A rotating machinery state perception method and system based on ultra-wideband radar micro-doppler dynamic structure, an electronic device, a medium and a product

Abstract

This invention provides a method, system, electronic device, medium, and product for state perception of rotating machinery based on the micro-Doppler dynamic structure of ultra-wideband radar, relating to the field of rotating machinery condition monitoring technology. The method includes: acquiring echo sequences of a rotating machinery target area collected by an ultra-wideband radar; filtering target range gates or target range gate groups to construct corresponding slow-time echo sequences; sequentially performing background suppression, phase stabilization, and micro-Doppler dynamic structure construction on the slow-time echo sequences to obtain a micro-Doppler dynamic structure representation; extracting explicit dynamic structure features based on the micro-Doppler dynamic structure representation, and / or constructing implicit representations; and obtaining the state perception result of the rotating machinery based on the dynamic structure features and / or the implicit representations. This invention overcomes the limitation of non-contact monitoring relying on a single feature and improves the robustness of state perception under complex operating conditions.

Description

Technical Field

[0001] This invention relates to the field of rotating machinery condition monitoring technology, and in particular to a rotating machinery condition sensing method, system, electronic device, medium, and product based on an ultra-wideband radar micro-Doppler dynamic structure. Background Technology

[0002] Rotating machinery is widely used in wind power generation, industrial motors, gearboxes, rail transportation, pump stations, compressors, petrochemical equipment, mining equipment, and high-end manufacturing equipment. Common abnormal conditions in rotating machinery during long-term operation include misalignment, eccentricity, loosening, rubbing, lubrication degradation, localized bearing damage, cage abnormalities, localized gear defects, and multi-component fusion failures. Once these abnormal conditions reach an uncontrolled stage, they often lead to unplanned downtime, decreased energy efficiency, increased maintenance costs, and heightened safety risks. Therefore, a sensing solution capable of providing non-contact early warning and continuous condition assessment is needed.

[0003] Existing condition monitoring solutions for rotating machinery mainly include contact and non-contact solutions. Contact solutions typically employ vibration sensors, acoustic emission sensors, temperature sensors, strain gauges, and oil monitoring devices. While they can achieve high signal-to-noise ratios at certain local installation locations, they suffer from drawbacks such as complex wiring, high maintenance costs, limited installation locations, high insulation requirements, poor adaptability to harsh environments, and difficulty in quickly replicating to multiple measurement points. The deployment and long-term maintenance of contact solutions are particularly challenging in high-voltage insulation, high-altitude engine rooms, sealed structures, hazardous areas, and external monitoring scenarios for inaccessible rotating equipment.

[0004] Existing non-contact solutions typically include laser vibrometers, visual micro-motion analysis, and millimeter-wave continuous-wave micro-Doppler monitoring. These solutions primarily focus on continuous displacement, continuous velocity, or the position of a fixed spectral peak on the surface of rotating components. While valuable for macroscopic continuous motion detection, they are susceptible to influences from installation geometry, surface reflection conditions, environmental multipath propagation, background vibration, and obstruction when dealing with low-speed, weak anomalies, multi-component coupling anomalies, or cross-viewpoint structural differences. Particularly under low-speed rotation conditions, anomalies often manifest not only as an enhancement at a single absolute frequency point, but also as changes in target gate position, spectral line reorganization, increased spectral complexity, and decreased consistency across multiple receiving viewpoints. If the solution remains focused on a single spectral peak, a single dominant frequency, or a single sensor combination, truly engineering-significant state information is easily lost.

[0005] In the aforementioned non-contact approaches, directly employing millimeter-wave continuous waves or methods that only estimate surface vibration velocities is generally more effective in scenarios with high directivity, stable reflection points, and a focus on continuous displacement or velocity measurements. However, in external observation of industrial rotating machinery, anomalous responses are often not merely manifested as an enhancement of a fixed frequency shift peak, but simultaneously include changes in target gate position, gate energy redistribution, spectral pattern variations, and cross-view consistency changes. Compared to approaches that only emphasize high carrier frequencies or single-point velocity readings, ultra-wideband radar, with its range gate resolution and target gate gating capabilities, can gate local echoes within the target area, making it more suitable for constructing slow-time dynamic structures around target gates or target gate groups. Summary of the Invention

[0006] To overcome the shortcomings of the prior art, the purpose of this invention is to provide a method, system, electronic device, medium and product for sensing the state of rotating machinery based on the dynamic structure of ultra-wideband radar micro-Doppler. This invention solves the technical problems of existing rotating machinery condition monitoring schemes, such as the difficulty in deployment and maintenance of contact schemes in complex environments, and the difficulty in accurately and stably sensing complex abnormal states such as low speed, weak anomalies and multi-component coupling under environmental interference due to over-reliance on single spectral peak characteristics.

[0007] To achieve the above objectives, the present invention provides the following solution: A method for sensing the state of rotating machinery based on the micro-Doppler dynamic structure of ultra-wideband radar includes: The echo sequence was acquired after an ultra-wideband radar transmitted a signal to the target area of ​​rotating machinery. Based on the echo sequence, a target range gate or a target range gate group is selected, and a slow-time echo sequence corresponding to the target range gate or the target range gate group is constructed. The slow-time echo sequence is preprocessed and a micro-Doppler dynamic structure is constructed to obtain a micro-Doppler dynamic structure characterization. Based on the micro-Doppler dynamic structure characterization, dynamic structural features related to the abnormal state of the rotating machinery are extracted, and / or an implicit characterization for characterizing the abnormal state of the rotating machinery is constructed based on the micro-Doppler dynamic structure characterization or the slow-time echo sequence. The state perception results of the rotating machinery are obtained based on the dynamic structural features and / or the implicit representations.

[0008] A rotating machinery state sensing system based on an ultra-wideband radar micro-Doppler dynamic structure includes: The data acquisition module is used to acquire the echo sequence obtained after the ultra-wideband radar transmits a signal to the target area of ​​the rotating machinery; The target gate construction module is used to filter out target range gates or target range gate groups based on the echo sequence, and construct a slow-time echo sequence corresponding to the target range gate or the target range gate group; The dynamic structure construction module is used to preprocess the slow-time echo sequence and construct the micro-Doppler dynamic structure to obtain the micro-Doppler dynamic structure characterization. The feature extraction module is used to extract dynamic structural features related to the abnormal state of the rotating machinery based on the micro-Doppler dynamic structural characterization, and / or to construct an implicit characterization of the abnormal state of the rotating machinery based on the micro-Doppler dynamic structural characterization or the slow-time echo sequence. The result generation module is used to obtain the state perception result of the rotating machinery based on the dynamic structural features and / or the implicit representation.

[0009] An electronic device, comprising: Ultra-wideband radar module; A processor; and a memory communicatively connected to the processor for storing instructions executable by the processor; The processor implements the method when executing the instructions.

[0010] A computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the method.

[0011] The present invention discloses the following technical effects: This invention provides a method, system, electronic device, medium, and product for rotating machinery state perception based on an ultra-wideband radar micro-Doppler dynamic structure. In scenarios such as wind turbine main shafts, high-voltage insulated rotating equipment, sealed housing equipment, and other environments where stable mounting of vibration sensors is difficult, this invention can achieve basic state perception and early anomaly enhancement detection solely relying on an ultra-wideband radar link, without requiring vibration sensors or other auxiliary sensors as prerequisites. Secondly, the micro-Doppler dynamic structure constructed around a target gate or target gate group can simultaneously characterize spectral complexity, gate group migration, and channel consistency changes, making it more suitable for low-speed, weak anomaly, and multi-component coupled scenarios than single-frequency readings. Thirdly, in implementations using readily available ultra-wideband chip platforms or impulse pulse front-ends, it can balance non-contact deployment, solution replicability, and engineering cost control. Fourthly, when vibration, temperature, load, or SCADA links are already available on-site, this application can use them as optional enhancement channels to trigger cross-validation, thus accommodating both independent single-link and multi-modal enhancement deployment methods.

[0012] This application does not simply interpret the ultra-wideband radar micro-Doppler route as the reading of a certain main frequency peak. Instead, it organizes target gate selection, phase stabilization, spectral structure construction, gate group migration, single receiver dynamic structure analysis, and optional multi-receiver averaging and cross-receiver channel consistency into a complete analysis link, thereby improving the engineering adaptability to low-speed, weak anomaly, and multi-component coupling anomaly scenarios. Attached Figure Description

[0013] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0014] Figure 1 A flowchart of a rotating machinery state sensing method based on an ultra-wideband radar micro-Doppler dynamic structure is provided in an embodiment of the present invention. Figure 2 A flowchart illustrating the construction of a target distance gate or a target distance gate group, provided in an embodiment of the present invention; Figure 3 A flowchart of phase stabilization and micro-Doppler dynamic structure construction provided for an embodiment of the present invention; Figure 4 A flowchart of corner domain or order domain resampling provided in an embodiment of the present invention; Figure 5 A flowchart for constructing multi-receiver collaborative features and cross-receiver channel correlation features is provided in an embodiment of the present invention; Figure 6 This invention provides a flowchart of explicit structural feature extraction and AI black-box end-to-end state reasoning. Figure 7 A flowchart of a multimodal triggered cross-validation method provided in an embodiment of the present invention; Figure 8 This is a schematic diagram illustrating a multi-type rotating machinery adaptation scenario provided by an embodiment of the present invention; Figure 9 This is an overview diagram of the measured dynamic structure of a micro-Doppler based on example data from two receiving channels, provided by an embodiment of the present invention. Figure 10 This invention provides a verification diagram of the dynamic structure of a real sample dual-receiver micro-Doppler system, as shown in the embodiment of the invention. Figure 10 (a) shows the dual-reception normalized slow-time spectra of normal / abnormal samples at the same distances of 70 cm and 90 cm. Figure 10 (b) shows the same-direction changes in dynamic structural indicators such as far-end dynamic proportion, amplitude correlation, high-frequency proportion and spectral entropy under the same distance control.

[0015] Figure 11 A simulation support diagram of the dynamic structure of an ultra-wideband radar under laboratory configuration constraints is provided for an embodiment of the present invention. Figure 12 This is a complementary schematic diagram of an ultra-wideband radar and a vibration sensor in the bearing fault characteristic stage, provided as an embodiment of the present invention. Detailed Implementation

[0016] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0017] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0018] like Figure 1 As shown, this invention provides a method for sensing the state of rotating machinery based on an ultra-wideband radar micro-Doppler dynamic structure, comprising: Step 100: Obtain the echo sequence acquired after the ultra-wideband radar transmits a signal to the target area of ​​the rotating machinery; Step 200: Based on the echo sequence, select the target range gate or target range gate group, and construct a slow-time echo sequence corresponding to the target range gate or target range gate group; Step 300: Preprocess the slow-time echo sequence and construct the micro-Doppler dynamic structure to obtain the micro-Doppler dynamic structure characterization; Step 400: Extract dynamic structural features related to the abnormal state of the rotating machinery based on the micro-Doppler dynamic structural characterization, and / or construct an implicit characterization for the abnormal state of the rotating machinery based on the micro-Doppler dynamic structural characterization or the slow-time echo sequence. Step 500: Obtain the state perception result of the rotating machinery based on the dynamic structural features and / or the implicit representation.

[0019] Furthermore, the dynamic structural features include: Spectral centroid, spectral broadening, spectral entropy, number of active spectral lines, high-frequency energy ratio, sideband structure, target gate migration characteristics, single receiver channel characteristics, multi-receiver average characteristics, and cross-receiver channel correlation characteristics.

[0020] Furthermore, the state perception results include: Anomaly alerts, anomaly types, anomaly location range, severity, health index, degradation trend, maintenance priority, and maintenance prompts.

[0021] Furthermore, the ultra-wideband radar is an impulse pulse radar, wherein the instantaneous bandwidth of the transmitted pulse of the impulse pulse radar is not less than 500MHz, and the pulse width is not greater than 2ns.

[0022] In one specific implementation, the ultra-wideband radar control unit transmits impulse pulse signals to the rotating machinery target area according to a preset pulse repetition frequency. The instantaneous bandwidth of the impulse pulse signal is not less than 500MHz and the pulse width is not greater than 2ns. The receiving end performs low-noise amplification, bandpass filtering, analog-to-digital conversion and baseband demodulation on the echo signal reflected or scattered by the target area and its adjacent structures to obtain a complex echo sequence containing in-phase component I and quadrature component Q.

[0023] For the k-th acquisition frame, the n-th range gate, and the c-th receiving channel, the complex echo can be denoted as x(k,n,c) = I(k,n,c) + jQ(k,n,c). Here, the range gate n can be determined by the echo delay τ_n and the equivalent propagation velocity, and its corresponding distance can be expressed as r_n = c0·τ_n / 2, where c0 is the electromagnetic wave propagation velocity. The system acquires x(k,n,c) sequentially from frame to frame, forming a raw echo tensor with frame number k as the slow time dimension, range gate n as the spatial gating dimension, and receiving channel c as the observation dimension.

[0024] Before proceeding to the subsequent target gate construction step, the system can also perform timestamp alignment, gain normalization, saturation frame removal, static background estimation, and sampling quality gating to ensure that the subsequent slow-time micro-Doppler dynamic structure is built on continuous, same-scale, and quality-controllable echo data.

[0025] Specifically, this parameter setting facilitates obtaining usable range gate gating capability and provides a foundation for subsequently constructing spatially constrained slow-time micro-Doppler structures within target range gates or target gate groups.

[0026] In an optional implementation, the ultra-wideband radar front end can be implemented using an off-the-shelf chip platform conforming to the IEEE 802.15.4z / 4ab standard, in order to reduce system cost and deployment complexity while maintaining target gating capability.

[0027] Furthermore, the step of filtering the target range gate or target range gate group based on the echo sequence includes: Based on the dynamic proportion of distance gates, gate group energy, gate group signal-to-noise ratio, dynamic centroid, gate group stability, and continuous hit rate of target gates, multiple candidate distance gates or candidate distance gate groups in the echo sequence are scored. The target distance gate or the target distance gate group is determined based on the scoring results.

[0028] Specifically, the goal is to ensure that subsequent micro-Doppler analysis is based on gate groups that truly reflect changes in the target state, rather than on arbitrary echo sequences. For multi-gate joint scenarios, the system can also perform weighted fusion of multiple adjacent gate groups to form a joint slow time series.

[0029] The original echo sequence can first undergo background cancellation or normalization to reduce the impact of static background and scale differences. Then, multiple candidate range gates or candidate gate groups are scored based on at least one of the following: range gate dynamic ratio, gate group energy, gate group signal-to-noise ratio, dynamic centroid, gate group stability, or target gate continuous hit rate. When the scoring results meet preset quality gating conditions, one or more candidate gate groups with higher scores are selected as the target gate group, and a slow-time echo sequence is constructed accordingly. For scenarios where adjacent gate groups have the same energy change direction or the target response extends across multiple gates, joint weighting can be performed on multiple adjacent gate groups to form a more stable slow-time input.

[0030] Specifically, in one particular implementation, let the complex echo after background suppression be z(k,n,c), where k=1,...,K represents the frame number, n represents the range gate number, and c represents the receiving channel number. For a candidate gate group G, the gate group energy E(G)=1 / K·Σ_kΣ_{n∈G}Σ_c|z(k,n,c)|^2 can be calculated first. Optionally, a reference gate group R_ref without target response is selected, and the gate group signal-to-noise ratio SNR(G)=10log10((E(G)+ε) / (E(R_ref)+ε)) is calculated, where ε is a constant to prevent the denominator from being zero.

[0031] The dynamic proportion of distance gates can be defined as R_dyn(G)=1 / (K·|G|)·Σ_kΣ_{n∈G}1(|d(k,n)|>θ_dyn), where d(k,n) is the dynamic component obtained after performing slow-time high-pass, sliding background subtraction, or inter-frame difference on |z(k,n,c)| or the multi-channel average amplitude, θ_dyn is the dynamic response threshold, and 1(·) is the indicator function. This index is used to characterize the proportion of distance gates that dynamically change with rotation within the candidate gate group.

[0032] The dynamic centroid can be defined as C(k,G)=Σ_{n∈G}n·A(k,n) / (Σ_{n∈G}A(k,n)+ε), where A(k,n)=Σ_c|z(k,n,c)|^2. Gate stability can be defined as Stab(G)=1 / (1+std_k(C(k,G))), or as the reciprocal of the variance of the center position of the candidate gate group exceeding the energy threshold in consecutive frames. The continuous hit rate of the target gate can be defined as Hit(G)=L_max / K, where L_max is the longest consecutive frame count in which the energy or dynamic proportion of the candidate gate group continuously meets the preset threshold.

[0033] When using multi-indicator weighted scoring, indicators such as E(G), SNR(G), R_dyn(G), Stab(G), and Hit(G) can be normalized to the 0-1 range to obtain e(G), s(G), r(G), st(G), and h(G). Then, the comprehensive score Score(G) = w1·r(G) + w2·s(G) + w3·e(G) + w4·st(G) + w5·h(G) can be calculated. For example, in the early anomaly detection scenario of low-speed rotating equipment, w1 = 0.25, w2 = 0.20, w3 = 0.20, w4 = 0.20, and w5 = 0.15 can be taken; in the field scenario with a low signal-to-noise ratio, the weights of w2 and w4 can be increased. The system selects one or more candidate gate groups that meet the preset quality gating threshold and are ranked high as the target gate group.

[0034] Determine the target gate group G Then, the system extracts G from the original or preprocessed complex echo according to the frame number k. The echo values ​​within the range are calculated, and a slow-time echo sequence is constructed. For a single-gate implementation, y(k,c) = z(k,n) can be set. c), where n Let y(k,c) = Σ_{n∈G} be the target distance gate; for a multi-gate joint implementation, let y(k,c) = Σ_{n∈G} }α_n·z(k,n,c), where α_n is the distance gate weight, which can be obtained by normalization based on the average energy within the gate, the signal-to-noise ratio, or the comprehensive score, and satisfies Σ_{n∈G} α_n=1.

[0035] When the target response expands across multiple adjacent distance gates, slow-time subsequences y_i(k,c) can be constructed from adjacent gate groups G1, G2, ..., Gm respectively. Then, the gate group weights β_i are determined based on the comprehensive score Score(G_i) or energy stability of each gate group, forming a joint slow-time sequence y_joint(k,c) = Σ_iβ_i·y_i(k,c). Thus, each acquisition frame k corresponds to one or more complex slow-time samples, and K consecutive frames form y(1,c), y(2,c), ..., y(K,c). This sequence retains the dynamic information of phase and amplitude changes with rotation state, which is used for subsequent phase stabilization and micro-Doppler dynamic structure construction.

[0036] like Figure 2 As shown, background suppression and normalization are first performed on the original echo sequence; then, the dynamic proportion, gate group energy, dynamic centroid, and consecutive hit rate of each distance gate are calculated; then, multiple candidate gate groups are scored; finally, one or more gate groups with the highest scores are selected as the target gate group. For cases where adjacent gate groups have strong and consistent energy, multiple gate groups can be jointly weighted to form a more stable slow-time input.

[0037] Furthermore, the step of sequentially performing background suppression and phase stabilization processing on the slow-time echo sequence includes: A reference reference for phase compensation is determined, the reference reference including: a statically stable reference gate, adjacent reference gate groups, hardware phase calibration parameters, a common reference channel, and normalized complex rotation compensation; Based on the reference benchmark, phase baseline correction is performed on the slow-time echo sequence to suppress slow-varying drift and non-target phase jitter in the slow-time echo sequence.

[0038] Specifically, the background suppression can include at least one of sliding background modeling, reference gate cancellation, bandpass filtering, high-pass filtering, or low-rank background separation; the phase stabilization can include at least one of statically stable reference gate phase compensation, complex normalized rotation, hardware phase calibration, or common reference channel compensation. After the above processing, discrete Fourier transform, short-time Fourier transform, wavelet transform, order transform, or angular domain transform can be performed on the target gate or target gate group to construct a micro-Doppler dynamic structure representation.

[0039] It should be noted that background suppression and phase stabilization are not required to be performed simultaneously, nor are they limited to a fixed order. When the static background of the echo sequence is weak, the hardware phase stability is high, or the target gate quality meets preset conditions, only background suppression or only phase stabilization can be performed, or one or more equivalent methods such as amplitude normalization, slow-time filtering, reference gate compensation, common channel compensation, complex normalization rotation, and phase alignment across receiving channels can be used to complete the preprocessing.

[0040] In the background suppression step, a sliding background b(k,n,c)=λ·b(k-1,n,c)+(1-λ)·x(k,n,c) can be established for each distance gate and receiving channel, and z(k,n,c)=x(k,n,c)-b(k,n,c) can be calculated, where λ is the background update coefficient; reference gate cancellation, slow-time high-pass filtering or low-rank background separation can also be used to suppress static structures, mounting brackets and slow environmental drift.

[0041] In the micro-Doppler dynamic structure construction step, a short-time Fourier transform Y(m,f,c)=Σ_{l=0}^{L-1}y(mH+l,c)·w(l)·exp(-j2πfl / L) can be performed on the phase-stabilized slow time series y(k,c), where L is the window length, H is the frame shift, and w(l) is the window function. The system can construct a time-frequency diagram based on |Y(m,f,c)| or log(1+|Y(m,f,c)|^2), and further calculate at least one of the following: spectral centroid, spectral broadening, spectral entropy, number of active spectral lines, high-frequency energy ratio, sideband spacing, sideband energy ratio, target gate shift, and cross-receiver channel correlation.

[0042] In the state reasoning step, the aforementioned explicit features can be input into a rule-based thresholding model, a statistical model, or a machine learning model. Alternatively, time-frequency graphs, angular domain spectra, or multi-gate joint spectral tensors can be input into a deep learning model. Thus, the specification discloses both an interpretable engineering algorithm chain and preserves the protection space for the end-to-end model to learn implicit representations.

[0043] It should be noted that the "micro-Doppler dynamic structure" in this application is not limited to a single frequency shift peak, but encompasses a whole set of structural information, including spectral centroid, spectral broadening, spectral entropy, number of active spectral lines, sideband organization, gate migration pattern, and consistency across receiving channels. For low-speed rotating equipment, this structured observation method is often more stable in characterizing abnormal states than a single main peak reading.

[0044] Specifically, such as Figure 3 As shown, phase compensation is first performed on the complex sequence of the target gate group based on statically stable reference gates, common reference channels, or hardware calibration parameters; then, spectral transformation, short-time Fourier transform, wavelet transform, or other time-frequency analysis are performed on the compensated slow-time sequence; finally, the slow-time spectrum, time-frequency plot, or joint spectral structure is output. For low-speed rotating devices, the preferred focus is not on a single main peak, but on the overall changes in spectral centroid, spectral broadening, spectral entropy, sideband organization, and spectral line number.

[0045] Furthermore, the method also includes: Obtain the phase reference data corresponding to the rotating machinery; Based on the phase reference data, the slow-time echo sequence is resampled to the angular domain or the order domain to obtain the micro-Doppler dynamic structure characterization in the angular domain or the order domain. The state perception results of the rotating machinery are obtained based on the micro-Doppler dynamic structure characterization of the angular domain or order domain.

[0046] Furthermore, the phase reference data includes: Encoder pulse, Hall phase signal, key phase signal, and rotation angle phase.

[0047] Furthermore, the microDoppler dynamic structure characterization includes: Slow-time spectrum, time-frequency graph, order spectrum, angular domain spectrum, time-order joint graph, and multi-gate joint spectrum of the target range gate group.

[0048] Specifically, one type of dynamic structural feature can originate from the target gate itself, such as the target gate spectral centroid, spectral broadening, spectral entropy, main peak frequency, mid-to-high frequency energy ratio, and number of active spectral lines. Another type of feature can originate from the gate group structure, such as dynamic proportion changes, dynamic centroid shift, gate group migration, and target gate switching stability. A third type of feature can originate from multi-channel relationships, such as multi-receiver average spectral centroid, multi-receiver average spectral entropy, amplitude correlation, phase correlation, negative correlation spectral line count, and cross-channel consistent direction change characteristics. For implementations with only a single receiving channel, the first two types of features can independently constitute the basis for state perception. In addition, the system can directly organize the micro-Doppler dynamic structural representation, time-frequency diagram, angular domain matrix, multi-gate joint spectral tensor, or target gate group slow-time complex sequence as model input, and the implicit representation can be automatically learned by a machine learning model or deep learning model. As long as the explicit or implicit representation can characterize the state changes of rotating machinery, they all fall within the scope of protection of this application.

[0049] like Figure 4 As shown, the encoder, Hall effect, key phase, or estimated phase is acquired as a phase reference, the slow-time echo sequence is resampled to the angular domain or order domain, and a micro-Doppler structure is constructed under a unified reference domain. This type of implementation is particularly suitable for scenarios such as local anomalies in the outer ring of bearings, gear meshing anomalies, and other scenarios with repetitive relationships to rotational angular positions.

[0050] Furthermore, the implementation path for obtaining the state perception result of the rotating machinery based on the dynamic structural features and / or the implicit representation includes: Path 1: Compare the dynamic structural features with preset rule thresholds or statistical models to output anomaly alarms, trend anomalies, or maintenance prompts; Path 2: Using the dynamic structural features as explicit input, reasoning is performed through a machine learning model to output the anomaly type, risk level, or health index of the rotating machinery; Path 3: Using the slow-time echo sequence, the micro-Doppler dynamic structure representation, or the generated tensor as input, the implicit representation is automatically extracted through a deep learning model, and the state-aware results are output end-to-end.

[0051] Furthermore, the training process of the machine learning model or deep learning model includes: A physical consistency constraint loss is introduced into the loss function to enable the model to learn features that conform to the motion laws of rotating machinery. The sample set is enhanced using simulated echo data generated based on a rotating machinery echo simulation model to optimize the model's recognition accuracy in scenarios with few labels.

[0052] Specifically, the methods for obtaining state-aware results based on dynamic structural features and / or implicit representations can include rule-based thresholding, statistical modeling, machine learning, or deep learning. For scenarios with limited samples and a strong emphasis on interpretability, it is preferable to output anomaly alerts, trend anomalies, or maintenance prompts based on rules and statistics. For scenarios with a large number of samples or more complex fault modes, it is preferable to use machine learning or deep learning models to output anomaly types, risk levels, or health indices based on explicit features, implicit representations, or a combination of both. For scenarios where it is desirable to reduce reliance on manual feature engineering, it is preferable to use end-to-end models to directly learn implicit representations from the target gate group's slow-time echo sequence, time-frequency plot, angular domain spectrum, or order spectrum and output state-aware results.

[0053] When the results from the ultra-wideband radar reach a preset threshold, show insufficient confidence, or require enhanced interpretation, cross-validation can be further triggered using vibration, temperature, load, SCADA, current, oil, or other auxiliary monitoring devices. The purpose of this design is not to make auxiliary sensors mandatory prerequisites, but rather to design multimodal fusion as an optional enhancement path led by the ultra-wideband radar.

[0054] Furthermore, when there are insufficient training samples or imbalanced label samples, the training sample set is enhanced using simulated echo samples generated by the rotating machinery echo simulation model to optimize the recognition accuracy of the machine learning model or deep learning model in scenarios with few labels; wherein, the simulated echo samples are used for model training, optimization or verification, and are not used as necessary inputs for outputting state perception results during runtime.

[0055] The rotating machinery echo simulation model refers to a model that generates complex echo samples that are consistent with or equivalent to the format of the actual acquired echo sequence, based on ultra-wideband radar transmitted signal parameters, received sampling parameters, target distance, rotational speed range, scattering point or scattering surface state, noise level, and abnormal state parameters. For example, the simulated echo sample can be represented as x_sim(k,n,c), where k represents the frame number, n represents the range gate number, and c represents the receiving channel number; x_sim(k,n,c) may include target scattering terms, background scattering terms, noise terms, and periodic or quasi-periodic modulation terms caused by eccentricity, loosening, local damage, lubrication degradation, or cage abnormalities.

[0056] Furthermore, the simulated echo samples can be processed using the same target gate screening, slow time series construction, phase stabilization, and micro-Doppler dynamic structure construction process as the measured samples, and can be used together with the measured samples for model training, pre-training, parameter optimization, or trend rationality verification. It should be noted that the simulated echo samples are only an optional technical means to improve the sufficiency of model training and verify the rationality of dynamic structure trends; the state perception results during the runtime of this application can still be obtained solely based on the echo sequences acquired in real time by the ultra-wideband radar.

[0057] Specifically, such as Figure 5 As shown, when there are at least two receiving channels, spectral centroid, spectral broadening, spectral entropy, number of active spectral lines, and target gate migration features can be extracted for different receiving channels, and multi-receiver average features can be further constructed. Simultaneously, the system calculates amplitude correlation, phase correlation, negative correlation spectral line counts, and cross-channel difference. For the currently verified two-receiver-channel data format, anomalies often manifest not only as increased multi-receiver average spectral complexity but also as a shift from relatively consistent cross-receiver channels to enhanced perspective differentiation. For single-transmit / single-receiver or single-receiver-channel implementations, the system can directly complete the state judgment based on the single-receiver dynamic structure features. Therefore, multi-receiver average features and cross-receiver-channel correlation features are enhanced anomaly evidence, not prerequisites for the validity of this application.

[0058] Furthermore, such as Figure 6 As shown, in this type of implementation, the system can perform state reasoning along two parallel paths: the first path extracts spectral centroid, spectral broadening, spectral entropy, high-frequency energy ratio, main peak frequency, sideband energy ratio, gate migration, and correlation change features from the target gate's micro-Doppler dynamic structure, and inputs these explicit features into a rule engine, statistical model, or neural network; the second path directly inputs the micro-Doppler dynamic structure representation, time-frequency plot, angular domain matrix, multi-gate joint spectral tensor, or target gate slow-time echo sequence into the end-to-end model, which automatically learns the implicit representation and outputs the state-aware results. The two paths can be used independently or fused in parallel to output anomaly alarms, anomaly types, health indices, maintenance priorities, or maintenance prompts.

[0059] like Figure 7 As shown, the ultra-wideband radar first independently performs state perception; when an anomaly probability increases, the level crosses the threshold, or the quality gating is insufficient, it then triggers temperature, vibration, load, SCADA, or other auxiliary links for cross-validation. Unlike writing auxiliary sensors as mandatory inputs, this application emphasizes a triggered fusion architecture dominated by the ultra-wideband radar micro-Doppler link.

[0060] like Figure 8 As shown, this embodiment can cover scenarios such as fan drive trains, industrial motors, pump sets, compressors, gearboxes, rail transit axle boxes, and high-voltage insulated rotating equipment. For different scenarios, a more suitable target gate assembly construction method, micro-Doppler structure representation method, and output aperture can be selected based on installation space, insulation requirements, observation geometry, speed range, and anomaly mechanism.

[0061] Furthermore, the method also includes: Acquire auxiliary monitoring data from the auxiliary monitoring device, including vibration data, temperature data, load data, SCADA data, current data, and oil data; The features extracted based on the auxiliary monitoring data are fused with the dynamic structural features; When the state perception results on the ultra-wideband radar side meet the preset conditions, the auxiliary monitoring device is activated and cross-validation or fusion decision is executed.

[0062] In one specific implementation, the ultra-wideband radar independently completes target gating, micro-Doppler dynamic structure construction, and initial state perception to obtain the initial anomaly probability P_uwb, health index HI_uwb, or maintenance level R_uwb. When P_uwb exceeds the warning threshold, HI_uwb shows a deteriorating trend, quality gating is insufficient, or enhanced fault interpretation is required, the system then restarts or reads auxiliary monitoring data such as vibration, temperature, load, SCADA, current, or oil levels.

[0063] Auxiliary monitoring data can be aligned with the UWB radar echo sequence according to timestamps, speed ranges, or operating conditions, and auxiliary features such as temperature rise rate, vibration peak value, envelope spectrum energy, load fluctuation, current harmonics, or oil particle trends can be extracted. Subsequently, the system can perform cross-validation or fusion decisions, such as increasing the maintenance level when the UWB dynamic structure is abnormal and the vibration envelope energy is rising synchronously, and outputting re-inspection suggestions or reducing the false alarm weight when the UWB is abnormal but the auxiliary data is inconsistent.

[0064] It should be noted that the auxiliary monitoring data is used for false alarm suppression, fault interpretation enhancement, level re-scoring, or maintenance strategy correction, and is an optional enhancement branch; the basic state awareness link of this application can be completed solely by relying on ultra-wideband radar echo sequences.

[0065] Furthermore, the execution of cross-validation or fusion decision includes: The dynamic structural features extracted by the ultra-wideband radar are used as the main diagnostic basis; The auxiliary monitoring data is used to perform false alarm suppression, fault interpretation enhancement, perception level re-scoring, or maintenance strategy correction to modify the state perception results output by the ultra-wideband radar.

[0066] Furthermore, the rotating machinery includes: Bearings, gears, couplings, rotors, spindles, motors, pump sets, compressors, gearboxes, and rail transit axle boxes.

[0067] In one specific implementation, the ultra-wideband radar independently completes target gating, micro-Doppler dynamic structure construction, and initial state perception to obtain the initial anomaly probability P_uwb, health index HI_uwb, or maintenance level R_uwb. When P_uwb exceeds the warning threshold, HI_uwb shows a deteriorating trend, quality gating is insufficient, or enhanced fault interpretation is required, the system then restarts or reads auxiliary monitoring data such as vibration, temperature, load, SCADA, current, or oil levels.

[0068] Auxiliary monitoring data can be aligned with the UWB radar echo sequence according to timestamps, speed ranges, or operating conditions, and auxiliary features such as temperature rise rate, vibration peak value, envelope spectrum energy, load fluctuation, current harmonics, or oil particle trends can be extracted. Subsequently, the system can perform cross-validation or fusion decisions, such as increasing the maintenance level when the UWB dynamic structure is abnormal and the vibration envelope energy is rising synchronously, and outputting re-inspection suggestions or reducing the false alarm weight when the UWB is abnormal but the auxiliary data is inconsistent.

[0069] It should be noted that the auxiliary monitoring data is used for false alarm suppression, fault interpretation enhancement, level re-scoring, or maintenance strategy correction, and is an optional enhancement branch; the basic state awareness link of this application can be completed solely by relying on ultra-wideband radar echo sequences.

[0070] The following specific scenarios provide explanations for the above content: Enhanced perception of geometric and status anomalies in low-speed rotating equipment: In low-speed rotating equipment scenarios, ultra-wideband radar is installed near the visible window of the equipment casing or in a safe observation area, aimed at structures adjacent to the main bearing, coupling housing, gearbox observation window, or other externally observable areas. The system focuses on extracting the target gate's dynamic proportion, dynamic center of gravity shift, spectral centroid, spectral broadening, spectral entropy, and, when multiple receiving channels are available, cross-receiver channel correlation changes. It also outputs misalignment correlation anomalies, eccentricity correlation anomalies, loosening, lubrication anomaly trends, and maintenance priorities. This implementation is particularly suitable as a non-contact enhanced sensing entry point in the current MVP stage.

[0071] Multi-receive channel consistency analysis: In a dual- or multi-receiver deployment scenario, the system extracts dynamic structural features for different receiving channels under the same operating conditions and constructs multi-receiver average spectral centroid, multi-receiver average spectral broadening, multi-receiver average spectral entropy, multi-receiver average number of active spectral lines, and cross-receiver channel correlation features. Under normal operating conditions, each receiving channel typically exhibits high consistency; under abnormal operating conditions, target gate shift, increased spectral structure complexity, and decreased amplitude correlation, decreased phase correlation, or increased negative correlation count often occur simultaneously. This implementation helps to incorporate "cross-view differentiation" itself into the anomaly judgment logic, but it does not mean that a single-transmitter, single-receiver implementation cannot independently fall within the scope of protection of this application.

[0072] Perception of repeating structures in angular or order domains: In scenarios with encoders, Hall effect sensors, or key-phase references, the system resamples the slow-time echo sequence to the angular or order domain and focuses on observing repetitive anomalies at fixed angular positions, repetition order stability, sideband energy ratio distribution, and angular gate migration structures. This implementation is particularly suitable for gears, bearings, and other periodic anomaly scenarios related to rotational angular positions.

[0073] Multimodal triggered rating rescoring: In pump sets, compressors, or complex industrial transmission chain scenarios, the system first performs anomaly warning based on the dynamic structure of ultra-wideband radar micro-Doppler; when the anomaly probability exceeds the threshold or when enhanced interpretation capability is required, it then triggers cross-validation or level re-scoring of temperature, vibration, load, or current links, thereby reducing the false alarm rate and improving the credibility of maintenance recommendations.

[0074] Fusion fault upgrade prompt: In multi-component coupled scenarios, the system detects multiple dynamic structural features shifting simultaneously in an abnormal direction, such as target gate shifting backward, spectral broadening and enhancement, spectral entropy increase, and cross-receiver channel correlation decrease occurring simultaneously. In this case, the system can avoid directly outputting a single fault type; instead, it can first output a fused fault escalation prompt, maintenance priority, or suggested re-inspection results. This type of output is particularly suitable for current engineering deployments and early warning phases in the field.

[0075] AI black box end-to-end state awareness path: In one type of engineering implementation, the system may not predefine a complete explicit feature set. Instead, it can directly input the slow-time complex echo sequence of the target gate or group of target gates, the time-frequency graph, order spectrum, angular domain spectrum, multi-gate joint spectral tensor, or multi-receiver structure tensor into a machine learning model or deep learning model. The model can be a convolutional neural network, a temporal convolutional network, a recurrent network, a Transformer, a graph neural network, an attention network, or a combination thereof. Through training, it automatically learns implicit representations that characterize the abnormal state of rotating machinery and directly outputs anomaly alarms, anomaly type, severity, health index, or maintenance prompts accordingly. This type of implementation is particularly suitable for complex scenarios, scenarios with inconsistent label formats, high costs of manual feature engineering, or scenarios requiring continuous online learning and optimization. In this type of implementation, the system can also introduce gate group attention, cross-channel attention, corner domain position encoding, physical consistency constraint loss, contrastive learning loss, or teacher distillation mechanism into the end-to-end model to enhance the model's efficiency in utilizing target gate transfer, multi-receiver decorrelation, and spectral complexity variations. Therefore, this application not only protects explicitly interpretable algorithm chains but also protects the implementation path of AI black-box directly learning implicit representations and completing state reasoning.

[0076] Before entering the deep learning model, the system can first divide the target gate's slow-time complex sequence y(k,n,c) into multiple slow-time segments according to a preset window length L and frame shift H, and perform a short-time Fourier transform or other time-frequency transform on each segment to obtain the complex spectrum Y(t,f,n,c). Then, the amplitude spectrum, power spectrum, or logarithmic power spectrum P(t,f,n,c)=log(1+|Y(t,f,n,c)| 2 ), and perform mean-variance normalization, max-min normalization or channel-based normalization on P(t,f,n,c).

[0077] For single-receiver, single-target gate implementations, P(t,f) can be reshaped into a T×F×1 input tensor; for multi-receiver implementations, different receiving channels can be concatenated along the channel dimension to form a T×F×C input tensor; for multi-target gate group implementations, the gate group dimension and the receiving channel dimension can be merged into the channel dimension to form a T×F×(|G|·C) multi-gate joint spectral tensor, or the gate group dimension can be retained as an independent dimension and processed using a three-dimensional convolutional network.

[0078] When the number of frames or frequency points of different samples are inconsistent, the system can use truncation, zero-padding, interpolation resampling, or unified frequency axis mapping to adjust the input tensor to a uniform size. After the tensor is input into the network, convolutional layers learn local time-frequency textures, pooling layers reduce dimensionality and enhance robustness, global average pooling or flattening layers form implicit representations, and fully connected layers output anomaly categories, health indices, or maintenance levels. The above tensor construction process gives the "tensor" in the claims a clear data source, dimensional meaning, and network input path.

[0079] More specifically, in one particular implementation, the deep learning model may employ a convolutional neural network structure to automatically learn implicit representations from a time-frequency graph tensor constructed from the slow-time echo sequence of the target gate group. For example, the input tensor size of the convolutional neural network may be T×F×C, where T represents the number of time frames, F represents the number of frequency points, and C represents the number of receiving channels. For a single receiving channel implementation, C can be 1; for a multi-receiver channel implementation, C can be greater than or equal to 2. However, in scenarios with multiple receiving channels and multi-gate joint spectrum, the channel dimension of the input tensor can be expanded to... ,in This represents the set of range gates contained in the target range gate group. This indicates the number of distance gates in the set (for example, if the gate group contains 3 distance gates, then...). The convolutional neural network may sequentially include an input layer, a first convolutional layer, a first pooling layer, a second convolutional layer, a third convolutional layer, a global average pooling layer, a flattening layer, a first fully connected layer, a second fully connected layer, and an output layer. Specifically, the first convolutional layer may have a 3×3 kernel size, a stride of 1, padding of 1, and 32 output channels; the first pooling layer may use max pooling with a 2×2 kernel size; the second convolutional layer may have a 3×3 kernel size and 64 output channels; the third convolutional layer may have a 3×3 kernel size and 128 output channels; the first fully connected layer may contain 128 neurons, and the second fully connected layer may contain 64 neurons; the output layer may use the Softmax activation function, and the number of output nodes corresponds to the number of state-aware result categories, where binary classification tasks may correspond to 2 output nodes, and multi-class classification tasks may correspond to multiple output nodes. The network structure described above is only an example of what can be implemented. In practical applications, the number of convolutional layers, kernel size, number of output channels, and number of neurons in fully connected layers can be adjusted according to computing resources, sample size, and deployment requirements.

[0080] For model training, a training dataset can be constructed first. This dataset includes multiple training samples, each of which can be composed of at least one of the following: a micro-Doppler dynamic structure representation or a time-frequency graph, angular domain spectrum, or multi-gate joint spectral tensor, along with corresponding annotation information. The annotation information can include at least one of abnormal state category, health index, and maintenance level. The training samples can originate from measured data and / or from samples generated based on simulated echo data enhancement. During training, supervised learning can be employed. For classification tasks, cross-entropy loss can be used; for regression tasks, mean squared error loss can be used. Alternatively, physical consistency constraint loss can be selectively introduced as a regularization term to constrain the degree of inconsistency between the model output and the preset dynamic structure index change pattern, target gate migration trend, channel correlation change direction, or other prior physical relationships. The optimizer can use Adam or its variants, with an initial learning rate of 0.001, a batch size of 16, 32, or 64 depending on the number of training samples, and a training epoch of 50 to 200. Early stopping can be used to prevent overfitting. After the model is trained, it can be deployed in electronic devices. During runtime, the micro-Doppler dynamic structure representation or the tensor constructed from it will be input into the model in real time, and the state perception result will be output through forward inference.

[0081] It should be noted that the above-described convolutional neural network structure, training parameters, loss function, and optimization method are merely examples. Those skilled in the art can, based on the disclosure of this application, select recurrent neural networks, Transformers, graph neural networks, temporal convolutional networks, attention networks, or combinations thereof according to specific application scenarios, and adjust the number of network layers, hyperparameters, and training strategies. All of these adjustments will not affect the core technical concept of this application, which revolves around end-to-end learning of implicit representations of the micro-Doppler dynamic structure and outputting the state perception results of rotating machinery.

[0082] More specifically, in the model training process, in addition to the conventional supervision loss (such as cross-entropy loss or mean squared error loss), physical consistency constraint loss can be introduced as a regularization term to make the model output conform to the known physical laws of the state evolution of rotating machinery.

[0083] For example, let the health index output by the model be... (0 represents perfect health, 1 represents severe failure), let a certain dynamic structural feature extracted from the input be... (For example, spectral entropy or target gate shift). Based on physical priors, in the evolution from normal to fault, Should follow Monotonically increasing (or decreasing). The loss due to physical consistency constraints can be defined as: in, These are sample pairs representing different fault levels under the same operating condition. This indicates the direction of feature change. When the direction of the health index change predicted by the model is opposite to the physical law, the above loss is positive, thus penalizing this inconsistency during training.

[0084] In other implementations, the physical consistency constraint loss may also take one of the following forms: Incorporate known physical equations (such as the linear relationship between degradation rate and characteristic change rate) as residual terms into the loss function; Use partial differential equation constraints (such as PDE loss in PINN); Use positive and negative sample pairs in contrastive learning to constrain (e.g., feature embeddings of normal samples should cluster, while faulty samples should disperse).

[0085] Algorithm alternative implementations: In one type of engineering implementation, the dynamic structure of micro-Doppler can be achieved through discrete Fourier transform, short-time Fourier transform, wavelet transform, spectrum construction after empirical mode decomposition, order tracking transform, synchronous compressed time-frequency analysis, or other processing methods that can construct a dynamic structure representation from the slow-time echo sequence of the target gate or target gate group. The determination of the target gate or target gate group can be achieved through rule scoring, cluster scoring, statistical testing, model scoring, or a combination thereof. Phase stabilization can be achieved through statically stable reference gate compensation, common reference channel compensation, complex normalized rotation, hardware calibration compensation, or other equivalent methods. State-aware results can be obtained through rule thresholding, statistical models, traditional machine learning models, deep learning models, or two-stage combined models. Therefore, the focus of this application is on the functional chain and module relationships, rather than limiting any function to a single algorithm implementation.

[0086] Verification of the dynamic structure of a multi-receiver micro-Doppler in practice: Please refer to Figure 9 as well as Figure 10 (a) Figure 10(b) In one implementation based on real samples, normal operating condition samples and abnormal operating condition samples under the same rotational speed and distance conditions can be selected as paired controls. The current example data uses a two-receiver channel configuration. Micro-Doppler dynamic structural features are extracted from the two receiving channels of each sample group, and then multi-receiver average features and cross-receiver channel correlation features are constructed. For example, under the same distance control conditions of 70 cm and 90 cm, abnormal samples, relative to normal samples, can simultaneously exhibit an increased proportion of multi-receiver average far-end dynamics, a shift in the center of gravity of multi-receiver average dynamics, enhanced multi-receiver average spectral complexity, and a decreased cross-receiver channel correlation. These experimental results demonstrate that this application does not merely remain at the level of principle derivation, but can observe structural changes with consistent directionality on the current real samples from two receiving channels; for the single-receiver channel implementation, state perception can still be completed based on the dynamic structural features of a single channel.

[0087] In another type of verification implementation, a rotating machinery echo simulation model can be constructed based on an ultra-wideband radar configuration consistent with or equivalent to the laboratory test platform. The simulation parameters include at least one of the following: center frequency, instantaneous bandwidth, pulse repetition frequency, target distance, rotational speed range, and noise level. A comparative analysis is performed on the target gate slow time series under normal operating conditions and under conditions of eccentricity, loosening, local damage, cage abnormalities, or lubrication degradation trends. These simulation results are used to verify the rationality of trends in dynamic structural characteristics such as spectral entropy, spectral broadening, target gate migration, and cross-channel correlation. They can serve as supplementary technical support beyond measured data, but this application is not limited to relying solely on simulation data.

[0088] Furthermore, in one specific implementation, the determination of the target gate or target gate group is not solely based on human experience, but rather on a structured scoring process of the original echo sequence. The system first performs background cancellation, amplitude normalization, and candidate gate group division on the original echo sequence. Then, for each candidate range gate or candidate gate group, it calculates at least one scoring index among range gate dynamic proportion, gate group energy, gate group signal-to-noise ratio, dynamic centroid, gate group stability, and target gate continuous hit rate. Based on the weighted scoring results, it selects one or more candidate gate groups with higher scores as the target gate group. For cases where the target response extends across multiple adjacent range gates, joint weighting can be performed on multiple adjacent gate groups to form a more stable slow-time input sequence.

[0089] In one specific implementation, phase stabilization can be achieved through statically stable reference gate phase compensation, common reference channel compensation, hardware phase calibration parameter compensation, complex normalization rotation, or a combination thereof. For example, the system can first select a reference distance gate with small amplitude-phase fluctuations under normal operating conditions as the statically stable reference gate, calculate the phase difference between the complex echo of the target gate group and the reference gate, and then perform phase baseline correction on the slow-time complex sequence of the target gate group to suppress slowly varying phase shifts introduced by environmental drift, hardware phase jitter, and non-target slowly varying disturbances. When multiple receiving channels are available, a common reference channel can be further used to perform cross-channel phase alignment on each receiving channel.

[0090] In one specific implementation, the system may select a statically stable reference gate group R_s and calculate the reference phase. _ref(k,c)=arg(Σ_{n∈R_s}z(k,n,c)). For the target gate group slow-time complex sequence y(k,c), phase compensation can be performed: y_p(k,c)=y(k,c)·exp(-j _ref(k,c)) to compensate for hardware phase drift, slow environmental disturbances, or common path phase shift.

[0091] When using complex-normalized rotation compensation, the reference phase θ(k,c) of the target gate group or reference gate group can be calculated first, and the complex sample can be normalized to u(k,c)=y(k,c) / (|y(k,c)|+ε), and then rotation compensation u_p(k,c)=u(k,c)·exp(-jθ(k,c)) can be performed; or the original complex sample can be directly subjected to y_p(k,c)=y(k,c)·exp(-jθ(k,c)). θ(k,c) can be determined by the statically stable reference gate phase, the phase difference of the previous frame, the phase of the common reference channel, or the hardware-calibrated phase.

[0092] In multi-receiver channel scenarios, one can also select a receiver channel as a common reference channel c0 to calculate the cross-channel phase deviation Δ. _c=median_k(arg(y(k,c))-arg(y(k,c0))), and execute y_align(k,c)=y(k,c)·exp(-jΔ) on each receiving channel. _c) to improve the stability of multi-receiver average characteristics and cross-receiver channel correlation characteristics.

[0093] In one specific implementation, the micro-Doppler dynamic structure can be constructed through discrete Fourier transform, short-time Fourier transform, wavelet transform, synchronous compressed time-frequency analysis, order tracking transform, angular domain resampling, or a combination thereof. For example, the system can perform frame-by-frame processing on the slow-time complex sequence of the target gate group and perform spectral transform on each frame to construct a time-frequency map, order spectrum, angular domain spectrum, or a multi-gate joint spectral tensor. Subsequently, the system can extract at least one of the following from the dynamic structure: spectral centroid, spectral broadening, spectral entropy, number of active spectral lines, high-frequency energy proportion, main peak frequency, sideband spacing, sideband energy ratio, target gate migration, dynamic centroid shift, multi-receiver average characteristics, and cross-receiver channel correlation characteristics.

[0094] In one specific implementation, state awareness results can be obtained through rule-based thresholding, statistical models, traditional machine learning models, deep learning models, or a two-stage combined model. For scenarios with few samples and an emphasis on interpretability, it is preferable to output anomaly alarms, trend anomalies, maintenance prompts, or level re-scoring based on rule-based thresholds and statistics. For scenarios with many samples, complex state categories, or requiring online optimization, it is preferable to output anomaly types, risk levels, health indices, or maintenance priorities based on explicit features, implicit representations, or a combination of both through machine learning models or deep learning models.

[0095] It should be noted that the above-described target gate construction, phase stabilization, dynamic structure construction, and state inference processes are only representative implementation methods. Those skilled in the art can make equivalent substitutions or combinations of the scoring indicators, compensation methods, spectrum analysis methods, model structures, and inference strategies based on different equipment types, observation geometry, number of channels, and computational resource conditions. All of these adjustments do not depart from the technical concept of this application, which revolves around constructing a micro-Doppler dynamic structure around an ultra-wideband radar target gate or target gate group and thereby outputting the state perception results of rotating machinery.

[0096] Please refer to Figure 11 , Figure 11 This is a simulation support diagram of the dynamic structure of an ultra-wideband radar under laboratory configuration constraints, provided for an embodiment of the present invention. Figure 11 Trend simulations were performed on normal, early pitting, lubrication degradation, and shaft bending / misalignment conditions under center frequency, bandwidth, and observation distance aperture consistent with or equivalent to the laboratory platform. It is evident that different anomalous states do not correspond to only a single spectral peak enhancement, but rather exhibit different combinations of dynamic structural indices such as eccentricity proxy, scattering roughness proxy, temperature rise rate, envelope entropy, and harmonic energy ratio. This figure is used to support the reasonableness and distinguishability of the trends in the dynamic structural characteristics described in this application, but is not intended to limit this application to a single fault model or a single numerical threshold.

[0097] Reference Figure 12 , Figure 12This is a complementary schematic diagram of an ultra-wideband radar and a vibration sensor in the bearing fault characteristic stage, provided as an embodiment of the present invention. Figure 12 This indicates that, during the normal baseline and early weak anomaly phases, ultra-wideband radar is more likely to obtain anomaly evidence from target gate drift, spectral entropy changes, and gate structure rearrangement, making it suitable for single-link deployments in low-speed, installation-constrained, sealed, or insulated scenarios. During the strong impact and high-energy fault phases, contact vibration sensors provide a more intuitive understanding of impact and high-energy mechanical responses, making them suitable as augmenting evidence. Furthermore, during the maintenance decision-making phase, the two can form a cross-validation relationship. Therefore, the technical solution of this application does not require a vibration link as a prerequisite, but rather allows the ultra-wideband radar to function independently and, when needed, to form a triggered complementarity with the vibration link.

[0098] In conclusion, Figure 9 as well as Figure 10 (a) Figure 10 (b) Provides experimental support based on real two-channel receiver samples. Figure 11 It provides trend simulation support with an equivalent scope to the laboratory platform. Figure 12 This further clarifies the differences in roles and complementary boundaries between ultra-wideband radar and vibration linkage in the bearing fault evolution stage. Therefore, this application is supported by real-world samples, trend simulations, and scenario-level complementary explanations, and does not belong to a purely theoretical abstract scheme.

[0099] It needs to be further explained that, Figure 9 as well as Figure 10 (a) Figure 10 (b) The corresponding experimental diagram is used to support the technical effect that "micro-Doppler dynamic structures can be stably constructed and can characterize abnormal state changes," and is not intended to limit this application to a specific test distance, a specific frequency axis scale, a specific equipment type, or a specific fine fault label. For different equipment, different observation geometry, and different signal processing configurations, those skilled in the art can choose appropriate diagram representation and feature output methods under the technical concept described in this application.

[0100] It should be noted that the micro-Doppler dynamic structure application concept referred to in this application focuses on protecting the slow-time dynamic structure analysis chain surrounding ultra-wideband radar target gates or target gate groups, rather than protecting any single industry scenario, any single frequency feature, or any single auxiliary sensor combination. Any equivalent technical solution that utilizes the echo sequence of ultra-wideband radar target gates or target gate groups to construct a micro-Doppler dynamic structure, and thereby perceives, assesses, locates, warns, re-evaluates, or provides maintenance prompts for abnormal states of rotating machinery, should fall within the scope of the protection concept of this application.

[0101] This embodiment also provides a rotating machinery state sensing system based on an ultra-wideband radar micro-Doppler dynamic structure, including: The data acquisition module is used to acquire the echo sequence obtained after the ultra-wideband radar transmits a signal to the target area of ​​the rotating machinery; The target gate construction module is used to filter out target range gates or target range gate groups based on the echo sequence, and construct a slow-time echo sequence corresponding to the target range gate or the target range gate group; The dynamic structure construction module is used to preprocess the slow-time echo sequence and construct the micro-Doppler dynamic structure to obtain the micro-Doppler dynamic structure characterization. The feature extraction module is used to extract dynamic structural features related to the abnormal state of the rotating machinery based on the micro-Doppler dynamic structural characterization, and / or to construct an implicit characterization of the abnormal state of the rotating machinery based on the micro-Doppler dynamic structural characterization or the slow-time echo sequence. The result generation module is used to obtain the state perception result of the rotating machinery based on the dynamic structural features and / or the implicit representation. The dynamic structural features are selected from one or more of the following features: spectral centroid, spectral broadening, spectral entropy, number of active spectral lines, high-frequency energy ratio, sideband structure, target gate migration features, single receiver channel features, multi-receiver average features, and cross-receiver channel correlation features. The status perception result is selected from one or more of the following: abnormal alarm, abnormal type, abnormal location range, severity, health index, degradation trend, maintenance priority, and maintenance prompt.

[0102] This embodiment also provides an electronic device, including: Ultra-wideband radar module; A processor; and a memory communicatively connected to the processor for storing instructions executable by the processor; Wherein, when the processor executes the instructions, it implements any one of the methods described.

[0103] This embodiment also provides a computer-readable storage medium storing a computer program thereon, characterized in that the computer program, when executed by a processor, implements any one of the methods described above.

[0104] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on the differences from other embodiments. The same or similar parts between the various embodiments can be referred to each other.

[0105] This document uses specific examples to illustrate the principles and implementation methods of the present invention. The descriptions of the above embodiments are only for the purpose of helping to understand the method and core ideas of the present invention. Furthermore, those skilled in the art will recognize that, based on the ideas of the present invention, there will be changes in the specific implementation methods and application scope. Therefore, the content of this specification should not be construed as a limitation of the present invention.

Claims

1. A method for sensing the state of rotating machinery based on an ultra-wideband radar micro-Doppler dynamic structure, characterized in that, include: The echo sequence was acquired after an ultra-wideband radar transmitted a signal to the target area of ​​rotating machinery. Based on the echo sequence, a target range gate or a target range gate group is selected, and a slow-time echo sequence corresponding to the target range gate or the target range gate group is constructed. The slow-time echo sequence is preprocessed and a micro-Doppler dynamic structure is constructed to obtain a micro-Doppler dynamic structure characterization. Based on the micro-Doppler dynamic structure characterization, dynamic structural features related to the abnormal state of the rotating machinery are extracted, and / or an implicit characterization for characterizing the abnormal state of the rotating machinery is constructed based on the micro-Doppler dynamic structure characterization or the slow-time echo sequence. The state perception results of the rotating machinery are obtained based on the dynamic structural features and / or the implicit representations.

2. The method for sensing the state of rotating machinery based on an ultra-wideband radar micro-Doppler dynamic structure according to claim 1, characterized in that, The dynamic structural features include: Spectral centroid, spectral broadening, spectral entropy, number of active spectral lines, high-frequency energy ratio, sideband structure, target gate migration characteristics, single receiver channel characteristics, multi-receiver average characteristics, and cross-receiver channel correlation characteristics.

3. The method for sensing the state of rotating machinery based on an ultra-wideband radar micro-Doppler dynamic structure according to claim 1, characterized in that, The state perception results include: Anomaly alerts, anomaly types, anomaly location range, severity, health index, degradation trend, maintenance priority, and maintenance prompts.

4. The method for sensing the state of rotating machinery based on an ultra-wideband radar micro-Doppler dynamic structure according to claim 1, characterized in that, The ultra-wideband radar is an impulse pulse radar, wherein the instantaneous bandwidth of the transmitted pulse of the impulse pulse radar is not less than 500MHz and the pulse width is not greater than 2ns.

5. The method for sensing the state of rotating machinery based on an ultra-wideband radar micro-Doppler dynamic structure according to claim 1, characterized in that, The step of filtering the target range gate or target range gate group based on the echo sequence includes: Based on the dynamic proportion of distance gates, gate group energy, gate group signal-to-noise ratio, dynamic centroid, gate group stability, and continuous hit rate of target gates, multiple candidate distance gates or candidate distance gate groups in the echo sequence are scored. The target distance gate or the target distance gate group is determined based on the scoring results.

6. The method for sensing the state of rotating machinery based on an ultra-wideband radar micro-Doppler dynamic structure according to claim 1, characterized in that, The preprocessing includes one or more of the following: background suppression processing, phase stabilization processing, amplitude normalization processing, filtering processing, reference gate compensation processing, and common channel compensation processing.

7. The method for sensing the state of rotating machinery based on an ultra-wideband radar micro-Doppler dynamic structure according to claim 1, characterized in that, The method further includes: Obtain the phase reference data corresponding to the rotating machinery; Based on the phase reference data, the slow-time echo sequence is resampled to the angular domain or the order domain to obtain the micro-Doppler dynamic structure characterization in the angular domain or the order domain. The state perception results of the rotating machinery are obtained based on the micro-Doppler dynamic structure characterization of the angular domain or order domain.

8. The method for sensing the state of rotating machinery based on an ultra-wideband radar micro-Doppler dynamic structure according to claim 7, characterized in that, The phase reference data includes: Encoder pulse, Hall phase signal, key phase signal, and rotation angle phase.

9. The method for sensing the state of rotating machinery based on an ultra-wideband radar micro-Doppler dynamic structure according to claim 1, characterized in that, The microDoppler dynamic structure characterization includes: Slow-time spectrum, time-frequency graph, order spectrum, angular domain spectrum, time-order joint graph, and multi-gate joint spectrum of the target range gate group.

10. The method for sensing the state of rotating machinery based on an ultra-wideband radar micro-Doppler dynamic structure according to claim 1, characterized in that, The implementation path for obtaining the state perception result of the rotating machinery based on the dynamic structural features and / or the implicit representation includes: Path 1: Compare the dynamic structural features with preset rule thresholds or statistical models to output anomaly alarms, trend anomalies, or maintenance prompts; Path 2: Using the dynamic structural features as explicit input, reasoning is performed through a machine learning model to output the anomaly type, risk level, or health index of the rotating machinery; Path 3: The slow-time echo sequence, the micro-Doppler dynamic structure representation, or the tensor obtained by converting the micro-Doppler dynamic structure representation is used as input. The implicit representation is automatically extracted through a deep learning model, and the state-aware results are output end-to-end.

11. The method for sensing the state of rotating machinery based on an ultra-wideband radar micro-Doppler dynamic structure according to claim 10, characterized in that, The training process of the machine learning model or deep learning model includes: A physical consistency constraint loss is introduced into the loss function to enable the model to learn features that conform to the motion laws of rotating machinery. The sample set is enhanced using simulated echo data generated based on a rotating machinery echo simulation model to optimize the model's recognition accuracy in scenarios with few labels.

12. The method for sensing the state of rotating machinery based on an ultra-wideband radar micro-Doppler dynamic structure according to claim 1, characterized in that, The method further includes: Acquire auxiliary monitoring data from the auxiliary monitoring device, including vibration data, temperature data, load data, SCADA data, current data, and oil data; The features extracted based on the auxiliary monitoring data are fused with the dynamic structural features; When the state perception results on the ultra-wideband radar side meet the preset conditions, the auxiliary monitoring device is activated and cross-validation or fusion decision is executed.

13. The method for sensing the state of rotating machinery based on an ultra-wideband radar micro-Doppler dynamic structure according to claim 12, characterized in that, The execution of cross-validation or fusion decisions includes: The dynamic structural features extracted by the ultra-wideband radar are used as the main diagnostic basis; The auxiliary monitoring data is used to perform false alarm suppression, fault interpretation enhancement, perception level re-scoring, or maintenance strategy correction to modify the state perception results output by the ultra-wideband radar.

14. The method for sensing the state of rotating machinery based on an ultra-wideband radar micro-Doppler dynamic structure according to claim 1, characterized in that, The rotating machinery includes: Bearings, gears, couplings, rotors, spindles, motors, pump sets, compressors, gearboxes, and rail transit axle boxes.

15. A rotating machinery state sensing system based on an ultra-wideband radar micro-Doppler dynamic structure, characterized in that, include: The data acquisition module is used to acquire the echo sequence obtained after the ultra-wideband radar transmits a signal to the target area of ​​the rotating machinery; The target gate construction module is used to filter out target range gates or target range gate groups based on the echo sequence, and construct a slow-time echo sequence corresponding to the target range gate or the target range gate group; The dynamic structure construction module is used to preprocess the slow-time echo sequence and construct the micro-Doppler dynamic structure to obtain the micro-Doppler dynamic structure characterization. The feature extraction module is used to extract dynamic structural features related to the abnormal state of the rotating machinery based on the micro-Doppler dynamic structural characterization, and / or to construct an implicit characterization of the abnormal state of the rotating machinery based on the micro-Doppler dynamic structural characterization or the slow-time echo sequence. The result generation module is used to obtain the state perception result of the rotating machinery based on the dynamic structural features and / or the implicit representation. The dynamic structural features are selected from one or more of the following features: spectral centroid, spectral broadening, spectral entropy, number of active spectral lines, high-frequency energy ratio, sideband structure, target gate migration features, single receiver channel features, multi-receiver average features, and cross-receiver channel correlation features. The status perception result is selected from one or more of the following: abnormal alarm, abnormal type, abnormal location range, severity, health index, degradation trend, maintenance priority, and maintenance prompt.

16. An electronic device, characterized in that, include: Ultra-wideband radar module; processor; And a memory communicatively connected to the processor for storing instructions that can be executed by the processor; Wherein, when the processor executes the instructions, it implements the method as described in any one of claims 1 to 14.

17. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the method as described in any one of claims 1 to 14.

Images