Helicopter medical equipment fixed bracket loosening detection method based on vibration characteristics

Abstract

The application discloses a helicopter medical equipment fixed support loosening detection method based on vibration characteristics, and relates to the technical field of electric digital data processing. Firstly, a sensor is installed to synchronously collect vibration data and flight parameters, the flight state is adaptively segmented, and isometric angle resampling is performed on the basis of speed fluctuation self-compensation; subsequently, a Teager energy operator is used to screen an impact active angle window and construct a loosening modulation impact spectrum tensor; then, a loosening state classification deep learning network based on asymmetric damping convolution is constructed, and a loosening damping physical regularization term is introduced on the basis of a classification cross-entropy loss to optimize network parameters; finally, the trained model is deployed to a monitoring device for real-time identification. The application effectively solves the problems in the prior art, such as unstable phase mapping, dilution of loosening characteristics, insufficient modeling of asymmetric response, and missing of physical priori, and improves detection precision and generalization ability.

Description

Technical Field

[0001] This invention relates to the field of electrical digital data processing technology, and specifically to a method for detecting loosening of fixation brackets in helicopter medical equipment based on vibration characteristics. Background Technology

[0002] The medical equipment mounting brackets within the helicopter medical bay are connected to the fuselage via connectors, and the reliability of this connection directly affects equipment safety and treatment effectiveness. However, the inherent rotor aerodynamic loads, engine vibrations, maneuvering flight loads, and takeoff and landing impacts subject the mounting brackets to a complex and wide-frequency vibration environment over extended periods. The rotor frequency and its higher-order harmonics continuously act on the bracket connection interface, causing it to be subjected to alternating stress. This can lead to a decrease in fastener preload, the appearance of minute gaps, and further development into periodic collisions, frictional slippage, and local structural resonance modulation, resulting in a loosening and degradation process of the mechanical connection. If this loosening is not detected in time, it can affect the working accuracy of the equipment, or even cause fixing failure, equipment detachment, and seriously threaten the safety of personnel on board.

[0003] In existing technologies, equal-angle resampling methods typically employ a fixed interpolation kernel width, failing to consider the impact of rotor speed variation rates on interpolation accuracy. This leads to unstable phase mapping under variable speed conditions, resulting in residual drift in the loosening impact order features. Furthermore, existing loosening detection methods often directly perform order analysis on the entire vibration signal segment, failing to distinguish between local segments containing loosening impacts and stationary vibration segments. A large number of invalid stationary components dilute the loosening modulation features, reducing the separability of slight loosening states. Simultaneously, existing convolutional neural networks lack the ability to model asymmetric mechanical responses at connection interfaces, making it difficult to express nonlinear features such as amplitude-dependent energy dissipation and direction-dependent contact stiffness changes caused by loosening. Finally, existing deep learning methods lack mechanisms for incorporating prior knowledge of loosening dynamics, potentially leading to inconsistencies between network features and actual loosening physical laws, affecting generalization ability and physical interpretability. Summary of the Invention

[0004] The technical problem to be solved by this invention is to provide a method for detecting loosening of a fixation bracket for helicopter medical equipment based on vibration characteristics, thereby solving the problems of unstable phase mapping, dilution of loosening characteristics, insufficient asymmetric response modeling, and lack of physical prior integration in the existing technology.

[0005] To address the aforementioned technical problems, the technical solution adopted by this invention is as follows:

[0006] A method for detecting loosening of fixation braces for helicopter medical equipment based on vibration characteristics includes the following steps:

[0007] S1. Install sensors on the fixed bracket of the helicopter medical equipment and collect vibration data and flight parameter information simultaneously;

[0008] S2. Adaptively segment the flight state and perform equal-angle resampling with self-compensation for speed fluctuations;

[0009] S3. Use the Teager energy operator to screen the impact active angle window and construct the loosening modulation impact spectrum tensor;

[0010] S4. Construct a deep learning network for loose state classification based on asymmetric damped convolution;

[0011] S5. Based on the classification cross-entropy loss, a loose damping physical regularization term is introduced to optimize the network parameters;

[0012] S6. Deploy the trained model to the monitoring device to identify the loosening status of the medical device fixation bracket in real time.

[0013] Furthermore, step S2 includes:

[0014] S201. Synthesize a comprehensive vibration modulus sequence using triaxial vibration acceleration components and segment the flight state;

[0015] S202. Adjust the interpolation kernel width according to the instantaneous frequency change rate and perform equal-angle resampling.

[0016] Furthermore, step S201 includes:

[0017] S20101. Synchronously read the X-axis vibration acceleration component, Y-axis vibration acceleration component and Z-axis vibration acceleration component according to a unified sampling clock, so that the three vibration acceleration components correspond one-to-one at the discrete sampling time.

[0018] S20102. Perform short-time Fourier transform on the comprehensive vibration modulus segment of continuous time frames, and extract the instantaneous frequency conversion curve by combining local spectrum peak search and autocorrelation period consistency verification.

[0019] S20103. Read the flight status tag output by the flight parameter recorder and align the flight status tag with the comprehensive vibration modulus sequence and instantaneous frequency curve according to a unified time reference.

[0020] S20104. Save the time-domain vibration signal and its instantaneous frequency curve in the takeoff state, the time-domain vibration signal and its instantaneous frequency curve in the cruise state, and the time-domain vibration signal and its instantaneous frequency curve in the landing state, respectively, and ensure that the discrete sampling time, the comprehensive vibration modulus value and the rotor instantaneous frequency correspond one-to-one within the same flight state.

[0021] Furthermore, step S202 includes:

[0022] S20201. For any flight state, read its time-domain vibration signal and instantaneous frequency sequence, and integrate the instantaneous frequency along the time direction to obtain the reference phase characterizing the cumulative rotor angle.

[0023] S20202. Establish an equal-angle sampling grid corresponding to the flight state according to a preset constant angle interval;

[0024] S20203. Based on the instantaneous frequency conversion curve of the flight state corresponding to the flight state, calculate the instantaneous frequency conversion rate at each discrete sampling time.

[0025] S20204. For each equal-angle sampling position, calculate its angular deviation from the reference phase, and use a kernel width adaptive sinc interpolation kernel to weight and reconstruct the adjacent time-domain vibration values ​​to obtain the flight state angular domain steady vibration signal corresponding to the flight state.

[0026] S20205. Perform the above-mentioned kernel width adaptive equal angle resampling on the time-domain vibration signal of takeoff, cruise and landing states respectively to obtain the angular domain stable vibration signal of takeoff, cruise and landing states.

[0027] Furthermore, step S3 includes:

[0028] S301. Using the Teager energy operator to filter impact active angle windows;

[0029] S302. Perform a local order Fourier transform on the impact active angle window and generate a multi-channel spectral tensor.

[0030] Furthermore, step S301 includes:

[0031] S30101. For the flight state angular domain steady vibration signal corresponding to the flight state, read the flight state angular domain steady vibration values ​​corresponding to the three adjacent equal angle sampling positions in sequence, and calculate the Teager energy value corresponding to the middle equal angle sampling position.

[0032] S30102. According to the fixed angle window length and fixed angle window step size, the flight status is... The corresponding non-negative Teager energy sequence is divided into multiple overlapping angle windows, and the overlapping angle windows are numbered according to the order of rotor rotation angle;

[0033] S30103. Calculate the maximum, minimum and average values ​​of non-negative Teager energy within each overlap angle window, and use the ratio of the energy variation range to the average value to obtain the shock activity index that measures the degree of transient change.

[0034] S30104. Compare the impact activity index corresponding to each overlapping angle window with the preset activity threshold, retain the overlapping angle windows whose impact activity index is greater than the preset activity threshold, and define the retained overlapping angle windows as impact active angle windows.

[0035] Furthermore, step S302 includes:

[0036] S30201. Read the non-negative Teager energy sequence within the impact active angle window, and multiply the non-negative Teager energy sequence point by point with the Hanning window of the same length;

[0037] S30202. Perform a discrete Fourier transform on the windowed energy sequence and take the square of the modulus to obtain the local loosening modulated impact spectrum that reflects the distribution of loosening impact energy at the rotor order position.

[0038] S30203. According to the original angular order of the active impact angle windows in the stable vibration signal of the flight state angular domain, arrange the local loosening modulation impact spectra corresponding to each active impact angle window in the same flight state along the dimension of the active impact angle window to obtain the loosening modulation impact spectrum matrix corresponding to the flight state.

[0039] S30204. Perform shock activity angle window dimension alignment on the takeoff state loose modulation shock spectrum matrix, cruise state loose modulation shock spectrum matrix and landing state loose modulation shock spectrum matrix.

[0040] S30205. Stack the aligned takeoff state loose modulation shock spectrum matrix, cruise state loose modulation shock spectrum matrix and landing state loose modulation shock spectrum matrix sequentially along the flight state channel direction to obtain the loose modulation shock spectrum tensor.

[0041] Furthermore, step S4 includes:

[0042] S401. Input the output feature map of the previous layer into the current asymmetric damped convolutional layer;

[0043] S402. Perform standard convolution operation on the input feature map using a set of two-dimensional convolution kernels to obtain a convolution response feature map that represents the changes in the combination of local order spectra.

[0044] S403. Extract the forward convolution response based on the convolution response feature map, calculate the exponential damping factor based on the absolute value of the convolution response and the trainable damping coefficient, and then multiply the forward convolution response and the exponential damping factor element by element to obtain the output of the damped convolution branch.

[0045] S404. Extract the forward convolution response based on the convolution response feature map, square the forward convolution response element by element, and adjust the intensity of the second response using a trainable asymmetric gain parameter; at the same time, construct the limiting denominator based on the absolute value of the convolution response and the trainable elastic limiting coefficient, and obtain the asymmetric second response branch output by dividing element by element.

[0046] S405. Add the output of the damped convolution branch to the output of the asymmetric quadratic response branch to obtain the output feature map of the fused amplitude-correlated damped response and the positive asymmetric quadratic response.

[0047] S406. Combine the feature maps of all output channels of the current layer, reduce the feature map spatial size and expand the receptive range of the next layer through max pooling operation;

[0048] S407: Stack multiple asymmetric damped convolutional layers and pooling layers, and output the predicted probabilities of normal, slight and severe loosening states through global average pooling and fully connected layers.

[0049] Furthermore, step S5 includes:

[0050] S501. Read a training batch from the training sample set, process each original loosening monitoring sample in the training batch through S2 and S3 in sequence to obtain the corresponding loosening modulation shock spectrum tensor, and then input each loosening modulation shock spectrum tensor into the loosening state classification deep learning network to obtain the predicted probabilities corresponding to normal state, slightly loosening state and severely loosening state.

[0051] S502. Read the convolution response and trainable damping coefficient of the last asymmetric damped convolutional layer, and calculate the average effective damping response that characterizes the energy dissipation intensity of the sample.

[0052] S503. Generate a loosening indicator function based on the sample's actual loosening state label, which is used to select the appropriate regularization term form for the sample in the physical constraint calculation;

[0053] S504. Set the physical reference threshold and regularization coefficient. Based on the true unique thermal coding label, predicted probability, loosening indicator function and average effective damping response of each original loosening monitoring sample, add the batch classification cross-entropy loss to the loosening damping physical regularization term to obtain the joint loss function.

[0054] S505. Mini-batch backpropagation is used to calculate the gradient of the joint loss function relative to each trainable parameter of the loose state classification deep learning network, and the network parameters are updated iteratively based on the gradient.

[0055] The beneficial effects of this invention compared with the prior art are: (1) This invention proposes a kernel width adaptive equal angle resampling method, which dynamically adjusts the sinc interpolation kernel width according to the instantaneous frequency change rate, so that the high angle resolution is maintained when the rotation speed is stable, and the kernel width is automatically increased when the rotation speed fluctuates violently to integrate a wider range of time domain information, thus solving the problem of unstable phase mapping of fixed kernel width interpolation under variable rotation speed conditions; (2) This invention proposes a local impact screening mechanism based on impact activity index, which uses the Teager energy operator to enhance the transient impact, and uses the ratio of peak-to-peak energy to average energy within the overlapping angle window to measure the impact significance, and uses relative index rather than absolute amplitude for screening, effectively overcoming the interference of the overall vibration amplitude difference in different flight states on the extraction of loosening features; (3) (3) This invention constructs an asymmetric damped convolutional layer and introduces an amplitude-related exponential damping branch and a positive quadratic response branch in parallel after the standard convolution. These branches simulate the characteristics of energy dissipation at the loose connection interface increasing with the increase of the response amplitude, as well as the asymmetric mechanical behavior of contact stiffness hardening under positive compression. This enables the network to learn loosening characteristics with clear physical meaning. (4) This invention designs a physical constraint training mechanism that integrates loosening priors. The average effective damping response formed by the trainable damping coefficient of the last layer of the network and the convolution response amplitude is used as the constraint object. The response of the loose sample is required to be higher than the physical reference threshold, and the response of the normal sample is close to zero. This achieves soft constraint of data adaptation while maintaining global parameter sharing, and avoids conflicts caused by directly constraining shared parameters according to sample labels. Attached Figure Description

[0056] Figure 1 This is an overall flowchart of the method for detecting loosening of the fixation bracket of helicopter medical equipment based on vibration characteristics according to the present invention.

[0057] Figure 2 This is a schematic diagram of the sensor installation location and data acquisition process in the embodiment;

[0058] Figure 3 This is a schematic diagram of the specific process of step S2 (adaptive segmentation of flight state and equal angle resampling) of the present invention;

[0059] Figure 4 This is a waveform diagram of the time-domain vibration signal before resampling in the takeoff state in the embodiment;

[0060] Figure 5 This is a waveform diagram of the stable vibration signal in the angular domain of takeoff state after kernel width adaptive equal angle resampling in the embodiment;

[0061] Figure 6 This is a schematic diagram of the non-negative Teager energy sequence calculated based on the angular domain stationary vibration signal in the embodiment.

[0062] Figure 7This is a schematic diagram of the impact activity index corresponding to each overlapping angle window in the embodiment;

[0063] Figure 8 This is a schematic diagram of the cruise state loosening modulation impact spectrum matrix, presented in the form of a three-dimensional curved surface.

[0064] Figure 9 This is a schematic diagram illustrating the specific process of constructing a loose state classification deep learning network based on asymmetric damped convolution in the embodiment.

[0065] Figure 10 This is a schematic diagram of the instantaneous envelope of the steady vibration signal in the angular domain during takeoff in the embodiment.

[0066] Figure 11 This is a schematic diagram illustrating the process of deploying the trained model to monitoring equipment for real-time identification according to the present invention. Detailed Implementation

[0067] The technical solution of the present invention will be further described below with reference to the accompanying drawings and specific embodiments. The accompanying drawings are for illustrative purposes only, representing schematic diagrams only, not actual physical objects, and should not be construed as limiting the scope of this patent. To better illustrate the embodiments of the present invention, some components in the drawings may be omitted, enlarged, or reduced, and do not represent the actual dimensions of the product. It is understandable to those skilled in the art that some well-known structures and their descriptions may be omitted in the drawings.

[0068] This invention proposes a method for detecting loosening of the fixation bracket in helicopter medical equipment based on vibration characteristics, such as... Figure 1 As shown, it includes the following steps:

[0069] S1. Install sensors on the fixed bracket of the helicopter medical equipment and collect vibration data and flight parameter information simultaneously.

[0070] A triaxial vibration sensor is installed on a medical equipment mounting bracket inside the helicopter medical cabin. The sensor is fixed near the bracket's connection interface, enabling it to simultaneously acquire vibration responses caused by interface collisions, gap crossings, friction, and structural resonance. The three sensing directions of the triaxial vibration sensor are orthogonal to each other, acquiring the X-axis, Y-axis, and Z-axis vibration acceleration components, respectively.

[0071] like Figure 2 As shown, the triaxial vibration sensor continuously collects vibration acceleration data according to a unified sampling clock, and the sampling frequency is denoted as . , This indicates the number of vibration acceleration sampling points collected per second. The sampling frequency should be higher than twice the highest analysis frequency corresponding to the rotor rotation frequency, the main order components of the rotor, and the local structural resonance frequency of the medical device fixation bracket. In one implementation, the sampling frequency... You can choose 5kHz, 10kHz or 20kHz.

[0072] The system synchronously reads rotor speed information, takeoff status label, cruise status label, and landing status label output by the flight parameter recorder, and aligns the data from the flight parameter recorder with the data collected by the triaxial vibration sensor using a unified timestamp. When the flight parameter recorder directly provides rotor speed, the rotor speed can be used as verification information for subsequent instantaneous frequency extraction results; when the flight parameter recorder does not continuously output rotor speed, the instantaneous rotor frequency is extracted based on the vibration signal.

[0073] Data acquisition was repeatedly performed under different loosened conditions of the fixation stents of medical devices, and the loosening status of the data corresponding to each flight mission was labeled, including normal, slightly loose, and severely loose states. A normal state refers to the fixation stent connectors being installed with the specified pre-tightening force and no observable gaps at the connection interface; a slightly loose state refers to a decrease in the pre-tightening force of the fixation stent connectors or the appearance of small gaps at the connection interface, but the medical device remains basically fixed; a severely loose state refers to a significant decrease in the pre-tightening force of the fixation stent connectors, with continuous collisions, slippage, or large gaps at the connection interface.

[0074] In one implementation, the loosening state can be labeled using pre-set fastener preload, tightening torque, connection gap, or structural test results. It should be noted that data for the same loosening state should cover different loads, different rotor speeds, and different flight missions to avoid subsequent loosening state classification models classifying based solely on the overall vibration amplitude of a single flight mission.

[0075] S2. Adaptively segment the flight state and perform equal-angle resampling with self-compensation for speed fluctuations.

[0076] Helicopters exhibit different rotor speed variation patterns during takeoff, cruise, and landing. Fluctuations in rotor speed can cause the collision cycle, friction modulation cycle, and order position of medical device fixation brackets to drift over time due to loosening. If vibration signals are directly intercepted at a fixed time length, the same loosening state will result in different frequency positions and impact intervals under different flight states, reducing the consistency of loosening characteristics.

[0077] This invention first synthesizes a comprehensive vibration modulus sequence and obtains the instantaneous frequency conversion curve by combining local spectral peaks and autocorrelation period consistency. Then, it divides the time-domain vibration signal into takeoff, cruise, and landing states based on flight state labels. Finally, it adjusts the width of the sinc (Singer) interpolation kernel according to the instantaneous frequency conversion rate of change, mapping each flight state time-domain vibration signal to an angular grid with constant angular intervals to obtain the corresponding flight state angular domain stationary vibration signal. The specific steps are as follows: Figure 3 As shown.

[0078] S201. Synthesize a comprehensive vibration modulus sequence using triaxial vibration acceleration components and segment the flight state.

[0079] A triaxial vibration sensor synchronously acquires the vibration acceleration components of the medical device fixation bracket in three orthogonal directions. A comprehensive vibration modulus sequence is obtained through vector synthesis. Instantaneous frequency response curves are then extracted from the comprehensive vibration modulus sequence, and continuous vibration data segmentation is performed based on flight status labels. The specific steps are as follows:

[0080] S20101. Synchronously read the X-axis vibration acceleration component, Y-axis vibration acceleration component and Z-axis vibration acceleration component according to a unified sampling clock, so that the three vibration acceleration components correspond one-to-one at the discrete sampling time.

[0081] In the specific implementation, the first The discrete sampling times are denoted as By performing squaring, summing, and square root operations on the three vibration acceleration components at the same discrete sampling time, the comprehensive vibration modulus sequence is obtained at the _th_ ... The comprehensive vibration modulus value at each discrete sampling time is expressed as:

[0082] ;

[0083] in, The comprehensive vibration modulus sequence represents the first... The comprehensive vibration modulus value at a discrete sampling moment refers to the vibration amplitude synthesized from the vibration acceleration components in three orthogonal directions at the same discrete sampling moment, which is used to characterize the overall vibration intensity of the medical device fixation bracket at that discrete sampling moment. Indicates the first X-axis vibration acceleration components at each discrete sampling time; Indicates the first Y-axis vibration acceleration components at each discrete sampling time; Indicates the first Z-axis vibration acceleration components at discrete sampling times; ; This represents the total number of sampling points included in the comprehensive vibration modulus sequence, for example, 65536.

[0084] It should be noted that the composite vibration modulus sequence retains vibration energy in three orthogonal directions simultaneously. When the installation direction of the triaxial vibration sensor or the helicopter attitude changes, the vibration amplitude in a single direction may change significantly, while the composite vibration modulus sequence is less sensitive to changes in vibration direction, thus providing a relatively stable input signal for subsequent instantaneous frequency extraction.

[0085] S20102. Perform a short-time Fourier transform on the comprehensive vibration modulus segment of the continuous time frame, and extract the instantaneous frequency conversion curve by combining local spectrum peak search and autocorrelation period consistency verification.

[0086] Perform a short-time Fourier transform on the comprehensive vibration modulus sequence according to continuous time frames to obtain the local spectrum corresponding to each time frame, and search for local spectrum peaks within the pre-calibrated rotor frequency range. Based on the local spectrum peaks, determine the candidate frequency for each time frame.

[0087] In the specific implementation, the first The comprehensive vibration modulus segment extracted from each time frame is denoted as ,in, Indicates the time frame index. This represents the index of the sampling point within a time frame; the time frame length can be 1024, 2048, or 4096 sampling points, and the overlap rate between adjacent time frames can be 50% or 75%. After windowing, a Fast Fourier Transform is performed to obtain the first... The local spectrum of each time frame. Furthermore, the pre-calibrated rotor frequency range is determined based on the normal rotational speed range of the helicopter rotor. For example, if the normal rotor frequency range is 4Hz to 8Hz, then local spectral peaks are searched only in the vicinity of 4Hz to 8Hz and its known integer multiples of harmonics.

[0088] It should be noted that windowing refers to the operation of multiplying a finite-length window function with a signal segment point by point in signal processing. The purpose is to smooth the boundaries of the signal segment, reduce spectral leakage caused by direct truncation, and make the spectrum obtained by the subsequent Fourier transform more concentrated and accurate.

[0089] Furthermore, to avoid the inability to extract the rotational frequency simply because the rotor's fundamental frequency is not obvious at certain installation locations, when the local spectral amplitude at the rotor's fundamental frequency exceeds a preset peak value threshold, the frequency containing the fundamental frequency peak is directly determined as the candidate rotational frequency. When the fundamental frequency amplitude is weak and the... When the amplitude of the first harmonic is strong, the first harmonic will be... Peak frequency of the first harmonic divided by the harmonic order This yields candidate frequency transitions. Among them, the harmonic order... You can choose 2, 3, or 4.

[0090] In one implementation, the method for determining whether the fundamental frequency amplitude is weak and the h-th harmonic amplitude is strong is as follows: The local spectral amplitude at the fundamental frequency corresponding to the pre-calibrated rotor frequency range is compared with the amplitude at the h-th harmonic frequency. If the fundamental frequency amplitude is lower than a preset peak value threshold, while the amplitude at the h-th harmonic frequency is significantly higher than this threshold, then the fundamental frequency amplitude is considered weak and the h-th harmonic amplitude is considered strong. The preset peak value threshold can be determined by environmental noise levels and historical vibration data statistics. For example, it can be three times the average amplitude of the local spectrum in the current time frame; that is, the average amplitude of all frequency points in the local spectrum of the m-th time frame is calculated. When the amplitude at the fundamental frequency or the h-th harmonic exceeds three times this average value, the peak value is considered significant.

[0091] In this embodiment, as an example, the pre-calibrated rotor frequency range is: to Within the local spectrum of a time frame, The amplitude did not exceed the preset peak value judgment threshold, but... Location (exactly) of If a significant peak occurs (times) and exceeds the threshold, then it can be determined that the first (times) is... The first harmonic is relatively strong. At this time, the peak harmonic frequency... Divide by harmonic order Candidate frequency switching was obtained .

[0092] Furthermore, regarding the first The autocorrelation sequence of the comprehensive vibration modulus segment within each time frame is calculated, and the autocorrelation peak at a non-zero time delay is searched within the time delay range corresponding to the pre-calibrated rotor frequency range. The sampling point delay corresponding to the autocorrelation peak is denoted as . , equal to the The number of sampling points between adjacent repetitive vibration cycles within a time frame. Based on the sampling frequency. and sampling point delay , obtained the Verification frequency corresponding to each time frame ,Right now .

[0093] In one implementation, the autocorrelation sequence is calculated using a method based on the Fast Fourier Transform. Specifically, the comprehensive vibration modulus segment is first calculated. The fast Fourier transform is performed to obtain its spectrum; then the power spectrum of this spectrum is calculated (i.e., the spectrum multiplied by its conjugate); finally, the inverse fast Fourier transform is performed on the power spectrum to obtain the autocorrelation sequence. In this implementation, searching for the autocorrelation peak at a non-zero time delay refers to finding the local maximum value of the autocorrelation sequence within the time delay interval corresponding to the pre-calibrated rotor frequency range. This interval is... (For example, frequency conversion) to correspond to ),in, This indicates the upper limit frequency of the pre-calibrated rotor speed range. This represents the lower limit frequency of the pre-calibrated rotor frequency range. Within this range, ignoring the global maximum value at zero delay, the first significant autocorrelation peak is found, and the delay time corresponding to this peak is one fundamental period of the vibration signal.

[0094] In this embodiment, as an example, the sampling frequency Frequency conversion range is to The search range for time delay is... The sampling point. If within this range, the autocorrelation sequence is at the sampling point. If the maximum peak value appears at the delay of each sampling point, then the frequency conversion is verified. .

[0095] Furthermore, the candidate frequency transitions obtained from the local spectral peak search are compared with the verification frequency transitions. A comparison is performed. If the relative deviation between the two is no greater than a preset consistency threshold, the candidate frequency is retained; if the relative deviation exceeds the preset consistency threshold, the candidate frequency is classified as an abnormal frequency. The preset consistency threshold can be set to 5% to 15%.

[0096] Simultaneously, the candidate frequency changes in adjacent time frames are compared. When the candidate frequency change in adjacent time frames exceeds the maximum allowable speed change of the helicopter rotor within the corresponding time interval, the candidate frequency that experiences a momentary jump is identified as an abnormal frequency. In one implementation, the abnormal frequency location is obtained by linear interpolation of the frequencies corresponding to adjacent valid time frames to obtain the frequency value at the center of each time frame. These frequency values ​​constitute a time frame-level sequence, called the instantaneous frequency sequence; this sequence is then analyzed at discrete time points. The sampled value is denoted as , indicating the first The instantaneous rotor frequency value corresponding to each discrete sampling moment is given by the set of all... The resulting curve is called the instantaneous frequency curve.

[0097] In one implementation, the maximum rotational speed variation can be preset based on the helicopter rotor's moment of inertia and engine speed regulation performance. As an example, for a certain type of helicopter, this value can be set to 0.5 Hz / s. When the rotational speed variation rate obtained by dividing the candidate rotational speed variation in adjacent time frames by the time frame step size exceeds 0.5 Hz / s, it is determined to be an abnormal rotational speed.

[0098] In this embodiment, as an example, assume that the first The candidate frequency shift in one time frame was determined to be abnormal, while the adjacent previous valid time frame The frequency conversion is The next valid time frame The frequency conversion is Through linear interpolation, the first... The frequency conversion results for each time frame will be corrected to .

[0099] Furthermore, the frequency conversion results of each time frame are interpolated to the discrete sampling times of the comprehensive vibration modulus sequence according to the center time of the time frame to obtain the instantaneous frequency conversion curve. The instantaneous frequency conversion curve at discrete times... The sampled value is denoted as , indicating the first The instantaneous rotor frequency value corresponding to each discrete sampling moment is a scalar value used to characterize the instantaneous rotational frequency of the rotor at that moment.

[0100] In practical implementation, the frequency conversion result for each time frame refers to a single numerical value, after local spectral peak search, autocorrelation period consistency verification, abnormal frequency conversion removal, and interpolation correction, that represents the instantaneous rotor rotation frequency at the center moment of each time frame (i.e., each short-time Fourier transform analysis window). This value can be denoted as: ,in, Indicates the first The center moment of each time frame, and this series This constitutes an instantaneous frequency conversion sequence at the time frame level.

[0101] In this embodiment, as an example, the sampling frequency Using a 10kHz frequency, a time frame length of 2048 sampling points, and a 75% overlap between adjacent time frames, the pre-calibrated rotor frequency range is set from 4Hz to 8Hz. If a significant peak appears at 12Hz in a given time frame, and this peak is identified as the second harmonic, then the candidate frequency is 6Hz. If the autocorrelation peak in that time frame corresponds to a sampling point delay of approximately 1667 sampling points, then the verified frequency is approximately... Since the two meet the consistency requirements, 6Hz is retained as the frequency conversion result for this time frame.

[0102] S20103. Read the flight status tag output by the flight parameter recorder and align the flight status tag with the comprehensive vibration modulus sequence and instantaneous frequency curve according to a unified time reference.

[0103] Specifically, the flight status label is recorded as , ;in, Indicates takeoff status; Indicates cruise status; Indicates landing status.

[0104] Furthermore, based on the start and end times corresponding to the takeoff status labels, the time-domain vibration signal of the takeoff status is extracted from the comprehensive vibration modulus sequence. Furthermore, a sequence of instantaneous frequency values ​​corresponding to the takeoff state is synchronously extracted from the instantaneous frequency curve. This sequence constitutes the instantaneous frequency curve for the takeoff state, and its value is determined at the [missing information - likely a specific value]. The value of each sampling point is denoted as .

[0105] Simultaneously, based on the start and end times corresponding to the cruise state labels, the time-domain vibration signal of the cruise state is extracted from the comprehensive vibration modulus sequence. Furthermore, a sequence of instantaneous frequency values ​​corresponding to the cruise state is synchronously extracted from the instantaneous frequency curve. This sequence constitutes the instantaneous frequency curve of the cruise state, and its value is determined at the [missing information - likely a date or time value]. The value of each sampling point is denoted as .

[0106] Simultaneously, based on the start and end times corresponding to the landing status labels, the time-domain vibration signal of the landing status is extracted from the comprehensive vibration modulus sequence. And synchronously extract the instantaneous frequency value sequence corresponding to the landing state from the instantaneous frequency curve. This sequence constitutes the instantaneous frequency curve of the landing state, which is the first... The value of each sampling point is denoted as .

[0107] In this embodiment, as an example, taking the takeoff state as an example, if the start time of the takeoff state tag "takeoff" provided by the flight parameter recorder is "the first time after the mission starts". "seconds", the end time is "the 1st second after the start of the mission". "seconds", then: from the comprehensive vibration modulus sequence In the middle, extract the timestamp. to All sampling points between these points constitute the time-domain vibration signal during takeoff. Meanwhile, from the instantaneous frequency conversion curve In the middle, the timestamp is captured synchronously. to The portion between these points constitutes the instantaneous frequency curve during takeoff. The interception operations for cruise and landing states are similar.

[0108] S20104. Save the time-domain vibration signal and its instantaneous frequency response curve during takeoff (both with a length of...). (vector), cruise state time-domain vibration signal and its cruise state instantaneous frequency conversion curve (both lengths are...) The vector), and the time-domain vibration signal of the landing state and its instantaneous frequency conversion curve of the landing state (both lengths are... The vector of the rotor is used to ensure that the discrete sampling time, the comprehensive vibration modulus value and the instantaneous rotor frequency correspond one-to-one within the same flight state.

[0109] It should be noted that processing takeoff, cruise, and landing states separately avoids simultaneously processing different speed change patterns such as continuously increasing speed, relatively stable speed, and continuously decreasing speed during the same equal-angle resampling process. This allows subsequent kernel width adaptive equal-angle resampling to independently perform phase compensation for instantaneous speed fluctuations in each flight state.

[0110] S202. Adjust the interpolation kernel width according to the instantaneous frequency change rate and perform equal-angle resampling.

[0111] To address the issue of different rotor speed variation patterns during takeoff, cruise, and landing, this invention calculates a reference phase based on the instantaneous rotational frequency curves corresponding to each flight state, and adjusts the sinc interpolation kernel width according to the instantaneous rotational frequency change rate. This converts the time-domain vibration signals of each flight state into stationary angular-domain vibration signals with constant angular intervals. The specific steps are as follows:

[0112] S20201. For any flight state, read its time-domain vibration signal and instantaneous frequency sequence, and integrate the instantaneous frequency along the time direction to obtain the reference phase characterizing the cumulative rotor angle.

[0113] Specifically, for any flight state Read flight status Corresponding flight state time-domain vibration signal and the values ​​at each sampling time The instantaneous frequency sequence of the flight state is constructed, and then the instantaneous frequency sequence is integrated along the time direction to obtain the flight state. The corresponding reference phase is represented as:

[0114] ;

[0115] in, Indicates flight status Next moment The corresponding reference phase, in radians, is used to characterize the rotor's transition from flight state. The start time to the end time The cumulative angle rotated; Indicates flight status Next moment The corresponding instantaneous rotor frequency; Indicates flight status The starting time; Represents the differential symbol; This represents the time variable during the integration process.

[0116] In the discrete implementation, the trapezoidal summation integral is performed using the average instantaneous rotation frequency of adjacent discrete sampling times. Specifically, let the flight state If the reference phase corresponding to the first discrete sampling time is 0, then the... The reference phase corresponding to each discrete sampling time can be derived using the following formula:

[0117] ;

[0118] in, This represents the reference phase corresponding to the previous discrete sampling time. This represents the time interval between adjacent discrete sampling moments. When vibration data uses a constant sampling frequency... During collection, .

[0119] It should be noted that the reference phase transforms the original sampling positions arranged according to time into non-uniform phase positions arranged according to the cumulative rotor angle, providing coordinate basis for the subsequent establishment of an angle grid with constant angular intervals.

[0120] S20202. Establish flight status according to preset constant angle intervals. The corresponding iso-angle sampling grid.

[0121] In practical implementation, the constant angle interval is denoted as... From flight status Starting from the initial reference phase, according to Equal-angle sampling positions are generated sequentially, i.e. .

[0122] in, Indicates the first The angular positions corresponding to each equal-angle sampling point; ; Indicates flight status The corresponding flight state angular domain steady vibration signal contains the number of equal-angle sampling points, and, According to flight status Cumulative reference phase range and constant angle interval Determine, for example, if the flight status The internal cumulative reference phase range is radian (i.e., the radius of rotation of the rotor) (Circle), preset constant angle interval Pick radian( Then, the number of equal-angle sampling points in this state is: indivual.

[0123] In this embodiment, as an example, a constant angular interval is used. Pick ,Right now The radian dimension ensures that each rotor revolution corresponds to 720 equal-angle sampling points. After mapping using this angular grid, the same-order vibration component at different rotor speeds can appear at the same angular periodic position. For example, when the angular interval is constant... Pick (Right now At radians, the rotor rotates for each revolution. correspond There are equiangular sampling points. If a certain loosening impact characteristic occurs once per rotor rotation, then in the angular domain steady vibration signal, this impact will stably occur at intervals. Each sampling point appears once, regardless of the rotor speed.

[0124] S20203, Based on Flight Status The instantaneous frequency change rate at each discrete sampling moment is calculated from the corresponding instantaneous frequency curve of the flight state.

[0125] In practical implementation, for discrete sampling moments located in the middle of the flight state, the central difference method can be used for calculation, expressed as:

[0126] ;

[0127] in, Indicates flight status Next The instantaneous frequency change rate corresponding to each discrete sampling moment is used to characterize how fast the rotor speed changes around that discrete sampling moment; and These represent the instantaneous rotor frequency values ​​corresponding to the next discrete sampling time and the previous discrete sampling time, respectively. For the discrete sampling times at both ends of the flight state, the instantaneous frequency change rate can be calculated using forward differential or backward differential.

[0128] Furthermore, based on constant angular intervals Instantaneous frequency change rate and nuclear width regulation scaling factor Determine the first The width of the adaptive interpolation kernel corresponding to each discrete sampling time is expressed as:

[0129] ;

[0130] in, Indicates flight status Next The adaptive interpolation kernel width corresponding to each discrete sampling time; This represents the nuclear width regulation scaling factor, whose dimensions are: It is a dimensionless value used to adjust the influence of the instantaneous frequency change rate on the interpolation kernel width.

[0131] It should be noted that when the instantaneous rate of change of frequency is small, Approximate constant angular interval The interpolation process primarily maintains local angular resolution; however, when the instantaneous frequency change rate is large, The corresponding increase enables the current equal-angle sampling points to integrate neighboring time-domain sampling information within a larger phase range, thereby reducing instantaneous frequency estimation errors and phase shifts caused by rotor speed fluctuations.

[0132] S20204. For each equal-angle sampling position, calculate its angular deviation from the reference phase, and use a kernel width adaptive sinc interpolation kernel to weight and reconstruct the adjacent time-domain vibration values ​​to obtain the flight state angular domain steady vibration signal corresponding to the flight state.

[0133] Specifically, for each isoangular sampling position Based on the reference phase corresponding to each discrete sampling time Calculate the sampling position at equal angles With reference phase The angular deviation between them was calculated, and a kernel width-adaptive sinc interpolation kernel was used to weight and reconstruct the nearby time-domain vibration values ​​to obtain the flight state. The corresponding steady-state vibration signal in the angular domain of the flight state is represented as:

[0134] ;

[0135] in, Indicates flight status Next The flight state angular domain steady vibration value corresponding to each equal angle sampling position; Indicates flight status Next The comprehensive vibration modulus value corresponding to each discrete sampling time; Indicates flight status Next The reference phase corresponding to each discrete sampling time; Indicates participation in the A set of indexes of neighboring time-domain sampling points calculated by interpolation of equal-angle sampling positions; This represents the normalized sinc interpolation kernel function, i.e. ,in, This represents the normalized angular deviation between the equal-angle sampling position and the reference phase of the time-domain sampling point.

[0136] In the actual implementation, for each isoangular sampling position First, search in the reference phase sequence for... The nearest time-domain sampling point is used as the index for selecting 4 to 8 time-domain sampling points before and after it to form a neighboring time-domain sampling point index set. This allows numerical computation to be performed using a finite-length sinc interpolation kernel. This finite-length processing only limits the number of neighboring time-domain sampling points involved in the computation, without changing the relationship that the width of the adaptive interpolation kernel increases with the rate of change of instantaneous frequency.

[0137] It should be noted that the kernel width adaptive equal-angle resampling does not set a fixed interpolation kernel width according to the flight state, but rather adjusts the interpolation kernel width based on the instantaneous rate of change of the rotational frequency at each discrete sampling moment within the same flight state. This design can maintain high angular resolution in the stable rotational speed region of cruise, and improve the stability of phase mapping in takeoff, landing, or instantaneous rotational speed fluctuation regions.

[0138] S20205. Perform the above kernel width adaptive equal-angle resampling on the time-domain vibration signals of takeoff, cruise, and landing states respectively to obtain the stable angular domain vibration signal of takeoff. Cruise state angular domain steady vibration signal and landing state angular domain steady vibration signal .

[0139] In this embodiment, as an example, a constant angular interval is used. Pick Nuclear width regulation scaling factor The value is set to 0.05. After kernel width adaptive angle resampling, the loosening collision period, which originally drifted in time and position with the rotor speed, is converted into a relatively stable periodic position within the angular domain, so that the same loosening state forms comparable order features in takeoff, cruise and landing states.

[0140] like Figure 4 As shown in this embodiment, the original waveform of the time-domain vibration signal in the takeoff state before resampling is displayed. The horizontal axis represents time in seconds, and the vertical axis represents the comprehensive vibration modulus in meters per second squared, reflecting the influence of rotor speed change on the period of the time-domain signal.

[0141] like Figure 5As shown in this embodiment, the stable vibration signal in the angular domain during takeoff is simultaneously displayed, i.e., the result after kernel width adaptive equal-angle resampling. The horizontal axis is the rotor angle in degrees; the vertical axis is the comprehensive vibration modulus in meters per second squared. After resampling, the impact intervals that originally varied with the rotational speed are mapped to constant angular intervals. Loosening impacts appear at fixed angular positions. Equal-angle resampling can eliminate order drift caused by rotational speed fluctuations.

[0142] S3. Use the Teager energy operator to screen the shock active angle window and construct the loose modulation shock spectrum tensor.

[0143] When the fixation bracket of medical equipment becomes loose, the connection interface will generate periodic collisions, friction and structural resonance modulation, which will cause localized transient impacts in the steady vibration signal of the flight state angular domain. The loosening impact does not occur continuously at different rotor angle positions and different flight states. If order analysis is performed directly on all steady vibration signals of the flight state angular domain, a large number of steady segments that do not contain obvious impacts will reduce the proportion of loosening features in the subsequent classification input.

[0144] This invention first utilizes the Teager energy operator to enhance the transient impact component in the stationary vibration signal in the angular domain of flight state. Then, the Teager energy sequence is divided into overlapping angular windows, and the impact activity index is calculated based on the relationship between the Teager energy variation range and the average energy within each overlapping angular window, thus selecting the impact-active angular windows. Next, a local order Fourier transform is performed on the impact-active angular windows, and the loosening modulation impact spectrum matrices corresponding to takeoff, cruise, and landing states are aligned and stacked to obtain the loosening modulation impact spectrum tensor. The specific steps are as follows:

[0145] S301. Use the Teager energy operator to screen for impact-active angle windows.

[0146] For each flight state, the non-negative Teager energy sequence is calculated from the stationary vibration signal in the angular domain. The impact-active angular window is then selected based on the dynamic change of the non-negative Teager energy within the overlapping angular window. The specific steps are as follows:

[0147] S30101, Regarding flight status Corresponding flight state angular domain steady vibration signal The flight state angular domain stationary vibration values ​​corresponding to three adjacent equal-angle sampling positions are read sequentially, and the Teager energy value corresponding to the middle equal-angle sampling position is calculated, as follows:

[0148] ;

[0149] in, Indicates flight status Next Teager energy values ​​corresponding to each equiangular sampling position; Indicates the first The flight state angular domain steady vibration value corresponding to each equal angle sampling position; This represents the steady-state vibration value in the angular domain of the flight state corresponding to the previous angular sampling position. This represents the steady-state vibration value in the angular domain of the flight state corresponding to the next equal-angle sampling position.

[0150] It should be noted that the Teager energy operator utilizes both the current vibration amplitude and its relationship with adjacent angular sampling positions. When a loosening or collision impact occurs in the steady-state vibration signal in the angular domain of flight, the change between the current vibration value and adjacent vibration values ​​increases, and the corresponding Teager energy value increases accordingly; when the steady-state vibration signal in the angular domain of flight is in a state of continuous steady change, the corresponding Teager energy value is relatively small.

[0151] Furthermore, considering that noise and discrete calculation errors may cause a small number of Teager energy values ​​to be less than 0, the Teager energy values ​​are non-negatively processed to obtain non-negative Teager energy values, i.e. .in, Indicates flight status Next The non-negative Teager energy values ​​corresponding to each equiangular sampling position are used to characterize the local transient energy intensity at that sampling position. Subsequent impact activity calculations and local order Fourier transforms are both based on the non-negative Teager energy values.

[0152] It should be noted that the Teager energy calculation only uses the effective angular positions with the previous and subsequent isoangular sampling positions. The first and last isoangular sampling positions of the steady vibration signal in the angular domain of the flight state are not included in the Teager energy calculation.

[0153] S30102. According to the fixed angle window length and fixed angle window step size, the flight status is... The corresponding non-negative Teager energy sequence is divided into multiple overlapping angle windows, and the overlapping angle windows are numbered according to the order of rotor rotation.

[0154] The number of non-negative Teager energy sampling points contained in each overlapping angle window is denoted as . The step size for moving adjacent overlapping angle windows is denoted as... ; will the first Within the overlapping angle window, the first The non-negative Teager energy value corresponding to each sampling point is denoted as . ,in, .

[0155] In this embodiment, as an example, the number of sampling points contained in each overlapping angle window is... If we take 256, and the overlap rate of adjacent overlapping angle windows is 50%, then the angle window movement step size is... Set to 128. Adjacent overlapping angle windows retain some of the same non-negative Teager energy sampling points, so that loose collision impacts located near the boundaries of the angle windows can be more fully represented in at least one overlapping angle window.

[0156] S30103. Calculate the maximum, minimum and average values ​​of non-negative Teager energy within each overlap angle window, and use the ratio of the energy variation range to the average value to obtain the shock activity index that measures the degree of transient change.

[0157] Specifically, calculate the first one separately. The maximum, minimum, and average values ​​of non-negative Teager energy within each overlapping angle window are calculated, and the difference between the maximum and minimum values ​​characterizes the dynamic range of transient energy within that overlapping angle window. The average value is used as the energy normalization benchmark between different vibration amplitude conditions to obtain the impact activity index, expressed as:

[0158] ;

[0159] in, Indicates flight status Next The impact activity index corresponding to each overlapping angle window is used to quantify the significance of loosening impact relative to the average vibration energy within that overlapping angle window. Indicates the first The maximum value of non-negative Teager energy within an overlapping angle window; Indicates the first The minimum value of non-negative Teager energy within an overlapping angle window; the difference between the two is used to characterize the range of variation of non-negative Teager energy within that overlapping angle window; Indicates the first The average value of non-negative Teager energy within each overlapping angle window.

[0160] It should be noted that when the first When the average non-negative Teager energy within a certain overlapping angle window is 0, it indicates that there is no effective vibration energy within that overlapping angle window, and the corresponding impact activity index will be... Setting it directly to 0 will prevent the division operation from being performed.

[0161] It should also be noted that the impact activity index is measured using the relative relationship between the energy variation range and the average energy. When there is a significant loose impact within the overlapping angle window, the difference between the local non-negative Teager energy peak and the background energy increases, and the impact activity index increases accordingly. When the overlapping angle window mainly contains continuous steady vibrations, the non-negative Teager energy variation range is relatively small, and the impact activity index decreases accordingly. This processing can reduce the impact of the overall vibration amplitude difference between takeoff, cruise, and landing states on the impact screening results.

[0162] In this embodiment, as an example, the maximum non-negative Teager energy within a certain overlap angle window is 2.4, the minimum is 0.2, and the average is 0.8. Therefore, the impact activity index is... The maximum non-negative Teager energy within the other overlapping angle window is 1.0, the minimum is 0.7, and the average is 0.8. Therefore, the impact activity index is... The previous overlapping angle window has more pronounced local energy abrupt changes and is more likely to contain loosening collision impacts.

[0163] S30104. Compare the impact activity index corresponding to each overlapping angle window with the preset activity threshold. For comparison, retain shock activity indices greater than the preset activity threshold. The overlapping angle windows are defined, and the retained overlapping angle windows are defined as impact active angle windows.

[0164] In one implementation, a preset activity threshold is used. Take 1.5. When At that time, the first A single overlapping angle window is determined as an impact-active angle window; when At that time, the first An overlapping angle window is used for subsequent loosening modulation shock spectrum construction.

[0165] Furthermore, the impact active angle windows are renumbered according to their original angular sequence in the steady vibration signal of the flight state angular domain, and the flight state is... The total number of corresponding impact active angle windows is denoted as .in, This represents the total number of impact active angle windows corresponding to the takeoff state. This indicates the total number of impact-active angle windows corresponding to the cruise state. This represents the total number of impact activity angle windows corresponding to the landing state. Because the intensity and frequency of loosening impacts vary under different flight states, , and They can be different.

[0166] like Figure 6As shown, in this embodiment, the non-negative Teager energy sequence calculated based on the stationary vibration signal in the angular domain is analyzed. The horizontal axis represents the equiangular sampling position, with the unit being radians; the vertical axis represents the non-negative Teager energy value, with its dimension being on the order of the square of the vibration value (if the unit of the vibration value is the unit of acceleration, then the unit of Teager energy is the square of the acceleration). At the angular position where a loosening collision impact occurs, the Teager energy value shows a significant peak, while the energy value is close to zero in the stationary vibration region.

[0167] like Figure 7 As shown in this embodiment, the impact activity index calculated for each overlapping angle window (window length 256 points, step size 128 points) is displayed. The horizontal axis represents the overlapping angle window number, which is unitless; the vertical axis represents the impact activity index, which is dimensionless and defined as the difference between the maximum and minimum non-negative Teager energy within the window divided by the average energy within the window. The red dashed line in the figure marks the preset activity threshold of 1.5, and the orange scatter dots mark the windows selected as impact-active angle windows. It can be seen that angle windows containing significant impacts have significantly higher impact activity values, and the impact activity index can measure the degree of transient change in a relative energy manner.

[0168] S302. Perform a local order Fourier transform on the impact active angle window and generate a multi-channel spectral tensor.

[0169] A windowed Fourier transform is performed on the non-negative Teager energy sequence within each impact active angle window to obtain the local loosening modulated impact spectrum reflecting the order distribution of loosening impact energy. Then, the spectrum matrix is ​​arranged, aligned, and stacked according to the flight state and angle evolution order. The specific steps are as follows:

[0170] S30201. Read the non-negative Teager energy sequence within the impact active angle window, and multiply the non-negative Teager energy sequence point by point with a Hanning window of the same length.

[0171] Specifically, regarding flight status Next A number of impact active angle windows are used to read the length within each impact active angle window. The non-negative Teager energy sequence is then multiplied pointwise by a Hanning window of the same length.

[0172] Hanning window in The window coefficient at each sampling point is denoted as . The calculation method is expressed as follows:

[0173] ;

[0174] in, The Hanning window is used to reduce spectral leakage caused by discontinuities at both ends of the impact active angle window, making the order energy distribution in the locally loosened modulated impact spectrum more concentrated.

[0175] S30202. Perform a discrete Fourier transform on the windowed energy sequence and take the square of the modulus to obtain the local loosening modulated impact spectrum that reflects the distribution of loosening impact energy at the rotor order position.

[0176] Specifically, a Discrete Fourier Transform is performed on the windowed non-negative Teager energy sequence, and the complex result obtained by the Discrete Fourier Transform is squared modulo the result to obtain the flight state. Next The local loosening modulation impact spectrum corresponding to each impact activity angle window is expressed as:

[0177] ;

[0178] in, Indicates flight status Next The first impact active angle window in the Loose modulation impulse spectrum values ​​at discrete frequency points; ; This indicates the number of non-negative discrete-order frequency points retained; ; Indicates flight status Next The first impact active angle window The non-negative Teager energy value corresponding to each sampling point (is the element value of the non-negative Teager energy sequence); Indicates the Hanning window is in Window coefficient at each sampling point; Represents the imaginary unit; This represents an exponential function.

[0179] Furthermore, the first The actual order corresponding to each discrete order frequency point is indexed by the discrete order frequency point. , Window length and constant angular interval Determined jointly. When the angular interval is constant. When expressed in radians, the first The actual order corresponding to each discrete order frequency point ;in, Indicates the first The number of cycles per revolution corresponding to the discrete-order frequency points represents the number of the first discrete-order frequency points. The actual order corresponding to each discrete order frequency point (in "cycles per revolution") is... ).

[0180] It should be noted that, It is used to characterize the distribution of loosening impact energy at rotor order positions, rather than at ordinary time frequency positions. Create discrete indexes The mapping relationship between the fundamental frequency (1st order) and the physical world's "order" with a definite physical meaning; for example, the fundamental frequency (1st order) represents an event that occurs once per revolution, which defines a matrix. The Middle The row's "row coordinate label" (i.e., physical order) and Together they described the order distribution characteristics of loosening impact energy.

[0181] In this embodiment, the length of the angle window Taking 256, and retaining the first 128 non-negative discrete-order frequency points in the discrete Fourier transform result, then... Take 128. When the angular interval is constant... Pick At that time, the first The actual order corresponding to each discrete order frequency point is: .

[0182] S30203. According to the original angular order of the impact active angle windows in the stable vibration signal of the flight state angular domain, arrange the local loosening modulation impact spectra corresponding to each impact active angle window in the same flight state sequentially along the dimension of the impact active angle window to obtain the flight state. The corresponding loosening modulation shock spectrum matrix .

[0183] Specifically, the loosening modulation shock spectrum matrix The size is The row direction corresponds to discrete-order frequency points, and the column direction corresponds to the evolution order of the impact activity angle window with rotor rotation angle. (Loosening modulation impact spectrum matrix) At the same time, it retains the order distribution information of loosening impact and the local evolution information of loosening impact as the rotor angle changes.

[0184] In this matrix, the row index corresponds to the discrete order frequency point. ( The column index corresponds to the impact active angle windows arranged in order of angle evolution under the corresponding flight state. ( ).

[0185] In one implementation, when no impact active angle window is selected for a certain flight state, the loosening modulation impact spectrum matrix corresponding to that flight state is set to a size of [size missing]. The all-zero matrix is ​​used to characterize the local impact that does not meet the preset activity threshold and is not detected under this flight condition.

[0186] S30204: Perform shock activity angle window dimension alignment on the takeoff state loose modulation shock spectrum matrix, cruise state loose modulation shock spectrum matrix and landing state loose modulation shock spectrum matrix.

[0187] In practical implementation, the maximum value of the total number of impact active angle windows corresponding to the three flight states in the same original loosening monitoring sample is determined as the number of impact active angle windows after in-sample alignment. .in, This indicates the number of impact active angle windows after aligning the original loosening monitoring samples; This indicates the operation of finding the maximum value. This represents the minimum number of impact active angle windows required to ensure that subsequent consecutive max pooling operations can be executed normally. When the network subsequently performs three consecutive 2x max pooling operations... We can choose 8.

[0188] In one implementation, for an impact active angle window number less than The loosened modulation shock spectrum matrix is ​​obtained by maintaining the existing arrangement order of the shock activity angle windows and adding spectral columns with values ​​of 0 at the end of the column directions until the matrix size reaches [value missing]. The supplementary spectral series with a value of 0 is used to characterize the absence of a preserved impact active angle window at the corresponding location, thus not introducing additional loosening impact energy.

[0189] It should be noted that different original loosening monitoring samples correspond to... They can differ. The deep learning network for loosening state classification only uses convolution, max pooling, and global average pooling in the order frequency direction and the impact activity angle window direction, and does not require all original loosening monitoring samples to have exactly the same characteristics. During mini-batch training, each original loosening monitoring sample within the same training batch is further padded with zeros to the maximum value in that training batch. The zero-padding spectrum is only used as the location of the unloosened impact energy in the subsequent global feature calculation.

[0190] S30205. Stack the aligned takeoff state loose modulation shock spectrum matrix, cruise state loose modulation shock spectrum matrix, and landing state loose modulation shock spectrum matrix sequentially along the flight state channel direction to obtain the loose modulation shock spectrum tensor. .

[0191] Specifically, loosening modulation shock spectrum tensor The size is The first dimension consists of three channels corresponding to takeoff, cruise, and landing states, respectively; the second dimension corresponds to discrete-order frequency points; and the third dimension corresponds to the aligned impact activity angle window position.

[0192] Furthermore, the loosening modulation shock spectrum tensor It can simultaneously express the energy distribution of loosening impact at different order positions, the local variation of loosening impact with rotor angle, and the differentiated response of the same loosening state in takeoff, cruise and landing states.

[0193] It should be noted that the local impact activity screening does not directly select the angle window with the largest vibration amplitude, but rather selects the angle window with obvious transient changes based on the ratio of the non-negative Teager energy change range to the average energy. This can reduce the impact of the overall vibration amplitude difference in different flight states on the screening results and increase the proportion of local collision and friction modulation features caused by slight loosening in the loosening modulated impact spectrum tensor.

[0194] like Figure 8 As shown in this embodiment, the loosening modulation shock spectrum matrix in cruise state is displayed in the form of a three-dimensional surface. The X-axis represents the discrete order frequency index, the Y-axis represents the shock activity angle window number, and the Z-axis represents the loosening modulation shock spectrum value. The surface plot clearly shows the peak structure in the order direction (corresponding to the rotor frequency and its harmonics) and the spectral fluctuations along the angle window direction, demonstrating that the loosening modulation shock spectrum tensor simultaneously retains order distribution information and local changes in shock intensity with angle evolution.

[0195] S4. Construct a deep learning network for loose state classification based on asymmetric damped convolution.

[0196] When a medical device's fixation bracket loosens, the connection interface exhibits different mechanical responses during forward compression and reverse separation. During forward compression, contact stiffness may increase, while during reverse movement, gap crossing and contact loss may occur. Simultaneously, energy dissipation caused by collisions and friction varies with the amplitude of the local vibration response. Conventional convolutional layers, employing linear convolution and fixed activation functions, struggle to directly express amplitude-dependent damping and asymmetric responses in both positive and negative directions.

[0197] This invention constructs an asymmetric damped convolutional layer using the loosening modulated impact spectrum tensor as input. Specifically, the asymmetric damped convolutional layer first obtains the local order spectral response through two-dimensional convolution, then applies amplitude-correlated exponential decay to the forward convolutional response using trainable damping coefficients, and applies secondary enhancement and elastic limiting to the forward convolutional response, resulting in an output feature map that simultaneously contains local spectral combination features, amplitude-correlated damping features, and forward contact hardening features. The specific steps are as follows: Figure 9 As shown.

[0198] S401. Input the output feature map of the previous layer into the current asymmetric damped convolutional layer.

[0199] For the deep learning network for loosening state classification, the first... Asymmetric damped convolutional layers are used to denot the input feature map as follows: The first is obtained through standard two-dimensional convolution, damped modulation, and asymmetric quadratic response processing. Layer output feature map .

[0200] Specifically, the first Feature maps output by the layer Enter the first Asymmetric damped convolutional layers. Input feature map. The size is ,in, Indicates the first The number of input channels in an asymmetric damped convolutional layer; This represents the size of the input feature map in the direction of the order frequency point; This represents the size of the input feature map in the direction of the impact active angle window.

[0201] when At that time, the input feature map of the first asymmetric damped convolutional layer For loosening the modulated shock spectrum tensor Number of input channels There are 3 input channels, corresponding to takeoff, cruise, and landing states respectively.

[0202] S402. Perform standard convolution operation on the input feature map using a set of two-dimensional convolution kernels to obtain a convolution response feature map that represents the changes in the combination of local order spectra.

[0203] Specifically, using the first Layer The set of two-dimensional convolutional kernels corresponding to each output channel, for the input feature map Perform standard two-dimensional convolution on all input channels to obtain the first... The convolutional response feature map corresponding to each output channel is represented as follows:

[0204] ;

[0205] in, Indicates the first Layer Convolutional response feature maps corresponding to each output channel; Indicates the use of generating the first A set of two-dimensional convolutional kernels for each output channel, this set of two-dimensional convolutional kernels covers all One input channel; This represents a standard two-dimensional convolution operation; ; Indicates the first The number of output channels of an asymmetric damped convolutional layer.

[0206] In one implementation, the size of the two-dimensional convolution kernel of each asymmetric damped convolutional layer is taken as... The stride of the two-dimensional convolution is set to 1, and zero padding is performed at one position in the order frequency direction and the impact active angle window direction, so that the convolution response feature map and the input feature map maintain the same spatial size.

[0207] It should be noted that the convolutional response feature map It is used to characterize the local combination changes between adjacent order frequency points and adjacent impact active angle windows in the loosening modulation impact spectrum tensor. Damping modulation and asymmetric quadratic response both act on the same convolutional response feature map, thereby ensuring the consistency of spatial size in subsequent element-wise operations.

[0208] S403, Based on Convolutional Response Feature Map Extract the forward convolution response, calculate the exponential damping factor based on the absolute value of the convolution response and the trainable damping coefficient, and then multiply the forward convolution response element-wise with the exponential damping factor to obtain the output of the damped convolution branch, expressed as:

[0209] ;

[0210] in, Indicates the first Layer Each output channel corresponds to a damped convolution branch output; This represents a modified linear activation function, used to retain positive values ​​in the convolution response and set negative values ​​to 0; Indicates the first Layer The trainable damping coefficients corresponding to each output channel; This indicates taking the absolute value of each element of the convolutional response feature map; This indicates element-wise exponentiation. This indicates element-wise multiplication.

[0211] It should be noted that the trainable damping coefficient The parameter is non-negative, and in the th... The frequency points of all orders and all impulse active angle window positions of each output channel are shared. When the absolute value of the convolution response increases, the exponential damping factor decreases, causing a more significant attenuation of the large-amplitude positive convolution response; when the convolution response is in the small-to-medium amplitude range, the exponential damping factor is close to 1, preserving the loose modulation characteristics of the small-to-medium amplitude.

[0212] In one implementation, an internal parameter that is not bound by sign can be set. Then, the trainable damping coefficient is obtained through the Softplus function, i.e. .in, This represents the internal damping parameters used for optimization; this parameterization method ensures trainable damping coefficients. It should always be no less than 0 to avoid the exponential damping factor from increasing abnormally as the amplitude of the convolution response increases.

[0213] It should be noted that the damped convolution branch simulates the energy dissipation effect of the loose connection interface of the medical device fixation bracket, which is enhanced by the increase of local response amplitude during collision and friction. This can avoid the strong background resonance response from completely covering the loosening modulation mode of small and medium amplitude.

[0214] S404, Based on Convolutional Response Feature Map The forward convolution response is extracted, and element-wise squared. The intensity of the quadratic response is then adjusted using a trainable asymmetric gain parameter. Simultaneously, a limiting denominator is constructed based on the absolute value of the convolution response and the trainable elastic limiting coefficient. The asymmetric quadratic response branch output is obtained by element-wise division, expressed as:

[0215] ;

[0216] in, Indicates the first Layer Each output channel corresponds to an asymmetric quadratic response branch output; Indicates the first Layer Each output channel has a corresponding trainable asymmetric gain parameter, which is used to adjust the enhancement level of the positive quadratic response. This means retaining positive values ​​in the convolution response element by element and setting negative values ​​to 0; This represents the element-wise square; Represents a matrix of all ones with the same spatial size as the convolutional response feature map; Indicates the first Layer Each output channel has a corresponding trainable elastic limiting coefficient, which is used to adjust the limiting degree of the second-order response based on the absolute value of the convolution response.

[0217] In one implementation, the asymmetric gain parameters can be trained. It can be initialized to 1; trainable elastic limiting coefficient It is a positive parameter, and in the th The output channels share this parameter across all spatial locations. In this implementation, the same Softplus parameterization method as the trainable damping coefficient can be used to ensure the trainable elastic limiting coefficient. Greater than 0.

[0218] It should be noted that secondary enhancement is performed only on the positive portion of the convolutional response, so that the asymmetric secondary response branch can correspond to the contact stiffness hardening characteristics of the medical device fixation bracket connection interface during positive compression. The negative portion of the convolutional response is not included in the secondary enhancement calculation; it is used to reflect the asymmetric difference between positive compression and reverse gap crossing. The amplitude-limiting denominator increases with the absolute value of the convolutional response, which can avoid excessively large element-wise squaring results that could cause numerical instability during training.

[0219] S405. Add the output of the damped convolution branch to the output of the asymmetric quadratic response branch to obtain the output feature map of the fused amplitude-correlated damped response and the positive asymmetric quadratic response.

[0220] Specifically, the output of the damped convolution branch... With asymmetric quadratic response branch output Adding each element one by one, we get the first... Layer The output feature map corresponding to each output channel is represented as follows:

[0221] ;

[0222] in, Indicates the first Layer The output feature map corresponding to each output channel refers to the local loosening feature map after fusing the amplitude-correlated damping response and the positive asymmetric quadratic response.

[0223] It should be noted that, , and All parameters are configured according to the output channel and broadcast along the order frequency direction and the impact activity angle window direction of the corresponding output channel, enabling element-wise operations to be performed on the convolutional response feature map. Different output channels can learn different damping strengths, positive secondary enhancement strengths, and elastic limiting strengths, thereby adapting to different order ranges and different local impact evolution modes.

[0224] S406. Combine the feature maps of all output channels of the current layer, reduce the feature map spatial size through max pooling operation, and expand the receptive range of the next layer.

[0225] Specifically, the first The output feature maps corresponding to all output channels of the layer are combined along the channel direction to obtain the first layer. Layer output feature map , No. Layer output feature map The size is .

[0226] Furthermore, regarding the first Layer output feature map implement Max pooling is used to obtain the input feature map for the next asymmetric damped convolutional layer. In this process, the step size of max pooling is set to 2 in both the order frequency direction and the shock activity angle window direction, which is used to reduce the feature map space size and expand the sensing range of the next asymmetric damped convolutional layer on the original loosened modulation shock spectrum tensor.

[0227] S407: Stack multiple asymmetric damped convolutional layers and pooling layers, and output the predicted probabilities of normal, slight and severe loosening states through global average pooling and fully connected layers.

[0228] Specifically, three asymmetric damped convolutional layers are stacked sequentially. The number of output channels for the first, second, and third asymmetric damped convolutional layers are set to 16, 32, and 64, respectively. Each asymmetric damped convolutional layer is followed by one... Max pooling layer.

[0229] Furthermore, the feature map obtained after max pooling in the third layer is input into a global average pooling layer. The average of all feature values ​​of each output channel in the order frequency direction and the impact activity angle window direction is calculated to obtain a global loosening feature vector of length 64. .in, This represents the global loosening feature vector extracted from three asymmetric damping convolutional layers, used to comprehensively characterize the loosening impact order distribution, impact angle evolution, amplitude-related damping, and positive contact hardening characteristics under different flight conditions.

[0230] Furthermore, the global loosening feature vector Input the classification fully connected layer to obtain the classification score vector. ,Right now .in, This represents the classification score vector, which has a dimension of 3. The weight matrix of the fully connected classification layer is a trainable parameter. This represents the bias vector of the fully connected classification layer, and is a trainable parameter. Classification score vector. The three elements correspond to the normal state, slightly loose state, and severely loose state, respectively.

[0231] Furthermore, regarding the classification score vector Perform the Softmax operation to obtain the predicted probability vector of the loosened state. , is represented as:

[0232] ;

[0233] in, Indicates the first Predicted probability of a loose state; Represents the classification score vector The first in Each category score, Represents the classification score vector The first in Each category score; Categories 1, 2, and 3 correspond to the normal state, slightly loose state, and severely loose state, respectively.

[0234] During inference, the probability vector for predicting the loosened state is... The category with the highest predicted probability was determined to be the loosening state of the medical device fixation stent.

[0235] It should be noted that the asymmetric damped convolutional layer directly introduces an amplitude-correlated exponential damping branch and a positive quadratic response branch after the standard two-dimensional convolution. The amplitude-correlated exponential damping branch is used to suppress excessively strong background resonance while preserving small-to-medium amplitude loosening modulation features, while the positive quadratic response branch is used to enhance the local spectral response related to positive contact hardening. Both types of responses act on the same convolutional response feature map, enabling the formation of nonlinear feature expressions with loosening dynamics without altering the original convolutional spatial relationships.

[0236] It should also be noted that the specific structure of the loose state classification deep learning network is shown in the table below:

[0237] level name Specific configuration Asymmetric damped convolution layer 1 Output channels = 16, 2D convolution kernel size = 3×3, convolution stride = 1, zero padding = 1 Max pooling layer 1 Pooling core size = 2 × 2, step size = 2 Asymmetric damped convolution layer 2 Output channels = 32, 2D convolution kernel size = 3×3, convolution stride = 1, zero padding = 1 Max pooling layer 2 Pooling core size = 2 × 2, step size = 2 Asymmetric damped convolution layer 3 Output channels = 64, 2D convolution kernel size = 3×3, convolution stride = 1, zero padding = 1 Max pooling layer 3 Pooling core size = 2 × 2, step size = 2 Global average pooling layer Perform global averaging on the feature map of each channel. Classification of fully connected layers Output dimension = 3 Softmax layer Mapping 3D classification scores to probability distributions

[0238] like Figure 10 As shown in this embodiment, the instantaneous envelope of the stable vibration signal in the angular domain during takeoff is displayed. The horizontal axis represents the rotor angle in degrees, and the vertical axis represents the comprehensive vibration modulus in meters per second squared. The blue-filled area represents the upper and lower envelopes of the vibration signal. The periodic fluctuations of the envelope range correspond to the amplitude modulation caused by the loosening impact, reflecting the intensity variation law of the loosening modulation feature in the angular domain.

[0239] S5. Based on the classification cross-entropy loss, a loosening damping physical regularization term is introduced to optimize the network parameters.

[0240] When training a deep learning network for classifying loose states using only classification cross-entropy loss, the trainable damping coefficients in the asymmetric damping convolutional layers may only be adjusted according to the classification error, making it difficult to guarantee that the damping modulation results are consistent with the loosening dynamics of medical device fixation stents. After a medical device fixation stent becomes loose, the collision and friction at the connection interface increase, and the effective damping response, formed by the trainable damping coefficients and the amplitude of the current sample convolutional response, is usually higher than that in the normal fastening state.

[0241] The joint loss function consists of batch classification cross-entropy loss and a loosening damping physical regularization term. This invention introduces a loosening damping physical regularization term on top of the classification cross-entropy loss. This term does not directly utilize the loosening state label to constrain the globally shared trainable damping coefficients. Instead, it calculates the average effective damping response related to the sample based on the convolution response amplitude of the trainable damping coefficients and the current original loosening monitoring sample. For slightly loosened and severely loosened states, a penalty is applied when the average effective damping response is lower than the physical reference threshold. For normal states, a penalty is applied based on the degree to which the average effective damping response deviates from 0. The specific steps are as follows:

[0242] S501. Read a training batch from the training sample set, process each original loosening monitoring sample in the training batch through S2 and S3 in sequence to obtain the corresponding loosening modulation shock spectrum tensor, and then input each loosening modulation shock spectrum tensor into the loosening state classification deep learning network to obtain the predicted probabilities corresponding to normal state, slightly loosening state and severely loosening state.

[0243] The number of original loosening monitoring samples contained in a training batch is denoted as . ; will the first The first original loosening monitoring sample was in the first The true one-hot encoded label on the class is denoted as A deep learning network classifies loose states for the first... The first original loosening monitoring sample predicted the first Class prediction probability is denoted as .in, ; Categories 1, 2, and 3 correspond to the normal state, slightly loose state, and severely loose state, respectively.

[0244] When the The first original loosening monitoring sample belongs to the first When class, The true one-hot encoded labels for the other categories are set to 0.

[0245] S502. Read the convolution response and trainable damping coefficient of the last asymmetric damped convolutional layer, and calculate the average effective damping response characterizing the energy dissipation intensity of the sample.

[0246] Specifically, read the first The convolution response feature map of the first original loosening monitoring sample in the last asymmetric damping convolutional layer is obtained, and the trainable damping coefficients corresponding to each output channel of the last asymmetric damping convolutional layer are read. Further, the trainable damping coefficients are multiplied element-wise by the absolute value of the convolution response in the corresponding output channel, and the average is calculated over all output channels and all spatial locations to obtain the first... The average effective damping response of the original loosening monitoring samples is expressed as:

[0247] ;

[0248] in, Indicates the first The average effective damping response of an original loosening monitoring sample in the last asymmetric damping convolutional layer refers to the average result across all effective output channels and all effective spatial locations after the trainable damping coefficient is applied to the convolutional response amplitude of the current original loosening monitoring sample. This represents the number of output channels in the last asymmetric damped convolutional layer, which is 64 in this embodiment. This represents the size of the last convolutional response feature map in the direction of the order frequency point; Indicates the first The effective impact activity angle window orientation size of each original loosening monitoring sample in the last layer of the convolutional response feature map; This indicates the last asymmetric damped convolutional layer. Each output channel has a corresponding trainable damping coefficient; Indicates the first The first original loosening monitoring sample was in the last layer. The output channel, the first The position of the first-order feature and the first The convolution response value at the feature location of the impact active angle window.

[0249] In one implementation, when the loosened modulated impact spectrum tensor within the same training batch is zero-padding to unify the impact activity angle window direction size, the average effective damping response is only summed and averaged over the corresponding effective spatial positions before zero-padding, without taking the batch zero-padding positions into account. To avoid the difference in the number of zeros padded changing the average effective damping response.

[0250] It should be noted that the average effective damping response is influenced by both the globally shared trainable damping coefficients and the convolutional response amplitude of the current original loosening monitoring sample. Therefore, even if different original loosening monitoring samples share the same set of trainable damping coefficients, the normal state, the slightly loosened state, and the severely loosened state can still form different average effective damping responses, avoiding directly constraining the same set of globally shared parameters to conflicting values ​​according to the sample labels.

[0251] It should also be noted that, Its dimensions are the same as those of the vibration response amplitude. Its value is positively correlated with the intensity of energy dissipation caused by loosening as perceived by the network. It does not represent the damping coefficient in the physical world, but rather serves as an abstract characteristic index constrained by physical laws.

[0252] S503. Generate a loosening indicator function based on the sample's actual loosening state label, which is used to select the regularization term form applicable to the physical constraint calculation for that sample.

[0253] Specifically, according to the first A loosening indicator function is generated from the true loosening status labels of the original loosening monitoring samples. ,Right now .in, Indicates the first The loosening indicator function corresponding to each original loosening monitoring sample is used to select the physical constraint form corresponding to that original loosening monitoring sample.

[0254] It should be noted that both the slightly loose and severely loose states are loose samples where the connection interface has experienced a decrease in preload, gaps, collisions, or friction. Therefore, they are both constrained by the lower limit of the average effective damping response. The difference in degree between the slightly loose and severely loose states is further learned by the 3-class cross-entropy loss.

[0255] S504, Set Physical Reference Threshold and regularization coefficient Based on the true unique thermal encoded labels, predicted probabilities, loosening indicator functions, and average effective damping responses of each original loosening monitoring sample, the batch classification cross-entropy loss is added to the loosening damping physical regularization term to obtain the joint loss function, expressed as:

[0256] ;

[0257] in, This represents the joint loss function, where the first term is the batch classification cross-entropy loss, used to improve the classification accuracy of normal, slightly loose, and severely loose states, and the second term is the loosening damping physical regularization term. This represents the damping physical regularization coefficient, used to adjust the relative weight between the batch classification cross-entropy loss and the loosening damping physical regularization term; This represents the effective damping physical reference threshold corresponding to the loosened sample, and .

[0258] In practical implementation, for original loosening monitoring samples in a slightly loose or severely loose state, when hour, This generates a physical penalty and increases the average effective damping response of the original loosening monitoring sample; when At that time, the corresponding loosening physical penalty is set to 0. Furthermore, for the original loosening monitoring sample in a normal state, A penalty is imposed on the non-zero average effective damping response, causing the average effective damping response corresponding to the normal state to approach 0.

[0259] In one implementation, the damping physical canonical coefficient Set to 0.1, physical reference threshold. Set to 0.5. Physical reference threshold. The lower limit of the average effective damping response in the loose state can be determined based on the structural dynamics test results of the medical device fixation bracket under normal tightened and known loose states, so that it matches the actual collision and friction energy dissipation law. For example, multiple sets of samples can be collected under normal tightened state, and their average effective damping response can be calculated. The statistical upper limit (e.g., the mean plus three standard deviations) is set, and this value is ensured to be less than the lowest average effective damping response produced by the sample under slightly loose conditions.

[0260] S505. Calculate the joint loss function using mini-batch backpropagation. The gradients of each trainable parameter of the loose state classification deep learning network are compared, and the network parameters are iteratively updated based on the gradients.

[0261] The updated network parameters include the set of two-dimensional convolution kernels for each asymmetric damped convolutional layer. Trainable damping coefficient Corresponding internal damping parameters, trainable asymmetric gain parameters Trainable elastic limiting coefficient The corresponding internal limiting parameters, and the weight matrix of the classification fully connected layer. and bias vector .

[0262] In one implementation, the Adam optimizer is used to update the network parameters. The initial learning rate can be 0.001, and the training batch contains the number of original loosening monitoring samples. The learning rate can be set to 16, 32, or 64, and the number of training epochs can be between 100 and 200. During training, the learning rate can be adjusted based on the joint loss function on the validation set or the classification accuracy on the validation set, and the network parameters corresponding to the highest classification performance on the validation set can be saved.

[0263] In practice, during training, the trainable damping coefficient is kept at least 0 and the trainable elastic limiting coefficient is kept greater than 0 through Softplus parameterization, ensuring that the exponential damping branch and the elastic limiting branch always have a stable and clear operational relationship. Training is iterated continuously until the joint loss function converges, resulting in the final loose state classification model.

[0264] It should be noted that the loosening damping physical regularization term constrains the average effective damping response formed by the trainable damping coefficient and the convolutional response of the current original loosening monitoring sample, rather than directly constraining the globally shared trainable damping coefficient according to the sample category. This allows the physical constraint to change with the input vibration characteristics, maintaining the network parameter sharing mechanism while enabling the normal state and the loosening state to form a characteristic expression that conforms to the difference in energy dissipation at the connection interface.

[0265] S6. Deploy the trained model to the monitoring device to identify the loosening status of the medical device fixation bracket in real time.

[0266] After completing the training of the final loosening state classification model, the instantaneous frequency extraction and kernel width adaptive equal angle resampling procedure in S2, the loosening modulation shock spectrum tensor construction procedure in S3, and the final loosening state classification model are deployed to helicopter airborne monitoring equipment or ground analysis equipment.

[0267] like Figure 11 As shown, during online operation, the X-axis vibration acceleration components, Y-axis vibration acceleration components, Z-axis vibration acceleration components, and flight status labels are continuously read using the same sensor configuration and sampling frequency as during the training phase. First, a comprehensive vibration modulus sequence is synthesized, and the instantaneous frequency conversion curve is obtained based on the consistency of local spectral peaks and autocorrelation periods. Then, based on the flight status labels, time-domain vibration signals for takeoff, cruise, and landing are generated respectively, and kernel width adaptive equal-angle resampling is performed to obtain the angular domain stable vibration signals for takeoff, cruise, and landing.

[0268] Three stationary vibration signals in the angular domain of flight states were subjected to Teager energy calculation, nonnegation processing, overlapping angle window division, impact activity screening, and local order Fourier transform to obtain the loosening modulated impact spectrum tensor to be identified. The loosening modulated impact spectrum tensor to be identified was input into the final loosening state classification model to obtain the predicted probabilities corresponding to normal state, slight loosening state, and severe loosening state. The category with the highest predicted probability was determined as the loosening state identification result of the current medical device fixation stent.

[0269] The final loosening state classification model eliminates the mixed influence of rotational speed variation patterns in different flight stages by segmenting the flight state, reduces the order drift caused by instantaneous rotor speed fluctuations by kernel width adaptive equal-angle resampling, increases the proportion of loosening collision and friction modulation features in the input tensor by local impact activity screening, and extracts classification features with the meaning of interface damping and positive contact hardening by asymmetric damping convolution and loosening damping physical regularization term. Thus, it can identify the normal state, slightly loosened state and severely loosened state of medical equipment fixation brackets in helicopters during takeoff, cruise and landing.

Claims

1. A method for detecting loosening of helicopter medical equipment fixation brackets based on vibration characteristics, characterized in that, Includes the following steps: S1. Install sensors on the fixed bracket of the helicopter medical equipment and collect vibration data and flight parameter information simultaneously; S2. Adaptively segment the flight state and perform equal-angle resampling with self-compensation for speed fluctuations; S3. Use the Teager energy operator to screen the impact active angle window and construct the loosening modulation impact spectrum tensor; S4. Construct a deep learning network for loose state classification based on asymmetric damped convolution; S5. Based on the classification cross-entropy loss, a loose damping physical regularization term is introduced to optimize the network parameters; S6. Deploy the trained model to the monitoring device to identify the loosening status of the medical device fixation bracket in real time.

2. The method for detecting loosening of helicopter medical equipment fixation brackets based on vibration characteristics according to claim 1, characterized in that, Step S2 includes: S201. Synthesize a comprehensive vibration modulus sequence using triaxial vibration acceleration components and segment the flight state; S202. Adjust the interpolation kernel width according to the instantaneous frequency change rate and perform equal-angle resampling.

3. The method for detecting loosening of helicopter medical equipment fixation brackets based on vibration characteristics according to claim 2, characterized in that, Step S201 includes: S20101. Synchronously read the X-axis vibration acceleration component, Y-axis vibration acceleration component and Z-axis vibration acceleration component according to a unified sampling clock, so that the three vibration acceleration components correspond one-to-one at the discrete sampling time. S20102. Perform short-time Fourier transform on the comprehensive vibration modulus segment of continuous time frames, and extract the instantaneous frequency conversion curve by combining local spectrum peak search and autocorrelation period consistency verification. S20103. Read the flight status tag output by the flight parameter recorder and align the flight status tag with the comprehensive vibration modulus sequence and instantaneous frequency curve according to a unified time reference. S20104. Save the time-domain vibration signal and its instantaneous frequency curve in the takeoff state, the time-domain vibration signal and its instantaneous frequency curve in the cruise state, and the time-domain vibration signal and its instantaneous frequency curve in the landing state, respectively, and ensure that the discrete sampling time, the comprehensive vibration modulus value and the rotor instantaneous frequency correspond one-to-one within the same flight state.

4. The method for detecting loosening of helicopter medical equipment fixation brackets based on vibration characteristics according to claim 2, characterized in that, Step S202 includes: S20201. For any flight state, read its time-domain vibration signal and instantaneous frequency sequence, and integrate the instantaneous frequency along the time direction to obtain the reference phase characterizing the cumulative rotor angle. S20202. Establish an equal-angle sampling grid corresponding to the flight state according to a preset constant angle interval; S20203. Based on the instantaneous frequency conversion curve of the flight state corresponding to the flight state, calculate the instantaneous frequency conversion rate at each discrete sampling time. S20204. For each equal-angle sampling position, calculate its angular deviation from the reference phase, and use a kernel width adaptive sinc interpolation kernel to weight and reconstruct the adjacent time-domain vibration values ​​to obtain the flight state angular domain steady vibration signal corresponding to the flight state. S20205. Perform the above-mentioned kernel width adaptive equal angle resampling on the time-domain vibration signal of takeoff, cruise and landing states respectively to obtain the angular domain stable vibration signal of takeoff, cruise and landing states.

5. The method for detecting loosening of helicopter medical equipment fixation brackets based on vibration characteristics according to claim 1, characterized in that, Step S3 includes: S301. Using the Teager energy operator to screen impact active angle windows; S302. Perform a local order Fourier transform on the impact active angle window and generate a multi-channel spectral tensor.

6. The method for detecting loosening of helicopter medical equipment fixation brackets based on vibration characteristics according to claim 5, characterized in that, Step S301 includes: S30101. For the flight state angular domain steady vibration signal corresponding to the flight state, read the flight state angular domain steady vibration values ​​corresponding to the three adjacent equal angle sampling positions in sequence, and calculate the Teager energy value corresponding to the middle equal angle sampling position. S30102. According to the fixed angle window length and fixed angle window step size, the flight status is... The corresponding non-negative Teager energy sequence is divided into multiple overlapping angle windows, and the overlapping angle windows are numbered according to the order of rotor rotation angle; S30103. Calculate the maximum, minimum and average values ​​of non-negative Teager energy within each overlap angle window, and use the ratio of the energy variation range to the average value to obtain the shock activity index that measures the degree of transient change. S30104. Compare the impact activity index corresponding to each overlapping angle window with the preset activity threshold, retain the overlapping angle windows whose impact activity index is greater than the preset activity threshold, and define the retained overlapping angle windows as impact active angle windows.

7. The method for detecting loosening of helicopter medical equipment fixation brackets based on vibration characteristics according to claim 5, characterized in that, Step S302 includes: S30201. Read the non-negative Teager energy sequence within the impact active angle window, and multiply the non-negative Teager energy sequence point by point with the Hanning window of the same length; S30202. Perform a discrete Fourier transform on the windowed energy sequence and take the square of the modulus to obtain the local loosening modulated impact spectrum that reflects the distribution of loosening impact energy at the rotor order position. S30203. According to the original angular order of the active impact angle windows in the stable vibration signal of the flight state angular domain, arrange the local loosening modulation impact spectra corresponding to each active impact angle window in the same flight state along the dimension of the active impact angle window to obtain the loosening modulation impact spectrum matrix corresponding to the flight state. S30204. Perform shock activity angle window dimension alignment on the takeoff state loose modulation shock spectrum matrix, cruise state loose modulation shock spectrum matrix and landing state loose modulation shock spectrum matrix. S30205. Stack the aligned takeoff state loose modulation shock spectrum matrix, cruise state loose modulation shock spectrum matrix and landing state loose modulation shock spectrum matrix sequentially along the flight state channel direction to obtain the loose modulation shock spectrum tensor.

8. The method for detecting loosening of helicopter medical equipment fixation brackets based on vibration characteristics according to claim 1, characterized in that, Step S4 includes: S401. Input the output feature map of the previous layer into the current asymmetric damped convolutional layer; S402. Perform standard convolution operation on the input feature map using a set of two-dimensional convolution kernels to obtain a convolution response feature map that represents the changes in the combination of local order spectra. S403. Extract the forward convolution response based on the convolution response feature map, calculate the exponential damping factor based on the absolute value of the convolution response and the trainable damping coefficient, and then multiply the forward convolution response and the exponential damping factor element by element to obtain the output of the damped convolution branch. S404. Extract the forward convolution response based on the convolution response feature map, square the forward convolution response element by element, and adjust the intensity of the second response using a trainable asymmetric gain parameter; at the same time, construct the limiting denominator based on the absolute value of the convolution response and the trainable elastic limiting coefficient, and obtain the asymmetric second response branch output by dividing element by element. S405. Add the output of the damped convolution branch to the output of the asymmetric quadratic response branch to obtain the output feature map of the fused amplitude-correlated damped response and the positive asymmetric quadratic response. S406. Combine the feature maps of all output channels of the current layer, reduce the feature map spatial size and expand the receptive range of the next layer through max pooling operation; S407: Stack multiple asymmetric damped convolutional layers and pooling layers, and output the predicted probabilities of normal, slight and severe loosening states through global average pooling and fully connected layers.

9. The method for detecting loosening of helicopter medical equipment fixation brackets based on vibration characteristics according to claim 1, characterized in that, Step S5 includes: S501. Read a training batch from the training sample set, process each original loosening monitoring sample in the training batch through S2 and S3 in sequence to obtain the corresponding loosening modulation shock spectrum tensor, and then input each loosening modulation shock spectrum tensor into the loosening state classification deep learning network to obtain the predicted probabilities corresponding to normal state, slightly loosening state and severely loosening state. S502. Read the convolution response and trainable damping coefficient of the last asymmetric damped convolutional layer, and calculate the average effective damping response that characterizes the energy dissipation intensity of the sample. S503. Generate a loosening indicator function based on the sample's actual loosening state label, which is used to select the appropriate regularization term form for the sample in the physical constraint calculation; S504. Set the physical reference threshold and regularization coefficient. Based on the true unique thermal coding label, predicted probability, loosening indicator function and average effective damping response of each original loosening monitoring sample, add the batch classification cross-entropy loss to the loosening damping physical regularization term to obtain the joint loss function. S505. Mini-batch backpropagation is used to calculate the gradient of the joint loss function relative to each trainable parameter of the loose state classification deep learning network, and the network parameters are updated iteratively based on the gradient.

Images