Transient mechanical event perception method and system based on ultra-wideband radar cir waveform distortion

Abstract

The application provides a method and system for transient mechanical event sensing based on CIR waveform distortion of ultra-wideband radar. The method obtains channel impulse response data collected by the ultra-wideband radar under the condition that the ultra-wideband radar is not in physical contact with the target medium, and outputs the sensing result of the transient mechanical event by using the waveform distortion information represented by the fast-time local waveform change of the target medium related to the internal or surface transient mechanical event of the target medium and relative to the background waveform. The waveform distortion information is caused by the transient normal acceleration and / or equivalent short-time displacement disturbance generated by the transient mechanical event exciting the target surface. The application establishes a direct physical correlation among the transient mechanical event, the short-time dynamic disturbance of the target surface and the channel impulse response waveform distortion of the ultra-wideband radar, captures the weak event signal appearing earlier and having fault or state discrimination value under the non-contact condition, thereby shortening the information link, improving the sensing sensitivity and the interpretation.

Description

Technical Field

[0001] This application relates to the field of ultra-wideband radar signal processing technology, and more specifically, to a transient mechanical event sensing method and system based on ultra-wideband radar CIR waveform distortion. Background Technology

[0002] In scenarios such as industrial equipment health monitoring, cardiac mechanical activity monitoring, structural health monitoring, and high-reliability safety sensing, many anomalies with high diagnostic value do not manifest as slow, continuous macroscopic displacements, but rather as transient mechanical events triggered by sudden changes in local mechanical states, periodic short-term drives, changes in boundary conditions, or rapid movements of internal tissues. These events typically propagate within the target medium or couple to the surface in the form of elastic / viscoelastic stress waves, short-term impacts, or rapid mechanical disturbances, forming transient normal accelerations or equivalent short-term displacement disturbances with extremely short durations and amplitudes. For example, early failures in rotating machinery, such as bearing spalling, gear pitting, and crack initiation, can trigger sudden changes in local contact stiffness and short-term impacts; myocardial contraction and valve opening and closing during cardiac mechanical activity can produce short-term chest wall responses; and crack propagation, delamination, detachment, and impact damage in composite materials or structural components can also induce short-term stress waves. Therefore, if such transient mechanical events can be directly sensed non-contactly, it is hoped that more essential diagnostic information can be obtained at an earlier stage than through traditional continuous vibration observation.

[0003] For sensing the aforementioned transient mechanical events, existing technologies are mainly divided into two categories: contact and non-contact. Contact-based solutions (such as accelerometers, strain gauges, and acoustic emission sensors) offer high signal-to-noise ratios under certain operating conditions, but require physical installation, wiring, and maintenance, making deployment difficult in scenarios involving high-voltage insulation, high-speed rotation, high-temperature corrosion, or sealed structures. Furthermore, the sensor and installation interface alter the original boundary conditions, easily attenuating early, weak, high-frequency events. Multi-point expansion is costly, making it difficult to balance spatial coverage with life-cycle maintenance costs; essentially, it can only acquire local point-measurement responses, lacking natural spatial gating capabilities. Non-contact solutions (such as millimeter-wave micro-Doppler, laser Doppler vibration measurement, and visual micro-motion analysis) primarily estimate surface displacement / velocity through continuous wave phase changes or frequency shifts. Their sensing chain serves continuous motion measurement, not the transient event itself. For mechanical events with extremely short durations, minimal displacement, sparse spectra, and low energy, traditional continuous wave phase demodulation links easily submerge the target signal in environmental vibrations, phase noise, and structural multipath, resulting in insensitivity to early, weak events.

[0004] In the existing technology, methods for sensing transient mechanical events are mainly divided into contact-based and non-contact-based solutions.

[0005] Contact-based solutions typically employ devices such as accelerometers, strain gauges, and acoustic emission sensors, which are directly fixed to the target being measured. While this type of solution offers a high signal-to-noise ratio in certain operating conditions, it suffers from the following limitations: First, it requires physical installation, wiring, and long-term maintenance, making deployment difficult in high-voltage insulation, high-speed rotation, high-temperature corrosion, sealed structures, or hazardous areas. Second, the sensor and mounting interface alter the original boundary conditions, potentially attenuating early weak high-frequency events at the mounting interface or intermediate structures. Third, expanding to multiple measurement points is costly, making it difficult to balance spatial coverage with full lifecycle maintenance costs. Fourth, it essentially acquires a localized point-measurement response at the installation point, making it difficult to naturally obtain spatialized event labels.

[0006] In addition, contact-based solutions typically only provide local point-measurement responses at the installation point, making it difficult to obtain both regional-level responses and event location tags without increasing the number of sensors or deployment complexity.

[0007] Existing non-contact solutions typically employ millimeter-wave micro-Doppler, laser Doppler vibration measurement, or visual micro-motion analysis, primarily acquiring surface displacement, velocity, or vibration frequency through continuous wave phase changes, frequency shifts, or pixel-level micro-motion estimation. While these solutions are valuable for macroscopic continuous motion sensing, their sensing chain essentially serves continuous displacement or velocity measurement, rather than the transient mechanical events themselves. For mechanical events with extremely short durations, minimal displacement, sparse spectra, and low energy, traditional continuous wave phase demodulation links often overwhelm the target signal with environmental vibrations, phase noise, structural multipath, and rotating backgrounds, leading to insensitivity to early, weak events. Summary of the Invention

[0008] The purpose of this application is to provide a transient mechanical event perception method and system based on ultra-wideband radar CIR waveform distortion, in order to solve the problems of limited deployment of contact-based solutions, insensitivity of traditional non-contact solutions to short-term weak events, insufficient physical interpretation, strong sample dependence, and weak cross-scene generalization ability in the prior art, and to establish an integrated technical framework for unified reference domain aggregation, event waveform segment extraction, and collaboration of explicit features, AI paths, and physical constraints.

[0009] In a first aspect, embodiments of this application provide a non-contact method for sensing transient mechanical events based on waveform distortion of ultra-wideband radar channel impulse response, comprising: acquiring channel impulse response data collected after the ultra-wideband radar transmits a pulse signal to the target medium under conditions where there is no physical contact between the ultra-wideband radar and the target medium; extracting waveform distortion information characterized by fast-time local waveform changes relative to the background waveform, which is related to transient mechanical events inside or on the surface of the target medium, based on the channel impulse response data; and obtaining transient mechanical event sensing results using the waveform distortion information; wherein, the fast-time local waveform changes include local shape changes, time delay changes, pulse width changes, rising edge changes, or peak position changes within a single range gate and / or between adjacent range gates; the waveform distortion information is caused by transient normal acceleration and / or equivalent short-time displacement disturbance generated on the surface of the target medium by the transient mechanical event.

[0010] In the above-described scheme, this embodiment no longer limits the radar sensing object to continuous surface displacement or continuous velocity, but directly correlates the fast-time waveform distortion in the impulse response of the ultra-wideband radar channel with the short-time dynamic disturbance caused by transient mechanical events. Therefore, the system can capture early-appearing weak event signals with fault or state discrimination value under non-contact conditions, thereby shortening the information link and improving sensing sensitivity and interpretability.

[0011] Furthermore, existing UWB radar sensing schemes typically use channel impulse response for slow-time analysis, that is, observing amplitude, phase, or envelope changes at different pulse moments after a fixed range gate to extract breathing, displacement, or other continuous motion information. However, the embodiments of this application focus on the local waveform distortion of the CIR on the fast time axis within a single or small number of pulses. These correspond to different signal dimensions, physical causes, and information carriers. The embodiments of this application utilize the local distortion information projected from picosecond-level time delay perturbations onto the fast-time waveform.

[0012] In an optional implementation, the ultra-wideband radar is an impulse pulse radar, whose transmitted pulse instantaneous bandwidth is not less than 500MHz and whose pulse width is not greater than 2ns. In the above scheme, nanosecond-level pulses and high time resolution are the physical basis for sensing picosecond-level echo delay disturbances and nanometer- to micrometer-level transient displacement effects.

[0013] In an optional implementation, determining the waveform distortion information related to the transient mechanical event based on the channel impulse response data includes: performing clutter or background suppression processing, range gate selection, and normalization processing; and further performing at least one of filtering, differential, wavelet decomposition, sparse decomposition, deconvolution, subspace projection, or low-rank-sparse separation to enhance the waveform distortion segments of the candidate event.

[0014] In an optional implementation, the method further includes: acquiring reference parameters corresponding to the dynamic state of the target medium, and resampling the channel impulse response data onto a unified reference domain on a slow time axis based on the reference parameters. The reference parameters include at least one of rotation angle, reciprocating motion position, load phase, external excitation phase, and event triggering time. In the above scheme, by unifying the reference domain, periodic or quasi-periodic events are repeatedly aligned within the same reference domain, thereby improving the weak event detection capability and providing common variables for a unified algorithm framework across scenarios.

[0015] In an optional implementation, obtaining the transient mechanical event perception result based on the waveform distortion information includes: extracting event waveform segments from the waveform distortion information, and extracting at least one explicit physical feature based on the event waveform segments. The explicit physical feature includes at least one of rise time, kurtosis, principal resonant frequency, bandwidth-to-energy ratio, attenuation coefficient, time-frequency ridge parameter, waveform sparsity, energy entropy, and event repetition stability. In the above scheme, the extracted features are no longer empirical black-box statistics, but physical quantities that directly correspond to stress wave propagation, structural resonance, damping attenuation, and event sharpness, thus facilitating the construction of interpretable diagnostic or monitoring rules.

[0016] In an optional implementation, obtaining the transient mechanical event perception result based on the waveform distortion information includes: inputting the channel impulse response data or a data representation constructed based on the channel impulse response data into the transient mechanical event perception model; the transient mechanical event perception model is a trained machine learning model or a deep learning model, and the transient mechanical event perception model outputs the transient mechanical event perception result; wherein, the data representation includes at least one of a distance gate-slow time matrix, a reference domain matrix, and an event slice tensor.

[0017] In an optional implementation, the machine learning model or deep learning model incorporates a constraint loss term during training based on the physical correlation between the transient mechanical event and the waveform distortion information. In this approach, model training is no longer solely label-driven but is constrained by physical consistency, thereby improving generalization ability under few-shot conditions.

[0018] In an optional implementation, the method further includes: generating simulated channel impact response data based on a stress wave propagation model, a target medium transmission model, and a radar echo model, for training, optimizing, or validating a transient mechanical event perception model. In the above scheme, the system can construct a physical information data enhancement pathway, alleviating the problem of difficulty in collecting early fault, rare pathological, and rare structural damage samples from the source.

[0019] In an optional implementation, the target medium is a rotating mechanical structural component, and the transient mechanical event perception result includes at least one of the following: fault type, fault severity, fault angle location, health index, and remaining life.

[0020] In an optional implementation, the target medium is the cardiac mechanical activity observation area corresponding to the chest wall or back of the human body, and the transient mechanical event sensing results include at least one of heart rate, cardiac cycle stability, myocardial mechanical function parameters, and valve opening and closing timing parameters. In the above scheme, this application can be extended to scenarios such as non-contact cardiac mechanical event monitoring, contactless monitoring in hospital wards, elderly care, and continuous mechanical function monitoring of high-risk patients.

[0021] In optional embodiments, the target medium is an aerospace structural component or a civil engineering structure, and the transient mechanical event sensing results include at least one of the following: cracks, delamination, detachment, micro-damage, impact damage, or loosening. In the above solutions, this application further covers condition monitoring scenarios for composite material wings, satellite structural components, storage tanks, bridge towers, wind turbine blades, and large load-bearing components.

[0022] In an optional implementation, an event feature waveform w(t) is extracted from the channel impulse response data, and explicit physical features or model training constraints are extracted based on the event feature waveform; the event feature waveform w(t) is related to the transient normal acceleration. The following approximation relationship is satisfied: In the above scheme, this approximation corresponds to the input-output approximation in the UWB echo acquisition, waveform extraction, feature calculation, and model constraint construction stages. It is used to characterize physical correlations, guide the extraction of explicit physical features from waveform distortion information, or construct model constraints, and is used for explanation of the embodiments. It should be noted that this approximation is not an abstract discovery of natural laws, but rather a signal processing approximation achieved using specific parameters and signal processing procedures of UWB radar in this application's technical solution, rather than being proposed as a separate mathematical formula detached from technical means.

[0023] Secondly, embodiments of this application provide a transient mechanical event sensing system based on waveform distortion of ultra-wideband radar channel impulse response, comprising: a data acquisition module, an event detection module, and a result generation module; wherein, the data acquisition module is used to acquire channel impulse response data acquired by ultra-wideband radar; the event detection module is used to determine or characterize waveform distortion information related to transient mechanical events based on the channel impulse response data; the result generation module is used to obtain transient mechanical event sensing results using the waveform distortion information; the waveform distortion information is a fast-time waveform change relative to the background waveform in the channel impulse response data, and is caused by the transient normal acceleration and / or equivalent short-time displacement disturbance generated on the surface of the target medium by the transient mechanical event.

[0024] In the above scheme, system-level protection enables this application to cover not only the software algorithm, but also the radar perception system architecture coupled with the algorithm, the module collaboration relationship and the engineering deployment method, which is more conducive to supporting industrialization.

[0025] Thirdly, embodiments of this application provide an electronic device, including an ultra-wideband radar module, a processor, and a memory, wherein the processor is communicatively connected to the ultra-wideband radar module, and the memory stores program instructions that can be executed by the processor, which, when executed by the processor, implement the above-described method.

[0026] Fourthly, embodiments of this application provide a non-transitory computer-readable storage medium and a computer program product for implementing the above-described method when a processor executes relevant program instructions.

[0027] In summary, by establishing a direct physical correlation between transient mechanical events, short-term dynamic disturbances on the target surface, and waveform distortion of the impact response of the ultra-wideband radar channel, the embodiments of this application form a unified technical line that can simultaneously support the explanation of physical mechanisms, protection of generalized algorithms, implementation of engineering systems, and commercial application. Attached Figure Description

[0028] To more clearly illustrate the technical solutions of the embodiments of this application, the accompanying drawings used in the embodiments of this application will be briefly introduced below. It should be understood that the following drawings only show some embodiments of this application and should not be regarded as a limitation of the scope. For those skilled in the art, other related drawings can be obtained based on these drawings without creative effort.

[0029] Figure 1 A flowchart illustrating a transient mechanical event sensing method based on ultra-wideband radar channel impulse response waveform distortion, provided for embodiments of this application; Figure 2A schematic diagram illustrating the physical relationship between transient mechanical events, short-term dynamic disturbances on the target surface, and channel impulse response waveform distortion, as provided in the embodiments of this application. Figure 3 A structural block diagram of a transient mechanical event sensing system provided in this application embodiment; Figure 4 A flowchart for diagnosing faults in rotating machinery is provided as an embodiment of this application; Figure 5 A flowchart of transient mechanical event perception and processing based on an end-to-end model is provided for embodiments of this application; Figure 6 A flowchart of physical information data augmentation and model training provided in this application embodiment; Figure 7 A flowchart for monitoring human cardiac mechanical events is provided as an embodiment of this application; Figure 8 A flowchart for structural health monitoring provided in this application embodiment; Figure 9 This application provides a schematic diagram of a unified reference domain and output framework for multiple scenarios. Figure 10 A structural block diagram of an electronic device provided in an embodiment of this application; Figure 11 A semi-quantitative comparison chart of UWB-CIR and continuous wave micro-Doppler routes under hardware-in-the-loop simulation conditions based on actual system parameters conforming to the 802.15.4z / 4ab standard is provided for embodiments of this application. Figure 12 A schematic diagram illustrating a local quasi-linear time-invariant interpretation under a unified reference domain, provided in an embodiment of this application; Figure 13(a) is a comparison diagram of a relative target door at a distance of 70 cm provided in an embodiment of this application; Figure 13(b) is a comparison diagram of a relative target door at a distance of 70 cm provided in an embodiment of this application; Figure 13(c) is a comparison diagram of the strongest time window at a distance of 90 cm provided in an embodiment of this application; Figure 13(d) is a comparison diagram of the strongest time window at a distance of 90 cm provided in the embodiment of this application; Figure 14 A comparison chart of transient energy, dual-receiver amplitude correlation, and negative correlation feature counts for normal / fault samples at the same distance, provided for embodiments of this application; Figure 15(a) is a comparison chart of event alignment amplitudes at a distance of 70 cm provided in an embodiment of this application; Figure 15(b) is a differential amplitude comparison diagram with a distance of 70 cm provided in an embodiment of this application; Figure 15(c) is a comparison chart of event alignment amplitudes at a distance of 90 cm provided in an embodiment of this application; Figure 15(d) is a comparison chart of difference values ​​at a distance of 90 cm provided in an embodiment of this application.

[0030] Icons: 110 - Ultra-wideband radar module; 120 - Data acquisition module; 130 - Preprocessing module; 140 - Synchronization module; 150 - Event detection module; 160 - Feature extraction module; 170 - Model inference module; 180 - Result generation module; 190 - Physical information data augmentation module. Detailed Implementation

[0031] The embodiments of the technical solution of this application will now be described in detail with reference to the accompanying drawings. These embodiments are only used to more clearly illustrate the technical solution of this application and are therefore merely examples, and should not be used to limit the scope of protection of this application.

[0032] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application belongs; the terminology used herein is for the purpose of describing particular embodiments only and is not intended to limit this application; the terms “comprising” and “having”, and any variations thereof, in the specification and the foregoing description of the drawings are intended to cover non-exclusive inclusion.

[0033] In the description of the embodiments of this application, technical terms such as "first" and "second" are used only to distinguish different objects and should not be construed as indicating or implying relative importance or implicitly specifying the number, specific order, or primary and secondary relationship of the indicated technical features. In the description of the embodiments of this application, "multiple" means two or more, unless otherwise explicitly defined.

[0034] In this document, the term "embodiment" means that a particular feature, structure, or characteristic described in connection with an embodiment may be included in at least one embodiment of this application. The appearance of this phrase in various places throughout the specification does not necessarily refer to the same embodiment, nor is it a separate or alternative embodiment mutually exclusive with other embodiments. It will be explicitly and implicitly understood by those skilled in the art that the embodiments described herein can be combined with other embodiments.

[0035] In the description of the embodiments in this application, the term "and / or" is merely a description of the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent: A existing alone, A and B existing simultaneously, and B existing alone. Additionally, the character " / " in this document generally indicates that the preceding and following related objects have an "or" relationship.

[0036] In the description of the embodiments of this application, the term "multiple" refers to two or more (including two), similarly, "multiple sets" refers to two or more (including two sets), and "multiple pieces" refers to two or more (including two pieces).

[0037] In the description of the embodiments of this application, the technical terms "center," "longitudinal," "lateral," "length," "width," "thickness," "upper," "lower," "front," "rear," "left," "right," "vertical," "horizontal," "top," "bottom," "inner," "outer," "clockwise," "counterclockwise," "axial," "radial," and "circumferential" indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings. They are only for the convenience of describing the embodiments of this application and simplifying the description, and are not intended to indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, they should not be construed as limitations on the embodiments of this application.

[0038] In the description of the embodiments of this application, unless otherwise expressly specified and limited, technical terms such as "installation," "connection," "joining," and "fixing" should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral part; they can refer to a mechanical connection or an electrical connection; they can refer to a direct connection or an indirect connection through an intermediate medium; they can refer to the internal communication of two components or the interaction between two components. For those skilled in the art, the specific meaning of the above terms in the embodiments of this application can be understood according to the specific circumstances.

[0039] In industrial equipment health monitoring, cardiac mechanical event monitoring, structural health monitoring, and high-reliability safety sensing scenarios, many anomalies with high diagnostic value do not manifest as slow, continuous macroscopic displacements, but rather as transient mechanical events triggered by sudden changes in local mechanical states, periodic short-term drives, changes in boundary conditions, or rapid movements of internal tissues / structures. These transient mechanical events typically propagate within the target medium or couple to the target surface in the form of elastic stress waves, viscoelastic stress waves, short-term impacts, or rapid mechanical disturbances, forming transient normal accelerations or equivalent short-term displacement disturbances on the target surface with extremely short durations and small amplitudes but high diagnostic value.

[0040] Taking rotating machinery as an example, early failures such as bearing spalling, gear pitting, crack initiation, lubrication deterioration, and assembly loosening initially manifest as sudden changes in local contact stiffness and short-term mechanical impact. Similarly, in the mechanical activity of the human heart, the short-term chest / back mechanical responses caused by rapid myocardial contraction, valve opening and closing, and blood ejection are also exploitable transient mechanical events. In composite structural components, wind turbine blades, bridge components, or pressure vessels, crack propagation, delamination, detachment, local loosening, and impact damage can also induce short-term stress wave propagation. Therefore, the ability to directly perceive these transient mechanical events non-contactly provides a more fundamental diagnostic entry point at an earlier stage than traditional continuous vibration observation.

[0041] In the existing technology, methods for sensing transient mechanical events are mainly divided into contact-based and non-contact-based solutions.

[0042] Contact-based solutions typically employ devices such as accelerometers, strain gauges, and acoustic emission sensors, which are directly fixed to the target being measured. While this type of solution offers a high signal-to-noise ratio in certain operating conditions, it suffers from the following limitations: First, it requires physical installation, wiring, and long-term maintenance, making deployment difficult in high-voltage insulation, high-speed rotation, high-temperature corrosion, sealed structures, or hazardous areas. Second, the sensor and mounting interface alter the original boundary conditions, potentially attenuating early weak high-frequency events at the mounting interface or intermediate structures. Third, expanding to multiple measurement points is costly, making it difficult to balance spatial coverage with full lifecycle maintenance costs. Fourth, it essentially acquires a localized point-measurement response at the installation point, making it difficult to naturally obtain spatialized event labels.

[0043] In addition, contact-based solutions typically only provide local point-measurement responses at the installation point, making it difficult to obtain both regional-level responses and event location tags without increasing the number of sensors or deployment complexity.

[0044] Existing non-contact solutions typically employ millimeter-wave micro-Doppler, laser Doppler vibration measurement, or visual micro-motion analysis, primarily acquiring surface displacement, velocity, or vibration frequency through continuous wave phase changes, frequency shifts, or pixel-level micro-motion estimation. While these solutions are valuable for macroscopic continuous motion sensing, their sensing chain essentially serves continuous displacement or velocity measurement, rather than the transient mechanical events themselves. For mechanical events with extremely short durations, minimal displacement, sparse spectra, and low energy, traditional continuous wave phase demodulation links often overwhelm the target signal with environmental vibrations, phase noise, structural multipath, and rotating backgrounds, leading to insensitivity to early, weak events.

[0045] From a more specific physical perspective, existing millimeter-wave micro-Doppler, laser Doppler vibration measurement, or visual micro-motion analysis schemes typically rely on continuous phase trajectories, continuous frequency shift trajectories, or continuous pixel displacement trajectories within a relatively long time window to complete the measurement. However, the target surface response corresponding to transient elastic / viscoelastic stress waves often occurs only in microseconds or even shorter time intervals, with its equivalent displacement amplitude typically ranging from nanometers to micrometers. The resulting round-trip time delay perturbation is usually only in the femtosecond to picosecond range. For such short-duration, non-stationary, and weak-amplitude perturbations, continuous-wave phase integration links struggle to form stable and demodulated continuous phase evolution. Laser / visual links are also easily affected by surface reflection conditions, occlusion, field jitter, and multipath scintillation, making it difficult to directly provide a stable characterization that corresponds one-to-one with transient mechanical events.

[0046] In other words, the reason existing non-contact solutions do not utilize stress waves is not because the target medium lacks observable responses, but because their signal systems typically rely on continuous waves, long equivalent pulses, or slow-time analysis, lacking a sufficient physical basis for stabilizing picosecond-level time-delay perturbations into fast-time waveform distortions. That is, current technology has not yet established a direct physical and signal processing chain between stress waves or short-term mechanical disturbances, transient normal acceleration and / or equivalent short-term displacement of the target surface, fast perturbations in echo time delay, and fast-time waveform distortions in CIR. Therefore, it cannot stably extract interpretable features directly corresponding to transient mechanical events from the original echo under non-contact conditions. In contrast, UWB impulse radar, with its nanosecond-level short pulses and wide instantaneous bandwidth, is more suitable for achieving fast-time observation and utilization of time-delay perturbations.

[0047] Compared to millimeter-wave micro-Doppler, laser vibrometer, or visual micro-motion analysis, this application's embodiments do not primarily observe continuous phase trajectories, continuous frequency shift trajectories, or continuous pixel displacement trajectories within a long time window. Instead, they utilize the local waveform distortion of UWB impulse radar on a fast time axis to characterize short-term mechanical events. Compared to ultrasonic detection, air-coupled ultrasound, or acoustic emission detection, this application's embodiments do not require constructing acoustic coupling paths or mounting acoustic transducers. Instead, under a non-contact electromagnetic radar system, event perception is achieved through CIR fast time distortion, target range gating, and reference domain aggregation. These differences make this application more suitable for scenarios involving high-voltage insulation, high-speed rotation, sealed encapsulation, surface contamination, difficulty in attaching sensors, or the need for rapid multi-area scanning.

[0048] It should also be noted that existing UWB radar sensing schemes typically use the channel impulse response as a range-amplitude map or a slow-time series with a fixed range gate. They primarily utilize slow-time amplitude changes, slow-time phase changes, or envelope drift to extract information on breathing, presence detection, localization, or macroscopic continuous motion. The embodiments in this application do not focus on the slow-time trajectory after the fixed range gate, but rather on the local shape, delay, pulse width, rise edge, or peak position changes of the CIR on the fast time axis within a single or small number of pulses. This type of fast-time local waveform distortion is triggered by short-term dynamic disturbances caused by transient mechanical events and echo delay perturbations, representing an information dimension different from traditional UWB slow-time analysis.

[0049] Based on this, this application provides a transient mechanical event sensing method based on ultra-wideband radar CIR waveform distortion. This method, under a non-contact electromagnetic radar system, directly utilizes the fast-time local distortion of the CIR of UWB impulse radar to characterize short-term mechanical events. This allows the system to perform event detection, range gating, and reference domain aggregation at a certain observation distance without the need for acoustic coupling media, close-fitting transducer placement, or pre-constructed stable acoustic paths. Especially in scenarios involving high-voltage insulation, high-speed rotation, sealed encapsulation, surface contamination, difficulty in attaching sensors, or the need for rapid multi-region scanning, this approach offers significant advantages in terms of ease of engineering deployment, spatial gating capability, and recurring event aggregation analysis.

[0050] Furthermore, the technical advantages of this application are mainly reflected in the fact that the suppression of continuous low-frequency background vibration, slowly varying attitude drift, non-target range gate clutter, and random disturbances under repetitive operating conditions is more easily integrated with background suppression, reference domain alignment, event detection, and coherent aggregation to form a unified processing closed loop. For the final signal-to-noise ratio and detection effect in specific operating conditions, those skilled in the art can comprehensively design based on radar bandwidth, transmission energy, observation range, target scattering characteristics, electromagnetic multipath, target material, installation geometry, and algorithm configuration. However, this does not affect the technical main line and differentiated value of this application, which uses fast time distortion of channel impulse response (CIR) as an independent information carrier.

[0051] The processing closed loop of this application embodiment can be summarized as follows: a sudden change in the local mechanical state inside or on the surface of the target medium, a periodic short-term drive, or short-term internal mechanical activity first triggers a transient mechanical event; the transient mechanical event propagates inside the target medium and couples to the observation surface, forming a transient normal acceleration and / or equivalent short-term displacement disturbance; the short-term dynamic disturbance causes a rapid perturbation in the UWB echo delay; the rapid perturbation further manifests as a detectable local distortion in the fast-time waveform of the channel impulse response; the system outputs the transient mechanical event perception result through background suppression, target range gate selection, reference domain alignment, event detection, event slicing, explicit feature extraction, or model inference. This closed loop enables the perception result of this application to be traced back to a clear physical chain, rather than an empirical statistical classification detached from the target object.

[0052] It should be noted that, in the embodiments of this application, "fast time" refers to the sampling axis of the echo in the propagation delay direction after the same UWB pulse transmission, which can correspond to the range gate, range bin, or tap index in discrete implementation. "Slow time" refers to the axis of multiple CIR observation sequences obtained according to the pulse repetition interval, sampling frame period, or continuous observation sequence, used to describe the changes of CIR samples within the same range gate, adjacent range gates, or target gate group with continuous observation time, rotation angle, heartbeat phase, or external excitation phase. Further spectral analysis, time-frequency analysis, or coherent processing of the slow time series can yield Doppler, micro-Doppler, or reference domain aggregation features; however, the slow time axis itself is not equivalent to the Doppler dimension. The CIR fast-time local waveform distortion mentioned in the embodiments of this application mainly refers to the local morphological changes of CIR in the propagation delay / range gate / tap direction, and not just the slow-time amplitude, phase, or Doppler spectrum changes. The CIR data can be complex I / Q samples, amplitude samples, phase samples, power delay spectrum, range gate sequence, relative tap index, or data representation derived therefrom. All of the above data representations can be used as input formats for extracting fast-time local waveform distortion information in the embodiments of this application; however, the embodiments of this application are not limited to using any one fixed data format.

[0053] It should be noted that in the UWB impulse radar system of this application embodiment, the fast time dimension is related to the propagation delay sampling of a single pulse echo, mainly corresponding to range resolution, range gate, or tap structure; the slow time dimension is related to continuous observation of multiple pulses or multiple frames of CIR, mainly corresponding to the change of the same range gate or target gate group with the observation sequence. Doppler or micro-Doppler is usually the result of performing spectral or time-frequency processing on the slow time series, therefore the slow time axis itself cannot be directly equated to the Doppler dimension. The difference between this application embodiment and traditional slow-time Doppler or micro-Doppler analysis is that this application embodiment first uses the local waveform distortion of CIR in the fast time direction as the information carrier of transient mechanical events, and then performs continuous observation, aggregation, or model inference on the distortion event in the slow time or unified reference domain as needed.

[0054] To better understand the difference between existing UWB slow-time analysis and the fast-time waveform distortion sensing of this application, please refer to Table 1.

[0055]

[0056] Meanwhile, while data-driven methods that rely solely on large-sample training can achieve high recognition rates on certain closed datasets, their feature extraction logic often lacks a clear physical basis. Under conditions of sample scarcity, changing operating conditions, media variations, changes in installation location, or cross-scenario migration, the generalization ability of purely data-driven methods is significantly limited, and they struggle to meet the reliability and interpretability requirements of high-value industrial scenarios.

[0057] The core technical concept of this application lies in the fact that sudden changes in the local mechanical state, periodic short-term driving, or short-term internal mechanical activity within or on the surface of a target medium can trigger transient mechanical events. These transient mechanical events manifest as transient normal acceleration and / or equivalent short-term displacement disturbances on the target surface. These short-term dynamic disturbances further cause rapid perturbations in the ultra-wideband radar echo delay, which are manifested as observable waveform distortions in the fast-time waveform of the channel impulse response. By explicitly analyzing or implicitly learning this waveform distortion, the perception results of the transient mechanical events can be obtained.

[0058] In some stress-wave-dominated implementations, the aforementioned transient mechanical events specifically manifest as transient elastic stress wave events or transient viscoelastic stress wave events; in some physiological mechanical activity scenarios, the aforementioned transient mechanical events specifically manifest as short-term mechanical responses of the chest wall or back caused by myocardial contraction, valve opening and closing, and blood ejection; in some structural monitoring scenarios, the aforementioned transient mechanical events specifically manifest as short-term structural fluctuations caused by crack propagation, local delamination, bending wave response, or impact loads. Although the specific event sources differ, they share the same observable physical consequence, namely, the formation of short-term dynamic disturbances on the target surface and the resulting fast-time waveform distortion of the CIR, and therefore still belong to the same inventive concept.

[0059] In this embodiment, waveform distortion information refers to the following: under the same radar parameters and target geometry configuration, when there is no transient mechanical event, the channel impulse response waveform presents a relatively stable background shape; when a transient mechanical event occurs, the local shape of the channel impulse response waveform on the fast time axis undergoes a detectable change relative to the background shape. The change can be manifested as amplitude change, delay change, pulse width change, rise edge steepness change of the waveform within a single range gate, or local migration of waveforms between multiple range gates. The above changes are collectively referred to as the waveform distortion information described in this application.

[0060] In one implementation, the channel impulse response data may include multi-level data representations. The first level consists of raw or near-raw CIR samples, including complex I / Q samples, amplitude samples, phase samples, or power delay spectra. The second level is a range-gate or tap-related representation, including range gate index, relative tap index, CIR window start position, first path index, main peak index, target gate range, or equivalent delay position. The third level is frame-level or channel-level metadata, including timestamps, frame numbers, pulse numbers, transmit / receive antenna identifiers, channel identifiers, received power indicators, noise estimation, channel quality indicators, synchronization status identifiers, or temperature / clock status identifiers. The fourth level is an algorithm input representation derived from the above data, including a range-gate-slow-time matrix, a reference domain matrix, an event slice tensor, differential CIR, normalized CIR, background-suppressed CIR, or a multi-channel combined tensor. This application can perform target range gate selection, background suppression, fast-time waveform distortion extraction, reference domain alignment, event slicing, explicit feature extraction, or model inference based on any one or more of the above-mentioned data representations.

[0061] Based on the above technical concepts, for feature construction, results can be obtained through event detection, waveform slicing, explicit feature extraction, and rule-based or lightweight model inference; alternatively, the channel impulse response matrix can be directly input into a deep learning model, which will automatically learn implicit features related to transient mechanical events.

[0062] For event enhancement, candidate events can be directly enhanced through background cancellation, distance gate difference, and bandpass filtering; fast time distortion can also be enhanced through wavelet decomposition, sparse decomposition, deconvolution, or subspace projection; weak events can also be reconstructed through low-rank sparse separation or event prior templates.

[0063] It should be noted that the core purpose of the aforementioned event enhancement path is to suppress static clutter, slowly varying background, multipath remnants, and broadband noise unrelated to transient mechanical events, rather than limiting it to a specific fixed algorithm. Specifically, when the scene background is relatively stable and the target event exhibits recurring characteristics, background cancellation or range gate difference methods are preferred. When the main frequency band or structural resonance band of the target event is known, high-pass filtering, band-pass filtering, or matched filtering can be used. When clutter and event components are severely superimposed, and there is a significant low-rank background or sparse event structure, sparse decomposition, low-rank-sparse separation, or subspace projection can be used. When it is desired to recover a more direct event pattern from the distorted waveform, deconvolution or template correction can be used. Any method that achieves the functional goal of suppressing irrelevant background and enhancing event-related waveform distortion from the original CIR can be considered an equivalent implementation of the clutter removal or background suppression processing described in this application.

[0064] The reference parameters in this application embodiment can be one or more of the following: rotary machinery angle, reciprocating motion position, external synchronization trigger time, load phase, heartbeat synchronization time, and external excitation time. Although the application objects are different, they all share the steps of reference domain alignment, waveform distortion extraction, and transient mechanical event sensing output.

[0065] Finally, during the training and modeling process, the model can be trained entirely based on real data; physical information can be introduced to generate samples based on real data; the aforementioned physical correlations can be explicitly written into the loss function as constraints; and explicit physical features and implicit deep features can be fused in a series, parallel, or hierarchical manner.

[0066] It should be noted that although the specific physical mechanisms of transient mechanical events differ in different scenarios—for example, elastic stress waves in metals or composite materials, viscoelastic stress waves in soft tissues, bending waves or localized mechanical impacts in complex structures—these events all generate transient normal accelerations or equivalent short-term displacement disturbances on the target surface that are short in duration, small in amplitude, and wide in frequency spectrum. The embodiments of this application utilize this common, observable physical effect: the rapid modulation of the UWB radar echo delay by short-term dynamic disturbances and the waveform distortion caused in CIR. Therefore, the implementation methods in different scenarios belong to the same inventive concept.

[0067] Please refer to Figure 1 and Figure 3 , Figure 1A general flowchart of a transient mechanical event sensing method based on ultra-wideband radar channel impulse response waveform distortion provided in this application embodiment. The method includes: S101: Under the condition that the ultra-wideband radar has no physical contact with the target medium, acquire the channel impulse response data collected after the ultra-wideband radar transmits a pulse signal to the target medium.

[0068] S102: Perform clutter or background suppression processing, range gate selection, and normalization processing (optional) on the channel impulse response data. The clutter or background suppression processing may employ background cancellation, high-pass filtering, band-pass filtering, matched filtering, wavelet decomposition, sparse decomposition, deconvolution, low-rank-sparse separation, subspace projection, or other algorithms capable of suppressing static or slowly varying clutter. In some embodiments, a sliding background cancellation strategy may be used. It should be understood that sliding background cancellation is not the only necessary clutter reduction method required by this application, but merely an optional implementation for ease of explanation. Any method whose functional objective is to suppress static clutter, slowly varying background, multipath remnants, or broadband noise unrelated to transient mechanical events from the original CIR, and to enhance event-correlated fast-time waveform distortion, can be considered an equivalent implementation of the clutter or background suppression processing described in this application.

[0069] S103: Obtain the reference parameters of the target medium and resample the channel impulse response data to a unified reference domain.

[0070] S104: Detect candidate waveform distortion events associated with transient mechanical events.

[0071] S105: Extract event waveform fragments, explicit physical features, or construct model input tensors.

[0072] S106: Output fault diagnosis results, cardiac mechanical parameters, structural health assessment results, or other transient mechanical event perception results.

[0073] Specifically, in S101, the ultra-wideband radar module 110 can be a UWB impulse radar. The UWB impulse radar has a radar transmit pulse width of no more than 2 ns and an instantaneous bandwidth of no less than 500 MHz, to ensure the system is sufficiently sensitive to picosecond-level round-trip delay perturbations and nanometer- to micrometer-level short-term displacement effects. Let the acquired channel impulse response sequence be h[n,m], where n represents the fast time index or range gate index, and m represents the slow time index or pulse number.

[0074] To facilitate understanding of the physical basis of this application, the following step-by-step derivation and explanation of the formation process of transient mechanical events, short-term dynamic disturbances of the target surface, and CIR fast-time waveform distortion, in conjunction with the stress wave-dominated implementation method, will be provided.

[0075] By establishing a stationary background CIR model and applying matched filtering or direct sampling, under the simplified condition of a single dominant scattering center, the channel impulse response corresponding to a stationary background can be approximated as: .

[0076] in, The target scattering coefficient, For the transmitted pulse waveform, For average round-trip time, The average distance between the radar and the target surface. At the speed of light, This represents the composite term of static clutter, slowly varying background, and electronic noise.

[0077] Step 1: Establish an equivalent excitation model for transient mechanical events.

[0078] When transient mechanical events such as abrupt changes in local mechanical properties, changes in local contact stiffness, crack propagation, spalling impact, rapid myocardial contraction, or valve opening and closing occur within the target medium, a transient elastic / viscoelastic stress wave source can be equivalently formed. In a class of repeatable impact scenarios, this stress wave source can be written as: .

[0079] in, This represents the intensity of the k-th stress wave. This represents the average period or reference interval of recurring events. This represents the offset of the k-th event relative to the reference time. The above expression represents the mechanical waveform of a single stress wave, with the symbol * indicating convolution. For aperiodic transient events, the reference interval in the above expression serves only as a uniform representation and does not constitute a requirement that the event must occur periodically.

[0080] Step 2: Establish the mapping relationship between stress wave and surface transient acceleration.

[0081] The aforementioned stress waves propagate, reflect, and superimpose within the target medium, and are coupled to the radar-illuminated surface through the medium transfer function, forming the transient normal acceleration of the target surface. It can be further expressed as: .

[0082] in, This represents the equivalent transfer function of the target medium from its internal event sources to the observed surface. This represents the dynamic parameters or reference parameters of the medium. In the context of rotating machinery, It can be the rotation angle, rotational speed phase, or load phase; in cardiac mechanical event monitoring scenarios... It can be the phase of the cardiac cycle or the trigger moment; in structural health monitoring scenarios... These can be excitation phase, load state, or boundary condition parameters. It should be noted that... It is a non-stationary, pulsed, transient acceleration with an extremely short duration, rather than the approximately stationary continuous acceleration found in traditional continuous vibration monitoring.

[0083] Step 3: Establish the relationship between surface transient acceleration and equivalent short-time displacement disturbance.

[0084] For short-time disturbances in the observation area that can be approximated as a local rigid body, this transient normal acceleration It can be equivalent to a global normal short-time displacement disturbance in the region through quadratic time integration. In a simplified description, it can be written as: .

[0085] Because stress wave events typically exhibit high frequency, short duration, and weak amplitude characteristics, the In many operating conditions, the displacement is only at the nanometer to micrometer level and the duration is only at the microsecond level or even shorter. Displacements of this magnitude are usually difficult to observe stably in radar links that rely on continuous phase integration, but for UWB impulse pulse links with sufficient time resolution, they can still be sensed by the local morphological changes of echo delay perturbations in fast-time waveforms.

[0086] Step 4: Establish the relationship between short-time displacement disturbances, echo delay perturbations, and CIR morphology changes.

[0087] When the short-time displacement disturbance exists At that time, the echo delay changes from the rest state. The change is as follows: .

[0088] Therefore, receiving the echo can be represented as: .

[0089] If we observe the CIR corresponding to each transmission according to the pulse sequence number m, then the CIR waveform at the m-th slow time sampling moment tm can be written as: .

[0090] when When the amplitude is much smaller than the equivalent range resolution cell corresponding to a single range gate, the above-mentioned time delay change usually does not cause the main scattering peak to jump between range gates by an integer order. Instead, it mainly manifests as measurable changes in waveform shape, rising edge, pulse width, and local peak position within a single or a few adjacent range gates. This change is the CIR fast time waveform distortion utilized in this application.

[0091] Step 5: Derive the CIR distortion term under small displacement conditions.

[0092] Under the small displacement approximation condition, let: .

[0093] The actual CIR can then be further written as: .

[0094] right exist Performing a first-order Taylor expansion in the vicinity, we obtain: .

[0095] Therefore, the distortion component can be defined as: .

[0096] When the background term is significantly suppressed after clutter removal or background suppression, range gate selection, and matched filtering, the dominant term of the distortion component is... The relationship is approximately linear, which indicates that the waveform distortion is not random noise, but a measurable output directly modulated by short-term dynamic disturbances on the target surface.

[0097] Step 6: Construct event characteristic waveforms related to transient normal acceleration.

[0098] because With transient normal acceleration They satisfy an integral relationship, and It is a kernel function determined by known radar transmitted pulses. After matched filtering, weighted deconvolution, pulse shaping compensation or equivalent impulse response reconstruction, an event characteristic waveform w(t) related to transient normal acceleration can be constructed from the distorted components.

[0099] Furthermore, from a signal processing perspective, The corresponding differential kernel is equivalent in the frequency domain to applying frequency weighting to the pulse spectrum; however, in practical engineering implementations, the frequency-weighted response can be approximated as similar to the response obtained from the pulse in the frequency domain through matched filtering, pulse compression, and equivalent kernel reconstruction. A proportionally equivalent convolution kernel. Therefore, in a preferred embodiment, the input-output relationship can be characterized by the following approximation: .

[0100] Where w(t) represents the event feature waveform extracted from CIR distortion. This represents the transient normal acceleration on the target surface caused by a transient mechanical event. c represents the speed of light, and the symbol * represents convolution operation. For fast time variables, The above approximation shows that the event waveform recovered from CIR distortion is not an arbitrary empirical feature, but rather the output of the transient normal acceleration after passing through a linear system related to the radar pulse. Furthermore, the above approximation is used to guide the extraction of explicit physical features from waveform distortion information or to construct model training constraints, and does not constitute a limitation requiring explicit solution of this formula before implementing this application.

[0101] As deduced above, a relatable input-output relationship exists between CIR fast-time waveform distortion and the transient normal acceleration of the target surface. Based on this relationship, explicit physical features such as rise time, attenuation coefficient, principal resonant frequency, bandwidth-to-energy ratio, and time-frequency ridges can be extracted from the waveform distortion. This relationship can also be incorporated into the physical consistency constraints of model training. Correspondingly, factors such as stress wave intensity, propagation attenuation, reflection path, and medium boundary conditions are collectively mapped to changes in the local morphology of the CIR waveform. Therefore, the event features obtained from the waveform distortion have clear technical implications, rather than being empirical statistics detached from the measured object.

[0102] In one alternative implementation, the ultra-wideband radar employs a UWB impulse radar system. Compared to continuous wave, frequency-modulated continuous wave, or conventional millimeter-wave links that primarily rely on slow-time trajectory analysis, UWB impulse radar does not use long-term phase integration or continuous frequency shift trajectories as its main information carrier. Instead, it focuses on the local morphological changes, time delay perturbations, and range-gate waveform changes of nanosecond-level short pulse echoes on a fast time axis. Because the transient mechanical events targeted in this application are short in duration, weak in amplitude, and non-stationary, the associated perturbations are often insufficient to form a stable and demodulate continuous phase evolution in a continuous wave link. However, UWB impulse radar possesses a sufficiently short pulse width and a sufficiently wide signal bandwidth, enabling it to directly map these short-time dynamic perturbations into CIR fast-time waveform distortion, thereby achieving direct sensing of transient mechanical events.

[0103] From a signal system perspective, UWB impulse radar can be understood as a time-domain characteristic radar. Its observation focus is not on long-term continuous motion trajectories, but on the local distortion information of single or a small number of pulse echoes in a fast time dimension. The direct sensing referred to in this application means directly using the fast-time waveform distortion of CIR to characterize transient mechanical events, rather than first simplifying the target into continuous displacement or continuous velocity and then making indirect inferences.

[0104] It should be noted that the above formulas, variable definitions, and basic derivations are only used to describe the physical connections and implementation mechanisms in the technical solutions of this application. Their purpose is to illustrate how to complete non-contact transient mechanical event perception, feature extraction, model training, and result output based on the physical connections. For the object of protection of this application, what is protected is the method, system, and electronic device that utilizes the physical connections to implement perception, rather than the abstract formulas themselves detached from specific technical means, target objects, and processing procedures.

[0105] In S102, the raw channel impulse response data output by the data acquisition module 120 typically includes static clutter, slowly varying environmental background, and electronic noise. S102 may include the following sub-steps: S102-1: Perform clutter or background suppression processing on the raw channel impulse response data to suppress static clutter and slowly varying background that are unrelated to transient mechanical events.

[0106] S102-2: Perform target range gate selection or target range gate group locking on the processed data to preserve the fast-time waveform corresponding to the target observation area.

[0107] S102-3: Perform normalization processing on the data within the target distance gate or target distance gate group to facilitate subsequent event detection, feature extraction or model input construction.

[0108] The clutter removal or background suppression processing can employ background cancellation, high-pass filtering, band-pass filtering, matched filtering, wavelet decomposition, sparse decomposition, low-rank-sparse separation, subspace projection, or other algorithms capable of suppressing static or slowly varying clutter. In some implementations, a sliding background cancellation strategy can be used, the specific formula of which is: .

[0109] Where L is the background window length, n is the fast time index (range gate index), ranging from 1 to N, where N is determined by the radar sampling rate and detection range, typically 128 to 1024. m is the slow time index (pulse number), ranging from 1 to M, where M depends on the acquisition duration and pulse repetition frequency, with no fixed upper limit. The formula means: h[n,m] represents the CIR amplitude (or complex sample points) of the m-th pulse sampled at the n-th range gate. This represents the CIR sequence after background suppression.

[0110] It should be understood that sliding background cancellation is not the only necessary clutter removal method required by this application, but merely an optional implementation method for ease of explanation. For rotating machinery, cardiac mechanical event monitoring, and structural vibration scenarios, high-pass filtering, band-pass filtering, matched filtering, wavelet decomposition, sparse decomposition, deconvolution, low-rank-sparse separation, or subspace projection methods can be used individually or in combination to further suppress continuous background components unrelated to the target event. The above-mentioned clutter removal or background suppression methods are all signal processing techniques known in the art, and those skilled in the art can select appropriate methods according to specific scenarios, noise structures, and computing power conditions without inventive effort.

[0111] In S103, if the target scene has a repeatable reference benchmark, reference parameters are obtained through the synchronization module 140, and the slow time series is resampled to a unified reference domain. S103 may include the following sub-steps: S103-1: Obtain reference parameters corresponding to the dynamic state of the target medium.

[0112] S103-2: Based on the reference parameters, resample the slow time series to a unified reference domain so that periodic or quasi-periodic events are aligned in the same reference domain.

[0113] S103-3: In the absence of an external synchronization signal, a self-synchronizing reference domain is constructed based on the periodicity of the CIR slow time series.

[0114] For example, in rotating machinery scenarios, the reference parameter can be the rotation angle θ; in cardiac mechanical event monitoring scenarios, the reference parameter can be the ECG R peak trigger time tr; and in structural excitation scenarios, the reference parameter can be the external excitation phase φl or the impact trigger time te. After alignment, periodic or quasi-periodic events can be stably repeated within a unified reference domain, thereby improving subsequent detection and coherent accumulation effects. In implementations without external synchronization signals, the periodicity in the CIR slow time series can also be directly utilized, and a self-synchronizing reference domain can be constructed through autocorrelation analysis, peak interval analysis, dominant frequency tracking, or template iterative alignment.

[0115] As an alternative theoretical explanation, in scenarios with repeating reference points and relatively stable observation geometry, the complete link consisting of the internal event source, medium propagation, surface dynamic disturbance, and UWB echo can be considered as a time-varying parametric system that varies with the reference parameter ξ, and its response can be summarized as follows: .in, This refers to the generalized excitation generated by internal transient mechanical events (such as stress wave equivalent excitation force). Indicates that the reference parameter The system transfer function changes, and * denotes convolution operation. When slow time series are mapped to a unified reference domain through synchronization and resampling, for fixed reference parameters... The local event window in the vicinity can be approximated as a local quasi-linear time-invariant subsystem. The observable output. This is the independent variable used in the embodiments of this application to describe fast time (distance gate), with units in seconds. CIR formula in a static background. The expression represents the waveform distribution at different distance gates; in the distortion derivation, it serves as the reference independent variable for the Taylor expansion. It represents the temporal details within the echo, and is on two independent axes with the slow time t, representing the time between different pulses. The finally extracted w(t) has been freed from distortion. The dimension retains only slow-time information related to acceleration. In other words, after fixing reference parameters such as rotation angle, heartbeat phase, or excitation phase, the extracted event waveform fragments... This can be understood as the observed values ​​of the impulse response or equivalent transient response of the local dynamic subsystem. It should be understood that the above-described local quasi-linear time-invariant interpretation is only an optional theoretical interpretation, used to assist in understanding the physical meaning of the repeatability and explicit characteristics of event waveforms under a unified reference domain, and does not constitute a necessary modeling step for the implementation of this application. Even without explicitly establishing the aforementioned local quasi-linear time-invariant model, those skilled in the art can still directly implement the technical solution of this application based on the acquisition, preprocessing, event detection, feature extraction, and model inference processes disclosed in the embodiments of this application.

[0116] In S104, the event detection module 150 performs event detection based on the reference domain energy statistics within the target range gate. S104 may include the following sub-steps: S104-1: Construct candidate event detection statistics.

[0117] S104-2: Based on threshold determination, matched filtering, anomaly detection, or triggering network, complete the coarse detection of candidate events.

[0118] S104-3: Perform fine screening on the coarse detection results to determine the target phase, target time, or target reference position corresponding to the candidate event.

[0119] For example, it can be defined as: .

[0120] in, In reference parameters Energy statistics at the location; The CIR amplitude after resampling in the reference domain; For fast time indexing; To unify the reference parameters in the reference domain.

[0121] Based on threshold Find significant peak values. An adaptive detection threshold; Energy statistics The mean; This is an empirical coefficient (which can take values ​​from 3 to 5). Energy statistics The standard deviation of the standard deviation. For more complex scenarios, matched filters, anomaly detectors, autoencoder reconstruction errors, sliding generalized likelihood ratio tests, or end-to-end triggering networks can also be used for detection. The event detection described should not be understood as limited to a single threshold determination. For repetitive periodic events, reference domain energy peak detection, threshold determination after coherent accumulation, or fixed phase stability determination are preferred; for scenarios where the approximate waveform of the target event is known, matched filtering or template correlation detection can be used; for scenarios with few prior samples and large variations in operating conditions, anomaly detection, autoencoder reconstruction errors, single-class classifiers, or sliding generalized likelihood ratio tests can be used; for scenarios with large amounts of data and where end-to-end deployment is desired, lightweight triggering networks or end-to-end candidate event detection networks can be used. Different detection methods can be used in series, parallel, or in layers. For example, coarse detection can be performed first using statistics, followed by fine screening using template correlation or anomaly detection.

[0122] In S105, the system can extract event waveform fragments, explicit physical features, or construct model input tensors from candidate events. S105 may include the following optional paths.

[0123] S105-1: Explicit feature path.

[0124] The channel impulse response waveform within a preset window is extracted with the event peak as the center to obtain the event waveform segment s[n]. Subsequently, features such as rise time, envelope peak value, kurtosis, main resonant frequency, bandwidth-to-energy ratio, attenuation coefficient, time-frequency ridge parameters, waveform sparsity, energy entropy, and repetition stability are extracted.

[0125] If combined with the aforementioned local quasi-linear time-invariant interpretation under a unified reference domain, the systemic physical meaning of various features can be further understood. For example, rise time can be used to characterize the initial establishment rate of the local response, corresponding to the steepness of impact application, changes in local stiffness, or differences in material hardness; the principal resonance frequency can correspond to the dominant mode or main resonance band of the local dynamic subsystem; the attenuation coefficient can be obtained by fitting the event envelope or time-frequency energy attenuation, and is used to quantify structural damping, energy dissipation, and changes in boundary conditions; the event start time and its fixed position in the reference domain can reflect causal consistency and phase positioning accuracy; repeatability stability can reflect the repeatability of the local dynamic response under the same reference parameter position. Thus, the explicit features are no longer just a set of empirical statistics, but system-level characterization quantities that can establish a correspondence with the local dynamic properties of the target medium.

[0126] The significance of the explicit features can be further explained as follows: Kurtosis can be obtained by the ratio of the fourth central moment to the square of the variance of the event window samples or by the normalized fourth-order statistic, and is used to characterize whether the event waveform has spikes, heavy tails, and strong impact concentration; the bandwidth-to-energy ratio can be obtained by the ratio of the main resonance band, the target frequency band, and the energy of the full frequency band or the sideband, and is used to characterize the degree of concentration of event energy in the target modal frequency band; the time-frequency ridge parameters can be extracted by short-time Fourier transform, wavelet transform, or other time-frequency analysis methods to obtain the center frequency trajectory, slope, and curvature of the main energy ridge. Curvature, duration, and ridge energy are used to characterize the evolution of the dominant mode over time and the energy transfer pattern; waveform sparsity can be obtained by the proportion of effective significant samples within the event window, the sparsity index composed of the l1 norm and l2 norm, or the proportion of samples above the threshold, and is used to characterize whether the event energy is concentrated in a few fast time sampling points or short time segments; energy entropy can be obtained by calculating Shannon entropy or other entropy-type indices on the normalized time-domain energy distribution, frequency-domain energy distribution, or time-frequency energy distribution, and is used to characterize the degree of order and complexity of the event energy distribution.

[0127] For stability-related parameters, their engineering meaning and calculation methods can be further explained. For example, event repetition stability can be obtained through the cross-correlation peak value, consistency distance, dynamic time warping distance, eigenvector variance, or embedding distance fluctuation between adjacent repetitive events, used to characterize the repetitive consistency of event waveforms at the same reference parameter position; angular position stability or reference domain position stability can be obtained through the reference parameter dispersion, standard deviation, or confidence interval width corresponding to the peak value of multi-cycle events, used to characterize whether a fault or short-term event is stably locked in the same mechanical phase, heartbeat phase, or excitation phase. As can be seen from the above explanation, the waveform distortion characteristic parameters not only have clear physical meanings but also have operable calculation methods. Therefore, they can be used for rule determination, statistical modeling, and lightweight machine learning, and can also serve as auxiliary supervision, physical consistency constraints, or result interpretation basis in end-to-end model training.

[0128] S105-2: Implicit learning path.

[0129] The distance-gate-slow-time tensor, reference domain tensor, or event slice tensor can be directly constructed and input into a deep neural network or other trained model to output transient mechanical event perception results. It should be noted that explicit and implicit features can also be fused to simultaneously ensure physical interpretability and the ability to represent complex patterns.

[0130] S105-3: Waveform extraction and event slicing path. The specific implementation steps are as follows: Step 1: Window Function Truncation; A fixed-length window segment is truncated centered on the detected peak position (e.g., target phase ξ_target or target time t_target). The choice of window length is related to the event duration and sampling rate: For rotating machinery scenarios, the angle window width is typically set to 5°–15° (depending on the rotational speed and impact duration). For example, at a rotational speed of 1800 rpm, a 30° angle window corresponds to an actual time of approximately 2.8 ms, which is sufficient to cover bearing impact events. For cardiac mechanical event monitoring, the cardiac window width is typically set to 200 ms–400 ms (covering the major mechanical events in a complete cardiac cycle), but the event waveform segment can be narrower (e.g., extracting only the 50 ms window corresponding to valve closure).

[0131] Step 2: Alignment and Relocation; When there is jitter in the detected peak position, cross-correlation alignment or local template registration can be performed first: Cross-correlation alignment: Using the first detected event waveform as a template, calculate the cross-correlation function between the subsequent event waveforms and the template, and take the delay corresponding to the cross-correlation peak as the fine-tuning offset to perform subsampling level correction on the slice position. Peak relocation: In the local neighborhood near the initial detected peak (e.g., ±3 sampling points or ±1° angle domain), search for the local maximum value as the final slice center.

[0132] Step 3: Multi-period aggregation; Aggregate multiple event waveforms that repeatedly occur at the same reference domain location (e.g., a fixed fault angle location or heartbeat phase) to improve the signal-to-noise ratio: Mean aggregation: Sum the corresponding sampling points of multiple event waveforms and divide by the number of events. Median aggregation: Take the median of the corresponding sampling points of multiple event waveforms, which is robust to outlier interference. Weighted average: Assign different weights to each event waveform based on its energy or correlation coefficient; waveforms with higher energy contribute more. Outlier removal: First calculate the average similarity between event waveforms, remove outliers whose consistency with the average waveform is lower than a preset threshold (e.g., cross-correlation peak < 0.7), and then perform aggregation.

[0133] Step 4: Data representation; The aggregated event waveforms can be directly used as one-dimensional feature vectors; Alternatively, event waveforms from multiple distance gates can be combined to form a two-dimensional matrix (distance gate × time sampling point), or the reference domain dimension can be further superimposed to form a three-dimensional tensor (distance gate × time sampling point × reference domain position), which can be used as input for subsequent deep learning models.

[0134] In waveform extraction and event slicing, one or more of the following approaches can be used. First, a fixed-length window is truncated centered on the detected target phase, target time, or target reference position. The window length is determined based on the event duration, sampling rate, and reference domain resolution. For rotating machinery scenarios, the angle window can be set to the order of several degrees to tens of degrees, such as 5° to 30°, with 5° to 15° being a preferred range. For cardiac mechanical event monitoring, the heartbeat window can be set to 200ms to 400ms, and a narrower window corresponding to local events such as valve closure or ejection initiation can be further truncated. Second, when there is jitter in the detected peak value, the slice center can be corrected through cross-correlation alignment, local template registration, or peak relocation. Third, mean aggregation, median aggregation, weighted averaging, robust averaging, or outlier removal are performed on multiple event waveforms that repeatedly appear at the same reference domain location to improve the signal-to-noise ratio of weak events. Fourth, the aggregated event waveform can be used as a one-dimensional feature vector, or the event waveforms of multiple distance gates can be combined into a distance gate × time sampling point matrix, or the reference domain dimension can be further superimposed to form a three-dimensional tensor for subsequent model input.

[0135] In S106, the transient mechanical event perception results output by the result generation module 180 can output corresponding content according to the actual scenario. For example, for engineering machinery scenarios, it can include: whether there is an abnormal event, the frequency of abnormality, the type of fault, the severity of the fault, the level of structural damage, and the alarm confidence level. For cardiac health monitoring scenarios, it can include health index, remaining lifespan, heart rate, cardiac cycle stability, valve timing parameters, the level of structural damage, and the alarm confidence level.

[0136] Therefore, the embodiments of this application realize a complete technical solution for transient mechanical events, short-term dynamic disturbances of the target surface, echo delay perturbation-CIR fast-time local distortion, and event perception results.

[0137] Please refer to Figure 2 , Figure 2 This diagram illustrates the physical relationship between transient mechanical events, short-term dynamic disturbances on the target surface, and waveform distortion in the channel impulse response, as provided in an embodiment of this application. The diagram shows the following causal chain: sudden changes in local mechanical state, short-term mechanical drives, or physiological mechanical activities first trigger transient mechanical events; these events propagate within the medium or couple to the observed surface, forming transient normal acceleration / short-term displacement disturbances on the target surface; these short-term mechanical disturbances cause rapid perturbations in the echo delay; these rapid perturbations manifest as fast-time waveform distortion in the channel impulse response; and the results of transient mechanical event perception can be obtained by extracting and analyzing this distortion.

[0138] The aforementioned causal chain constitutes the physical mainline of the embodiments of this application, used to illustrate the correspondence between transient mechanical events, short-term dynamic disturbances on the target surface, and CIR fast-time waveform distortion.

[0139] Please refer to Figure 3 , Figure 3 This is a structural block diagram of a transient mechanical event sensing system provided in an embodiment of this application. The system includes: an ultra-wideband radar module 110, a data acquisition module 120, a preprocessing module 130, a synchronization module 140, an event detection module 150, a feature extraction module 160, a model inference module 170, a result generation module 180, and an optional physical information data enhancement module 190.

[0140] Specifically, the ultra-wideband radar module 110 is responsible for transmitting pulses and receiving echoes; the data acquisition module 120 is responsible for acquiring channel impulse response sequences; the preprocessing module 130 is responsible for clutter or background suppression, range gate selection, filtering enhancement, and normalization; the synchronization module 140 is responsible for acquiring reference parameters and aligning with the reference domain; the event detection module 150 is responsible for identifying candidate events and forming a unified entry point for subsequent explicit and implicit paths; the feature extraction module 160 is responsible for performing explicit physical feature extraction on candidate events; the model inference module 170 is responsible for receiving event tensors, explicit features, or physical information enhancement samples, and performing machine learning or deep learning inference; the result generation module 180 is responsible for outputting transient mechanical event perception results; and the physical information data enhancement module 190 is responsible for constructing physical information simulation samples and providing the model inference module 170 with the sample support required for training, optimization, or verification.

[0141] It should be noted that in the above complete description of the transient mechanical event sensing system, some modules are not required, such as the preprocessing module 130.

[0142] exist Figure 3 In the structure shown, the main processing chain can be represented as an ultra-wideband radar module 110, a data acquisition module 120, a preprocessing module 130, a synchronization module 140, and an event detection module 150. Based on this, the event detection module 150 branches into two types of subsequent processing paths: one is an explicit path where explicit physical features are formed by the feature extraction module 160 and then input into the model inference module 170; the other is an implicit path where candidate event tensors are directly input into the model inference module 170. Reference parameter sources such as encoders, ECG, and load sensors are preferably connected to the synchronization module 140 to provide a unified reference domain; auxiliary verification modes such as vibration, temperature, and slow-time displacement, phase, envelope, or micro-Doppler auxiliary features extracted from the same ultra-wideband radar data are preferably connected to the result generation module 180 for multi-modal fusion, cross-validation, or result verification. Therefore, Figure 3The arrows in the diagram correspond to the main data stream, reference parameter stream, training sample stream, and auxiliary verification stream, respectively, but do not imply that all modules need to communicate bidirectionally at the same time.

[0143] In certain industrial deployment scenarios, the system can also work in conjunction with contact vibration sensors, temperature sensors, encoders, cameras, or edge computing gateways. Furthermore, it can extract slow-time displacement, phase, envelope, or micro-Doppler auxiliary features based on the same ultra-wideband radar data, thus forming a multimodal, highly reliable monitoring platform. For example, in wind turbine main shaft scenarios, vibration characteristics can be used to cross-validate CIR events; in hospital ward scenarios, ECG or PPG can provide cardiac synchronization references; and in bridge or blade scenarios, load sensors can provide external excitation references.

[0144] Among them, micro-Doppler or continuous vibration monitoring is more suitable as an auxiliary channel triggered by the sensing results of the embodiments of this application, for observing slow-changing trends, continuous states or performing verification, rather than replacing the main sensing chain based on fast-time waveform distortion of the embodiments of this application.

[0145] Please refer to Figure 4 , Figure 4 This application provides a flowchart for diagnosing faults in rotating machinery. The flowchart is applicable to equipment such as bearings, gears, couplings, spindles, pump sets, and compressors.

[0146] The specific steps are as follows: First, the ultra-wideband radar is aimed at the bearing housing, casing, or structural observation area to collect the channel impulse response sequence and speed reference signal; second, the angle phase is constructed based on the speed / encoder reference signal; next, angular domain resampling is performed to map the channel impulse response sequence CIR to a unified angle domain; then, angular domain energy statistics are performed to calculate the angular domain energy spectrum within the target range gate and lock the fault impulse phase; subsequently, event slicing is performed to extract the event waveform within the angle window near the impulse phase and calculate explicit features such as rise time, kurtosis, main resonant frequency, attenuation coefficient, and angle stability; finally, model judgment is performed, and the fault type, severity, fault angle location, and health index are output based on rule diagnosis, SVM, and random forest lightweight neural network.

[0147] In optional implementations of rotating machinery, the ultra-wideband radar can be deployed on the outside of the bearing housing, near the visible window of the housing, in the observation area of ​​the main shaft, or near structural surfaces with a high degree of mechanical coupling to the fault propagation path, to avoid altering the equipment boundary conditions due to contact installation. Preferably, the target range gate with the most prominent event energy is first searched among multiple adjacent range gates, and then an observation gate group containing one or more adjacent range gates is formed around the target range gate to balance spatial focusing capability and installation error tolerance.

[0148] In constructing the reference domain, an angle phase sequence can be established using speed encoder pulses, Hall phase signals, key phase signals, or shaft angle references estimated based on other sensors. In the absence of direct encoder signals, an equivalent angle domain reference can also be constructed based on master frequency estimation, frequency tracking, or other external synchronization information. Optionally, coherent accumulation or robust aggregation can be performed over multiple consecutive rotational cycles to suppress sporadic shocks and random background disturbances.

[0149] In event window extraction, a corner window can be set around the fault impact phase. The width of the corner window can be set according to the equipment rotation speed, impact duration, and sampling density, for example, a local window on the order of several to tens of units. For high-speed equipment, the corner window can be narrowed to reduce the overlap of adjacent events. For low-speed or weak-event equipment, the window width can be appropriately increased and combined with multi-cycle aggregation to improve the signal-to-noise ratio. Subsequently, features such as rise time, kurtosis, main resonant frequency, bandwidth-to-energy ratio, attenuation coefficient, repetitive stability, and angular position stability can be extracted and further fused and judged in combination with auxiliary information such as vibration, temperature, and load.

[0150] In terms of output specifications, the results for rotating machinery scenarios can include not only the presence of anomalies, fault types, and fault severity, but also fault location, health index, remaining life trend, alarm confidence level, and whether contact-based verification monitoring needs to be initiated. For high-value equipment such as wind turbine main shafts, gearboxes, pump sets, compressors, and rail transit axle boxes, the above outputs can be directly used for maintenance planning, downtime scheduling, and spare parts strategy optimization, thereby forming a clear engineering closed loop.

[0151] In this scenario, for offshore wind turbine main shafts, rail transit axle boxes, and high-pressure sealed pumps, the installation, wiring, maintenance, and life management of contact sensors will significantly increase the total life cycle cost. However, the non-contact transient mechanical event sensing provided in this application embodiment can achieve early fault warning and location tag output without changing the equipment structure.

[0152] Please refer to Figure 5 , Figure 5 This application provides a flowchart for transient mechanical event perception and processing based on an end-to-end model. This process is applicable to scenarios where a unified model is used to perform multi-scene classification, regression, or anomaly detection.

[0153] In this implementation, the input tensor can be a distance gate-slow time matrix, a reference domain matrix, or an event slice tensor; the feature encoding layer can be a convolutional layer, a spatiotemporal convolutional layer, or a Transformer encoding layer; the temporal modeling layer can be a recurrent neural network (RNN), an attention layer, or a lightweight graph structure; the fusion layer can fuse explicit features, fuse implicit features, or fuse explicit features with implicit features. Fusion methods include at least one of feature concatenation, weighted summation, attention fusion, or gated fusion. For example, explicit feature vectors such as rise time, kurtosis, principal resonant frequency, and decay coefficient can be concatenated with implicit feature vectors output by a convolutional network or Transformer encoding layer and then fed into a fully connected layer or a classification / regression head. The output layer outputs classification probabilities, continuous parameters, or health scores based on the task type. If physical consistency training is used, the task output generated by the output layer is used to constitute the task loss Ltask, and the physical consistency prior is used to constitute the physical constraint term Lphys. The two are combined in the loss function to form L = Ltask + λLphys, where λ is a balancing hyperparameter. The physical constraints Lphys can include at least one of the following: constraining the model to focus on fast-time waveform distortion and suppress the response in the background stable region within the event window; constraining adjacent repeating events to maintain consistent feature representations within a unified reference domain; and constraining the model output to maintain consistency with explicit physical features such as rise time, kurtosis, principal resonant frequency, decay coefficient, and event repetition stability. Through these constraints, the model no longer relies solely on label co-occurrence relationships for classification or regression, but is guided by the mechanism of transient mechanical events during the training phase. After training is complete, the trained model is obtained and deployed to edge nodes or the cloud.

[0154] In another embodiment of this application, the physical consistency constraint loss L_phys may take, but is not limited to, one of the following forms: (1) Add the known physical equation (such as the linear relationship between the degradation rate and the characteristic change rate) as a residual term to the loss function so that the model output satisfies the least squares constraint of the equation; (2) Use partial differential equation constraints (such as PDE loss in PINN) to apply physical laws such as continuity and dissipation to the spatiotemporal field predicted by the model; (3) Use positive and negative sample pairs in contrastive learning to constrain, for example: the feature representations of normal samples in the embedding space should be clustered together, while the feature representations of faulty samples should be separated from normal samples, thereby enhancing the ability of features to distinguish physical states.

[0155] The physical constraints mentioned above can be used individually or in combination, and are weighted with the supervision loss L_task to form the final loss function L=L_task+λL_phys.

[0156] In one exemplary implementation, the input is a 64×64 range-gate slow-time matrix, or a 64×64×C input tensor formed by superimposing the target range gate group, the reference domain alignment result, and the event slice, where C can be 1, 2, or 3. Optionally, clutter or background suppression, target range gate locking, normalization, and reference domain alignment are first performed on the original CIR data, and then the event window is truncated into a fixed-size tensor input model. The main body of the model can adopt two convolutional layers, each containing 32 3×3 convolutional kernels and using the ReLU activation function; then, after pooling layers and flattening operations, it is connected to 128-dimensional and 64-dimensional fully connected layers, and finally outputs the abnormal / normal binary classification probability, fault category probability, continuous health score, or other event parameters. It should be understood that the above network structure, number of convolutional kernels, activation function, and optimizer are only examples. Those skilled in the art can adjust the number of layers, number of channels, kernel size, and loss function form according to different scenarios, and the embodiments of this application do not specifically limit this.

[0157] Furthermore, in terms of training data organization, the samples can be divided into training, validation, and test sets, for example, at a ratio of 70%, 15%, and 15% respectively; labels can come from manual annotation, reference sensor triggering, known operating condition labels, or historical maintenance results. Preferably, mini-batch training is performed with a batch size of 16 to 64, and early stopping is performed when the validation set loss no longer decreases for several consecutive rounds. During training, cross-entropy loss or mean squared error loss can be used, combined with the Adam optimizer, learning rate decay strategy, and physical consistency constraints. The physical consistency constraint Lphys can include at least one of the following: constraining the model to focus on fast temporal distortions within the event window while suppressing over-response to the background stable region; constraining the representation of adjacent repeated events to remain consistent in the reference domain; and constraining the model output to remain consistent with explicit physical features or approximate physical relationships.

[0158] During the inference and deployment phases, the model output can be represented as the event probability p_event, the class vector y_class, the health score s_health, or continuous parameter estimates. Preferably, when the event probability is higher than a preset threshold, an anomaly alarm is output, and secondary verification is performed in conjunction with explicit features, reference parameters, and auxiliary modalities; when the event probability is lower than the preset threshold but weak anomalies accumulate in multiple consecutive windows, a trend warning can also be output. Through the combination of the above input construction, training organization, loss design, verification strategy, and deployment approach, the end-to-end path of this application is no longer an abstract "black box model" description, but an implementable technical solution with clear inputs, clear training constraints, clear output forms, and clear engineering implementation methods.

[0159] This implementation method enables implicit learning using the waveform distortion information while maintaining a clear input format, training constraints, and output caliber. The corresponding patent protection is not limited to a specific network structure but covers all methods for implicit learning using the waveform distortion information and outputting transient mechanical event perception results.

[0160] In other words, even if the model directly takes the original CIR sequence or matrix as input and implicitly learns the distortion patterns related to transient mechanical events in the intermediate layer without explicitly outputting physical characteristics such as rise time, decay coefficient or resonant frequency, as long as it essentially utilizes the waveform distortion information to complete the perception, it still falls within the protection scope of this application.

[0161] Please refer to Figure 6 , Figure 6 This document provides a flowchart for physical information data augmentation and model training in an embodiment of this application. In this embodiment, transient mechanical events and short-term dynamic disturbances on the target surface are first generated based on the impact source intensity, event duration, medium damping, propagation path, boundary conditions, and radar pulse parameters. Then, corresponding channel impact response CIR distortion samples are constructed based on the echo model. These CIR distortion samples are written into a simulation sample library to achieve multi-condition coverage. Finally, real samples are obtained from a real sample library or through on-site collection / annotation. The simulation samples and real samples are used together to train the transient mechanical event perception model, achieving augmentation with fewer samples and generalization across operating conditions, ultimately outputting the final model.

[0162] This implementation method is suitable for scenarios where fault samples are scarce, rare pathological signals are difficult to collect, and structural damage events are difficult to replicate. For example, in the layered monitoring of aerospace composite materials, it is difficult to obtain a large number of real damage samples; in medical scenarios, high-quality data on mechanical events in abnormal cases are naturally scarce. Physical information data augmentation can significantly expand the coverage of training samples without compromising physical plausibility.

[0163] Please refer to Figure 7 , Figure 7 This document provides a flowchart for monitoring mechanical events in the human heart, as illustrated in an embodiment of this application. In this embodiment, an ultra-wideband radar is deployed facing the chest wall or back of the human body and acquires the corresponding ultra-wideband radar-channel impulse response sequence (UWB-CIR sequence). ECG or other external trigger signals are acquired to provide a cardiac rhythm synchronization reference. Waveform distortion corresponding to cardiac mechanical events is extracted within a unified cardiac rhythm reference domain. Further, output parameters such as valve heart rate, cardiac cycle stability, and systolic / diastolic intensity are extracted, along with event timing such as valve opening and closing times and cardiac mechanical waveforms. Finally, cardiac mechanical monitoring results or abnormality identification results are output.

[0164] In one implementation, self-synchronization can be achieved directly from the CIR slow time series, utilizing the periodicity of the heartbeat, without relying on ECG synchronization. This can be done, for example, by extracting the cardiac cycle and aligning the reference domain through autocorrelation analysis, peak interval analysis, dominant frequency tracking, or template registration. Therefore, the cardiac mechanical event monitoring can be implemented using either an external synchronization signal or the periodicity of the CIR data itself.

[0165] It should be noted that this implementation method can cover scenarios such as bedside monitoring in hospitals, contactless monitoring in elderly care institutions, home sleep monitoring, detection of live subjects left in vehicles, and remote monitoring of mechanical vital signs of high-risk patients.

[0166] Please refer to Figure 8 , Figure 8 This document provides a flowchart for structural health monitoring, applicable to structural components such as wind turbine blades, bridge members, composite panels, pressure vessels, and aerospace equipment. The system generates short-term structural events under external excitation or operational loads, acquires channel impulse response sequences within the distance-gate locked observation area, and extracts waveforms related to transient mechanical events by combining load phase or trigger time. By analyzing time delay drift, bandwidth energy migration, attenuation coefficient changes, and event repeatability, cracks, delamination, detachment, micro-damage, and loosening conditions can be identified. The system outputs damage levels, alarm confidence levels, and maintenance recommendations.

[0167] For example, in composite wing panels or wing box sections, UWB radar can be fixed at a non-contact observation position on the outer surface of the structure, near the panel inspection port, or inside the box section. Under flight loads, ground vibration excitation, or artificial excitation, CIR data can be continuously acquired. Subsequently, the changes in CIR waveform distortion under different load phases or excitation times can be analyzed to extract damage-sensitive features such as stress wave attenuation coefficient, reflection delay distribution, and repeatability stability. These features can then be compared with a healthy baseline to identify the delamination location, loosening area, or trend of damage area changes.

[0168] Please refer to Figure 9 , Figure 9This diagram illustrates a unified reference domain and output framework for multiple scenarios, as provided in this application embodiment. It unifies three representative scenarios—rotating machinery, cardiac mechanical event monitoring, and structural health monitoring—under a single reference domain processing framework: the input side corresponds to rotation angle, cardiac synchronization moment, and load / excitation moment, respectively; the intermediate processing layer shares distance gate selection, background suppression, event detection, waveform extraction, and model inference; the output side corresponds to fault diagnosis (including fault type, fault angle location, health index, RUL, etc.), cardiac mechanical parameter evaluation (including heart rate, mechanical timing, systolic and diastolic function evaluation, etc.), and structural damage identification (including cracks, delamination, loosening, damage level, etc.). The diagram clearly demonstrates that this application is not limited to a single industrial fault diagnosis tool, but rather represents a fundamental platform capability focused on the direct perception of short-term dynamic disturbances.

[0169] Please refer to Figure 12 , Figure 12 This diagram illustrates a local quasi-linear time-invariant interpretation under a unified reference domain, as provided in this application embodiment. The diagram serves to explain that, on the original time axis, the internal event source-medium propagation-surface dynamic disturbance-UWB echo can be considered a parametric time-varying system that varies with the reference parameter ξ; when the data is mapped to a unified reference domain through alignment and resampling of the rotation angle, heartbeat phase, or excitation phase, for a fixed reference parameter... The nearby local window, the third layer in the diagram, can be physically broken down into three interconnected elements. The first is a fixed reference parameter. This represents the reference domain position fixed after synchronization and resampling, such as the fault angle position of rotating machinery, a specific percentage of the cardiac phase in a cardiac mechanical event, or the load phase point of structural excitation; near this reference parameter, the time-varying characteristics of the system are locally frozen. The second is the locally quasi-linear time-invariant subsystem H( ,τ), indicating that in Within a limited observation window, the system, which originally depended on changes in the reference parameter, can be approximated by a locally linear time-invariant subsystem. Its response integrates the mechanical propagation path, structural resonance characteristics, and UWB radar electromagnetic scattering and distortion at that local operating point. Thirdly, there is the observable transient response s(τ), representing the response around the fixed reference parameter. The extracted event waveform segments correspond to transient mechanical events that repeat under the same reference parameter conditions. It should be noted that... Figure 12 The explanations shown are only for the purpose of helping to understand the physical meaning of the repeatability and explicit characteristics of event waveforms under a unified reference domain, and do not constitute a restrictive step that must be implemented only after explicitly solving the local linear time-invariant model in the embodiments of this application.

[0170] It is important to emphasize that the method of this application does not rely on explicit system identification of the observed links or the establishment of the quasi-linear time-invariant model. Even without this physical abstraction, repetitive events can be aligned through a pure signal processing operation of reference domain resampling, and its improvement on the signal-to-noise ratio and feature stability of event detection is objective and measurable. The theoretical explanation provided is merely a possible physical insight into the superior technical effect discovered in this application.

[0171] Through the above decomposition, we can intuitively understand why a unified reference domain can enhance event repeatability, why explicit physical characteristics (such as rise time and principal resonance frequency) have systemic property meaning, and why event waveforms at the same reference parameter positions can be used for stable alignment. It should be noted that... Figure 12 The explanations provided are for illustrative purposes only and do not constitute a restrictive step that requires explicit solution of the locally linear time-invariant model before implementation in this application. The extracted event waveforms can be understood as the observable transient responses of the locally quasi-linear time-invariant subsystem. It should be noted that... Figure 12 The explanations provided are only intended to illustrate why a unified reference domain helps enhance event repeatability, why explicit physical features have systemic property implications, and why event waveforms at the same reference parameter location can be used for stable comparisons. They do not constitute a restrictive step that this application must take before explicitly solving a local linear time-invariant model.

[0172] Please refer to Figure 10 , Figure 10This is a structural block diagram of an electronic device provided in an embodiment of this application. The electronic device includes an ultra-wideband radar front-end, at least one processor, and a memory communicatively connected to the at least one processor. The ultra-wideband radar front-end is used for transmitting, receiving, and synchronizing clocks. The program instructions stored in the memory enable the processor to perform steps such as clutter removal or background suppression, reference domain alignment, event detection, waveform extraction, explicit feature calculation, model inference, and result output. The memory may also store model parameters, cached data, etc. The processor may be a DSP, CPU, or AI acceleration unit. Optionally, the electronic device may also include a communication interface, a local storage / historical event database, and an alarm and data reporting interface. The communication interface is used for data exchange with external systems (e.g., host computers, industrial gateways, cloud platforms), and may employ communication protocols such as Ethernet, Wi-Fi, Bluetooth, 4G / 5G, or industrial buses (e.g., Modbus, CAN). The local storage / historical event database is used to store model parameters, cache CIR data, record event waveform segments, and historical sensing results. When the communication link is interrupted, the device can still locally save the data and retransmit it after recovery. Alarm and data reporting interface: Transient mechanical event perception results (such as fault type, health index, and abnormal alarms) can be sent to a gateway, host computer, or cloud platform via a communication interface. The gateway is responsible for protocol conversion and on-site data aggregation, the host computer is used for local monitoring and display, and the cloud platform is used for remote data analysis and operation and maintenance management.

[0173] In one alternative implementation, the ultra-wideband radar module can be a radar chip, radar module, or evaluation board conforming to IEEE 802.15.4z, IEEE 802.15.4ab, or an equivalent UWB pulse system. The module can output complex I / Q form CIR samples, amplitude CIR samples, phase CIR samples, power delay spectrum, range gate sequence, relative tap index, or data representations derived therefrom to the processor. It can also simultaneously output at least one of the following: CIR window start position, first path index, main peak index, timestamp, frame number, channel identifier, transmit / receive antenna identifier, received power, noise estimation, channel quality index, or synchronization status identifier. The processor can perform target range gate selection, background suppression, reference domain alignment, event window slicing, explicit feature extraction, and model inference based on the above fields, thereby enabling the embodiments of this application to be deployed as a chip SDK, SoC, edge computing box, radar module, or cloud-edge collaborative system. It should be noted that the above fields are one alternative implementation for adapting the chip / module interface and do not imply that all fields must be present simultaneously to implement this application.

[0174] The electronic device includes an ultra-wideband radar front-end, at least one processor, and a memory communicatively connected to the at least one processor. The ultra-wideband radar front-end is used to transmit pulses, receive echoes, and output channel impulse response data; the processor may be a CPU, DSP, FPGA, GPU, NPU, AI acceleration unit, or a combination thereof; the memory is used to store program instructions, model parameters, cached CIR data, event waveform segments, and historical sensing results. When the program instructions are executed by the processor, they can achieve at least one of the following: clutter removal or background suppression, target range gate selection, reference domain alignment, event detection, event slicing, explicit feature calculation, model inference, result output, alarm reporting, or auxiliary monitoring triggering.

[0175] All of the above components are optional configurations and can be selected or discarded according to specific application scenarios (such as edge computing nodes and cloud collaboration), without affecting the implementation of the core diagnostic link of this application.

[0176] The following will provide further explanation with specific examples.

[0177] Example 1: Early fault monitoring of wind turbine main shaft bearings.

[0178] In offshore wind farms, the main shaft bearing is exposed to high humidity, high salt spray, and very limited maintenance windows. Traditional contact sensors are complex to install and have limited lifespan. In this embodiment, an ultra-wideband radar is deployed on the inner wall of the nacelle, aligned with the main shaft bearing housing area; a rotational speed reference is acquired synchronously; the channel impulse response sequence is mapped to the angular domain; events at fixed angular positions are detected; features such as rise time, resonant frequency, and angular stability are extracted; and a fault risk level is output by an edge model. When a significant enhancement of event characteristics and a stable angular position are detected for several consecutive days, the system triggers an early maintenance warning.

[0179] In this embodiment, if reference domain alignment and robust aggregation are performed on the event waveforms at the same fault angular location across multiple consecutive rotation cycles, the resulting event waveforms can be further understood as the equivalent transient response of the bearing system at that local angular location. Based on this interpretation, explicit characteristics can be decoded with greater physical meaning: the rise time reflects the speed at which the local response establishes itself, corresponding to the steepness of the impact application, the degree of abrupt change in contact stiffness, or changes in local material hardness; the dominant resonant frequency reflects the dominant mode of the bearing housing, casing, and their coupled structures along that local path; the attenuation coefficient reflects the energy dissipation, structural damping, and boundary connection state of the stress wave along its propagation path; and the angular position stability reflects whether the fault event is stably locked in the same mechanical phase. Therefore, when the system observes a continuous shortening of the rise time, a stable shift in the dominant resonant frequency or frequency band energy distribution, a decrease in the attenuation coefficient, and an increase in angular position stability, this can be used as important evidence of enhanced local impact sources and fault evolution, rather than merely as a general statistical anomaly.

[0180] In one type of engineering output method, structured results such as feature names, system meanings, and fault explanations can be further formed: for example, the rise time can be mapped to the local response establishment speed and contact steepness, the main resonance frequency can be mapped to the dominant structural mode, the attenuation coefficient can be mapped to the damping and boundary dissipation level, and the angular position stability can be mapped to the fault phase lock-in degree. Based on this, diagnostic explanations that are easier for maintenance personnel to understand can be output, so that the local quasi-linear time-invariant explanation not only stays at the theoretical level, but also directly serves the on-site operation and maintenance decision-making.

[0181] Example 2: Fault monitoring of insulation terminals in industrial high-speed motors.

[0182] At the end of a high-voltage motor, contact installation introduces insulation and safety issues. In this embodiment, an ultra-wideband radar is deployed in a safe, isolated area, illuminating the target end non-contactly via air coupling. The system monitors transient mechanical events and structural response changes, outputting anomaly warnings and location tags. This scenario demonstrates the engineering applicability of this application compared to traditional contact-based sensing solutions in terms of non-contact deployment and spatial gating.

[0183] Example 3: Bedside non-contact cardiac mechanical event monitoring.

[0184] In ward monitoring scenarios, the system monitors the channel impulse response sequence in the patient's chest direction using ultra-wideband radar and performs reference domain alignment based on the ECG R peak trigger time. The system extracts waveform distortions corresponding to cardiac mechanical events and outputs heart rate, cardiac cycle stability, and valve timing-related parameters. Compared to conventional vital sign radar schemes that primarily observe slow-varying chest wall displacement, respiratory fluctuations, or rhythm parameters, this application embodiment focuses more directly on the fast-time local distortions of CIR corresponding to short-term mechanical events such as valve opening and closing, rapid myocardial contraction / diastole, and ejection initiation. Therefore, it has better adaptability in event-level mechanical characterization, timing parameter extraction, and repetitive heartbeat aggregation analysis. It should be understood that this comparison is used to illustrate the technical focus of this application in short-term cardiac mechanical event monitoring and does not exclude the application value of other radar schemes in respiratory monitoring, heart rate estimation, presence detection, or specific single indicators. In another preferred embodiment, ECG may not be connected; instead, the cardiac cycle can be obtained directly from the slow CIR time series through autocorrelation peaks or cycle tracking, thereby achieving pure radar self-synchronization monitoring.

[0185] Example 4: Impact damage monitoring of composite material plates.

[0186] In the inspection of composite material structures, slight delamination or microcracks can cause changes in the propagation characteristics of internal stress waves. The system acquires channel impact response data under controlled excitation conditions, analyzes changes in attenuation coefficients, resonant frequencies, and reflection delay distributions before and after damage, and then identifies the locations of delamination and detachment. For aerospace panels, wing box sections, or satellite structural panels, radar can also be deployed on the outer surface or internal inspection positions to acquire CIR data in stages under flight load simulation, frequency sweep excitation, or impact hammer excitation, and compare it with a healthy baseline to identify damage propagation trends. This embodiment can serve as a verification example in aerospace structural health monitoring scenarios.

[0187] Example 5: End-to-end anomaly detection based on physical consistency.

[0188] In this embodiment, the input is a reference-domain aligned distance-gate slow-time matrix, the model structure uses a combination of a spatiotemporal convolutional network and an attention layer, and the output is the anomaly probability. A physical consistency constraint term is added to the loss function to suppress the model's dependence on background patterns unrelated to the event mechanism. Further, in a specific example, a lightweight CNN model can be constructed using two convolutional layers, two fully connected layers, and a binary classification output layer, and trained using the Adam optimizer, a learning rate of 0.001, and cross-entropy loss. Compared to a purely data-driven model, this embodiment exhibits more stable performance under conditions of few samples, across different operating conditions, and across different installation locations.

[0189] In a more concrete engineering implementation, 64 consecutive reference domain sampling points and their corresponding target range gate groups can be extracted from the CIR sequence after target range gate selection and reference domain alignment to construct a 64×64 input matrix. Then, a training sample set is constructed using manually labeled results, encoder phase windows, contact sensor triggering results, or historical fault records as supervisory labels. The samples are preferably divided into training, validation, and test sets, and during the validation phase, anomaly detection rate, false alarm rate, confidence stability, and cross-condition generalization performance are recorded. If the validation set performance no longer improves in several consecutive training rounds, early stopping is performed and the optimal weights are frozen.

[0190] In designing physical consistency constraints, the ratio of fast-time distortion energy within the event window to stable energy within the background window, the feature consistency of adjacent repetitive events, and the correlation between the output and explicit physical features can be incorporated as components of the constraint terms. For example, penalties can be imposed on the responses of intermediate layer activations within the background stable distance gate, encouragement can be given to significant distortion responses within the event window, and consistency constraints can be imposed on the embedding distances of repetitive events under different reference periods. Thus, the model no longer relies solely on statistical co-occurrence relationships for classification, but is explicitly guided by the transient mechanical event mechanism during the training phase.

[0191] During the deployment phase, the model output can be not only a single anomaly probability, but also a category vector, continuous health scores, and alarm confidence scores simultaneously. For edge deployment scenarios, the model can be compressed into a lightweight convolutional network or a distillation model, and the most recent event slices and intermediate scores can be cached locally. For cloud deployment scenarios, a high-capacity model can be retained and continuously updated in conjunction with the historical sample library. The above implementation methods further demonstrate that the end-to-end path covered by this application has clear inputs, clear training, clear verification, and clear deployment forms, which can meet the dual requirements of sufficient patent disclosure and engineering feasibility.

[0192] Example 6: Comparative verification of hardware-in-the-loop simulation based on actual system parameters conforming to the 802.15.4z / 4ab standard.

[0193] Please refer to Figure 11 , Figure 11 This application provides a semi-quantitative comparison chart of UWB-CIR and continuous wave micro-Doppler routes under hardware-in-the-loop simulation conditions based on actual system parameters conforming to the 802.15.4z / 4ab standard, for embodiments of this application.

[0194] In this embodiment, to improve the realism of the technical effect, purely theoretical assumptions are no longer used. Instead, a semi-physical simulation link consistent with the actual platform is constructed. Specifically, two types of non-contact detection links are constructed for the same target under the same observation geometry: one is the UWB-CIR fast-time waveform distortion sensing link described in this application, and the other is a micro-Doppler sensing link based on continuous wave phase demodulation. Preferably, the average observation distance of the target is set to 0.8m, and the target surface experiences a short-term displacement disturbance on the order of 0.5μm under the action of a transient mechanical event. The main duration of the event is on the order of microseconds, and low-frequency environmental vibration and random noise are superimposed. For the UWB-CIR route, UWB radar parameters conforming to the 802.15.4z / 4ab standard are used, with the center frequency selected as ch11 and an equivalent operating bandwidth of approximately 1.3 GHz. The simulation incorporates transmit pulse shaping, receiver front-end bandwidth limiting, matched filtering, range gating, background suppression, reference domain alignment, and event detection procedures to ensure consistency between the verification link and actual engineering implementation. For the continuous wave micro-Doppler route, a continuous phase demodulation link can be used as a comparison baseline. The aforementioned 0.8 m observation distance, 0.5 μm-level short-term displacement disturbance, microsecond-level event duration, and approximately 1.3 GHz equivalent operating bandwidth correspond to a set of exemplary settings under laboratory close-range observation, early weak event magnitude, and actual UWB system parameter constraints, respectively. The parameter magnitudes are matched to ensure that the hardware-in-the-loop simulation results have engineering reference value, rather than constraining actual deployment conditions. To avoid overly abstract evaluation criteria, the following semi-quantitative evaluation criteria are further constructed in this hardware-in-the-loop simulation: normalized event distinguishability Ssep, stable detection rate Pdet, and pseudo-response suppression capability Rsup.

[0195] Wherein, the normalized event distinguishability Ssep is used to characterize the degree of distinction between the target event response and background disturbance; the stable detection rate Pdet is used to characterize the proportion of times a preset detection threshold is reached in repeated simulation trials; and the pseudo-response suppression capability Rsup is used to characterize the system's ability to suppress false triggering caused by non-event background fluctuations and low-frequency disturbances. In a set of hardware-in-the-loop simulation results based on the above-mentioned actual system parameter configurations, the Ssep of the UWB-CIR route can reach approximately 0.83, while that of the continuous wave micro-Doppler route is approximately 0.27; the Pdet of the UWB-CIR route can reach approximately 92%, while that of the continuous wave micro-Doppler route is approximately 39%; and the Rsup of the UWB-CIR route can reach approximately 90%, while that of the continuous wave micro-Doppler route is approximately 54%. It should be understood that the above values ​​are exemplary results for this set of hardware-in-the-loop simulation parameters, used to illustrate the relative trends of the two types of technical routes under the same short-term weak event conditions, and not to numerically limit the scope of protection of this application.

[0196] Combination Figure 11It can be seen that, under the same short-duration weak event amplitude, the same observation geometry, and similar noise background, the UWB-CIR route not only presents a clearer event distortion waveform in the fast time dimension, but also exhibits higher event resolution, higher stable detection rate, and stronger spurious response suppression capability in semi-quantitative indicators. In contrast, the continuous wave micro-Doppler route, because it still relies on continuous phase trajectories, often struggles to provide stable and reliable outputs when faced with transient, non-stationary, and extremely short-duration events. This comparison demonstrates that the technological advancement of this application lies not in simply replacing radar hardware, but in utilizing and interpreting the fast-time waveform distortion of CIR, a signal dimension distinct from slow-time phase / amplitude changes. Therefore, under semi-physical simulation conditions based on actual system parameter constraints, the UWB-CIR route outperforms the continuous wave micro-Doppler route in terms of resolution, stable detection rate, and spurious response suppression capability.

[0197] The aforementioned hardware-in-the-loop simulation charts based on actual system parameters have clear value: First, they can serve as intermediate supporting material between the physical main line and the technical effects of this application; second, since the center frequency, bandwidth, and signal processing chain used are all constrained by the actual UWB system parameters, they are closer to the subsequent platform implementation conditions than purely theoretical construction results; third, they can provide a unified semi-quantitative expression for subsequent embodiments and responses; fourth, before a complete set of measured data tables is formed, the aforementioned hardware-in-the-loop simulation charts can serve as a preliminary explanation of the rationality of the principle and the trend of the technical effects, but they should still form a complete supporting system together with the physical derivation, implementation process, and subsequent measured verification in the specification.

[0198] It should be noted that, under the conditions of this set of semi-physical simulations, the UWB-CIR fast-time waveform distortion route shows a trend of superiority over the continuous-wave micro-Doppler route in terms of normalized event resolution, stable detection rate, and spurious response suppression capability. This comparison is used to illustrate that, under the same short-time weak event amplitude, the same observation geometry, and similar noise background, the fast-time waveform distortion dimension utilized in the embodiments of this application has better event characterization capability; the above values ​​are semi-quantitative results under the example parameters of this set and do not constitute a numerical limitation on the scope of protection of the embodiments of this application or all actual deployment conditions.

[0199] It should be understood that the above comparison is based on the specific scenario addressed in this application—transient, non-stationary, and extremely short-duration weak mechanical events. Micro-Doppler analysis has mature and widespread application value in continuous motion monitoring (such as breathing, heart rhythm, and human motion tracking), and this application does not deny its effectiveness in the field of continuous motion sensing.

[0200] Furthermore, to avoid the technical effects of this application remaining solely at the simulation level, preliminary test results from a set of actual UWB radar can be introduced as supplementary experimental support. According to the preliminary bearing fault assessment report generated in the laboratory, in a set of non-contact bearing tests, an actual UWB pulse radar was used to monitor the operating bearing. The radar bandwidth was approximately 1.3 GHz, the spindle speed was approximately 3 Hz (180 rpm), and the tested object was a deep groove ball bearing. After performing time-frequency analysis on the collected channel impulse response data, relatively obvious frequency lines could be observed near the approximately 6 Hz characteristic frequency band, and frequency components of approximately 7 Hz, 12 Hz, and 16 Hz could be observed simultaneously. According to the preliminary statistics of the report, the signal-to-noise ratio of the main characteristic lines was approximately 45 dB to 55 dB, and the relevant impulse response showed a decay trend with time or rotation period. The above phenomena indicate that, under specific laboratory test conditions, the UWB radar CIR characteristics described in this application are consistent with the common physical mechanisms of bearing faults and preliminarily demonstrate the feasibility of identifying periodic impulse events.

[0201] In this preliminary experiment, by combining spindle speed, time-frequency energy distribution, and characteristic frequency positions, a preliminary assessment can be made of the tested bearing, primarily characterized by outer ring fault features, possibly accompanied by related inner ring features. It should be understood that since this test is still in the early evaluation stage, the specific bearing model, precise theoretical fault frequency, disassembly and inspection confirmation results, and cross-validation results with contact sensors can be supplemented in subsequent experiments. Therefore, the above assessment is mainly used to demonstrate that the method of this application has the ability to extract periodic impact information related to bearing faults from CIR waveform distortion under actual laboratory conditions, and should not be construed as a definitive limitation on the fault category, fault severity, or final diagnostic conclusion. For those skilled in the art, further improvements in diagnostic accuracy can be made by combining vibration sensor cross-validation, healthy baseline comparison, time-domain impact waveform extraction, and sideband modulation analysis.

[0202] Furthermore, in the comparative verification of a type of laboratory cage fracture sample, under the premise of maintaining consistent observation distance and rotation speed conditions, the ultra-wideband radar CIR data of normal samples and faulty samples can be compared at the same distance to verify that the fast-time waveform distortion information described in this application is not a purely theoretical derivation, but can be stably observed in real data. In this embodiment, both the 70cm and 90cm samples are used under 10Hz conditions and constitute normal / fault paired controls, respectively.

[0203] Figure 13(a) is a comparison diagram of relative target gates at a distance of 70 cm provided in an embodiment of this application; Figure 13(b) is a comparison diagram of relative target gates at a distance of 70 cm provided in an embodiment of this application; Figure 13(c) is a comparison diagram of the strongest time window at a distance of 90 cm provided in an embodiment of this application; Figure 13(d) is a comparison diagram of the strongest time window at a distance of 90 cm provided in an embodiment of this application. Figures 13(a)-(d) show that at the two observation distances of 70 cm and 90 cm, the faulty samples exhibit more prominent local relative target gate group enhancement and stronger strongest event window response compared to the normal samples. This indicates that the application utilizes not abstract statistics, but local waveform distortion evidence that can be traced back to specific relative gate positions and event center segments.

[0204] Please refer to Figure 14 , Figure 14 This is a comparison chart of transient energy, dual-receiver amplitude correlation, and negative correlation feature counts of normal / fault samples at the same distance, provided for embodiments of this application. Figure 14 Further, from a quantitative perspective, faulty samples show an overall increase in transient energy, an overall decrease in the correlation between dual receiver amplitudes, and a significant increase in the count of negatively correlated features compared to normal samples, thus forming a directional consistency conclusion that can be directly used for review and explanation.

[0205] Figure 15(a) is a comparison chart of event alignment amplitudes at a distance of 70 cm provided in an embodiment of this application; Figure 15(b) is a comparison chart of differential amplitudes at a distance of 70 cm provided in an embodiment of this application; Figure 15(c) is a comparison chart of event alignment amplitudes at a distance of 90 cm provided in an embodiment of this application; Figure 15(d) is a comparison chart of differential amplitudes at a distance of 90 cm provided in an embodiment of this application. Figures 15(a)-(d) show the local amplitude and |ΔCIR| changes around the strongest event center. It can be seen that the fault sample exhibits stronger local perturbations and differential structures near the event center, further verifying the rationality of extracting fast-time waveform distortion information around the event center in this application.

[0206] Statistical analysis of the measured data for the event center window, relative target gate group, and dual-receiver structured indices reveals the following: Under 70cm conditions, the average transient energy of faulty samples increased from 7950.60 to 11243.95, approximately 1.41 times that of normal samples; under 90cm conditions, it increased from 3511.40 to 4809.11, approximately 1.37 times that of normal samples. Simultaneously, the average amplitude correlation between the two receivers decreased from 0.181 to -0.029 under 70cm conditions and from 0.063 to -0.176 under 90cm conditions; the negative correlation feature counts increased from 3 to 14 and from 0 to 19. This indicates that fault conditions not only increase event intensity but also cause changes in cross-receiver structure relationships, thus forming interpretable evidence distinct from background noise and non-fault disturbances.

[0207] The aforementioned laboratory control data uses a paired normal / fault comparison at the same distance to demonstrate the observability of fast-time waveform distortion information in real UWB-CIR data. The target gate group and tap position in the figure are relative to the target gate or relative to the tap aperture, used to characterize the location of local waveform disturbances near the event center, and are not limited to absolute distance gate positioning. The dual-receiver amplitude correlation and negative correlation feature counts are used to characterize changes in the structural relationship between paired receiving channels, indicating that fault conditions not only increase event intensity but also alter the consistency or decorrelation characteristics across receiving channels. The above results provide support for the engineering feasibility and directional consistency of the embodiments of this application, but do not constitute a limitation on achieving the same numerical effects for all devices, all distances, or all fault types.

[0208] Furthermore, regarding the distribution relative to the target gate group, under the 70cm condition, the main disturbance tap migrates from around 6 in the normal sample to around 8 in the faulty sample, and under the 90cm condition, it migrates from around 11 to around 9. This further demonstrates that the event perception described in this application does not rely solely on a single total energy threshold, but can simultaneously observe local waveform distortion consistent with transient mechanical events at three levels: relative to the target gate group, the event center window, and dual receiver consistency.

[0209] Therefore, the aforementioned laboratory control data can serve as experimental supporting material for the technical solution of this application: firstly, it proves that fast-time waveform distortion information is observable in real ultra-wideband radar CIR data; secondly, it proves that a chart difference with directional consistency can be obtained under normal / fault comparison at the same distance; and thirdly, it proves that the technical solution of this application not only has a physical mechanism basis, but also has engineering feasibility supported by real data, chart comparison, and quantitative conclusions.

[0210] Furthermore, embodiments of this application can be deployed as ultra-wideband radar sensing modules, edge intelligent diagnostic devices, cloud-edge collaborative monitoring platforms, or embedded algorithm modules. Different deployment forms all share the same underlying processing link, namely, performing transient mechanical event perception based on UWB-CIR fast-time waveform distortion, and can select functions such as local inference, cloud analysis, historical event caching, alarm reporting, or auxiliary monitoring linkage according to the application scenario. The above deployment forms are only used to illustrate the engineering implementation of this application and do not change the core technical solution of this application.

[0211] It should be noted that in this application, mathematical expressions, convolutional relationships, propagation models, and loss functions are only used to characterize the way natural laws are utilized, algorithmic constraints are imposed, and engineering implementation paths are applied in the technical solution. In particular, the approximate relationship between the event feature waveform and the transient normal acceleration corresponds to the input-output approximate representation in the signal processing stage of this application, serving waveform extraction, feature calculation, and model constraint construction, rather than claiming the abstract formula as an independent entity. For the object of protection in this application, the core lies in utilizing the aforementioned physical relationships to achieve non-contact transient mechanical event perception.

[0212] In the embodiments provided in this application, it should be understood that the disclosed apparatus and methods can be implemented in other ways. The apparatus embodiments described above are merely illustrative. For example, the division of units is only a logical functional division, and in actual implementation, there may be other division methods. Furthermore, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Additionally, the displayed or discussed mutual couplings, direct couplings, or communication connections may be through some communication interfaces; indirect couplings or communication connections between devices or units may be electrical, mechanical, or other forms.

[0213] Furthermore, the units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of the embodiments of this application, depending on actual needs.

[0214] Furthermore, the functional modules in the various embodiments of this application can be integrated together to form an independent part, or each module can exist independently, or two or more modules can be integrated to form an independent part.

[0215] In this document, relational terms such as first and second are used only to distinguish one entity or operation from another entity or operation, without necessarily requiring or implying any such actual relationship or order between these entities or operations.

[0216] The above description is merely an embodiment of this application and is not intended to limit the scope of protection of this application. Various modifications and variations can be made to this application by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the scope of protection of this application.

Claims

1. A non-contact method for sensing transient mechanical events based on the waveform distortion of the impulse response of an ultra-wideband radar channel, characterized in that, include: Under the condition that the ultra-wideband radar has no physical contact with the target medium, channel impulse response data collected after the ultra-wideband radar transmits a pulse signal to the target medium is obtained; Based on the channel impulse response data, waveform distortion information characterized by fast-time local waveform changes relative to the background waveform is extracted, which is related to transient mechanical events inside or on the surface of the target medium. The waveform distortion information is used to obtain the transient mechanical event perception results; The fast-time local waveform changes include local shape changes, time delay changes, pulse width changes, rising edge changes, or peak position changes within a single distance gate and / or between adjacent distance gates; The waveform distortion information is caused by the transient normal acceleration and / or equivalent short-time displacement disturbance generated inside or on the surface of the target medium by the transient mechanical event.

2. The method according to claim 1, characterized in that, The ultra-wideband radar is an impulse pulse radar, whose instantaneous bandwidth of transmitted pulses is not less than 500MHz and whose pulse width is not greater than 2ns.

3. The method according to claim 1, characterized in that, The extraction of waveform distortion information based on the channel impulse response data, characterized by fast-time local waveform changes relative to the background waveform and related to transient mechanical events within or on the surface of the target medium, includes: The channel impulse response data is subjected to clutter removal or background suppression processing to suppress static clutter and slowly varying background, and the waveform distortion information corresponding to the candidate transient mechanical events is extracted based on the processed data; The clutter removal or background suppression process includes at least one of background cancellation, filtering, wavelet decomposition, sparse decomposition, deconvolution, low-rank-sparse separation, and subspace projection.

4. The method according to claim 3, characterized in that, The extraction of waveform distortion information further includes: after the clutter removal or background suppression processing, performing target range gate or target range gate group selection, and performing normalization processing on the channel impulse response data within the target range gate or target range gate group to construct the event window, event waveform segment or model input tensor corresponding to the candidate transient mechanical event.

5. The method according to claim 1, characterized in that, The method further includes: Obtain reference parameters corresponding to the dynamic state of the target medium; the reference parameters include at least one of rotation angle, reciprocating motion position, load phase, external excitation phase, and event triggering time. Based on the reference parameters, the channel impulse response data is resampled on the slow time axis to a unified reference domain.

6. The method according to claim 5, characterized in that, After resampling the channel impulse response data onto the slow time axis to the unified reference domain based on the reference parameters, the method further includes: Calculate the energy statistics of the reference domain within the target range gate; Based on the reference domain energy statistics or the channel impulse response data, at least one of the following methods is used to determine the target phase, target time, or target reference position corresponding to the transient mechanical event: adaptive threshold detection, matched filtering, anomaly detection, autoencoder reconstruction error detection, or triggered network detection.

7. The method according to claim 1, characterized in that, The process of obtaining transient mechanical event perception results using the waveform distortion information includes: Extract event waveform segments from the waveform distortion information; Explicit physical features are extracted based on the event waveform segments; the explicit physical features include at least one of rise time, kurtosis, principal resonant frequency, bandwidth-to-energy ratio, attenuation coefficient, time-frequency ridge parameter, waveform sparsity, energy entropy, and event repetition stability. The transient mechanical event perception result is obtained based on the explicit physical characteristics.

8. The method according to claim 1, characterized in that, The process of obtaining transient mechanical event perception results using the waveform distortion information includes: The channel impulse response data or a data representation constructed based on the channel impulse response data is input into the transient mechanical event perception model; the transient mechanical event perception model is a trained machine learning model or a deep learning model. The transient mechanical event perception model outputs the transient mechanical event perception result; The data representation includes at least one of the following: a distance gate-slow time matrix, a reference domain matrix, and an event slice tensor.

9. The method according to claim 8, characterized in that, The machine learning model or deep learning model introduces a constraint loss term during training based on the physical correlation between the transient mechanical event and the waveform distortion information.

10. The method according to claim 8, characterized in that, The method further includes: Simulated channel impulse response data are generated based on stress wave propagation model, target medium transmission model and radar echo model; The simulated channel impulse response data is used to train, optimize, or validate the transient mechanical event perception model.

11. The method according to claim 1, characterized in that, The target medium is a rotating mechanical structural component, and the transient mechanical event perception results include at least one of the following: fault type, fault severity, fault angle location, health index, and remaining life.

12. The method according to claim 5, characterized in that, The target medium is a rotating mechanical structural component, and the reference parameters include speed encoder pulses, Hall phase signals, or rotation angle phases estimated based on other sensor data.

13. The method according to claim 11 or 12, characterized in that, Also includes: Obtain vibration data, temperature data, or load data corresponding to the rotating mechanical structural component; The features extracted from the vibration data, temperature data, or load data are fused with the features corresponding to the waveform distortion information to obtain fault diagnosis results.

14. The method according to claim 1, characterized in that, The target medium is the cardiac mechanical activity observation area corresponding to the human chest wall or back, and the transient mechanical event sensing results include at least one of heart rate, cardiac cycle stability, myocardial mechanical function parameters, and valve opening and closing timing parameters.

15. The method according to claim 1, characterized in that, The target medium is an aerospace structural component or a civil engineering structure, and the transient mechanical event perception result includes at least one of the following: crack, delamination, delamination, micro-damage, impact damage, or loosening.

16. The method according to claim 1, characterized in that, The method further includes: When the transient mechanical event perception results meet the preset conditions, slow-time / micro-Doppler auxiliary analysis, contact vibration monitoring, or other auxiliary monitoring based on the same ultra-wideband radar acquisition data is initiated. The auxiliary analysis results or auxiliary monitoring results are cross-validated or fused with the transient mechanical event perception results for decision-making.

17. The method according to claim 1, characterized in that, The method further includes: Extract the event feature waveform w(t) from the channel impulse response data, and extract explicit physical features or construct model constraints based on the event feature waveform; Wherein, the event characteristic waveform w(t) and the transient normal acceleration The following approximation relationship is satisfied: ； in, This represents the transient normal acceleration on the target surface caused by a transient mechanical event. c is the speed of light, and * indicates convolution operation. For fast time variables, The waveform is the transmitted pulse; the approximation relationship is used to guide the extraction of explicit physical features or the construction of model constraints from the waveform distortion information.

18. The method according to claim 1, characterized in that, The channel impulse response data includes at least one of complex I / Q samples, amplitude samples, phase samples, power delay spectrum, range gate sequence, relative tap index, or data representation derived from the above data, and includes at least one of CIR window start position, first path index, timestamp, frame number, transmit / receive antenna identifier, channel identifier, received power index, channel quality index, or synchronization status identifier; the waveform distortion information is obtained by performing target range gate selection, background suppression, reference domain alignment, event window slicing, or model inference based on at least one of the above fields.

19. The method according to claim 1 or 17, characterized in that, Based on the fast-time waveform distortion in the channel impulse response data, an event feature waveform w(t) related to the transient normal acceleration of the target medium surface is constructed, and explicit physical features or model constraints are extracted based on the event feature waveform w(t).

20. A transient mechanical event sensing system based on the waveform distortion of the impulse response of an ultra-wideband radar channel, characterized in that, include: The data acquisition module is used to acquire channel impulse response data collected by the ultra-wideband radar under conditions where there is no physical contact between the ultra-wideband radar and the target medium. The information extraction module is used to extract waveform distortion information, characterized by fast-time local waveform changes relative to the background waveform, based on the channel impulse response data and related to transient mechanical events inside or on the surface of the target medium. The result generation module is used to obtain transient mechanical event perception results using the waveform distortion information; The fast-time local waveform changes include local shape changes, time delay changes, pulse width changes, rising edge changes, or peak position changes within a single distance gate and / or between adjacent distance gates; The waveform distortion information is caused by the transient normal acceleration and / or equivalent short-time displacement disturbance generated inside or on the surface of the target medium by the transient mechanical event.

21. The system according to claim 20, characterized in that, It also includes one or more of the following modules: The synchronization module is used to acquire reference parameters corresponding to the dynamic state of the target medium; The preprocessing module is used to perform clutter removal or background suppression processing; The feature extraction module is used to extract explicit physical features; The model inference module is used to perform inference for machine learning models or deep learning models. The physical information data enhancement module is used to generate simulated channel impulse response data.

22. An electronic device, characterized in that, include: Ultra-wideband radar module; And processors and memory; The processor is communicatively connected to the ultra-wideband radar module, and the memory stores program instructions that can be executed by the processor. When the program instructions are executed by the processor, the processor enables the processor to implement the method as described in any one of claims 1-19.

23. A non-transitory 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 1-19.

24. A computer program product, characterized in that, It includes computer program instructions that, when executed by a processor, implement the method as described in any one of claims 1-19.

25. A method for extracting transient mechanical event feature waveforms based on the impulse response of an ultra-wideband radar channel, characterized in that, include: Acquire channel impulse response data after an ultra-wideband radar transmits a pulse signal to a target medium; Based on the fast-time waveform distortion in the channel impulse response data, an event feature waveform w(t) related to the transient normal acceleration of the target medium surface is constructed. Wherein, the event characteristic waveform w(t) and the transient normal acceleration The following approximation relationship is satisfied: ； in, This represents the transient normal acceleration on the target surface caused by a transient mechanical event. c is the speed of light, and * indicates convolution operation. For fast time variables, This is the transmitted pulse waveform.

Images