provided by an embodiment of the present invention. Figure 1 As shown, the method specifically includes the following steps:

[0024] Step S102: Obtain the time T before the current moment for the hypersonic vehicle to be diagnosed. w The time-series observations of the target sensor over a period of time. Wherein, T w This is a preset time period.

[0025] Optionally, the target sensor includes: an inertial navigation system sensor and an embedded atmospheric data sensing system sensor; the timing observation includes: an angle-of-attack timing observation of the inertial navigation system sensor, an angle-of-attack timing observation of the embedded atmospheric data sensing system sensor, and a joint timing observation of a complementary filter. The joint timing observation of the complementary filter is obtained by passing the angle-of-attack timing observations of the inertial navigation system sensor and the embedded atmospheric data sensing system sensor through a complementary filter.

[0026] Step S104: Determine the target state variable set by the residual between the time-series observations and the corresponding observations of the nominal system of the hypersonic vehicle to be diagnosed.

[0027] Step S106: The target state variable set is refined by the encoding module in the Transformer model to obtain refined temporal features.

[0028] Step S108: The refined temporal features are used as input data for the trained temporal convolutional network. The trained temporal convolutional network is used to perform fault diagnosis on the hypersonic vehicle to be diagnosed, and the fault type sequence is output.

[0029] Step S110: Determine the time of occurrence of the fault based on the fault type sequence output.

[0030] This invention provides a deep learning-based method for diagnosing minor faults that considers fault timing. By fusing the encoder structure of a Transformer model with Temporal Convolutional Networks (TCNs), it achieves a fundamental shift from sequential distributed processing to an attention-based memory mechanism, constructing an end-to-end deep learning network. This invention leverages the parallel processing capabilities of TCNs to obtain a sequence output containing high-precision fault type and fault occurrence time information, alleviating the technical problems of low accuracy and inability to pinpoint fault occurrence time in existing technologies.

[0031] In this embodiment of the invention, the invention further includes: establishing a nominal system model, a sensor model, and a minor fault model for the hypersonic vehicle to be diagnosed.

[0032] Specifically, in order to obtain sensor data under normal and various minor fault conditions that approximate the actual working conditions of the aircraft, a high-fidelity model of the hypersonic aircraft control system and a common fault model are established.

[0033] Hypersonic vehicles are characterized by complex flight environments, highly variable aerodynamic parameters, and high Mach numbers. A control-oriented attitude system model for a hypersonic vehicle was established.

[0034]

[0035] In the formula, g is the local gravitational acceleration of the hypersonic vehicle, α, β, and μ are the angle of attack, sideslip angle, and velocity roll angle, respectively, and ω x ω y ω z Let L and C represent the three-axis attitude angular velocities of the hypersonic vehicle, and L and C represent the lift and lateral forces, respectively. x I y I z These are the three-axis rotational inertia of the hypersonic vehicle. l A m A n A These represent the roll, yaw, and pitch aerodynamic moments experienced by a hypersonic vehicle, M. RCSx M RCSy M RCSz These are the roll, yaw, and pitch moments provided by the Reaction Control System (RCS), respectively. V is the flight speed of the hypersonic vehicle, and θ is the speed of the hypersonic vehicle. c , where is the trajectory inclination angle of the hypersonic vehicle in the inertial coordinate system, and m is the mass of the hypersonic vehicle. , , , , , α, β, μ, ω respectively x ω y ω z The derivative of .

[0036] The attitude state formula is rewritten as an affine nonlinear attitude model and linearized with small perturbations to obtain a linearized state-space model:

[0037]

[0038] Where y=x=[α,β,μ,ω] x ,ω y ,ω z ] T u=[ M RCSxM RCSy M RCSz ] T Both A and B are Jacobian matrices, which can be represented as: and , where A 11 A 12 A 21 A 22 B2 and B2 are both known quantities.

[0039] Organized into a hypersonic vehicle attitude multiple-input multiple-output (MIMO) system model:

[0040]

[0041] Where x1=[α, β, μ]T, x2= [ω x , ω y , ω z ] T.

[0042] Based on the attitude linear model, an RCS attitude control system for a hypersonic vehicle was designed. The RCS system is used as the actuator, and the terminal sliding mode control attitude control algorithm and the control allocation group method are adopted for attitude control. The RCS system consists of eight on / off jet thrusters with constant thrust. Among them, thrusters 1, 2, 3, and 4 are parallel to the coordinate axis, while thrusters 5, 6, 7, and 8 are oblique thrusters.

[0043] The PWPF modulation method is applied to perform command modulation for RCS system control allocation, and the thruster model for RCS system control allocation is established as follows:

[0044]

[0045] Where F C It is the ideal variable thrust obtained in the control allocation algorithm, F RCS This represents the constant thrust generated by the thruster. E is the input of the PWPF modulation, UM is the amplitude of the modulated pulse, Km and τm are the transfer function constants of the inertial element, and Uon and Uoff are the switching voltages. Ton and Toff are the modulated start-up and shutdown times.

[0046] Further design of the terminal sliding mode attitude controller for the hypersonic vehicle, with the sliding surface as follows:

[0047] S=C1·[α,β,μ] T +C2·[ω x ,ω y ,ω z ] T + C3([α,β,μ] T ) q / p

[0048] The sliding mode control law is:

[0049] M c =-(C2B2) -1 [C1A 11 [α,β,μ] T + C1A 12 [ω x ,ω y ,ω z ] T + C2A 21 [ω x ,ω y ,ω z ] T + C2A 22 [ω x ,ω y ,ω z ] T +(q / p)C3diag(α (q-p) / p β (q-p) / p μ (q-p) / p (A) 11 [α,β,μ] T + A 12 [ω x ,ω y ,ω z ] T )+ κ 1(S / ||S||)]

[0050] Where p and q are both odd numbers, and 2q>p. C1, C2, and C3 are coefficient matrices, and ||| represents the norm. κ 1 is a positive number.

[0051] In the attitude control system, the observations are sensor outputs. Without loss of generality, for the longitudinal channel of a hypersonic vehicle, the sensor system considering the measurement angle of attack consists of an Inertial Navigation System (INS) and an Embedded Air Data Sensing System (FADS).

[0052] INS utilizes the velocity components measured in the geographic coordinate system. The attitude angles are used to calculate the three-axis velocity components in the carrier coordinate system, and finally, the angle of attack α is calculated based on the velocity in the carrier coordinate system. INS :

[0053]

[0054] In the formula, L represents the velocity components in the carrier coordinate system. n Let be the direction cosine matrix. For sensor measurements, rand(1)∈[0,1] represents random values, and V noise This represents the amplitude of random noise.

[0055] FADS calculates the aircraft's attitude information by measuring the total pressure at different positions on the nose. The FADS sensor model is established as follows: α FADS =α+(2·rand(1)-1)·α noise ;

[0056] In the formula, α noise This represents the amplitude of random noise.

[0057] The sensor system employs a complementary filter to fuse FADS and INS data, obtaining a true and unbiased atmospheric dataset. The complementary filter is designed as follows:

[0058]

[0059] Where τ is the time constant of the complementary filter, determined by the natural frequencies of INS and FADS. α f The output of the complementary filter is s, where s is a complex parameter variable.

[0060] Establish a micro-fault model for a hypersonic vehicle, where t f The time of the failure. , , It is a constant. , For constant gain:

[0061] INS sensor data deviation fault:

[0062]

[0063] INS sensor gain variation fault:

[0064]

[0065] FADS sensor data deviation fault:

[0066]

[0067] RCS thruster thrust slowly decreasing fault:

[0068]

[0069] RCS thruster switch delay increased fault:

[0070]

[0071] The impact of minor faults on the control system is reflected in changes in the observed quantities. Therefore, in this invention, the set of state variables is defined as the residuals between the joint sensor observations (INS sensor angle of attack observations, FADS sensor angle of attack observations, and complementary filter observations) and the corresponding observations of the nominal system. The nominal system does not consider system disturbances or sensor noise. Specifically, by using the wavelet packet energy characteristics of the sensor time-series data within the Tw time interval before the current time (t=0) as the multidimensional time-frequency characteristics of the current time, preliminary mechanistic features are generated. Then, the set of state variables at time t can be expressed as:

[0072] X(t)=[E(α INS (t),n), E(α FADS (t),n), E(α f (t),n)]

[0073] Here, E(α,n) represents the energy feature of the wavelet packet decomposition to the nth layer. Thus, the fusion of sensor data at the feature layer is achieved.

[0074] Optionally, step S106 further includes the following steps:

[0075] Step S1061: Through linear projection operation, the target state variable set is mapped to a high-dimensional feature space to obtain spatial channel refinement features.

[0076] Step S1062: Add corresponding timestamps to the spatial channel refinement features, and perform multi-granularity scanning and filtering operations to obtain elite features.

[0077] Step S1063: Perform multi-head self-attention processing on the elite features to obtain refined temporal features.

[0078] Specifically, in this embodiment of the invention, after obtaining the multi-source fusion sensor feature input of the observation residuals, the original input data is mapped to a high-dimensional feature space through a linear projection operation to achieve feature refinement in the spatial channel:

[0079]

[0080] Among them W x Let be the weight coefficient matrix, and i be the index of the data in the dataset. t Let be the set of state variables at time t.

[0081] The Encoder distinguishes the temporal features of parallel processing through a "positional encoding" operation. This operation is applied to encode each historical and future moment, adding a corresponding timestamp to each residual data to be embedded. The timestamp is defined as follows:

[0082]

[0083] This ensures the uniqueness of a specific moment within the total time step length of the data. D represents the dimension of the Encoder output vector space, d = 1, 2, ..., D. The timestamp has different values ​​at different positions in the output vector, k... c Indicates the position of the timestamp, therefore This indicates that the output vector contains vectors located at 2k... c The timestamp value of the location, This indicates that the output vector contains vectors located at 2k... c The value of the timestamp at position -1.

[0084] Minor faults are not obvious in their early stages and are easily overlapped with noise characteristics in the time and frequency domains. Fault diagnosis systems are expected to detect early signs of faults using limited observational data. To achieve this goal, this invention designs a multi-granularity scanning layer, which scans data through a sliding window of length T. w Energy input characteristics of the signal obtained by intercepting the time window By refining the time channels, the problem of the encoder only focusing on the temporal data relationships at the entire data scale is solved, and the encoder's performance in terms of the correlation of temporal data at different scales is enhanced.

[0085] Multi-granularity scanning layer utilizes scanning filters A 1D convolution operation is performed on the input timestamped tensor along the time dimension. k and p are granularity parameters, where k represents the kernel size and p represents the padding size. For a specific set of s granularities, the scanning filter processes the input tensor at each scale to obtain a tensor with a time dimension of... The tensor is used, with D filters set simultaneously at each granularity to ensure multi-sensor feature dimensionality. After the temporal features are refined through multi-granularity, elite features are selected as the final processing result.

[0086]

[0087] It should be noted that the above formula restricts the correspondence between k and p: .

[0088] After passing through a linear projection layer and a multi-granularity scanning layer, the input temporal features are refined in both the spatiotemporal domains. The resulting distributed representation is then subjected to multi-head self-attention processing, implemented through parallel computation of h self-attention modules. For j self-attention modules, the trainable hyperparameter query vector... Key point vector and value vector Each is determined by the query matrix W Q Key point matrix WK And value matrix W V This determines and together constitutes the weighting mechanism for attention:

[0089]

[0090] In the formula, d k =D / h represents the dimension of the hyperparameter matrix. After each self-attention module is computed, parallel attention computation is applied to integrate information from different representation subspaces:

[0091]

[0092] In the formula, W A Let be the attention matrix, and Concat(·) be the tensor concatenation. h represents the number of self-attention modules in the multi-head self-attention module. The feedforward fully connected module consists of a linear transformation and a ReLU activation function, applied with equal weights to each observation time step.

[0093]

[0094] In the formula, The feature vectors processed by the multi-head self-attention module have now been refined, resulting in the refined features. .

[0095] In this embodiment of the invention, the TCN layer consists of L residual blocks, and each residual block consists of two dilated-causal TCN base layers. The calculations within the residual blocks can be expressed as follows:

[0096] o=Activation(s (i) +F(s (i) ))

[0097] Each residual network contains two dilated-causal TCN base layers and a nonlinear mapping. Weight normalization (WeightNorm) and Dropout layers are added after the output of each base layer for network regularization.

[0098] Specifically, when performing convolution calculations, the TCN network constructed from residual blocks extracts feature information from the test data using the following formula:

[0099]

[0100] in, Represents fault characteristic data, k TCN This represents the kernel size, and j represents the cumulative iteration factor. This indicates a filter with ReLU activation, where l represents a feature data point in the feature data set, and d represents the inflation factor, which grows exponentially, d=[2,4,…,2]. L ].

[0101] Adding a fully connected layer in the spatial dimension maps the output from the high-dimensional feature space back to the one-dimensional output space:

[0102]

[0103] Furthermore, to obtain the categorical integer output for each category, a threshold θ is set. THR ∈(0,1), the final fault type sequence is output, and the output value at each time step is:

[0104]

[0105] Similarly, the time of failure can be determined based on the output of the failure type sequence:

[0106]

[0107] At this point, the end-to-end fault diagnosis model based on multi-granularity encoder and TCN has been completed. Figure 2 This is a schematic diagram of a deep learning fault diagnosis model structure provided according to an embodiment of the present invention.

[0108] The hypersonic vehicle micro-fault diagnosis algorithm learns the nonlinear relationship between fault type and fault occurrence time and multi-element fusion observations from INS and FADS sensors by offline training of the deep learning fault diagnosis model constructed in this embodiment of the invention. The trained model will be directly applied to the online diagnosis stage to achieve accurate real-time fault separation and fault time localization in the attitude control process of hypersonic vehicles.

[0109] During offline training, the backpropagation algorithm uses the Adam optimizer to minimize the error and achieve nonlinear fitting. The loss function is defined as follows:

[0110]

[0111] In the formula Output the network classification sequence. For the category label, PairwiseDistance(X,Y) is the pixel-level Euclidean distance.

[0112] To clearly describe the purpose and technical solution of this invention, the method provided by the embodiments of this invention will be further described in detail with reference to the following simulation examples.

[0113] The initial conditions for the simulation example are x=z=0, h=33.5km, V=15Ma, and the initial values ​​of the system attitude variables are [ω]. x ,ω y ,ω z [α,β,μ]=[0,0,0,2°,0.5°,0.5°]. The desired attitude is an angle of attack, sideslip angle, and velocity roll angle of 0°. The RCS system uses eight symmetrically mounted thrusters. In the yoz plane along the body axis, the angle between the thrust F and the Z-axis is θ, and the vertical distance of the thrust F from the center of mass is d; in the xoy plane, the distance between the thrust F and the center of mass is L. Where θ=60°, d=0.3 m, L=3 m, and the fixed thrust F=1500N. The design parameters for PWPF modulation are Km=7.7856, τ m =0.12976, Uon=0.45, Uoff=0.15. INS noise V noise =0.1m / s, FADS noise α noise =0.022°.

[0114] The fault parameters are set as follows:

[0115] Under the above conditions, multiple simulations were run, with a simulation time of 3 seconds and a sampling time of 0.001 seconds. Different types of faults were injected into the simulation model at 1.5 seconds to obtain training data, and different types of faults were injected into the simulation model at different times to obtain test data.

[0116] Using a sliding window of length T w =0.6s, sliding distance is T s A sliding window of 0.02s is used to extract training data and add labels to form training and test sets. The input layer uses db1 wavelets to perform three-layer wavelet packet decomposition on the sensor signal of length and then concatenates the packets for multi-source fusion of sensor data features. The parameters in the Encoder feature extraction layer are set to... The multi-granularity scanning filter selects three granularities k=1, 3, 5, with corresponding padding sizes p=0, 1, 2. The entire encoder layer consists of four basic layers, and each multi-head self-attention module has eight attention heads. The TCN classifier consists of convolutional kernels with a size of k... TCN The model consists of 8 residual modules with a value of 3, and the hidden layer has 25 channels. During offline training, the batch size is set to 30, the maximum number of iterations is 50, and the training set to test set size is set to 8:2.

[0117] Experimental results show that the method proposed in this invention can diagnose faults and separate minor faults within two sliding time intervals. The average algorithm running time is 0.1274s, which is less than the sliding distance T. sThis avoids data stacking and meets real-time requirements. The fault location latency is 0.04s, and the fault diagnosis accuracy is 86.48%.

[0118] As described above, this invention provides a novel real-time diagnosis method for minor faults using deep learning, considering fault time localization. By integrating the Transformer's Encoder structure with a TCN, a fundamental shift from sequential distributed processing to an attention-based memory mechanism is achieved, constructing an end-to-end deep learning network. Superior to existing technologies, this invention introduces a multi-scale scanning mechanism into the Encoder to further refine spatiotemporal features. Combining the Encoder's powerful multi-dimensional spatial mapping and generalization capabilities, a distributed semantic representation of sensitive features is obtained. Furthermore, the parallel processing capabilities of the TCN are utilized to obtain a sequence output containing high-precision fault type and fault occurrence time information.

[0119] Example 2:

[0120] Figure 3 This is a schematic diagram of a deep learning-based micro-fault diagnosis system that considers fault time localization according to an embodiment of the present invention. Figure 3 As shown, the system includes: an acquisition module 10, a determination module 20, a feature refinement module 30, and a fault diagnosis module 40.

[0121] Specifically, the acquisition module 10 is used to acquire the hypersonic vehicle to be diagnosed at the current time T. w The time-series observations of the target sensor over a period of time. Wherein, T w This is a preset time period.

[0122] Optionally, the target sensor includes: an inertial navigation system sensor and an embedded atmospheric data sensing system sensor; the timing observations include: the angle-of-attack timing observations of the inertial navigation system sensor, the angle-of-attack timing observations of the embedded atmospheric data sensing system sensor, and the joint timing observations of the complementary filter.

[0123] The determination module 20 is used to determine the target state variable set by comparing the residuals between the time-series observations and the corresponding observations of the nominal system of the hypersonic vehicle to be diagnosed.

[0124] The feature refinement module 30 is used to refine the target state variable set with sensitive features through the encoding module in the Transformer model to obtain refined temporal features.

[0125] Optionally, the feature refinement module 30 is further configured to: map the target state variable set to a high-dimensional feature space through a linear projection operation to obtain spatial channel refined features; add corresponding timestamps to the spatial channel refined features and perform multi-granularity scanning and filtering operations to obtain elite features; and perform multi-head self-attention processing on the elite features to obtain refined temporal features.

[0126] The fault diagnosis module 40 is used to take the refined temporal features as input data of the trained temporal convolutional network, use the trained temporal convolutional network to perform fault diagnosis on the hypersonic vehicle to be diagnosed, and output the fault type sequence.

[0127] The fault diagnosis module 40 is also used to determine the time of fault occurrence based on the fault type sequence output.

[0128] This invention provides a deep learning-based micro-fault diagnosis system that considers fault timing. By fusing the encoder structure of the Transformer model with Temporal Convolutional Networks (TCNs), it achieves a fundamental shift from sequential distributed processing to an attention-based memory mechanism, constructing an end-to-end deep learning network. This invention utilizes the parallel processing capabilities of TCNs to obtain a sequence output containing high-precision fault type and fault occurrence time information, alleviating the technical problems of low accuracy and inability to pinpoint fault occurrence time in existing technologies.

[0129] Figure 4 This is a schematic diagram of another deep learning-based micro-fault diagnosis system that considers fault time localization according to an embodiment of the present invention. Figure 4 As shown, it also includes a training module 50, which is used to train a preset temporal convolutional network to obtain a trained temporal convolutional network.

[0130] This invention also provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements the steps of the method in Embodiment 1 described above.

[0131] This invention also provides a computer-readable medium having processor-executable non-volatile program code that causes the processor to perform the method described in Embodiment 1 above.

[0132] 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 deep learning-based method for diagnosing minute faults that considers fault time-based localization, characterized in that, include: Obtain the hypersonic vehicle under diagnosis at the current time T. w Temporal observations of the target sensor over a period of time; The target state variable set is determined by the residual between the time-series observations and the corresponding observations of the nominal system of the hypersonic vehicle to be diagnosed. The target state variable set is refined by using the encoding module in the Transformer model to obtain refined temporal features. The refined temporal features are used as input data for a trained temporal convolutional network. The trained temporal convolutional network is then used to diagnose the hypersonic vehicle to be diagnosed, and a fault type sequence is output. Based on the output of the fault type sequence, the time of fault occurrence is determined; The refinement of the sensitive features of the target state variable set includes: By using linear projection, the target state variable set is mapped to a high-dimensional feature space to obtain spatial channel refinement features; The spatial channel refinement features are stamped with corresponding timestamps, and multi-granularity scanning and filtering operations are performed to obtain elite features; The refined temporal features are obtained by performing multi-head self-attention processing on the elite features. The fault type sequence output includes: The refined temporal features are extracted using a temporal convolutional network to obtain fault feature data. A fully connected layer is added in the spatial dimension to map the fault feature data from the high-dimensional feature space back to the one-dimensional output space, resulting in a one-dimensional output sequence. Furthermore, a threshold θ is set to obtain a categorical integer output for each category. THR ∈(0,1), the final fault type sequence is output, and the output value at each time step is: ； Similarly, based on the fault type sequence output, the fault occurrence time is determined: 。 2. The method according to claim 1, characterized in that, The target sensors include: an inertial navigation system sensor and an embedded atmospheric data sensing system sensor; The timing observations include: the angle-of-attack timing observations of the inertial navigation system sensor, the angle-of-attack timing observations of the embedded atmospheric data sensing system sensor, and the joint timing observations of the complementary filter.

3. The method according to claim 1, characterized in that, Also includes: Establish the nominal system model, sensor model, and minor fault model of the hypersonic vehicle to be diagnosed.

4. A deep learning-based micro-fault diagnosis system considering fault time localization, characterized in that, include: The module includes an acquisition module, a determination module, a feature refinement module, and a fault diagnosis module; among which, The acquisition module is used to acquire the hypersonic vehicle to be diagnosed at the current time T. w Temporal observations of the target sensor over a period of time; The determining module is used to determine the target state variable set by the residual between the time-series observations and the nominal system corresponding observations of the hypersonic vehicle to be diagnosed. The feature refinement module is used to refine the target state variable set with sensitive features through the encoding module in the Transformer model to obtain refined temporal features; The fault diagnosis module is used to take the refined temporal features as input data of the trained temporal convolutional network, and use the trained temporal convolutional network to perform fault diagnosis on the hypersonic vehicle to be diagnosed, and output the fault type sequence. The fault diagnosis module is also used to determine the fault occurrence time based on the fault type sequence output; The feature refinement module is further configured to: map the target state variable set to a high-dimensional feature space through a linear projection operation to obtain spatial channel refinement features; add a corresponding timestamp to the spatial channel refinement features and perform multi-granularity scanning and filtering operations to obtain elite features; and perform multi-head self-attention processing on the elite features to obtain the refined temporal features. The fault diagnosis module is specifically used for: The refined temporal features are extracted using a temporal convolutional network to obtain fault feature data. A fully connected layer is added in the spatial dimension to map the fault feature data from the high-dimensional feature space back to the one-dimensional output space, resulting in a one-dimensional output sequence. Furthermore, a threshold θ is set to obtain a categorical integer output for each category. THR ∈(0,1), the final fault type sequence is output, and the output value at each time step is: ； Similarly, based on the fault type sequence output, the fault occurrence time is determined: 。 5. The system according to claim 4, characterized in that, The target sensors include: an inertial navigation system sensor and an embedded atmospheric data sensing system sensor; The timing observations include: the angle-of-attack timing observations of the inertial navigation system sensor, the angle-of-attack timing observations of the embedded atmospheric data sensing system sensor, and the joint timing observations of the complementary filter.

6. The system according to claim 4, characterized in that, It also includes a training module for training a preset temporal convolutional network to obtain the trained temporal convolutional network.

7. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the computer program, it implements the steps of the method described in any one of claims 1 to 3.

8. A computer-readable medium having processor-executable non-volatile program code, characterized in that, The program code causes the processor to execute the method according to any one of claims 1-3.

Images