Bridge support accumulated slip long-term monitoring device and method based on microwave vibration measurement

Abstract

This application discloses a long-term monitoring device and method for cumulative slippage of bridge bearings based on microwave vibration measurement, including: a microwave vibration radar module; a data acquisition and preprocessing module; a time-frequency feature extraction module; a cumulative slippage calculation module connected to the time-frequency feature extraction module, used to calculate the absolute cumulative slippage and net cumulative slippage based on the low-frequency displacement components, and to identify friction anomalies and jamming events based on the high-frequency vibration components; and a condition assessment and early warning module. This application introduces a slippage trend prediction network based on gated cyclic units. This network integrates multi-dimensional features such as cumulative slippage, slippage rate, high-frequency energy, and slippage entropy, enabling it to learn the long-term evolution law of bearing slippage and predict the slippage trend over a future period. Combined with preset multi-level thresholds, it can issue early, graded, and targeted early warning signals, thereby achieving prediction-based graded early warning and providing data support for bridge maintenance decisions.

Description

Technical Field

[0001] This application belongs to the field of bridge structural health monitoring technology, specifically relating to a long-term monitoring device and method for cumulative slippage of bridge bearings based on microwave vibration measurement. Background Technology

[0002] Bridge bearings, as key components connecting the superstructure and substructure of a bridge, directly affect the overall safety and operational life of the bridge. Bearing slippage (especially the cumulative slippage of the PTFE sliding plate) is a core indicator for assessing bearing wear and determining whether limit failure or jamming has occurred. Long-term, reliable monitoring of bearing cumulative slippage is crucial for preventative bridge maintenance and ensuring traffic safety.

[0003] Currently, the following technologies are mainly used for monitoring bridge bearing slippage: Traditional displacement sensor methods, such as wire-type displacement gauges and laser displacement gauges, require the sensor to be directly mounted between the upper and lower plates of the support. This process is complex and costly, and the sensors are susceptible to corrosion, aging, and mechanical fatigue due to long-term exposure to outdoor environments, resulting in insufficient reliability and durability. Furthermore, contact-based measurements may interfere with the free sliding of the support.

[0004] Image / visual measurement method: This method involves taking pictures of the target on the support with a fixed camera and calculating the displacement using image processing algorithms. However, this method is susceptible to interference from environmental factors such as changes in lighting, rain, fog, and dust. Furthermore, in long-term monitoring, it suffers from significant problems such as camera parameter drift, large image data volume, and limited real-time processing capabilities, making it difficult to achieve continuous, automated monitoring around the clock.

[0005] Acceleration integration method: This method obtains displacement by integrating the acceleration signal twice using an accelerometer installed on the support. However, this method suffers from low-frequency drift and cumulative error, and cannot accurately reflect the quasi-static slow slippage and long-term cumulative slippage of the support.

[0006] In summary, existing monitoring technologies generally suffer from the following problems: inconvenient installation and maintenance, susceptibility to environmental influences, difficulty in accurately capturing both high-frequency dynamic responses and long-term slow slippage, lack of effective differentiation between slippage types (normal expansion and contraction versus abnormal sliding), and inability to predict support wear trends and potential faults in advance. Therefore, there is an urgent need for a non-contact, high-precision, robust device and method capable of long-term, automated cumulative slippage monitoring and intelligent early warning. Summary of the Invention

[0007] This application provides a long-term monitoring device and method for cumulative slippage of bridge bearings based on microwave vibration measurement, aiming to solve the problems in the prior art that bridge bearing slippage monitoring is difficult to achieve non-contact, high-precision long-term measurement, cannot effectively distinguish between normal expansion and contraction caused by temperature and abnormal slippage events such as friction and jamming, and lacks comprehensive quantitative assessment and graded early warning of cumulative slippage wear, net offset and future slippage trend.

[0008] Firstly, a long-term monitoring device for cumulative slippage of bridge bearings based on microwave vibration measurement includes: The microwave vibration radar module is used to transmit and receive frequency-modulated continuous wave signals, and outputs I / Q baseband signals after orthogonal mixing; The data acquisition and preprocessing module is connected to the microwave vibration radar module and is used to perform analog-to-digital conversion and preprocessing on the I / Q baseband signal to reconstruct the original displacement signal sequence. The time-frequency feature extraction module, connected to the data acquisition and preprocessing module, is used to perform time-frequency decomposition on the original displacement signal sequence to separate the low-frequency displacement component reflecting the normal expansion and contraction of the support and the high-frequency vibration component reflecting dynamic events. The cumulative slip calculation module, connected to the time-frequency feature extraction module, is used to calculate the absolute cumulative slip and net cumulative slip based on the low-frequency displacement component, and to identify friction anomalies and jamming events based on the high-frequency vibration component. The condition assessment and early warning module, connected to the cumulative slip calculation module, is used to assess the health status of the support and selectively issue graded early warning signals based at least on the identification results of the absolute cumulative slip, net cumulative slip, friction anomalies, and jamming events.

[0009] Optionally, the data acquisition and preprocessing module is specifically used to: perform analog-to-digital conversion on the I / Q baseband signal, and then sequentially perform DC offset correction, amplitude-phase imbalance compensation, phase unwrapping, and displacement conversion to generate the original displacement signal sequence; and the data acquisition and preprocessing module is also used to acquire the reference channel displacement pointing to the fixed pier wall, and subtract the reference channel displacement from the support channel displacement to eliminate common-mode vibration interference.

[0010] Optionally, the time-frequency feature extraction module has a built-in wavelet packet decomposition unit and a short-time Fourier transform unit, and adaptively selects the decomposition method according to the motion state of the original displacement signal sequence: by default, the wavelet packet decomposition unit is used to decompose the original displacement signal sequence into multiple frequency band subspaces, and the low-frequency displacement component and the high-frequency vibration component are obtained by setting the coefficients of the specified sub-bands to zero and reconstructing them. When an intermittent strong impact is detected in the displacement signal, the system temporarily switches to the short-time Fourier transform unit to output high-frequency transient features in the form of a time-spectrum matrix and impact feature flags.

[0011] Optionally, the cumulative slip calculation module includes: a direction determination unit, which determines the slip direction based on the comparison result of the instantaneous rate of the low-frequency displacement component and the velocity threshold, combined with hysteresis logic; The absolute cumulative slip calculation unit accumulates the absolute displacement difference between adjacent sampling points when it is determined that the support has actually slipped, and outputs the total slip path length. The net cumulative slip calculation unit calculates the algebraic sum of the low-frequency displacement components relative to the initial moment and outputs the overall offset of the support. A short-time energy calculation unit calculates the short-time energy of the high-frequency vibration component; The anomaly identification unit, based on the dynamic baseline threshold method, determines static friction anomalies, sliding friction anomalies, and jamming-release events according to the short-term energy and the instantaneous rate.

[0012] Optionally, the state assessment and early warning module incorporates a slip trend prediction network based on a gated recurrent unit. The input feature vector of the slip trend prediction network includes at least the absolute cumulative slip, net cumulative slip, average slip rate, high-frequency energy, and slip entropy. The state assessment and early warning module is used to predict the amount of slip in the future time period based on the slip trend prediction network output, and trigger the corresponding level of early warning signal in advance by combining the preset level threshold.

[0013] Optionally, the state assessment and early warning module is also used to execute an incremental learning self-update mechanism: when the prediction error continuously exceeds a set threshold, an online fine-tuning program is triggered to update the final layer weights of the gating recurrent unit's slip trend prediction network with an update step size lower than the initial learning rate; and it is used to merge all newly accumulated monitoring data with the initial training set every quarter to perform a complete incremental retraining.

[0014] Optionally, it may also include a data storage and communication module; The data storage and communication module is connected to the cumulative slip calculation module and the status assessment and early warning module respectively. It is used to store the original displacement data, characteristic parameters, health status records and early warning event logs with different priorities, and send differentiated early warning messages to the remote monitoring platform by wired or wireless communication according to the early warning level.

[0015] Secondly, a long-term monitoring method for cumulative slippage of bridge bearings based on microwave vibration measurement includes the following steps: S1: I / Q baseband signals are acquired by a microwave vibration radar module installed on the side of the bridge pier; S2: Perform analog-to-digital conversion, DC offset correction, imbalance compensation, phase unwrapping and displacement conversion on the I / Q baseband signal, and use the reference channel to eliminate common-mode vibration interference to reconstruct the original displacement signal sequence; S3: Perform time-frequency decomposition on the original displacement signal sequence, using wavelet packet decomposition or short-time Fourier transform to separate the low-frequency displacement component reflecting the normal expansion and contraction of the support and the high-frequency vibration component reflecting dynamic events. S4: Based on the low-frequency displacement components, calculate the absolute cumulative slip and net cumulative slip through direction discrimination and threshold logic, and calculate the slip entropy; S5: Calculate short-time energy based on the high-frequency vibration components, and identify friction anomalies and jamming-release events using the dynamic baseline threshold method; S6: Based at least on the identification results of the absolute cumulative slip, net cumulative slip, friction anomaly and jamming event, combined with the predicted value output by the slip trend prediction network based on the gated loop unit, assess the health status of the support and issue a graded early warning signal according to the graded threshold.

[0016] Optionally, the phase unwrapping in S2 employs a cumulative offset correction method: setting a jump detection threshold. When the absolute value of the phase difference between adjacent sampling points is greater than At that time, the subsequent phase is increased or decreased as a whole according to the direction of the difference. In the displacement conversion, the instantaneous radial displacement is calculated by multiplying the unwrapped continuous phase by a coefficient. Received, among which This is the radar carrier wavelength.

[0017] Compared with the prior art, this application has at least the following beneficial effects: This application employs millimeter-wave frequency-modulated continuous wave (FMCW) radar vibration measurement technology, reconstructing displacement from phase information by transmitting and receiving microwave signals. This non-contact measurement method avoids interference from traditional sensor installations on support movement and is unaffected by environmental factors such as rain, fog, and dust. Combined with reference channel common-mode cancellation technology, it effectively suppresses low-frequency swaying of the bridge structure itself and platform micro-vibration interference, enabling high-precision displacement measurement and providing a reliable data foundation for long-term accurate monitoring.

[0018] This application utilizes a built-in time-frequency feature extraction module to employ wavelet packet decomposition or short-time Fourier transform, enabling the adaptive decomposition of the original displacement signal into low-frequency components reflecting temperature expansion and contraction and normal slip, and high-frequency components reflecting dynamic events such as frictional vibration and jamming impact. This signal decoupling process makes the analysis of support sliding behavior more physically meaningful and can accurately distinguish between normal service wear and fault symptoms.

[0019] This application differs from simply measuring total displacement. It utilizes the decomposed low-frequency displacement components to simultaneously calculate both "absolute cumulative slip" and "net cumulative slip." The former characterizes the total wear history of the sliding plate material, while the latter reflects whether the support has experienced overall displacement. By combining direction discrimination and threshold logic, minor reciprocating vibration interference caused by thermal expansion and contraction is eliminated, making the slip wear statistics more accurate and providing crucial quantitative evidence for evaluating support life and the effectiveness of the limiting device.

[0020] This application effectively identifies static friction anomalies, sliding friction anomalies, and highly dangerous jamming-release events by calculating the short-time energy of high-frequency vibration components and employing a dynamic baseline thresholding method. In particular, the precise capture of jamming impacts (including start and end times, peak acceleration, and impact energy) provides evidence for detecting internal structural damage to supports, overcoming the limitation of traditional displacement monitoring in detecting vibration characteristics.

[0021] This application introduces a slip trend prediction network based on gated recurrent units (GRUs). This network integrates multi-dimensional features such as cumulative slip, slip rate, high-frequency energy, and slip entropy, enabling it to learn the long-term evolution of bearing slip and predict slip trends over a future period. Combined with preset multi-level thresholds—based on the bearing's design allowable wear and offset—it is used to grade and assess the bearing's health status and issue corresponding early warning signals. The system can issue early, graded, and targeted early warning signals, thereby achieving prediction-based graded early warning and providing data support for bridge maintenance decisions. Attached Figure Description

[0022] Figure 1 This is a schematic diagram of the module connection of a long-term monitoring device for cumulative slippage of bridge bearings based on microwave vibration measurement, provided in one embodiment of this application.

[0023] Figure 2 This is a flowchart illustrating a long-term monitoring method for cumulative slippage of bridge bearings based on microwave vibration measurement, provided as an embodiment of this application. Detailed Implementation

[0024] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments.

[0025] The long-term monitoring device for cumulative slippage of bridge bearings based on microwave vibration measurement provided in this application, such as... Figure 1 As shown, the device includes: a microwave vibration radar module, a data acquisition and preprocessing module, a time-frequency feature extraction module, a cumulative slip calculation module, a state assessment and early warning module, and a data storage and communication module; The microwave vibration radar module adopts a frequency-modulated continuous wave system and consists of five parts: a millimeter-wave transceiver chip, a transmitting antenna, a receiving antenna, an intermediate frequency signal conditioning circuit, and an embedded microcontroller.

[0026] The transceiver chip is a single-chip millimeter-wave transceiver with a working frequency band of 76-81GHz. It integrates two transmit channels and four receive channels on the chip, and can realize a multi-input multi-output virtual antenna array. In support sliding tracking, it forms a horizontal angular resolution of about 2° and an elevation angular resolution of about 5°. Based on this, the radial displacement along the support sliding direction can be separated from the lateral swing perpendicular to the sliding direction. The antenna employs a microstrip patch array, printed on the same substrate as the transceiver chip's circuit board, eliminating the need for additional RF cables and reducing signal insertion loss and parasitic radiation. The intermediate frequency signal conditioning circuit sequentially performs low-noise amplification, bandpass filtering, and analog-to-digital conversion on the beat signal output from the receiving channel. The embedded microcontroller is responsible for configuring the timing parameters of the control chip (such as starting and stopping radar transmission, adjusting transmission power, setting the sweep slope and sampling frequency), and sending the intermediate frequency signal data after analog-to-digital conversion to the next-level data acquisition and preprocessing module through a serial peripheral interface or a universal asynchronous transceiver. The microwave vibration radar module is encapsulated within an aluminum alloy housing, with an antenna window mounted on the front and electrical and mechanical mounting interfaces on the rear. The housing has a protection rating of at least IP67 to withstand the humid, condensation, and salt spray environments of bridge substructures. The mechanical mounting interface consists of a two-degree-of-freedom precision turntable, vibration-damping connectors, and a rigid mounting bracket. The vibration damping connector adopts a corrugated steel plate and rubber composite laminate structure. Bridge environmental excitation tests show that it can attenuate the vibration transmission from the bridge pier by about 60% in the 1-200Hz range, ensuring that the phase modulation of the millimeter wave signal mainly comes from the support movement rather than platform jitter. The pitch and horizontal adjustment ranges of the two-degree-of-freedom precision turntable are ±15° and ±30°, respectively, with an adjustment resolution of less than 0.1°, so that the radar beam main lobe can be precisely aligned with the reflector on the support sliding surface after adjustment. After the initial on-site installation, the conversion relationship between the radar coordinate system and the overall bridge coordinate system was calibrated using a total station. Three locating pins evenly distributed on the turntable base were used as physical references for subsequent repeated installations, ensuring that the repositioning error of the beam alignment after replacing or moving the radar equipment did not exceed 2mm. The rigid fixing bracket was fixed to the side wall of the pier or the bottom of the cap beam using chemical anchors or mechanical expansion bolts. The bottom of the bracket was equipped with a horizontal adjustment screw, which could adjust the relative height between the radar front end and the installation reference surface within a range of ±10mm. The radar operating frequency band was selected as 76-81 GHz, mainly based on the following considerations: First, the shorter carrier wavelength (approximately 3.7-3.9 mm) imparts higher displacement sensitivity to the phase extraction stage, based on the ratio of displacement to phase change. The theoretical relationship is that a 1° phase change corresponds to a displacement of about 10μm, which provides sufficient measurement resolution for identifying early fretting wear of the PTFE slide plate of the support. Secondly, the atmospheric attenuation in this frequency band is in a relatively low window of the millimeter wave spectrum (about 0.2-0.5 dB / km), which is almost negligible within the line-of-sight distance of most bridge supports (usually 5-30 m). Third, the 77-81GHz frequency resources have been made widely available for vehicle-mounted radar, and the supply chain for related transceiver chips is mature and cost-controllable, avoiding the obstacles to productization caused by using closed-source or prohibited frequency bands. The sweep bandwidth is set to 3.5GHz, corresponding to a theoretical distance resolution of approximately 4.3cm, which is sufficient to distinguish the support reflector from background reflection points such as bridge pier walls and guardrails in terms of distance. The radar horizontal beamwidth is controlled within ±3°, and the vertical beamwidth is controlled within ±3°. A narrower beam helps to reduce sidelobe coverage and thus minimize false echo interference, while also facilitating precise target locking via the main lobe. The sampling rate is designed according to the required upper limit of the bearing slip rate: the maximum slip rate of bridge bearings under normal operating conditions typically does not exceed 5 mm / s, and the daily cyclic displacement period caused by temperature changes is approximately 10-12 hours, but the natural frequency caused by sudden release of the jamming can reach tens to hundreds of hertz. Therefore, the pulse repetition frequency of the intermediate frequency signal is set to be no less than 1kHz, so that the maximum measurable vibration frequency corresponding to the Nyquist frequency reaches more than 500Hz, which is sufficient to cover the frequency band of the support locking release impact. The microwave vibration radar module transmits an FMCW signal and receives the echo modulated by the reflecting target. After quadrature mixing, it outputs two baseband signals, I and Q. After level conversion and analog-to-digital conversion, it generates a raw intermediate frequency data stream, which is transmitted to the data acquisition and preprocessing module via an Ethernet interface. This data stream fully realizes the conversion from microwave physical quantities to digital domain quantized signals, providing the raw input for subsequent phase-based displacement reconstruction.

[0027] The data acquisition and preprocessing module adopts a plug-in box structure with a backplane and function daughter cards, and is installed in the bridge pier power distribution box or in a protective enclosure shared with the radar module. The data acquisition and preprocessing module includes the following sub-units: high-speed analog-to-digital converter, programmable anti-aliasing filter, embedded signal processor, synchronous buffer memory, and power management circuit; Specifically, the high-speed analog-to-digital converter (ADC) is a dual-channel synchronous sampling successive approximation ADC with a quantization bit depth of 16 bits. The maximum sampling rate of a single channel is set to match the intermediate frequency signal output bandwidth of the microwave vibration radar module, and is no less than 2MHz. The two channels correspond to the analog-to-digital conversion of the I and Q baseband signals, respectively. The sampling time is triggered by the frame synchronization pulse output by the radar module to ensure the accuracy of the I / Q phase quadrature relationship during displacement reconstruction. The analog-to-digital converter integrates a differential input driver amplifier at the front end, which adjusts the differential I / Q signal output by the radar module into a single-ended signal (typical range ±2.5V) within the full-scale range of the analog-to-digital converter input. A DC bias calibration register is also introduced, which can drive the internal digital-to-analog converter to perform DC bias zeroing during the power-on self-test phase by the embedded processor, eliminating static errors caused by temperature drift or channel inconsistency.

[0028] The programmable anti-aliasing filter is a fourth-order Butterworth low-pass filter, and its cutoff frequency can be adjusted via a digital potentiometer set by the embedded processor. The default cutoff frequency is set to 0.4 times the sampling rate (i.e., 0.8MHz), which suppresses high-frequency noise above this frequency and interference components from the external electromagnetic environment, preventing spectral aliasing during analog-to-digital conversion. The filter's gain flatness within the passband is better than ±0.1dB, ensuring amplitude consistency between the I and Q signals. The synchronous buffer memory employs a two-level cache architecture consisting of a double data rate synchronous dynamic random access memory (DRAM) and a ping-pong static random access memory (SRAM). After analog-to-digital conversion, the data is first stored in a set of buffer pages in the ping-pong SRAM. When a page is full, the embedded processor automatically switches to another page for subsequent writing, while simultaneously initiating a direct memory access channel to transfer the data from the full page to the double data rate synchronous dynamic random access memory. The capacity of the double data rate synchronous dynamic random access memory (DRAM) is sufficient to buffer raw data continuously for 24 hours without data loss. Based on a typical sampling rate of 1kHz for bridge bearings and 4 bytes per sample (dual-channel 16-bit), the 24-hour data volume is approximately 345MB. Therefore, a 512MB DRAM configuration provides sufficient margin. This design allows the module to independently retain all raw data during temporary communication interruptions, and retransmit it after communication is restored, avoiding data loss due to intermittent network failures. The embedded processor uses an ARM Cortex-M7 with a floating-point unit or an equivalent real-time processor with a clock speed of at least 400MHz. It integrates tightly coupled on-chip memory to store the real-time intermediate frequency (IF) signal preprocessing algorithm code. The processor reads the sampling results from the analog-to-digital converter via a serial peripheral interface or parallel bus, and sends the I / Q data stream to the preprocessing pipeline. The input to the preprocessing pipeline is the unprocessed I / Q baseband sampling sequence, and the output is the discrete sequence of the original displacement signal obtained through phase calculation. The time stamp of this sequence is strictly aligned with the radar sampling period interval and is pushed to the next-stage time-frequency feature extraction module at a data rate of 1200Hz (i.e., 1.2kHz). The preprocessing procedure sequentially executes four sub-steps: DC offset correction, unbalance compensation, phase unwrapping, and displacement conversion, as detailed below: (i) DC offset correction: Before the radar starts monitoring, the embedded processor shuts down the transmit channel and only collects an environmental noise sample of length N (N≥512) through the receive channel to calculate the average DC component of the I and Q paths. After the signal is officially transmitted, this average DC value is subtracted from the sampled value in real time to eliminate the fixed bias introduced by mixer self-mixing and circuit background noise; (ii) Unbalanced compensation: Due to potential gain differences between the analog mixer and analog-to-digital converter channels, the reconstructed phase trajectory may appear elliptical rather than a theoretical circle. The processor has a built-in calibration program that, during initial installation or quarterly routine calibration, drives the radar to collect sufficiently long I / Q data from a stationary reflective target. The least-squares elliptic fitting method is used to extract the gain mismatch coefficient and orthogonal phase deviation. During actual monitoring, this coefficient is used to perform a linear transformation on the real-time I / Q samples, restoring the ellipse to a standard circle, thus ensuring the linearity of the arctangent demodulation.

[0029] (iii) Phase unwrapping: The corrected I / Q data is used to obtain the wrapped phase through arctangent operation. Its range is When the true phase change caused by the support motion exceeds At that time, the wrapped phase will produce A jump occurs. The processor employs a phase unwrapping algorithm based on cumulative correction: a dynamic threshold is set. When the absolute value of the phase difference between two adjacent sampling points is greater than When a transition occurs, the phase value is determined to increase or decrease accordingly based on the direction of the transition. Continuous phase after untangling The change in radial distance between the support and the radar is linearly related.

[0030] (iv) Displacement conversion, continuous phase Through formula Converted to instantaneous radial displacement, where The radar carrier wavelength is approximately 3.95 mm at 76 GHz and approximately 3.70 mm at 81 GHz. The processor calculates the precise wavelength in real time based on the current channel center frequency. value); To eliminate common-mode displacement caused by low-frequency swaying of the bridge as a whole or micro-vibration of the radar platform, the processor has an additional environmental compensation channel (if the radar module has more than two receiving channels, the echo pointing towards the fixed pier can be selected as the reference). The relative displacement is obtained by subtracting the reference channel displacement from the support channel displacement. ; After preprocessing, the embedded processor packages the raw displacement signal sequence into binary data blocks. Each data block contains the following fields: timestamp (accurate to milliseconds, synchronized using GPS or Network Time Protocol), sampling period (microseconds), radar configuration parameter number, displacement sequence array (each displacement value is a 32-bit floating-point number in millimeters), and cyclic redundancy check (CRC) code. While the data blocks are stored in a double data rate synchronous dynamic random access memory, a set of data copies is sent out via the communication module to ensure synchronization between real-time transmission and local redundant backup. The output displacement sequence... This serves as the input for the subsequent time-frequency feature extraction module; The embedded processor incorporates gross error detection logic: when the displacement change rate of three consecutive sampling points exceeds ten times the maximum physical sliding rate of the support (at 5 mm / s and a sampling rate of 1.2 kHz, meaning the change rate between adjacent points should not exceed approximately 4.2 μm), or when the absolute displacement value abruptly exceeds the range setting (e.g., a short-term jump exceeding 2 mm), the data point is determined to be an abnormal pulse, marked as invalid, and replaced with a linear interpolation of the preceding and following valid values. This mechanism can suppress outliers caused by transient interference such as birds flying across the beam path or large vehicles obstructing the beam, ensuring the stability of subsequent feature extraction. The I / Q baseband signal received from the microwave vibration radar module via gigabit Ethernet is sampled by an analog-to-digital converter to generate a digital I / Q stream. This stream then undergoes DC offset correction, imbalance compensation, arctangent demodulation, phase unwrapping, and displacement conversion in the embedded processor, before outputting the original displacement signal sequence. The sequence is stored in a synchronization buffer as data blocks and transmitted in real time to the time-frequency feature extraction module via an internal data bus.

[0031] The time-frequency feature extraction module receives the raw displacement signal sequence from the data acquisition and preprocessing module. (sampling rate) The data block length is preset to 30 seconds per segment, meaning each segment contains 36,000 sampling points. The module's main function is to decompose the one-dimensional displacement signal into low-frequency components reflecting slow temperature changes and normal support expansion and contraction, and high-frequency components reflecting dynamic events such as frictional vibration and jamming impacts during support sliding. The output consists of two reconstructed time-domain signals (low-frequency components). High-frequency components The output includes a time-frequency feature matrix F, where each row of F corresponds to a time window and each column corresponds to a frequency band energy or statistical feature value. The output is then passed to the cumulative slip calculation module and the state assessment and early warning module via memory sharing or a message queue. The time-frequency feature extraction module incorporates two time-frequency decomposition methods: wavelet packet decomposition and short-time Fourier transform, which are adaptively selected by the embedded processor based on the support's motion state. Wavelet packet decomposition is used by default because it has better time-frequency localization capabilities for non-stationary signals, and the decomposed frequency band components can be directly used for time-domain reconstruction, facilitating subsequent cumulative slip calculations. When the processor detects intermittent strong impacts in the displacement signal (such as a stuck release spike, whose time-domain peak value exceeds 5 times the normal fluctuation amplitude), it temporarily switches to short-time Fourier transform mode to capture the start and end times and frequency spread range of the impact with higher time resolution. The feature dimensions output by the two methods are aligned and then uniformly input into subsequent modules to ensure that the evaluation model is not affected by the switching of decomposition methods. Specifically, the implementation process of wavelet packet decomposition is as follows: Wavelet basis function and decomposition level selection. Based on the analysis and comparison of numerous measured vibration signals from bridge bearings, the db10 wavelet (Daubechies10), with its good compact support and regularity, was selected as the mother wavelet. The decomposition level was set to 4 levels, which divides the signal frequency band into 2... 4 =16 equal-width sub-bands. At a sampling rate of 1200Hz, the Nyquist frequency is 600Hz, and the bandwidth of each sub-band is 600 / 16 = 37.5Hz. The first three sub-bands (0-37.5Hz, 37.5-75Hz, 75-112.5Hz) cover the main frequency bands of slow temperature changes and normal expansion and contraction (usually below 5Hz); sub-bands 4 to 16 (112.5-600Hz) cover support friction, jamming release impact, and high-frequency environmental disturbances. This division method retains sufficient low-frequency resolution (37.5Hz is sufficient to resolve temperature changes at the 0.01Hz level) while subdividing the high-frequency part to distinguish different types of fault characteristics (for example, the main frequency of jamming release is generally 150-300Hz). Decomposition and Reconstruction Process. A 36,000-point displacement data block undergoes a four-level wavelet packet decomposition to obtain wavelet packet coefficients for 16 nodes. For the reconstruction of the low-frequency scaling component: the wavelet packet coefficients of the first sub-band (0-37.5Hz) are selected, the coefficients of the remaining sub-bands are set to zero, and an inverse wavelet packet transform is performed to obtain the low-frequency reconstructed signal. For the reconstruction of high-frequency vibration components: select the wavelet packet coefficients of sub-bands 4 to 16, set the coefficients of sub-bands 1 to 3 to zero, and perform an inverse transform to obtain the high-frequency reconstructed signal. To ensure that the decomposition boundary does not introduce spurious oscillations, 200 sampling points are extended at each end of each data block (using a symmetrical extension method), and the extended portion is cut off after reconstruction. Frequency band energy feature extraction. The root mean square value of the wavelet packet coefficients for each sub-band is calculated. (i=1,...,16, (where N is the number of coefficients, and the i-th subband is the k-th coefficient). Construct the time-frequency characteristic matrix. The rows correspond to time windows (every 30 seconds), and the columns correspond to the energy values ​​of 16 sub-bands. To reduce the feature dimensionality, sub-bands 4-16 are merged into frequency groups of 3 sub-bands each (4 frequency groups in total). The total energy within each frequency group is calculated, resulting in a 5-dimensional feature vector consisting of low-frequency energy (1-dimensional) and high-frequency energy of the sub-groups (4-dimensional), which serves as one of the inputs to the subsequent state evaluation module. The specific process of the short-time Fourier transform mode is as follows: Upon detecting an impact event, the module switches to short-time Fourier transform mode. The Hanning window is used, with a window length L = 256 sampling points (approximately 0.213 seconds), an overlap ratio of 75%, and a frame shift R = 64 sampling points. This window length strikes a balance between frequency resolution and temporal resolution: frequency resolution... This is sufficient to distinguish similar fault frequency components; the time window covers approximately 0.2 seconds, capturing the transient process of impact events. The short-time Fourier transform points are set to 512, and zero-padding is applied to the signal within the window to improve spectral smoothness. The time-spectrum matrix is ​​constructed by performing a Fast Fourier Transform (FFT) on each frame to calculate the amplitude spectrum S(m,k), where m is the frame index and k is the frequency index. The frequency axis is divided into five bands of interest: ultra-low frequency band (0-5Hz, normal stretching), low frequency band (5-50Hz, structural response), mid frequency band (50-200Hz, frictional vibration), high frequency band (200-400Hz, jamming impact), and ultra-high frequency band (400-600Hz, noise floor). The amplitude spectrum values ​​within each band are summed to obtain the frame-level band energy sequence. The band energy sequences of all frames are then concatenated on the time axis to form the time-spectrum feature matrix. Its size is (number of frames M) × (5 frequency bands). Since the detected impact events are typically short-lived (approximately 1-2 seconds), they generate a maximum of about 10-20 frames. The amount of data is much smaller It is suitable for detailed analysis of transient events; The impact characteristic flag is generated after the module completes the short-time Fourier transform mode, extracting the peak frequency, maximum energy band, and impact duration within the impact period, and generating a custom-formatted impact characteristic flag (containing three fields: peak frequency value, maximum impact amplitude, and start and end timestamps). This flag is related to... They are stored together in the cache so that the status assessment module can read them first when calling them, so as to quickly trigger the abnormal warning logic.

[0032] The time-frequency feature extraction module runs in a dedicated core of a standalone digital signal processor or embedded processor, exchanging data with the data acquisition and preprocessing module via a dual-port random access memory (DRAM). The raw shift sequence is written to the input buffer of the DRAM (capacity of two data blocks, i.e., 60 seconds of raw data). The module employs pipelined processing: while the current block undergoes wavelet packet decomposition and feature extraction, the next block of raw data is simultaneously prepared in another buffer, ensuring zero waiting overhead during processing. After processing each data block, the low-frequency components are... and high frequency components Each component is written to its respective output buffer, and the time-frequency feature matrix FF is written to the circular queue. The cumulative slip calculation module reads the high-frequency and low-frequency components from the output buffer via interrupt, while the state evaluation module pulls the feature matrix from the circular queue as needed.

[0033] Furthermore, the module performs a self-check daily at dawn during the period of lowest bridge traffic: collecting 10 seconds of static displacement data (at which time the supports are essentially still) and calculating its frequency band energy distribution. If the background noise energy in the ultra-high frequency band (400-600Hz) is found to increase by more than 30% compared to the previous day, it is determined that there is an abnormal gain change in the radar front end or this module, and a complete unbalance compensation calibration is automatically triggered. If the energy of the low-frequency stretching component fluctuates abnormally during periods without temperature changes, the low-frequency sub-band truncation boundary of the wavelet packet decomposition is automatically adjusted (moderately widened from the default 0-37.5Hz to 0-50Hz or narrowed to 0-30Hz) to adapt to the differences in the dynamic characteristics of different bridge structures. The adjusted parameters are stored in non-volatile memory and loaded each time the power is turned on.

[0034] The cumulative slip calculation module receives two reconstructed signals from the time-frequency feature extraction module: a low-frequency stretching displacement component. (Sampling rate 1200Hz, components above 112.5Hz have been filtered out, mainly reflecting slow expansion and contraction caused by temperature and normal slippage of the support) and high-frequency vibration components. (Sampling rate 1200Hz, retaining the 112.5-600Hz frequency band, mainly including dynamic information such as friction and impact). Module output includes: absolute cumulative slip. Net cumulative slip Instantaneous slip rate High-frequency short-time energy sequence Friction abnormality marker position The system also records stuck events (start and end times, peak impact time, and duration). These outputs are stored in shared memory as a structure for periodic reading by the status assessment and early warning module (default reading interval is 1 second). The specific workflow of the cumulative slip calculation module is as follows: 1) The calculation process for the low-frequency component slip is as follows: 1.1) Direction determination logic, The radial displacement of the support relative to the radar is considered, with the positive direction defined as the support moving closer to the radar (corresponding to the sliding plate sliding in a certain direction, increasing the displacement). Since temperature-induced expansion and contraction have diurnal reciprocating characteristics, while net slippage caused by faults often exhibits unidirectional accumulation, it is necessary to accurately determine the sign of each displacement change. A velocity threshold is set. When instantaneous rate When the state is considered static, the slip is not accumulated. At that time, according to The sign determines the direction of accumulation: if Then the positive slip increment is included, if The negative slip increment is then included. To avoid frequent direction misjudgments caused by noise, a hysteresis logic is introduced: after a direction change, three consecutive sampling points must remain in the same direction before the direction flag is actually flipped. After flipping, it must remain flipped for at least 10ms (i.e., 12 sampling points) and is not allowed to flip again. This direction discrimination logic outputs two cumulative counters: a positive cumulative distance counter and a negative cumulative distance counter. and negative cumulative distance ; 1.2) Calculation of Absolute Cumulative Slip. Absolute cumulative slip is defined as the total sliding path length, disregarding direction, and characterizes the cumulative wear degree of the sliding plate material. The calculation formula is:

[0035] In the discrete implementation, the absolute value of the absolute displacement difference between adjacent sampling points is calculated every millisecond and accumulated to... :

[0036] when At that time, it is assumed that no actual sliding has occurred in the support, and therefore no accumulation is performed. This method can eliminate the interference of minute reciprocating vibrations caused by thermal expansion and contraction on wear statistics, making... More accurately reflects the effective wear history of the skateboard; 1.3) Calculation of net cumulative slip: Net cumulative slip is defined as the algebraic sum of forward and reverse slip, characterizing whether the support has experienced overall displacement.

[0037] Discrete implementation can directly use the initial value of the low-frequency displacement signal after it has been reset to zero. The net slip can be used to determine whether the support limit device has failed or whether the slide has slipped out of the design range. When A pre-alarm is triggered when the offset exceeds 80% of the design allowable offset (±30mm for this background bridge). 1.4) Slip rate calculation, instantaneous slip rate Depend on Obtained through central difference: ,in To reduce the impact of high-frequency noise on the speed, further... Perform a 10-point moving average to obtain a smoothed rate sequence. The sequence is averaged per second to form a second-level rate array, which is then stored in a rate history queue (storing data from the most recent 7 days) for trend analysis. 2) High-frequency component energy statistics and anomaly identification, the specific workflow of which is as follows: 2.1) Short-time energy calculation. For high-frequency components... According to fixed window length =1200 (corresponding to 1 second) is used for non-overlapping segmentation, and the short-time energy of each time window is calculated:

[0038] The unit is mm. 2 Simultaneously calculate the peak amplitude within this time window. . and A high-frequency feature sequence is constructed, with a cache length of 1440 points (corresponding to 24 hours), for comparison with historical baselines; 2.2) Friction Anomaly Identification: During normal bearing sliding, the friction between the PTFE sliding plate and the stainless steel mirror panel generates stable high-frequency background energy. When the sliding plate wears or lubrication fails, the friction coefficient increases, and the high-frequency energy rises significantly. This module uses a dynamic baseline threshold method: taking time windows from the same period within the past 24 hours (e.g., 2:00-4:00 AM, when the temperature is stable and the vehicle load is minimal). The median was used as the baseline for that day. And calculate its standard deviation. .when And the duration exceeded 10 seconds, and at this time If the support does not actively slide, it is determined to be an abnormal static friction, and the abnormal friction flag is set. If during the sliding process If the speed fluctuates drastically, it is determined to be an abnormal sliding friction, and a friction abnormality event record is generated (timestamp, abnormal energy value, corresponding slip rate). 2.3) Jamming Event Identification: Jamming refers to a situation where the support experiences resistance to movement at a certain position, followed by sudden release under external force accompanied by impact vibration. The identification logic consists of two steps: First, detecting abrupt changes in slip rate. This involves processing a smoothed rate sequence. Perform the difference test; when the absolute value of the average rate of change of adjacent seconds is greater than 5 mm / s 2 (i.e., the rate change exceeds 5 mm / s), and the rate jumps from near zero to greater than 2 mm / s, then it is initially determined to be a potential release event; The second step is to examine the high-frequency energy within one second before and after the event. If At the moment of release, it exceeded the historical baseline by more than 10 times, and The presence of a single spike (peak amplitude exceeding four times the normal fluctuation range) confirms a jamming event. The jamming event log includes: jamming start time (the point where the velocity drops below 0.05 mm / s), release time (the point where the velocity suddenly increases), jamming duration, release peak acceleration (estimated by the second difference of displacement), and impact energy. All jamming events are stored in a non-volatile memory event log, with a maximum of 2000 entries; exceeding this limit will overwrite the oldest record. 2.4) Slip Entropy and Irregularity Index: To quantify the degree of anomalous support slippage, this module additionally calculates a slip entropy index. This index is calculated by representing the absolute cumulative slippage. Divide the data into intervals of 10mm each, and calculate the percentage of time the slide falls into each interval over the past hour. Calculate Shannon entropy Normal support sliding exhibits a regular reciprocating motion with a low entropy value (typically 0.5-1.0); when random vibrations or unexpected sliding occur, the entropy value increases (>1.5). Entropy values ​​exceeding the threshold can help determine if the slide plate is worn or the limit switch is loose. Furthermore, the cumulative slip calculation module automatically performs zero-position calibration once a week: after confirming that the support has no movement (through polling within 10 seconds). Under the condition that the maximum value is <0.01mm / s and the temperature change is less than 0.5℃, the current net cumulative slip amount will be... Forced to zero, and at the same time The current value is used as the new zero point. This operation eliminates the long-term cumulative error of small DC drift in radar ranging systems, ensuring that the net slip always reflects the offset relative to the calibration zero point, rather than the absolute distance. Calibration events are logged, and remote manual calibration can be triggered.

[0039] The condition assessment and early warning module receives a comprehensive feature vector (including absolute cumulative slip, net cumulative slip, instantaneous slip rate, high-frequency vibration energy, slip entropy, friction anomaly flag, and jamming event flag) from the cumulative slip calculation module, and fuses it with the time-frequency feature matrix output by the time-frequency feature extraction module to achieve multimodal data fusion. This module has the dual tasks of real-time assessment of the current state and prediction of future trends: on the one hand, it classifies and assesses the current health status of the support based on the multidimensional feature fusion results; on the other hand, it constructs a time-series prediction model based on recurrent neural networks to predict the evolution trend of the support slip in advance, and issues an early warning signal when an anomaly occurs based on an adaptive classification threshold.

[0040] Specifically, the state assessment and early warning module incorporates a glide trend prediction network based on gated recurrent units (GRUs). As a simplified variant of long short-term memory networks, GRUs retain the ability to handle long-term dependencies while having fewer parameters and higher computational efficiency, making them more suitable for long-term operation on edge-side embedded processors. Compared to LSTM, GRU replaces the input, forget, and output gates of LSTM with a dual-gating mechanism of reset and update gates. GRU integrates the forget and input gates into a single update gate, merging the cell state and hidden state, thus reducing the parameter size by approximately 25% compared to LSTM. This meets the real-time and resource-constrained requirements of edge computing devices, making it a better choice for embedded platforms that need to run continuously on edge nodes for several years. Simultaneously, the update gate can directly pass historical hidden states to the current time step, effectively mitigating gradient decay and preserving long-term dependency information of support slippage trends. The input layer dimension of the prediction network corresponds to the dimension of the output feature vector of the cumulative slip calculation module (including eight normalized features: absolute cumulative slip, net cumulative slip, slip rate, high-frequency energy, and slip entropy). The hidden layer adopts a two-layer stacked bidirectional GRU architecture, with each layer containing 128 hidden units. The forward GRU processes the feature sequence in forward chronological order, while the backward GRU captures future contextual information in reverse chronological order. This bidirectional structure can more fully exploit the bidirectional temporal dependencies in the slip feature sequence, improving prediction accuracy. The forward hidden state sequence is denoted as The backward hidden state sequence is denoted as The final output is a concatenation of bidirectional hidden states. The hidden layer is followed by a fully connected layer with 64 neurons, which uses the ReLU activation function for nonlinear transformation. The final prediction result is output through the output layer. The model has approximately 400,000 parameters, which can meet the real-time and resource constraints of edge computing devices during embedded inference, making it easy to deploy on edge computing devices at bridge sites. To support the real-time inference requirements of the aforementioned gating trend prediction network based on gated recurrent units (GRUs), the state assessment and early warning module of this application is preferably deployed on an edge computing hardware platform with a certain computing power. In one specific embodiment, the module operates on an ARM Cortex-M7 or equivalent microcontroller with a clock speed of at least 800MHz, at least 512KB of tightly coupled memory (TCM) integrated on-chip, and support for a single-precision floating-point unit (FPU). In another preferred embodiment, a low-power application processor with an integrated neural network accelerator (NPU) can be used, with typical power consumption controlled below 2W. The selected hardware platform must be equipped with at least 4MB of off-chip flash memory for storing model parameters and at least 8MB of dynamic random access memory (DRAM) to cache input feature sequences and intermediate calculation results. By employing model compression techniques such as fixed-point quantization (e.g., INT8), the requirements for memory bandwidth and storage space can be further reduced, ensuring that the entire monitoring device can achieve long-term, stable, and low-power autonomous operation within an embedded chassis that meets IP67 protection standards.

[0041] Furthermore, the training process and dataset construction steps of the status assessment and early warning module are as follows: offline pre-training. When the system is first deployed, the training data comes from three sources: historical monitoring data accumulated by the bridge (if any), normal slip data samples of other bridges with similar span and bearing types, and accelerated degradation test data obtained by applying standard fatigue load cycles to the bearing of this type in a closed environment. After feature extraction from the above dataset, a batch of time-labeled feature vector sequences were obtained. Training samples were constructed using a sliding window method: the input window length was set to 24 hours (corresponding to 8640 time steps at a sampling period of 1Hz), and the prediction window length was set to 6 hours (2160 time steps, of which the focus was only on the trend of the sliding amount change in the initial 4 hours). The training and validation sets are divided in an 8:2 ratio. Mean squared error is used as the loss function, and the Adam optimizer is used for parameter updates. The initial learning rate is set to 0.001, and the learning rate is reduced to 0.8 times the original rate every 10 training epochs when the loss no longer decreases. The batch size is set to 64, and training lasts for a maximum of 200 epochs. An early stopping mechanism is used to terminate training if the loss does not improve after 20 consecutive validation epochs to control overfitting. The loss function is constructed by adding a trend consistency loss term to the standard mean squared error loss. Predicting the slip sequence. Compared with the true slip sequence If the trend directions (monotonically increasing, monotonically decreasing, or stationary) are consistent, the trend loss term is zero; otherwise, an additional penalty is applied based on the degree of inconsistency. After offline optimization, the weights of the two losses are set to... , This composite loss function ensures that the model not only focuses on the accurate fitting of numerical amplitudes, but also takes into account the directional judgment of the slip change trend, which is of substantial significance for the early warning of support wear trends; Furthermore, to adapt to the differences in characteristics under different bridge structures and operating environments, the module dynamically normalizes the features from the cumulative slip calculation module. Using historical data from the past 30 days as a reference window, the mean and standard deviation of each feature dimension are calculated, and the features are transformed to a standard normal distribution in real time. This normalization parameter is automatically updated every morning to reflect the natural drift and seasonal changes in the support state. Features that are extremely sensitive to numerical range, such as slip entropy and high-frequency vibration energy, are first logarithmically transformed and then normalized to compress the dynamic range and reduce the impact of outliers.

[0042] Furthermore, the working principle of the status assessment and early warning module of this application will be described below with reference to a specific embodiment. It should be emphasized that the specific values ​​mentioned below (such as 3000mm, 30mm, etc.) are merely exemplary thresholds calculated for the design parameters of specific bridge bearings and should not be considered as limitations on the scope of protection of this application. Those skilled in the art can adjust these thresholds accordingly based on the actual design parameters of different bearings. Combining the bearing design parameters at the engineering site with actual operation and maintenance experience, this module divides the bearing health status into four levels: (i) Normal state (green): The load meets all of the following conditions: the absolute cumulative slip is less than 50% of the design allowable wear (the design allowable wear depth of the PTFE sliding plate of this bridge bearing is 4mm, and the recommended value of absolute cumulative slip is approximately 3000mm after conversion based on the circumferential wear relationship of the sliding plate); the net slip is less than 30% of the design allowable offset (the design allowable offset is ±30mm, therefore within 9mm); the high-frequency vibration energy is less than the historical baseline plus 3 standard deviations; the slip entropy is less than 1.0; and no abnormal flags are triggered. This state indicates that the bearing is in a healthy service stage and only requires continuous monitoring without any intervention. (II) Warning Status (Yellow Alert): Triggered by any of the following conditions: Absolute cumulative slip reaches 50% to 75% of the design allowable wear (i.e., 1500-2250 mm); Net slip reaches 30% to 60% of the design allowable offset (i.e., 9-18 mm); High-frequency vibration energy exceeds the historical baseline plus 3 standard deviations but is below plus 5 standard deviations for 7 consecutive days; Slip entropy is between 1.0 and 1.5; or the GRU model predicts that the yellow alert threshold will be reached within the next 72 hours. This status indicates that the support is showing initial signs of degradation, and the frequency of manual inspections should be increased and the lubrication status should be rechecked. (III) Abnormal State (Orange Warning) is triggered when any of the following conditions occur: the absolute cumulative slip reaches or exceeds 75% of the design allowable wear amount, but does not exceed 100% (i.e., ≥2250mm and <3000mm); the net slip reaches 60% to 80% (18-24mm) of the design allowable offset; the high-frequency vibration energy exceeds the historical baseline plus 5 times the standard deviation for 3 consecutive days; the friction abnormality flag or the jamming event flag is set for more than 24 hours; or the GRU model predicts that the orange warning threshold will be reached within the next 48 hours. (iv) Severe state (red alert) is triggered when any of the following conditions occur: the absolute cumulative slip reaches or exceeds its corresponding design allowable wear amount (≥3000mm); the net slip exceeds 80% of the design allowable offset (>24mm); the high-frequency vibration energy exceeds the historical baseline plus 8 times the standard deviation; the jamming event repeats more than 3 times within 24 hours; the slip entropy is greater than 1.8; the cumulative slip rate continues to rise within 72 hours and the current rate exceeds 2mm / day; or it is predicted that the red alert threshold will be reached within the next 24 hours. The warning signal triggers a differentiated output strategy based on the current health status level: (i) Yellow warning output: Send a warning message (including warning level, triggering feature item, and current feature value) to the remote monitoring platform through the communication module; save the complete feature sequence of 72 hours before and after the warning time in the local data storage for subsequent traceability analysis; light up the yellow indicator light in the local LED status indicator system; the trigger frequency is limited to no more than once every 24 hours to avoid repeated status fluctuations in a short period of time that lead to reporting redundancy; (ii) Orange warning output: Based on the aforementioned yellow warning output, an MHTML format report containing trend charts of related features (including the cumulative slip curve, net slip curve, high-frequency energy curve, and rate curve for the past 7 days) is sent; an undeletable warning event record is generated in the local database, which is forcibly write-protected to prevent subsequent rotation overwriting or accidental deletion; at the same time, the warning level upgrade label is pushed to all monitoring subsystems of the entire bridge (such as the main beam expansion joint monitoring module, cable force monitoring module, etc.) to achieve cross-system linkage monitoring.

[0043] (III) Red warning output: Based on the aforementioned orange warning output, the wireless communication module is forcibly triggered to send an emergency notification to the designated maintenance person in the form of a text message or application push; the built-in DO output interface of the module is used to drive the external buzzer alarm to issue an audible and visual alarm; if the bridge site is equipped with a variable message sign, the red warning and bearing number can be displayed on the sign through the reserved RS485 interface to remind on-site personnel to stay away from the danger zone; at the same time, the high-frequency impact peak acceleration recorded in the jamming event is analyzed. If it exceeds the bearing design acceleration limit (0.5g can be taken as a reference for the general design standard of similar bearings), the warning message will be additionally labeled with "impact overload" to indicate that the bearing may have potential hidden structural damage.

[0044] Furthermore, during long-term monitoring, the support state may slowly drift due to environmental changes, skateboard wear, or maintenance interventions, making it difficult for the offline-trained static GRU model to maintain prediction accuracy. Therefore, this module introduces a self-updating mechanism based on incremental learning. The specific strategy is as follows: Strategy 1: Online fine-tuning triggered by low-confidence predictions. When the model's predicted 6-hour slip value is verified by subsequent actual measurements and found to have an error exceeding a set threshold (relative error greater than 30% and absolute error greater than 10mm) for at least three consecutive times, the model is determined to deviate from the actual operating conditions. The online fine-tuning procedure is triggered by feeding new feature samples collected over the next 7 days into the model in batches of 32 for small-batch gradient updates, with no more than 5 update rounds. Fine-tuning uses an update step size one order of magnitude lower than the initial learning rate (1e-4), updating only the weights of the last fully connected layer and the GRU output layer, without affecting the underlying feature extraction layer. This allows the model to adapt to changes in new operating conditions while retaining basic memory of normal slip patterns, preventing catastrophic forgetting. Strategy 2: Periodic incremental training. Every quarter, all newly accumulated monitoring data and the initial training set are combined for a complete incremental retraining. The parameters from the initial offline training are used as the starting point for retraining. The training process is then executed again on the complete dataset with a low initial learning rate (1e-4) until convergence. After retraining, the currently running online model is replaced with the updated model parameters. Strategy 3: Manually intervened forced calibration. After a bridge undergoes major maintenance (such as bearing replacement, slide plate replacement, or expansion joint overhaul), maintenance personnel can manually trigger a model reset via remote commands to clear accumulated historical feature data, restore the initial pre-trained model, and start incremental learning again with data under new operating conditions. All model versions and update records are stored in local storage, retaining version logs and corresponding timestamps, allowing remote operation and maintenance platforms to retrieve and compare the predictive performance of different versions as needed.

[0045] The data storage and communication module consists of three parts: a local non-volatile storage unit, a communication interface controller, and a protocol stack processor. The local non-volatile storage unit is responsible for storing raw displacement data, time-frequency characteristic parameters, cumulative slip sequence, high-frequency energy sequence, health status level, early warning event log, and model parameter files. The communication interface controller manages both wired (Ethernet, RS485) and wireless (4G / 5G, NB-IoT, Wi-Fi) transmission channels, while the protocol stack processor implements data packetization, compression, encryption, and breakpoint resumption functions. The module employs an independent power supply circuit (provided from the bridge distribution box via DC-DC isolation conversion to 5V / 2A) and automatically switches to a backup battery during mains power outages to ensure local storage is not lost during communication failures. The data storage and communication module is connected to the status assessment and early warning module, the cumulative slip calculation module, and the time-frequency feature extraction module via internal buses (SPI and SDIO), respectively. Whenever the status assessment and early warning module generates a new health status record or early warning event, it sends a storage request to this module via a message queue. This module writes the request to the database and triggers communication to send the data when needed. The raw displacement data blocks and feature matrices generated by the cumulative slip calculation module and the time-frequency feature extraction module are stored in the eMMC via a direct memory access channel, without consuming processor resources. All modules can receive parameter query and configuration commands from the remote platform through the remote procedure call interface provided by this module.

[0046] In one embodiment, a long-term monitoring method for cumulative slippage of bridge bearings based on microwave vibration measurement is provided, such as... Figure 2 As shown, the method includes the following steps: S1: Installation and signal acquisition of microwave vibration radar. The microwave vibration radar module is fixed on the rigid support on the side of the pier cap beam. The straight distance between the radar front end and the support slide plate is 8.5m. The horizontal angle between the main axis of the radar beam and the sliding direction of the support is controlled within 5° to reduce cosine error. A 0.5mm thick stainless steel triangular cone reflective target is glued to the center of the moving end of the support slide. The cone's base side length is 30mm, and the edges are chamfered to reduce the impact of wind-induced flutter. After the radar module is powered on, it first executes a self-test program: the transmit power is set to 12dBm, the sweep bandwidth to 3.5GHz, and the pulse repetition frequency to 1.2kHz. The built-in noise figure test function is used to confirm that the signal-to-noise ratio of the receiving channel is not less than 32dB. After passing the self-test, the radar continuously transmits frequency-modulated continuous wave signals with a sawtooth waveform and a sweep period of 256μs. After the receiving antenna captures the target's reflected signal, it passes through a low-noise amplifier, mixer, and quadrature demodulator to output I / Q dual-channel baseband analog signals. The baseband signal is transmitted to the data acquisition and preprocessing module via shielded twisted-pair cable. In this embodiment, when the ambient temperature drops sharply at night, the support contracts and shifts, and the radar echo phase changes linearly accordingly. The I / Q Lissajous figure shows an approximately circular trajectory, indicating that the antenna beam is well aligned and there is no obvious amplitude and phase imbalance distortion. S2: Displacement signal reconstruction and preprocessing. The data acquisition and preprocessing module performs 16-bit analog-to-digital conversion on the received I / Q baseband signal at a sampling rate of 2.4MHz. The sampling clock is generated by the frame synchronization pulse phase-locked loop frequency multiplication of the radar module to ensure strict synchronization of the I / Q channels. The discretized I / Q digital signals first enter the DC offset correction stage: 1024 sampling points are collected with the radar transmission channel closed, and the average value of the DC components of the I and Q paths is calculated separately. This DC value is subtracted in real time during subsequent formal measurements. Next, amplitude and phase imbalance compensation is performed—the least squares ellipse fitting method is used to fit the ellipse parameters from 2000 I / Q sample points collected in the static state, and the gain mismatch coefficient is found to be 1.03 and the orthogonal phase deviation is 2.1°. Then, a linear transformation is performed on the real-time data to correct the ellipse to the theoretical circle. Perform a four-quadrant arctangent operation on the corrected I / Q data to obtain the wrapped phase sequence. Its range is limited to Internally, because the maximum cumulative displacement of the support can reach tens of millimeters, which is much larger than the radar wavelength (the wavelength of the 76GHz download wave is about 3.95mm), phase jumps occur frequently; This embodiment uses the cumulative offset correction method to achieve phase unwrapping: setting a jump detection threshold. When the absolute value of the phase difference between adjacent sampling points is greater than If the difference is positive, it is considered a positive transition, and the subsequent phases are subtracted. If it is negative, the phase jumps in the opposite direction, and the subsequent phases are added together. Continuous phase after untangling Multiply by a coefficient Converted to instantaneous radial displacement Where λ is calculated to be 3.87 mm based on the current center frequency of the radar (77.5 GHz). Considering the low-frequency micro-vibrations of the bridge pier (measured main frequency of approximately 2.5 Hz), this embodiment utilizes the auxiliary receiving channel in the radar module—which is fixedly pointed to a stationary structural edge on the surface of the bridge pier—to synchronously acquire the reference displacement. Then, the differential displacement is obtained by subtracting the reference channel displacement from the support channel displacement. This eliminates common-mode vibration interference. Finally, for Median filtering (window length 5 points) is performed to remove transient outliers, followed by a fourth-order Butterworth low-pass filter with a cutoff frequency of 50Hz to retain the main frequency band of the support's mechanical motion, and the preprocessed real-time displacement sequence is output. The sampling rate was reduced to 1.2 kHz; S3: Time-frequency decomposition and feature extraction, using wavelet packet decomposition method for displacement sequences. Time-frequency analysis was performed. The db10 wavelet was selected as the mother wavelet, and the decomposition level was set to 4 levels, resulting in 16 frequency band subspaces. At a sampling rate of 1.2kHz, the Nyquist frequency was 600Hz, and the nominal bandwidth of each subband was 37.5Hz. Statistical analysis of extensive measured data revealed that the energy of the normal expansion and contraction displacement of the support caused by temperature changes is concentrated in the 0–5 Hz range. Therefore, all low-frequency coefficients of the first sub-band (0–37.5 Hz) were retained, and the coefficients of the remaining sub-bands were set to zero before performing an inverse transform to reconstruct the low-frequency displacement components. This component is mainly used for calculating cumulative slip. Meanwhile, the coefficients of sub-bands 4 to 16 (corresponding to 112.5–600 Hz) are retained, while the coefficients of sub-bands 1 to 3 are set to zero, and the high-frequency vibration components are reconstructed. Used for identifying friction anomalies and jamming events; To reduce boundary effects, each data block is extended by 200 sampling points at both the beginning and end (using symmetrical extension), and the extended portion is removed after reconstruction. For high-frequency components... Frame segmentation is performed, with a frame length of 1 second (1200 points), and frame shifts are non-overlapping. The short-time energy of each frame is calculated. Simultaneously, amplitude spectrum analysis was performed on the signal frame (Fast Fourier Transform with 2048 points, Hanning window), and the frequency corresponding to the peak value of the amplitude spectrum was extracted as the peak frequency. If a certain frame More than 6 times the background noise baseline and If the frequency is concentrated in the 150-300Hz range, it is preliminarily determined that there is an impact excitation event. S4: Cumulative slip calculation, using the low-frequency displacement component obtained from S3. As input, the instantaneous slip rate is first calculated. Discretization uses central difference with an interval of 1 / 1200 seconds. A direction discrimination threshold is set. When the support is considered stationary, slip accumulation is not considered; when When the positive sliding is included; when When the reverse slip is included; To avoid frequent direction jumps caused by noise, a hysteresis logic is introduced: the direction flag bit must be flipped for three consecutive sampling points in the same direction, and it cannot be flipped again for at least 10ms after the flip. An absolute cumulative slip is defined. Discretely accumulate the absolute value of the absolute displacement difference between adjacent sampling points every millisecond. Define the net cumulative slip. This refers to the real-time displacement value after the initial zeroing point. In this embodiment, zero-point drift calibration is automatically performed daily from 2:00 AM to 3:00 AM (when the bridge temperature is stable and traffic volume is minimal): if within 10 consecutive seconds... If the maximum value is less than 0.01 mm / s and the temperature change is less than 0.3℃, then the current... Forced to set to the new zero point, at the same time Zeroing out. This mechanism eliminates the long-term cumulative error of minute DC drift in radar ranging systems; In addition, to characterize the randomness of slip behavior, slip entropy is calculated: the absolute cumulative slip over the past hour is divided into intervals of 10 mm, and the percentage of time the slip displacement falls into each interval is statistically analyzed. Calculate Shannon entropy The normal bearing H value is usually between 0.5 and 1.0. When the slide plate is worn or irregular vibration occurs, H will rise to more than 1.5. S5: Abnormal slip and jamming identification, utilizing the high-frequency short-time energy obtained from S3. and the net slip obtained in step 4 Jointly determine abnormal conditions. First, establish a dynamic baseline: based on the time period from 2:00 AM to 4:00 AM each day over the past 24 hours (when temperatures are lowest and vehicle load is negligible). The median serves as the baseline for the background energy of the day. And calculate its standard deviation. If a certain frame And at this time, the instantaneous slip rate of the support If the support is basically stationary, it is determined to be an abnormal static friction, and the abnormal friction flag is set. If during the movement of the support exceeding 10 consecutive seconds If so, it is determined to be an abnormal sliding friction. A two-order method is used for jamming event identification: first, a sudden change in slip rate is detected—the absolute value of the average rate of change of slip rate over an adjacent second is greater than 5 mm / s. 2 The velocity jumps from less than 0.05 mm / s to greater than 2 mm / s; then the impact energy is verified—within 1 second before and after the velocity abrupt change, the instantaneous peak value of the high-frequency energy exceeds... More than 10 times, and It exhibits a single peak shape (peak amplitude exceeds 4 times the normal fluctuation range). A jamming-release event is confirmed when both conditions are met simultaneously. Record the jamming initiation time (the moment the velocity first drops below 0.05 mm / s), the release time (the moment the velocity abruptly rises), the jamming duration, and the release peak acceleration (estimated by the second difference of displacement). For net slip, when If the change is monotonically increased (or decreased) over 72 consecutive hours and the cumulative change exceeds 5 mm, while the temperature change during the period is less than 5℃ (indicating that the influence of temperature is excluded), then it is determined that the skateboard limiter is in failure or the overall offset trend is abnormal. S6: Condition Assessment and Early Warning. A gated recurrent unit (GRU) is used to construct a slip trend prediction model. The model structure is a two-layer bidirectional GRU with 128 hidden units in each layer, followed by a fully connected layer to output the predicted value. The input feature vector includes the following eight normalized indicators: absolute cumulative slip increment, net cumulative slip, average slip rate, average high-frequency energy, slip entropy, friction anomaly indicator, jamming event count (within the past 24 hours), and bearing temperature (obtained from the bridge health monitoring system). The input window length is 24 hours (corresponding to one feature point per second for a total of 8640 steps), and the prediction window is 6 hours, focusing on outputting the net slip and absolute cumulative slip for the next 4 hours. The model training adopts an offline pre-training + online incremental update strategy: the pre-training data comes from the historical monitoring data of this bridge over the past 2 years and fatigue test data of similar bearings; during online operation, whenever the prediction error (the difference between the predicted net slip in 6 hours and the measured value) exceeds 3mm for 3 consecutive times, a small batch fine-tuning is triggered (using the data of the most recent 7 days, learning rate 1e-4, 5 rounds of updates). In one embodiment, the output health status can be divided into four levels: green (normal), yellow (caution), orange (abnormal), and red (severe). The red warning can be triggered when the absolute cumulative slip reaches or exceeds its corresponding design allowable wear level. An orange alert can be triggered when the absolute cumulative slip reaches or exceeds 75% but not 100% of the design allowable wear. A yellow alert can be triggered when the absolute cumulative slip reaches between 50% and 75% of the design allowable wear. The specific threshold can be set according to actual engineering requirements. When the status reaches yellow or above, the communication module issues a warning signal (yellow: local LED indicator, platform message push; orange: additional trend report; red: additional SMS notification to the maintenance person in charge, and triggering a buzzer alarm). Simultaneously, from the moment of the abnormal status, the module automatically saves the original displacement data, characteristic parameters, and model input snapshots from 72 hours prior to the current time to 24 hours prior to the current time to a protected area in the local memory (which cannot be overwritten) for subsequent professional detection and analysis.

[0047] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

Claims

1. A long-term monitoring device for accumulated slip of bridge support based on microwave vibration measurement, characterized in that, include: The microwave vibration radar module is used to transmit and receive frequency-modulated continuous wave signals, and outputs I / Q baseband signals after orthogonal mixing; The data acquisition and preprocessing module is connected to the microwave vibration radar module and is used to perform analog-to-digital conversion and preprocessing on the I / Q baseband signal to reconstruct the original displacement signal sequence. The time-frequency feature extraction module, connected to the data acquisition and preprocessing module, is used to perform time-frequency decomposition on the original displacement signal sequence to separate the low-frequency displacement component reflecting the normal expansion and contraction of the support and the high-frequency vibration component reflecting dynamic events. The cumulative slip calculation module, connected to the time-frequency feature extraction module, is used to calculate the absolute cumulative slip and net cumulative slip based on the low-frequency displacement component, and to identify friction anomalies and jamming events based on the high-frequency vibration component. The condition assessment and early warning module, connected to the cumulative slip calculation module, is used to assess the health status of the support and selectively issue graded early warning signals based at least on the identification results of the absolute cumulative slip, net cumulative slip, friction anomalies, and jamming events.

2. The long-term monitoring device for cumulative slippage of bridge bearings based on microwave vibration measurement according to claim 1, characterized in that, The data acquisition and preprocessing module is specifically used to: perform analog-to-digital conversion on the I / Q baseband signal, and then sequentially perform DC offset correction, amplitude-phase imbalance compensation, phase unwrapping, and displacement conversion to generate the original displacement signal sequence; and the data acquisition and preprocessing module is also used to acquire the reference channel displacement pointing to the fixed pier wall, and subtract the reference channel displacement from the support channel displacement to eliminate common-mode vibration interference.

3. The long-term monitoring device for cumulative slippage of bridge bearings based on microwave vibration measurement according to claim 1, characterized in that, The time-frequency feature extraction module has a built-in wavelet packet decomposition unit and a short-time Fourier transform unit, and adaptively selects the decomposition method according to the motion state of the original displacement signal sequence: the wavelet packet decomposition unit is used by default to decompose the original displacement signal sequence into multiple frequency band subspaces, and the low-frequency displacement component and the high-frequency vibration component are obtained by setting the coefficients of the specified sub-bands to zero and reconstructing them. When an intermittent strong impact is detected in the displacement signal, the system temporarily switches to the short-time Fourier transform unit to output high-frequency transient features in the form of a time-spectrum matrix and impact feature flags.

4. The long-term monitoring device for cumulative slippage of bridge bearings based on microwave vibration measurement according to claim 1, characterized in that, The cumulative slip calculation module includes: a direction determination unit, which determines the slip direction based on the comparison result of the instantaneous rate of the low-frequency displacement component and the velocity threshold, combined with hysteresis logic; The absolute cumulative slip calculation unit accumulates the absolute displacement difference between adjacent sampling points when it is determined that the support has actually slipped, and outputs the total slip path length. The net cumulative slip calculation unit calculates the algebraic sum of the low-frequency displacement components relative to the initial moment and outputs the overall offset of the support. A short-time energy calculation unit calculates the short-time energy of the high-frequency vibration component; The anomaly identification unit, based on the dynamic baseline threshold method, determines static friction anomalies, sliding friction anomalies, and jamming-release events according to the short-term energy and the instantaneous rate.

5. The long-term monitoring device for cumulative slippage of bridge bearings based on microwave vibration measurement according to claim 1, characterized in that, The state assessment and early warning module incorporates a slip trend prediction network based on a gated recurrent unit. The input feature vector of the slip trend prediction network includes at least the absolute cumulative slip, net cumulative slip, average slip rate, high-frequency energy, and slip entropy. The state assessment and early warning module is used to predict the amount of slip in the future time period based on the slip trend prediction network output, and trigger the corresponding level of early warning signal in advance by combining the preset level threshold.

6. The long-term monitoring device for cumulative slippage of bridge bearings based on microwave vibration measurement according to claim 5, characterized in that, The state assessment and early warning module is also used to execute an incremental learning self-update mechanism: when the prediction error continuously exceeds the set threshold, an online fine-tuning program is triggered to update the final layer weights of the gating recurrent unit's slip trend prediction network with an update step size lower than the initial learning rate; and it is used to merge all newly accumulated monitoring data with the initial training set every quarter to perform a complete incremental retraining.

7. The long-term monitoring device for cumulative slippage of bridge bearings based on microwave vibration measurement according to claim 1, characterized in that, It also includes a data storage and communication module; The data storage and communication module is connected to the cumulative slip calculation module and the status assessment and early warning module respectively. It is used to store the original displacement data, characteristic parameters, health status records and early warning event logs with different priorities, and send differentiated early warning messages to the remote monitoring platform by wired or wireless communication according to the early warning level.

8. A long-term monitoring method for cumulative slippage of bridge bearings based on microwave vibration measurement, characterized in that, Includes the following steps: S1: I / Q baseband signals are acquired by a microwave vibration radar module installed on the side of the bridge pier; S2: Perform analog-to-digital conversion, DC offset correction, imbalance compensation, phase unwrapping and displacement conversion on the I / Q baseband signal, and use the reference channel to eliminate common-mode vibration interference to reconstruct the original displacement signal sequence; S3: Perform time-frequency decomposition on the original displacement signal sequence, using wavelet packet decomposition or short-time Fourier transform to separate the low-frequency displacement component reflecting the normal expansion and contraction of the support and the high-frequency vibration component reflecting dynamic events. S4: Based on the low-frequency displacement components, calculate the absolute cumulative slip and net cumulative slip through direction discrimination and threshold logic, and calculate the slip entropy; S5: Calculate short-time energy based on the high-frequency vibration components, and use the dynamic baseline threshold method to identify friction anomalies and jam-release events; S6: Based at least on the identification results of the absolute cumulative slip, net cumulative slip, friction anomaly and jamming event, combined with the predicted value output by the slip trend prediction network based on the gated loop unit, assess the health status of the support and issue a graded early warning signal according to the graded threshold.

9. The long-term monitoring method for cumulative slippage of bridge bearings based on microwave vibration measurement according to claim 8, characterized in that, The phase unwrapping in S2 employs a cumulative offset correction method: setting a jump detection threshold. When the absolute value of the phase difference between adjacent sampling points is greater than At that time, the subsequent phase is increased or decreased as a whole according to the direction of the difference. In the displacement conversion, the instantaneous radial displacement is calculated by multiplying the unwrapped continuous phase by a coefficient. Received, among which This is the radar carrier wavelength.

Images